From b3204e76712b8f66218affdd4ad44ec221c6dcb6 Mon Sep 17 00:00:00 2001 From: giuros01 Date: Mon, 1 Apr 2019 13:50:22 +0100 Subject: COMPMID-2093: Implement CLGEMMNative Change-Id: I347130f6b5ae8d08b7c5c101b523b158565874a1 Signed-off-by: giuros01 Reviewed-on: https://review.mlplatform.org/c/1114 Comments-Addressed: Arm Jenkins Reviewed-by: Gian Marco Iodice Tested-by: Arm Jenkins --- src/core/CL/cl_kernels/gemm.cl | 319 ++++++++++++++++++++++++++++++++++++++++- 1 file changed, 318 insertions(+), 1 deletion(-) (limited to 'src/core/CL/cl_kernels/gemm.cl') diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl index c3107a20f2..da45d0fc18 100644 --- a/src/core/CL/cl_kernels/gemm.cl +++ b/src/core/CL/cl_kernels/gemm.cl @@ -1535,7 +1535,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs), #endif // M0 > 6 #if M0 > 7 VEC_DATA_TYPE(DATA_TYPE, 2) - a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin)); + a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7)); #endif // M0 > 7 LD_RHS_VFMA_M0xN0(0, a, c); @@ -1886,8 +1886,325 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), #undef RHS_OFFSET_X #undef RHS_STEP_X } + #endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(K) && defined(DATA_TYPE) +#if defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE) + +#define VFMA(a, b, c) \ + ({ \ + c = fma(a, b, c); \ + }) + +#if M0 == 1 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + }) +#elif M0 == 2 // M0 == 2 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + }) +#elif M0 == 3 // M0 == 3 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + }) +#elif M0 == 4 // M0 == 4 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + }) +#elif M0 == 5 // M0 == 5 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + }) +#elif M0 == 6 // M0 == 6 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ + }) +#elif M0 == 7 // M0 == 7 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ + }) +#elif M0 == 8 // M0 == 8 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \ + }) +#else // M0 not supported +#error "M0 not supported" +#endif // M0 not supported + +/** This OpenCL kernel computes the matrix multiplication between 2 matrices. + * The LHS matrix is NOT reshaped + * The RHS matrix is NOT reshaped + * + * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. + * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (i.e. -DM=52, -DN=30 and -DK=90) + * @note The number of columns of LHS matrix must be passed at compile time using -DK (i.e. -DK=64) + * @note The number of M0 rows to process must be passed at compile time using -DM0 (i.e. -DM0=2) + * @note The number of K0 partial accumulations must be passed at compile time using -DK0 (i.e., -DK0=2) + * @note The number of N0 columns to process must be passed at compile time using -DN0 (i.e. -DN0=2) + * @note Only the following configurations of M0, N0 and K0 are currently supported: + * - M0 = 1, 2, 3, 4, 5, 6, 7, 8 + * - N0 = 2, 3, 4, 8, 16 + * - K0 = 2, 3, 4, 8, 16 + * + * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: + * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D + * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D + * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. + * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor + * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix + * + * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F16/F32 + * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) + * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) + * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix + * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr + * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) + * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) + * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix + * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr + * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix + * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) + * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) + * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + */ +__kernel void gemm_mm_native(IMAGE_DECLARATION(lhs), + IMAGE_DECLARATION(rhs), + IMAGE_DECLARATION(dst), + uint lhs_stride_z, + uint rhs_stride_z, + uint dst_stride_z +#if defined(REINTERPRET_INPUT_AS_3D) + , + uint lhs_cross_plane_pad +#endif // REINTERPRET_INPUT_AS_3D +#if defined(REINTERPRET_OUTPUT_AS_3D) + , + uint dst_cross_plane_pad +#endif // REINTERPRET_OUTPUT_AS_3D + ) +{ + // Block size +#define RHS_BLOCK_SIZE ((K0) * (N0)) + + // RHS offset and step X +#define RHS_OFFSET_X (RHS_BLOCK_SIZE) + + uint x = get_global_id(0); + uint y = get_global_id(1); + uint z = get_global_id(2); + +#if defined(DUMMY_WORK_ITEMS) + if((x * N0 >= N) || (y * M0 >= M)) + { + return; + } +#endif // defined(DUMMY_WORK_ITEMS) + + // Compute LHS matrix address + uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + + // Compute RHS matrix address + uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE); + +#if defined(MATRIX_B_DEPTH) + // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 + rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z; +#else // defined(MATRIX_B_DEPTH) + rhs_offset += z * rhs_stride_z; +#endif // defined(MATRIX_B_DEPTH) + + REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0; + REPEAT_VAR_INIT_TO_CONST(16, uint, zrhs, 0); + +#if defined(REINTERPRET_INPUT_AS_3D) + // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zlhs, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply lhs_stride_z by DEPTH_GEMM3D + lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_INPUT_AS_3D) + + // Add offset for batched GEMM + lhs_offset += z * lhs_stride_z; + +#endif // defined(REINTERPRET_INPUT_AS_3D) + + // Initialize the accumulators + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(N0-1)=0; + + int i = 0; + for(; i <= (K - K0); i += K0) + { + // Supported cases (M0, K0): + // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 + // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 + // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 + // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 + // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 + // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 + // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 + // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 + // Load values from LHS matrix + LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); + + // Load values from RHS matrix + LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zrhs); + + RHS_VFMA_M0xN0(0, a, b0, c); + RHS_VFMA_M0xN0(1, a, b1, c); +#if K0 > 2 + RHS_VFMA_M0xN0(2, a, b2, c); +#endif // K0 > 2 +#if K0 > 3 + RHS_VFMA_M0xN0(3, a, b3, c); +#endif // K0 > 3 +#if K0 > 4 + RHS_VFMA_M0xN0(4, a, b4, c); + RHS_VFMA_M0xN0(5, a, b5, c); + RHS_VFMA_M0xN0(6, a, b6, c); + RHS_VFMA_M0xN0(7, a, b7, c); +#endif // K0 > 4 +#if K0 > 8 + RHS_VFMA_M0xN0(8, a, b8, c); + RHS_VFMA_M0xN0(9, a, b9, c); + RHS_VFMA_M0xN0(A, a, b10, c); + RHS_VFMA_M0xN0(B, a, b11, c); + RHS_VFMA_M0xN0(C, a, b12, c); + RHS_VFMA_M0xN0(D, a, b13, c); + RHS_VFMA_M0xN0(E, a, b14, c); + RHS_VFMA_M0xN0(F, a, b15, c); +#endif // K0 > 8 + + lhs_offset += K0 * sizeof(DATA_TYPE); + rhs_offset += K0 * rhs_stride_y; + } + + // Left-over accumulations + for(; i < K; ++i) + { + // Load values from LHS matrix + VEC_DATA_TYPE(DATA_TYPE, 2) + a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0)); +#if M0 > 1 + VEC_DATA_TYPE(DATA_TYPE, 2) + a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1)); +#endif // M0 > 1 +#if M0 > 2 + VEC_DATA_TYPE(DATA_TYPE, 2) + a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2)); +#endif // M0 > 2 +#if M0 > 3 + VEC_DATA_TYPE(DATA_TYPE, 2) + a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3)); +#endif // M0 > 3 +#if M0 > 4 + VEC_DATA_TYPE(DATA_TYPE, 2) + a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4)); +#endif // M0 > 4 +#if M0 > 5 + VEC_DATA_TYPE(DATA_TYPE, 2) + a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5)); +#endif // M0 > 5 +#if M0 > 6 + VEC_DATA_TYPE(DATA_TYPE, 2) + a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6)); +#endif // M0 > 6 +#if M0 > 7 + VEC_DATA_TYPE(DATA_TYPE, 2) + a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7)); +#endif // M0 > 7 + + VEC_DATA_TYPE(DATA_TYPE, N0) + b = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * rhs_stride_y)); + RHS_VFMA_M0xN0(0, a, b, c); + + lhs_offset += sizeof(DATA_TYPE); + rhs_offset += rhs_stride_y; + } + + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * dst_stride_y); + + REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; + +#if defined(REINTERPRET_OUTPUT_AS_3D) + // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply dst_stride_z by DEPTH_GEMM3D + dst_addr += z * dst_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_OUTPUT_AS_3D) + + // Add offset for batched GEMM + dst_addr += z * dst_stride_z; + +#endif // defined(REINTERPRET_OUTPUT_AS_3D) + + // Multiply by the weight of matrix-matrix product and store the result + // Multiply by the weight of matrix-matrix product and store the result +#if defined(ALPHA) + SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); +#endif // defined(ALPHA) + + // Store output block + STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout); + +#undef RHS_BLOCK_SIZE +#undef RHS_OFFSET_X +#undef RHS_STEP_X +} +#endif // defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE) + #if defined(TRANSPOSE_W) && defined(MULT_TRANSPOSE1XW_WIDTH) #if ELEMENT_SIZE == 1 -- cgit v1.2.1