diff options
Diffstat (limited to 'src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp')
-rw-r--r-- | src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp | 63 |
1 files changed, 41 insertions, 22 deletions
diff --git a/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp b/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp index 66331bc818..0bb6b0c083 100644 --- a/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp +++ b/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp @@ -29,17 +29,16 @@ #include "arm_compute/core/QuantizationInfo.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/utils/ActivationFunctionUtils.h" -#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/core/utils/helpers/AdjustVecSize.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "arm_compute/core/utils/StringUtils.h" #include "src/common/utils/Log.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "src/gpu/cl/ClCompileContext.h" #include "src/gpu/cl/kernels/helpers/MatMulKernelHelpers.h" - #include "support/Cast.h" #include "support/StringSupport.h" @@ -62,51 +61,62 @@ Status validate_matmul_kernel_info(const MatMulKernelInfo &matmul_kernel_info) // Validate M0 ARM_COMPUTE_RETURN_ERROR_ON_MSG(m0 < 1, "Only positive integers are supported for M0"); - if(adj_lhs) + if (adj_lhs) { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((m0 & (m0 - 1)) && (m0 != 3)) || (m0 > 16), "Only 1,2,3,4,8,16 are supported for M0 for Lhs transposed"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((m0 & (m0 - 1)) && (m0 != 3)) || (m0 > 16), + "Only 1,2,3,4,8,16 are supported for M0 for Lhs transposed"); } // Validate N0 ARM_COMPUTE_RETURN_ERROR_ON_MSG(n0 < 1, "Only positive integers are supported for N0"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((n0 & (n0 - 1)) && (n0 != 3)) || (n0 > 16), "Only 1,2,3,4,8,16 are supported for N0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((n0 & (n0 - 1)) && (n0 != 3)) || (n0 > 16), + "Only 1,2,3,4,8,16 are supported for N0"); // Validate K0 ARM_COMPUTE_RETURN_ERROR_ON_MSG(k0 < 1, "Only positive integers are supported for K0"); - if(!adj_lhs || adj_rhs) + if (!adj_lhs || adj_rhs) { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((k0 & (k0 - 1)) && (k0 != 3)) || (k0 > 16), "Only 1,2,3,4,8,16 are supported for K0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((k0 & (k0 - 1)) && (k0 != 3)) || (k0 > 16), + "Only 1,2,3,4,8,16 are supported for K0"); } return Status{}; } -} +} // namespace ClMatMulLowpNativeKernel::ClMatMulLowpNativeKernel() { _type = CLKernelType::GEMM; } -Status ClMatMulLowpNativeKernel::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *bias, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info, +Status ClMatMulLowpNativeKernel::validate(const ITensorInfo *lhs, + const ITensorInfo *rhs, + const ITensorInfo *bias, + const ITensorInfo *dst, + const MatMulKernelInfo &matmul_kernel_info, const ActivationLayerInfo &act_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs, dst); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lhs, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, rhs); ARM_COMPUTE_RETURN_ON_ERROR(validate_matmul_kernel_info(matmul_kernel_info)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_matmul_input_shapes(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info)); + ARM_COMPUTE_RETURN_ON_ERROR( + validate_matmul_input_shapes(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info)); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.activation() != ActivationFunction::IDENTITY && act_info.activation() != ActivationFunction::RELU - && act_info.activation() != ActivationFunction::LU_BOUNDED_RELU && act_info.activation() != ActivationFunction::BOUNDED_RELU), + ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.activation() != ActivationFunction::IDENTITY && + act_info.activation() != ActivationFunction::RELU && + act_info.activation() != ActivationFunction::LU_BOUNDED_RELU && + act_info.activation() != ActivationFunction::BOUNDED_RELU), "Activation Function specified is unsupported."); - const TensorShape expected_output_shape = misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info); + const TensorShape expected_output_shape = + misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info); - if(dst->total_size() != 0) + if (dst->total_size() != 0) { const TensorInfo tensor_info_output = dst->clone()->set_tensor_shape(expected_output_shape); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, dst); } - if(bias != nullptr) + if (bias != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); @@ -115,7 +125,12 @@ Status ClMatMulLowpNativeKernel::validate(const ITensorInfo *lhs, const ITensorI return Status{}; } -void ClMatMulLowpNativeKernel::configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *bias, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info, +void ClMatMulLowpNativeKernel::configure(const ClCompileContext &compile_context, + ITensorInfo *lhs, + ITensorInfo *rhs, + ITensorInfo *bias, + ITensorInfo *dst, + const MatMulKernelInfo &matmul_kernel_info, const ActivationLayerInfo &act_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst, &compile_context, &matmul_kernel_info); @@ -123,7 +138,8 @@ void ClMatMulLowpNativeKernel::configure(const ClCompileContext &compile_context ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, bias, dst, matmul_kernel_info)); // dst tensor auto initialization if not yet initialized - auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info))); + auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape( + lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info))); const int m = dst->dimension(1); const int n = dst->dimension(0); @@ -217,10 +233,13 @@ void ClMatMulLowpNativeKernel::run_op(ITensorPack &tensors, const Window &window ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - const ICLTensor *lhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); - const ICLTensor *rhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); - const ICLTensor *bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2)); - ICLTensor *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); + const ICLTensor *lhs = + utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); + const ICLTensor *rhs = + utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); + const ICLTensor *bias = + utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2)); + ICLTensor *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst); ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst); @@ -229,7 +248,7 @@ void ClMatMulLowpNativeKernel::run_op(ITensorPack &tensors, const Window &window add_3d_tensor_nhw_argument(idx, lhs); add_3d_tensor_nhw_argument(idx, rhs); - if(bias != nullptr) + if (bias != nullptr) { add_3d_tensor_nhw_argument(idx, bias); } |