Improve cuBLAS performance by dequantizing on the GPU (#1065)

CMakeLists.txt

@@ -110,6 +110,7 @@ if (APPLE AND LLAMA_ACCELERATE)
         message(WARNING "Accelerate framework not found")
     endif()
 endif()
+
 if (LLAMA_OPENBLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
@@ -150,6 +151,10 @@ if (LLAMA_CUBLAS)
     if (CUDAToolkit_FOUND)
         message(STATUS "cuBLAS found")
 
+        enable_language(CUDA)
+
+        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+
         add_compile_definitions(GGML_USE_CUBLAS)
 
         if (LLAMA_STATIC)
@@ -241,21 +246,26 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
     message(STATUS "x86 detected")
     if (MSVC)
         if (LLAMA_AVX512)
-            add_compile_options(/arch:AVX512)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX512>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)
             # MSVC has no compile-time flags enabling specific
             # AVX512 extensions, neither it defines the
             # macros corresponding to the extensions.
             # Do it manually.
             if (LLAMA_AVX512_VBMI)
-                add_compile_definitions(__AVX512VBMI__)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
             endif()
             if (LLAMA_AVX512_VNNI)
-                add_compile_definitions(__AVX512VNNI__)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
             endif()
         elseif (LLAMA_AVX2)
-            add_compile_options(/arch:AVX2)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX2>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX2>)
         elseif (LLAMA_AVX)
-            add_compile_options(/arch:AVX)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX>)
         endif()
     else()
         if (LLAMA_F16C)
@@ -292,7 +302,8 @@ endif()
 
 add_library(ggml OBJECT
             ggml.c
-            ggml.h)
+            ggml.h
+            ${GGML_CUDA_SOURCES})
 
 target_include_directories(ggml PUBLIC .)
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
@@ -314,6 +325,14 @@ if (BUILD_SHARED_LIBS)
     target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
 endif()
 
+if (GGML_CUDA_SOURCES)
+    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
+    set_property(TARGET ggml PROPERTY CUDA_ARCHITECTURES OFF)
+    set_property(TARGET ggml PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
+    set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES OFF)
+endif()
+
+
 #
 # programs, examples and tests
 #

Makefile

@@ -1,3 +1,6 @@
+# Define the default target now so that it is always the first target
+default: main quantize quantize-stats perplexity embedding vdot
+
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
@@ -100,6 +103,9 @@ endif
 ifdef LLAMA_CUBLAS
 	CFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include
 	LDFLAGS += -lcublas_static -lculibos -lcudart_static -lcublasLt_static -lpthread -ldl -L/usr/local/cuda/lib64
+	OBJS	+= ggml-cuda.o
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	nvcc -arch=native -c -o $@ $<
 endif
 ifdef LLAMA_GPROF
 	CFLAGS   += -pg
@@ -137,8 +143,6 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main quantize quantize-stats perplexity embedding vdot
-
 #
 # Build library
 #
@@ -155,35 +159,35 @@ common.o: examples/common.cpp examples/common.h
 clean:
 	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-q4_0-matmult
 
-main: examples/main/main.cpp ggml.o llama.o common.o
+main: examples/main/main.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-quantize: examples/quantize/quantize.cpp ggml.o llama.o
+quantize: examples/quantize/quantize.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o
+quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
+perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
+embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-vdot: pocs/vdot/vdot.cpp ggml.o
+vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-libllama.so: llama.o ggml.o
+libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 #
 # Tests
 #
 
-benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o
+benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o benchmark-q4_0-matmult $(LDFLAGS)
 	./benchmark-q4_0-matmult
 

ggml-cuda.cu

@@ -0,0 +1,116 @@
+#include <stdint.h>
+#include <cuda_fp16.h>
+#include "ggml-cuda.h"
+
+typedef uint16_t ggml_fp16_t;
+static_assert(sizeof(__half) == sizeof(ggml_fp16_t), "wrong fp16 size");
+
+#define QK4_0 32
+typedef struct {
+    float   d;              // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    float   d;              // delta
+    float   m;              // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK4_2 16
+typedef struct {
+    __half d;               // delta
+    uint8_t qs[QK4_2 / 2];  // nibbles / quants
+} block_q4_2;
+static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding");
+
+
+static __global__ void dequantize_block_q4_0(const void * vx, float * y) {
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_0; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = (vi0 - 8)*d;
+        const float v1 = (vi1 - 8)*d;
+
+        y[i*QK4_0 + l + 0] = v0;
+        y[i*QK4_0 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q4_1(const void * vx, float * y) {
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+    const float m = x[i].m;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_1; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = vi0*d + m;
+        const float v1 = vi1*d + m;
+
+        y[i*QK4_1 + l + 0] = v0;
+        y[i*QK4_1 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q4_2(const void * vx, float * y) {
+    const block_q4_2 * x = (const block_q4_2 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_2; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = (vi0 - 8)*d;
+        const float v1 = (vi1 - 8)*d;
+
+        y[i*QK4_2 + l + 0] = v0;
+        y[i*QK4_2 + l + 1] = v1;
+    }
+}
+
+extern "C" {
+    __host__ void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_0;
+        dequantize_block_q4_0<<<nb, 1, 0, stream>>>(vx, y);
+    }
+
+    __host__ void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_1;
+        dequantize_block_q4_1<<<nb, 1, 0, stream>>>(vx, y);
+    }
+
+    __host__ void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_2;
+        dequantize_block_q4_2<<<nb, 1, 0, stream>>>(vx, y);
+    }
+}

ggml-cuda.h

@@ -0,0 +1,11 @@
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+
+#ifdef  __cplusplus
+}
+#endif

ggml.c

@@ -150,6 +150,8 @@ inline static void* ggml_aligned_malloc(size_t size) {
 #elif defined(GGML_USE_CUBLAS)
 #include <cublas_v2.h>
 #include <cuda_runtime.h>
+#include "ggml-cuda.h"
+
 #define CUDA_CHECK(err)                                                        \
     do {                                                                       \
         cudaError_t err_ = (err);                                              \
@@ -177,6 +179,7 @@ static void init_cublas(void) {
         CUBLAS_CHECK(cublasCreate(&cublasH));
 
         CUDA_CHECK(cudaStreamCreateWithFlags(&cudaStream, cudaStreamNonBlocking));
+
         CUBLAS_CHECK(cublasSetStream(cublasH, cudaStream));
 
         // configure logging to stdout
@@ -7311,7 +7314,6 @@ static void ggml_compute_forward_mul_mat_f32(
 
                 // copy data to host
                 CUDA_CHECK(cudaMemcpyAsync(d, d_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
-                CUDA_CHECK(cudaStreamSynchronize(cudaStream));
 #else
                 // zT = y * xT
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
@@ -7323,6 +7325,7 @@ static void ggml_compute_forward_mul_mat_f32(
             }
         }
 #if defined(GGML_USE_CUBLAS)
+        CUDA_CHECK(cudaStreamSynchronize(cudaStream));
         CUDA_CHECK(cudaFree(d_X));
         CUDA_CHECK(cudaFree(d_Y));
         CUDA_CHECK(cudaFree(d_D));
@@ -7535,7 +7538,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
 
                 // copy data to host
                 CUDA_CHECK(cudaMemcpyAsync(d, d_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
-                CUDA_CHECK(cudaStreamSynchronize(cudaStream));
 #else
                 const float * x = wdata;
                 const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
@@ -7553,6 +7555,7 @@ static void ggml_compute_forward_mul_mat_f16_f32(
         }
 
 #if defined(GGML_USE_CUBLAS)
+        CUDA_CHECK(cudaStreamSynchronize(cudaStream));
         CUDA_CHECK(cudaFree(d_X));
         CUDA_CHECK(cudaFree(d_Y));
         CUDA_CHECK(cudaFree(d_D));
@@ -7722,13 +7725,11 @@ static void ggml_compute_forward_mul_mat_q_f32(
             return;
         }
 
-        float * const wdata = params->wdata;
-        dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q;
-
 #if defined(GGML_USE_CUBLAS)
         float *d_X = NULL;
         float *d_Y = NULL;
         float *d_D = NULL;
+        float *d_Q = NULL;
         const float alpha = 1.0f;
         const float beta = 0.0f;
         const int x_ne = ne01 * ne10;
@@ -7738,10 +7739,41 @@ static void ggml_compute_forward_mul_mat_q_f32(
         CUDA_CHECK(cudaMalloc((void **)(&d_X), sizeof(float) * x_ne));
         CUDA_CHECK(cudaMalloc((void **)(&d_Y), sizeof(float) * y_ne));
         CUDA_CHECK(cudaMalloc((void **)(&d_D), sizeof(float) * d_ne));
+        CUDA_CHECK(cudaMalloc((void **)(&d_Q), GGML_TYPE_SIZE[type] * x_ne / GGML_BLCK_SIZE[type]));
+
+        void (*dequantize_row_q_cuda)(const void * x, float * y, int k, cudaStream_t stream)  = NULL;
+        if (type == GGML_TYPE_Q4_0) {
+            dequantize_row_q_cuda = dequantize_row_q4_0_cuda;
+        }
+        else if (type == GGML_TYPE_Q4_1) {
+            dequantize_row_q_cuda = dequantize_row_q4_1_cuda;
+        }
+        else if (type == GGML_TYPE_Q4_2) {
+            dequantize_row_q_cuda = dequantize_row_q4_2_cuda;
+        }
+        else {
+            GGML_ASSERT(false);
+        }
+#else
+        float * const wdata = params->wdata;
+        dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q;
 #endif
 
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
+                const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
+
+                float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+
+#if defined(GGML_USE_CUBLAS)
+                // copy and dequantize on device
+                CUDA_CHECK(
+                    cudaMemcpyAsync(d_Q, (char *) src0->data + i03*nb03 + i02*nb02,
+                        GGML_TYPE_SIZE[type] * x_ne / GGML_BLCK_SIZE[type], cudaMemcpyHostToDevice, cudaStream));
+
+                dequantize_row_q_cuda(d_Q, d_X, ne01 * ne00, cudaStream);
+                CUDA_CHECK(cudaGetLastError());
+#else
                 {
                     size_t id = 0;
                     for (int64_t i01 = 0; i01 < ne01; ++i01) {
@@ -7749,15 +7781,12 @@ static void ggml_compute_forward_mul_mat_q_f32(
                         id += ne00;
                     }
                 }
-
                 const float * x = wdata;
-                const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
+#endif
 
-                float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
 
 #if defined(GGML_USE_CUBLAS)
                 // copy data to device
-                CUDA_CHECK(cudaMemcpyAsync(d_X, x, sizeof(float) * x_ne, cudaMemcpyHostToDevice, cudaStream));
                 CUDA_CHECK(cudaMemcpyAsync(d_Y, y, sizeof(float) * y_ne, cudaMemcpyHostToDevice, cudaStream));
 
                 // compute
@@ -7770,7 +7799,6 @@ static void ggml_compute_forward_mul_mat_q_f32(
 
                 // copy data to host
                 CUDA_CHECK(cudaMemcpyAsync(d, d_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
-                CUDA_CHECK(cudaStreamSynchronize(cudaStream));
 #else
                 // zT = y * xT
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
@@ -7783,9 +7811,11 @@ static void ggml_compute_forward_mul_mat_q_f32(
         }
 
 #if defined(GGML_USE_CUBLAS)
+        CUDA_CHECK(cudaStreamSynchronize(cudaStream));
         CUDA_CHECK(cudaFree(d_X));
         CUDA_CHECK(cudaFree(d_Y));
         CUDA_CHECK(cudaFree(d_D));
+        CUDA_CHECK(cudaFree(d_Q));
 #endif
         //printf("CBLAS = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);