diff options
author | Gunes Bayir <gunes.bayir@arm.com> | 2023-03-20 10:19:10 +0000 |
---|---|---|
committer | Gunes Bayir <gunes.bayir@arm.com> | 2023-03-24 11:35:03 +0000 |
commit | bbeef721c285d467d003a739a1e68b2c86899750 (patch) | |
tree | b298e2df7eacfa50ce3824a400c8c1ac82c5ebe9 /src/core | |
parent | 20cfa45faefbf56f62c8b1aa95dfd0b4f52e5641 (diff) | |
download | ComputeLibrary-bbeef721c285d467d003a739a1e68b2c86899750.tar.gz |
Add Texture Pipe Support for Matmul Lhs T/NT Rhs NT kernels
Resolves: COMPMID-5945, COMPMID-5954
Change-Id: I7b27021d21f8e08c4896f6b1f595a75125064f9e
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9356
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core')
-rw-r--r-- | src/core/CL/CLHelpers.cpp | 6 | ||||
-rw-r--r-- | src/core/CL/cl_kernels/common/mat_mul.cl | 28 |
2 files changed, 17 insertions, 17 deletions
diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp index b31864211c..6b011f1f7c 100644 --- a/src/core/CL/CLHelpers.cpp +++ b/src/core/CL/CLHelpers.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2022 Arm Limited. + * Copyright (c) 2016-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -443,7 +443,7 @@ void set_wbsm(cl::Kernel &kernel, cl_int wbsm_hint) bool export_to_cl_image(const ITensorInfo *tensor) { - if(tensor->tensor_shape()[0] % 4) + if(tensor->tensor_shape()[0] % 4 != 0) { return false; } @@ -467,7 +467,7 @@ bool export_to_cl_image(const ITensorInfo *tensor) } const size_t image_w = tensor->tensor_shape()[0] / 4; - const size_t image_h = tensor->tensor_shape()[1] * tensor->tensor_shape()[2] * tensor->tensor_shape()[3]; + const size_t image_h = tensor->tensor_shape().total_size() / tensor->tensor_shape()[0]; const size_t max_image_w = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>(); const size_t max_image_h = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>(); diff --git a/src/core/CL/cl_kernels/common/mat_mul.cl b/src/core/CL/cl_kernels/common/mat_mul.cl index 956d37a9d8..90ebf80a6a 100644 --- a/src/core/CL/cl_kernels/common/mat_mul.cl +++ b/src/core/CL/cl_kernels/common/mat_mul.cl @@ -33,10 +33,11 @@ * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) + * @note The tensor type ("BUFFER" or "IMAGE") of the rhs tensor must be passed at compile time using -DRHS_TENSOR_TYPE (e.g. -DRHS_TENSOR_TYPE=BUFFER) * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_NT_NT) * @note Only the following configurations of M0, N0 and K0 are currently supported: * - M0 > 0 - * - N0 = 1, 2, 3, 4, 8, 16 + * - N0 = 1, 2, 3, 4, 8, 16 (only 4, 8, 16 if RHS_TENSOR_TYPE=IMAGE) * - K0 = 1, 2, 3, 4, 8, 16 * @note Values > 8 for M0 are not expected to be efficient * @@ -47,6 +48,7 @@ * @param[in] lhs_h The height of the lhs tensor * @param[in] lhs_n Number of the matrices (buffers) in the batch * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix + * @param[in] rhs_img (Optional) Read only cl_image object for the rhs tensor. Included when RHS_TENSOR_TYPE=IMAGE * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) @@ -64,7 +66,7 @@ */ __kernel void mat_mul_native_nt_nt( TENSOR3D_T(lhs, BUFFER), - TENSOR3D_T(rhs, BUFFER), + TENSOR3D_T(rhs, RHS_TENSOR_TYPE), TENSOR3D_T(dst, BUFFER)) { const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0); @@ -73,7 +75,6 @@ __kernel void mat_mul_native_nt_nt( // Compute LHS/RHS/DST matrix address lhs_offset_first_element_in_bytes += y * lhs_stride_y + z * lhs_stride_z; - rhs_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + z * rhs_stride_z; dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z; // Initialize the accumulators @@ -84,6 +85,7 @@ __kernel void mat_mul_native_nt_nt( acc[i].v = 0.f; }) + const int rhs_z = z * rhs_h; int k; for(k = 0; k <= K - K0; k += K0) { @@ -102,12 +104,11 @@ __kernel void mat_mul_native_nt_nt( // Load tile from the lhs/rhs tensors T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); - T_LOAD(DATA_TYPE, K0, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); + T_LOAD(DATA_TYPE, K0, N0, RHS_TENSOR_TYPE, rhs, x, k + rhs_z, 1, rhs_stride_y, b); T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, NT, a, b, acc); lhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE); - rhs_offset_first_element_in_bytes += K0 * rhs_stride_y; } #ifdef K % K0 != 0 @@ -129,12 +130,11 @@ __kernel void mat_mul_native_nt_nt( // Load tile from the lhs/rhs tensors T_LOAD(DATA_TYPE, M0, 1, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); - T_LOAD(DATA_TYPE, 1, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); + T_LOAD(DATA_TYPE, 1, N0, BUFFER, rhs, x, k + rhs_z, 1, rhs_stride_y, b); T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, NT, a, b, acc); lhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE); - rhs_offset_first_element_in_bytes += 1 * rhs_stride_y; } #endif // K % K0 != 0 @@ -314,10 +314,11 @@ __kernel void mat_mul_native_nt_t(TENSOR3D_T(lhs, BUFFER), * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) + * @note The tensor type ("BUFFER" or "IMAGE") of the rhs tensor must be passed at compile time using -DRHS_TENSOR_TYPE (e.g. -DRHS_TENSOR_TYPE=BUFFER) * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_T_NT) * @note Only the following configurations of M0, N0 and K0 are currently supported: * - M0 = 1, 2, 3, 4, 8, 16 - * - N0 = 1, 2, 3, 4, 8, 16 + * - N0 = 1, 2, 3, 4, 8, 16 (only 4, 8, 16 if RHS_TENSOR_TYPE=IMAGE) * - K0 > 0 * * @note Values > 8 for M0, and K0 are not expected to be efficient * @@ -328,6 +329,7 @@ __kernel void mat_mul_native_nt_t(TENSOR3D_T(lhs, BUFFER), * @param[in] lhs_h The height of the lhs tensor * @param[in] lhs_n Number of the matrices (buffers) in the batch * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix + * @param[in] rhs_img (Optional) Read only cl_image object for the rhs tensor. Included when RHS_TENSOR_TYPE=IMAGE * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) @@ -345,7 +347,7 @@ __kernel void mat_mul_native_nt_t(TENSOR3D_T(lhs, BUFFER), */ __kernel void mat_mul_native_t_nt( TENSOR3D_T(lhs, BUFFER), - TENSOR3D_T(rhs, BUFFER), + TENSOR3D_T(rhs, RHS_TENSOR_TYPE), TENSOR3D_T(dst, BUFFER)) { const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0); @@ -354,7 +356,6 @@ __kernel void mat_mul_native_t_nt( // Compute LHS/RHS/DST matrix address lhs_offset_first_element_in_bytes += y * sizeof(DATA_TYPE) + z * lhs_stride_z; - rhs_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + z * rhs_stride_z; dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z; // Initialize the accumulators @@ -365,6 +366,7 @@ __kernel void mat_mul_native_t_nt( acc[i].v = 0.f; }) + const int rhs_z = z * rhs_h; int k; for(k = 0; k <= K - K0; k += K0) { @@ -383,7 +385,7 @@ __kernel void mat_mul_native_t_nt( // Load tile from the lhs/rhs tensors T_LOAD(DATA_TYPE, K0, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); - T_LOAD(DATA_TYPE, K0, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); + T_LOAD(DATA_TYPE, K0, N0, RHS_TENSOR_TYPE, rhs, x, k + rhs_z, 1, rhs_stride_y, b); #if GPU_ARCH == GPU_ARCH_MIDGARD // For explanation, see mat_mul_native_nt_t @@ -401,7 +403,6 @@ __kernel void mat_mul_native_t_nt( #endif // GPU_ARCH == GPU_ARCH_MIDGARD lhs_offset_first_element_in_bytes += K0 * lhs_stride_y; - rhs_offset_first_element_in_bytes += K0 * rhs_stride_y; } #ifdef K % K0 != 0 @@ -423,7 +424,7 @@ __kernel void mat_mul_native_t_nt( // Load tile from the lhs/rhs tensors T_LOAD(DATA_TYPE, 1, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); - T_LOAD(DATA_TYPE, 1, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); + T_LOAD(DATA_TYPE, 1, N0, BUFFER, rhs, x, k + rhs_z, 1, rhs_stride_y, b); #if GPU_ARCH == GPU_ARCH_MIDGARD // For explanation, see mat_mul_native_nt_t @@ -438,7 +439,6 @@ __kernel void mat_mul_native_t_nt( #endif // GPU_ARCH == GPU_ARCH_MIDGARD lhs_offset_first_element_in_bytes += 1 * lhs_stride_y; - rhs_offset_first_element_in_bytes += 1 * rhs_stride_y; } #endif // K % K0 != 0 |