From e84b71c2c6da6e69c8f815168ea836f9716a325e Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 23 May 2024 10:00:21 +0300 Subject: [PATCH] ggml : drop support for QK_K=64 (#7473) * ggml : drop support for QK_K=64 ggml-ci * opencl : restore QK_K=256 define --- CMakeLists.txt | 5 - Makefile | 4 - ci/run.sh | 3 +- ggml-common.h | 54 - ggml-cuda/convert.cu | 138 -- ggml-cuda/dmmv.cu | 151 --- ggml-cuda/mmq.cu | 6 - ggml-cuda/vecdotq.cuh | 126 -- ggml-metal.m | 17 - ggml-metal.metal | 398 +----- ggml-opencl.cpp | 2 +- ggml-quants.c | 2672 +------------------------------------ ggml-sycl.cpp | 472 +------ ggml.c | 12 - gguf-py/gguf/constants.py | 3 +- llama.cpp | 12 +- 16 files changed, 26 insertions(+), 4049 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index c09d834fb..ef02ff669 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -124,7 +124,6 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING set(LLAMA_METAL_STD "" CACHE STRING "llama: metal standard version (-std flag)") option(LLAMA_KOMPUTE "llama: use Kompute" OFF) option(LLAMA_RPC "llama: use RPC" OFF) -option(LLAMA_QKK_64 "llama: use super-block size of 64 for k-quants" OFF) option(LLAMA_SYCL "llama: use SYCL" OFF) option(LLAMA_SYCL_F16 "llama: use 16 bit floats for sycl calculations" OFF) set(LLAMA_SYCL_TARGET "INTEL" CACHE STRING "llama: sycl target device") @@ -384,10 +383,6 @@ if (LLAMA_LLAMAFILE) set(GGML_SOURCES_LLAMAFILE sgemm.cpp) endif() -if (LLAMA_QKK_64) - add_compile_definitions(GGML_QKK_64) -endif() - if (LLAMA_CUBLAS) message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead") set(LLAMA_CUDA ON) diff --git a/Makefile b/Makefile index 6b7c853b3..fe63cbd60 100644 --- a/Makefile +++ b/Makefile @@ -389,10 +389,6 @@ else MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d endif -ifdef LLAMA_QKK_64 - MK_CPPFLAGS += -DGGML_QKK_64 -endif - ifndef LLAMA_NO_ACCELERATE # Mac OS - include Accelerate framework. # `-framework Accelerate` works both with Apple Silicon and Mac Intel diff --git a/ci/run.sh b/ci/run.sh index d5972480b..79dcd0772 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -606,7 +606,8 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then if [ -z ${GG_BUILD_CUDA} ]; then - test $ret -eq 0 && gg_run open_llama_3b_v2 + #test $ret -eq 0 && gg_run open_llama_3b_v2 + date # dummy else test $ret -eq 0 && gg_run open_llama_7b_v2 fi diff --git a/ggml-common.h b/ggml-common.h index 43c7978a0..77e6bfba4 100644 --- a/ggml-common.h +++ b/ggml-common.h @@ -65,13 +65,8 @@ typedef sycl::half2 ggml_half2; // QK = number of values after dequantization // QK_K = super-block size -#ifdef GGML_QKK_64 -#define QK_K 64 -#define K_SCALE_SIZE 4 -#else #define QK_K 256 #define K_SCALE_SIZE 12 -#endif // GGML_QKK_64 #if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL) // QR = QK / number of values before dequantization @@ -131,13 +126,8 @@ typedef sycl::half2 ggml_half2; #define QI4_NL (QK4_NL / (4*QR4_NL)) #define QR4_NL 2 -#if QK_K == 64 -#define QI4_XS QI4_NL -#define QR4_XS QR4_NL -#else #define QI4_XS (QK_K / (4*QR4_XS)) #define QR4_XS 8 -#endif #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP @@ -228,15 +218,6 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wro // weight is represented as x = a * q // 16 blocks of 16 elements each // Effectively 3.4375 bits per weight -#ifdef GGML_QKK_64 -typedef struct { - uint8_t hmask[QK_K/8]; // quants - high bit - uint8_t qs[QK_K/4]; // quants - low 2 bits - uint8_t scales[2]; - ggml_half d; // super-block scale -} block_q3_K; -static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding"); -#else typedef struct { uint8_t hmask[QK_K/8]; // quants - high bit uint8_t qs[QK_K/4]; // quants - low 2 bits @@ -244,20 +225,11 @@ typedef struct { ggml_half d; // super-block scale } block_q3_K; static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding"); -#endif // 4-bit quantization // 8 blocks of 32 elements each // weight is represented as x = a * q + b // Effectively 4.5 bits per weight -#ifdef GGML_QKK_64 -typedef struct { - ggml_half d[2]; // super-block scales/mins - uint8_t scales[2]; // 4-bit block scales/mins - uint8_t qs[QK_K/2]; // 4--bit quants -} block_q4_K; -static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding"); -#else typedef struct { union { struct { @@ -270,21 +242,11 @@ typedef struct { uint8_t qs[QK_K/2]; // 4--bit quants } block_q4_K; static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding"); -#endif // 5-bit quantization // 8 blocks of 32 elements each // weight is represented as x = a * q + b // Effectively 5.5 bits per weight -#ifdef GGML_QKK_64 -typedef struct { - ggml_half d; // super-block scale - int8_t scales[QK_K/16]; // 8-bit block scales - uint8_t qh[QK_K/8]; // quants, high bit - uint8_t qs[QK_K/2]; // quants, low 4 bits -} block_q5_K; -static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding"); -#else typedef struct { union { struct { @@ -298,7 +260,6 @@ typedef struct { uint8_t qs[QK_K/2]; // quants, low 4 bits } block_q5_K; static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding"); -#endif // 6-bit quantization // weight is represented as x = a * q @@ -356,11 +317,7 @@ typedef struct { static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding"); // 3.4375 bpw -#if QK_K == 64 -#define IQ3S_N_SCALE 2 -#else #define IQ3S_N_SCALE QK_K/64 -#endif typedef struct { ggml_half d; uint8_t qs[QK_K/4]; @@ -381,16 +338,9 @@ static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wron typedef struct { uint8_t qs[QK_K/8]; // grid index, low 8 bits uint8_t qh[QK_K/16]; // grid index, high 3 bits + grid shift bit (for two groups of 8) -#if QK_K == 64 - ggml_half d; -#endif uint8_t scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64) } block_iq1_m; -#if QK_K == 64 -static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding"); -#else static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding"); -#endif // Used by IQ1_M quants typedef union { @@ -406,9 +356,6 @@ typedef struct { } block_iq4_nl; static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding"); -#if QK_K == 64 -#define block_iq4_xs block_iq4_nl -#else typedef struct { ggml_half d; uint16_t scales_h; @@ -416,7 +363,6 @@ typedef struct { uint8_t qs[QK_K/2]; } block_iq4_xs; static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding"); -#endif #endif // GGML_COMMON_DECL #endif // GGML_COMMON_DECL diff --git a/ggml-cuda/convert.cu b/ggml-cuda/convert.cu index 830e2d756..c0a444707 100644 --- a/ggml-cuda/convert.cu +++ b/ggml-cuda/convert.cu @@ -131,7 +131,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t const block_q2_K * x = (const block_q2_K *) vx; const int64_t tid = threadIdx.x; -#if QK_K == 256 const int64_t n = tid/32; const int64_t l = tid - 32*n; const int64_t is = 8*n + l/16; @@ -145,17 +144,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4); y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4); y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4); -#else - const int64_t is = tid/16; // 0 or 1 - const int64_t il = tid%16; // 0...15 - const uint8_t q = x[i].qs[il] >> (2*is); - dst_t * y = yy + i*QK_K + 16*is + il; - float dall = __low2half(x[i].dm); - float dmin = __high2half(x[i].dm); - y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); - y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4); -#endif - } template @@ -164,7 +152,6 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t const int64_t i = blockIdx.x; const block_q3_K * x = (const block_q3_K *) vx; -#if QK_K == 256 const int64_t r = threadIdx.x/4; const int64_t tid = r/2; const int64_t is0 = r%2; @@ -188,31 +175,8 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t const uint8_t * hm = x[i].hmask; for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)); -#else - const int64_t tid = threadIdx.x; - const int64_t is = tid/16; // 0 or 1 - const int64_t il = tid%16; // 0...15 - const int64_t im = il/8; // 0...1 - const int64_t in = il%8; // 0...7 - - dst_t * y = yy + i*QK_K + 16*is + il; - - const uint8_t q = x[i].qs[il] >> (2*is); - const uint8_t h = x[i].hmask[in] >> (2*is + im); - const float d = (float)x[i].d; - - if (is == 0) { - y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); - y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); - } else { - y[ 0] = d * ((x[i].scales[0] >> 4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); - y[32] = d * ((x[i].scales[1] >> 4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); - } -#endif - } -#if QK_K == 256 static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { if (j < 4) { d = q[j] & 63; m = q[j + 4] & 63; @@ -221,7 +185,6 @@ static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); } } -#endif template static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { @@ -229,7 +192,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t const int64_t i = blockIdx.x; -#if QK_K == 256 // assume 32 threads const int64_t tid = threadIdx.x; const int64_t il = tid/8; @@ -253,15 +215,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t y[l + 0] = d1 * (q[l] & 0xF) - m1; y[l +32] = d2 * (q[l] >> 4) - m2; } -#else - const int64_t tid = threadIdx.x; - const uint8_t * q = x[i].qs; - dst_t * y = yy + i*QK_K; - const float d = (float)x[i].dm[0]; - const float m = (float)x[i].dm[1]; - y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4); - y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4); -#endif } template @@ -270,7 +223,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t const int64_t i = blockIdx.x; -#if QK_K == 256 // assume 64 threads - this is very slightly better than the one below const int64_t tid = threadIdx.x; const int64_t il = tid/16; // il is in 0...3 @@ -297,18 +249,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t hm <<= 1; y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2; y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2; -#else - const int64_t tid = threadIdx.x; - const uint8_t q = x[i].qs[tid]; - const int64_t im = tid/8; // 0...3 - const int64_t in = tid%8; // 0...7 - const int64_t is = tid/16; // 0 or 1 - const uint8_t h = x[i].qh[in] >> im; - const float d = x[i].d; - dst_t * y = yy + i*QK_K + tid; - y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16)); - y[32] = d * x[i].scales[is+2] * ((q >> 4) - ((h >> 4) & 1 ? 0 : 16)); -#endif } template @@ -316,7 +256,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t const block_q6_K * x = (const block_q6_K *) vx; const int64_t i = blockIdx.x; -#if QK_K == 256 // assume 64 threads - this is very slightly better than the one below const int64_t tid = threadIdx.x; @@ -336,24 +275,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32); y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32); y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32); -#else - - // assume 32 threads - const int64_t tid = threadIdx.x; - const int64_t ip = tid/16; // 0 or 1 - const int64_t il = tid - 16*ip; // 0...15 - - dst_t * y = yy + i*QK_K + 16*ip + il; - - const float d = x[i].d; - - const uint8_t ql = x[i].ql[16*ip + il]; - const uint8_t qh = x[i].qh[il] >> (2*ip); - const int8_t * sc = x[i].scales; - - y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32); - y[32] = d * sc[ip+2] * ((int8_t)((ql >> 4) | (((qh >> 4) & 3) << 4)) - 32); -#endif } template @@ -363,7 +284,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds const block_iq2_xxs * x = (const block_iq2_xxs *) vx; const int64_t tid = threadIdx.x; -#if QK_K == 256 const int64_t il = tid/8; // 0...3 const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -374,10 +294,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f; const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127]; for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); -#else - NO_DEVICE_CODE; -#endif - } template @@ -387,7 +303,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst const block_iq2_xs * x = (const block_iq2_xs *) vx; const int64_t tid = threadIdx.x; -#if QK_K == 256 const int64_t il = tid/8; // 0...3 const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -396,10 +311,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; const uint8_t signs = ksigns_iq2xs[q2[il] >> 9]; for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); -#else - NO_DEVICE_CODE; -#endif - } template @@ -409,7 +320,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_ const block_iq2_s * x = (const block_iq2_s *) vx; const int64_t tid = threadIdx.x; -#if QK_K == 256 const int64_t il = tid/8; // 0...3 const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -417,10 +327,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_ const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; const uint8_t signs = x[i].qs[QK_K/8+4*ib+il]; for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); -#else - NO_DEVICE_CODE; -#endif - } template @@ -430,7 +336,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds const block_iq3_xxs * x = (const block_iq3_xxs *) vx; const int64_t tid = threadIdx.x; -#if QK_K == 256 const int64_t il = tid/8; // 0...3 const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -445,10 +350,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); } -#else - NO_DEVICE_CODE; -#endif - } template @@ -458,7 +359,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_ const block_iq3_s * x = (const block_iq3_s *) vx; const int64_t tid = threadIdx.x; -#if QK_K == 256 const int64_t il = tid/8; // 0...3 const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -471,10 +371,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_ y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); } -#else - NO_DEVICE_CODE; -#endif - } template @@ -484,7 +380,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_ const block_iq1_s * x = (const block_iq1_s *) vx; const int64_t tid = threadIdx.x; -#if QK_K == 256 const int64_t il = tid/8; // 0...3 const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -497,10 +392,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_ for (int j = 0; j < 8; ++j) { y[j] = d * (q[j] + delta); } -#else - NO_DEVICE_CODE; -#endif - } template @@ -510,7 +401,6 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_ const block_iq1_m * x = (const block_iq1_m *) vx; const int64_t tid = threadIdx.x; -#if QK_K == 256 const int64_t il = tid/8; // 0...3 const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -527,13 +417,8 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_ for (int j = 0; j < 8; ++j) { y[j] = d * (q[j] + delta); } -#else - NO_DEVICE_CODE; -#endif - } - template static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) { @@ -550,10 +435,8 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf]; y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; } - } -#if QK_K != 64 template static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) { const int64_t i = blockIdx.x; @@ -570,7 +453,6 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; } } -#endif template static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) { @@ -592,21 +474,13 @@ static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * template static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; -#if QK_K == 256 dequantize_block_q2_K<<>>(vx, y); -#else - dequantize_block_q2_K<<>>(vx, y); -#endif } template static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; -#if QK_K == 256 dequantize_block_q3_K<<>>(vx, y); -#else - dequantize_block_q3_K<<>>(vx, y); -#endif } template @@ -632,21 +506,13 @@ static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k template static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; -#if QK_K == 256 dequantize_block_q5_K<<>>(vx, y); -#else - dequantize_block_q5_K<<>>(vx, y); -#endif } template static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = k / QK_K; -#if QK_K == 256 dequantize_block_q6_K<<>>(vx, y); -#else - dequantize_block_q6_K<<>>(vx, y); -#endif } template @@ -700,11 +566,7 @@ static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t template static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { const int nb = (k + QK_K - 1) / QK_K; -#if QK_K == 64 - dequantize_block_iq4_nl<<>>(vx, y); -#else dequantize_block_iq4_xs<<>>(vx, y); -#endif } template diff --git a/ggml-cuda/dmmv.cu b/ggml-cuda/dmmv.cu index 7313e3e17..47d4d5d9e 100644 --- a/ggml-cuda/dmmv.cu +++ b/ggml-cuda/dmmv.cu @@ -22,7 +22,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, float tmp = 0; // partial sum for thread in warp -#if QK_K == 256 const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...15 const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1 @@ -71,37 +70,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, tmp += dall * sum1 - dmin * sum2; } -#else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 - const int offset = tid * K_QUANTS_PER_ITERATION; - - uint32_t uaux[2]; - const uint8_t * d = (const uint8_t *)uaux; - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + offset; - const uint8_t * q = x[i].qs + offset; - const uint32_t * s = (const uint32_t *)x[i].scales; - - uaux[0] = s[0] & 0x0f0f0f0f; - uaux[1] = (s[0] >> 4) & 0x0f0f0f0f; - - const float2 dall = __half22float2(x[i].dm); - - float sum1 = 0, sum2 = 0; - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - const uint8_t ql = q[l]; - sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3) - + y[l+16] * d[1] * ((ql >> 2) & 3) - + y[l+32] * d[2] * ((ql >> 4) & 3) - + y[l+48] * d[3] * ((ql >> 6) & 3); - sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7]; - } - tmp += dall.x * sum1 - dall.y * sum2; - } -#endif // sum up partial sums and write back result tmp = warp_reduce_sum(tmp); @@ -123,8 +91,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, float tmp = 0; // partial sum for thread in warp -#if QK_K == 256 - const uint16_t kmask1 = 0x0303; const uint16_t kmask2 = 0x0f0f; @@ -175,34 +141,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, tmp += d * sum; } -#else - - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 - const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14 - const int in = offset/8; // 0 or 1 - const int im = offset%8; // 0...7 - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + offset; - const uint8_t * q = x[i].qs + offset; - const uint8_t * s = x[i].scales; - - const float dall = (float)x[i].d; - - float sum = 0; - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - const uint8_t hl = x[i].hmask[im+l] >> in; - const uint8_t ql = q[l]; - sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4)) - + y[l+16] * dall * ((s[0] >> 4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4)) - + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4)) - + y[l+48] * dall * ((s[1] >> 4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4)); - } - tmp += sum; - } -#endif // sum up partial sums and write back result tmp = warp_reduce_sum(tmp); @@ -221,7 +159,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const block_q4_K * x = (const block_q4_K *)vx + ib0; -#if QK_K == 256 const uint16_t kmask1 = 0x3f3f; const uint16_t kmask2 = 0x0f0f; const uint16_t kmask3 = 0xc0c0; @@ -306,36 +243,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, #endif } -#else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); - - const int step = tid * K_QUANTS_PER_ITERATION; - - uint16_t aux16[2]; - const uint8_t * s = (const uint8_t *)aux16; - - float tmp = 0; - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - const uint8_t * q = x[i].qs + step; - const float * y = yy + i*QK_K + step; - const uint16_t * a = (const uint16_t *)x[i].scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - const float d = (float)x[i].dm[0]; - const float m = (float)x[i].dm[1]; - float sum = 0.f; - for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { - sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2]) - + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2]) - + y[j+32] * (d * s[1] * (q[j+ 0] >> 4) - m * s[3]) - + y[j+48] * (d * s[1] * (q[j+16] >> 4) - m * s[3]); - } - tmp += sum; - } - -#endif // sum up partial sums and write back result tmp = warp_reduce_sum(tmp); @@ -355,7 +262,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, float tmp = 0; // partial sum for thread in warp -#if QK_K == 256 const uint16_t kmask1 = 0x3f3f; const uint16_t kmask2 = 0x0f0f; const uint16_t kmask3 = 0xc0c0; @@ -426,30 +332,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin; } -#else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); - const int step = tid * K_QUANTS_PER_ITERATION; - const int im = step/8; - const int in = step%8; - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - const uint8_t * q = x[i].qs + step; - const int8_t * s = x[i].scales; - const float * y = yy + i*QK_K + step; - const float d = x[i].d; - float sum = 0.f; - for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { - const uint8_t h = x[i].qh[in+j] >> im; - sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16)) - + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16)) - + y[j+32] * d * s[2] * ((q[j+ 0] >> 4) - ((h >> 4) & 1 ? 0 : 16)) - + y[j+48] * d * s[3] * ((q[j+16] >> 4) - ((h >> 6) & 1 ? 0 : 16)); - } - tmp += sum; - } -#endif - // sum up partial sums and write back result tmp = warp_reduce_sum(tmp); @@ -470,8 +352,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const block_q6_K * x = (const block_q6_K *)vx + ib0; -#if QK_K == 256 - const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1 @@ -526,37 +406,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, } -#else - - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...7 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0...3 - - const int step = tid * K_QUANTS_PER_ITERATION; - - float tmp = 0; // partial sum for thread in warp - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + step; - const uint8_t * ql = x[i].ql + step; - const uint8_t * qh = x[i].qh + step; - const int8_t * s = x[i].scales; - - const float d = x[i+0].d; - - float sum = 0; - for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { - sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32) - + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32) - + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >> 4) | ((qh[j] & 0x30) >> 0)) - 32) - + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >> 4) | ((qh[j] & 0xc0) >> 2)) - 32); - } - tmp += sum; - - } - -#endif - // sum up partial sums and write back result tmp = warp_reduce_sum(tmp); diff --git a/ggml-cuda/mmq.cu b/ggml-cuda/mmq.cu index 933d799ce..c0a66d9b6 100644 --- a/ggml-cuda/mmq.cu +++ b/ggml-cuda/mmq.cu @@ -826,11 +826,7 @@ template static __device__ __forceinlin const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd; -#if QK_K == 256 x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm; -#else - x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]}; -#endif } #pragma unroll @@ -933,9 +929,7 @@ template static __device__ __forceinlin const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd; -#if QK_K == 256 x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm; -#endif } #pragma unroll diff --git a/ggml-cuda/vecdotq.cuh b/ggml-cuda/vecdotq.cuh index 86b87fa93..5ebdddcc7 100644 --- a/ggml-cuda/vecdotq.cuh +++ b/ggml-cuda/vecdotq.cuh @@ -712,7 +712,6 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1( static __device__ __forceinline__ float vec_dot_q4_K_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#ifndef GGML_QKK_64 const block_q4_K * bq4_K = (const block_q4_K *) vbq; int v[2]; @@ -754,58 +753,11 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1( } return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8); - -#else - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_q4_K * bq4_K = (const block_q4_K *) vbq; - - float sumf_d = 0.0f; - float sumf_m = 0.0f; - - uint16_t aux16[2]; - const uint8_t * s = (const uint8_t *)aux16; - - const uint16_t * a = (const uint16_t *)bq4_K->scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - - const float dall = bq4_K->dm[0]; - const float dmin = bq4_K->dm[1]; - - const float d8_1 = __low2float(bq8_1[0].ds); - const float d8_2 = __low2float(bq8_1[1].ds); - - const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); - const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); - const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); - const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); - - const int * q4 = (const int *)bq4_K->qs + (iqs/2); - const int v1 = q4[0]; - const int v2 = q4[4]; - - const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0)); - const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0)); - const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0)); - const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0)); - - sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]); - sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]); - - return dall * sumf_d - dmin * sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A - -#endif } static __device__ __forceinline__ float vec_dot_q5_K_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#ifndef GGML_QKK_64 const block_q5_K * bq5_K = (const block_q5_K *) vbq; int vl[2]; @@ -847,48 +799,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1( } return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8); - -#else - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_q5_K * bq5_K = (const block_q5_K *) vbq; - - const int8_t * s = bq5_K->scales; - - const float d = bq5_K->d; - - const float d8_1 = __low2half(bq8_1[0].ds); - const float d8_2 = __low2half(bq8_1[1].ds); - - const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); - const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); - const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); - const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); - - const int * ql = (const int *)bq5_K->qs + (iqs/2); - const int vl1 = ql[0]; - const int vl2 = ql[4]; - - const int step = 4 * (iqs/2); // 0, 4, 8, 12 - const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6 - const int in = step%8; // 0, 4, 0, 4 - const int vh = (*((const int *)(bq5_K->qh + in))) >> im; - - const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f); - const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f); - const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f); - const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f); - - const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1]) - + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]); - - return d * sumf_d; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A - -#endif } static __device__ __forceinline__ float vec_dot_q6_K_q8_1( @@ -919,7 +829,6 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1( static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if QK_K == 256 const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq; #if QR2_XXS == 8 @@ -960,15 +869,11 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( } return d * (sumi1 + sumi2); #endif -#else - NO_DEVICE_CODE; -#endif } static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq; const int ib32 = iqs; @@ -1002,17 +907,12 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( GGML_UNUSED(ksigns64); NO_DEVICE_CODE; #endif -#else - GGML_UNUSED(ksigns64); - NO_DEVICE_CODE; -#endif } // TODO static __device__ __forceinline__ float vec_dot_iq2_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 const block_iq2_s * bq2 = (const block_iq2_s *) vbq; const int ib32 = iqs; @@ -1048,16 +948,11 @@ static __device__ __forceinline__ float vec_dot_iq2_s_q8_1( GGML_UNUSED(ksigns64); NO_DEVICE_CODE; #endif -#else - GGML_UNUSED(ksigns64); - NO_DEVICE_CODE; -#endif } static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq; const int ib32 = iqs; @@ -1082,16 +977,12 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1( #else NO_DEVICE_CODE; #endif -#else - NO_DEVICE_CODE; -#endif } // TODO: don't use lookup table for signs static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 const block_iq3_s * bq2 = (const block_iq3_s *) vbq; const int ib32 = iqs; @@ -1114,14 +1005,10 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( #else NO_DEVICE_CODE; #endif -#else - NO_DEVICE_CODE; -#endif } static __device__ __forceinline__ float vec_dot_iq1_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if QK_K == 256 const block_iq1_s * bq1 = (const block_iq1_s *) vbq; const int ib32 = iqs; @@ -1149,14 +1036,10 @@ static __device__ __forceinline__ float vec_dot_iq1_s_q8_1( const float d = d1q * __low2float (bq8_1[ib32].ds); const float m = d1q * __high2float(bq8_1[ib32].ds); return d * sumi + m * delta; -#else - NO_DEVICE_CODE; -#endif } static __device__ __forceinline__ float vec_dot_iq1_m_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if QK_K == 256 const block_iq1_m * bq1 = (const block_iq1_m *) vbq; const int ib32 = iqs; @@ -1192,9 +1075,6 @@ static __device__ __forceinline__ float vec_dot_iq1_m_q8_1( scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds); return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1)); -#else - NO_DEVICE_CODE; -#endif } #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics @@ -1250,9 +1130,7 @@ static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1( static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if QK_K == 256 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq; const uint8_t * values = (const uint8_t *)kvalues_iq4nl; @@ -1270,10 +1148,6 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1( sumi2 = __dp4a(v2, q8[j+4], sumi2); } return d * (sumi1 + sumi2); - -#else - NO_DEVICE_CODE; -#endif #else return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs); #endif diff --git a/ggml-metal.m b/ggml-metal.m index 5d5ad20ad..c9e570dbf 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -381,10 +381,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { // dictionary of preprocessor macros NSMutableDictionary * prep = [NSMutableDictionary dictionary]; -#ifdef GGML_QKK_64 - prep[@"GGML_QKK_64"] = @(1); -#endif - MTLCompileOptions* options = [MTLCompileOptions new]; options.preprocessorMacros = prep; @@ -1773,11 +1769,7 @@ static enum ggml_status ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else if (src0t == GGML_TYPE_Q3_K) { -#ifdef GGML_QKK_64 - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; -#else [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; -#endif } else if (src0t == GGML_TYPE_Q5_K) { [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; @@ -2018,12 +2010,7 @@ static enum ggml_status ggml_metal_graph_compute( { nth0 = 4; nth1 = 16; - #if QK_K == 64 - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline; - #else pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline; - #endif - } break; default: { @@ -2088,11 +2075,7 @@ static enum ggml_status ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else if (src0t == GGML_TYPE_Q3_K) { -#ifdef GGML_QKK_64 - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; -#else [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; -#endif } else if (src0t == GGML_TYPE_Q5_K) { [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; diff --git a/ggml-metal.metal b/ggml-metal.metal index c5eb25280..8ff70d7a7 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -3386,7 +3386,6 @@ void kernel_mul_mv_q2_K_f32_impl( const int step = sizeof(block_q2_K) * nb; -#if QK_K == 256 const int ix = tiisg/8; // 0...3 const int it = tiisg%8; // 0...7 const int iq = it/4; // 0 or 1 @@ -3438,57 +3437,6 @@ void kernel_mul_mv_q2_K_f32_impl( y4 += 4 * QK_K; } -#else - const int ix = tiisg/2; // 0...15 - const int it = tiisg%2; // 0...1 - - device const float * y4 = y + ix * QK_K + 8 * it; - - for (int ib = ix; ib < nb; ib += 16) { - - float4 sumy = {0.f, 0.f, 0.f, 0.f}; - for (int i = 0; i < 8; ++i) { - yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0]; - yl[i+ 8] = y4[i+16]; sumy[1] += yl[i+ 8]; - yl[i+16] = y4[i+32]; sumy[2] += yl[i+16]; - yl[i+24] = y4[i+48]; sumy[3] += yl[i+24]; - } - - device const uint8_t * sc = (device const uint8_t *)x[ib].scales; - device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it; - device const half * dh = &x[ib].d; - - for (int row = 0; row < N_DST; row++) { - - float4 acc1 = {0.f, 0.f, 0.f, 0.f}; - float4 acc2 = {0.f, 0.f, 0.f, 0.f}; - for (int i = 0; i < 8; i += 2) { - acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003); - acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300); - acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c); - acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00); - acc1[2] += yl[i+16] * (qs[i/2] & 0x0030); - acc2[2] += yl[i+17] * (qs[i/2] & 0x3000); - acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0); - acc2[3] += yl[i+25] * (qs[i/2] & 0xc000); - } - - float dall = dh[0]; - float dmin = dh[1]; - sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f + - (acc1[1] + 1.f/256.f * acc2[1]) * (sc[1] & 0xF) * 1.f/ 4.f + - (acc1[2] + 1.f/256.f * acc2[2]) * (sc[2] & 0xF) * 1.f/16.f + - (acc1[3] + 1.f/256.f * acc2[3]) * (sc[3] & 0xF) * 1.f/64.f) - - dmin * (sumy[0] * (sc[0] >> 4) + sumy[1] * (sc[1] >> 4) + sumy[2] * (sc[2] >> 4) + sumy[3] * (sc[3] >> 4)); - - qs += step/2; - sc += step; - dh += step/2; - } - - y4 += 16 * QK_K; - } -#endif for (int row = 0; row < N_DST; ++row) { all_sum = simd_sum(sumf[row]); @@ -3526,7 +3474,6 @@ kernel void kernel_mul_mv_q2_K_f32( kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg); } -#if QK_K == 256 void kernel_mul_mv_q3_K_f32_impl( device const void * src0, device const float * src1, @@ -3685,84 +3632,6 @@ void kernel_mul_mv_q3_K_f32_impl( } } } -#else -void kernel_mul_mv_q3_K_f32_impl( - device const void * src0, - device const float * src1, - device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne10, - constant int64_t & ne12, - constant int64_t & ne0, - constant int64_t & ne1, - constant uint & r2, - constant uint & r3, - threadgroup int8_t * shared_values [[threadgroup(0)]], - uint3 tgpig[[threadgroup_position_in_grid]], - uint tiisg[[thread_index_in_simdgroup]], - uint sgitg[[simdgroup_index_in_threadgroup]]) { - - const int nb = ne00/QK_K; - - const int64_t r0 = tgpig.x; - const int64_t r1 = tgpig.y; - const int64_t im = tgpig.z; - - const int row = 2 * r0 + sgitg; - - const uint i12 = im%ne12; - const uint i13 = im/ne12; - - const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); - - device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb + offset0; - device const float * yy = (device const float *) src1 + r1*ne10 + im*ne00*ne1; - - const int ix = tiisg/4; - const int il = 4 * (tiisg%4);// 0, 4, 8, 12 - const int iq = il/8; // 0, 0, 1, 1 - const int in = il%8; // 0, 4, 0, 4 - - float2 sum = {0.f, 0.f}; - - for (int i = ix; i < nb; i += 8) { - - const float d_all = (float)(x[i].d); - - device const uint16_t * q = (device const uint16_t *)(x[i].qs + il); - device const uint16_t * h = (device const uint16_t *)(x[i].hmask + in); - device const uint16_t * s = (device const uint16_t *)(x[i].scales); - device const float * y = yy + i * QK_K + il; - - const float d1 = d_all * ((int32_t)(s[0] & 0x000F) - 8); - const float d2 = d_all * ((int32_t)(s[0] & 0x00F0) - 128) * 1.f/64.f; - const float d3 = d_all * ((int32_t)(s[0] & 0x0F00) - 2048) * 1.f/4096.f; - const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f; - - for (int l = 0; l < 4; l += 2) { - const uint16_t hm = h[l/2] >> iq; - sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 : 4)) - + y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16)) - + y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64)) - + y[l+48] * d4 * ((int32_t)(q[l/2] & 0x00c0) - ((hm & 0x0040) ? 0 : 256)); - sum[1] += y[l+ 1] * d1 * ((int32_t)(q[l/2] & 0x0300) - ((hm & 0x0100) ? 0 : 1024)) - + y[l+17] * d2 * ((int32_t)(q[l/2] & 0x0c00) - ((hm & 0x0400) ? 0 : 4096)) - + y[l+33] * d3 * ((int32_t)(q[l/2] & 0x3000) - ((hm & 0x1000) ? 0 : 16384)) - + y[l+49] * d4 * ((int32_t)(q[l/2] & 0xc000) - ((hm & 0x4000) ? 0 : 65536)); - } - - } - const float sumf = sum[0] + sum[1] * 1.f/256.f; - - const float tot = simd_sum(sumf); - if (tiisg == 0) { - dst[r1*ne0 + im*ne0*ne1 + row] = tot; - } - -} -#endif [[host_name("kernel_mul_mv_q3_K_f32")]] kernel void kernel_mul_mv_q3_K_f32( @@ -3792,7 +3661,6 @@ kernel void kernel_mul_mv_q3_K_f32( kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg); } -#if QK_K == 256 void kernel_mul_mv_q4_K_f32_impl( device const void * src0, device const float * src1, @@ -3906,103 +3774,6 @@ void kernel_mul_mv_q4_K_f32_impl( } } } -#else -void kernel_mul_mv_q4_K_f32_impl( - device const void * src0, - device const float * src1, - device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne10, - constant int64_t & ne12, - constant int64_t & ne0, - constant int64_t & ne1, - constant uint & r2, - constant uint & r3, - threadgroup int8_t * shared_values [[threadgroup(0)]], - uint3 tgpig[[threadgroup_position_in_grid]], - uint tiisg[[thread_index_in_simdgroup]], - uint sgitg[[simdgroup_index_in_threadgroup]]) { - - const int ix = tiisg/4; // 0...7 - const int it = tiisg%4; // 0...3 - - const int nb = ne00/QK_K; - const int r0 = tgpig.x; - const int r1 = tgpig.y; - const int im = tgpig.z; - const int first_row = r0 * N_DST; - const int ib_row = first_row * nb; - - const uint i12 = im%ne12; - const uint i13 = im/ne12; - - const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); - - device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0; - device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1; - - float yl[8]; - float yh[8]; - float sumf[N_DST]={0.f}, all_sum; - - const int step = sizeof(block_q4_K) * nb / 2; - - device const float * y4 = y + ix * QK_K + 8 * it; - - uint16_t sc16[4]; - - for (int ib = ix; ib < nb; ib += 8) { - - float2 sumy = {0.f, 0.f}; - for (int i = 0; i < 8; ++i) { - yl[i] = y4[i+ 0]; sumy[0] += yl[i]; - yh[i] = y4[i+32]; sumy[1] += yh[i]; - } - - device const uint16_t * sc = (device const uint16_t *)x[ib].scales; - device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it; - device const half * dh = x[ib].d; - - for (int row = 0; row < N_DST; row++) { - - sc16[0] = sc[0] & 0x000f; - sc16[1] = sc[0] & 0x0f00; - sc16[2] = sc[0] & 0x00f0; - sc16[3] = sc[0] & 0xf000; - - float2 acc1 = {0.f, 0.f}; - float2 acc2 = {0.f, 0.f}; - for (int i = 0; i < 8; i += 2) { - acc1[0] += yl[i+0] * (qs[i/2] & 0x000F); - acc1[1] += yl[i+1] * (qs[i/2] & 0x0F00); - acc2[0] += yh[i+0] * (qs[i/2] & 0x00F0); - acc2[1] += yh[i+1] * (qs[i/2] & 0xF000); - } - - float dall = dh[0]; - float dmin = dh[1]; - sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc16[0] + - (acc2[0] + 1.f/256.f * acc2[1]) * sc16[1] * 1.f/4096.f) - - dmin * 1.f/16.f * (sumy[0] * sc16[2] + sumy[1] * sc16[3] * 1.f/256.f); - - qs += step; - sc += step; - dh += step; - } - - y4 += 8 * QK_K; - } - - for (int row = 0; row < N_DST; ++row) { - all_sum = simd_sum(sumf[row]); - if (tiisg == 0) { - dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum; - } - } -} -#endif [[host_name("kernel_mul_mv_q4_K_f32")]] kernel void kernel_mul_mv_q4_K_f32( @@ -4070,8 +3841,6 @@ void kernel_mul_mv_q5_K_f32_impl( const int step = sizeof(block_q5_K) * nb; -#if QK_K == 256 -# float yl[16], yh[16]; const uint16_t kmask1 = 0x3f3f; @@ -4154,54 +3923,6 @@ void kernel_mul_mv_q5_K_f32_impl( y1 += 4 * QK_K; } -#else - float yl[8], yh[8]; - - const int il = 4 * (tiisg/8); // 0, 4, 8, 12 - const int ix = tiisg%8; - const int iq = il/8; // 0, 0, 1, 1 - const int in = il%8; // 0, 4, 0, 4 - - device const float * y = yy + ix*QK_K + il; - - for (int i = ix; i < nb; i += 8) { - - for (int l = 0; l < 4; ++l) { - yl[l+0] = y[l+ 0]; - yl[l+4] = y[l+16]; - yh[l+0] = y[l+32]; - yh[l+4] = y[l+48]; - } - - device const half * dh = &x[i].d; - device const uint8_t * q = x[i].qs + il; - device const uint8_t * h = x[i].qh + in; - device const int8_t * s = x[i].scales; - - for (int row = 0; row < 2; ++row) { - - const float d = dh[0]; - - float2 acc = {0.f, 0.f}; - for (int l = 0; l < 4; ++l) { - const uint8_t hl = h[l] >> iq; - acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16)) - + yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16)); - acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256)) - + yh[l+4] * s[3] * ((int16_t)(q[l+16] & 0xF0) - (hl & 0x40 ? 0 : 256)); - } - sumf[row] += d * (acc[0] + 1.f/16.f * acc[1]); - - q += step; - h += step; - s += step; - dh += step/2; - - } - - y += 8 * QK_K; - } -#endif for (int row = 0; row < 2; ++row) { const float tot = simd_sum(sumf[row]); @@ -4280,7 +4001,6 @@ void kernel_mul_mv_q6_K_f32_impl( float sumf = 0; -#if QK_K == 256 const int tid = tiisg/2; const int ix = tiisg%2; const int ip = tid/8; // 0 or 1 @@ -4316,30 +4036,6 @@ void kernel_mul_mv_q6_K_f32_impl( } -#else - const int ix = tiisg/4; - const int il = 4*(tiisg%4); - - for (int i = ix; i < nb; i += 8) { - device const float * y = yy + i * QK_K + il; - device const uint8_t * ql = x[i].ql + il; - device const uint8_t * qh = x[i].qh + il; - device const int8_t * s = x[i].scales; - - const float d = x[i].d; - - float4 sums = {0.f, 0.f, 0.f, 0.f}; - for (int l = 0; l < 4; ++l) { - sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32); - sums[1] += y[l+16] * ((int8_t)((ql[l+16] & 0xF) | ((qh[l] & kmask2) << 2)) - 32); - sums[2] += y[l+32] * ((int8_t)((ql[l+ 0] >> 4) | ((qh[l] & kmask3) >> 0)) - 32); - sums[3] += y[l+48] * ((int8_t)((ql[l+16] >> 4) | ((qh[l] & kmask4) >> 2)) - 32); - } - sumf += d * (sums[0] * s[0] + sums[1] * s[1] + sums[2] * s[2] + sums[3] * s[3]); - } - -#endif - const float tot = simd_sum(sumf); if (tiisg == 0) { dst[r1*ne0 + im*ne0*ne1 + row] = tot; @@ -5173,9 +4869,7 @@ void kernel_mul_mv_iq1_m_f32_impl( device const float * y4 = y + 32 * ix; -#if QK_K != 64 iq1m_scale_t scale; -#endif for (int ib32 = ix; ib32 < nb32; ib32 += 32) { @@ -5196,10 +4890,7 @@ void kernel_mul_mv_iq1_m_f32_impl( device const uint16_t * sc = (device const uint16_t *)xr->scales; for (int row = 0; row < N_DST; row++) { - -#if QK_K != 64 scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); -#endif constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700))); constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700))); @@ -5215,14 +4906,9 @@ void kernel_mul_mv_iq1_m_f32_impl( } const float delta1 = sumy[0] * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[1] * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA); const float delta2 = sumy[2] * (qh[1] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[3] * (qh[1] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA); -#if QK_K == 64 - const float d = (float) *((device const half *)(sc - 1)); - sumf[row] += d * ((sum[0] + delta1) * (2*((sc[0] >> (8*(ib%2)+0)) & 0xf) + 1) + - (sum[1] + delta2) * (2*((sc[0] >> (8*(ib%2)+4)) & 0xf) + 1)); -#else + sumf[row] += (float)scale.f16 * ((sum[0] + delta1) * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 7) + 1) + (sum[1] + delta2) * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 7) + 1)); -#endif sc += nb*sizeof(block_iq1_m)/2; qs += nb*sizeof(block_iq1_m); @@ -5334,7 +5020,6 @@ void kernel_mul_mv_iq4_nl_f32_impl( } } -#if QK_K != 64 void kernel_mul_mv_iq4_xs_f32_impl( device const void * src0, device const float * src1, @@ -5429,7 +5114,6 @@ void kernel_mul_mv_iq4_xs_f32_impl( } } } -#endif [[host_name("kernel_mul_mv_iq1_s_f32")]] kernel void kernel_mul_mv_iq1_s_f32( @@ -5542,11 +5226,7 @@ kernel void kernel_mul_mv_iq4_xs_f32( uint tiisg[[thread_index_in_simdgroup]], uint sgitg[[simdgroup_index_in_threadgroup]]) { -#if QK_K == 64 - kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg); -#else kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg); -#endif } //============================= templates and their specializations ============================= @@ -5672,10 +5352,9 @@ void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg float dl, ml; uint8_t sc = xb->scales[il]; -#if QK_K == 256 q = q + 32*(il/8) + 16*(il&1); il = (il/2)%4; -#endif + half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h); uchar mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4); @@ -5691,7 +5370,6 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg device const uint8_t * h = (device const uint8_t *)xb->hmask; device const int8_t * scales = (device const int8_t *)xb->scales; -#if QK_K == 256 q = q + 32 * (il/8) + 16 * (il&1); h = h + 16 * (il&1); uint8_t m = 1 << (il/2); @@ -5712,17 +5390,6 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg for (int i = 0; i < 16; ++i) { reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml); } -#else - float kcoef = il&1 ? 1.f/16.f : 1.f; - uint16_t kmask = il&1 ? 0xF0 : 0x0F; - float dl = d_all * ((scales[il/2] & kmask) * kcoef - 8); - float coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h); - uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); - uint8_t m = 1<<(il*2); - for (int i = 0; i < 16; ++i) { - reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i%8] & (m * (1 + i/8))) ? 0 : 4.f/coef)); - } -#endif } static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) { @@ -5734,7 +5401,6 @@ template void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) { device const uchar * q = xb->qs; -#if QK_K == 256 short is = (il/4) * 2; q = q + (il/4) * 32 + 16 * (il&1); il = il & 3; @@ -5743,16 +5409,7 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg const float min = xb->dmin; const float dl = d * sc[0]; const float ml = min * sc[1]; -#else - (void) get_scale_min_k4_just2; - q = q + 16 * (il&1); - device const uint8_t * s = xb->scales; - device const half2 * dh = (device const half2 *)xb->d; - const float2 d = (float2)dh[0]; - const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h; - const float ml = il<2 ? d[1] * (s[0]>>4) : d[1] * (s[1]>>4); -#endif const ushort mask = il<2 ? 0x0F : 0xF0; for (int i = 0; i < 16; ++i) { reg[i/4][i%4] = dl * (q[i] & mask) - ml; @@ -5764,7 +5421,6 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg device const uint8_t * q = xb->qs; device const uint8_t * qh = xb->qh; -#if QK_K == 256 short is = (il/4) * 2; q = q + 32 * (il/4) + 16 * (il&1); qh = qh + 16 * (il&1); @@ -5781,17 +5437,6 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg for (int i = 0; i < 16; ++i) { reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml; } -#else - q = q + 16 * (il&1); - device const int8_t * s = xb->scales; - const float dl = xb->d * s[il]; - uint8_t m = 1<<(il*2); - const float coef = il<2 ? 1.f : 1.f/16.f; - const ushort mask = il<2 ? 0x0F : 0xF0; - for (int i = 0; i < 16; ++i) { - reg[i/4][i%4] = coef * dl * ((q[i] & mask) - (qh[i%8] & (m*(1+i/8)) ? 0.f : 16.f/coef)); - } -#endif } template @@ -5801,15 +5446,11 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg device const uint8_t * qh = (device const uint8_t *)xb->qh; device const int8_t * scales = (device const int8_t *)xb->scales; -#if QK_K == 256 ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1); qh = qh + 32*(il/8) + 16*(il&1); float sc = scales[(il%2) + 2 * ((il/2))]; il = (il/2) & 3; -#else - ql = ql + 16 * (il&1); - float sc = scales[il]; -#endif + const uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); const uint16_t kmask2 = il>1 ? 0xF0 : 0x0F; const float coef = il>1 ? 1.f/16.f : 1.f; @@ -5966,20 +5607,15 @@ void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & const int ib32 = il/2; il = il%2; device const uint16_t * sc = (device const uint16_t *)xb->scales; -#if QK_K == 64 - const float d = xb->d; -#else + iq1m_scale_t scale; scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); const float d = scale.f16; -#endif + device const uint8_t * qs = xb->qs + 4*ib32 + 2*il; device const uint8_t * qh = xb->qh + 2*ib32 + il; -#if QK_K == 64 - const float dl = d * (2*((sc[ib32/2] >> (8*(ib32%2)+4*il)) & 0xf) + 1); -#else + const float dl = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1); -#endif const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA); const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA); constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700))); @@ -6009,9 +5645,6 @@ void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 template void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) { -#if QK_K == 64 - dequantize_iq4_nl(xb, il, reg); -#else // il is 0...15 for QK_K = 256 => index of block of 32 is il/2 const int ib32 = il/2; il = il%2; @@ -6028,7 +5661,6 @@ void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 reg[i][2] = d * kvalues_iq4nl_f[q8[2]]; reg[i][3] = d * kvalues_iq4nl_f[q8[3]]; } -#endif } template @@ -6533,11 +6165,7 @@ kernel void kernel_mul_mm_id( sgitg); } -#if QK_K == 256 #define QK_NL 16 -#else -#define QK_NL 4 -#endif // // get rows @@ -6577,11 +6205,7 @@ template [[host_name("kernel_get_rows_iq2_s")]] kernel get_rows_t kernel_get_r template [[host_name("kernel_get_rows_iq1_s")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_iq1_m")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_iq4_nl")]] kernel get_rows_t kernel_get_rows; -#if QK_K == 64 -template [[host_name("kernel_get_rows_iq4_xs")]] kernel get_rows_t kernel_get_rows; -#else template [[host_name("kernel_get_rows_iq4_xs")]] kernel get_rows_t kernel_get_rows; -#endif // // matrix-matrix multiplication @@ -6609,11 +6233,7 @@ template [[host_name("kernel_mul_mm_iq2_s_f32")]] kernel mat_mm_t kernel_mul_m template [[host_name("kernel_mul_mm_iq1_s_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_iq1_m_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_iq4_nl_f32")]] kernel mat_mm_t kernel_mul_mm; -#if QK_K == 64 -template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mat_mm_t kernel_mul_mm; -#else template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mat_mm_t kernel_mul_mm; -#endif // // indirect matrix-matrix multiplication @@ -6641,11 +6261,7 @@ template [[host_name("kernel_mul_mm_id_iq2_s_f32")]] kernel mat_mm_id_t kernel template [[host_name("kernel_mul_mm_id_iq1_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_iq1_m_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -#if QK_K == 64 -template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -#else template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -#endif // // matrix-vector multiplication @@ -6854,7 +6470,5 @@ template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t template [[host_name("kernel_mul_mv_id_iq3_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; template [[host_name("kernel_mul_mv_id_iq2_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -#if QK_K != 64 template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -#endif diff --git a/ggml-opencl.cpp b/ggml-opencl.cpp index 922f24837..e28566a7b 100644 --- a/ggml-opencl.cpp +++ b/ggml-opencl.cpp @@ -1,4 +1,4 @@ -#include "ggml.h" +#include "ggml.h" #include "ggml-opencl.h" #include "ggml-backend-impl.h" diff --git a/ggml-quants.c b/ggml-quants.c index ed40ca74a..88f58a339 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -1888,7 +1888,6 @@ static float make_qkx2_quants(int n, int nmax, const float * restrict x, const f return scale; } -#if QK_K == 256 static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) { if (j < 4) { *d = q[j] & 63; *m = q[j + 4] & 63; @@ -1897,7 +1896,6 @@ static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * *m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); } } -#endif //========================- 2-bit (de)-quantization @@ -1961,20 +1959,13 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict } } -#if QK_K == 256 for (int j = 0; j < QK_K; j += 128) { for (int l = 0; l < 32; ++l) { y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6); } } -#else - for (int l = 0; l < 16; ++l) { - y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6); - } -#endif x += QK_K; - } } @@ -1989,7 +1980,6 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int6 const uint8_t * q = x[i].qs; -#if QK_K == 256 int is = 0; float dl, ml; for (int n = 0; n < QK_K; n += 128) { @@ -2008,19 +1998,6 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int6 } q += 32; } -#else - float dl1 = d * (x[i].scales[0] & 0xF), ml1 = min * (x[i].scales[0] >> 4); - float dl2 = d * (x[i].scales[1] & 0xF), ml2 = min * (x[i].scales[1] >> 4); - float dl3 = d * (x[i].scales[2] & 0xF), ml3 = min * (x[i].scales[2] >> 4); - float dl4 = d * (x[i].scales[3] & 0xF), ml4 = min * (x[i].scales[3] >> 4); - for (int l = 0; l < 16; ++l) { - y[l+ 0] = dl1 * ((int8_t)((q[l] >> 0) & 3)) - ml1; - y[l+16] = dl2 * ((int8_t)((q[l] >> 2) & 3)) - ml2; - y[l+32] = dl3 * ((int8_t)((q[l] >> 4) & 3)) - ml3; - y[l+48] = dl4 * ((int8_t)((q[l] >> 6) & 3)) - ml4; - } - y += QK_K; -#endif } } @@ -2211,36 +2188,9 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri } float dm, mm; -#if QK_K == 64 - float max_scale = 0, max_min = 0; - for (int j = 0; j < QK_K/16; ++j) { - max_scale = MAX(max_scale, scales[j]); - max_min = MAX(max_min, mins[j]); - } - dm = max_scale/15; - mm = max_min/15; - if (max_scale) { - float id = 1/dm; - for (int j = 0; j < QK_K/16; ++j) { - int l = nearest_int(id*scales[j]); - Ls[j] = MAX(0, MIN(15, l)); - } - } else { - memset(Ls, 0, QK_K/16); - } - if (max_min) { - float id = 1/mm; - for (int j = 0; j < QK_K/16; ++j) { - int l = nearest_int(id*mins[j]); - Lm[j] = MAX(0, MIN(15, l)); - } - } else { - memset(Lm, 0, QK_K/16); - } -#else dm = make_qp_quants(QK_K/16, 15, scales, Ls, sw); mm = make_qp_quants(QK_K/16, 15, mins, Lm, sw); -#endif + y[i].d = GGML_FP32_TO_FP16(dm); y[i].dmin = GGML_FP32_TO_FP16(mm); dm = GGML_FP16_TO_FP32(y[i].d); @@ -2263,20 +2213,13 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri } } -#if QK_K == 256 for (int j = 0; j < QK_K; j += 128) { for (int l = 0; l < 32; ++l) { y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6); } } -#else - for (int l = 0; l < 16; ++l) { - y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6); - } -#endif x += QK_K; - } } @@ -2317,7 +2260,6 @@ void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict } } -#if QK_K == 256 memset(y[i].scales, 0, 12); if (max_scale) { float iscale = -32.f/max_scale; @@ -2351,36 +2293,6 @@ void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict L[16*j + ii] = l + 4; } } -#else - if (max_scale) { - float iscale = -8.f/max_scale; - for (int j = 0; j < QK_K/16; j+=2) { - int l1 = nearest_int(iscale*scales[j]); - l1 = 8 + MAX(-8, MIN(7, l1)); - int l2 = nearest_int(iscale*scales[j+1]); - l2 = 8 + MAX(-8, MIN(7, l2)); - y[i].scales[j/2] = l1 | (l2 << 4); - } - y[i].d = GGML_FP32_TO_FP16(1/iscale); - } else { - for (int j = 0; j < QK_K/16; j+=2) { - y[i].scales[j/2] = 0; - } - y[i].d = GGML_FP32_TO_FP16(0.f); - } - for (int j = 0; j < QK_K/16; ++j) { - int s = j%2 == 0 ? y[i].scales[j/2] & 0xF : y[i].scales[j/2] >> 4; - float d = GGML_FP16_TO_FP32(y[i].d) * (s - 8); - if (!d) { - continue; - } - for (int ii = 0; ii < 16; ++ii) { - int l = nearest_int(x[16*j + ii]/d); - l = MAX(-4, MIN(3, l)); - L[16*j + ii] = l + 4; - } - } -#endif memset(y[i].hmask, 0, QK_K/8); // We put the high-bit for the 1st 8 quants into bit 0, the next 8 into bit 1, etc. @@ -2395,23 +2307,16 @@ void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict m = 0; hm <<= 1; } } -#if QK_K == 256 for (int j = 0; j < QK_K; j += 128) { for (int l = 0; l < 32; ++l) { y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6); } } -#else - for (int l = 0; l < 16; ++l) { - y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6); - } -#endif x += QK_K; } } -#if QK_K == 256 void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -2461,49 +2366,12 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int6 } } -#else -void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int64_t k) { - assert(k % QK_K == 0); - assert(QK_K == 64); - const int nb = k / QK_K; - - for (int i = 0; i < nb; i++) { - - const float d_all = GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict q = x[i].qs; - const uint8_t * restrict hm = x[i].hmask; - - const float d1 = d_all * ((x[i].scales[0] & 0xF) - 8); - const float d2 = d_all * ((x[i].scales[0] >> 4) - 8); - const float d3 = d_all * ((x[i].scales[1] & 0xF) - 8); - const float d4 = d_all * ((x[i].scales[1] >> 4) - 8); - - for (int l=0; l<8; ++l) { - uint8_t h = hm[l]; - y[l+ 0] = d1 * ((int8_t)((q[l+0] >> 0) & 3) - ((h & 0x01) ? 0 : 4)); - y[l+ 8] = d1 * ((int8_t)((q[l+8] >> 0) & 3) - ((h & 0x02) ? 0 : 4)); - y[l+16] = d2 * ((int8_t)((q[l+0] >> 2) & 3) - ((h & 0x04) ? 0 : 4)); - y[l+24] = d2 * ((int8_t)((q[l+8] >> 2) & 3) - ((h & 0x08) ? 0 : 4)); - y[l+32] = d3 * ((int8_t)((q[l+0] >> 4) & 3) - ((h & 0x10) ? 0 : 4)); - y[l+40] = d3 * ((int8_t)((q[l+8] >> 4) & 3) - ((h & 0x20) ? 0 : 4)); - y[l+48] = d4 * ((int8_t)((q[l+0] >> 6) & 3) - ((h & 0x40) ? 0 : 4)); - y[l+56] = d4 * ((int8_t)((q[l+8] >> 6) & 3) - ((h & 0x80) ? 0 : 4)); - } - y += QK_K; - } -} -#endif void quantize_row_q3_K(const float * restrict x, void * restrict vy, int64_t k) { quantize_row_q3_K_reference(x, vy, k); } static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int64_t n_per_row, const float * restrict quant_weights) { -#if QK_K != 256 - (void)quant_weights; - quantize_row_q3_K_reference(x, y, n_per_row); -#else assert(n_per_row % QK_K == 0); const int nb = n_per_row / QK_K; @@ -2585,7 +2453,6 @@ static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restri x += QK_K; } -#endif } size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { @@ -2617,7 +2484,6 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict float scales[QK_K/32]; for (int i = 0; i < nb; i++) { - float max_scale = 0; // as we are deducting the min, scales are always positive float max_min = 0; for (int j = 0; j < QK_K/32; ++j) { @@ -2637,7 +2503,6 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict } } -#if QK_K == 256 float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f; float inv_min = max_min > 0 ? 63.f/max_min : 0.f; for (int j = 0; j < QK_K/32; ++j) { @@ -2669,39 +2534,7 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict L[32*j + ii] = l; } } -#else - const float s_factor = 15.f; - float inv_scale = max_scale > 0 ? s_factor/max_scale : 0.f; - float inv_min = max_min > 0 ? s_factor/max_min : 0.f; - int d1 = nearest_int(inv_scale*scales[0]); - int m1 = nearest_int(inv_min*mins[0]); - int d2 = nearest_int(inv_scale*scales[1]); - int m2 = nearest_int(inv_min*mins[1]); - y[i].scales[0] = d1 | (m1 << 4); - y[i].scales[1] = d2 | (m2 << 4); - y[i].d[0] = GGML_FP32_TO_FP16(max_scale/s_factor); - y[i].d[1] = GGML_FP32_TO_FP16(max_min/s_factor); - float sumlx = 0; - int suml2 = 0; - for (int j = 0; j < QK_K/32; ++j) { - const uint8_t sd = y[i].scales[j] & 0xF; - const uint8_t sm = y[i].scales[j] >> 4; - const float d = GGML_FP16_TO_FP32(y[i].d[0]) * sd; - if (!d) continue; - const float m = GGML_FP16_TO_FP32(y[i].d[1]) * sm; - for (int ii = 0; ii < 32; ++ii) { - int l = nearest_int((x[32*j + ii] + m)/d); - l = MAX(0, MIN(15, l)); - L[32*j + ii] = l; - sumlx += (x[32*j + ii] + m)*l*sd; - suml2 += l*l*sd*sd; - } - } - if (suml2) { - y[i].d[0] = GGML_FP32_TO_FP16(sumlx/suml2); - } -#endif uint8_t * q = y[i].qs; for (int j = 0; j < QK_K; j += 64) { for (int l = 0; l < 32; ++l) q[l] = L[j + l] | (L[j + l + 32] << 4); @@ -2709,7 +2542,6 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict } x += QK_K; - } } @@ -2718,11 +2550,8 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int6 const int nb = k / QK_K; for (int i = 0; i < nb; i++) { - const uint8_t * q = x[i].qs; -#if QK_K == 256 - const float d = GGML_FP16_TO_FP32(x[i].d); const float min = GGML_FP16_TO_FP32(x[i].dmin); @@ -2737,18 +2566,6 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int6 for (int l = 0; l < 32; ++l) *y++ = d2 * (q[l] >> 4) - m2; q += 32; is += 2; } -#else - const float dall = GGML_FP16_TO_FP32(x[i].d[0]); - const float mall = GGML_FP16_TO_FP32(x[i].d[1]); - const float d1 = dall * (x[i].scales[0] & 0xF), m1 = mall * (x[i].scales[0] >> 4); - const float d2 = dall * (x[i].scales[1] & 0xF), m2 = mall * (x[i].scales[1] >> 4); - for (int l = 0; l < 32; ++l) { - y[l+ 0] = d1 * (q[l] & 0xF) - m1; - y[l+32] = d2 * (q[l] >> 4) - m2; - } - y += QK_K; -#endif - } } @@ -2759,10 +2576,6 @@ void quantize_row_q4_K(const float * restrict x, void * restrict vy, int64_t k) } static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int64_t n_per_row, const float * quant_weights) { -#if QK_K != 256 - (void)quant_weights; - quantize_row_q4_K_reference(x, y, n_per_row); -#else assert(n_per_row % QK_K == 0); const int64_t nb = n_per_row / QK_K; @@ -2833,7 +2646,6 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri x += QK_K; } -#endif } size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { @@ -2858,21 +2670,13 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict assert(k % QK_K == 0); const int64_t nb = k / QK_K; -#if QK_K == 256 uint8_t L[QK_K]; float mins[QK_K/32]; float scales[QK_K/32]; float weights[32]; uint8_t Laux[32]; -#else - int8_t L[QK_K]; - float scales[QK_K/16]; -#endif for (int i = 0; i < nb; i++) { - -#if QK_K == 256 - float max_scale = 0; // as we are deducting the min, scales are always positive float max_min = 0; for (int j = 0; j < QK_K/32; ++j) { @@ -2944,55 +2748,8 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict m1 <<= 2; m2 <<= 2; ql += 32; } -#else - float max_scale = 0, amax = 0; - for (int j = 0; j < QK_K/16; ++j) { - scales[j] = make_qx_quants(16, 16, x + 16*j, L + 16*j, 1, NULL); - float abs_scale = fabsf(scales[j]); - if (abs_scale > amax) { - amax = abs_scale; - max_scale = scales[j]; - } - } - - float iscale = -128.f/max_scale; - for (int j = 0; j < QK_K/16; ++j) { - int l = nearest_int(iscale*scales[j]); - y[i].scales[j] = MAX(-128, MIN(127, l)); - } - y[i].d = GGML_FP32_TO_FP16(1/iscale); - - for (int j = 0; j < QK_K/16; ++j) { - const float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j]; - if (!d) continue; - for (int ii = 0; ii < 16; ++ii) { - int l = nearest_int(x[16*j + ii]/d); - l = MAX(-16, MIN(15, l)); - L[16*j + ii] = l + 16; - } - } - - uint8_t * restrict qh = y[i].qh; - uint8_t * restrict ql = y[i].qs; - memset(qh, 0, QK_K/8); - - for (int j = 0; j < 32; ++j) { - int jm = j%8; - int is = j/8; - int l1 = L[j]; - if (l1 > 15) { - l1 -= 16; qh[jm] |= (1 << is); - } - int l2 = L[j + 32]; - if (l2 > 15) { - l2 -= 16; qh[jm] |= (1 << (4 + is)); - } - ql[j] = l1 | (l2 << 4); - } -#endif x += QK_K; - } } @@ -3001,12 +2758,9 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int6 const int64_t nb = k / QK_K; for (int i = 0; i < nb; i++) { - const uint8_t * ql = x[i].qs; const uint8_t * qh = x[i].qh; -#if QK_K == 256 - const float d = GGML_FP16_TO_FP32(x[i].d); const float min = GGML_FP16_TO_FP32(x[i].dmin); @@ -3023,21 +2777,6 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int6 ql += 32; is += 2; u1 <<= 2; u2 <<= 2; } -#else - float d = GGML_FP16_TO_FP32(x[i].d); - const int8_t * restrict s = x[i].scales; - for (int l = 0; l < 8; ++l) { - y[l+ 0] = d * s[0] * ((ql[l+ 0] & 0xF) - (qh[l] & 0x01 ? 0 : 16)); - y[l+ 8] = d * s[0] * ((ql[l+ 8] & 0xF) - (qh[l] & 0x02 ? 0 : 16)); - y[l+16] = d * s[1] * ((ql[l+16] & 0xF) - (qh[l] & 0x04 ? 0 : 16)); - y[l+24] = d * s[1] * ((ql[l+24] & 0xF) - (qh[l] & 0x08 ? 0 : 16)); - y[l+32] = d * s[2] * ((ql[l+ 0] >> 4) - (qh[l] & 0x10 ? 0 : 16)); - y[l+40] = d * s[2] * ((ql[l+ 8] >> 4) - (qh[l] & 0x20 ? 0 : 16)); - y[l+48] = d * s[3] * ((ql[l+16] >> 4) - (qh[l] & 0x40 ? 0 : 16)); - y[l+56] = d * s[3] * ((ql[l+24] >> 4) - (qh[l] & 0x80 ? 0 : 16)); - } - y += QK_K; -#endif } } @@ -3048,10 +2787,6 @@ void quantize_row_q5_K(const float * restrict x, void * restrict vy, int64_t k) } static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int64_t n_per_row, const float * quant_weights) { -#if QK_K != 256 - (void)quant_weights; - quantize_row_q5_K_reference(x, y, n_per_row); -#else assert(n_per_row % QK_K == 0); const int64_t nb = n_per_row / QK_K; @@ -3142,7 +2877,6 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri x += QK_K; } -#endif } size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { @@ -3215,7 +2949,6 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict uint8_t * restrict ql = y[i].ql; uint8_t * restrict qh = y[i].qh; -#if QK_K == 256 for (int j = 0; j < QK_K; j += 128) { for (int l = 0; l < 32; ++l) { const uint8_t q1 = L[j + l + 0] & 0xF; @@ -3229,19 +2962,8 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict ql += 64; qh += 32; } -#else - for (int l = 0; l < 32; ++l) { - const uint8_t q1 = L[l + 0] & 0xF; - const uint8_t q2 = L[l + 32] & 0xF; - ql[l] = q1 | (q2 << 4); - } - for (int l = 0; l < 16; ++l) { - qh[l] = (L[l] >> 4) | ((L[l + 16] >> 4) << 2) | ((L[l + 32] >> 4) << 4) | ((L[l + 48] >> 4) << 6); - } -#endif x += QK_K; - } } @@ -3250,14 +2972,12 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int6 const int64_t nb = k / QK_K; for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict ql = x[i].ql; const uint8_t * restrict qh = x[i].qh; const int8_t * restrict sc = x[i].scales; -#if QK_K == 256 for (int n = 0; n < QK_K; n += 128) { for (int l = 0; l < 32; ++l) { int is = l/16; @@ -3275,20 +2995,6 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int6 qh += 32; sc += 8; } -#else - for (int l = 0; l < 16; ++l) { - const int8_t q1 = (int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; - const int8_t q2 = (int8_t)((ql[l+16] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; - const int8_t q3 = (int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; - const int8_t q4 = (int8_t)((ql[l+16] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; - y[l+ 0] = d * sc[0] * q1; - y[l+16] = d * sc[1] * q2; - y[l+32] = d * sc[2] * q3; - y[l+48] = d * sc[3] * q4; - } - y += 64; -#endif - } } @@ -3299,10 +3005,6 @@ void quantize_row_q6_K(const float * restrict x, void * restrict vy, int64_t k) } static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int64_t n_per_row, const float * quant_weights) { -#if QK_K != 256 - (void)quant_weights; - quantize_row_q6_K_reference(x, y, n_per_row); -#else assert(n_per_row % QK_K == 0); const int64_t nb = n_per_row / QK_K; @@ -3384,7 +3086,6 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri x += QK_K; } -#endif } size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { @@ -3801,30 +3502,21 @@ void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, in float delta[4]; uint16_t idx[4]; -#if QK_K != 64 iq1m_scale_t scale; -#endif for (int i = 0; i < nb; i++) { const uint16_t * sc = (const uint16_t *)x[i].scales; -#if QK_K == 64 - const float d = GGML_FP16_TO_FP32(x[i].d); -#else scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); const float d = GGML_FP16_TO_FP32(scale.f16); -#endif + const uint8_t * qs = x[i].qs; const uint8_t * qh = x[i].qh; for (int ib = 0; ib < QK_K/32; ++ib) { -#if QK_K == 64 - const float dl1 = d * (2*((sc[ib/2] >> (8*(ib%2)+0)) & 0xf) + 1); - const float dl2 = d * (2*((sc[ib/2] >> (8*(ib%2)+4)) & 0xf) + 1); -#else const float dl1 = d * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1); const float dl2 = d * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1); -#endif + idx[0] = qs[0] | ((qh[0] << 8) & 0x700); idx[1] = qs[1] | ((qh[0] << 4) & 0x700); idx[2] = qs[2] | ((qh[1] << 8) & 0x700); @@ -3875,9 +3567,6 @@ void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y, void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y, int64_t k) { assert(k % QK_K == 0); -#if QK_K == 64 - dequantize_row_iq4_nl((const block_iq4_nl *)x, y, k); -#else const int64_t nb = k / QK_K; for (int i = 0; i < nb; i++) { @@ -3897,7 +3586,6 @@ void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y, qs += 16; } } -#endif } //===================================== Q8_K ============================================== @@ -5849,7 +5537,6 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r #endif } -#if QK_K == 256 void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(nrc == 1); UNUSED(nrc); @@ -6433,410 +6120,6 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r #endif } -#else - -void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { - assert(nrc == 1); - UNUSED(nrc); - UNUSED(bx); - UNUSED(by); - UNUSED(bs); - - const block_q2_K * restrict x = vx; - const block_q8_K * restrict y = vy; - - const int nb = n / QK_K; - -#ifdef __ARM_NEON - const uint8x16_t m3 = vdupq_n_u8(0x3); - - const int32x4_t vzero = vdupq_n_s32(0); - - ggml_int8x16x4_t q2bytes; - - uint32_t aux32[2]; - const uint8_t * scales = (const uint8_t *)aux32; - - float sum = 0; - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - const uint32_t * restrict sc = (const uint32_t *)x[i].scales; - - aux32[0] = sc[0] & 0x0f0f0f0f; - aux32[1] = (sc[0] >> 4) & 0x0f0f0f0f; - - sum += dmin * (scales[4] * y[i].bsums[0] + scales[5] * y[i].bsums[1] + scales[6] * y[i].bsums[2] + scales[7] * y[i].bsums[3]); - - int isum1 = 0, isum2 = 0; - - const uint8x16_t q2bits = vld1q_u8(q2); - - const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); - - q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits, m3)); - q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 2), m3)); - q2bytes.val[2] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 4), m3)); - q2bytes.val[3] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 6), m3)); - - isum1 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * scales[0]; - isum2 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * scales[1]; - isum1 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[2], q8bytes.val[2])) * scales[2]; - isum2 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[3], q8bytes.val[3])) * scales[3]; - - sum += d * (isum1 + isum2); - } - - *s = sum; - -#elif defined __AVX2__ - - const __m256i m3 = _mm256_set1_epi8(3); - - __m256 acc = _mm256_setzero_ps(); - - uint32_t ud, um; - const uint8_t * restrict db = (const uint8_t *)&ud; - const uint8_t * restrict mb = (const uint8_t *)&um; - - float summs = 0; - - // TODO: optimize this - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const uint32_t * restrict sc = (const uint32_t *)x[i].scales; - ud = (sc[0] >> 0) & 0x0f0f0f0f; - um = (sc[0] >> 4) & 0x0f0f0f0f; - - int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3]; - summs += dmin * smin; - - const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); - const __m256i q2_0 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 2), q2bits), m3); - const __m256i q2_1 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3); - - const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); - - const __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0); - const __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1); - - const __m256i p_0 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p0, 0)); - const __m256i p_1 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p0, 1)); - const __m256i p_2 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p1, 0)); - const __m256i p_3 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p1, 1)); - - acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0), acc); - acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1), acc); - acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2), acc); - acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3), acc); - } - - *s = hsum_float_8(acc) + summs; - -#elif defined __AVX__ - - const __m128i m3 = _mm_set1_epi8(3); - - __m256 acc = _mm256_setzero_ps(); - - uint32_t ud, um; - const uint8_t * restrict db = (const uint8_t *)&ud; - const uint8_t * restrict mb = (const uint8_t *)&um; - - float summs = 0; - - // TODO: optimize this - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const uint32_t * restrict sc = (const uint32_t *)x[i].scales; - ud = (sc[0] >> 0) & 0x0f0f0f0f; - um = (sc[0] >> 4) & 0x0f0f0f0f; - - int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3]; - summs += dmin * smin; - - const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); - const __m128i q2_0 = _mm_and_si128(q2bits, m3); - const __m128i q2_1 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3); - const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3); - const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3); - - const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); - - const __m128i p0 = _mm_maddubs_epi16(q2_0, _mm256_extractf128_si256(q8_0, 0)); - const __m128i p1 = _mm_maddubs_epi16(q2_1, _mm256_extractf128_si256(q8_0, 1)); - const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0)); - const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1)); - - const __m256i p_0 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0)); - const __m256i p_1 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1)); - const __m256i p_2 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2)); - const __m256i p_3 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3)); - - acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc); - acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc); - acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2)), acc); - acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3)), acc); - } - - *s = hsum_float_8(acc) + summs; - -#elif defined __riscv_v_intrinsic - - uint32_t aux32[2]; - const uint8_t * scales = (const uint8_t *)aux32; - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - const uint32_t * restrict sc = (const uint32_t *)x[i].scales; - - aux32[0] = sc[0] & 0x0f0f0f0f; - aux32[1] = (sc[0] >> 4) & 0x0f0f0f0f; - - sumf += dmin * (scales[4] * y[i].bsums[0] + scales[5] * y[i].bsums[1] + scales[6] * y[i].bsums[2] + scales[7] * y[i].bsums[3]); - - int isum1 = 0; - int isum2 = 0; - - size_t vl = 16; - - vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1); - - // load Q2 - vuint8mf2_t q2_x = __riscv_vle8_v_u8mf2(q2, vl); - - vint8mf2_t q2_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q2_x, 0x03, vl)); - vint8mf2_t q2_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x2, vl), 0x03 , vl)); - vint8mf2_t q2_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x4, vl), 0x03 , vl)); - vint8mf2_t q2_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x6, vl), 0x03 , vl)); - - // load Q8, and take product with Q2 - vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q2_0, __riscv_vle8_v_i8mf2(q8, vl), vl); - vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q2_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl); - vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q2_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl); - vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q2_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl); - - vint16m1_t vs_0 = __riscv_vredsum_vs_i16m1_i16m1(p0, vzero, vl); - vint16m1_t vs_1 = __riscv_vredsum_vs_i16m1_i16m1(p1, vzero, vl); - vint16m1_t vs_2 = __riscv_vredsum_vs_i16m1_i16m1(p2, vzero, vl); - vint16m1_t vs_3 = __riscv_vredsum_vs_i16m1_i16m1(p3, vzero, vl); - - isum1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[0]; - isum2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[1]; - isum1 += __riscv_vmv_x_s_i16m1_i16(vs_2) * scales[2]; - isum2 += __riscv_vmv_x_s_i16m1_i16(vs_3) * scales[3]; - - sumf += d * (isum1 + isum2); - - } - - *s = sumf; - - -#elif defined(__POWER9_VECTOR__) - const vector signed char lowMask = vec_splats((signed char)0x3); - const vector signed char lowScaleMask = vec_splats((signed char)0xF); - const vector unsigned char v2 = vec_splats((unsigned char)0x2); - const vector unsigned char v4 = vec_splats((unsigned char)0x4); - const vector unsigned char v6 = vec_splats((unsigned char)0x6); - - vector float vsumf0 = vec_splats(0.0f); - vector float vsumf1 = vec_splats(0.0f); - vector float vsumf2 = vec_splats(0.0f); - vector float vsumf3 = vec_splats(0.0f); - -#pragma GCC unroll 2 - for (int i = 0; i < nb; ++i) { - __builtin_prefetch(x[i].qs, 0, 1); - __builtin_prefetch(y[i].qs, 0, 1); - - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); - vector float vyd = vec_splats(y[i].d); - vector float vd = vec_mul(vxd, vyd); - - vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); - vector float vdmin = vec_mul(vxmin, vyd); - - vector signed short q8ysums0 = vec_xl_len(y[i].bsums, 8); - - vector signed char q2xmins = (vector signed char)vec_xl_len(x[i].scales, 4); - vector signed char vscales = vec_and(q2xmins, lowScaleMask); - - q2xmins = vec_sr(q2xmins, v4); - vector signed short q2xmins0 = vec_unpackh((vector signed char)q2xmins); - - vector signed int prod0 = vec_mule(q2xmins0, q8ysums0); - vector signed int prod1 = vec_mulo(q2xmins0, q8ysums0); - - vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0); - vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1); - - vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].qs); - vector signed char q2x00 = vec_and(qxs0, lowMask); - vector signed char q2x01 = vec_and(vec_sr(qxs0, v2), lowMask); - vector signed char q2x02 = vec_and(vec_sr(qxs0, v4), lowMask); - vector signed char q2x03 = vec_and(vec_sr(qxs0, v6), lowMask); - - vector signed char q8y00 = vec_xl( 0, y[i].qs); - vector signed char q8y01 = vec_xl( 16, y[i].qs); - vector signed char q8y02 = vec_xl( 32, y[i].qs); - vector signed char q8y03 = vec_xl( 48, y[i].qs); - - vector signed short qv0 = vec_add(vec_mule(q2x00, q8y00), vec_mulo(q2x00, q8y00)); - vector signed short qv1 = vec_add(vec_mule(q2x01, q8y01), vec_mulo(q2x01, q8y01)); - vector signed short qv2 = vec_add(vec_mule(q2x02, q8y02), vec_mulo(q2x02, q8y02)); - vector signed short qv3 = vec_add(vec_mule(q2x03, q8y03), vec_mulo(q2x03, q8y03)); - - vector signed short vscales_h = vec_unpackh(vscales); - vector signed short vs0 = vec_splat(vscales_h, 0); - vector signed short vs1 = vec_splat(vscales_h, 1); - vector signed short vs2 = vec_splat(vscales_h, 2); - vector signed short vs3 = vec_splat(vscales_h, 3); - - vector signed int vsumi0 = vec_add(vec_mule(qv0, vs0), vec_mulo(qv0, vs0)); - vector signed int vsumi1 = vec_add(vec_mule(qv1, vs1), vec_mulo(qv1, vs1)); - vector signed int vsumi2 = vec_add(vec_mule(qv2, vs2), vec_mulo(qv2, vs2)); - vector signed int vsumi3 = vec_add(vec_mule(qv3, vs3), vec_mulo(qv3, vs3)); - - vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); - vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); - vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); - vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); - } - - vsumf0 = vec_add(vsumf0, vsumf2); - vsumf1 = vec_add(vsumf1, vsumf3); - - vsumf0 = vec_add(vsumf0, vsumf1); - - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); - - *s = vec_extract(vsumf0, 0); - -#elif defined __loongarch_asx - - const __m256i m3 = __lasx_xvreplgr2vr_b(3); - - __m256 acc = (__m256)__lasx_xvldi(0); - - uint32_t ud, um; - const uint8_t * restrict db = (const uint8_t *)&ud; - const uint8_t * restrict mb = (const uint8_t *)&um; - - float summs = 0; - - // TODO: optimize this - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const uint32_t * restrict sc = (const uint32_t *)x[i].scales; - ud = (sc[0] >> 0) & 0x0f0f0f0f; - um = (sc[0] >> 4) & 0x0f0f0f0f; - - int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3]; - summs += dmin * smin; - - const __m128i q2bits = __lsx_vld((const __m128i*)q2, 0); - const __m256i q2_0 = __lasx_xvand_v(lasx_insertf128(__lsx_vsrli_h(q2bits, 2), q2bits), m3); - const __m256i q2_1 = __lasx_xvand_v(lasx_insertf128(__lsx_vsrli_h(q2bits, 6), __lsx_vsrli_h(q2bits, 4)), m3); - - const __m256i q8_0 = __lasx_xvld((const __m256i*)(q8+ 0), 0); - const __m256i q8_1 = __lasx_xvld((const __m256i*)(q8+32), 0); - - const __m256i p0 = lasx_maddubs_h(q2_0, q8_0); - const __m256i p1 = lasx_maddubs_h(q2_1, q8_1); - - const __m256i p_0 = lasx_ext16_32(lasx_extracti128(p0, 0)); - const __m256i p_1 = lasx_ext16_32(lasx_extracti128(p0, 1)); - const __m256i p_2 = lasx_ext16_32(lasx_extracti128(p1, 0)); - const __m256i p_3 = lasx_ext16_32(lasx_extracti128(p1, 1)); - - ft_union t0, t1, t2, t3; - t0.f = d * db[0]; - t1.f = d * db[1]; - t2.f = d * db[2]; - t3.f = d * db[3]; - acc = __lasx_xvfmadd_s(__lasx_xvreplgr2vr_w(t0.i), __lasx_xvffint_s_w(p_0), acc); - acc = __lasx_xvfmadd_s(__lasx_xvreplgr2vr_w(t1.i), __lasx_xvffint_s_w(p_1), acc); - acc = __lasx_xvfmadd_s(__lasx_xvreplgr2vr_w(t2.i), __lasx_xvffint_s_w(p_2), acc); - acc = __lasx_xvfmadd_s(__lasx_xvreplgr2vr_w(t3.i), __lasx_xvffint_s_w(p_3), acc); - } - - *s = hsum_float_8(acc) + summs; - -#else - - float sumf = 0; - - int isum[QK_K/16]; - - for (int i = 0; i < nb; ++i) { - - const uint8_t * q2 = x[i].qs; - const int8_t * q8 = y[i].qs; - const uint8_t * sc = x[i].scales; - - int summs = 0; - for (int j = 0; j < QK_K/16; ++j) { - summs += y[i].bsums[j] * (sc[j] >> 4); - } - - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - - memset(isum, 0, (QK_K/16)*sizeof(int)); - for (int l = 0; l < 16; ++l) { - isum[0] += q8[l+ 0] * ((q2[l] >> 0) & 3); - isum[1] += q8[l+16] * ((q2[l] >> 2) & 3); - isum[2] += q8[l+32] * ((q2[l] >> 4) & 3); - isum[3] += q8[l+48] * ((q2[l] >> 6) & 3); - } - for (int l = 0; l < QK_K/16; ++l) { - isum[l] *= (sc[l] & 0xF); - } - sumf += dall * (isum[0] + isum[1] + isum[2] + isum[3]) - dmin * summs; - } - *s = sumf; -#endif -} -#endif - -#if QK_K == 256 void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); @@ -7616,512 +6899,6 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r } -#else - -void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { - assert(n % QK_K == 0); - assert(nrc == 1); - UNUSED(nrc); - UNUSED(bx); - UNUSED(by); - UNUSED(bs); - - const block_q3_K * restrict x = vx; - const block_q8_K * restrict y = vy; - - const int nb = n / QK_K; - -#ifdef __ARM_NEON - const int32x4_t vzero = vdupq_n_s32(0); - - const uint8x16_t m3b = vdupq_n_u8(0x3); - const uint8x16_t mh = vdupq_n_u8(4); - - ggml_int8x16x4_t q3bytes; - - uint16_t aux16[2]; - int8_t * scales = (int8_t *)aux16; - - float sum = 0; - - for (int i = 0; i < nb; ++i) { - - ggml_uint8x16x4_t q3h; - - const uint8x8_t hbits = vld1_u8(x[i].hmask); - const uint8x16_t q3bits = vld1q_u8(x[i].qs); - const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(y[i].qs); - - const uint16_t a = *(const uint16_t *)x[i].scales; - aux16[0] = a & 0x0f0f; - aux16[1] = (a >> 4) & 0x0f0f; - - for (int j = 0; j < 4; ++j) scales[j] -= 8; - - int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]); - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const uint8x16_t htmp = vcombine_u8(hbits, vshr_n_u8(hbits, 1)); - q3h.val[0] = vandq_u8(mh, vshlq_n_u8(htmp, 2)); - q3h.val[1] = vandq_u8(mh, htmp); - q3h.val[2] = vandq_u8(mh, vshrq_n_u8(htmp, 2)); - q3h.val[3] = vandq_u8(mh, vshrq_n_u8(htmp, 4)); - - q3bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q3bits, m3b), q3h.val[0])); - q3bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 2), m3b), q3h.val[1])); - q3bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 4), m3b), q3h.val[2])); - q3bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q3bits, 6), q3h.val[3])); - - isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes.val[0])) * scales[0]; - isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes.val[1])) * scales[2]; - isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes.val[2])) * scales[1]; - isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes.val[3])) * scales[3]; - - sum += d * isum; - - } - - *s = sum; - -#elif defined __AVX2__ - - const __m256i m3 = _mm256_set1_epi8(3); - const __m256i m1 = _mm256_set1_epi8(1); - - __m256 acc = _mm256_setzero_ps(); - - uint64_t aux64; - - uint16_t aux16[2]; - const int8_t * aux8 = (const int8_t *)aux16; - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const uint16_t a = *(const uint16_t *)x[i].scales; - aux16[0] = a & 0x0f0f; - aux16[1] = (a >> 4) & 0x0f0f; - - const __m256i scale_0 = MM256_SET_M128I(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8)); - const __m256i scale_1 = MM256_SET_M128I(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8)); - - memcpy(&aux64, x[i].hmask, 8); - - const __m128i haux = _mm_set_epi64x(aux64 >> 1, aux64 >> 0); - __m256i q3h_0 = MM256_SET_M128I(_mm_srli_epi16(haux, 2), haux); - __m256i q3h_1 = _mm256_srli_epi16(q3h_0, 4); - q3h_0 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_0, m1), 2); - q3h_1 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_1, m1), 2); - - // load low 2 bits - const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3); - - // prepare low and high bits - const __m256i q3aux = MM256_SET_M128I(_mm_srli_epi16(q3bits, 2), q3bits); - const __m256i q3l_0 = _mm256_and_si256(q3aux, m3); - const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3aux, 4), m3); - - // load Q8 quants - const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); - - // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16, - // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set, - // and 2 if the high bit was set) - const __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0); - const __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1); - - __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0); - __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1); - - p16_0 = _mm256_sub_epi16(p16_0, q8s_0); - p16_1 = _mm256_sub_epi16(p16_1, q8s_1); - - // multiply with scales - p16_0 = _mm256_madd_epi16(scale_0, p16_0); - p16_1 = _mm256_madd_epi16(scale_1, p16_1); - - p16_0 = _mm256_add_epi32(p16_0, p16_1); - - // multiply with block scale and accumulate - acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16_0), acc); - - } - - *s = hsum_float_8(acc); - -#elif defined __AVX__ - - const __m128i m3 = _mm_set1_epi8(3); - const __m128i m1 = _mm_set1_epi8(1); - - __m256 acc = _mm256_setzero_ps(); - - uint64_t aux64; - - uint16_t aux16[2]; - const int8_t * aux8 = (const int8_t *)aux16; - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const uint16_t a = *(const uint16_t *)x[i].scales; - aux16[0] = a & 0x0f0f; - aux16[1] = (a >> 4) & 0x0f0f; - - const __m128i scale_0 = _mm_set1_epi16(aux8[0] - 8); - const __m128i scale_1 = _mm_set1_epi16(aux8[2] - 8); - const __m128i scale_2 = _mm_set1_epi16(aux8[1] - 8); - const __m128i scale_3 = _mm_set1_epi16(aux8[3] - 8); - - memcpy(&aux64, x[i].hmask, 8); - - __m128i q3h_0 = _mm_set_epi64x(aux64 >> 1, aux64 >> 0); - __m128i q3h_1 = _mm_srli_epi16(q3h_0, 2); - __m128i q3h_2 = _mm_srli_epi16(q3h_0, 4); - __m128i q3h_3 = _mm_srli_epi16(q3h_0, 6); - q3h_0 = _mm_slli_epi16(_mm_andnot_si128(q3h_0, m1), 2); - q3h_1 = _mm_slli_epi16(_mm_andnot_si128(q3h_1, m1), 2); - q3h_2 = _mm_slli_epi16(_mm_andnot_si128(q3h_2, m1), 2); - q3h_3 = _mm_slli_epi16(_mm_andnot_si128(q3h_3, m1), 2); - - // load low 2 bits - const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3); - - // prepare low and high bits - const __m128i q3l_0 = _mm_and_si128(q3bits, m3); - const __m128i q3l_1 = _mm_and_si128(_mm_srli_epi16(q3bits, 2), m3); - const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits, 4), m3); - const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits, 6), m3); - - // load Q8 quants - const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); - - // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm_maddubs_epi16, - // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set, - // and 2 if the high bit was set) - const __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, _mm256_extractf128_si256(q8_0, 0)); - const __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, _mm256_extractf128_si256(q8_0, 1)); - const __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, _mm256_extractf128_si256(q8_1, 0)); - const __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, _mm256_extractf128_si256(q8_1, 1)); - - __m128i p16_0 = _mm_maddubs_epi16(q3l_0, _mm256_extractf128_si256(q8_0, 0)); - __m128i p16_1 = _mm_maddubs_epi16(q3l_1, _mm256_extractf128_si256(q8_0, 1)); - __m128i p16_2 = _mm_maddubs_epi16(q3l_2, _mm256_extractf128_si256(q8_1, 0)); - __m128i p16_3 = _mm_maddubs_epi16(q3l_3, _mm256_extractf128_si256(q8_1, 1)); - - p16_0 = _mm_sub_epi16(p16_0, q8s_0); - p16_1 = _mm_sub_epi16(p16_1, q8s_1); - p16_2 = _mm_sub_epi16(p16_2, q8s_2); - p16_3 = _mm_sub_epi16(p16_3, q8s_3); - - // multiply with scales - p16_0 = _mm_madd_epi16(scale_0, p16_0); - p16_1 = _mm_madd_epi16(scale_1, p16_1); - p16_2 = _mm_madd_epi16(scale_2, p16_2); - p16_3 = _mm_madd_epi16(scale_3, p16_3); - - p16_0 = _mm_add_epi32(p16_0, p16_2); - p16_1 = _mm_add_epi32(p16_1, p16_3); - __m256i p16 = MM256_SET_M128I(p16_1, p16_0); - - // multiply with block scale and accumulate - acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc); - - } - - *s = hsum_float_8(acc); - -#elif defined __riscv_v_intrinsic - - uint16_t aux16[2]; - int8_t * scales = (int8_t *)aux16; - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const uint16_t a = *(const uint16_t *)x[i].scales; - aux16[0] = a & 0x0f0f; - aux16[1] = (a >> 4) & 0x0f0f; - - for (int j = 0; j < 4; ++j) scales[j] -= 8; - - int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]); - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); - - // load qh - vuint8mf4_t qh_x1 = __riscv_vle8_v_u8mf4(x[i].hmask, 8); - vuint8mf2_t qh_x2 = __riscv_vlmul_ext_v_u8mf4_u8mf2(__riscv_vsrl_vx_u8mf4(qh_x1, 1, 8)); - - size_t vl = 16; - - // extend and combine both qh_x1 and qh_x2 - vuint8mf2_t qh_x = __riscv_vslideup_vx_u8mf2(__riscv_vlmul_ext_v_u8mf4_u8mf2(qh_x1), qh_x2, vl/2, vl); - - vuint8mf2_t qh_0 = __riscv_vand_vx_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x2, vl), 0x4, vl); - vuint8mf2_t qh_1 = __riscv_vand_vx_u8mf2(qh_x, 0x4, vl); - vuint8mf2_t qh_2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl), 0x4, vl); - vuint8mf2_t qh_3 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x4, vl), 0x4, vl); - - // load Q3 - vuint8mf2_t q3_x = __riscv_vle8_v_u8mf2(q3, vl); - - vuint8mf2_t q3h_0 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q3_x, 0x3, vl), qh_0, vl); - vuint8mf2_t q3h_1 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 2, vl), 0x3, vl), qh_1, vl); - vuint8mf2_t q3h_2 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 4, vl), 0x3, vl), qh_2, vl); - vuint8mf2_t q3h_3 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 0x6, vl), qh_3, vl); - - vint8mf2_t q3_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_0); - vint8mf2_t q3_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_1); - vint8mf2_t q3_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_2); - vint8mf2_t q3_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_3); - - // load Q8 and take product with Q3 - vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q3_0, __riscv_vle8_v_i8mf2(q8, vl), vl); - vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q3_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl); - vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q3_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl); - vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q3_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl); - - vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl); - vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl); - vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl); - vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl); - - isum += __riscv_vmv_x_s_i32m1_i32(vs_0) * scales[0]; - isum += __riscv_vmv_x_s_i32m1_i32(vs_1) * scales[2]; - isum += __riscv_vmv_x_s_i32m1_i32(vs_2) * scales[1]; - isum += __riscv_vmv_x_s_i32m1_i32(vs_3) * scales[3]; - - sumf += d * isum; - - } - - *s = sumf; - -#elif defined(__POWER9_VECTOR__) - const vector signed char lowMask = vec_splats((signed char)0x3); - const vector signed char v1 = vec_splats((signed char)0x1); - const vector unsigned char v2 = vec_splats((unsigned char)0x2); - const vector unsigned char v4 = vec_splats((unsigned char)0x4); - const vector unsigned char v6 = vec_splats((unsigned char)0x6); - const vector signed char off = vec_splats((signed char)0x8); - - vector float vsumf0 = vec_splats(0.0f); - vector float vsumf1 = vec_splats(0.0f); - vector float vsumf2 = vec_splats(0.0f); - vector float vsumf3 = vec_splats(0.0f); - -#pragma GCC unroll 2 - for (int i = 0; i < nb; ++i) { - __builtin_prefetch(x[i].qs, 0, 1); - __builtin_prefetch(y[i].qs, 0, 1); - - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); - vector float vyd = vec_splats(y[i].d); - vector float vd = vec_mul(vxd, vyd); - - uint16_t aux16[2]; - int8_t * scales = (int8_t *)aux16; - - const uint16_t a = *(const uint16_t *)x[i].scales; - aux16[0] = a & 0x0f0f; - aux16[1] = (a >> 4) & 0x0f0f; - - vector signed char vscales = (vector signed char)vec_xl_len(scales, 8); - vector signed char qxhs0 = (vector signed char)vec_xl_len(x[i].hmask, 8); - qxhs0 = vec_or(qxhs0, vec_sr(vec_sld(qxhs0, qxhs0, 8), (vector unsigned char)v1)); - - vscales = vec_sub(vscales, off); - - vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].qs); - vector signed char qxs00 = vec_and(qxs0, lowMask); - vector signed char qxs01 = vec_and(vec_sr(qxs0, v2), lowMask); - vector signed char qxs10 = vec_and(vec_sr(qxs0, v4), lowMask); - vector signed char qxs11 = vec_and(vec_sr(qxs0, v6), lowMask); - - //the 3rd bit - vector signed char qxh00 = vec_sl(vec_andc(v1, qxhs0), v2); - vector signed char qxh01 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v2)), v2); - vector signed char qxh02 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v4)), v2); - vector signed char qxh03 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v6)), v2); - qxhs0 = vec_sr(qxhs0, v4); - - vector signed char q3x00 = vec_sub(qxs00, qxh00); - vector signed char q3x01 = vec_sub(qxs01, qxh01); - vector signed char q3x10 = vec_sub(qxs10, qxh02); - vector signed char q3x11 = vec_sub(qxs11, qxh03); - - vector signed char q8y00 = vec_xl( 0, y[i].qs); - vector signed char q8y01 = vec_xl( 16, y[i].qs); - vector signed char q8y10 = vec_xl( 32, y[i].qs); - vector signed char q8y11 = vec_xl( 48, y[i].qs); - - vector signed short vscales_h = vec_unpackh(vscales); - vector signed short vs0 = vec_splat(vscales_h, 0); - vector signed short vs1 = vec_splat(vscales_h, 1); - vector signed short vs2 = vec_splat(vscales_h, 2); - vector signed short vs3 = vec_splat(vscales_h, 3); - - vector signed short qv00 = vec_add(vec_mule(q3x00, q8y00), vec_mulo(q3x00, q8y00)); - vector signed short qv10 = vec_add(vec_mule(q3x10, q8y10), vec_mulo(q3x10, q8y10)); - vector signed short qv01 = vec_add(vec_mule(q3x01, q8y01), vec_mulo(q3x01, q8y01)); - vector signed short qv11 = vec_add(vec_mule(q3x11, q8y11), vec_mulo(q3x11, q8y11)); - - vector signed int vsumi0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0)); - vector signed int vsumi1 = vec_add(vec_mule(qv10, vs1), vec_mulo(qv10, vs1)); - vector signed int vsumi2 = vec_add(vec_mule(qv01, vs2), vec_mulo(qv01, vs2)); - vector signed int vsumi3 = vec_add(vec_mule(qv11, vs3), vec_mulo(qv11, vs3)); - - vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); - vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); - vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); - vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); - } - - vsumf0 = vec_add(vsumf0, vsumf2); - vsumf1 = vec_add(vsumf1, vsumf3); - - vsumf0 = vec_add(vsumf0, vsumf1); - - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); - - *s = vec_extract(vsumf0, 0); - -#elif defined __loongarch_asx - - const __m256i m3 = __lasx_xvreplgr2vr_b(3); - const __m256i m1 = __lasx_xvreplgr2vr_b(1); - - __m256 acc = (__m256)__lasx_xvldi(0); - - uint64_t aux64; - - uint16_t aux16[2]; - const int8_t * aux8 = (const int8_t *)aux16; - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - const __m256i scale_0 = lasx_insertf128(__lasx_xvreplgr2vr_h(aux8[2] - 8), __lasx_xvreplgr2vr_h(aux8[0] - 8)); - const __m256i scale_1 = lasx_insertf128(__lasx_xvreplgr2vr_h(aux8[3] - 8), __lasx_xvreplgr2vr_h(aux8[1] - 8)); - - memcpy(&aux64, x[i].hmask, 8); - - __m128i haux = __lsx_vinsgr2vr_d(haux, aux64, 0); - haux = __lsx_vinsgr2vr_d(haux, aux64 >> 1, 1); - __m256i q3h_0 = lasx_insertf128(__lsx_vsrli_h(haux, 2), haux); - __m256i q3h_1 = __lasx_xvsrli_h(q3h_0, 4); - q3h_0 = __lasx_xvslli_h(__lasx_xvandn_v(q3h_0, m1), 2); - q3h_1 = __lasx_xvslli_h(__lasx_xvandn_v(q3h_1, m1), 2); - - // load low 2 bits - const __m128i q3bits = __lsx_vld((const __m128i*)q3, 0); - - // prepare low and high bits - const __m256i q3aux = lasx_insertf128(__lsx_vsrli_h(q3bits, 2), q3bits); - const __m256i q3l_0 = __lasx_xvand_v(q3aux, m3); - const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3aux, 4), m3); - - // load Q8 quants - const __m256i q8_0 = __lasx_xvld((const __m256i*)(q8+ 0), 0); - const __m256i q8_1 = __lasx_xvld((const __m256i*)(q8+32), 0); - - // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use lasx_maddubs_h, - // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set, - // and 2 if the high bit was set) - const __m256i q8s_0 = lasx_maddubs_h(q3h_0, q8_0); - const __m256i q8s_1 = lasx_maddubs_h(q3h_1, q8_1); - - __m256i p16_0 = lasx_maddubs_h(q3l_0, q8_0); - __m256i p16_1 = lasx_maddubs_h(q3l_1, q8_1); - - p16_0 = __lasx_xvsub_h(p16_0, q8s_0); - p16_1 = __lasx_xvsub_h(p16_1, q8s_1); - - // multiply with scales - p16_0 = lasx_madd_h(scale_0, p16_0); - p16_1 = lasx_madd_h(scale_1, p16_1); - - p16_0 = __lasx_xvadd_w(p16_0, p16_1); - - // multiply with block scale and accumulate - acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(p16_0), acc); - } - - *s = hsum_float_8(acc); - -#else - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - int32_t scales[4]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict hm = x[i].hmask; - const int8_t * restrict q8 = y[i].qs; - int8_t * restrict a = aux8; - for (int l = 0; l < 8; ++l) { - a[l+ 0] = (int8_t)((q3[l+0] >> 0) & 3) - (hm[l] & 0x01 ? 0 : 4); - a[l+ 8] = (int8_t)((q3[l+8] >> 0) & 3) - (hm[l] & 0x02 ? 0 : 4); - a[l+16] = (int8_t)((q3[l+0] >> 2) & 3) - (hm[l] & 0x04 ? 0 : 4); - a[l+24] = (int8_t)((q3[l+8] >> 2) & 3) - (hm[l] & 0x08 ? 0 : 4); - a[l+32] = (int8_t)((q3[l+0] >> 4) & 3) - (hm[l] & 0x10 ? 0 : 4); - a[l+40] = (int8_t)((q3[l+8] >> 4) & 3) - (hm[l] & 0x20 ? 0 : 4); - a[l+48] = (int8_t)((q3[l+0] >> 6) & 3) - (hm[l] & 0x40 ? 0 : 4); - a[l+56] = (int8_t)((q3[l+8] >> 6) & 3) - (hm[l] & 0x80 ? 0 : 4); - } - - scales[0] = (x[i].scales[0] & 0xF) - 8; - scales[1] = (x[i].scales[0] >> 4) - 8; - scales[2] = (x[i].scales[1] & 0xF) - 8; - scales[3] = (x[i].scales[1] >> 4) - 8; - - memset(aux32, 0, 8*sizeof(int32_t)); - for (int j = 0; j < QK_K/16; ++j) { - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] += q8[l] * a[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux32[l] += scales[j] * aux16[l]; - } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; - -#endif - -} -#endif - -#if QK_K == 256 void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); @@ -8679,381 +7456,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; #endif } -#else -void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { - assert(n % QK_K == 0); - assert(nrc == 1); - UNUSED(nrc); - UNUSED(bx); - UNUSED(by); - UNUSED(bs); - const block_q4_K * restrict x = vx; - const block_q8_K * restrict y = vy; - - const int nb = n / QK_K; - -#ifdef __ARM_NEON - const uint8x16_t m4b = vdupq_n_u8(0xf); - - const int32x4_t mzero = vdupq_n_s32(0); - - float sumf = 0; - - ggml_int8x16x2_t q4bytes; - ggml_int8x16x4_t q8bytes; - - float sum_mins = 0.f; - - uint16_t aux16[2]; - const uint8_t * restrict scales = (const uint8_t *)aux16; - - for (int i = 0; i < nb; ++i) { - - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const uint16_t * restrict a = (const uint16_t *)x[i].scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - - const int32_t summi = scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]); - sum_mins += y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * summi; - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]); - - const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); - - q8bytes = ggml_vld1q_s8_x4(q8); - q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); - q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); - - const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]); - const int32_t sumi1 = vaddvq_s32(p1) * scales[0]; - - q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4)); - q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4)); - - const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[2]), q4bytes.val[1], q8bytes.val[3]); - const int32_t sumi2 = vaddvq_s32(p2) * scales[1]; - - sumf += d * (sumi1 + sumi2); - } - - *s = sumf - sum_mins; - -#elif defined __AVX2__ - - const __m256i m4 = _mm256_set1_epi8(0xF); - - __m256 acc = _mm256_setzero_ps(); - - float summs = 0; - - uint16_t aux16[2]; - const uint8_t * scales = (const uint8_t *)aux16; - - for (int i = 0; i < nb; ++i) { - - const float d = GGML_FP16_TO_FP32(x[i].d[0]) * y[i].d; - const float m = GGML_FP16_TO_FP32(x[i].d[1]) * y[i].d; - const __m256 vd = _mm256_set1_ps(d); - - const uint16_t * a = (const uint16_t *)x[i].scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - - summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3])); - - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); - const __m256i q4l = _mm256_and_si256(q4bits, m4); - const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4); - - const __m256i q8l = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8h = _mm256_loadu_si256((const __m256i*)(q8+32)); - - const __m256i p16l = _mm256_maddubs_epi16(q4l, q8l); - const __m256i p16h = _mm256_maddubs_epi16(q4h, q8h); - - const __m256i p32l = _mm256_madd_epi16(_mm256_set1_epi16(scales[0]), p16l); - acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(p32l), acc); - - const __m256i p32h = _mm256_madd_epi16(_mm256_set1_epi16(scales[1]), p16h); - acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(p32h), acc); - - } - - *s = hsum_float_8(acc) - summs; - -#elif defined __AVX__ - - const __m128i m4 = _mm_set1_epi8(0xF); - - __m256 acc = _mm256_setzero_ps(); - - float summs = 0; - - uint16_t aux16[2]; - const uint8_t * scales = (const uint8_t *)aux16; - - for (int i = 0; i < nb; ++i) { - - const float d = GGML_FP16_TO_FP32(x[i].d[0]) * y[i].d; - const float m = GGML_FP16_TO_FP32(x[i].d[1]) * y[i].d; - const __m256 vd = _mm256_set1_ps(d); - - const uint16_t * a = (const uint16_t *)x[i].scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - - summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3])); - - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); - const __m128i q4bits_0 = _mm256_extractf128_si256(q4bits, 0); - const __m128i q4bits_1 = _mm256_extractf128_si256(q4bits, 1); - const __m128i q4_0 = _mm_and_si128(q4bits_0, m4); - const __m128i q4_1 = _mm_and_si128(q4bits_1, m4); - const __m128i q4_2 = _mm_and_si128(_mm_srli_epi16(q4bits_0, 4), m4); - const __m128i q4_3 = _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4); - - const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); - - const __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0)); - const __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1)); - const __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0)); - const __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1)); - - const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0); - const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1); - acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_1, p32_0))), acc); - - const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2); - const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3); - acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_3, p32_2))), acc); - - } - - *s = hsum_float_8(acc) - summs; - -#elif defined __riscv_v_intrinsic - - uint16_t s16[2]; - const uint8_t * restrict scales = (const uint8_t *)s16; - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const uint16_t * restrict b = (const uint16_t *)x[i].scales; - s16[0] = b[0] & 0x0f0f; - s16[1] = (b[0] >> 4) & 0x0f0f; - - sumf -= y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3])); - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]); - - size_t vl = 32; - - vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1); - - // load Q4 - vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl); - - // load Q8 and multiply it with lower Q4 nibble - vint8m1_t q4_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl)); - vint16m2_t va_0 = __riscv_vwmul_vv_i16m2(q4_a, __riscv_vle8_v_i8m1(q8, vl), vl); - vint16m1_t aux1 = __riscv_vredsum_vs_i16m2_i16m1(va_0, vzero, vl); - - sumf += d*scales[0]*__riscv_vmv_x_s_i16m1_i16(aux1); - - // load Q8 and multiply it with upper Q4 nibble - vint8m1_t q4_s = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl)); - vint16m2_t va_1 = __riscv_vwmul_vv_i16m2(q4_s, __riscv_vle8_v_i8m1(q8+32, vl), vl); - vint16m1_t aux2 = __riscv_vredsum_vs_i16m2_i16m1(va_1, vzero, vl); - - sumf += d*scales[1]*__riscv_vmv_x_s_i16m1_i16(aux2); - - } - - *s = sumf; - -#elif defined(__POWER9_VECTOR__) - const vector signed char lowMask = vec_splats((signed char)0xF); - const vector unsigned char v4 = vec_splats((unsigned char)0x4); - - vector float vsumf0 = vec_splats(0.0f); - vector float vsumf1 = vec_splats(0.0f); - vector float vsumf2 = vec_splats(0.0f); - vector float vsumf3 = vec_splats(0.0f); - -#pragma GCC unroll 2 - for (int i = 0; i < nb; ++i) { - __builtin_prefetch(x[i].qs, 0, 1); - __builtin_prefetch(y[i].qs, 0, 1); - - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d[1])); - vector float vyd = vec_splats(y[i].d); - vector float vd= vec_mul(vxd, vyd); - - uint16_t s16[2]; - const uint8_t * scales = (const uint8_t *)s16; - - const uint16_t * restrict b = (const uint16_t *)x[i].scales; - s16[0] = b[0] & 0x0f0f; - s16[1] = (b[0] >> 4) & 0x0f0f; - - vector signed char utmps = (vector signed char)vec_xl_len(scales, 4); - vector signed short vscales = (vector signed short)vec_unpackh(utmps); - vector signed short q4xmins0 = vec_mergeh(vscales, vscales); - q4xmins0 = vec_sld(q4xmins0, q4xmins0, 8); - - vector signed short q8ysums0 = vec_xl_len((const int16_t *)(y[i].bsums), 8); - - vector signed int prod0 = vec_mule(q4xmins0, q8ysums0); - vector signed int prod1 = vec_mulo(q4xmins0, q8ysums0); - - vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vd, vsumf0); - vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vd, vsumf1); - - vd = vec_mul(vyd, vec_splats(GGML_FP16_TO_FP32(x[i].d[0]))); - - vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].qs); - vector signed char qxs1 = (vector signed char)vec_xl(16, x[i].qs); - vector signed char q4x00 = vec_and(qxs0, lowMask); - vector signed char q4x01 = vec_sr(qxs0, v4); - vector signed char q4x10 = vec_and(qxs1, lowMask); - vector signed char q4x11 = vec_sr(qxs1, v4); - - vector signed char q8y00 = vec_xl( 0, y[i].qs); - vector signed char q8y10 = vec_xl(16, y[i].qs); - vector signed char q8y01 = vec_xl(32, y[i].qs); - vector signed char q8y11 = vec_xl(48, y[i].qs); - - vector signed short qv00 = vec_add(vec_mule(q4x00, q8y00), vec_mulo(q4x00, q8y00)); - vector signed short qv01 = vec_add(vec_mule(q4x01, q8y01), vec_mulo(q4x01, q8y01)); - vector signed short qv10 = vec_add(vec_mule(q4x10, q8y10), vec_mulo(q4x10, q8y10)); - vector signed short qv11 = vec_add(vec_mule(q4x11, q8y11), vec_mulo(q4x11, q8y11)); - - vector signed short vs0 = vec_splat(vscales, 0); - vector signed short vs1 = vec_splat(vscales, 1); - - vector signed int vsumi0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0)); - vector signed int vsumi1 = vec_add(vec_mule(qv10, vs0), vec_mulo(qv10, vs0)); - vector signed int vsumi2 = vec_add(vec_mule(qv01, vs1), vec_mulo(qv01, vs1)); - vector signed int vsumi3 = vec_add(vec_mule(qv11, vs1), vec_mulo(qv11, vs1)); - - vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); - vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); - vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); - vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); - } - - vsumf0 = vec_add(vsumf0, vsumf2); - vsumf1 = vec_add(vsumf1, vsumf3); - - vsumf0 = vec_add(vsumf0, vsumf1); - - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); - - *s = vec_extract(vsumf0, 0); - -#elif defined __loongarch_asx - - const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); - - __m256 acc = (__m256)__lasx_xvldi(0); - - float summs = 0; - - uint16_t aux16[2]; - const uint8_t * scales = (const uint8_t *)aux16; - - for (int i = 0; i < nb; ++i) { - - const float d = GGML_FP16_TO_FP32(x[i].d[0]) * y[i].d; - const float m = GGML_FP16_TO_FP32(x[i].d[1]) * y[i].d; - const __m256 vd = __lasx_xvreplfr2vr_s(d); - - const uint16_t * a = (const uint16_t *)x[i].scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - - summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3])); - - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const __m256i q4bits = __lasx_xvld((const __m256i*)q4, 0); - const __m256i q4l = __lasx_xvand_v(q4bits, m4); - const __m256i q4h = __lasx_xvand_v(__lasx_xvsrli_h(q4bits, 4), m4); - - const __m256i q8l = __lasx_xvld((const __m256i*)(q8+ 0), 0); - const __m256i q8h = __lasx_xvld((const __m256i*)(q8+32), 0); - - const __m256i p16l = lasx_maddubs_h(q4l, q8l); - const __m256i p16h = lasx_maddubs_h(q4h, q8h); - - const __m256i p32l = lasx_madd_h(__lasx_xvreplgr2vr_h(scales[0]), p16l); - acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(p32l), acc); - - const __m256i p32h = lasx_madd_h(__lasx_xvreplgr2vr_h(scales[1]), p16h); - acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(p32h), acc); - } - - *s = hsum_float_8(acc) - summs; - -#else - - uint8_t aux8[QK_K]; - int16_t aux16[16]; - float sums [8]; - memset(sums, 0, 8*sizeof(float)); - - uint16_t s16[2]; - const uint8_t * restrict scales = (const uint8_t *)s16; - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - uint8_t * restrict a = aux8; - for (int l = 0; l < 32; ++l) a[l+ 0] = q4[l] & 0xF; - for (int l = 0; l < 32; ++l) a[l+32] = q4[l] >> 4; - - const uint16_t * restrict b = (const uint16_t *)x[i].scales; - s16[0] = b[0] & 0x0f0f; - s16[1] = (b[0] >> 4) & 0x0f0f; - - sumf -= y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3])); - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]); - - for (int j = 0; j < QK_K/32; ++j) { - for (int l = 0; l < 16; ++l) aux16[l] = q8[l] * a[l]; - q8 += 16; a += 16; - for (int l = 0; l < 16; ++l) aux16[l] += q8[l] * a[l]; - q8 += 16; a += 16; - const float dl = d * scales[j]; - for (int l = 0; l < 8; ++l) sums[l] += dl * (aux16[l] + aux16[l+8]); - } - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; -#endif -} -#endif - -#if QK_K == 256 void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); @@ -9151,12 +7554,10 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r float summs = 0.f; - for (int i = 0; i < nb; ++i) { - + for (int i = 0; i < nb; ++i) { const uint8_t * restrict q5 = x[i].qs; const int8_t * restrict q8 = y[i].qs; -#if QK_K == 256 const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); @@ -9166,10 +7567,6 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); utmp[2] = uaux; utmp[0] &= kmask1; -#else - // TODO - const float d = 0, dmin = 0; -#endif const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0])); @@ -9552,15 +7949,10 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r const uint8_t * restrict q5 = x[i].qs; const int8_t * restrict q8 = y[i].qs; -#if QK_K == 256 const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); -#else - // TODO - const float d = 0, dmin = 0; -#endif const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0])); @@ -9679,402 +8071,6 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r #endif } -#else - -void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { - assert(n % QK_K == 0); - assert(nrc == 1); - UNUSED(nrc); - UNUSED(bx); - UNUSED(by); - UNUSED(bs); - - const block_q5_K * restrict x = vx; - const block_q8_K * restrict y = vy; - - const int nb = n / QK_K; - -#ifdef __ARM_NEON - const uint8x16_t m4b = vdupq_n_u8(0xf); - const uint8x16_t mh = vdupq_n_u8(16); - const int32x4_t mzero = vdupq_n_s32(0); - - ggml_int8x16x4_t q5bytes; - ggml_uint8x16x4_t q5h; - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const int8_t * sc = x[i].scales; - - const uint8_t * restrict q5 = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - - const uint8x8_t qhbits = vld1_u8(qh); - - const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); - const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); - - const uint8x16_t htmp = vcombine_u8(qhbits, vshr_n_u8(qhbits, 1)); - q5h.val[0] = vbicq_u8(mh, vshlq_n_u8(htmp, 4)); - q5h.val[1] = vbicq_u8(mh, vshlq_n_u8(htmp, 2)); - q5h.val[2] = vbicq_u8(mh, htmp); - q5h.val[3] = vbicq_u8(mh, vshrq_n_u8(htmp, 2)); - - q5bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q5bits.val[0], m4b)), vreinterpretq_s8_u8(q5h.val[0])); - q5bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q5bits.val[1], m4b)), vreinterpretq_s8_u8(q5h.val[1])); - q5bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[0], 4)), vreinterpretq_s8_u8(q5h.val[2])); - q5bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[1], 4)), vreinterpretq_s8_u8(q5h.val[3])); - - int32_t sumi1 = sc[0] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0])); - int32_t sumi2 = sc[1] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[1], q8bytes.val[1])); - int32_t sumi3 = sc[2] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2])); - int32_t sumi4 = sc[3] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[3], q8bytes.val[3])); - - sumf += d * (sumi1 + sumi2 + sumi3 + sumi4); - } - - *s = sumf; - -#elif defined __AVX2__ - - const __m256i m4 = _mm256_set1_epi8(0xF); - const __m256i mone = _mm256_set1_epi8(1); - - __m256 acc = _mm256_setzero_ps(); - - for (int i = 0; i < nb; ++i) { - - const uint8_t * restrict q5 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); - - const __m256i scale_l = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0])); - const __m256i scale_h = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2])); - - int64_t aux64; - memcpy(&aux64, x[i].qh, 8); - const __m128i haux128 = _mm_set_epi64x(aux64 >> 1, aux64); - const __m256i haux256 = MM256_SET_M128I(_mm_srli_epi16(haux128, 2), haux128); - - const __m256i q5h_0 = _mm256_slli_epi16(_mm256_andnot_si256(haux256, mone), 4); - const __m256i q5h_1 = _mm256_slli_epi16(_mm256_andnot_si256(_mm256_srli_epi16(haux256, 4), mone), 4); - - const __m256i q5l_0 = _mm256_and_si256(q5bits, m4); - const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4); - - const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); - - const __m256i p16_0 = _mm256_madd_epi16(scale_l, _mm256_maddubs_epi16(q5l_0, q8_0)); - const __m256i p16_1 = _mm256_madd_epi16(scale_h, _mm256_maddubs_epi16(q5l_1, q8_1)); - const __m256i s16_0 = _mm256_madd_epi16(scale_l, _mm256_maddubs_epi16(q5h_0, q8_0)); - const __m256i s16_1 = _mm256_madd_epi16(scale_h, _mm256_maddubs_epi16(q5h_1, q8_1)); - - const __m256i dot = _mm256_sub_epi32(_mm256_add_epi32(p16_0, p16_1), _mm256_add_epi32(s16_0, s16_1)); - - acc = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(dot), acc); - - } - - *s = hsum_float_8(acc); - -#elif defined __AVX__ - - const __m128i m4 = _mm_set1_epi8(0xF); - const __m128i mone = _mm_set1_epi8(1); - - __m256 acc = _mm256_setzero_ps(); - - for (int i = 0; i < nb; ++i) { - - const uint8_t * restrict q5 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); - - const __m128i scale_0 = _mm_set1_epi16(x[i].scales[0]); - const __m128i scale_1 = _mm_set1_epi16(x[i].scales[1]); - const __m128i scale_2 = _mm_set1_epi16(x[i].scales[2]); - const __m128i scale_3 = _mm_set1_epi16(x[i].scales[3]); - - int64_t aux64; - memcpy(&aux64, x[i].qh, 8); - const __m128i haux128_0 = _mm_set_epi64x(aux64 >> 1, aux64); - const __m128i haux128_1 = _mm_srli_epi16(haux128_0, 2); - - const __m128i q5h_0 = _mm_slli_epi16(_mm_andnot_si128(haux128_0, mone), 4); - const __m128i q5h_1 = _mm_slli_epi16(_mm_andnot_si128(haux128_1, mone), 4); - const __m128i q5h_2 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_0, 4), mone), 4); - const __m128i q5h_3 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_1, 4), mone), 4); - - const __m128i q5l_0 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 0), m4); - const __m128i q5l_1 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 1), m4); - const __m128i q5l_2 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 0), 4), m4); - const __m128i q5l_3 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 1), 4), m4); - - const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); - - const __m128i p16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5l_0, _mm256_extractf128_si256(q8_0, 0))); - const __m128i p16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5l_1, _mm256_extractf128_si256(q8_0, 1))); - const __m128i p16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5l_2, _mm256_extractf128_si256(q8_1, 0))); - const __m128i p16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5l_3, _mm256_extractf128_si256(q8_1, 1))); - const __m128i s16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5h_0, _mm256_extractf128_si256(q8_0, 0))); - const __m128i s16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5h_1, _mm256_extractf128_si256(q8_0, 1))); - const __m128i s16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5h_2, _mm256_extractf128_si256(q8_1, 0))); - const __m128i s16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5h_3, _mm256_extractf128_si256(q8_1, 1))); - - const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2)); - const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3)); - - acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(dot_1, dot_0))), acc); - - } - - *s = hsum_float_8(acc); - -#elif defined __riscv_v_intrinsic - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const int8_t * sc = x[i].scales; - - const uint8_t * restrict q5 = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - - vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); - - // load qh - vuint8mf4_t qh_x1 = __riscv_vle8_v_u8mf4(qh, 8); - vuint8mf2_t qh_x2 = __riscv_vlmul_ext_v_u8mf4_u8mf2(__riscv_vsrl_vx_u8mf4(qh_x1, 1, 8)); - - size_t vl = 16; - - // combine both qh_1 and qh_2 - vuint8mf2_t qh_x = __riscv_vslideup_vx_u8mf2(__riscv_vlmul_ext_v_u8mf4_u8mf2(qh_x1), qh_x2, vl/2, vl); - - vuint8mf2_t qh_h0 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x4, vl), vl), 16, vl); - vuint8mf2_t qh_h1 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x2, vl), vl), 16, vl); - vuint8mf2_t qh_h2 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(qh_x, vl), 16, vl); - vuint8mf2_t qh_h3 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x4, vl), vl), 16, vl); - - vint8mf2_t qh_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h0); - vint8mf2_t qh_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h1); - vint8mf2_t qh_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h2); - vint8mf2_t qh_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h3); - - // load q5 - vuint8mf2_t q5_x1 = __riscv_vle8_v_u8mf2(q5, vl); - vuint8mf2_t q5_x2 = __riscv_vle8_v_u8mf2(q5+16, vl); - - vint8mf2_t q5s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q5_x1, 0xF, vl)); - vint8mf2_t q5s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q5_x2, 0xF, vl)); - vint8mf2_t q5s_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(q5_x1, 0x4, vl)); - vint8mf2_t q5s_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(q5_x2, 0x4, vl)); - - vint8mf2_t q5_0 = __riscv_vsub_vv_i8mf2(q5s_0, qh_0, vl); - vint8mf2_t q5_1 = __riscv_vsub_vv_i8mf2(q5s_1, qh_1, vl); - vint8mf2_t q5_2 = __riscv_vsub_vv_i8mf2(q5s_2, qh_2, vl); - vint8mf2_t q5_3 = __riscv_vsub_vv_i8mf2(q5s_3, qh_3, vl); - - // load Q8 and multiply it with Q5 - vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q5_0, __riscv_vle8_v_i8mf2(q8, vl), vl); - vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q5_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl); - vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q5_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl); - vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q5_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl); - - vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl); - vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl); - vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl); - vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl); - - int32_t sumi1 = sc[0] * __riscv_vmv_x_s_i32m1_i32(vs_0); - int32_t sumi2 = sc[1] * __riscv_vmv_x_s_i32m1_i32(vs_1); - int32_t sumi3 = sc[2] * __riscv_vmv_x_s_i32m1_i32(vs_2); - int32_t sumi4 = sc[3] * __riscv_vmv_x_s_i32m1_i32(vs_3); - - sumf += d * (sumi1 + sumi2 + sumi3 + sumi4); - - } - - *s = sumf; - -#elif defined(__POWER9_VECTOR__) - const vector signed char lowMask = vec_splats((signed char)0xF); - const vector unsigned char v1 = vec_splats((unsigned char)0x1); - const vector unsigned char v2 = vec_splats((unsigned char)0x2); - const vector unsigned char v4 = vec_splats((unsigned char)0x4); - - vector float vsumf0 = vec_splats(0.0f); - vector float vsumf1 = vec_splats(0.0f); - vector float vsumf2 = vec_splats(0.0f); - vector float vsumf3 = vec_splats(0.0f); - -#pragma GCC unroll 2 - for (int i = 0; i < nb; ++i) { - __builtin_prefetch(x[i].qs, 0, 1); - __builtin_prefetch(y[i].qs, 0, 1); - - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); - vector float vyd = vec_splats(y[i].d); - vector float vd= vec_mul(vxd, vyd); - - vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].qs); - vector signed char qxs1 = (vector signed char)vec_xl(16, x[i].qs); - vector signed char qxs00 = (vector signed char)vec_and(qxs0, lowMask); - vector signed char qxs01 = (vector signed char)vec_sr(qxs0, v4); - vector signed char qxs10 = (vector signed char)vec_and(qxs1, lowMask); - vector signed char qxs11 = (vector signed char)vec_sr(qxs1, v4); - - vector signed char qxhs = (vector signed char)vec_xl_len(x[i].qh, 8); - vector signed char qxhs0 = vec_or(qxhs, vec_sr(vec_sld(qxhs, qxhs, 8), v1)); - vector signed char qxhs1 = vec_sr(qxhs0, v2); - vector signed char qxh00 = vec_sl(vec_andc((vector signed char)v1, qxhs0), v4); - vector signed char qxh10 = vec_sl(vec_andc((vector signed char)v1, qxhs1), v4); - vector signed char qxh01 = vec_sl(vec_andc((vector signed char)v1, vec_sr(qxhs0, v4)), v4); - vector signed char qxh11 = vec_sl(vec_andc((vector signed char)v1, vec_sr(qxhs1, v4)), v4); - - vector signed char q5x00 = vec_sub(qxs00, qxh00); - vector signed char q5x10 = vec_sub(qxs10, qxh10); - vector signed char q5x01 = vec_sub(qxs01, qxh01); - vector signed char q5x11 = vec_sub(qxs11, qxh11); - - vector signed char q8y00 = vec_xl( 0, y[i].qs); - vector signed char q8y10 = vec_xl(16, y[i].qs); - vector signed char q8y01 = vec_xl(32, y[i].qs); - vector signed char q8y11 = vec_xl(48, y[i].qs); - - vector signed short qv00 = vec_add(vec_mule(q5x00, q8y00), vec_mulo(q5x00, q8y00)); - vector signed short qv01 = vec_add(vec_mule(q5x01, q8y01), vec_mulo(q5x01, q8y01)); - vector signed short qv10 = vec_add(vec_mule(q5x10, q8y10), vec_mulo(q5x10, q8y10)); - vector signed short qv11 = vec_add(vec_mule(q5x11, q8y11), vec_mulo(q5x11, q8y11)); - - vector signed short vs = (vector signed short)vec_unpackh(vec_xl_len(x[i].scales, 4)); - vector signed short vs0 = vec_splat(vs, 0); - vector signed short vs1 = vec_splat(vs, 1); - vector signed short vs2 = vec_splat(vs, 2); - vector signed short vs3 = vec_splat(vs, 3); - - vector signed int vsumi0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0)); - vector signed int vsumi1 = vec_add(vec_mule(qv10, vs1), vec_mulo(qv10, vs1)); - vector signed int vsumi2 = vec_add(vec_mule(qv01, vs2), vec_mulo(qv01, vs2)); - vector signed int vsumi3 = vec_add(vec_mule(qv11, vs3), vec_mulo(qv11, vs3)); - - vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); - vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); - vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); - vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); - } - - vsumf0 = vec_add(vsumf0, vsumf2); - vsumf1 = vec_add(vsumf1, vsumf3); - - vsumf0 = vec_add(vsumf0, vsumf1); - - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); - - *s = vec_extract(vsumf0, 0); - -#elif defined __loongarch_asx - - const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); - const __m256i mone = __lasx_xvreplgr2vr_b(1); - - __m256 acc = (__m256)__lasx_xvldi(0); - - for (int i = 0; i < nb; ++i) { - - const uint8_t * restrict q5 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); - - const __m256i scale_l = lasx_insertf128(__lsx_vreplgr2vr_h(x[i].scales[1]), __lsx_vreplgr2vr_h(x[i].scales[0])); - const __m256i scale_h = lasx_insertf128(__lsx_vreplgr2vr_h(x[i].scales[3]), __lsx_vreplgr2vr_h(x[i].scales[2])); - - int64_t aux64; - memcpy(&aux64, x[i].qh, 8); - __m128i haux128 = __lsx_vinsgr2vr_d(haux128, aux64, 0); - haux128 = __lsx_vinsgr2vr_d(haux128, aux64 >> 1, 1); - const __m256i haux256 = lasx_insertf128(__lsx_vsrli_h(haux128, 2), haux128); - - const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvandn_v(haux256, mone), 4); - const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvandn_v(__lasx_xvsrli_h(haux256, 4), mone), 4); - - const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4); - const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4); - - const __m256i q8_0 = __lasx_xvld((const __m256i*)(q8+ 0), 0); - const __m256i q8_1 = __lasx_xvld((const __m256i*)(q8+32), 0); - - const __m256i p16_0 = lasx_madd_h(scale_l, lasx_maddubs_h(q5l_0, q8_0)); - const __m256i p16_1 = lasx_madd_h(scale_h, lasx_maddubs_h(q5l_1, q8_1)); - const __m256i s16_0 = lasx_madd_h(scale_l, lasx_maddubs_h(q5h_0, q8_0)); - const __m256i s16_1 = lasx_madd_h(scale_h, lasx_maddubs_h(q5h_1, q8_1)); - - const __m256i dot = __lasx_xvsub_w(__lasx_xvadd_w(p16_0, p16_1), __lasx_xvadd_w(s16_0, s16_1)); - - acc = __lasx_xvfmadd_s((__m256)__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(dot), acc); - } - - *s = hsum_float_8(acc); - -#else - - int8_t aux8[QK_K]; - int16_t aux16[16]; - float sums [8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q4 = x[i].qs; - const uint8_t * restrict hm = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - int8_t * restrict a = aux8; - for (int l = 0; l < 32; ++l) { - a[l+ 0] = q4[l] & 0xF; - a[l+32] = q4[l] >> 4; - } - for (int is = 0; is < 8; ++is) { - uint8_t m = 1 << is; - for (int l = 0; l < 8; ++l) a[8*is + l] -= (hm[l] & m ? 0 : 16); - } - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const int8_t * restrict sc = x[i].scales; - - for (int j = 0; j < QK_K/16; ++j) { - const float dl = d * sc[j]; - for (int l = 0; l < 16; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) sums[l] += dl * (aux16[l] + aux16[8+l]); - q8 += 16; a += 16; - } - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; -#endif -} -#endif - - -#if QK_K == 256 void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); @@ -10733,446 +8729,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r #endif } -#else - -void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { - assert(n % QK_K == 0); - assert(nrc == 1); - UNUSED(nrc); - UNUSED(bx); - UNUSED(by); - UNUSED(bs); - - const block_q6_K * restrict x = vx; - const block_q8_K * restrict y = vy; - - const int nb = n / QK_K; - -#ifdef __ARM_NEON - float sum = 0; - - const uint8x16_t m4b = vdupq_n_u8(0xF); - const int8x16_t m32s = vdupq_n_s8(32); - const int32x4_t vzero = vdupq_n_s32(0); - - const uint8x16_t mone = vdupq_n_u8(3); - - ggml_int8x16x4_t q6bytes; - ggml_uint8x16x4_t q6h; - - for (int i = 0; i < nb; ++i) { - - const float d_all = GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict q6 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - - const int8_t * restrict scale = x[i].scales; - - int32_t isum = 0; - - uint8x16_t qhbits = vld1q_u8(qh); - ggml_uint8x16x2_t q6bits = ggml_vld1q_u8_x2(q6); - ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); - - q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits), 4); - uint8x16_t shifted = vshrq_n_u8(qhbits, 2); - q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4); - shifted = vshrq_n_u8(qhbits, 4); - q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4); - shifted = vshrq_n_u8(qhbits, 6); - q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4); - - q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s); - q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s); - q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[2])), m32s); - q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[3])), m32s); - - isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] + - vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] + - vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] + - vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3]; - - sum += isum * d_all * y[i].d; - - } - *s = sum; - -#elif defined __AVX2__ - - const __m256i m4 = _mm256_set1_epi8(0xF); - const __m256i m2 = _mm256_set1_epi8(3); - const __m256i m32s = _mm256_set1_epi8(32); - - __m256 acc = _mm256_setzero_ps(); - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict q4 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - - const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]); - const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]); - const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]); - const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]); - - __m256i sumi = _mm256_setzero_si256(); - - const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1); - const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3); - - const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); - const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh); - - const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4); - const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4); - - const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0); - const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_1); - - const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); - - __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0); - __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1); - - __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0); - __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1); - - p16_0 = _mm256_sub_epi16(p16_0, q8s_0); - p16_1 = _mm256_sub_epi16(p16_1, q8s_1); - - p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0); - p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1); - - sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1)); - - acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc); - } - - *s = hsum_float_8(acc); - -#elif defined __AVX__ - - const __m128i m4 = _mm_set1_epi8(0xF); - const __m128i m2 = _mm_set1_epi8(3); - const __m128i m32s = _mm_set1_epi8(32); - - __m256 acc = _mm256_setzero_ps(); - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict q4 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - - const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]); - const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]); - const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]); - const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]); - - __m128i sumi_0 = _mm_setzero_si128(); - __m128i sumi_1 = _mm_setzero_si128(); - - const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1); - const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3); - - const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); - const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh); - - const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH, m2), 4); - const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 2), m2), 4); - const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 4), m2), 4); - const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 6), m2), 4); - - const __m128i q4_0 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 0), m4), q4h_0); - const __m128i q4_1 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 1), m4), q4h_1); - const __m128i q4_2 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 0), 4), m4), q4h_2); - const __m128i q4_3 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 1), 4), m4), q4h_3); - - const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); - const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); - - __m128i q8s_0 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 0)); - __m128i q8s_1 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 1)); - __m128i q8s_2 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 0)); - __m128i q8s_3 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 1)); - - __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0)); - __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1)); - __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0)); - __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1)); - - p16_0 = _mm_sub_epi16(p16_0, q8s_0); - p16_1 = _mm_sub_epi16(p16_1, q8s_1); - p16_2 = _mm_sub_epi16(p16_2, q8s_2); - p16_3 = _mm_sub_epi16(p16_3, q8s_3); - - p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0); - p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1); - p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2); - p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3); - - sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2)); - sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3)); - - acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi_1, sumi_0))), acc); - } - - *s = hsum_float_8(acc); - -#elif defined __riscv_v_intrinsic - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const float d_all = GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict q6 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - - const int8_t * restrict scale = x[i].scales; - - int32_t isum = 0; - - size_t vl = 16; - - vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); - - // load Q6 - vuint8mf2_t q6_0 = __riscv_vle8_v_u8mf2(q6, vl); - vuint8mf2_t q6_1 = __riscv_vle8_v_u8mf2(q6+16, vl); - - // load qh - vuint8mf2_t qh_x = __riscv_vle8_v_u8mf2(qh, vl); - - vuint8mf2_t qh0 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl); - qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl); - vuint8mf2_t qh1 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl); - qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl); - vuint8mf2_t qh2 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl); - qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl); - vuint8mf2_t qh3 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl); - - vuint8mf2_t q6h_0 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q6_0, 0xF, vl), qh0, vl); - vuint8mf2_t q6h_1 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q6_1, 0xF, vl), qh1, vl); - vuint8mf2_t q6h_2 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q6_0, 0x4, vl), qh2, vl); - vuint8mf2_t q6h_3 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q6_1, 0x4, vl), qh3, vl); - - vint8mf2_t q6v_0 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_0), 32, vl); - vint8mf2_t q6v_1 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_1), 32, vl); - vint8mf2_t q6v_2 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_2), 32, vl); - vint8mf2_t q6v_3 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_3), 32, vl); - - // load Q8 and take product - vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q6v_0, __riscv_vle8_v_i8mf2(q8, vl), vl); - vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q6v_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl); - vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q6v_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl); - vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q6v_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl); - - vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl); - vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl); - vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl); - vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl); - - isum += __riscv_vmv_x_s_i32m1_i32(vs_0) * scale[0]; - isum += __riscv_vmv_x_s_i32m1_i32(vs_1) * scale[1]; - isum += __riscv_vmv_x_s_i32m1_i32(vs_2) * scale[2]; - isum += __riscv_vmv_x_s_i32m1_i32(vs_3) * scale[3]; - - sumf += isum * d_all * y[i].d; - - } - - *s = sumf; - -#elif defined(__POWER9_VECTOR__) - const vector signed char lowMask = vec_splats((signed char)0xF); - const vector unsigned char v2 = vec_splats((unsigned char)0x2); - const vector unsigned char v3 = vec_splats((unsigned char)0x3); - const vector unsigned char v4 = vec_splats((unsigned char)0x4); - const vector unsigned char v6 = vec_splats((unsigned char)0x6); - const vector signed char off = vec_splats((signed char)0x20); - - vector float vsumf0 = vec_splats(0.0f); - vector float vsumf1 = vec_splats(0.0f); - vector float vsumf2 = vec_splats(0.0f); - vector float vsumf3 = vec_splats(0.0f); - -#pragma GCC unroll 2 - for (int i = 0; i < nb; ++i) { - __builtin_prefetch(x[i].ql, 0, 1); - __builtin_prefetch(x[i].qh, 0, 1); - __builtin_prefetch(y[i].qs, 0, 1); - - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); - vector float vyd = vec_splats(y[i].d); - vector float vd= vec_mul(vxd, vyd); - - vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].ql); - vector signed char qxs1 = (vector signed char)vec_xl(16, x[i].ql); - vector signed char qxs00 = vec_and(qxs0, lowMask); - vector signed char qxs01 = vec_sr(qxs0, v4); - vector signed char qxs10 = vec_and(qxs1, lowMask); - vector signed char qxs11 = vec_sr(qxs1, v4); - - vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh); - - vector signed char qxh00 = vec_sl(vec_and((vector signed char)v3, qxhs0), v4); - vector signed char qxh01 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v4)), v4); - vector signed char qxh10 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v2)), v4); - vector signed char qxh11 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v6)), v4); - - vector signed char q6x00 = vec_sub(vec_or(qxh00, qxs00), off); - vector signed char q6x01 = vec_sub(vec_or(qxh01, qxs01), off); - vector signed char q6x10 = vec_sub(vec_or(qxh10, qxs10), off); - vector signed char q6x11 = vec_sub(vec_or(qxh11, qxs11), off); - - vector signed char q8y00 = vec_xl( 0, y[i].qs); - vector signed char q8y10 = vec_xl(16, y[i].qs); - vector signed char q8y01 = vec_xl(32, y[i].qs); - vector signed char q8y11 = vec_xl(48, y[i].qs); - - vector signed short qv00 = vec_add(vec_mule(q6x00, q8y00), vec_mulo(q6x00, q8y00)); - vector signed short qv10 = vec_add(vec_mule(q6x10, q8y10), vec_mulo(q6x10, q8y10)); - vector signed short qv01 = vec_add(vec_mule(q6x01, q8y01), vec_mulo(q6x01, q8y01)); - vector signed short qv11 = vec_add(vec_mule(q6x11, q8y11), vec_mulo(q6x11, q8y11)); - - vector signed short vs = (vector signed short)vec_unpackh(vec_xl_len(x[i].scales, 4)); - vector signed short vs0 = vec_splat(vs, 0); - vector signed short vs1 = vec_splat(vs, 1); - vector signed short vs2 = vec_splat(vs, 2); - vector signed short vs3 = vec_splat(vs, 3); - - vector signed int vsumi0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0)); - vector signed int vsumi1 = vec_add(vec_mule(qv10, vs1), vec_mulo(qv10, vs1)); - vector signed int vsumi2 = vec_add(vec_mule(qv01, vs2), vec_mulo(qv01, vs2)); - vector signed int vsumi3 = vec_add(vec_mule(qv11, vs3), vec_mulo(qv11, vs3)); - - vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); - vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); - vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); - vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); - } - - vsumf0 = vec_add(vsumf0, vsumf2); - vsumf1 = vec_add(vsumf1, vsumf3); - - vsumf0 = vec_add(vsumf0, vsumf1); - - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); - vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); - - *s = vec_extract(vsumf0, 0); - -#elif defined __loongarch_asx - - const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); - const __m256i m2 = __lasx_xvreplgr2vr_b(3); - const __m256i m32s = __lasx_xvreplgr2vr_b(32); - - __m256 acc = (__m256)__lasx_xvldi(0); - - for (int i = 0; i < nb; ++i) { - - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict q4 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - - const __m64 scales_1 = __lasx_xvreplgr2vr_b(x[i].scales[0]); - const __m64 scales_2 = __lasx_xvreplgr2vr_b(x[i].scales[1]); - const __m64 scales_3 = __lasx_xvreplgr2vr_b(x[i].scales[2]); - const __m64 scales_4 = __lasx_xvreplgr2vr_b(x[i].scales[3]); - - __m256i sumi = __lasx_xvldi(0); - - __m128i scale_0 = __lsx_vinsgr2vr_d(scale_0, scales_1, 0); - scale_0 = __lsx_vinsgr2vr_d(scale_0, scales_2, 1); - __m128i scale_1 = __lsx_vinsgr2vr_d(scale_1, scales_3, 0); - scale_1 = __lsx_vinsgr2vr_d(scale_1, scales_4, 1); - - const __m256i q4bits1 = __lasx_xvld((const __m256i*)q4, 0); - const __m128i q4bitsH = __lsx_vld((const __m128i*)qh, 0); - - const __m256i q4h_0 = __lasx_xvslli_h(__lasx_xvand_v(lasx_insertf128(__lasx_xvsrli_h(q4bitsH, 2), q4bitsH), m2), 4); - const __m256i q4h_1 = __lasx_xvslli_h(__lasx_xvand_v(lasx_insertf128(__lasx_xvsrli_h(q4bitsH, 6), __lasx_xvsrli_h(q4bitsH, 4)), m2), 4); - - const __m256i q4_0 = __lasx_xvor_v(__lasx_xvand_v(q4bits1, m4), q4h_0); - const __m256i q4_1 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits1, 4), m4), q4h_1); - - const __m256i q8_0 = __lasx_xvld((const __m256i*)(q8+ 0), 0); - const __m256i q8_1 = __lasx_xvld((const __m256i*)(q8+32), 0); - - __m256i q8s_0 = lasx_maddubs_h(m32s, q8_0); - __m256i q8s_1 = lasx_maddubs_h(m32s, q8_1); - - __m256i p16_0 = lasx_maddubs_h(q4_0, q8_0); - __m256i p16_1 = lasx_maddubs_h(q4_1, q8_1); - - p16_0 = __lasx_xvsub_h(p16_0, q8s_0); - p16_1 = __lasx_xvsub_h(p16_1, q8s_1); - - p16_0 = lasx_madd_h(lasx_ext8_16(scale_0), p16_0); - p16_1 = lasx_madd_h(lasx_ext8_16(scale_1), p16_1); - - sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1)); - - acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc); - } - - *s = hsum_float_8(acc); - -#else - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q4 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * restrict a = aux8; - for (int l = 0; l < 16; ++l) { - a[l+ 0] = (int8_t)((q4[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; - a[l+16] = (int8_t)((q4[l+16] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; - a[l+32] = (int8_t)((q4[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; - a[l+48] = (int8_t)((q4[l+16] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; - } - int is = 0; - for (int j = 0; j < QK_K/16; ++j) { - int scale = x[i].scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; -#endif -} - -#endif - #if defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) || defined(__loongarch_asx) static const int8_t keven_signs_q2xs[1024] = { 1, 1, 1, 1, 1, 1, 1, 1, -1, 1, 1, 1, 1, 1, 1, -1, 1, -1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, 1, 1, @@ -11564,64 +9120,6 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * const __m256i block_sign_shuffle_1 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_1); const __m256i block_sign_shuffle_2 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_2); -#if QK_K == 64 - static const uint8_t k_bit_helper[16] = { - 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00, - }; - const __m128i bit_helper = _mm_loadu_si128((const __m128i*)k_bit_helper); - const __m128i m511 = _mm_set1_epi16(511); - typedef union { - __m128i vec_index; - uint16_t index[8]; - } index_t; - - index_t idx; - __m256 accumf = _mm256_setzero_ps(); - for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const __m128i q2_data = _mm_loadu_si128((const __m128i*)x[i].qs); - idx.vec_index = _mm_and_si128(q2_data, m511); - - const __m128i partial_sign_bits = _mm_srli_epi16(q2_data, 9); - const __m128i partial_sign_bits_upper = _mm_srli_epi16(q2_data, 13); - const __m128i partial_sign_bits_for_counting = _mm_xor_si128(partial_sign_bits, partial_sign_bits_upper); - - const __m128i odd_bits = _mm_shuffle_epi8(bit_helper, partial_sign_bits_for_counting); - const __m128i full_sign_bits = _mm_or_si128(partial_sign_bits, odd_bits); - const __m256i full_signs = MM256_SET_M128I(full_sign_bits, full_sign_bits); - - const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)y[i].qs); - const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)(y[i].qs+32)); - - const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[idx.index[3]], iq2xs_grid[idx.index[2]], - iq2xs_grid[idx.index[1]], iq2xs_grid[idx.index[0]]); - const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[idx.index[7]], iq2xs_grid[idx.index[6]], - iq2xs_grid[idx.index[5]], iq2xs_grid[idx.index[4]]); - - __m256i signs; - signs = _mm256_shuffle_epi8(full_signs, block_sign_shuffle_1); - signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask); - const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone)); - - signs = _mm256_shuffle_epi8(full_signs, block_sign_shuffle_2); - signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask); - const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone)); - - const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); - const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); - - const __m256i sc1 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[0] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[0] & 0xf)+1)); - const __m256i sc2 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[1] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[1] & 0xf)+1)); - - const __m256i sum = _mm256_add_epi32(_mm256_madd_epi16(sc1, dot1), _mm256_madd_epi16(sc2, dot2)); - - accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sum), accumf); - - } - - *s = 0.125f * hsum_float_8(accumf); -#else - static const uint8_t k_bit_helper[32] = { 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00, @@ -11719,7 +9217,6 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * } *s = 0.125f * hsum_float_8(accumf); -#endif #elif defined(__loongarch_asx) const __m256i mone = __lasx_xvreplgr2vr_b(1); @@ -11740,62 +9237,6 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * const __m256i block_sign_shuffle_1 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_1, 0); const __m256i block_sign_shuffle_2 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_2, 0); -#if QK_K == 64 - static const uint8_t k_bit_helper[16] = { - 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00, - }; - const __m128i bit_helper = __lsx_vld((const __m128i*)k_bit_helper, 0); - const __m128i m511 = __lsx_vreplgr2vr_h(511); - typedef union { - __m128i vec_index; - uint16_t index[8]; - } index_t; - - index_t idx; - __m256 accumf = (__m256)__lasx_xvldi(0); - for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const __m128i q2_data = __lsx_vld((const __m128i*)x[i].qs, 0); - idx.vec_index = __lsx_vand_v(q2_data, m511); - - const __m128i partial_sign_bits = __lsx_vsrli_h(q2_data, 9); - const __m128i partial_sign_bits_upper = __lsx_vsrli_h(q2_data, 13); - const __m128i partial_sign_bits_for_counting = __lsx_vxor_v(partial_sign_bits, partial_sign_bits_upper); - - const __m128i odd_bits = lsx_shuffle_b(bit_helper, partial_sign_bits_for_counting); - const __m128i full_sign_bits = __lsx_vor_v(partial_sign_bits, odd_bits); - const __m256i full_signs = lasx_insertf128(full_sign_bits, full_sign_bits); - - const __m256i q8_1 = __lasx_xvld((const __m256i *)y[i].qs, 0); - const __m256i q8_2 = __lasx_xvld((const __m256i *)(y[i].qs+32), 0); - - const __m256i q2_1 = lasx_set_d(iq2xs_grid[idx.index[3]], iq2xs_grid[idx.index[2]], - iq2xs_grid[idx.index[1]], iq2xs_grid[idx.index[0]]); - const __m256i q2_2 = lasx_set_d(iq2xs_grid[idx.index[7]], iq2xs_grid[idx.index[6]], - iq2xs_grid[idx.index[5]], iq2xs_grid[idx.index[4]]); - __m256i signs; - signs = lasx_shuffle_b(full_signs, block_sign_shuffle_1); - signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask); - const __m256i q8s_1 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_1); - - signs = lasx_shuffle_b(full_signs, block_sign_shuffle_2); - signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask); - const __m256i q8s_2 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_2); - - const __m256i dot1 = lasx_maddubs_h(q2_1, q8s_1); - const __m256i dot2 = lasx_maddubs_h(q2_2, q8s_2); - - const __m256i sc1 = lasx_insertf128(_mm_set1_epi16(2*(x[i].scales[0] >> 4)+1), __lsx_vreplgr2vr_h(2*(x[i].scales[0] & 0xf)+1)); - const __m256i sc2 = lasx_insertf128(_mm_set1_epi16(2*(x[i].scales[1] >> 4)+1), __lsx_vreplgr2vr_h(2*(x[i].scales[1] & 0xf)+1)); - - const __m256i sum = __lasx_xvadd_w(lasx_madd_h(sc1, dot1), lasx_madd_h(sc2, dot2)); - - accumf = __lasx_vfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sum), accumf); - } - - *s = 0.125f * hsum_float_8(accumf); -#else - static const uint8_t k_bit_helper[32] = { 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00, @@ -11893,9 +9334,6 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * } *s = 0.125f * hsum_float_8(accumf); -#endif - - #elif defined(__POWER9_VECTOR__) vector float vsumf0 = vec_splats(0.0f); vector float vsumf1 = vec_splats(0.0f); @@ -12748,10 +10186,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * ggml_int8x16x4_t q8b; vec_index_t idx; -#if QK_K == 256 uint32_t scales32[2]; const uint8_t * scales8 = (const uint8_t *)scales32; -#endif float sumf = 0; for (int i = 0; i < nb; ++i) { @@ -12761,11 +10197,9 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * const uint16_t * restrict signs = (const uint16_t *)x[i].signs; const int8_t * restrict q8 = y[i].qs; -#if QK_K == 256 memcpy(scales32, x[i].scales, 4); scales32[1] = (((scales32[0] >> 4) & 0x0f0f0f0f) << 1) | 0x01010101; scales32[0] = ((scales32[0] & 0x0f0f0f0f) << 1) | 0x01010101; -#endif int sumi1 = 0, sumi2 = 0; for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { @@ -12806,13 +10240,9 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]); const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]); -#if QK_K == 256 + sumi1 += vaddvq_s32(p1) * scales8[ib32/2+0]; sumi2 += vaddvq_s32(p2) * scales8[ib32/2+4]; -#else - sumi1 += vaddvq_s32(p1) * (1 + 2*(x[i].scales[ib32/2] & 0xf)); - sumi2 += vaddvq_s32(p2) * (1 + 2*(x[i].scales[ib32/2] >> 4)); -#endif } sumf += d*(sumi1 + sumi2); } @@ -13476,17 +10906,10 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void const int nb = n / QK_K; -#if QK_K != 64 iq1m_scale_t scale; -#endif #if defined __ARM_NEON - -#if QK_K == 64 - const int32x4_t mask = vdupq_n_s32(0xf); -#else const int32x4_t mask = vdupq_n_s32(0x7); -#endif const int32x4_t mone = vdupq_n_s32(1); const int32x4_t mzero = vdupq_n_s32(0); @@ -13510,9 +10933,7 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void const uint8_t * qh = x[i].qh; const uint16_t * sc = (const uint16_t *)x[i].scales; -#if QK_K != 64 scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); -#endif int32x4_t sumi1 = mzero; int32x4_t sumi2 = mzero; @@ -13541,11 +10962,8 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void const int32x4_t p4 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[2]], q8b.val[2]), ggml_vdotq_s32(mzero, deltas.val[aux8[3]], q8b.val[3])); const int32x4_t p34 = vpaddq_s32(p3, p4); -#if QK_K == 64 - int32x4_t scales_4 = ggml_vld1q_u32(sc[0] >> 0, sc[0] >> 4, sc[0] >> 8, sc[0] >> 12); -#else int32x4_t scales_4 = ggml_vld1q_u32(sc[ib/2] >> 0, sc[ib/2] >> 3, sc[ib/2] >> 6, sc[ib/2] >> 9); -#endif + scales_4 = vaddq_s32(vshlq_n_s32(vandq_s32(scales_4, mask), 1), mone); sumi1 = vmlaq_s32(sumi1, scales_4, p12); @@ -13555,22 +10973,14 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void } -#if QK_K == 64 - sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2)); -#else sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2)); -#endif } *s = sumf; #elif defined __AVX2__ -#if QK_K == 64 - const __m256i mask = _mm256_set1_epi16(0xf); -#else const __m256i mask = _mm256_set1_epi16(0x7); -#endif const __m256i mone = _mm256_set1_epi16(1); __m256 accum1 = _mm256_setzero_ps(); @@ -13582,9 +10992,7 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void const uint8_t * qh = x[i].qh; const uint16_t * sc = (const uint16_t *)x[i].scales; -#if QK_K != 64 scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); -#endif __m256i sumi1 = _mm256_setzero_si256(); __m256i sumi2 = _mm256_setzero_si256(); @@ -13614,13 +11022,10 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void const __m256i dot3 = mul_add_epi8(delta1, q8b_1); const __m256i dot4 = mul_add_epi8(delta2, q8b_2); -#if QK_K == 64 - __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[0] >> 4), _mm_set1_epi16(sc[0] >> 0)); - __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[0] >> 12), _mm_set1_epi16(sc[0] >> 8)); -#else + __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 3), _mm_set1_epi16(sc[ib/2] >> 0)); __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 9), _mm_set1_epi16(sc[ib/2] >> 6)); -#endif + scale1 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale1, mask), 1), mone); scale2 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale2, mask), 1), mone); const __m256i p1 = _mm256_madd_epi16(dot1, scale1); @@ -13634,14 +11039,10 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void qs += 8; qh += 4; } -#if QK_K == 64 - const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d)); -#else const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16)); -#endif + accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1); accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2); - } *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2); @@ -13658,9 +11059,7 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void const uint8_t * qh = x[i].qh; const uint16_t * sc = (const uint16_t *)x[i].scales; -#if QK_K != 64 scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); -#endif int sumi1 = 0, sumi2 = 0; for (int ib = 0; ib < QK_K/32; ++ib) { @@ -13680,24 +11079,17 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void sum1[l/2] += lsum1; sum2[l/2] += lsum2*delta[l]; } -#if QK_K == 64 - const int ls1 = 2*((sc[0] >> (8*(ib%2)+0)) & 0xf) + 1; - const int ls2 = 2*((sc[0] >> (8*(ib%2)+4)) & 0xf) + 1; -#else + const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1; const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1; -#endif + sumi1 += sum1[0] * ls1 + sum1[1] * ls2; sumi2 += sum2[0] * ls1 + sum2[1] * ls2; qs += 4; qh += 2; } -#if QK_K == 64 - sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); -#else sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); -#endif } *s = sumf; @@ -13885,9 +11277,6 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * UNUSED(by); UNUSED(bs); assert(n % QK_K == 0); -#if QK_K == 64 - ggml_vec_dot_iq4_nl_q8_0(n, s, bs, vx, bx, vy, by, nrc); -#else const block_iq4_xs * restrict x = vx; const block_q8_K * restrict y = vy; @@ -14180,7 +11569,6 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * } *s = sumf; #endif -#endif } // ================================ IQ2 quantization ============================================= @@ -15998,10 +13386,6 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy const float * xx; for (int ibl = 0; ibl < nbl; ++ibl) { - -#if QK_K == 64 - y[ibl].d = GGML_FP32_TO_FP16(0.f); -#endif memset(y[ibl].qs, 0, QK_K/8); memset(y[ibl].qh, 0, QK_K/16); memset(y[ibl].scales, 0, QK_K/32); @@ -16176,22 +13560,13 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy } uint16_t * sc = (uint16_t *)y[ibl].scales; -#if QK_K == 64 - float d = max_scale/31; -#else float d = max_scale/15; -#endif float id = 1/d; float sumqx_f = 0, sumq2_f = 0; for (int ib = 0; ib < QK_K/block_size; ++ib) { int l = nearest_int(0.5f*(id*scales[ib+0]-1)); -#if QK_K == 64 - l = MAX(0, MIN(15, l)); - sc[ib/4] |= (l << 4*(ib%4)); -#else l = MAX(0, MIN(7, l)); sc[ib/4] |= (l << 3*(ib%4)); -#endif y[ibl].qh[ib] |= masks[shifts[ib]]; const float * xb = xbl + block_size*ib; if (quant_weights) { @@ -16214,14 +13589,10 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy } if (sumq2_f > 0) d = sumqx_f/sumq2_f; s.f16 = GGML_FP32_TO_FP16(d*1.1125f); // 1.1125f is another fudge factor. Don't ask me why it is needed. -#if QK_K == 64 - y[ibl].d = s.f16; -#else sc[0] |= ((s.u16 & 0x000f) << 12); sc[1] |= ((s.u16 & 0x00f0) << 8); sc[2] |= ((s.u16 & 0x0f00) << 4); sc[3] |= ((s.u16 & 0xf000) << 0); -#endif } } @@ -16410,9 +13781,6 @@ void quantize_row_iq4_nl_reference(const float * restrict x, block_iq4_nl * rest } size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { -#if QK_K == 64 - return quantize_iq4_nl(src, dst, nrow, n_per_row, quant_weights); -#else GGML_ASSERT(n_per_row%QK_K == 0); int64_t nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -16430,7 +13798,6 @@ size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t qrow += nblock*sizeof(block_iq4_xs); } return nrow * nblock * sizeof(block_iq4_xs); -#endif } void quantize_row_iq4_xs(const float * restrict x, void * restrict vy, int64_t k) { @@ -16842,19 +14209,11 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte } break; case GGML_TYPE_Q4_K: { - #ifdef GGML_QKK_64 - VALIDATE_ROW_DATA_DM_F16_IMPL(block_q4_K, data, nb, d[0], d[1]); - #else VALIDATE_ROW_DATA_DM_F16_IMPL(block_q4_K, data, nb, d, dmin); - #endif } break; case GGML_TYPE_Q5_K: { - #ifdef GGML_QKK_64 - VALIDATE_ROW_DATA_D_F16_IMPL(block_q5_K, data, nb); - #else VALIDATE_ROW_DATA_DM_F16_IMPL(block_q5_K, data, nb, d, dmin); - #endif } break; case GGML_TYPE_Q6_K: { @@ -16877,18 +14236,12 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte { const block_iq1_m * q = (const block_iq1_m *) data; for (size_t i = 0; i < nb; ++i) { - #if QK_K == 64 - if (!validate_fp16(q[i].d, i)) { - return false; - } - #else iq1m_scale_t scale; const uint16_t * sc = (const uint16_t *)q[i].scales; scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); if (!validate_fp16(scale.f16, i)) { return false; } - #endif } } break; case GGML_TYPE_IQ2_XXS: @@ -16913,12 +14266,9 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte VALIDATE_ROW_DATA_D_F16_IMPL(block_iq3_s, data, nb); } break; case GGML_TYPE_IQ4_XS: - #if QK_K != 64 { VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_xs, data, nb); } break; - #endif - // with QK_K == 64, iq4_xs is iq4_nl case GGML_TYPE_IQ4_NL: { VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb); diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index f486b6c0a..496ec61c3 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -4197,7 +4197,6 @@ static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restri const block_q2_K * x = (const block_q2_K *) vx; const int tid = item_ct1.get_local_id(2); -#if QK_K == 256 const int n = tid/32; const int l = tid - 32*n; const int is = 8*n + l/16; @@ -4211,18 +4210,6 @@ static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restri y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4); y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4); y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4); -#else - const int is = tid/16; // 0 or 1 - const int il = tid%16; // 0...15 - const uint8_t q = x[i].qs[il] >> (2*is); - dst_t * y = yy + i*QK_K + 16*is + il; - - float dall = x[i].dm[0]; - float dmin = x[i].dm[1]; - y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); - y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4); -#endif - } template @@ -4232,7 +4219,6 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri const int i = item_ct1.get_group(2); const block_q3_K * x = (const block_q3_K *) vx; -#if QK_K == 256 const int r = item_ct1.get_local_id(2) / 4; const int tid = r/2; const int is0 = r%2; @@ -4256,31 +4242,8 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri const uint8_t * hm = x[i].hmask; for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)); -#else - const int tid = item_ct1.get_local_id(2); - const int is = tid/16; // 0 or 1 - const int il = tid%16; // 0...15 - const int im = il/8; // 0...1 - const int in = il%8; // 0...7 - - dst_t * y = yy + i*QK_K + 16*is + il; - - const uint8_t q = x[i].qs[il] >> (2*is); - const uint8_t h = x[i].hmask[in] >> (2*is + im); - const float d = (float)x[i].d; - - if (is == 0) { - y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); - y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); - } else { - y[ 0] = d * ((x[i].scales[0] >> 4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); - y[32] = d * ((x[i].scales[1] >> 4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); - } -#endif - } -#if QK_K == 256 static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { if (j < 4) { d = q[j] & 63; m = q[j + 4] & 63; @@ -4289,7 +4252,6 @@ static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8 m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); } } -#endif template static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy, @@ -4298,7 +4260,6 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri const int i = item_ct1.get_group(2); -#if QK_K == 256 // assume 32 threads const int tid = item_ct1.get_local_id(2); const int il = tid/8; @@ -4322,15 +4283,6 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri y[l + 0] = d1 * (q[l] & 0xF) - m1; y[l +32] = d2 * (q[l] >> 4) - m2; } -#else - const int tid = item_ct1.get_local_id(2); - const uint8_t * q = x[i].qs; - dst_t * y = yy + i*QK_K; - const float d = (float)x[i].dm[0]; - const float m = (float)x[i].dm[1]; - y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4); - y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4); -#endif } template @@ -4340,7 +4292,6 @@ static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restri const int i = item_ct1.get_group(2); -#if QK_K == 256 // assume 64 threads - this is very slightly better than the one below const int tid = item_ct1.get_local_id(2); const int il = tid/16; // il is in 0...3 @@ -4367,18 +4318,6 @@ static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restri hm <<= 1; y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2; y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2; -#else - const int tid = item_ct1.get_local_id(2); - const uint8_t q = x[i].qs[tid]; - const int im = tid/8; // 0...3 - const int in = tid%8; // 0...7 - const int is = tid/16; // 0 or 1 - const uint8_t h = x[i].qh[in] >> im; - const float d = x[i].d; - dst_t * y = yy + i*QK_K + tid; - y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16)); - y[32] = d * x[i].scales[is+2] * ((q >> 4) - ((h >> 4) & 1 ? 0 : 16)); -#endif } template @@ -4387,7 +4326,6 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri const block_q6_K * x = (const block_q6_K *) vx; const int i = item_ct1.get_group(2); -#if QK_K == 256 // assume 64 threads - this is very slightly better than the one below const int tid = item_ct1.get_local_id(2); @@ -4407,24 +4345,6 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32); y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32); y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32); -#else - - // assume 32 threads - const int tid = item_ct1.get_local_id(2); - const int ip = tid/16; // 0 or 1 - const int il = tid - 16*ip; // 0...15 - - dst_t * y = yy + i*QK_K + 16*ip + il; - - const float d = x[i].d; - - const uint8_t ql = x[i].ql[16*ip + il]; - const uint8_t qh = x[i].qh[il] >> (2*ip); - const int8_t * sc = x[i].scales; - - y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32); - y[32] = d * sc[ip+2] * ((int8_t)((ql >> 4) | (((qh >> 4) & 3) << 4)) - 32); -#endif } template @@ -4438,7 +4358,6 @@ static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __res const block_iq2_xxs * x = (const block_iq2_xxs *) vx; const int tid = item_ct1.get_local_id(2); -#if QK_K == 256 const int il = tid/8; // 0...3 const int ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -4449,10 +4368,6 @@ static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __res const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f; const uint8_t signs = ksigns_iq2xs_ptr[(aux32 >> 7*il) & 127]; for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs_ptr[j] ? -1.f : 1.f); -#else - assert(false); -#endif - } template @@ -4466,7 +4381,6 @@ static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __rest const block_iq2_xs * x = (const block_iq2_xs *) vx; const int tid = item_ct1.get_local_id(2); -#if QK_K == 256 const int il = tid/8; // 0...3 const int ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -4475,10 +4389,6 @@ static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __rest const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; const uint8_t signs = ksigns_iq2xs[q2[il] >> 9]; for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); -#else - assert(false); -#endif - } template @@ -4490,7 +4400,6 @@ dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy, const block_iq2_s * x = (const block_iq2_s *) vx; const int tid = item_ct1.get_local_id(2); -#if QK_K == 256 const int il = tid/8; // 0...3 const int ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -4498,13 +4407,9 @@ dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy, const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; const uint8_t signs = x[i].qs[QK_K/8+4*ib+il]; #pragma unroll - for (int j = 0; j < 8; ++j) + for (int j = 0; j < 8; ++j) { y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); -#else - assert(false); - -#endif - + } } template @@ -4518,7 +4423,6 @@ static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __res const block_iq3_xxs * x = (const block_iq3_xxs *) vx; const int tid = item_ct1.get_local_id(2); -#if QK_K == 256 const int il = tid/8; // 0...3 const int ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -4533,10 +4437,6 @@ static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __res y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); } -#else - assert(false); -#endif - } template @@ -4549,7 +4449,6 @@ dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy, const block_iq3_s * x = (const block_iq3_s *) vx; const int tid = item_ct1.get_local_id(2); -#if QK_K == 256 const int il = tid/8; // 0...3 const int ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -4563,10 +4462,6 @@ dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy, y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); } -#else - assert(false); -#endif - } template @@ -4579,7 +4474,6 @@ dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy, const block_iq1_s * x = (const block_iq1_s *) vx; const int tid = item_ct1.get_local_id(2); -#if QK_K == 256 const int il = tid/8; // 0...3 const int ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -4593,10 +4487,6 @@ dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy, for (int j = 0; j < 8; ++j) { y[j] = d * (q[j] + delta); } -#else - assert(false); -#endif - } template @@ -4609,7 +4499,6 @@ dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy, const block_iq1_m * x = (const block_iq1_m *) vx; const int tid = item_ct1.get_local_id(2); -#if QK_K == 256 const int il = tid/8; // 0...3 const int ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; @@ -4627,10 +4516,6 @@ dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy, for (int j = 0; j < 8; ++j) { y[j] = d * (q[j] + delta); } -#else - assert(false); -#endif - } template @@ -4704,7 +4589,6 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, float tmp = 0; // partial sum for thread in warp -#if QK_K == 256 const int tid = item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...15 const int ix = @@ -4755,42 +4639,6 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, tmp += dall * sum1 - dmin * sum2; } -#else - const int tid = item_ct1.get_local_id(2) / - (2 * K_QUANTS_PER_ITERATION); // 0...15 or 0...7 - const int ix = item_ct1.get_local_id(2) % - (2 * K_QUANTS_PER_ITERATION); // 0....1 or 0...3 - const int offset = tid * K_QUANTS_PER_ITERATION; - - uint32_t uaux[2]; - const uint8_t * d = (const uint8_t *)uaux; - - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + offset; - const uint8_t * q = x[i].qs + offset; - const uint32_t * s = (const uint32_t *)x[i].scales; - - uaux[0] = s[0] & 0x0f0f0f0f; - uaux[1] = (s[0] >> 4) & 0x0f0f0f0f; - - const sycl::float2 dall = - x[i].dm.convert(); - - float sum1 = 0, sum2 = 0; - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - const uint8_t ql = q[l]; - sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3) - + y[l+16] * d[1] * ((ql >> 2) & 3) - + y[l+32] * d[2] * ((ql >> 4) & 3) - + y[l+48] * d[3] * ((ql >> 6) & 3); - sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7]; - } - tmp += dall.x() * sum1 - dall.y() * sum2; - } - -#endif // sum up partial sums and write back result #pragma unroll @@ -4828,8 +4676,6 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx, float tmp = 0; // partial sum for thread in warp -#if QK_K == 256 - const uint16_t kmask1 = 0x0303; const uint16_t kmask2 = 0x0f0f; @@ -4882,34 +4728,6 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx, tmp += d * sum; } -#else - - const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 - const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 - const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14 - const int in = offset/8; // 0 or 1 - const int im = offset%8; // 0...7 - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + offset; - const uint8_t * q = x[i].qs + offset; - const uint8_t * s = x[i].scales; - - const float dall = (float)x[i].d; - - float sum = 0; - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - const uint8_t hl = x[i].hmask[im+l] >> in; - const uint8_t ql = q[l]; - sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4)) - + y[l+16] * dall * ((s[0] >> 4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4)) - + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4)) - + y[l+48] * dall * ((s[1] >> 4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4)); - } - tmp += sum; - } -#endif // sum up partial sums and write back result #pragma unroll @@ -4944,7 +4762,6 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx, const block_q4_K * x = (const block_q4_K *)vx + ib0; -#if QK_K == 256 const uint16_t kmask1 = 0x3f3f; const uint16_t kmask2 = 0x0f0f; const uint16_t kmask3 = 0xc0c0; @@ -5033,36 +4850,6 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx, #endif } -#else - const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 - const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); - - const int step = tid * K_QUANTS_PER_ITERATION; - - uint16_t aux16[2]; - const uint8_t * s = (const uint8_t *)aux16; - - float tmp = 0; - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - const uint8_t * q = x[i].qs + step; - const float * y = yy + i*QK_K + step; - const uint16_t * a = (const uint16_t *)x[i].scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - const float d = (float)x[i].dm[0]; - const float m = (float)x[i].dm[1]; - float sum = 0.f; - for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { - sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2]) - + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2]) - + y[j+32] * (d * s[1] * (q[j+ 0] >> 4) - m * s[3]) - + y[j+48] * (d * s[1] * (q[j+16] >> 4) - m * s[3]); - } - tmp += sum; - } - -#endif // sum up partial sums and write back result #pragma unroll @@ -5097,7 +4884,6 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx, float tmp = 0; // partial sum for thread in warp -#if QK_K == 256 const uint16_t kmask1 = 0x3f3f; const uint16_t kmask2 = 0x0f0f; const uint16_t kmask3 = 0xc0c0; @@ -5174,30 +4960,6 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx, dmin * smin; } -#else - const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 - const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); - const int step = tid * K_QUANTS_PER_ITERATION; - const int im = step/8; - const int in = step%8; - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - const uint8_t * q = x[i].qs + step; - const int8_t * s = x[i].scales; - const float * y = yy + i*QK_K + step; - const float d = x[i].d; - float sum = 0.f; - for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { - const uint8_t h = x[i].qh[in+j] >> im; - sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16)) - + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16)) - + y[j+32] * d * s[2] * ((q[j+ 0] >> 4) - ((h >> 4) & 1 ? 0 : 16)) - + y[j+48] * d * s[3] * ((q[j+16] >> 4) - ((h >> 6) & 1 ? 0 : 16)); - } - tmp += sum; - } -#endif - // sum up partial sums and write back result #pragma unroll for (int mask = 16; mask > 0; mask >>= 1) { @@ -5224,8 +4986,6 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa const block_q6_K * x = (const block_q6_K *)vx + ib0; -#if QK_K == 256 - const int tid = item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 const int ix = @@ -5282,37 +5042,6 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa } -#else - - const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...7 - const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); // 0...3 - - const int step = tid * K_QUANTS_PER_ITERATION; - - float tmp = 0; // partial sum for thread in warp - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + step; - const uint8_t * ql = x[i].ql + step; - const uint8_t * qh = x[i].qh + step; - const int8_t * s = x[i].scales; - - const float d = x[i+0].d; - - float sum = 0; - for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { - sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32) - + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32) - + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >> 4) | ((qh[j] & 0x30) >> 0)) - 32) - + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >> 4) | ((qh[j] & 0xc0) >> 2)) - 32); - } - tmp += sum; - - } - -#endif - // sum up partial sums and write back result #pragma unroll for (int mask = 16; mask > 0; mask >>= 1) { @@ -6857,7 +6586,6 @@ static __dpct_inline__ float vec_dot_q4_K_q8_1(const void *__restrict__ vbq, const block_q8_1 *__restrict__ bq8_1, const int &iqs) { -#ifndef GGML_QKK_64 const block_q4_K * bq4_K = (const block_q4_K *) vbq; int v[2]; @@ -6899,52 +6627,6 @@ vec_dot_q4_K_q8_1(const void *__restrict__ vbq, } return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8); - -#else - -#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics - const block_q4_K * bq4_K = (const block_q4_K *) vbq; - - float sumf_d = 0.0f; - float sumf_m = 0.0f; - - uint16_t aux16[2]; - const uint8_t * s = (const uint8_t *)aux16; - - const uint16_t * a = (const uint16_t *)bq4_K->scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - - const float dall = bq4_K->dm[0]; - const float dmin = bq4_K->dm[1]; - - const float d8_1 = bq8_1[0].ds[0]; - const float d8_2 = bq8_1[1].ds[1]; - - const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); - const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); - const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); - const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); - - const int * q4 = (const int *)bq4_K->qs + (iqs/2); - const int v1 = q4[0]; - const int v2 = q4[4]; - - const int dot1 = dpct::dp4a(ui2, v2 & 0x0f0f0f0f, dpct::dp4a(ui1, v1 & 0x0f0f0f0f, 0)); - const int dot2 = dpct::dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, dpct::dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0)); - const int dot3 = dpct::dp4a(0x01010101, ui2, dpct::dp4a(0x01010101, ui1, 0)); - const int dot4 = dpct::dp4a(0x01010101, ui4, dpct::dp4a(0x01010101, ui3, 0)); - - sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]); - sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]); - - return dall * sumf_d - dmin * sumf_m; - -#else - bad_arch(); -#endif // __SYCL_ARCH__ >= VER_4VEC - -#endif } template @@ -7003,11 +6685,7 @@ load_tiles_q4_K(const void *__restrict__ vx, int *__restrict__ x_ql, const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd; -#if QK_K == 256 x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm; -#else - x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]}; -#endif } #pragma unroll @@ -7050,7 +6728,6 @@ static __dpct_inline__ float vec_dot_q5_K_q8_1(const void *__restrict__ vbq, const block_q8_1 *__restrict__ bq8_1, const int &iqs) { -#ifndef GGML_QKK_64 const block_q5_K * bq5_K = (const block_q5_K *) vbq; int vl[2]; @@ -7092,48 +6769,6 @@ vec_dot_q5_K_q8_1(const void *__restrict__ vbq, } return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8); - -#else - -#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics - const block_q5_K * bq5_K = (const block_q5_K *) vbq; - - const int8_t * s = bq5_K->scales; - - const float d = bq5_K->d; - - const float d8_1 = bq8_1[0].ds[0]; - const float d8_2 = bq8_1[1].ds[1]; - - const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); - const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); - const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); - const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); - - const int * ql = (const int *)bq5_K->qs + (iqs/2); - const int vl1 = ql[0]; - const int vl2 = ql[4]; - - const int step = 4 * (iqs/2); // 0, 4, 8, 12 - const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6 - const int in = step%8; // 0, 4, 0, 4 - const int vh = (*((const int *)(bq5_K->qh + in))) >> im; - - const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f); - const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f); - const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f); - const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f); - - const float sumf_d = d8_1 * (dpct::dp4a(ui1, v1, 0) * s[0] + dpct::dp4a(ui2, v2, 0) * s[1]) - + d8_2 * (dpct::dp4a(ui3, v3, 0) * s[2] + dpct::dp4a(ui4, v4, 0) * s[3]); - - return d * sumf_d; - -#else - bad_arch(); -#endif // __SYCL_ARCH__ >= VER_4VEC - -#endif } template @@ -7205,9 +6840,7 @@ load_tiles_q5_K(const void *__restrict__ vx, int *__restrict__ x_ql, const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd; -#if QK_K == 256 x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm; -#endif } #pragma unroll @@ -7387,7 +7020,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq, const block_q8_1 *__restrict__ bq8_1, const int &iqs, const uint64_t *iq2xxs_grid, const uint8_t *ksigns_iq2xs, const uint8_t *kmask_iq2xs) { -#if QK_K == 256 const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq; #if QR2_XXS == 8 @@ -7428,10 +7060,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq, } return d * (sumi1 + sumi2); #endif -#else - assert(false); - return 0.f; -#endif } static __dpct_inline__ float @@ -7440,7 +7068,6 @@ vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq, const uint64_t *iq2xs_grid, const uint64_t *ksigns64) { #if DPCT_COMPATIBILITY_TEMP >= \ MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq; const int ib32 = iqs; @@ -7478,16 +7105,11 @@ vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq, assert(false); return 0.f; #endif -#else - assert(false); - return 0.f; -#endif } static __dpct_inline__ float vec_dot_iq2_s_q8_1(const void *__restrict__ vbq, const block_q8_1 *__restrict__ bq8_1, const int &iqs) { -#if QK_K == 256 const block_iq2_s * bq2 = (const block_iq2_s *) vbq; const int ib32 = iqs; @@ -7531,9 +7153,6 @@ vec_dot_iq2_s_q8_1(const void *__restrict__ vbq, } const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f; return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); -#else - assert(false); -#endif } static __dpct_inline__ float @@ -7542,7 +7161,6 @@ vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq, const uint32_t *iq3xxs_grid, const uint64_t *ksigns64) { #if DPCT_COMPATIBILITY_TEMP >= \ MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq; const int ib32 = iqs; @@ -7570,17 +7188,12 @@ vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq, assert(false); return 0.f; #endif -#else - assert(false); - return 0.f; -#endif } static __dpct_inline__ float vec_dot_iq3_s_q8_1(const void *__restrict__ vbq, const block_q8_1 *__restrict__ bq8_1, const int &iqs, const uint32_t *iq3s_grid) { -#if QK_K == 256 const block_iq3_s * bq2 = (const block_iq3_s *) vbq; const int ib32 = iqs; @@ -7609,16 +7222,12 @@ vec_dot_iq3_s_q8_1(const void *__restrict__ vbq, (1 + 2 * ((bq2->scales[ib32 / 2] >> 4 * (ib32 % 2)) & 0xf)) * bq8_1[ib32].ds[0]; return d * sumi; -#else - assert(false); -#endif } static __dpct_inline__ float vec_dot_iq1_s_q8_1(const void *__restrict__ vbq, const block_q8_1 *__restrict__ bq8_1, const int &iqs, const uint32_t *iq1s_grid_gpu) { -#if QK_K == 256 const block_iq1_s * bq1 = (const block_iq1_s *) vbq; const int ib32 = iqs; @@ -7637,15 +7246,11 @@ vec_dot_iq1_s_q8_1(const void *__restrict__ vbq, const float d = d1q * bq8_1[ib32].ds[0]; const float m = d1q * bq8_1[ib32].ds[1]; return d * sumi + m * delta; -#else - assert(false); -#endif } static __dpct_inline__ float vec_dot_iq1_m_q8_1(const void *__restrict__ vbq, const block_q8_1 *__restrict__ bq8_1, const int &iqs) { -#if QK_K == 256 const block_iq1_m * bq1 = (const block_iq1_m *) vbq; const int ib32 = iqs; @@ -7670,9 +7275,6 @@ vec_dot_iq1_m_q8_1(const void *__restrict__ vbq, scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); const float d = (float)scale.f16 * bq8_1[ib32].ds[0]; return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1)); -#else - assert(false); -#endif } static __dpct_inline__ void get_int_from_table_16(const uint32_t &q4, @@ -7720,7 +7322,6 @@ static __dpct_inline__ float vec_dot_iq4_xs_q8_1(const void *__restrict__ vbq, const block_q8_1 *__restrict__ bq8_1, const int &iqs) { -#if QK_K == 256 const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq; const uint8_t * values = (const uint8_t *)kvalues_iq4nl; @@ -7738,9 +7339,6 @@ vec_dot_iq4_xs_q8_1(const void *__restrict__ vbq, sumi2 = dpct::dp4a(v2, q8[j + 4], sumi2); } return d * (sumi1 + sumi2); -#else - assert(false); -#endif } template static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k, dpct::queue_ptr stream) { const int nb = k / QK_K; -#if QK_K == 256 { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); @@ -10215,27 +9812,12 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k, dequantize_block_q2_K(vx, y, item_ct1); }); } -#else - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q2_K(vx, y, item_ct1); - }); - } - -#endif } template static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k, dpct::queue_ptr stream) { const int nb = k / QK_K; -#if QK_K == 256 { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); @@ -10247,19 +9829,6 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k, dequantize_block_q3_K(vx, y, item_ct1); }); } -#else - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q3_K(vx, y, item_ct1); - }); - } -#endif } template @@ -10320,7 +9889,6 @@ template static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k, dpct::queue_ptr stream) { const int nb = k / QK_K; -#if QK_K == 256 { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); @@ -10332,27 +9900,12 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k, dequantize_block_q5_K(vx, y, item_ct1); }); } -#else - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q5_K(vx, y, item_ct1); - }); - } - -#endif } template static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k, dpct::queue_ptr stream) { const int nb = k / QK_K; -#if QK_K == 256 { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); @@ -10364,20 +9917,6 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k, dequantize_block_q6_K(vx, y, item_ct1); }); } -#else - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q6_K(vx, y, item_ct1); - }); - } - -#endif } template @@ -10529,9 +10068,6 @@ template static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int k, dpct::queue_ptr stream) { const int nb = (k + QK_K - 1) / QK_K; -#if QK_K == 64 - dequantize_row_iq4_nl_sycl(vx, y, k, stream); -#else { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); @@ -10546,7 +10082,6 @@ static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int k, }); }); } -#endif } @@ -12051,8 +11586,6 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, const int nrows_y, const int nrows_dst, dpct::queue_ptr stream) try { -#if QK_K == 256 - int id; SYCL_CHECK( CHECK_TRY_ERROR(id = get_current_device_id())); @@ -12167,7 +11700,6 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, }); } } -#endif } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ diff --git a/ggml.c b/ggml.c index d316e3d31..673c47748 100644 --- a/ggml.c +++ b/ggml.c @@ -871,22 +871,14 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { }, [GGML_TYPE_IQ4_XS] = { .type_name = "iq4_xs", -#if QK_K == 64 - .blck_size = QK4_NL, -#else .blck_size = QK_K, -#endif .type_size = sizeof(block_iq4_xs), .is_quantized = true, .to_float = (ggml_to_float_t) dequantize_row_iq4_xs, .from_float = quantize_row_iq4_xs, .from_float_reference = (ggml_from_float_t)quantize_row_iq4_xs_reference, .vec_dot = ggml_vec_dot_iq4_xs_q8_K, -#if QK_K == 64 - .vec_dot_type = GGML_TYPE_Q8_0, -#else .vec_dot_type = GGML_TYPE_Q8_K, -#endif .nrows = 1, }, [GGML_TYPE_Q8_K] = { @@ -22117,11 +22109,7 @@ size_t ggml_quantize_chunk( case GGML_TYPE_IQ1_S: result = quantize_iq1_s (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ1_M: result = quantize_iq1_m (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_NL: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; -#if QK_K == 64 - case GGML_TYPE_IQ4_XS: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; -#else case GGML_TYPE_IQ4_XS: result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; -#endif case GGML_TYPE_F16: { size_t elemsize = sizeof(ggml_fp16_t); diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 42df2e4d0..67e23dcc1 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -905,9 +905,8 @@ class GGUFValueType(IntEnum): raise ValueError(f"Unknown type: {type(val)}") -# Note: Does not support GGML_QKK_64 -QK_K = 256 # Items here are (block size, type size) +QK_K = 256 GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = { GGMLQuantizationType.F32: (1, 4), GGMLQuantizationType.F16: (1, 2), diff --git a/llama.cpp b/llama.cpp index 34137c7ad..37b3d58c6 100644 --- a/llama.cpp +++ b/llama.cpp @@ -26,13 +26,9 @@ #ifdef GGML_USE_METAL # include "ggml-metal.h" #endif -#ifndef QK_K -# ifdef GGML_QKK_64 -# define QK_K 64 -# else -# define QK_K 256 -# endif -#endif + +// TODO: replace with ggml API call +#define QK_K 256 #ifdef __has_include #if __has_include() @@ -14308,8 +14304,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) && use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; - else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) && - (qs.i_attention_wv < qs.n_attention_wv/8 || qs.i_attention_wv >= 7*qs.n_attention_wv/8)) new_type = GGML_TYPE_Q6_K; if (qs.model.type == MODEL_70B) { // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with