cuda : performance optimizations (#1530)

* xor hack

* block y dim

* loop unrolling

* Fixed cmake LLAMA_CUDA_BY option

* Removed hipblas compatibility code

* Define GGML_CUDA_DMMV_BLOCK_Y if not defined

* Fewer iters, more ops per iter

* Renamed DMMV X/Y compilation options

CMakeLists.txt

@@ -68,6 +68,8 @@ option(LLAMA_ACCELERATE             "llama: enable Accelerate framework"
 option(LLAMA_BLAS                       "llama: use BLAS"                                       OFF)
 option(LLAMA_BLAS_VENDOR                "llama: BLA_VENDOR from https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors" Generic)
 option(LLAMA_CUBLAS                     "llama: use cuBLAS"                                     OFF)
+set(LLAMA_CUDA_DMMV_X "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
+set(LLAMA_CUDA_DMMV_Y "1" CACHE STRING  "llama: y block size for dmmv CUDA kernels")
 option(LLAMA_CLBLAST                    "llama: use CLBlast"                                    OFF)
 
 option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
@@ -184,6 +186,8 @@ if (LLAMA_CUBLAS)
         set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        add_compile_definitions(GGML_CUDA_DMMV_Y=${LLAMA_CUDA_DMMV_Y})
 
         if (LLAMA_STATIC)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)

Makefile

@@ -133,9 +133,19 @@ ifdef LLAMA_CUBLAS
 	OBJS      += ggml-cuda.o
 	NVCC      = nvcc
 	NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native
+ifdef LLAMA_CUDA_DMMV_X
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+else
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+endif # LLAMA_CUDA_DMMV_X
+ifdef LLAMA_CUDA_DMMV_Y
+	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=$(LLAMA_CUDA_DMMV_Y)
+else
+	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=1
+endif # LLAMA_CUDA_DMMV_Y
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
-endif
+endif # LLAMA_CUBLAS
 ifdef LLAMA_CLBLAST
 	CFLAGS  += -DGGML_USE_CLBLAST
 	CXXFLAGS  += -DGGML_USE_CLBLAST

ggml-cuda.cu

@@ -83,9 +83,19 @@ typedef struct {
 } block_q8_0;
 static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
 
+#define WARP_SIZE 32
+
 #define CUDA_MUL_BLOCK_SIZE 256
+
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
-#define CUDA_DMMV_BLOCK_SIZE 32 // dmmv = dequantize_mul_mat_vec
+
+// dmmv = dequantize_mul_mat_vec
+#ifndef GGML_CUDA_DMMV_X
+#define GGML_CUDA_DMMV_X 32
+#endif
+#ifndef GGML_CUDA_DMMV_Y
+#define GGML_CUDA_DMMV_Y 1
+#endif
 
 static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -200,41 +210,51 @@ static __global__ void dequantize_block(const void * vx, float * y, const int k)
     dequantize_kernel(vx, ib, iqs, v0, v1);
 }
 
-template <int block_size, int qk, int qr, dequantize_kernel_t dequantize_kernel>
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
 static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y, float * dst, const int ncols) {
-    const int row = blockIdx.x;
+    // qk = quantized weights per x block
+    // qr = number of quantized weights per data value in x block
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
+    const int iter_stride = 2*GGML_CUDA_DMMV_X;
+    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
     const int y_offset = qr == 1 ? 1 : qk/2;
 
-    __shared__ float tmp[block_size]; // separate sum for each thread
+    float tmp = 0; // partial sum for thread in warp
-    tmp[tid] = 0;
 
-    for (int i = 0; i < ncols/block_size; i += 2) {
+    for (int i = 0; i < ncols; i += iter_stride) {
-        const int col = i*block_size + 2*tid;
+        const int col = i + vals_per_iter*tid;
-        const int ib = (row*ncols + col)/qk; // block index
+        const int ib = (row*ncols + col)/qk; // x block index
-        const int iqs = (col%qk)/qr; // quant index
+        const int iqs = (col%qk)/qr; // x quant index
         const int iybs = col - col%qk; // y block start index
 
+// processing >2 values per i iter is faster for fast GPUs
+#pragma unroll
+        for (int j = 0; j < vals_per_iter; j += 2) {
+            // process 2 vals per j iter
+
             // dequantize
             float v0, v1;
-        dequantize_kernel(vx, ib, iqs, v0, v1);
+            dequantize_kernel(vx, ib, iqs + j/qr, v0, v1);
+            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
 
             // matrix multiplication
-        tmp[tid] += v0 * y[iybs + iqs + 0];
+            tmp += v0 * y[iybs + iqs + j/qr + 0];
-        tmp[tid] += v1 * y[iybs + iqs + y_offset];
+            tmp += v1 * y[iybs + iqs + j/qr + y_offset];
+            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
+        }
     }
 
     // sum up partial sums and write back result
     __syncthreads();
-    for (int s=block_size/2; s>0; s>>=1) {
+#pragma unroll
-        if (tid < s) {
+    for (int mask = 16; mask > 0; mask >>= 1) {
-            tmp[tid] += tmp[tid + s];
+        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-        }
-        __syncthreads();
     }
+
     if (tid == 0) {
-        dst[row] = tmp[0];
+        dst[row] = tmp;
     }
 }
 
@@ -269,33 +289,43 @@ static void dequantize_row_q8_0_cuda(const void * vx, float * y, const int k, cu
 }
 
 static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK4_0, QR4_0, dequantize_q4_0>
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK4_1, QR4_1, dequantize_q4_1>
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK5_0, QR5_0, dequantize_q5_0>
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK5_1, QR5_1, dequantize_q5_1>
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK8_0, QR8_0, dequantize_q8_0>
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
@@ -304,9 +334,11 @@ static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, c
 }
 
 static void convert_mul_mat_vec_f16_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, 32, 1, convert_f16>
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<1, 1, convert_f16>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {