diff options
Diffstat (limited to 'src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp')
| -rw-r--r-- | src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp | 42 |
1 files changed, 30 insertions, 12 deletions
diff --git a/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp b/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp index 1e06e84d4d..89cad30214 100644 --- a/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp +++ b/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp @@ -26,6 +26,7 @@ #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/TensorShape.h" + #include "src/gpu/cl/kernels/ClMatMulNativeKernel.h" #include <limits> @@ -37,22 +38,32 @@ namespace cl_matmul { MatMulKernelInfo select_info(const MatMulKernelInfo &info0, const MatMulKernelInfo &info1, - unsigned int m, unsigned int n, unsigned int k, unsigned int b, DataType data_type, bool rhs_lock_padding) + unsigned int m, + unsigned int n, + unsigned int k, + unsigned int b, + DataType data_type, + bool rhs_lock_padding) { - ARM_COMPUTE_ERROR_ON_MSG(info1.export_rhs_to_cl_image == true, "The fallback MatMul configuration cannot have export_to_cl_image = true"); - ARM_COMPUTE_ERROR_ON_MSG(info0.adj_lhs != info1.adj_lhs, "The MatMul configurations must have the same adj_lhs value"); - ARM_COMPUTE_ERROR_ON_MSG(info0.adj_rhs != info1.adj_rhs, "The MatMul configurations must have the same adj_rhs value"); + ARM_COMPUTE_ERROR_ON_MSG(info1.export_rhs_to_cl_image == true, + "The fallback MatMul configuration cannot have export_to_cl_image = true"); + ARM_COMPUTE_ERROR_ON_MSG(info0.adj_lhs != info1.adj_lhs, + "The MatMul configurations must have the same adj_lhs value"); + ARM_COMPUTE_ERROR_ON_MSG(info0.adj_rhs != info1.adj_rhs, + "The MatMul configurations must have the same adj_rhs value"); const bool adj_lhs = info0.adj_lhs; const bool adj_rhs = info0.adj_rhs; - TensorInfo lhs_info = !adj_lhs ? TensorInfo(TensorShape(k, m, b), 1, data_type) : TensorInfo(TensorShape(m, k, b), 1, data_type); - TensorInfo rhs_info = !adj_rhs ? TensorInfo(TensorShape(n, k, b), 1, data_type) : TensorInfo(TensorShape(k, n, b), 1, data_type); + TensorInfo lhs_info = + !adj_lhs ? TensorInfo(TensorShape(k, m, b), 1, data_type) : TensorInfo(TensorShape(m, k, b), 1, data_type); + TensorInfo rhs_info = + !adj_rhs ? TensorInfo(TensorShape(n, k, b), 1, data_type) : TensorInfo(TensorShape(k, n, b), 1, data_type); TensorInfo dst_info; - if(rhs_lock_padding == false) + if (rhs_lock_padding == false) { - if(bool(opencl::kernels::ClMatMulNativeKernel::validate(&lhs_info, &rhs_info, nullptr, &dst_info, info0))) + if (bool(opencl::kernels::ClMatMulNativeKernel::validate(&lhs_info, &rhs_info, nullptr, &dst_info, info0))) { return info0; } @@ -67,7 +78,13 @@ MatMulKernelInfo select_info(const MatMulKernelInfo &info0, } } -MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, bool adj_lhs, bool adj_rhs, unsigned int m, unsigned int n, unsigned int k, unsigned int b) +MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, + bool adj_lhs, + bool adj_rhs, + unsigned int m, + unsigned int n, + unsigned int k, + unsigned int b) { size_t min_acc = std::numeric_limits<size_t>::max(); size_t min_idx = 0; @@ -76,12 +93,13 @@ MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, bool adj_lh const size_t num_rows = configs.size(); const size_t num_cols = configs[0].size(); - ARM_COMPUTE_ERROR_ON_MSG(num_cols != 8U, "The entry should have 8 integer values representing: M, N, K, B, M0, N0. K0, IMG_RHS"); + ARM_COMPUTE_ERROR_ON_MSG(num_cols != 8U, + "The entry should have 8 integer values representing: M, N, K, B, M0, N0. K0, IMG_RHS"); ARM_COMPUTE_UNUSED(num_cols); // Find nearest GeMM workload // Note: the workload does not depend on the K dimension - for(size_t y = 0; y < num_rows; ++y) + for (size_t y = 0; y < num_rows; ++y) { size_t mc0 = static_cast<size_t>(configs[y][0]); size_t nc0 = static_cast<size_t>(configs[y][1]); @@ -94,7 +112,7 @@ MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, bool adj_lh acc += (k - kc0) * (k - kc0); acc += (b - bc0) * (b - bc0); acc = std::sqrt(acc); - if(acc < min_acc) + if (acc < min_acc) { min_acc = acc; min_idx = y; |