diff options
Diffstat (limited to 'tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h')
| -rw-r--r-- | tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h | 204 |
1 files changed, 131 insertions, 73 deletions
diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h index e30a564930..1f4e223b93 100644 --- a/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h +++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,13 +21,12 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE -#define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE_H #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" - #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" #include "arm_compute/dynamic_fusion/sketch/attributes/Conv2dAttributes.h" @@ -38,9 +37,9 @@ #include "tests/CL/CLAccessor.h" #include "tests/framework/Fixture.h" #include "tests/framework/Macros.h" -#include "tests/validation/Validation.h" #include "tests/validation/reference/ConvolutionLayer.h" #include "tests/validation/reference/Permute.h" +#include "tests/validation/Validation.h" using namespace arm_compute::experimental::dynamic_fusion; @@ -55,11 +54,11 @@ namespace template <typename U> void fill(U &&tensor, int i) { - switch(tensor.data_type()) + switch (tensor.data_type()) { case DataType::F16: { - arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f }; + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{-1.0f, 1.0f}; library->fill(tensor, distribution, i); break; } @@ -84,12 +83,21 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DynamicFusionGpuConv2dValidationGenericFixture : public framework::Fixture { public: - using TBias = typename std::conditional < std::is_same<typename std::decay<T>::type, uint8_t>::value - || std::is_same<typename std::decay<T>::type, int8_t>::value, - int32_t, T >::type; // If T: uint8_t or int8_t then TBias: int32_t, otherwise TBias: T - - void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, const PadStrideInfo &info, const Size2D &dilation, DataType data_type, - DataLayout data_layout, QuantizationInfo quantization_info, QuantizationInfo weight_quantization_info) + using TBias = typename std::conditional<std::is_same<typename std::decay<T>::type, uint8_t>::value || + std::is_same<typename std::decay<T>::type, int8_t>::value, + int32_t, + T>::type; // If T: uint8_t or int8_t then TBias: int32_t, otherwise TBias: T + + void setup(TensorShape input_shape, + TensorShape weights_shape, + TensorShape bias_shape, + TensorShape output_shape, + const PadStrideInfo &info, + const Size2D &dilation, + DataType data_type, + DataLayout data_layout, + QuantizationInfo quantization_info, + QuantizationInfo weight_quantization_info) { ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); // Dynamic fusion conv2d only supports NHWC layout const Conv2dAttributes conv2d_attr = convert_pad_stride_info_to_conv_attr(info, dilation); @@ -100,12 +108,15 @@ public: _weight_quantization_info = weight_quantization_info; _bias_data_type = _is_quantized ? DataType::S32 : data_type; _target = compute_target(input_shape, weights_shape, bias_shape, conv2d_attr); - _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, conv2d_attr); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, conv2d_attr); } protected: // Given input is in nchw format - TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, Conv2dAttributes conv2d_attr) + TensorType compute_target(TensorShape input_shape, + TensorShape weights_shape, + const TensorShape &bias_shape, + Conv2dAttributes conv2d_attr) { ARM_COMPUTE_ERROR_ON(_data_layout != DataLayout::NHWC); permute(input_shape, PermutationVector(2U, 0U, 1U)); @@ -114,23 +125,23 @@ protected: // Create a new workload sketch auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); - auto context = GpuWorkloadContext{ &cl_compile_ctx }; - GpuWorkloadSketch sketch{ &context }; + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; // Create sketch tensors - TensorInfo input_info = context.create_tensor_info(TensorInfo(input_shape, 1, _data_type, _data_layout)); - TensorInfo weight_info = context.create_tensor_info(TensorInfo(weights_shape, 1, _data_type, _data_layout)); - TensorInfo bias_info = context.create_tensor_info(TensorInfo(bias_shape, 1, _data_type, _data_layout)); - TensorInfo dst_info = context.create_tensor_info(); + ITensorInfo *input_info = context.create_tensor_info(TensorInfo(input_shape, 1, _data_type, _data_layout)); + ITensorInfo *weight_info = context.create_tensor_info(TensorInfo(weights_shape, 1, _data_type, _data_layout)); + ITensorInfo *bias_info = context.create_tensor_info(TensorInfo(bias_shape, 1, _data_type, _data_layout)); + ITensorInfo *dst_info = context.create_tensor_info(); - ITensorInfo *ans_info = FunctionType::create_op(sketch, &input_info, &weight_info, &bias_info, conv2d_attr); - GpuOutput::create_op(sketch, ans_info, &dst_info); + ITensorInfo *ans_info = FunctionType::create_op(sketch, input_info, weight_info, bias_info, conv2d_attr); + GpuOutput::create_op(sketch, ans_info, dst_info); // Configure runtime ClWorkloadRuntime runtime; runtime.configure(sketch); // (Important) Allocate auxiliary tensor memory if there are any - for(auto &data : runtime.get_auxiliary_tensors()) + for (auto &data : runtime.get_auxiliary_tensors()) { CLTensor *tensor = std::get<0>(data); TensorInfo info = std::get<1>(data); @@ -145,10 +156,10 @@ protected: TensorType t_dst{}; // Initialize user tensors - t_input.allocator()->init(input_info); - t_weight.allocator()->init(weight_info); - t_bias.allocator()->init(bias_info); - t_dst.allocator()->init(dst_info); + t_input.allocator()->init(*input_info); + t_weight.allocator()->init(*weight_info); + t_bias.allocator()->init(*bias_info); + t_dst.allocator()->init(*dst_info); // Allocate and fill user tensors t_input.allocator()->allocate(); @@ -161,17 +172,20 @@ protected: fill(AccessorType(t_bias), 2); // Run runtime - runtime.run({ &t_input, &t_weight, &t_bias, &t_dst }); + runtime.run({&t_input, &t_weight, &t_bias, &t_dst}); return t_dst; } - SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, - const TensorShape &output_shape, Conv2dAttributes conv2d_attr) + SimpleTensor<T> compute_reference(const TensorShape &input_shape, + const TensorShape &weights_shape, + const TensorShape &bias_shape, + const TensorShape &output_shape, + Conv2dAttributes conv2d_attr) { // Create reference - SimpleTensor<T> src{ input_shape, _data_type, 1, _quantization_info }; - SimpleTensor<T> weight{ weights_shape, _data_type, 1, _weight_quantization_info }; - SimpleTensor<TBias> bias{ bias_shape, _data_type, 1, _quantization_info }; + SimpleTensor<T> src{input_shape, _data_type, 1, _quantization_info}; + SimpleTensor<T> weight{weights_shape, _data_type, 1, _weight_quantization_info}; + SimpleTensor<TBias> bias{bias_shape, _data_type, 1, _quantization_info}; fill(src, 0); fill(weight, 1); @@ -182,9 +196,11 @@ protected: auto bias_nchw = bias; auto output_shape_nchw = output_shape; - PadStrideInfo legacy_pad_stride(conv2d_attr.stride().x(), conv2d_attr.stride().y(), conv2d_attr.pad().left, conv2d_attr.pad().right, conv2d_attr.pad().top, conv2d_attr.pad().bottom, + PadStrideInfo legacy_pad_stride(conv2d_attr.stride().x(), conv2d_attr.stride().y(), conv2d_attr.pad().left, + conv2d_attr.pad().right, conv2d_attr.pad().top, conv2d_attr.pad().bottom, DimensionRoundingType{}); - auto dst_nchw = reference::convolution_layer(src_nchw, weights_nchw, bias_nchw, output_shape_nchw, legacy_pad_stride, conv2d_attr.dilation()); + auto dst_nchw = reference::convolution_layer(src_nchw, weights_nchw, bias_nchw, output_shape_nchw, + legacy_pad_stride, conv2d_attr.dilation()); return dst_nchw; } @@ -199,14 +215,23 @@ protected: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class DynamicFusionGpuConv2dValidationFixture : public DynamicFusionGpuConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +class DynamicFusionGpuConv2dValidationFixture + : public DynamicFusionGpuConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape output_shape, TensorShape bias_shape, - const PadStrideInfo &info, const Size2D &dialation, DataType data_type, DataLayout data_layout, QuantizationInfo quantization_info) + void setup(TensorShape input_shape, + TensorShape weights_shape, + TensorShape output_shape, + TensorShape bias_shape, + const PadStrideInfo &info, + const Size2D &dialation, + DataType data_type, + DataLayout data_layout, + QuantizationInfo quantization_info) { - DynamicFusionGpuConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, output_shape, bias_shape, info, dialation, - data_type, data_layout, quantization_info, quantization_info); + DynamicFusionGpuConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + input_shape, weights_shape, output_shape, bias_shape, info, dialation, data_type, data_layout, + quantization_info, quantization_info); } }; @@ -218,10 +243,19 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DynamicFusionDirectConv2dValidationGenericFixture : public framework::Fixture { public: - using TBias = typename std::conditional < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int32_t, T >::type; - - void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, - DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout) + using TBias = + typename std::conditional<std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int32_t, T>::type; + + void setup(TensorShape input_shape, + int stride_x, + int stride_y, + int pad_x, + int pad_y, + unsigned int kernel_size, + unsigned int num_kernels, + DataType data_type, + QuantizationInfo quantization_info, + DataLayout data_layout) { ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); // Dynamic fusion conv2d only supports NHWC layout @@ -230,20 +264,30 @@ public: const PadStrideInfo info(stride_x, stride_y, pad_x, pad_y, DimensionRoundingType::FLOOR); const DataType bias_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type; - const Conv2dAttributes conv2d_attr = convert_pad_stride_info_to_conv_attr(info, { 1U, 1U } /* dilation */); + const Conv2dAttributes conv2d_attr = convert_pad_stride_info_to_conv_attr(info, {1U, 1U} /* dilation */); TensorInfo input_info = TensorInfo(input_shape, 1, data_type); TensorInfo weights_info = TensorInfo(weights_shape, 1, data_type); - const TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(input_info, weights_info, info); + const TensorShape output_shape = + misc::shape_calculator::compute_deep_convolution_shape(input_info, weights_info, info); - _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, conv2d_attr, data_type, bias_data_type, quantization_info, data_layout); - _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, quantization_info); + _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, conv2d_attr, data_type, + bias_data_type, quantization_info, data_layout); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, + bias_data_type, quantization_info); } protected: - TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, TensorShape output_shape, const Conv2dAttributes &conv2d_attr, - DataType data_type, DataType bias_data_type, QuantizationInfo quantization_info, const DataLayout &data_layout) + TensorType compute_target(TensorShape input_shape, + TensorShape weights_shape, + const TensorShape &bias_shape, + TensorShape output_shape, + const Conv2dAttributes &conv2d_attr, + DataType data_type, + DataType bias_data_type, + QuantizationInfo quantization_info, + const DataLayout &data_layout) { ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); ARM_COMPUTE_UNUSED(quantization_info); @@ -253,8 +297,8 @@ protected: permute(output_shape, PermutationVector(2U, 0U, 1U)); auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); - auto context = GpuWorkloadContext{ &cl_compile_ctx }; - GpuWorkloadSketch sketch{ &context }; + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; // Create sketch tensors auto input_info = context.create_tensor_info(TensorInfo(input_shape, 1, data_type, data_layout)); @@ -262,14 +306,14 @@ protected: auto bias_info = context.create_tensor_info(TensorInfo(bias_shape, 1, bias_data_type, data_layout)); auto dst_info = context.create_tensor_info(); - ITensorInfo *ans_info = FunctionType::create_op(sketch, &input_info, &weight_info, &bias_info, conv2d_attr); - GpuOutput::create_op(sketch, ans_info, &dst_info); + ITensorInfo *ans_info = FunctionType::create_op(sketch, input_info, weight_info, bias_info, conv2d_attr); + GpuOutput::create_op(sketch, ans_info, dst_info); // Configure runtime ClWorkloadRuntime runtime; runtime.configure(sketch); - for(auto &data : runtime.get_auxiliary_tensors()) + for (auto &data : runtime.get_auxiliary_tensors()) { CLTensor *tensor = std::get<0>(data); TensorInfo info = std::get<1>(data); @@ -284,10 +328,10 @@ protected: TensorType t_dst{}; // Initialize user tensors - t_input.allocator()->init(input_info); - t_weight.allocator()->init(weight_info); - t_bias.allocator()->init(bias_info); - t_dst.allocator()->init(dst_info); + t_input.allocator()->init(*input_info); + t_weight.allocator()->init(*weight_info); + t_bias.allocator()->init(*bias_info); + t_dst.allocator()->init(*dst_info); ARM_COMPUTE_ASSERT(t_input.info()->is_resizable()); ARM_COMPUTE_ASSERT(t_weight.info()->is_resizable()); @@ -310,17 +354,23 @@ protected: fill(AccessorType(t_bias), 2); // Run runtime - runtime.run({ &t_input, &t_weight, &t_bias, &t_dst }); + runtime.run({&t_input, &t_weight, &t_bias, &t_dst}); return t_dst; } - SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info, - DataType data_type, DataType bias_data_type, QuantizationInfo quantization_info) + SimpleTensor<T> compute_reference(const TensorShape &input_shape, + const TensorShape &weights_shape, + const TensorShape &bias_shape, + const TensorShape &output_shape, + const PadStrideInfo &info, + DataType data_type, + DataType bias_data_type, + QuantizationInfo quantization_info) { // Create reference - SimpleTensor<T> src{ input_shape, data_type, 1, quantization_info }; - SimpleTensor<T> weights{ weights_shape, data_type, 1, quantization_info }; - SimpleTensor<TBias> bias{ bias_shape, bias_data_type, 1, quantization_info }; + SimpleTensor<T> src{input_shape, data_type, 1, quantization_info}; + SimpleTensor<T> weights{weights_shape, data_type, 1, quantization_info}; + SimpleTensor<TBias> bias{bias_shape, bias_data_type, 1, quantization_info}; // Fill reference fill(src, 0); @@ -335,19 +385,27 @@ protected: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class DynamicFusionDirectConv2dValidationFixture : public DynamicFusionDirectConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +class DynamicFusionDirectConv2dValidationFixture + : public DynamicFusionDirectConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, DataType data_type, - DataLayout data_layout) + void setup(TensorShape input_shape, + int stride_x, + int stride_y, + int pad_x, + int pad_y, + unsigned int kernel_size, + unsigned int num_kernels, + DataType data_type, + DataLayout data_layout) { - DynamicFusionDirectConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type, - QuantizationInfo(), - data_layout); + DynamicFusionDirectConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type, QuantizationInfo(), + data_layout); } }; } // namespace validation } // namespace test } // namespace arm_compute -#endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE_H |