From 71cbd28b7cf5115b0451d43e5c84cce4ae4d8ec7 Mon Sep 17 00:00:00 2001 From: SiCongLi Date: Wed, 3 Nov 2021 12:17:06 +0000 Subject: Fix out-of-bound reads in cl gemm kernels * Revert "Remove padding in FP Cl Gemm kernels" This reverts commit 48717a3d38fef8d316cd4b9fd9a3bc1a43db736b. * Allow different boundary row handling strategies across native, reshaped and reshaped_only_rhs kernels by introducing a ELTWISE_OPERAND_ROW parameter to the macro Resolves COMPMID-4919 Change-Id: Icefc23c0760a6abb838fef1d0d5bda06b07c79e3 Signed-off-by: SiCongLi Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6569 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice --- .../fp_post_ops_act_eltwise_op_act.h | 20 +++--- .../act_eltwise_op_act/gemm_mm_native.cl | 24 ++++--- .../act_eltwise_op_act/gemm_mm_reshaped.cl | 8 +-- .../gemm_mm_reshaped_only_rhs.cl | 76 ++++++++++------------ 4 files changed, 62 insertions(+), 66 deletions(-) (limited to 'src/core/CL/cl_kernels/common/experimental') diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h index fc9704f13b..070c47bd7f 100644 --- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h +++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h @@ -54,6 +54,9 @@ * @param[in] N0 The number of consecutive columns * @param[in] BASENAME The basename of the result variables * @param[in] ELTWISE_OPERAND_NAME The basename of the other operand variables + * @param[in] ELTWISE_OPERAND_ROW The starting row of the other operand variables. Required as different boundary handling strategies are used by different kernels + * E.g. reshaped_only_rhs and native kernels shifts rows (by using COMPUTE_M0_START_ROW) to handle boundary rows, + * whereas reshaped kernels do not shift rows * @param[in] DATA_TYPE Data type of the result variables * @param[in] DATA_TYPE_ACCUMULATR Higher-precision accumulator data type in case of mixed-precision op * @param[in] ZERO Zero vector for z offset @@ -66,26 +69,25 @@ #if defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) #if P2_ELTWISE_ARG1_HEIGHT == 1 #if P2_ELTWISE_ARG1_WIDTH == 1 // Case 1: Broadcasting in both X and Y; op2 arg tile shape[YxX] == [1x1] -#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ - __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z; \ - VEC_DATA_TYPE(DATA_TYPE, 1) \ - ELTWISE_OPERAND_NAME##0 = VLOAD(1)(0, (__global DATA_TYPE *)ELTWISE_OPERAND_NAME##_addr); \ +#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_ROW, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ + __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z; \ + VEC_DATA_TYPE(DATA_TYPE, 1) \ + ELTWISE_OPERAND_NAME##0 = VLOAD(1)(0, (__global DATA_TYPE *)ELTWISE_OPERAND_NAME##_addr); \ MIXED_PRECISION_ELTWISE_OP_BLOCK_BROADCAST(OP, M0, 1, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE_ACCUMULATOR, ELTWISE_OPERAND_NAME##_hp); #else // P2_ELTWISE_ARG1_WIDTH == 1; Case 2: Broadcasting in only Y; op2 arg tile shape[YxX] == [1xN0] -#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ +#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_ROW, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z; \ LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_NAME##_addr, 0, ELTWISE_OPERAND_NAME##_stride_y, ZERO, 1, PARTIAL_LOAD_N0, false, PARTIAL_COND_X); \ MIXED_PRECISION_ELTWISE_OP_BLOCK_BROADCAST(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE_ACCUMULATOR, ELTWISE_OPERAND_NAME##_hp); #endif // P2_ELTWISE_ARG1_WIDTH == 1 #else // P2_ELTWISE_ARG1_HEIGHT == 1; Case 3: No broadcasting; op2 arg tile shape[YxX] == [M0xN0] -#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ - __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * ELTWISE_OPERAND_NAME##_stride_y) + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z; \ - LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_NAME##_addr, 0, ELTWISE_OPERAND_NAME##_stride_y, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X); \ +#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_ROW, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ + __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (ELTWISE_OPERAND_ROW * ELTWISE_OPERAND_NAME##_stride_y) + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z; \ + LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_NAME##_addr, 0, ELTWISE_OPERAND_NAME##_stride_y, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X); \ MIXED_PRECISION_ELTWISE_OP_BLOCK(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE_ACCUMULATOR, ELTWISE_OPERAND_NAME##_hp); #endif // P2_ELTWISE_ARG1_HEIGHT == 1 #endif // defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) /** @} */ // end of group POST_OP2_ELTWISE_OP - /** Post Op 3: Activation Block (Optional) * @name POST_OP3_ACTIVATION_OPTIONAL * Toggled by -DP3_ACTIVATION_TYPE diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl index e53ce3d1b2..bbe97b2781 100644 --- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl +++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl @@ -170,7 +170,7 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), #endif // defined(DUMMY_WORK_ITEMS) // Compute LHS matrix address - uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; // Compute RHS matrix address uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE); @@ -187,7 +187,7 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), #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 * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -294,17 +294,13 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), rhs_offset += rhs_stride_y; } - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y); + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); - // Boundary conditions: detect if current block is at the "bottom" or "right" boundary - const bool cond_y = ((y + 1) * M0 >= M); - const bool cond_x = ((x + 1) * N0 >= N); - #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 * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -327,7 +323,7 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), #if defined(BROADCAST_BIAS) __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); - LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x); + LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(1, DATA_TYPE, bias, BETA); @@ -337,10 +333,9 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), ADD_BLOCK_BROADCAST(M0, c, bias0); #else // defined(BROADCAST_BIAS) - __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id( - 2) * bias_stride_z; + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; - LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); @@ -352,10 +347,13 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) + const bool cond_y = y == 0; + const bool cond_x = ((x + 1) * N0 >= N); + // c = act(c) POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) - POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x); // c = act(c) POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl index 758fd327fe..9e9a73ccf6 100644 --- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl +++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl @@ -409,7 +409,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_post_act_eltwise_op_act(IMAGE_DECLAR // c = act(c) POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) - POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, get_global_id(1) * (uint)M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); // c = act(c) POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); @@ -626,7 +626,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture_post_act_eltwise_op_act(IMAG // c = act(c) POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) - POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, get_global_id(1) * (uint)M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); // c = act(c) POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); @@ -1069,7 +1069,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLAR // c = act(c) POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) - POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, get_global_id(1) * (uint)M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); // c = act(c) POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); @@ -1384,7 +1384,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture_post_act_eltwise_op_act(IMAG // c = act(c) POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) - POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, get_global_id(1) * (uint)M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); // c = act(c) POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl index 508ee96d2d..fe2d103de5 100644 --- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl +++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl @@ -214,6 +214,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT uint y = get_global_id(1); uint z = get_global_id(2); + const bool cond_y = y == 0; + const bool cond_x = ((x + 1) * N0 >= N); + #if defined(DUMMY_WORK_ITEMS) if((x * N0 >= N) || (y * M0 >= M)) { @@ -222,7 +225,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT #endif // defined(DUMMY_WORK_ITEMS) // Compute LHS matrix address - uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; // Compute RHS reshaped matrix address uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y; @@ -239,7 +242,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT #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 * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -338,18 +341,14 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT rhs_offset += sizeof(DATA_TYPE); } - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y); + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; - // Boundary conditions: detect if current block is at the "bottom" or "right" boundary - const bool cond_y = ((y + 1) * M0 >= M); - const bool cond_x = ((x + 1) * N0 >= N); - #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 * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -382,7 +381,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT ADD_BLOCK_BROADCAST(M0, c, bias0); #else // defined(BROADCAST_BIAS) - __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z; + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); @@ -399,7 +398,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT // c = act(c) POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) - POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x); // c = act(c) POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); @@ -480,6 +479,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ uint y = get_global_id(1); uint z = get_global_id(2); + const bool cond_y = y == 0; + const bool cond_x = ((x + 1) * N0 >= N); + #if defined(DUMMY_WORK_ITEMS) if((x * N0 >= N) || (y * M0 >= M)) { @@ -488,7 +490,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ #endif // defined(DUMMY_WORK_ITEMS) // Compute LHS matrix address - uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 @@ -506,7 +508,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ #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 * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -654,18 +656,14 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ #endif // LEFTOVER_K != 0 - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y); + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; - // Boundary conditions: detect if current block is at the "bottom" or "right" boundary - const bool cond_y = ((y + 1) * M0 >= M); - const bool cond_x = ((x + 1) * N0 >= N); - #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 * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -698,7 +696,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ ADD_BLOCK_BROADCAST(M0, c, bias0); #else // defined(BROADCAST_BIAS) - __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z; + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); @@ -715,7 +713,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ // c = act(c) POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) - POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x); // c = act(c) POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); @@ -870,6 +868,9 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA uint y = get_global_id(1); uint z = get_global_id(2); + const bool cond_y = y == 0; + const bool cond_x = ((x + 1) * N0 >= N); + #if defined(DUMMY_WORK_ITEMS) if((x * N0 >= N) || (y * M0 >= M)) { @@ -878,7 +879,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA #endif // defined(DUMMY_WORK_ITEMS) // Compute LHS matrix address - uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; // Compute RHS reshaped matrix address uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y; @@ -896,7 +897,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA #if defined(REINTERPRET_INPUT_AS_3D) // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D - CALCULATE_Z_OFFSET(M0, uint, zin, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -1020,17 +1021,13 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA rhs_offset += RHS_STEP_X * sizeof(DATA_TYPE); } - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y); + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; - // Boundary conditions: detect if current block is at the "bottom" or "right" boundary - const bool cond_y = ((y + 1) * M0 >= M); - const bool cond_x = ((x + 1) * N0 >= N); - #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 * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -1063,7 +1060,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA ADD_BLOCK_BROADCAST(M0, c, bias0); #else // defined(BROADCAST_BIAS) - __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z; + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); @@ -1080,7 +1077,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA // c = act(c) POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) - POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x); // c = act(c) POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); @@ -1157,6 +1154,9 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE uint y = get_global_id(1); uint z = get_global_id(2); + const bool cond_y = y == 0; + const bool cond_x = ((x + 1) * N0 >= N); + #if defined(DUMMY_WORK_ITEMS) if((x * N0 >= N) || (y * M0 >= M)) { @@ -1165,7 +1165,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE #endif // defined(DUMMY_WORK_ITEMS) // Compute LHS matrix address - uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 @@ -1184,7 +1184,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE #if defined(REINTERPRET_INPUT_AS_3D) // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D - CALCULATE_Z_OFFSET(M0, uint, zin, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -1299,17 +1299,13 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE x_rhs += RHS_STEP_X; } - __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y); + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; - // Boundary conditions: detect if current block is at the "bottom" or "right" boundary - const bool cond_y = ((y + 1) * M0 >= M); - const bool cond_x = ((x + 1) * N0 >= N); - #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 * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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 @@ -1342,7 +1338,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE ADD_BLOCK_BROADCAST(M0, c, bias0); #else // defined(BROADCAST_BIAS) - __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z; + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); @@ -1359,7 +1355,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE // c = act(c) POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) - POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x); // c = act(c) POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); -- cgit v1.2.1