2024-06-09 09:42:25 +02:00
#pragma once
2024-03-25 13:50:23 +01:00
#include "common.cuh"
2024-06-05 16:53:00 +02:00
#include "vecdotq.cuh"
2024-06-10 11:45:13 +02:00
#include "mma.cuh"
2024-06-05 16:53:00 +02:00
#include <climits>
#include <cstdint>
2024-06-24 17:43:42 +02:00
#define MMQ_DP4A_MAX_BATCH_SIZE 64 // Max. batch size to use for dp4a MMQ kernels when FP16 tensor cores are available.
2024-06-24 12:41:23 +02:00
typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int & kbx0, const int & i_max, const int & stride);
typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0);
2024-06-20 14:39:21 +02:00
typedef void (*mmq_write_back_t)(const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max);
2024-06-09 09:42:25 +02:00
struct block_q8_1_mmq {
half2 ds[4];
int8_t qs[4*QK8_1];
};
static_assert(sizeof(block_q8_1_mmq) == 4*QK8_1 + 4*sizeof(half2), "Unexpected block_q8_1_mmq size");
static_assert(sizeof(block_q8_1_mmq) == 4*sizeof(block_q8_1), "Unexpected block_q8_1_mmq size");
2024-06-05 16:53:00 +02:00
struct tile_x_sizes {
2024-06-14 18:41:49 +02:00
int qs;
2024-06-05 16:53:00 +02:00
int dm;
int sc;
};
2024-06-24 17:43:42 +02:00
static constexpr int get_mmq_x_max_host(const int cc) {
return int8_mma_available(cc) ? 128 :
#ifdef GGML_CUDA_FORCE_MMQ
cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? 128 : 64;
#else
cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? MMQ_DP4A_MAX_BATCH_SIZE : 64;
#endif // GGML_CUDA_FORCE_MMQ
}
2024-06-05 16:53:00 +02:00
static constexpr __device__ int get_mmq_x_max_device() {
2024-06-24 17:43:42 +02:00
#ifdef INT8_MMA_AVAILABLE
return 128;
#else // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
2024-06-24 17:43:42 +02:00
return 128;
#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
2024-06-05 16:53:00 +02:00
#if __CUDA_ARCH__ >= CC_VOLTA
2024-06-24 17:43:42 +02:00
#ifdef GGML_CUDA_FORCE_MMQ
return MMQ_DP4A_MAX_BATCH_SIZE;
#else // GGML_CUDA_FORCE_MMQ
2024-06-05 16:53:00 +02:00
return 128;
2024-06-24 17:43:42 +02:00
#endif // GGML_CUDA_FORCE_MMQ
#else // __CUDA_ARCH__ >= CC_VOLTA
2024-06-05 16:53:00 +02:00
return 64;
#endif // __CUDA_ARCH__ >= CC_VOLTA
2024-06-24 17:43:42 +02:00
2024-06-05 16:53:00 +02:00
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
2024-06-24 17:43:42 +02:00
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
2024-06-24 17:43:42 +02:00
static constexpr int get_mmq_y_host(const int cc) {
return int8_mma_available(cc) || cc >= CC_VOLTA ? 128 : 64;
}
2024-06-05 16:53:00 +02:00
2024-06-20 14:39:21 +02:00
static constexpr __device__ int get_mmq_y_device() {
2024-06-05 16:53:00 +02:00
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
2024-06-20 14:39:21 +02:00
return 128;
2024-06-05 16:53:00 +02:00
#else
#if __CUDA_ARCH__ >= CC_VOLTA
2024-06-20 14:39:21 +02:00
return 128;
2024-06-05 16:53:00 +02:00
#else
return 64;
#endif // __CUDA_ARCH__ >= CC_VOLTA
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
2024-06-20 14:39:21 +02:00
}
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
#define MMQ_DP4A_TXS_Q4_0 tile_x_sizes{mmq_y*WARP_SIZE + mmq_y, mmq_y*WARP_SIZE/QI4_0 + mmq_y/QI4_0, 0}
#define MMQ_DP4A_TXS_Q4_1 tile_x_sizes{mmq_y*WARP_SIZE + mmq_y, mmq_y*WARP_SIZE/QI4_1 + mmq_y/QI4_1, 0}
#define MMQ_DP4A_TXS_Q5_0 tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI5_0 + mmq_y/QI5_0, 0}
#define MMQ_DP4A_TXS_Q5_1 tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI5_1 + mmq_y/QI5_1, 0}
#define MMQ_DP4A_TXS_Q8_0 tile_x_sizes{mmq_y*WARP_SIZE + mmq_y, mmq_y*WARP_SIZE/QI8_0 + mmq_y/QI8_0, 0}
#define MMQ_DP4A_TXS_Q2_K tile_x_sizes{mmq_y*WARP_SIZE + mmq_y, mmq_y*WARP_SIZE + mmq_y, 0}
#define MMQ_DP4A_TXS_Q3_K tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI3_K + mmq_y/QI3_K, mmq_y*WARP_SIZE/4 + mmq_y/4}
#define MMQ_DP4A_TXS_Q4_K tile_x_sizes{mmq_y*WARP_SIZE + mmq_y, mmq_y*WARP_SIZE/QI4_K + mmq_y/QI4_K, mmq_y*WARP_SIZE/8 + mmq_y/8}
#define MMQ_DP4A_TXS_Q5_K tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI5_K + mmq_y/QI5_K, mmq_y*WARP_SIZE/8 + mmq_y/8}
#define MMQ_DP4A_TXS_Q6_K tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI6_K + mmq_y/QI6_K, mmq_y*WARP_SIZE/8 + mmq_y/8}
static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int mmq_y) {
return type == GGML_TYPE_Q4_0 ? MMQ_DP4A_TXS_Q4_0 :
type == GGML_TYPE_Q4_1 ? MMQ_DP4A_TXS_Q4_1 :
type == GGML_TYPE_Q5_0 ? MMQ_DP4A_TXS_Q5_0 :
type == GGML_TYPE_Q5_1 ? MMQ_DP4A_TXS_Q5_1 :
type == GGML_TYPE_Q8_0 ? MMQ_DP4A_TXS_Q8_0 :
type == GGML_TYPE_Q2_K ? MMQ_DP4A_TXS_Q2_K :
type == GGML_TYPE_Q3_K ? MMQ_DP4A_TXS_Q3_K :
type == GGML_TYPE_Q4_K ? MMQ_DP4A_TXS_Q4_K :
type == GGML_TYPE_Q5_K ? MMQ_DP4A_TXS_Q5_K :
type == GGML_TYPE_Q6_K ? MMQ_DP4A_TXS_Q6_K :
tile_x_sizes{0, 0, 0};
2024-06-05 16:53:00 +02:00
}
2024-06-24 12:41:23 +02:00
#define MMQ_MMA_TILE_X_K_Q4_0 (1*WARP_SIZE + WARP_SIZE/QI4_0 + 4)
#define MMQ_MMA_TILE_X_K_Q4_1 (1*WARP_SIZE + WARP_SIZE/QI4_1 + 4)
#define MMQ_MMA_TILE_X_K_Q5_0 (2*WARP_SIZE + WARP_SIZE/QI5_0 + 4)
#define MMQ_MMA_TILE_X_K_Q5_1 (2*WARP_SIZE + WARP_SIZE/QI5_1 + 4)
#define MMQ_MMA_TILE_X_K_Q8_0 (1*WARP_SIZE + WARP_SIZE/QI8_0 + 0)
#define MMQ_MMA_TILE_X_K_Q2_K (1*WARP_SIZE + WARP_SIZE + 4)
#define MMQ_MMA_TILE_X_K_Q3_K (2*WARP_SIZE + WARP_SIZE/QI3_K + WARP_SIZE/4 + 2)
#define MMQ_MMA_TILE_X_K_Q4_K (1*WARP_SIZE + WARP_SIZE/QI4_K + WARP_SIZE/8 + 7)
#define MMQ_MMA_TILE_X_K_Q5_K (2*WARP_SIZE + WARP_SIZE/QI5_K + WARP_SIZE/8 + 7)
#define MMQ_MMA_TILE_X_K_Q6_K (2*WARP_SIZE + WARP_SIZE/QI6_K + WARP_SIZE/8 + 7)
static_assert(MMQ_MMA_TILE_X_K_Q4_0 % 8 == 4, "Wrong padding.");
static_assert(MMQ_MMA_TILE_X_K_Q4_1 % 8 == 4, "Wrong padding.");
static_assert(MMQ_MMA_TILE_X_K_Q5_0 % 8 == 4, "Wrong padding.");
static_assert(MMQ_MMA_TILE_X_K_Q5_1 % 8 == 4, "Wrong padding.");
static_assert(MMQ_MMA_TILE_X_K_Q8_0 % 8 == 4, "Wrong padding.");
static_assert(MMQ_MMA_TILE_X_K_Q2_K % 8 == 4, "Wrong padding.");
static_assert(MMQ_MMA_TILE_X_K_Q3_K % 8 == 4, "Wrong padding.");
static_assert(MMQ_MMA_TILE_X_K_Q4_K % 8 == 4, "Wrong padding.");
static_assert(MMQ_MMA_TILE_X_K_Q5_K % 8 == 4, "Wrong padding.");
static_assert(MMQ_MMA_TILE_X_K_Q6_K % 8 == 4, "Wrong padding.");
static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
return type == GGML_TYPE_Q4_0 ? MMQ_MMA_TILE_X_K_Q4_0 :
type == GGML_TYPE_Q4_1 ? MMQ_MMA_TILE_X_K_Q4_1 :
type == GGML_TYPE_Q5_0 ? MMQ_MMA_TILE_X_K_Q5_0 :
type == GGML_TYPE_Q5_1 ? MMQ_MMA_TILE_X_K_Q5_1 :
type == GGML_TYPE_Q8_0 ? MMQ_MMA_TILE_X_K_Q8_0 :
type == GGML_TYPE_Q2_K ? MMQ_MMA_TILE_X_K_Q2_K :
type == GGML_TYPE_Q3_K ? MMQ_MMA_TILE_X_K_Q3_K :
type == GGML_TYPE_Q4_K ? MMQ_MMA_TILE_X_K_Q4_K :
type == GGML_TYPE_Q5_K ? MMQ_MMA_TILE_X_K_Q5_K :
type == GGML_TYPE_Q6_K ? MMQ_MMA_TILE_X_K_Q6_K :
0;
}
#define MMQ_TILE_Y_K (WARP_SIZE + WARP_SIZE/QI8_1)
#define MMQ_NWARPS 8
static int mmq_get_granularity_host(const int mmq_x, const int cc) {
return int8_mma_available(cc) && mmq_x >= 48 ? 16 : 8;
}
#ifdef INT8_MMA_AVAILABLE
static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) {
return mmq_x >= 48 ? 16 : 8;
2024-06-05 16:53:00 +02:00
}
2024-06-24 12:41:23 +02:00
#else
static constexpr __device__ int mmq_get_granularity_device(const int /* mmq_x */) {
return 8;
}
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
// ------------------------------------------------------------
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + WARP_SIZE);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + txs.qs);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = threadIdx.x / QI4_0;
const int kqsx = threadIdx.x % QI4_0;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q4_0 + threadIdx.x] = get_int_from_uint8(bxi->qs, kqsx);
#else
x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_from_uint8(bxi->qs, kqsx);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
const int blocks_per_tile_x_row = WARP_SIZE / QI4_0;
const int kbxd = threadIdx.x % blocks_per_tile_x_row;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) {
int i = i0 + threadIdx.y * QI4_0 + threadIdx.x / blocks_per_tile_x_row;
if (need_check) {
i = min(i, i_max);
}
const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_df[i*MMQ_MMA_TILE_X_K_Q4_0 + kbxd] = bxi->d;
#else
x_df[i*(WARP_SIZE/QI4_0) + i/QI4_0 + kbxd] = bxi->d;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-10 11:45:13 +02:00
static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + txs.qs;
2024-06-10 11:45:13 +02:00
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-09 09:42:25 +02:00
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const int kyqs = k0 % (QI8_1/2) + QI8_1 * (k0 / (QI8_1/2));
int u[2*VDR_Q4_0_Q8_1_MMQ];
#pragma unroll
for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) {
2024-06-09 09:42:25 +02:00
u[2*l+0] = y_qs[j*MMQ_TILE_Y_K + (kyqs + l) % WARP_SIZE];
u[2*l+1] = y_qs[j*MMQ_TILE_Y_K + (kyqs + l + QI4_0) % WARP_SIZE];
2024-06-05 16:53:00 +02:00
}
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMQ>
2024-06-14 18:41:49 +02:00
(&x_qs[i*(WARP_SIZE + 1) + k0], u, x_df[i*(WARP_SIZE/QI4_0) + i/QI4_0 + k0/QI4_0],
2024-06-09 09:42:25 +02:00
y_ds[j*MMQ_TILE_Y_K + (2*k0/QI8_1) % (WARP_SIZE/QI8_1)]);
2024-06-05 16:53:00 +02:00
}
}
}
2024-06-10 11:45:13 +02:00
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q4_0_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
typedef mma_int_A_I16K8 mma_A;
typedef mma_int_B_J8K8 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + WARP_SIZE;
2024-06-10 11:45:13 +02:00
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-24 12:41:23 +02:00
mma_A A[ntx];
float dA[ntx][mma_C::ne/2];
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
#pragma unroll
for (int l = 0; l < mma_A::ne; ++l) {
const int i = i0 + n*mma_A::I + mma_A::get_i(l);
const int k = k0 + mma_A::get_k(l) % QI4_0;
const int shift = 4*(mma_A::get_k(l) / QI4_0);
A[n].x[l] = __vsubss4((x_qs[i*MMQ_MMA_TILE_X_K_Q4_0 + k] >> shift) & 0x0F0F0F0F, 0x08080808);
}
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q4_0 + k0/QI4_0];
}
2024-06-10 11:45:13 +02:00
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
mma_B B;
float dB[mma_C::ne/2];
B.load(y_qs + j0*MMQ_TILE_Y_K + (2*k0) % WARP_SIZE, MMQ_TILE_Y_K);
2024-06-10 11:45:13 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
2024-06-24 12:41:23 +02:00
dB[l] = __low2float(y_ds[j*MMQ_TILE_Y_K + (2*k0/QI8_1) % (WARP_SIZE/QI8_1)]);
2024-06-10 11:45:13 +02:00
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int n = 0; n < ntx; ++n) {
mma_C C;
C.mma_K8(A[n], B);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne; ++l) {
sum[(j0/mma_C::J + n)*mma_C::ne + l] += dA[n][l/2]*dB[l%2]*C.x[l];
}
2024-06-10 11:45:13 +02:00
}
}
2024-06-14 18:41:49 +02:00
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
}
2024-06-05 16:53:00 +02:00
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + WARP_SIZE);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + txs.qs);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = threadIdx.x / QI4_1;
const int kqsx = threadIdx.x % QI4_1;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q4_1 + threadIdx.x] = get_int_from_uint8_aligned(bxi->qs, kqsx);
#else
x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_from_uint8_aligned(bxi->qs, kqsx);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
const int blocks_per_tile_x_row = WARP_SIZE / QI4_1;
const int kbxd = threadIdx.x % blocks_per_tile_x_row;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) {
int i = i0 + threadIdx.y * QI4_1 + threadIdx.x / blocks_per_tile_x_row;
if (need_check) {
i = min(i, i_max);
}
const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_dm[i*MMQ_MMA_TILE_X_K_Q4_1 + kbxd] = bxi->dm;
#else
x_dm[i*(WARP_SIZE/QI4_1) + i/QI4_1 + kbxd] = bxi->dm;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-10 11:45:13 +02:00
static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + txs.qs;
2024-06-09 09:42:25 +02:00
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const int kyqs = k0 % (QI8_1/2) + QI8_1 * (k0 / (QI8_1/2));
int u[2*VDR_Q4_1_Q8_1_MMQ];
#pragma unroll
for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
2024-06-09 09:42:25 +02:00
u[2*l+0] = y_qs[j*MMQ_TILE_Y_K + (kyqs + l) % WARP_SIZE];
u[2*l+1] = y_qs[j*MMQ_TILE_Y_K + (kyqs + l + QI4_1) % WARP_SIZE];
2024-06-05 16:53:00 +02:00
}
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
2024-06-14 18:41:49 +02:00
(&x_qs[i*(WARP_SIZE + 1) + k0], u, x_dm[i*(WARP_SIZE/QI4_1) + i/QI4_1 + k0/QI4_1],
2024-06-09 09:42:25 +02:00
y_ds[j*MMQ_TILE_Y_K + (2*k0/QI8_1) % (WARP_SIZE/QI8_1)]);
2024-06-05 16:53:00 +02:00
}
}
}
2024-06-10 11:45:13 +02:00
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q4_1_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
typedef mma_int_A_I16K8 mma_A;
2024-06-24 12:41:23 +02:00
typedef mma_int_A_I16K4 mma_A_K4;
2024-06-10 11:45:13 +02:00
typedef mma_int_B_J8K8 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + WARP_SIZE;
2024-06-10 11:45:13 +02:00
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-24 12:41:23 +02:00
mma_A A[ntx];
half2 dmA[ntx][mma_C::ne/2];
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
((mma_A_K4 *) &A[n])[0].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q4_1 + k0, MMQ_MMA_TILE_X_K_Q4_1);
A[n].x[2] = (A[n].x[0] >> 4) & 0x0F0F0F0F;
A[n].x[3] = (A[n].x[1] >> 4) & 0x0F0F0F0F;
A[n].x[0] &= 0x0F0F0F0F;
A[n].x[1] &= 0x0F0F0F0F;
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
dmA[n][l] = x_dm[i*MMQ_MMA_TILE_X_K_Q4_1 + k0/QI4_1];
}
2024-06-10 11:45:13 +02:00
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
2024-06-10 11:45:13 +02:00
mma_B B;
half2 dsB[mma_C::ne/2];
2024-06-24 12:41:23 +02:00
B.load(y_qs + j0*MMQ_TILE_Y_K + (2*k0) % WARP_SIZE, MMQ_TILE_Y_K);
2024-06-10 11:45:13 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
dsB[l] = y_ds[j*MMQ_TILE_Y_K + (2*k0/QI8_1) % (WARP_SIZE/QI8_1)];
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int n = 0; n < ntx; ++n) {
mma_C C;
C.mma_K8(A[n], B);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne; ++l) {
const half2 dmA_dsB = dmA[n][l/2]*dsB[l%2];
sum[(j0/mma_C::J + n)*mma_C::ne + l] += __low2float(dmA_dsB)*C.x[l] + __high2float(dmA_dsB);
}
2024-06-10 11:45:13 +02:00
}
}
2024-06-14 18:41:49 +02:00
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
}
2024-06-05 16:53:00 +02:00
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + WARP_SIZE*2);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, mmq_y);
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + txs.qs);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = threadIdx.x / QI5_0;
const int kqsx = threadIdx.x % QI5_0;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbx;
const int ql = get_int_from_uint8(bxi->qs, kqsx);
const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_0));
int qs0 = (ql >> 0) & 0x0F0F0F0F;
qs0 |= (qh << 4) & 0x00000010; // 0 -> 4
qs0 |= (qh << 11) & 0x00001000; // 1 -> 12
qs0 |= (qh << 18) & 0x00100000; // 2 -> 20
qs0 |= (qh << 25) & 0x10000000; // 3 -> 28
qs0 = __vsubss4(qs0, 0x10101010); // subtract 16
int qs1 = (ql >> 4) & 0x0F0F0F0F;
qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4
qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12
qs1 |= (qh << 2) & 0x00100000; // 18 -> 20
qs1 |= (qh << 9) & 0x10000000; // 19 -> 28
qs1 = __vsubss4(qs1, 0x10101010); // subtract 16
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q5_0 + kbx*(2*QI5_0) + kqsx + 0] = qs0;
x_qs[i*MMQ_MMA_TILE_X_K_Q5_0 + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
#else
x_qs[i*(2*WARP_SIZE + 1) + kbx*(2*QI5_0) + kqsx + 0] = qs0;
x_qs[i*(2*WARP_SIZE + 1) + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
const int blocks_per_tile_x_row = WARP_SIZE / QI5_0;
const int kbxd = threadIdx.x % blocks_per_tile_x_row;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) {
int i = i0 + threadIdx.y * QI5_0 + threadIdx.x / blocks_per_tile_x_row;
if (need_check) {
i = min(i, i_max);
}
const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_df[i*MMQ_MMA_TILE_X_K_Q5_0 + kbxd] = bxi->d;
#else
x_df[i*(WARP_SIZE/QI5_0) + i/QI5_0 + kbxd] = bxi->d;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-10 11:45:13 +02:00
static __device__ __forceinline__ void vec_dot_q5_0_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, mmq_y);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + txs.qs;
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-09 09:42:25 +02:00
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q8_0_q8_1_impl<float, QR5_0*VDR_Q5_0_Q8_1_MMQ>
2024-06-24 12:41:23 +02:00
(&x_qs[i*(2*WARP_SIZE + 1) + 2*k0], &y_qs[j*MMQ_TILE_Y_K + (2*k0) % WARP_SIZE],
x_df[i*(WARP_SIZE/QI5_0) + i/QI5_0 + k0/QI5_0], y_df[j*MMQ_TILE_Y_K + (2*k0/QI8_1) % (WARP_SIZE/QI8_1)]);
2024-06-05 16:53:00 +02:00
}
}
}
2024-06-10 11:45:13 +02:00
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q5_0_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
typedef mma_int_A_I16K8 mma_A;
typedef mma_int_B_J8K8 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + WARP_SIZE*2;
2024-06-10 11:45:13 +02:00
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-24 12:41:23 +02:00
mma_A A[ntx];
float dA[ntx][mma_C::ne/2];
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
A[n].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q5_0 + QR5_1*k0, MMQ_MMA_TILE_X_K_Q5_0);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + mma_C::get_i(2*l) + n*mma_C::I;
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q5_0 + k0/QI5_0];
}
2024-06-10 11:45:13 +02:00
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
2024-06-10 11:45:13 +02:00
mma_B B;
float dB[mma_C::ne/2];
2024-06-24 12:41:23 +02:00
B.load(y_qs + j0*MMQ_TILE_Y_K + (2*k0) % WARP_SIZE, MMQ_TILE_Y_K);
2024-06-10 11:45:13 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
dB[l] = y_df[j*MMQ_TILE_Y_K + (2*k0/QI8_1) % (WARP_SIZE/QI8_1)];
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int n = 0; n < ntx; ++n) {
mma_C C;
C.mma_K8(A[n], B);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne; ++l) {
sum[(j0/mma_C::J + n)*mma_C::ne + l] += dA[n][l/2]*dB[l%2]*C.x[l];
}
2024-06-10 11:45:13 +02:00
}
}
2024-06-14 18:41:49 +02:00
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
}
2024-06-05 16:53:00 +02:00
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_1(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + 2*WARP_SIZE);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + txs.qs);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = threadIdx.x / QI5_1;
const int kqsx = threadIdx.x % QI5_1;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbx;
const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_1));
int qs0 = (ql >> 0) & 0x0F0F0F0F;
qs0 |= (qh << 4) & 0x00000010; // 0 -> 4
qs0 |= (qh << 11) & 0x00001000; // 1 -> 12
qs0 |= (qh << 18) & 0x00100000; // 2 -> 20
qs0 |= (qh << 25) & 0x10000000; // 3 -> 28
int qs1 = (ql >> 4) & 0x0F0F0F0F;
qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4
qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12
qs1 |= (qh << 2) & 0x00100000; // 18 -> 20
qs1 |= (qh << 9) & 0x10000000; // 19 -> 28
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q5_1 + kbx*(2*QI5_1) + kqsx + 0] = qs0;
x_qs[i*MMQ_MMA_TILE_X_K_Q5_1 + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
#else
x_qs[i*(2*WARP_SIZE + 1) + kbx*(2*QI5_1) + kqsx + 0] = qs0;
x_qs[i*(2*WARP_SIZE + 1) + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
const int blocks_per_tile_x_row = WARP_SIZE / QI5_1;
const int kbxd = threadIdx.x % blocks_per_tile_x_row;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) {
int i = i0 + threadIdx.y * QI5_1 + threadIdx.x / blocks_per_tile_x_row;
if (need_check) {
i = min(i, i_max);
}
const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_dm[i*MMQ_MMA_TILE_X_K_Q5_1 + kbxd] = bxi->dm;
#else
x_dm[i*(WARP_SIZE/QI5_1) + i/QI5_1 + kbxd] = bxi->dm;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-10 11:45:13 +02:00
static __device__ __forceinline__ void vec_dot_q5_1_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + txs.qs;
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-09 09:42:25 +02:00
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q8_1_q8_1_impl<QR5_1*VDR_Q5_1_Q8_1_MMQ>
2024-06-24 12:41:23 +02:00
(&x_qs[i*(2*WARP_SIZE + 1) + 2*k0], &y_qs[j*MMQ_TILE_Y_K + (2*k0) % WARP_SIZE],
x_dm[i*(WARP_SIZE/QI5_1) + i/QI5_1 + k0/QI5_1], y_ds[j*MMQ_TILE_Y_K + (2*k0/QI8_1) % (WARP_SIZE/QI8_1)]);
2024-06-05 16:53:00 +02:00
}
}
}
2024-06-10 11:45:13 +02:00
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q5_1_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
typedef mma_int_A_I16K8 mma_A;
typedef mma_int_B_J8K8 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + 2*WARP_SIZE;
2024-06-10 11:45:13 +02:00
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-24 12:41:23 +02:00
mma_A A[ntx];
half2 dmA[ntx][mma_C::ne/2];
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
A[n].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q5_1 + QR5_1*k0, MMQ_MMA_TILE_X_K_Q5_1);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + mma_C::get_i(2*l) + n*mma_C::I;
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
dmA[n][l] = x_dm[i*MMQ_MMA_TILE_X_K_Q5_1 + k0/QI5_1];
}
2024-06-10 11:45:13 +02:00
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
2024-06-10 11:45:13 +02:00
mma_B B;
half2 dsB[mma_C::ne/2];
2024-06-24 12:41:23 +02:00
B.load(y_qs + j0*MMQ_TILE_Y_K + (2*k0) % WARP_SIZE, MMQ_TILE_Y_K);
2024-06-10 11:45:13 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
dsB[l] = y_ds[j*MMQ_TILE_Y_K + (2*k0/QI8_1) % (WARP_SIZE/QI8_1)];
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int n = 0; n < ntx; ++n) {
mma_C C;
C.mma_K8(A[n], B);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne; ++l) {
const half2 dmA_dsB = dmA[n][l/2]*dsB[l%2];
sum[(j0/mma_C::J + n)*mma_C::ne + l] += __low2float(dmA_dsB)*C.x[l] + __high2float(dmA_dsB);
}
2024-06-10 11:45:13 +02:00
}
}
2024-06-14 18:41:49 +02:00
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
}
2024-06-05 16:53:00 +02:00
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_tile + WARP_SIZE);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + txs.qs);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = threadIdx.x / QI8_0;
const int kqsx = threadIdx.x % QI8_0;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + threadIdx.x] = get_int_from_int8(bxi->qs, kqsx);
#else
x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_from_int8(bxi->qs, kqsx);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
const int blocks_per_tile_x_row = WARP_SIZE / QI8_0;
const int kbxd = threadIdx.x % blocks_per_tile_x_row;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) {
int i = i0 + threadIdx.y * QI8_0 + threadIdx.x / blocks_per_tile_x_row;
if (need_check) {
i = min(i, i_max);
}
const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d;
#else
x_df[i*(WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-10 11:45:13 +02:00
static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + txs.qs;
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-09 09:42:25 +02:00
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q8_0_q8_1_impl<float, VDR_Q8_0_Q8_1_MMQ>
2024-06-24 12:41:23 +02:00
(&x_qs[i*(WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k0], x_df[i*(WARP_SIZE/QI8_0) + i/QI8_0 + k0/QI8_0],
2024-06-09 09:42:25 +02:00
y_df[j*MMQ_TILE_Y_K + k0/QI8_1]);
2024-06-05 16:53:00 +02:00
}
}
}
2024-06-10 11:45:13 +02:00
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
typedef mma_int_A_I16K8 mma_A;
typedef mma_int_B_J8K8 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + WARP_SIZE;
2024-06-10 11:45:13 +02:00
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-24 12:41:23 +02:00
mma_A A[ntx];
float dA[ntx][mma_C::ne/2];
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y/ntx)*rows_per_warp;
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
A[n].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_A::I + mma_C::get_i(2*l);
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + k0/QI8_0];
}
2024-06-10 11:45:13 +02:00
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
2024-06-10 11:45:13 +02:00
mma_B B;
float dB[mma_C::ne/2];
2024-06-24 12:41:23 +02:00
B.load(y_qs + j0*MMQ_TILE_Y_K + k0, MMQ_TILE_Y_K);
2024-06-10 11:45:13 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
dB[l] = y_df[j*MMQ_TILE_Y_K + k0/QI8_1];
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int n = 0; n < ntx; ++n) {
mma_C C;
C.mma_K8(A[n], B);
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne; ++l) {
sum[(j0/mma_C::J + n)*mma_C::ne + l] += C.x[l]*dA[n][l/2]*dB[l%2];
}
2024-06-10 11:45:13 +02:00
}
}
2024-06-14 18:41:49 +02:00
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
}
2024-06-05 16:53:00 +02:00
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + WARP_SIZE);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + txs.qs);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = threadIdx.x / QI2_K;
const int kqsx = threadIdx.x % QI2_K;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q2_K * bxi = (const block_q2_K *) x + kbx0 + i*stride + kbx;
2024-06-14 18:41:49 +02:00
const int x_ql_0 = get_int_from_uint8(bxi->qs, kqsx);
2024-06-05 16:53:00 +02:00
#pragma unroll
2024-06-14 18:41:49 +02:00
for (int l = 0; l < QR2_K; ++l) {
const int k = kbx*QI2_K + (kqsx/8)*8 + l*2 + (kqsx % 8)/4;
2024-06-05 16:53:00 +02:00
2024-06-14 18:41:49 +02:00
int x_qs_k = ((x_ql_0 >> (2*l)) & 0x03030303) << (2*(kqsx % 4));
x_qs_k |= __shfl_xor_sync(0xFFFFFFFF, x_qs_k, 1, WARP_SIZE);
x_qs_k |= __shfl_xor_sync(0xFFFFFFFF, x_qs_k, 2, WARP_SIZE);
2024-06-05 16:53:00 +02:00
2024-06-14 18:41:49 +02:00
if (kqsx % QR2_K != 0) {
continue;
}
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k] = x_qs_k;
#else
x_qs[i*(WARP_SIZE + 1) + k] = x_qs_k;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
2024-06-14 18:41:49 +02:00
const int sc_m = bxi->scales[kqsx];
#ifdef FAST_FP16_AVAILABLE
const half2 x_dm_ik = __hmul2(bxi->dm, make_half2(sc_m & 0x0F, sc_m >> 4));
#else
const float2 bxi_dmf = __half22float2(bxi->dm);
const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4));
#endif // FAST_FP16_AVAILABLE
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + threadIdx.x] = x_dm_ik;
#else
x_dm[i*(WARP_SIZE + 1) + threadIdx.x] = x_dm_ik;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-14 18:41:49 +02:00
static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + txs.qs;
2024-06-14 18:41:49 +02:00
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-09 09:42:25 +02:00
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
2024-06-14 18:41:49 +02:00
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq(
&x_qs[i*(WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + (QR2_K*k0) % WARP_SIZE],
&x_dm[i*(WARP_SIZE + 1) + k0], y_df[j*MMQ_TILE_Y_K + ((QR2_K*k0) % WARP_SIZE)/QI8_1]);
}
}
}
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
typedef mma_int_A_I16K4 mma_A;
typedef mma_int_B_J8K4 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + WARP_SIZE;
2024-06-14 18:41:49 +02:00
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
mma_A A[ntx][2];
float dA[ntx][mma_C::ne/2][2];
float mA[ntx][mma_C::ne/2][2];
2024-06-05 16:53:00 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
#pragma unroll
for (int l = 0; l < mma_A::ne; ++l) {
const int i = i0 + n*mma_A::I + mma_A::get_i(l);
const int shift = 2*mma_A::get_k(l);
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
A[n][0].x[l] = (x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k0 + 0] >> shift) & 0x03030303;
A[n][1].x[l] = (x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k0 + 1] >> shift) & 0x03030303;
}
2024-06-05 16:53:00 +02:00
2024-06-14 18:41:49 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
2024-06-14 18:41:49 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int kdm = 0; kdm < 2; ++kdm) {
const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0 + kdm]);
2024-06-14 18:41:49 +02:00
2024-06-24 12:41:23 +02:00
dA[n][l][kdm] = dm.x;
mA[n][l][kdm] = dm.y;
}
2024-06-05 16:53:00 +02:00
}
}
2024-06-14 18:41:49 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
2024-06-14 18:41:49 +02:00
mma_B B[2];
float dB[mma_C::ne/2];
2024-06-24 12:41:23 +02:00
B[0].load(y_qs + j0*MMQ_TILE_Y_K + (QR2_K*k0 + 0) % WARP_SIZE, MMQ_TILE_Y_K);
B[1].load(y_qs + j0*MMQ_TILE_Y_K + (QR2_K*k0 + mma_B::K) % WARP_SIZE, MMQ_TILE_Y_K);
2024-06-14 18:41:49 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
dB[l] = y_df[j*MMQ_TILE_Y_K + ((4*k0)/QI8_1) % (WARP_SIZE/QI8_1)];
}
2024-06-24 12:41:23 +02:00
mma_C Cm[2];
2024-06-14 18:41:49 +02:00
mma_A A1;
A1.x[0] = 0x01010101;
A1.x[1] = 0x01010101;
Cm[0].mma_K4(A1, B[0]);
Cm[1].mma_K4(A1, B[1]);
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
mma_C Cd[2];
Cd[0].mma_K4(A[n][0], B[0]);
Cd[1].mma_K4(A[n][1], B[1]);
#pragma unroll
for (int l = 0; l < mma_C::ne; ++l) {
sum[(j0/mma_C::J + n)*mma_C::ne + l] += (
Cd[0].x[l]*dA[n][l/2][0] + Cd[1].x[l]*dA[n][l/2][1] - Cm[0].x[l]*mA[n][l/2][0] - Cm[1].x[l]*mA[n][l/2][1])*dB[l%2];
}
2024-06-14 18:41:49 +02:00
}
}
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + WARP_SIZE*2);
int * x_sc = (int *) (x_df + WARP_SIZE/QI3_K);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + txs.qs);
int * x_sc = (int *) (x_df + txs.dm);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = threadIdx.x / QI3_K;
const int kqsx = threadIdx.x % QI3_K;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride + kbx;
2024-06-14 18:41:49 +02:00
const int x_ql_0 = get_int_from_uint8(bxi->qs, kqsx);
const int x_qh_0 = get_int_from_uint8(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2)));
#pragma unroll
for (int l = 0; l < QR3_K; ++l) {
const int k = kbx*(QR3_K*QI3_K) + (kqsx/8)*32 + l*8 + kqsx % 8;
const int x_ql_k = (x_ql_0 >> (2*l)) & 0x03030303;
const int x_qh_k = ((x_qh_0 >> l) << 2) & 0x04040404;
int x_qs_k = (x_ql_k | x_qh_k) << (4*(k%2));
x_qs_k |= __shfl_xor_sync(0xFFFFFFFF, x_qs_k, 1, WARP_SIZE);
if (kqsx % 2 != 0) {
continue;
}
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k/2] = x_qs_k;
#else
x_qs[i*(2*WARP_SIZE + 1) + k/2] = x_qs_k;
#endif // INT8_MMA_AVAILABLE
2024-06-14 18:41:49 +02:00
}
2024-06-05 16:53:00 +02:00
}
const int blocks_per_tile_x_row = WARP_SIZE / QI3_K;
const int kbxd = threadIdx.x % blocks_per_tile_x_row;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) {
int i = (i0 + threadIdx.y * QI3_K + threadIdx.x / blocks_per_tile_x_row) % mmq_y;
if (need_check) {
i = min(i, i_max);
}
const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride + kbxd;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_df[i*MMQ_MMA_TILE_X_K_Q3_K + kbxd] = bxi->d;
#else
x_df[i*(WARP_SIZE/QI3_K) + i/QI3_K + kbxd] = bxi->d;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
int i = i0 + threadIdx.y * 4 + threadIdx.x / (WARP_SIZE/4);
if (need_check) {
i = min(i, i_max);
}
const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride + (threadIdx.x % (WARP_SIZE/4)) / (QI3_K/4);
const int ksc = threadIdx.x % (QI3_K/4);
const int ksc_low = ksc % (QI3_K/8);
const int shift_low = 4 * (ksc / (QI3_K/8));
const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F;
const int ksc_high = QI3_K/8;
const int shift_high = 2 * ksc;
const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030;
const int sc = __vsubss4(sc_low | sc_high, 0x20202020);
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_sc[i*MMQ_MMA_TILE_X_K_Q3_K + threadIdx.x % (WARP_SIZE/4)] = sc;
#else
x_sc[i*(WARP_SIZE/4) + i/4 + threadIdx.x % (WARP_SIZE/4)] = sc;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-14 18:41:49 +02:00
static __device__ __forceinline__ void vec_dot_q3_K_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-09 09:42:25 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + txs.qs;
const int * x_sc = (const int *) x_df + txs.dm;
2024-06-14 18:41:49 +02:00
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const int kbx = k0 / QI3_K;
const int ky = (k0 % QI3_K) * QR3_K;
const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4;
2024-06-14 18:41:49 +02:00
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q3_K_q8_1_impl_mmq(
&x_qs[i*(2*WARP_SIZE + 1) + 2*k0], &y_qs[j*MMQ_TILE_Y_K + (k0*QR3_K) % WARP_SIZE], scales,
x_df[i*(WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[j*MMQ_TILE_Y_K + ((k0*QR3_K) % WARP_SIZE)/QI8_1]);
}
}
}
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q3_K_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
typedef mma_int_A_I16K4 mma_A;
typedef mma_int_B_J8K4 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + WARP_SIZE*2;
const int * x_sc = (const int *) x_df + WARP_SIZE/QI3_K;
2024-06-14 18:41:49 +02:00
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
2024-06-14 18:41:49 +02:00
2024-06-24 12:41:23 +02:00
mma_A A[ntx][2];
int scA[ntx][mma_C::ne/2][2];
float dA[ntx][mma_C::ne/2];
2024-06-05 16:53:00 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
#pragma unroll
for (int l = 0; l < mma_A::ne; ++l) {
const int i = i0 + n*mma_A::I + mma_A::get_i(l);
const int k = QR3_K*k0 + mma_A::get_k(l);
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
A[n][0].x[l] = (x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k/2 + 0] >> (4*(k%2))) & 0x0F0F0F0F;
A[n][1].x[l] = (x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k/2 + mma_A::K/2] >> (4*(k%2))) & 0x0F0F0F0F;
A[n][0].x[l] = __vsubss4(A[n][0].x[l], 0x04040404);
A[n][1].x[l] = __vsubss4(A[n][1].x[l], 0x04040404);
}
2024-06-05 16:53:00 +02:00
2024-06-14 18:41:49 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
const int kbx = k0 / QI3_K;
const int ky = (k0 % QI3_K) * QR3_K;
const int8_t * sc = ((const int8_t *) (x_sc + i*MMQ_MMA_TILE_X_K_Q3_K + kbx*4)) + ky/4;
2024-06-14 18:41:49 +02:00
2024-06-24 12:41:23 +02:00
scA[n][l][0] = sc[0];
scA[n][l][1] = sc[1];
}
2024-06-14 18:41:49 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
2024-06-14 18:41:49 +02:00
2024-06-24 12:41:23 +02:00
dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/QI3_K];
}
2024-06-14 18:41:49 +02:00
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
2024-06-14 18:41:49 +02:00
mma_B B[2];
float dB[mma_C::ne/2];
2024-06-24 12:41:23 +02:00
B[0].load(y_qs + j0*MMQ_TILE_Y_K + (QR3_K*k0 + 0) % WARP_SIZE, MMQ_TILE_Y_K);
B[1].load(y_qs + j0*MMQ_TILE_Y_K + (QR3_K*k0 + mma_B::K) % WARP_SIZE, MMQ_TILE_Y_K);
2024-06-14 18:41:49 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
dB[l] = y_df[j*MMQ_TILE_Y_K + ((4*k0)/QI8_1) % (WARP_SIZE/QI8_1)];
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int n = 0; n < ntx; ++n) {
mma_C C[2];
C[0].mma_K4(A[n][0], B[0]);
C[1].mma_K4(A[n][1], B[1]);
2024-06-14 18:41:49 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne; ++l) {
sum[(j0/mma_C::J + n)*mma_C::ne + l] += (C[0].x[l]*scA[n][l/2][0] + C[1].x[l]*scA[n][l/2][1])*dA[n][l/2]*dB[l%2];
}
2024-06-05 16:53:00 +02:00
}
}
2024-06-14 18:41:49 +02:00
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_K(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + WARP_SIZE);
int * x_sc = (int *) (x_dm + WARP_SIZE/QI4_K);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y);
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + txs.qs);
int * x_sc = (int *) (x_dm + txs.dm);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = 0; // threadIdx.x / QI4_K
const int kqsx = threadIdx.x; // threadIdx.x % QI4_K
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + kbx;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q4_K + threadIdx.x] = get_int_from_uint8_aligned(bxi->qs, kqsx);
#else
x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_from_uint8_aligned(bxi->qs, kqsx);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256
const int kbxd = threadIdx.x % blocks_per_tile_x_row; // == 0 if QK_K == 256
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) {
int i = (i0 + threadIdx.y * QI4_K + threadIdx.x / blocks_per_tile_x_row) % mmq_y;
if (need_check) {
i = min(i, i_max);
}
const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + kbxd;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_dm[i*MMQ_MMA_TILE_X_K_Q4_K + kbxd] = bxi->dm;
#else
x_dm[i*(WARP_SIZE/QI4_K) + i/QI4_K + kbxd] = bxi->dm;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
int i = (i0 + threadIdx.y * 8 + threadIdx.x / (WARP_SIZE/8)) % mmq_y;
if (need_check) {
i = min(i, i_max);
}
const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + (threadIdx.x % (WARP_SIZE/8)) / (QI4_K/8);
const int * scales = (const int *) bxi->scales;
const int ksc = threadIdx.x % (WARP_SIZE/8);
// scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_sc[i*MMQ_MMA_TILE_X_K_Q4_K + ksc] = scales8;
#else
x_sc[i*(WARP_SIZE/8) + i/8 + ksc] = scales8;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-11 08:26:07 +02:00
static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + txs.qs;
const int * x_sc = (const int *) x_dm + txs.dm;
2024-06-09 09:42:25 +02:00
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k0/16]) + 2*((k0 % 16) / 8);
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q4_K_q8_1_impl_mmq(
2024-06-14 18:41:49 +02:00
&x_qs[i*(WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + (QR4_K*k0) % WARP_SIZE], sc, sc+8,
2024-06-09 09:42:25 +02:00
x_dm[i*(WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[j*MMQ_TILE_Y_K + ((QR4_K*k0) % WARP_SIZE)/QI8_1]);
2024-06-05 16:53:00 +02:00
}
}
}
2024-06-11 08:26:07 +02:00
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q4_K_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
2024-06-11 08:26:07 +02:00
typedef mma_int_A_I16K8 mma_A;
typedef mma_int_B_J8K8 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + WARP_SIZE;
const int * x_sc = (const int *) x_dm + WARP_SIZE/QI4_K;
2024-06-11 08:26:07 +02:00
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
mma_A A[ntx][2];
int scA[ntx][mma_C::ne/2][2];
int mA[ntx][mma_C::ne/2][2];
half2 dmA[ntx][mma_C::ne/2];
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int kvdr = 0; kvdr < VDR_Q4_K_Q8_1_MMQ; kvdr += 8) {
A[n][kvdr/4 + 0].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q4_K + k0, MMQ_MMA_TILE_X_K_Q4_K);
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int l = 0; l < mma_A::ne; ++l) {
A[n][kvdr/4 + 1].x[l] = (A[n][kvdr/4 + 0].x[l] >> 4) & 0x0F0F0F0F;
A[n][kvdr/4 + 0].x[l] &= 0x0F0F0F0F;
}
2024-06-11 08:26:07 +02:00
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int kvdr = 0; kvdr < VDR_Q4_K_Q8_1_MMQ; kvdr += 4) {
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_A::I + mma_C::get_i(2*l);
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
const uint8_t * sc = ((const uint8_t *) &x_sc[i*MMQ_MMA_TILE_X_K_Q4_K + k0/16]) + 2 * ((k0 % 16) / 8);
const uint8_t * m = sc + 8;
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
scA[n][l][kvdr/4] = sc[kvdr/4];
mA[n][l][kvdr/4] = m[kvdr/4];
}
2024-06-11 08:26:07 +02:00
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_A::I + mma_C::get_i(2*l);
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
dmA[n][l] = x_dm[i*MMQ_MMA_TILE_X_K_Q4_K + k0/QI4_K];
}
2024-06-11 08:26:07 +02:00
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
float tmpd[ntx][mma_C::ne] = {{0.0f}};
float tmpm[ntx][mma_C::ne] = {{0.0f}};
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int kvdr = 0; kvdr < VDR_Q4_K_Q8_1_MMQ; kvdr += 4) {
2024-06-11 08:26:07 +02:00
mma_B B;
half2 dsB[mma_C::ne/2];
2024-06-24 12:41:23 +02:00
B.load(y_qs + j0*MMQ_TILE_Y_K + (2*k0 + 2*kvdr) % WARP_SIZE, MMQ_TILE_Y_K);
2024-06-11 08:26:07 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
dsB[l] = y_ds[j*MMQ_TILE_Y_K + ((2*k0 + 2*kvdr)/QI8_1) % (WARP_SIZE/QI8_1)];
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int n = 0; n < ntx; ++n) {
mma_C C;
C.mma_K8(A[n][kvdr/4], B);
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne; ++l) {
tmpd[n][l] += (C.x[l]*scA[n][l/2][kvdr/4]) * __low2float(dsB[l%2]);
tmpm[n][l] += mA[n][l/2][kvdr/4] * __high2float(dsB[l%2]);
}
2024-06-11 08:26:07 +02:00
}
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
#pragma unroll
for (int l = 0; l < mma_C::ne; ++l) {
sum[(j0/mma_C::J + n)*mma_C::ne + l] += __low2float(dmA[n][l/2])*tmpd[n][l] - __high2float(dmA[n][l/2])*tmpm[n][l];
}
2024-06-11 08:26:07 +02:00
}
}
2024-06-14 18:41:49 +02:00
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-11 08:26:07 +02:00
}
2024-06-05 16:53:00 +02:00
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_K(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + WARP_SIZE*2);
int * x_sc = (int *) (x_dm + WARP_SIZE/QI5_K);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y);
int * x_qs = (int *) x_tile;
half2 * x_dm = (half2 *) (x_qs + txs.qs);
int * x_sc = (int *) (x_dm + txs.dm);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = 0; // threadIdx.x / QI5_K
const int kqsx = threadIdx.x; // threadIdx.x % QI5_K
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride + kbx;
const int ky = QR5_K*kqsx;
const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
const int ql0 = (ql >> 0) & 0x0F0F0F0F;
const int ql1 = (ql >> 4) & 0x0F0F0F0F;
const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4));
const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010;
const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010;
const int kq0 = ky - ky % (QI5_K/2) + threadIdx.x % (QI5_K/4) + 0;
const int kq1 = ky - ky % (QI5_K/2) + threadIdx.x % (QI5_K/4) + (QI5_K/4);
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q5_K + kq0] = ql0 | qh0;
x_qs[i*MMQ_MMA_TILE_X_K_Q5_K + kq1] = ql1 | qh1;
#else
x_qs[i*(2*WARP_SIZE + 1) + kq0] = ql0 | qh0;
x_qs[i*(2*WARP_SIZE + 1) + kq1] = ql1 | qh1;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256
const int kbxd = threadIdx.x % blocks_per_tile_x_row; // == 0 if QK_K == 256
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) {
int i = (i0 + threadIdx.y * QI5_K + threadIdx.x / blocks_per_tile_x_row) % mmq_y;
if (need_check) {
i = min(i, i_max);
}
const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride + kbxd;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_dm[i*MMQ_MMA_TILE_X_K_Q5_K + kbxd] = bxi->dm;
#else
x_dm[i*(WARP_SIZE/QI5_K) + i/QI5_K + kbxd] = bxi->dm;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
int i = (i0 + threadIdx.y * 8 + threadIdx.x / (WARP_SIZE/8)) % mmq_y;
if (need_check) {
i = min(i, i_max);
}
const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride + (threadIdx.x % (WARP_SIZE/8)) / (QI5_K/8);
const int * scales = (const int *) bxi->scales;
const int ksc = threadIdx.x % (WARP_SIZE/8);
// scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_sc[i*MMQ_MMA_TILE_X_K_Q5_K + ksc] = scales8;
#else
x_sc[i*(WARP_SIZE/8) + i/8 + ksc] = scales8;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-11 08:26:07 +02:00
static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + txs.qs;
const int * x_sc = (const int *) x_dm + txs.dm;
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-09 09:42:25 +02:00
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k0/16]) + 2 * ((k0 % 16) / 8);
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q5_K_q8_1_impl_mmq(
2024-06-14 18:41:49 +02:00
&x_qs[i*(QR5_K*WARP_SIZE + 1) + QR5_K*k0], &y_qs[j*MMQ_TILE_Y_K + (QR5_K*k0) % WARP_SIZE], sc, sc+8,
2024-06-09 09:42:25 +02:00
x_dm[i*(WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[j*MMQ_TILE_Y_K + ((QR5_K*k0) % WARP_SIZE)/QI8_1]);
2024-06-05 16:53:00 +02:00
}
}
}
2024-06-11 08:26:07 +02:00
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q5_K_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
2024-06-11 08:26:07 +02:00
typedef mma_int_A_I16K8 mma_A;
typedef mma_int_B_J8K8 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const half2 * x_dm = (const half2 *) x_qs + WARP_SIZE*2;
const int * x_sc = (const int *) x_dm + WARP_SIZE/QI5_K;
2024-06-11 08:26:07 +02:00
const int * y_qs = (const int *) y + 4;
const half2 * y_ds = (const half2 *) y;
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
mma_A A[ntx][2];
int scA[ntx][mma_C::ne/2][2];
int mA[ntx][mma_C::ne/2][2];
half2 dmA[ntx][mma_C::ne/2];
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int kvdr = 0; kvdr < VDR_Q5_K_Q8_1_MMQ; kvdr += 4) {
A[n][kvdr/4].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q5_K + (QR5_K*k0 + QR5_K*kvdr), MMQ_MMA_TILE_X_K_Q5_K);
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
const uint8_t * sc = ((const uint8_t *) &x_sc[i*MMQ_MMA_TILE_X_K_Q5_K + k0/16]) + 2 * ((k0 % 16) / 8);
const uint8_t * m = sc + 8;
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
scA[n][l][kvdr/4] = sc[kvdr/4];
mA[n][l][kvdr/4] = m[kvdr/4];
}
2024-06-11 08:26:07 +02:00
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
dmA[n][l] = x_dm[i*MMQ_MMA_TILE_X_K_Q5_K + k0/QI5_K];
}
2024-06-11 08:26:07 +02:00
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
float tmpd[ntx][mma_C::ne] = {{0.0f}};
float tmpm[ntx][mma_C::ne] = {{0.0f}};
2024-06-11 08:26:07 +02:00
#pragma unroll
for (int kvdr = 0; kvdr < VDR_Q5_K_Q8_1_MMQ; kvdr += 4) {
mma_B B;
half2 dsB[mma_C::ne/2];
2024-06-24 12:41:23 +02:00
B.load(y_qs + j0*MMQ_TILE_Y_K + (2*k0 + 2*kvdr) % WARP_SIZE, MMQ_TILE_Y_K);
2024-06-11 08:26:07 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
dsB[l] = y_ds[j*MMQ_TILE_Y_K + ((2*k0 + 2*kvdr)/QI8_1) % (WARP_SIZE/QI8_1)];
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int n = 0; n < ntx; ++n) {
mma_C C;
C.mma_K8(A[n][kvdr/4], B);
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne; ++l) {
tmpd[n][l] += (C.x[l]*scA[n][l/2][kvdr/4]) * __low2float(dsB[l%2]);
tmpm[n][l] += mA[n][l/2][kvdr/4] * __high2float(dsB[l%2]);
}
2024-06-11 08:26:07 +02:00
}
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
#pragma unroll
for (int l = 0; l < mma_C::ne; ++l) {
sum[(j0/mma_C::J + n)*mma_C::ne + l] += __low2float(dmA[n][l/2])*tmpd[n][l] - __high2float(dmA[n][l/2])*tmpm[n][l];
}
2024-06-11 08:26:07 +02:00
}
}
2024-06-14 18:41:49 +02:00
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-11 08:26:07 +02:00
}
2024-06-05 16:53:00 +02:00
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q6_K(
2024-06-24 12:41:23 +02:00
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + WARP_SIZE*2);
int * x_sc = (int *) (x_df + WARP_SIZE/QI6_K);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y);
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + txs.qs);
int * x_sc = (int *) (x_df + txs.dm);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
const int kbx = 0; // threadIdx.x / QI6_K
const int kqsx = threadIdx.x; // threadIdx.x % QI6_K
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
if (need_check) {
i = min(i, i_max);
}
const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + kbx;
const int ky = QR6_K*kqsx;
const int ql = get_int_from_uint8(bxi->ql, kqsx);
const int ql0 = (ql >> 0) & 0x0F0F0F0F;
const int ql1 = (ql >> 4) & 0x0F0F0F0F;
const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4));
const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030;
const int qh1 = (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) & 0x30303030;
const int kq0 = ky - ky % QI6_K + threadIdx.x % (QI6_K/2) + 0;
const int kq1 = ky - ky % QI6_K + threadIdx.x % (QI6_K/2) + (QI6_K/2);
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
#else
x_qs[i*(2*WARP_SIZE + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
x_qs[i*(2*WARP_SIZE + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256
const int kbxd = threadIdx.x % blocks_per_tile_x_row; // == 0 if QK_K == 256
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) {
int i = (i0 + threadIdx.y * QI6_K + threadIdx.x / blocks_per_tile_x_row) % mmq_y;
if (need_check) {
i = min(i, i_max);
}
const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + kbxd;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_df[i*MMQ_MMA_TILE_X_K_Q6_K + kbxd] = bxi->d;
#else
x_df[i*(WARP_SIZE/QI6_K) + i/QI6_K + kbxd] = bxi->d;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
int i = (i0 + threadIdx.y * 8 + threadIdx.x / (WARP_SIZE/8)) % mmq_y;
if (need_check) {
i = min(i, i_max);
}
const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (WARP_SIZE/8)) / 4;
2024-06-24 12:41:23 +02:00
#ifdef INT8_MMA_AVAILABLE
x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, threadIdx.x % (QI6_K/8));
#else
x_sc[i*(WARP_SIZE/8) + i/8 + threadIdx.x % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, threadIdx.x % (QI6_K/8));
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
}
}
template <int mmq_x, int mmq_y, int nwarps>
2024-06-11 08:26:07 +02:00
static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + txs.qs;
const int * x_sc = (const int *) x_df + txs.dm;
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-09 09:42:25 +02:00
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k0/8]);
sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q6_K_q8_1_impl_mmq(
2024-06-14 18:41:49 +02:00
&x_qs[i*(QR6_K*WARP_SIZE + 1) + QR6_K*k0], &y_qs[j*MMQ_TILE_Y_K + (QR6_K*k0) % WARP_SIZE], sc,
2024-06-24 12:41:23 +02:00
x_df[i*(WARP_SIZE/QI6_K) + i/QI6_K], &y_df[j*MMQ_TILE_Y_K + ((QR6_K*k0) % WARP_SIZE)/QI8_1]);
2024-06-05 16:53:00 +02:00
}
}
}
2024-06-11 08:26:07 +02:00
template <int mmq_x, int mmq_y, int nwarps>
static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
2024-06-24 12:41:23 +02:00
const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k0) {
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
2024-06-11 08:26:07 +02:00
typedef mma_int_A_I16K4 mma_A;
typedef mma_int_B_J8K4 mma_B;
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
const int * x_qs = (const int *) x;
const float * x_df = (const float *) x_qs + WARP_SIZE*2;
const int * x_sc = (const int *) x_df + WARP_SIZE/QI6_K;
2024-06-11 08:26:07 +02:00
const int * y_qs = (const int *) y + 4;
const float * y_df = (const float *) y;
2024-06-24 12:41:23 +02:00
const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
mma_A A[ntx][4];
int scA[ntx][mma_C::ne/2][4];
float dA[ntx][mma_C::ne/2];
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int kvdr = 0; kvdr < VDR_Q6_K_Q8_1_MMQ; kvdr += 4) {
A[n][kvdr/2 + 0].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (QR6_K*k0 + QR6_K*kvdr + 0), MMQ_MMA_TILE_X_K_Q6_K);
A[n][kvdr/2 + 1].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (QR6_K*k0 + QR6_K*kvdr + mma_A::K), MMQ_MMA_TILE_X_K_Q6_K);
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
const int8_t * sc = ((const int8_t *) &x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + k0/8]);
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
scA[n][l][kvdr/2 + 0] = sc[kvdr/2 + 0];
scA[n][l][kvdr/2 + 1] = sc[kvdr/2 + 1];
}
2024-06-11 08:26:07 +02:00
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne/2; ++l) {
const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
2024-06-11 08:26:07 +02:00
2024-06-24 12:41:23 +02:00
dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q6_K + k0/QI6_K];
}
2024-06-11 08:26:07 +02:00
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
float tmp[ntx][mma_C::ne] = {{0.0f}};
2024-06-11 08:26:07 +02:00
#pragma unroll
for (int kvdr = 0; kvdr < VDR_Q6_K_Q8_1_MMQ; kvdr += 4) {
mma_B B[2];
float dB[mma_C::ne/2];
2024-06-24 12:41:23 +02:00
const int k0B = (2*k0 + 2*kvdr) % WARP_SIZE;
B[0].load(y_qs + j0*MMQ_TILE_Y_K + 0 + k0B, MMQ_TILE_Y_K);
B[1].load(y_qs + j0*MMQ_TILE_Y_K + mma_B::K + k0B, MMQ_TILE_Y_K);
2024-06-11 08:26:07 +02:00
#pragma unroll
for (int l = 0; l < mma_C::ne/2; ++l) {
const int j = j0 + mma_C::get_j(l);
dB[l] = y_df[j*MMQ_TILE_Y_K + ((2*k0 + 2*kvdr)/QI8_1) % (WARP_SIZE/QI8_1)];
}
2024-06-24 12:41:23 +02:00
#pragma unroll
for (int n = 0; n < ntx; ++n) {
mma_C C[2];
C[0].mma_K4(A[n][kvdr/2 + 0], B[0]);
C[1].mma_K4(A[n][kvdr/2 + 1], B[1]);
2024-06-11 08:26:07 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int l = 0; l < mma_C::ne; ++l) {
tmp[n][l] += (C[0].x[l]*scA[n][l/2][kvdr/2 + 0] + C[1].x[l]*scA[n][l/2][kvdr/2 + 1])*dB[l%2];
}
2024-06-11 08:26:07 +02:00
}
}
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
#pragma unroll
for (int l = 0; l < mma_C::ne; ++l) {
sum[(j0/mma_C::J + n)*mma_C::ne + l] += tmp[n][l]*dA[n][l/2];
}
2024-06-11 08:26:07 +02:00
}
}
2024-06-14 18:41:49 +02:00
#else
2024-06-24 12:41:23 +02:00
GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
2024-06-14 18:41:49 +02:00
NO_DEVICE_CODE;
#endif // INT8_MMA_AVAILABLE
2024-06-11 08:26:07 +02:00
}
2024-06-10 11:45:13 +02:00
template<int mmq_x, int mmq_y, int nwarps, bool need_check>
2024-06-20 14:39:21 +02:00
static __device__ __forceinline__ void mmq_write_back_dp4a(
const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
2024-06-10 11:45:13 +02:00
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
2024-06-20 14:39:21 +02:00
const int j = j0 + threadIdx.y;
2024-06-10 11:45:13 +02:00
2024-06-20 14:39:21 +02:00
if (j > j_max) {
2024-06-10 11:45:13 +02:00
return;
}
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
2024-06-20 14:39:21 +02:00
const int i = i0 + threadIdx.x;
2024-06-10 11:45:13 +02:00
2024-06-20 14:39:21 +02:00
if (need_check && i > i_max) {
2024-06-10 11:45:13 +02:00
continue;
}
2024-06-20 14:39:21 +02:00
dst[j*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
2024-06-10 11:45:13 +02:00
}
}
}
template<int mmq_x, int mmq_y, int nwarps, bool need_check>
2024-06-20 14:39:21 +02:00
static __device__ __forceinline__ void mmq_write_back_mma(
const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
2024-06-10 11:45:13 +02:00
typedef mma_int_C_I16J8 mma_C;
2024-06-24 12:41:23 +02:00
constexpr int granularity = mmq_get_granularity_device(mmq_x);
constexpr int rows_per_warp = 2 * granularity;
constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
const int i0 = (threadIdx.y / ntx) * (ntx*mma_C::I);
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
static_assert(nwarps*mma_C::I == mmq_y, "nwarps*mma_C::I != mmq_y");
2024-06-14 18:41:49 +02:00
#endif // INT8_MMA_AVAILABLE
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
2024-06-10 11:45:13 +02:00
#pragma unroll
2024-06-24 12:41:23 +02:00
for (int n = 0; n < ntx; ++n) {
#pragma unroll
for (int l = 0; l < mma_C::ne; ++l) {
2024-06-24 22:15:33 +02:00
const int j = j0 + (threadIdx.y % ntx) * mma_C::J + mma_C::get_j(l);
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
if (j > j_max) {
continue;
}
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
const int i = i0 + n*mma_C::I + mma_C::get_i(l);
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
if (need_check && i > i_max) {
continue;
}
2024-06-10 11:45:13 +02:00
2024-06-24 12:41:23 +02:00
dst[j*stride + i] = sum[(j0/mma_C::J + n)*mma_C::ne + l];
}
2024-06-10 11:45:13 +02:00
}
}
}
2024-06-05 16:53:00 +02:00
// -------------------------------------------------------------------------------------------------------------------------------------
template <int mmq_x, int mmq_y, int nwarps, bool need_check, ggml_type type>
struct mmq_type_traits;
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q4_0> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q4_0_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q4_0<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q4_0_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q4_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q4_1> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q4_1_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q4_1<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q4_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q4_1_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q5_0> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q5_0_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q5_0<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q5_0_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q5_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q5_1> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q5_1_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q5_1<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q5_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q5_1_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q8_0> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q8_0_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q8_0<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q2_K> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q2_K_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q2_K<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q2_K_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q2_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q3_K> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q3_K_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q3_K<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q3_K_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q3_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q4_K> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q4_K_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q4_K<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q4_K_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q4_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q5_K> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q5_K_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q5_K<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q5_K_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q5_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q6_K> {
2024-06-14 18:41:49 +02:00
static constexpr int vdr = VDR_Q6_K_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_q6_K<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q6_K_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q6_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
2024-06-05 16:53:00 +02:00
};
2024-06-14 18:41:49 +02:00
static bool mmq_need_sum(const ggml_type type_x) {
2024-06-09 09:42:25 +02:00
switch (type_x) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
return true;
case GGML_TYPE_Q5_0:
return false;
case GGML_TYPE_Q5_1:
return true;
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
return false;
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
return true;
case GGML_TYPE_Q6_K:
return false;
default:
GGML_ASSERT(false);
break;
}
return false;
}
2024-06-20 14:39:21 +02:00
template <ggml_type type, int mmq_x, int nwarps, bool need_check, bool fixup>
static __device__ void mul_mat_q_process_tile(
const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
const int & ne00, const int & ne01, const int & stride01, const int & ne10, const int & ne11, const int & stride11, const int & ne0,
const int & it, const int & jt, const int & kb0_start, const int & kb0_stop) {
2024-06-05 16:53:00 +02:00
constexpr int qk = ggml_cuda_type_traits<type>::qk;
constexpr int qr = ggml_cuda_type_traits<type>::qr;
constexpr int qi = ggml_cuda_type_traits<type>::qi;
2024-06-20 14:39:21 +02:00
constexpr int mmq_y = get_mmq_y_device();
2024-06-05 16:53:00 +02:00
constexpr int vdr = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::vdr;
constexpr load_tiles_mmq_t load_tiles = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::load_tiles;
2024-06-14 18:41:49 +02:00
2024-06-24 12:41:23 +02:00
extern __shared__ char data_mul_mat_q[];
int * tile_y = (int *) data_mul_mat_q;
int * tile_x = tile_y + GGML_PAD(mmq_x*(WARP_SIZE + WARP_SIZE/QI8_1), nwarps*WARP_SIZE);
2024-06-14 18:41:49 +02:00
#ifdef INT8_MMA_AVAILABLE
constexpr vec_dot_mmq_t vec_dot = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::vec_dot_mma;
constexpr mmq_write_back_t write_back = mmq_write_back_mma<mmq_x, mmq_y, nwarps, need_check>;
#else
constexpr vec_dot_mmq_t vec_dot = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::vec_dot_dp4a;
constexpr mmq_write_back_t write_back = mmq_write_back_dp4a<mmq_x, mmq_y, nwarps, need_check>;
#endif // INT8_MMA_AVAILABLE
2024-06-05 16:53:00 +02:00
2024-06-20 14:39:21 +02:00
constexpr int blocks_per_warp = WARP_SIZE / qi;
2024-06-05 16:53:00 +02:00
2024-06-20 14:39:21 +02:00
float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
2024-06-05 16:53:00 +02:00
2024-06-20 14:39:21 +02:00
const int tile_x_max_i = ne01 - it*mmq_y - 1;
const int tile_y_max_j = ne11 - jt*mmq_x - 1;
2024-06-09 09:42:25 +02:00
2024-06-20 14:39:21 +02:00
const int * y = (const int *) yc + jt*(mmq_x*sizeof(block_q8_1_mmq)/sizeof(int));
2024-06-05 16:53:00 +02:00
2024-06-20 14:39:21 +02:00
for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_warp) {
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
load_tiles(x, tile_x, stride01*it*mmq_y + kb0, tile_x_max_i, stride01);
2024-06-05 16:53:00 +02:00
#pragma unroll
for (int kr = 0; kr < qr; ++kr) {
2024-06-09 09:42:25 +02:00
const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + kr*sizeof(block_q8_1_mmq)/sizeof(int));
2024-06-05 16:53:00 +02:00
#pragma unroll
2024-06-09 09:42:25 +02:00
for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
2024-06-05 16:53:00 +02:00
2024-06-09 09:42:25 +02:00
tile_y[l] = by0[l];
2024-06-05 16:53:00 +02:00
}
__syncthreads();
// #pragma unroll // unrolling this loop causes too much register pressure
for (int k0 = kr*WARP_SIZE/qr; k0 < (kr+1)*WARP_SIZE/qr; k0 += vdr) {
2024-06-24 12:41:23 +02:00
vec_dot(tile_x, tile_y, sum, k0);
2024-06-05 16:53:00 +02:00
}
__syncthreads();
}
}
2024-06-20 14:39:21 +02:00
if (fixup) {
write_back(sum, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x);
} else {
write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j);
}
}
// The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598
template <ggml_type type, int mmq_x, int nwarps, bool need_check>
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(RDNA3) || defined(RDNA2)
__launch_bounds__(WARP_SIZE*nwarps, 2)
#endif // defined(RDNA3) || defined(RDNA2)
#else
#if __CUDA_ARCH__ >= CC_VOLTA
__launch_bounds__(WARP_SIZE*nwarps, 1)
#else
__launch_bounds__(WARP_SIZE*nwarps, 2)
#endif // __CUDA_ARCH__ >= CC_VOLTA
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
static __global__ void mul_mat_q(
const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
// Skip unused template specializations for faster compilation:
2024-06-24 12:41:23 +02:00
if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) {
2024-06-20 14:39:21 +02:00
NO_DEVICE_CODE;
return;
}
constexpr int qk = ggml_cuda_type_traits<type>::qk;
constexpr int qi = ggml_cuda_type_traits<type>::qi;
constexpr int mmq_y = get_mmq_y_device();
// On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
{
constexpr bool fixup = false;
mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
(x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
blockIdx.x, blockIdx.y, 0, ne00/qk);
return;
}
#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
const int64_t blocks_per_ne00 = ne00 / qk;
constexpr int blocks_per_warp = WARP_SIZE / qi;
const int ntx = (ne11 + mmq_x - 1) / mmq_x; // Number of tiles x
const int nty = (ne01 + mmq_y - 1) / mmq_y; // Number of tiles y
// kbc == k block continuous, current index in continuous ijk space.
int64_t kbc = GGML_PAD((int64_t) blockIdx.x *blocks_per_ne00*ntx*nty / gridDim.x, blocks_per_warp);
const int64_t kbc_stop = GGML_PAD((int64_t)(blockIdx.x + 1)*blocks_per_ne00*ntx*nty / gridDim.x, blocks_per_warp);
// kb0 == k index when doing the matrix multiplication for an output tile.
int kb0_start = kbc % blocks_per_ne00;
int kb0_stop = min(blocks_per_ne00, kb0_start + kbc_stop - kbc);
while (kbc < kbc_stop && kb0_stop == blocks_per_ne00) {
const int jt = kbc / (blocks_per_ne00*nty); // j index of current tile.
const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; // i index of current tile.
constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
(x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
it, jt, kb0_start, kb0_stop);
kbc += blocks_per_ne00;
kbc -= kbc % blocks_per_ne00;
kb0_start = 0;
kb0_stop = min(blocks_per_ne00, kbc_stop - kbc);
}
if (kbc >= kbc_stop) {
return;
}
const int jt = kbc / (blocks_per_ne00*nty);
const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
constexpr bool fixup = true; // Last index writes it data to fixup buffer to avoid data races with other blocks.
mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
(x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
it, jt, kb0_start, kb0_stop);
}
template <ggml_type type, int mmq_x, int nwarps, bool need_check>
static __global__ void mul_mat_q_stream_k_fixup(
float * __restrict__ dst, const float * __restrict__ tmp_last_tile, const int ne00, const int ne01, const int ne11, const int ne0, const int block_num_mmq) {
constexpr int mmq_y = get_mmq_y_device();
constexpr int qk = ggml_cuda_type_traits<type>::qk;
constexpr int qi = ggml_cuda_type_traits<type>::qi;
constexpr int blocks_per_warp = WARP_SIZE / qi;
const int64_t blocks_per_ne00 = ne00 / qk;
float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
const int ntx = (ne11 + mmq_x - 1) / mmq_x;
const int nty = (ne01 + mmq_y - 1) / mmq_y;
bool any_fixup = false;
const int bidx_start = (blockIdx.y*nty + blockIdx.x) * block_num_mmq / (gridDim.y*gridDim.x);
const int bidx_stop = (blockIdx.y*nty + blockIdx.x + 1) * block_num_mmq / (gridDim.y*gridDim.x) + 1;
for (int bidx = bidx_start; bidx < bidx_stop; ++bidx) {
const int64_t kbc = GGML_PAD((int64_t) bidx *blocks_per_ne00*ntx*nty / block_num_mmq, blocks_per_warp);
const int64_t kbc_stop = GGML_PAD((int64_t)(bidx + 1)*blocks_per_ne00*ntx*nty / block_num_mmq, blocks_per_warp);
// Skip fixup tile if the MMQ CUDA block never wrote anything to it:
if (kbc == kbc_stop || kbc_stop % blocks_per_ne00 == 0) {
continue;
}
const int jt = kbc_stop / (blocks_per_ne00*nty);
const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
// Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block:
if (it != blockIdx.x || jt != blockIdx.y) {
continue;
}
any_fixup = true;
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i];
}
}
}
if (!any_fixup) {
return;
}
dst += blockIdx.y*mmq_x*ne0 + blockIdx.x*mmq_y;
const int i_max = ne01 - blockIdx.x*mmq_y - 1;
const int j_max = ne11 - blockIdx.y*mmq_x - 1;
#pragma unroll
for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
const int j = j0 + threadIdx.y;
if (j > j_max) {
return;
}
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
if (need_check && i > i_max) {
continue;
}
dst[j*ne0 + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
}
}
2024-06-05 16:53:00 +02:00
}
struct mmq_args {
const char * x; const char * y; float * dst;
2024-06-09 09:42:25 +02:00
int64_t ne00; int64_t ne01; int64_t stride01;
int64_t ne10; int64_t ne11; int64_t stride11;
2024-06-05 16:53:00 +02:00
int64_t ne0;
};
2024-06-24 12:41:23 +02:00
template<ggml_type type>
static int mmq_get_shmem(const int mmq_x, const int mmq_y, const int cc) {
const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y);
const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
const int shmem_x = int8_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
const int shmem_y = mmq_x*sizeof(block_q8_1_mmq);
2024-06-20 14:39:21 +02:00
return shmem_x + GGML_PAD(shmem_y, MMQ_NWARPS*WARP_SIZE*sizeof(int));
2024-06-14 18:41:49 +02:00
}
2024-06-20 14:39:21 +02:00
template <ggml_type type, int mmq_x>
static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
2024-06-05 16:53:00 +02:00
const int id = ggml_cuda_get_device();
const int cc = ggml_cuda_info().devices[id].cc;
2024-06-20 14:39:21 +02:00
const int nsm = ggml_cuda_info().devices[id].nsm;
const int mmq_y = get_mmq_y_host(cc);
2024-06-05 16:53:00 +02:00
2024-06-20 14:39:21 +02:00
const dim3 block_dims(WARP_SIZE, MMQ_NWARPS, 1);
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
2024-06-05 16:53:00 +02:00
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
if (!shmem_limit_raised[id]) {
2024-06-20 14:39:21 +02:00
CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
2024-06-05 16:53:00 +02:00
shmem_limit_raised[id] = true;
}
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
2024-06-20 14:39:21 +02:00
const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
const dim3 block_nums_xy_tiling(nty, ntx, 1);
const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
if (!use_stream_k) {
if (args.ne01 % mmq_y == 0) {
constexpr bool need_check = false;
mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
(args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
} else {
constexpr bool need_check = true;
mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
(args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
}
return;
}
const dim3 block_nums_mmq(nsm, 1, 1);
ggml_cuda_pool & pool = ctx.pool();
ggml_cuda_pool_alloc<float> tmp_fixup(pool, block_nums_mmq.x * mmq_x*mmq_y);
2024-06-05 16:53:00 +02:00
if (args.ne01 % mmq_y == 0) {
2024-06-20 14:39:21 +02:00
constexpr bool need_check = false;
mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
(args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
(args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
2024-06-05 16:53:00 +02:00
} else {
2024-06-20 14:39:21 +02:00
constexpr bool need_check = true;
mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
(args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
(args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
2024-06-05 16:53:00 +02:00
}
}
template <ggml_type type>
2024-06-20 14:39:21 +02:00
void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
2024-06-14 18:41:49 +02:00
const int id = ggml_cuda_get_device();
const int nsm = ggml_cuda_info().devices[id].nsm;
const int cc = ggml_cuda_info().devices[id].cc;
const int smpbo = ggml_cuda_info().devices[id].smpbo;
2024-06-05 16:53:00 +02:00
const int mmq_x_max = get_mmq_x_max_host(cc);
2024-06-20 14:39:21 +02:00
const int mmq_y = get_mmq_y_host(cc);
2024-06-05 16:53:00 +02:00
const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
2024-06-20 14:39:21 +02:00
const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
2024-06-05 16:53:00 +02:00
int mmq_x_best = 0;
2024-06-20 14:39:21 +02:00
int nparts_best = INT_MAX;
for (int mmq_x = 8; mmq_x <= mmq_x_max && nparts_best > 1; mmq_x += 8) {
2024-06-24 12:41:23 +02:00
const int granularity = mmq_get_granularity_host(mmq_x, cc);
if (mmq_x % granularity != 0 || mmq_get_shmem<type>(mmq_x, mmq_y, cc) > smpbo) {
continue;
}
2024-06-20 14:39:21 +02:00
const int ntiles_x = (args.ne11 + mmq_x - 1) / mmq_x;
const int nwaves_xy_tiling = ntiles_x*block_num_y;
const int nparts = use_stream_k ? ntiles_x : nwaves_xy_tiling;
2024-06-05 16:53:00 +02:00
2024-06-24 12:41:23 +02:00
if (nparts < nparts_best) {
2024-06-05 16:53:00 +02:00
mmq_x_best = mmq_x;
2024-06-20 14:39:21 +02:00
nparts_best = nparts;
2024-06-05 16:53:00 +02:00
}
}
switch (mmq_x_best) {
case 8:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 8>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 16:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 16>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 24:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 24>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 32:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 32>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 40:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 40>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 48:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 48>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 56:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 56>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 64:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 64>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 72:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 72>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 80:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 80>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 88:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 88>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 96:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 96>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 104:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 104>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 112:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 112>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 120:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 120>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
case 128:
2024-06-20 14:39:21 +02:00
launch_mul_mat_q<type, 128>(ctx, args, stream);
2024-06-05 16:53:00 +02:00
break;
default:
2024-06-14 18:41:49 +02:00
fprintf(stderr, "mmq_x_best=%d\n", mmq_x_best);
2024-06-05 16:53:00 +02:00
GGML_ASSERT(false);
break;
}
}
#define DECL_MMQ_CASE(type) \
2024-06-20 14:39:21 +02:00
template void mul_mat_q_case<type>(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) \
2024-06-05 16:53:00 +02:00
extern DECL_MMQ_CASE(GGML_TYPE_Q4_0);
extern DECL_MMQ_CASE(GGML_TYPE_Q4_1);
extern DECL_MMQ_CASE(GGML_TYPE_Q5_0);
extern DECL_MMQ_CASE(GGML_TYPE_Q5_1);
extern DECL_MMQ_CASE(GGML_TYPE_Q8_0);
extern DECL_MMQ_CASE(GGML_TYPE_Q2_K);
extern DECL_MMQ_CASE(GGML_TYPE_Q3_K);
extern DECL_MMQ_CASE(GGML_TYPE_Q4_K);
extern DECL_MMQ_CASE(GGML_TYPE_Q5_K);
extern DECL_MMQ_CASE(GGML_TYPE_Q6_K);
// -------------------------------------------------------------------------------------------------------------------------
2024-03-25 13:50:23 +01:00
void ggml_cuda_op_mul_mat_q(
ggml_backend_cuda_context & ctx,
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
const int64_t src1_padded_row_size, cudaStream_t stream);
2024-06-24 17:43:42 +02:00
bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11);