From 19884630c37ae9de2f65a88ea2cda5630a551bad Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 16 Aug 2021 12:38:54 +0100 Subject: Rename [Cl|Cpu]GemmConvolution to [Cl|Gpu]GemmConv2d Renaming the gemm-based convolution operators to accomodate for new operators with higher convolution dimensonality Signed-off-by: Georgios Pinitas Change-Id: Id2f2cf11404221f0e87baa0e5d08ad5d63eaf78e Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6113 Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins --- Android.bp | 4 +- .../runtime/CL/functions/CLConvolutionLayer.h | 4 +- .../runtime/CL/functions/CLGEMMConvolutionLayer.h | 2 +- .../NEON/functions/NEGEMMConvolutionLayer.h | 2 +- filelist.json | 4 +- .../CL/functions/CLGEMMConvolutionLayer.cpp | 24 +- src/runtime/NEON/functions/NEConvolutionLayer.cpp | 2 +- .../NEON/functions/NEGEMMConvolutionLayer.cpp | 22 +- src/runtime/cpu/operators/CpuConv2d.cpp | 6 +- src/runtime/cpu/operators/CpuGemmConv2d.cpp | 612 ++++++++++++++++++++ src/runtime/cpu/operators/CpuGemmConv2d.h | 203 +++++++ src/runtime/cpu/operators/CpuGemmConvolution.cpp | 612 -------------------- src/runtime/cpu/operators/CpuGemmConvolution.h | 203 ------- src/runtime/gpu/cl/operators/ClConv2d.cpp | 6 +- src/runtime/gpu/cl/operators/ClConv2d.h | 2 +- src/runtime/gpu/cl/operators/ClGemmConv2d.cpp | 628 +++++++++++++++++++++ src/runtime/gpu/cl/operators/ClGemmConv2d.h | 185 ++++++ src/runtime/gpu/cl/operators/ClGemmConvolution.cpp | 628 --------------------- src/runtime/gpu/cl/operators/ClGemmConvolution.h | 185 ------ tests/validation/NEON/ConvolutionLayer.cpp | 6 +- 20 files changed, 1670 insertions(+), 1670 deletions(-) create mode 100644 src/runtime/cpu/operators/CpuGemmConv2d.cpp create mode 100644 src/runtime/cpu/operators/CpuGemmConv2d.h delete mode 100644 src/runtime/cpu/operators/CpuGemmConvolution.cpp delete mode 100644 src/runtime/cpu/operators/CpuGemmConvolution.h create mode 100644 src/runtime/gpu/cl/operators/ClGemmConv2d.cpp create mode 100644 src/runtime/gpu/cl/operators/ClGemmConv2d.h delete mode 100644 src/runtime/gpu/cl/operators/ClGemmConvolution.cpp delete mode 100644 src/runtime/gpu/cl/operators/ClGemmConvolution.h diff --git a/Android.bp b/Android.bp index 3a3f1db334..c1adf32893 100644 --- a/Android.bp +++ b/Android.bp @@ -647,7 +647,7 @@ cc_library_static { "src/runtime/cpu/operators/CpuFloor.cpp", "src/runtime/cpu/operators/CpuFullyConnected.cpp", "src/runtime/cpu/operators/CpuGemm.cpp", - "src/runtime/cpu/operators/CpuGemmConvolution.cpp", + "src/runtime/cpu/operators/CpuGemmConv2d.cpp", "src/runtime/cpu/operators/CpuGemmDirectConv2d.cpp", "src/runtime/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp", "src/runtime/cpu/operators/CpuGemmLowpOutputStage.cpp", @@ -679,7 +679,7 @@ cc_library_static { "src/runtime/gpu/cl/operators/ClFloor.cpp", "src/runtime/gpu/cl/operators/ClFullyConnected.cpp", "src/runtime/gpu/cl/operators/ClGemm.cpp", - "src/runtime/gpu/cl/operators/ClGemmConvolution.cpp", + "src/runtime/gpu/cl/operators/ClGemmConv2d.cpp", "src/runtime/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.cpp", "src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.cpp", "src/runtime/gpu/cl/operators/ClLogicalNot.cpp", diff --git a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h index 12b3ca1fd2..0f092bdbc2 100644 --- a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h +++ b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h @@ -24,9 +24,9 @@ #ifndef ARM_COMPUTE_CLCONVOLUTIONLAYER_H #define ARM_COMPUTE_CLCONVOLUTIONLAYER_H -#include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/core/CL/CLCompileContext.h" #include "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/IFunction.h" #include "arm_compute/runtime/IMemoryManager.h" @@ -40,7 +40,7 @@ class ITensorInfo; /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions: * - * -# @ref opencl::ClGemmConvolution + * -# @ref opencl::ClGemmConv2d * -# @ref opencl::ClWinogradConv2d * -# @ref opencl::ClDirectConv2d * -# @ref CLFFTConvolutionLayer diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h index 3075465ef7..d7a4e7f944 100644 --- a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h +++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h @@ -41,7 +41,7 @@ class ITensorInfo; /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions: * - * -# @ref opencl::ClGemmConvolution + * -# @ref opencl::ClGemmConv2d */ class CLGEMMConvolutionLayer : public IFunction { diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h index fe866dd941..cf5fb82398 100644 --- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h +++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h @@ -41,7 +41,7 @@ class ITensorInfo; /** Basic function to compute the convolution layer. This function calls the following kernels/functions: * - * -# @ref cpu::CpuGemmConvolution + * -# @ref cpu::CpuGemmConv2d * */ class NEGEMMConvolutionLayer : public IFunction diff --git a/filelist.json b/filelist.json index c5abc620cd..d9c67126c6 100644 --- a/filelist.json +++ b/filelist.json @@ -288,7 +288,7 @@ "files": { "operator": [ "src/runtime/gpu/cl/operators/ClGemm.cpp", - "src/runtime/gpu/cl/operators/ClGemmConvolution.cpp" + "src/runtime/gpu/cl/operators/ClGemmConv2d.cpp" ], "kernel": [ "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp", @@ -1240,7 +1240,7 @@ ], "files": { "operator": [ - "src/runtime/cpu/operators/CpuGemmConvolution.cpp" + "src/runtime/cpu/operators/CpuGemmConv2d.cpp" ], "kernel": [ "src/core/cpu/kernels/CpuWeightsReshapeKernel.cpp" diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp index 75ca77dbe2..563dbd414f 100644 --- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp @@ -32,7 +32,7 @@ #include "arm_compute/core/utils/quantization/AsymmHelpers.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "src/core/helpers/MemoryHelpers.h" -#include "src/runtime/gpu/cl/operators/ClGemmConvolution.h" +#include "src/runtime/gpu/cl/operators/ClGemmConv2d.h" #include "support/Cast.h" #include @@ -47,15 +47,15 @@ using namespace arm_compute::experimental; struct CLGEMMConvolutionLayer::Impl { - const ITensor *weights{ nullptr }; - std::unique_ptr op{ nullptr }; - ITensorPack run_pack{}; - ITensorPack prep_pack{}; - MemoryGroup memory_group{}; - IWeightsManager *weights_manager{ nullptr }; - MemoryRequirements aux_mem_req{}; - WorkspaceData workspace_tensors{}; - bool is_prepared{ false }; + const ITensor *weights{ nullptr }; + std::unique_ptr op{ nullptr }; + ITensorPack run_pack{}; + ITensorPack prep_pack{}; + MemoryGroup memory_group{}; + IWeightsManager *weights_manager{ nullptr }; + MemoryRequirements aux_mem_req{}; + WorkspaceData workspace_tensors{}; + bool is_prepared{ false }; }; CLGEMMConvolutionLayer::CLGEMMConvolutionLayer(std::shared_ptr memory_manager, IWeightsManager *weights_manager) @@ -79,7 +79,7 @@ void CLGEMMConvolutionLayer::configure(const CLCompileContext &compile_context, { ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); _impl->weights = weights; - _impl->op = std::make_unique(); + _impl->op = std::make_unique(); const Conv2dInfo conv2d_info = Conv2dInfo(conv_info, dilation, act_info, false, num_groups); _impl->op->configure(compile_context, input->info(), weights->info(), (biases != nullptr ? biases->info() : nullptr), output->info(), conv2d_info, weights_info); @@ -103,7 +103,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, unsigned int num_groups) { const Conv2dInfo conv2d_info = Conv2dInfo(conv_info, dilation, act_info, false, num_groups); - return opencl::ClGemmConvolution::validate(input, weights, biases, output, conv2d_info, weights_info); + return opencl::ClGemmConv2d::validate(input, weights, biases, output, conv2d_info, weights_info); } void CLGEMMConvolutionLayer::run() diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp index 6e25b429d4..8bd1119a69 100644 --- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp @@ -30,7 +30,7 @@ #include "src/core/helpers/MemoryHelpers.h" #include "src/runtime/cpu/operators/CpuConv2d.h" #include "src/runtime/cpu/operators/CpuDirectConv2d.h" -#include "src/runtime/cpu/operators/CpuGemmConvolution.h" +#include "src/runtime/cpu/operators/CpuGemmConv2d.h" #include "src/runtime/cpu/operators/CpuGemmDirectConv2d.h" #include "src/runtime/cpu/operators/CpuWinogradConv2d.h" diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp index c32584ec0d..47ab16816a 100644 --- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp @@ -28,7 +28,7 @@ #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/Tensor.h" #include "src/core/helpers/MemoryHelpers.h" -#include "src/runtime/cpu/operators/CpuGemmConvolution.h" +#include "src/runtime/cpu/operators/CpuGemmConv2d.h" using namespace arm_compute::experimental; @@ -36,14 +36,14 @@ namespace arm_compute { struct NEGEMMConvolutionLayer::Impl { - const ITensor *weights{ nullptr }; - std::unique_ptr op{ nullptr }; - ITensorPack run_pack{}; - MemoryGroup memory_group{}; - IWeightsManager *weights_manager{ nullptr }; - MemoryRequirements aux_mem_req{}; - WorkspaceData workspace_tensors{}; - bool is_prepared{ false }; + const ITensor *weights{ nullptr }; + std::unique_ptr op{ nullptr }; + ITensorPack run_pack{}; + MemoryGroup memory_group{}; + IWeightsManager *weights_manager{ nullptr }; + MemoryRequirements aux_mem_req{}; + WorkspaceData workspace_tensors{}; + bool is_prepared{ false }; }; NEGEMMConvolutionLayer::NEGEMMConvolutionLayer(const std::shared_ptr &memory_manager, IWeightsManager *weights_manager) @@ -59,7 +59,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig { ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); _impl->weights = weights; - _impl->op = std::make_unique(); + _impl->op = std::make_unique(); _impl->op->configure(input->info(), weights->info(), (biases != nullptr ? biases->info() : nullptr), output->info(), conv_info, weights_info, dilation, act_info, enable_fast_math, num_groups); _impl->run_pack = @@ -76,7 +76,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups) { - return cpu::CpuGemmConvolution::validate(input, weights, biases, output, conv_info, weights_info, dilation, act_info, enable_fast_math, num_groups); + return cpu::CpuGemmConv2d::validate(input, weights, biases, output, conv_info, weights_info, dilation, act_info, enable_fast_math, num_groups); } void NEGEMMConvolutionLayer::run() diff --git a/src/runtime/cpu/operators/CpuConv2d.cpp b/src/runtime/cpu/operators/CpuConv2d.cpp index 809663a918..cff9238308 100644 --- a/src/runtime/cpu/operators/CpuConv2d.cpp +++ b/src/runtime/cpu/operators/CpuConv2d.cpp @@ -26,7 +26,7 @@ #include "arm_compute/runtime/NEON/functions/NEFFTConvolutionLayer.h" #include "src/runtime/cpu/operators/CpuDirectConv2d.h" #include "src/runtime/cpu/operators/CpuGemm.h" -#include "src/runtime/cpu/operators/CpuGemmConvolution.h" +#include "src/runtime/cpu/operators/CpuGemmConv2d.h" #include "src/runtime/cpu/operators/CpuGemmDirectConv2d.h" #include "src/runtime/cpu/operators/CpuWinogradConv2d.h" @@ -62,7 +62,7 @@ void CpuConv2d::configure(ITensorInfo *input, ITensorInfo *weights, const ITenso } case ConvolutionMethod::GEMM: { - auto f = std::make_unique(); + auto f = std::make_unique(); f->configure(input, weights, biases, output, conv_info, weights_info, dilation, act_info, enable_fast_math); _function = std::move(f); break; @@ -101,7 +101,7 @@ Status CpuConv2d::validate(const ITensorInfo *input, const ITensorInfo *weights, ARM_COMPUTE_RETURN_ON_ERROR(CpuWinogradConv2d::validate(input, weights, biases, output, conv_info, act_info, enable_fast_math)); break; case ConvolutionMethod::GEMM: - ARM_COMPUTE_RETURN_ON_ERROR(CpuGemmConvolution::validate(input, weights, biases, output, conv_info, weights_info, dilation, act_info, enable_fast_math)); + ARM_COMPUTE_RETURN_ON_ERROR(CpuGemmConv2d::validate(input, weights, biases, output, conv_info, weights_info, dilation, act_info, enable_fast_math)); break; case ConvolutionMethod::GEMM_CONV2D: ARM_COMPUTE_RETURN_ON_ERROR(CpuGemmDirectConv2d::validate(input, weights, biases, output, info)); diff --git a/src/runtime/cpu/operators/CpuGemmConv2d.cpp b/src/runtime/cpu/operators/CpuGemmConv2d.cpp new file mode 100644 index 0000000000..a81dd8a661 --- /dev/null +++ b/src/runtime/cpu/operators/CpuGemmConv2d.cpp @@ -0,0 +1,612 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "src/runtime/cpu/operators/CpuGemmConv2d.h" + +#include "arm_compute/core/Size2D.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "arm_compute/runtime/NEON/NEScheduler.h" + +#include "src/core/cpu/kernels/CpuCol2ImKernel.h" +#include "src/core/cpu/kernels/CpuIm2ColKernel.h" +#include "src/core/cpu/kernels/CpuReshapeKernel.h" +#include "src/core/cpu/kernels/CpuWeightsReshapeKernel.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/runtime/cpu/operators/CpuGemm.h" +#include "src/runtime/cpu/operators/CpuGemmLowpMatrixMultiplyCore.h" +#include "src/runtime/cpu/operators/CpuGemmLowpOutputStage.h" +#include "src/runtime/cpu/utils/CpuAuxTensorHandler.h" + +#include +#include + +using namespace arm_compute::misc::shape_calculator; +using namespace arm_compute::experimental; + +namespace arm_compute +{ +namespace cpu +{ +CpuGemmConv2d::CpuGemmConv2d() + : _weights_reshape_kernel(nullptr), _im2col_kernel(), _mm_gemm(), _mm_gemmlowp(), _col2im_kernel(), _reshape_kernel(), _im2col_output(), _weights_reshaped(), _gemm_output(), _gemm_output_3d(), + _data_layout(DataLayout::NCHW), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _is_prepared(false), _aux_mem(AuxTensorIdx::Count) +{ +} +CpuGemmConv2d::~CpuGemmConv2d() = default; + +void CpuGemmConv2d::configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act_info, + bool enable_fast_math, int gemm_3d_depth) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights); + ARM_COMPUTE_ERROR_THROW_ON(validate_mm(src, weights, biases, dst, act_info, enable_fast_math, gemm_3d_depth, _skip_im2col)); + + // Create GEMMInfo structure + const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, + gemm_3d_depth, _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */, + false, GEMMLowpOutputStageInfo(), false, enable_fast_math, false, act_info); + + // Supported activations in GEMM + const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, + ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, + ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU + }; + + if(_is_quantized) + { + TensorInfo tmp_src{ *src }; + TensorInfo tmp_weights{ *weights }; + // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() + // Extract and negate input and weights offset + const QuantizationInfo iqinfo = src->quantization_info(); + const QuantizationInfo wqinfo = weights->quantization_info(); + const QuantizationInfo oqinfo = (dst->total_size() == 0) ? iqinfo : dst->quantization_info(); + const UniformQuantizationInfo uiqinfo = iqinfo.uniform(); + const UniformQuantizationInfo uoqinfo = oqinfo.uniform(); + const DataType data_type = src->data_type(); + + tmp_src.set_quantization_info(QuantizationInfo(uiqinfo.scale, -uiqinfo.offset)); + if(!is_data_type_quantized_per_channel(tmp_weights.data_type())) + { + const UniformQuantizationInfo uwqinfo = wqinfo.uniform(); + tmp_weights.set_quantization_info(QuantizationInfo(uwqinfo.scale, -uwqinfo.offset)); + } + + // Merge activation with output stage + PixelValue type_min{}; + PixelValue type_max{}; + std::tie(type_min, type_max) = get_min_max(data_type); + int32_t min_activation = type_min.get(); + int32_t max_activation = type_max.get(); + + if(supported_acts.count(act_info.activation()) != 0) + { + std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act_info, data_type, uoqinfo); + } + + GEMMLowpOutputStageInfo output_info; + output_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; + output_info.gemmlowp_offset = uoqinfo.offset; + output_info.gemmlowp_min_bound = min_activation; + output_info.gemmlowp_max_bound = max_activation; + output_info.is_quantized_per_channel = (tmp_weights.data_type() == DataType::QSYMM8_PER_CHANNEL); + quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info); + + _mm_gemmlowp = std::make_unique(); + _mm_gemmlowp->configure(&tmp_src, &tmp_weights, biases, dst, GEMMInfo(false, false, true, gemm_3d_depth, _skip_im2col, false, output_info, false, enable_fast_math, false, act_info)); + + auto mm_mem_req = _mm_gemmlowp->workspace(); + for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) + { + _aux_mem[cont] = mm_mem_req[cont]; + } + } + else + { + // Configure matrix multiply function + _mm_gemm = std::make_unique(); + _mm_gemm->configure(src, weights, biases, dst, 1.0f, 0.0f, gemm_info); + auto mm_mem_req = _mm_gemm->workspace(); + for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) + { + _aux_mem[cont] = mm_mem_req[cont]; + } + } +} + +Status CpuGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, + const ActivationLayerInfo &act_info, bool enable_fast_math, int gemm_3d_depth, bool skip_im2col) +{ + const DataType data_type = src->data_type(); + const bool is_quantized = is_data_type_quantized_asymmetric(data_type); + const bool is_activation_enabled = act_info.enabled(); + + // Create GEMMInfo structure + const GEMMInfo gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, + gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */, + false, GEMMLowpOutputStageInfo(), false, enable_fast_math, false, act_info); + + if(is_quantized) + { + // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() + // Extract and negate input and weights offset + const QuantizationInfo &iqinfo = src->quantization_info(); + const QuantizationInfo &wqinfo = weights->quantization_info(); + const QuantizationInfo &oqinfo = (dst->total_size() == 0) ? iqinfo : dst->quantization_info(); + const UniformQuantizationInfo uoqinfo = oqinfo.uniform(); + + // Merge activation with output stage + PixelValue type_min{}; + PixelValue type_max{}; + std::tie(type_min, type_max) = get_min_max(data_type); + int32_t min_activation = type_min.get(); + int32_t max_activation = type_max.get(); + + const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, + ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, + ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU + }; + if(is_activation_enabled && supported_acts.count(act_info.activation()) != 0) + { + std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act_info, data_type, uoqinfo); + } + + GEMMLowpOutputStageInfo output_info; + output_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; + output_info.gemmlowp_offset = uoqinfo.offset; + output_info.gemmlowp_min_bound = min_activation; + output_info.gemmlowp_max_bound = max_activation; + output_info.is_quantized_per_channel = (weights->data_type() == DataType::QSYMM8_PER_CHANNEL); + ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info)); + + // Perform validation step on GEMMLowp + std::unique_ptr input_qa = src->clone(); + std::unique_ptr weights_qa = weights->clone(); + input_qa->set_quantization_info(QuantizationInfo(iqinfo.uniform().scale, -iqinfo.uniform().offset)); + weights_qa->set_quantization_info(QuantizationInfo(wqinfo.uniform().scale, -wqinfo.uniform().offset)); + return CpuGemmLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, dst, GEMMInfo(false, false, true, gemm_3d_depth, skip_im2col, false, output_info, + false, enable_fast_math, false, act_info)); + } + else + { + // Perform validation step on Matrix multiply function + return CpuGemm::validate(src, weights, nullptr, dst, 1.0f, 0.0f, gemm_info); + } +} + +Status CpuGemmConv2d::validate_gemm3d(const ITensorInfo *input_info, const ITensorInfo *weights_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col) +{ + const DataType data_type = input_info->data_type(); + const unsigned int mult_y = skip_im2col ? 1U : gemm_3d_depth; + const unsigned int mult_z = skip_im2col ? gemm_3d_depth : 1U; + + // Set dummy tensor shapes for the validation + const TensorInfo dummy_input_info(TensorShape(4U, 4U * mult_y, 1U * mult_z), 1, data_type, input_info->quantization_info()); + const TensorInfo dummy_weights_info(TensorShape(4U, 4U), 1, data_type, weights_info->quantization_info()); + const TensorInfo dummy_output_info(TensorShape(4U, 4U, gemm_3d_depth), 1, data_type, input_info->quantization_info()); + + return validate_mm(&dummy_input_info, &dummy_weights_info, nullptr, &dummy_output_info, act_info, false, gemm_3d_depth, skip_im2col); +} + +void CpuGemmConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, + const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); + ARM_COMPUTE_UNUSED(num_groups, weights_info); + ARM_COMPUTE_ERROR_THROW_ON(CpuGemmConv2d::validate(src, + weights, + biases, + dst, + conv_info, + weights_info, + dilation, + act_info, + enable_fast_math, + num_groups)); + + const DataType data_type = src->data_type(); + const DataLayout data_layout = src->data_layout(); + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); + + const unsigned int kernel_width = weights->dimension(idx_width); + const unsigned int kernel_height = weights->dimension(idx_height); + + _is_prepared = weights_info.retain_internal_weights(); + _is_quantized = is_data_type_quantized_asymmetric(src->data_type()); + _data_layout = data_layout; + _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1); + + const ITensorInfo *gemm_input_to_use = src; + ITensorInfo *gemm_output_to_use = dst; + + // Get convolved dimensions + unsigned int conv_w = 0; + unsigned int conv_h = 0; + std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), + src->dimension(idx_height), + kernel_width, + kernel_height, + conv_info, + dilation); + ARM_COMPUTE_ERROR_ON_MSG((dst->dimension(idx_width) != conv_w) || (dst->dimension(idx_height) != conv_h), + "Output shape does not match the expected one"); + + // Check if GEMM3D is supported + if(data_layout == DataLayout::NHWC) + { + _skip_col2im = bool(validate_gemm3d(src, weights, act_info, conv_h, true)); + // If not supported, we need to perform im2col and col2im (or reshape layer) + if(!_skip_col2im) + { + _skip_im2col = false; + } + } + else + { + _skip_col2im = false; + } + + // Get parameters from conv_info + unsigned int stride_x = 0; + unsigned int stride_y = 0; + std::tie(stride_x, stride_y) = conv_info.stride(); + + unsigned int mat_weights_cols = weights->dimension(idx_kernels); + + // _weights_reshaped will be auto configured in the kernel. + // Just append biases and do not transpose 1xW as it will be reshaped in CpuGemm + _weights_reshape_kernel = std::make_unique(); + _weights_reshape_kernel->configure(weights, nullptr, &_weights_reshaped); + _weights_reshaped.set_quantization_info(weights->quantization_info()); + + // Create tensor to store im2col reshaped inputs + if(!_skip_im2col) + { + // Configure + _im2col_kernel = std::make_unique(); + _im2col_kernel->configure(src, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, false, dilation); + + // Update GEMM input + gemm_input_to_use = &_im2col_output; + } + + // Create temporary GEMM output tensor in case we cannot skip col2im + const DataType output_data_type = data_type == DataType::BFLOAT16 ? DataType::F32 : data_type; + if(!_skip_col2im) + { + TensorShape shape_gemm; + + // Calculate GEMM output shape + shape_gemm = _im2col_output.tensor_shape(); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, conv_w * conv_h); + + _gemm_output = TensorInfo(shape_gemm, 1, output_data_type); + _gemm_output.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); + _gemm_output_3d = TensorInfo(_gemm_output); + + // Update GEMM output + gemm_output_to_use = &_gemm_output; + } + else + { + _gemm_output_3d = TensorInfo(*dst); + _gemm_output_3d.set_data_type(output_data_type).set_data_layout(src->data_layout()).set_is_resizable(true); + _gemm_output = TensorInfo(_gemm_output_3d); + + // Update GEMM output + gemm_output_to_use = &_gemm_output_3d; + } + + // Configure GEMM + // In case we need to skip col2im, GEMM3D (gemm_3d_depth != 0) must be called in order to avoid reshaping the output matrix + const unsigned int gemm_3d_depth = _skip_col2im ? conv_h : 0; + configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, act_info, enable_fast_math, gemm_3d_depth); + + if(!_skip_col2im && _data_layout == DataLayout::NCHW) + { + // Configure col2im + _col2im_kernel = std::make_unique(); + _col2im_kernel->configure(gemm_output_to_use, dst, Size2D(conv_w, conv_h)); + } + else + { + // Configure reshape layer + _reshape_kernel = std::make_unique(); + _reshape_kernel->configure(gemm_output_to_use, dst); + } + + // Check if GEMM transforms weights + // Modernise through COMPMID-4535 + bool gemm_trans_wei = _aux_mem[1].size > 0; // Asm Pretranspose + gemm_trans_wei = _mm_gemm != nullptr ? _aux_mem[3].size > 0 : gemm_trans_wei; // Tranpose RHS + gemm_trans_wei = _mm_gemmlowp != nullptr ? _aux_mem[5].size > 0 : gemm_trans_wei; // Transpose RHS + + // Check lifetime + _aux_mem[Im2ColOutput] = MemoryInfo(offset_int_vec(Im2ColOutput), MemoryLifetime::Temporary, _im2col_output.total_size()); + _aux_mem[WeightsReshaped] = MemoryInfo(offset_int_vec(WeightsReshaped), gemm_trans_wei ? MemoryLifetime::Prepare : MemoryLifetime::Persistent, _weights_reshaped.total_size()); + _aux_mem[GemmOutput] = MemoryInfo(offset_int_vec(GemmOutput), MemoryLifetime::Temporary, _gemm_output.total_size()); +} + +Status CpuGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!"); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::BFLOAT16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::BFLOAT16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, weights); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups > 1, "Grouping (num_groups != 1) is not supported"); + + const DataLayout data_layout = src->data_layout(); + const DataType data_type = src->data_type(); + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); + const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); + + const unsigned int kernel_width = weights->dimension(idx_width); + const unsigned int kernel_height = weights->dimension(idx_height); + + TensorInfo im2col_reshaped_info{}; + TensorInfo info_gemm{}; + TensorInfo tmp_info{}; + TensorInfo weights_reshaped_info{}; + const ITensorInfo *gemm_input_to_use = src; + const ITensorInfo *gemm_output_to_use = dst; + const ITensorInfo *weights_to_use = weights; + + const bool append_bias = false; + const bool is_quantized = is_data_type_quantized_asymmetric(data_type); + const bool is_bf16 = data_type == DataType::BFLOAT16; + bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1); + + // Get convolved dimensions + unsigned int conv_w = 0; + unsigned int conv_h = 0; + + std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), + src->dimension(idx_height), + kernel_width, + kernel_height, + conv_info, + dilation); + + // Check if GEMM3D is supported + bool skip_col2im = false; + if(data_layout == DataLayout::NHWC) + { + skip_col2im = bool(validate_gemm3d(src, weights, act_info, conv_h, true)); + // If not supported, we need to perform im2col and col2im (or reshape layer) + if(!skip_col2im) + { + skip_im2col = false; + } + } + + if(skip_col2im) + { + // If not supported, we need to perform im2col and col2im (or reshape layer) + if(!bool(validate_gemm3d(src, weights, act_info, conv_h, skip_im2col))) + { + skip_im2col = false; + skip_col2im = false; + } + } + + ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_channel) != src->dimension(idx_channel)); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); + + // Validate biases + if(biases != nullptr) + { + if(is_quantized) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); + } + else if(is_bf16) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases); + } + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(idx_kernels)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + unsigned int mat_weights_cols = weights->dimension(idx_kernels); + unsigned int mat_weights_rows = weights->dimension(idx_width) * weights->dimension(idx_height) * weights->dimension(idx_channel); + + weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, append_bias), 1, data_type); + weights_reshaped_info.set_quantization_info(weights->quantization_info()); + weights_to_use = &weights_reshaped_info; + + if(!skip_im2col) + { + // Create tensor info for im2col reshaped inputs + // For CPU, the batch size is on the fourth dimension + TensorShape shape_im2col = src->tensor_shape(); + shape_im2col.set(0, mat_weights_rows); + shape_im2col.set(1, conv_w * conv_h); + shape_im2col.set(2, 1); + + im2col_reshaped_info = TensorInfo(shape_im2col, 1, data_type); + im2col_reshaped_info.set_quantization_info(src->quantization_info()); + ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuIm2ColKernel::validate(src, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation)); + gemm_input_to_use = &im2col_reshaped_info; + } + + // Create temporary GEMM output tensor in case we cannot skip col2im + const DataType output_data_type = data_type == DataType::BFLOAT16 ? DataType::F32 : data_type; + if(!skip_col2im) + { + TensorShape shape_gemm = gemm_input_to_use->tensor_shape(); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, conv_w * conv_h); + info_gemm = TensorInfo(shape_gemm, 1, output_data_type); + } + else + { + info_gemm = TensorInfo(dst->tensor_shape(), 1, output_data_type); + } + info_gemm.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); + gemm_output_to_use = &info_gemm; + ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases, gemm_output_to_use, act_info, enable_fast_math, skip_col2im ? conv_h : 0, skip_im2col)); + + // Validate Col2Im/ReshapeLayer + if(!skip_col2im && (data_layout == DataLayout::NCHW)) + { + ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuCol2ImKernel::validate(gemm_output_to_use, dst, Size2D(conv_w, conv_h))); + } + + return Status{}; +} + +void CpuGemmConv2d::run(ITensorPack &tensors) +{ + prepare(tensors); + + auto src = tensors.get_const_tensor(ACL_SRC_0); + auto dst = tensors.get_tensor(ACL_DST); + auto gemm_input_to_use = src; + + CpuAuxTensorHandler im2col_output(offset_int_vec(Im2ColOutput), _im2col_output, tensors, false); + CpuAuxTensorHandler gemm_output(offset_int_vec(GemmOutput), _gemm_output, tensors, false); + CpuAuxTensorHandler reshaped_wei(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors, false); + + bool out_has_padding = _skip_col2im && (dst->info()->padding().bottom != 0 || dst->info()->padding().top != 0); + if(!_skip_im2col) + { + // Run input reshaping + unsigned int y_dim = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); + ITensorPack pack = + { + { TensorType::ACL_SRC, src }, + { TensorType::ACL_DST, im2col_output.get() } + }; + NEScheduler::get().schedule_op(_im2col_kernel.get(), y_dim, _im2col_kernel->window(), pack); + gemm_input_to_use = im2col_output.get(); + } + + // Handle the case where output has top/bottom padding + const ITensor *out_to_use = out_has_padding ? gemm_output.get() : dst; + Tensor gemm3d; + _gemm_output_3d.extend_padding(out_to_use->info()->padding()); + gemm3d.allocator()->soft_init(_gemm_output_3d); + gemm3d.allocator()->import_memory(out_to_use->buffer()); + auto gemm_output_to_use = gemm_output.get(); + + if(_skip_im2col) + { + gemm_output_to_use = &gemm3d; + } + if(_skip_col2im && !out_has_padding) + { + gemm_output_to_use = dst; + } + + // Runs CpuGemm or CpuGemmLowpMatrixMultiplyCore functions + ITensorPack pack_mm = tensors; + pack_mm.add_const_tensor(TensorType::ACL_SRC_0, gemm_input_to_use); + pack_mm.add_const_tensor(TensorType::ACL_SRC_1, reshaped_wei.get()); + pack_mm.add_tensor(TensorType::ACL_DST, gemm_output_to_use); + if(_is_quantized) + { + // Run gemmlowp + _mm_gemmlowp->run(pack_mm); + } + else + { + // Run gemm + _mm_gemm->run(pack_mm); + } + + // Reshape output matrix + if(!_skip_col2im) + { + if(_data_layout == DataLayout::NCHW) + { + ITensorPack pack = + { + { TensorType::ACL_SRC, gemm_output.get() }, + { TensorType::ACL_DST, dst } + }; + NEScheduler::get().schedule_op(_col2im_kernel.get(), Window::DimY, _col2im_kernel->window(), pack); + } + else + { + ITensorPack pack = + { + { TensorType::ACL_SRC, gemm_output_to_use }, + { TensorType::ACL_DST, dst } + }; + NEScheduler::get().schedule_op(_reshape_kernel.get(), Window::DimY, _reshape_kernel->window(), pack); + } + } + else if(out_has_padding) + { + ITensorPack pack = + { + { TensorType::ACL_SRC, gemm_output_to_use }, + { TensorType::ACL_DST, dst } + }; + NEScheduler::get().schedule_op(_reshape_kernel.get(), Window::DimY, _reshape_kernel->window(), pack); + } +} + +void CpuGemmConv2d::prepare(ITensorPack &tensors) +{ + if(!_is_prepared) + { + // Run weights reshaping and mark original weights tensor as unused + CpuAuxTensorHandler weights_reshaped(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors); + auto weights = tensors.get_const_tensor(TensorType::ACL_SRC_1); + ITensorPack pack = + { + { TensorType::ACL_SRC, weights }, + { TensorType::ACL_DST, weights_reshaped.get() } + }; + NEScheduler::get().schedule_op(_weights_reshape_kernel.get(), 3, _weights_reshape_kernel->window(), pack); + weights->mark_as_unused(); + + // Prepare GEMM + ITensorPack gemm_pack = tensors; + gemm_pack.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get()); + _is_quantized ? _mm_gemmlowp->prepare(gemm_pack) : _mm_gemm->prepare(gemm_pack); + + _is_prepared = true; + } +} +experimental::MemoryRequirements CpuGemmConv2d::workspace() const +{ + return _aux_mem; +} +} // namespace cpu +} // namespace arm_compute \ No newline at end of file diff --git a/src/runtime/cpu/operators/CpuGemmConv2d.h b/src/runtime/cpu/operators/CpuGemmConv2d.h new file mode 100644 index 0000000000..529256594f --- /dev/null +++ b/src/runtime/cpu/operators/CpuGemmConv2d.h @@ -0,0 +1,203 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_CPU_GEMM_CONV2D_H +#define ARM_COMPUTE_CPU_GEMM_CONV2D_H + +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "src/runtime/cpu/ICpuOperator.h" + +#include + +namespace arm_compute +{ +namespace cpu +{ +class CpuGemm; +class CpuGemmLowpMatrixMultiplyCore; +class CpuGemmLowpOutputStage; +namespace kernels +{ +class CpuWeightsReshapeKernel; +class CpuIm2ColKernel; +class CpuCol2ImKernel; +class CpuReshapeKernel; +} // namespace kernels + +/** Basic function to compute the convolution layer. This function calls the following kernels/functions: + * + * -# @ref cpu::kernels::CpuIm2ColKernel + * -# @ref CpuGemm (if the data type is BFLOAT16/FP16/FP32) + * -# @ref CpuGemmLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED) + * -# @ref CpuGemmLowpOutputStage (if the data type is QASYMM8/QASYMM8_SIGNED) + * -# @ref cpu::kernels::CpuCol2ImKernel (if NCHW data layout) + * -# @ref kernels::CpuWeightsReshapeKernel + * + */ +class CpuGemmConv2d : public ICpuOperator +{ +public: + /** Constructor */ + CpuGemmConv2d(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuGemmConv2d(const CpuGemmConv2d &) = delete; + /** Prevent instances of this class from being moved (As this class contains non movable objects) */ + CpuGemmConv2d(CpuGemmConv2d &&) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuGemmConv2d &operator=(const CpuGemmConv2d &) = delete; + /** Prevent instances of this class from being moved (As this class contains non movable objects) */ + CpuGemmConv2d &operator=(CpuGemmConv2d &&) = delete; + /** Destructor */ + ~CpuGemmConv2d(); + /** Set the input and output tensors. + * + * Valid data layouts: + * - NHWC + * - NCHW + * + * Valid data type configurations: + * |src0 |src1 |src2 |dst | + * |:--------------|:------------------|:--------|:--------------| + * |F16 |F16 |F16 |F16 | + * |F32 |F32 |F32 |F32 | + * |BFLOAT16 |BFLOAT16 |BFLOAT16 |BFLOAT16 | + * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | + * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | + * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | + * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | + * + * @param[in] src Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. + * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. + * Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32. + * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. + * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. + * @param[out] dst Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * Data types supported: Same as @p input. + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. + * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights + * tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. + * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation + * available which may introduce a drop of accuracy as well. Default is false + * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported + */ + void configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(), + const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false, unsigned int num_groups = 1); + /** Static function to check if given info will lead to a valid configuration + * + * Similar to CpuGemmConvolution::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), + bool enable_fast_math = false, unsigned int num_groups = 1); + + // Inherited methods overridden: + void run(ITensorPack &tensors) override; + void prepare(ITensorPack &tensors) override; + experimental::MemoryRequirements workspace() const override; + +private: + /** Configures the appropriate matrix multiply routine + * + * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. + * @param[in] weights Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32. + * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. + * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. + * @param[out] dst Output tensor info. Data types supported: Same as @p input, + * except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. + * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation + * available which may introduce a drop of accuracy as well. Default is false + * @param[in] gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1) + */ + void configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo(), + bool enable_fast_math = false, int gemm_3d_depth = 1); + /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer matrix multiply routines + * + * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. + * @param[in] weights Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32. + * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. + * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. + * @param[in] dst Output tensor info. Data types supported: Same as @p input, + * except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. + * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation + * available which may introduce a drop of accuracy as well. Default is false + * @param[in] gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1) + * @param[in] skip_im2col (Optional) Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout. (Default to false) + * + * @return a status + */ + static Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo(), + bool enable_fast_math = false, int gemm_3d_depth = 1, bool skip_im2col = false); + /** Static function to check if GEMM3D is supported in @ref NEGEMM or in @ref CpuGemmMLowpMatrixMultiplyCore + * + * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. + * @param[in] weights Weights tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. + * @param[in] act_info Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. + * @param[in] gemm_3d_depth Depth of GEMM 3D + * @param[in] skip_im2col Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout + * + * @return a status + */ + static Status validate_gemm3d(const ITensorInfo *src, const ITensorInfo *weights, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col); + + enum AuxTensorIdx + { + // CpuGemmLowpMatrixMultiplyCore has up to 8 internal tensors + Im2ColOutput = 9, + WeightsReshaped, + GemmOutput, + Count + }; + + std::unique_ptr _weights_reshape_kernel; + std::unique_ptr _im2col_kernel; + std::unique_ptr _mm_gemm; + std::unique_ptr _mm_gemmlowp; + std::unique_ptr _col2im_kernel; + std::unique_ptr _reshape_kernel; + + TensorInfo _im2col_output; + TensorInfo _weights_reshaped; + TensorInfo _gemm_output; + TensorInfo _gemm_output_3d; + + DataLayout _data_layout; + + bool _skip_im2col; + bool _skip_col2im; + bool _is_quantized; + bool _is_prepared; + + experimental::MemoryRequirements _aux_mem{ Count }; +}; +} // namespace cpu +} // namespace arm_compute +#endif /* ARM_COMPUTE_CPU_GEMM_CONV2D_H */ diff --git a/src/runtime/cpu/operators/CpuGemmConvolution.cpp b/src/runtime/cpu/operators/CpuGemmConvolution.cpp deleted file mode 100644 index 81d656c905..0000000000 --- a/src/runtime/cpu/operators/CpuGemmConvolution.cpp +++ /dev/null @@ -1,612 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/runtime/cpu/operators/CpuGemmConvolution.h" - -#include "arm_compute/core/Size2D.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/core/utils/quantization/AsymmHelpers.h" -#include "arm_compute/runtime/NEON/NEScheduler.h" - -#include "src/core/cpu/kernels/CpuCol2ImKernel.h" -#include "src/core/cpu/kernels/CpuIm2ColKernel.h" -#include "src/core/cpu/kernels/CpuReshapeKernel.h" -#include "src/core/cpu/kernels/CpuWeightsReshapeKernel.h" -#include "src/core/helpers/MemoryHelpers.h" -#include "src/runtime/cpu/operators/CpuGemm.h" -#include "src/runtime/cpu/operators/CpuGemmLowpMatrixMultiplyCore.h" -#include "src/runtime/cpu/operators/CpuGemmLowpOutputStage.h" -#include "src/runtime/cpu/utils/CpuAuxTensorHandler.h" - -#include -#include - -using namespace arm_compute::misc::shape_calculator; -using namespace arm_compute::experimental; - -namespace arm_compute -{ -namespace cpu -{ -CpuGemmConvolution::CpuGemmConvolution() - : _weights_reshape_kernel(nullptr), _im2col_kernel(), _mm_gemm(), _mm_gemmlowp(), _col2im_kernel(), _reshape_kernel(), _im2col_output(), _weights_reshaped(), _gemm_output(), _gemm_output_3d(), - _data_layout(DataLayout::NCHW), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _is_prepared(false), _aux_mem(AuxTensorIdx::Count) -{ -} -CpuGemmConvolution::~CpuGemmConvolution() = default; - -void CpuGemmConvolution::configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act_info, - bool enable_fast_math, int gemm_3d_depth) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights); - ARM_COMPUTE_ERROR_THROW_ON(validate_mm(src, weights, biases, dst, act_info, enable_fast_math, gemm_3d_depth, _skip_im2col)); - - // Create GEMMInfo structure - const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, - gemm_3d_depth, _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */, - false, GEMMLowpOutputStageInfo(), false, enable_fast_math, false, act_info); - - // Supported activations in GEMM - const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, - ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, - ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU - }; - - if(_is_quantized) - { - TensorInfo tmp_src{ *src }; - TensorInfo tmp_weights{ *weights }; - // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() - // Extract and negate input and weights offset - const QuantizationInfo iqinfo = src->quantization_info(); - const QuantizationInfo wqinfo = weights->quantization_info(); - const QuantizationInfo oqinfo = (dst->total_size() == 0) ? iqinfo : dst->quantization_info(); - const UniformQuantizationInfo uiqinfo = iqinfo.uniform(); - const UniformQuantizationInfo uoqinfo = oqinfo.uniform(); - const DataType data_type = src->data_type(); - - tmp_src.set_quantization_info(QuantizationInfo(uiqinfo.scale, -uiqinfo.offset)); - if(!is_data_type_quantized_per_channel(tmp_weights.data_type())) - { - const UniformQuantizationInfo uwqinfo = wqinfo.uniform(); - tmp_weights.set_quantization_info(QuantizationInfo(uwqinfo.scale, -uwqinfo.offset)); - } - - // Merge activation with output stage - PixelValue type_min{}; - PixelValue type_max{}; - std::tie(type_min, type_max) = get_min_max(data_type); - int32_t min_activation = type_min.get(); - int32_t max_activation = type_max.get(); - - if(supported_acts.count(act_info.activation()) != 0) - { - std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act_info, data_type, uoqinfo); - } - - GEMMLowpOutputStageInfo output_info; - output_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; - output_info.gemmlowp_offset = uoqinfo.offset; - output_info.gemmlowp_min_bound = min_activation; - output_info.gemmlowp_max_bound = max_activation; - output_info.is_quantized_per_channel = (tmp_weights.data_type() == DataType::QSYMM8_PER_CHANNEL); - quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info); - - _mm_gemmlowp = std::make_unique(); - _mm_gemmlowp->configure(&tmp_src, &tmp_weights, biases, dst, GEMMInfo(false, false, true, gemm_3d_depth, _skip_im2col, false, output_info, false, enable_fast_math, false, act_info)); - - auto mm_mem_req = _mm_gemmlowp->workspace(); - for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) - { - _aux_mem[cont] = mm_mem_req[cont]; - } - } - else - { - // Configure matrix multiply function - _mm_gemm = std::make_unique(); - _mm_gemm->configure(src, weights, biases, dst, 1.0f, 0.0f, gemm_info); - auto mm_mem_req = _mm_gemm->workspace(); - for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) - { - _aux_mem[cont] = mm_mem_req[cont]; - } - } -} - -Status CpuGemmConvolution::validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, - const ActivationLayerInfo &act_info, bool enable_fast_math, int gemm_3d_depth, bool skip_im2col) -{ - const DataType data_type = src->data_type(); - const bool is_quantized = is_data_type_quantized_asymmetric(data_type); - const bool is_activation_enabled = act_info.enabled(); - - // Create GEMMInfo structure - const GEMMInfo gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, - gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */, - false, GEMMLowpOutputStageInfo(), false, enable_fast_math, false, act_info); - - if(is_quantized) - { - // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() - // Extract and negate input and weights offset - const QuantizationInfo &iqinfo = src->quantization_info(); - const QuantizationInfo &wqinfo = weights->quantization_info(); - const QuantizationInfo &oqinfo = (dst->total_size() == 0) ? iqinfo : dst->quantization_info(); - const UniformQuantizationInfo uoqinfo = oqinfo.uniform(); - - // Merge activation with output stage - PixelValue type_min{}; - PixelValue type_max{}; - std::tie(type_min, type_max) = get_min_max(data_type); - int32_t min_activation = type_min.get(); - int32_t max_activation = type_max.get(); - - const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, - ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, - ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU - }; - if(is_activation_enabled && supported_acts.count(act_info.activation()) != 0) - { - std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act_info, data_type, uoqinfo); - } - - GEMMLowpOutputStageInfo output_info; - output_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; - output_info.gemmlowp_offset = uoqinfo.offset; - output_info.gemmlowp_min_bound = min_activation; - output_info.gemmlowp_max_bound = max_activation; - output_info.is_quantized_per_channel = (weights->data_type() == DataType::QSYMM8_PER_CHANNEL); - ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info)); - - // Perform validation step on GEMMLowp - std::unique_ptr input_qa = src->clone(); - std::unique_ptr weights_qa = weights->clone(); - input_qa->set_quantization_info(QuantizationInfo(iqinfo.uniform().scale, -iqinfo.uniform().offset)); - weights_qa->set_quantization_info(QuantizationInfo(wqinfo.uniform().scale, -wqinfo.uniform().offset)); - return CpuGemmLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, dst, GEMMInfo(false, false, true, gemm_3d_depth, skip_im2col, false, output_info, - false, enable_fast_math, false, act_info)); - } - else - { - // Perform validation step on Matrix multiply function - return CpuGemm::validate(src, weights, nullptr, dst, 1.0f, 0.0f, gemm_info); - } -} - -Status CpuGemmConvolution::validate_gemm3d(const ITensorInfo *input_info, const ITensorInfo *weights_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col) -{ - const DataType data_type = input_info->data_type(); - const unsigned int mult_y = skip_im2col ? 1U : gemm_3d_depth; - const unsigned int mult_z = skip_im2col ? gemm_3d_depth : 1U; - - // Set dummy tensor shapes for the validation - const TensorInfo dummy_input_info(TensorShape(4U, 4U * mult_y, 1U * mult_z), 1, data_type, input_info->quantization_info()); - const TensorInfo dummy_weights_info(TensorShape(4U, 4U), 1, data_type, weights_info->quantization_info()); - const TensorInfo dummy_output_info(TensorShape(4U, 4U, gemm_3d_depth), 1, data_type, input_info->quantization_info()); - - return validate_mm(&dummy_input_info, &dummy_weights_info, nullptr, &dummy_output_info, act_info, false, gemm_3d_depth, skip_im2col); -} - -void CpuGemmConvolution::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, - const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); - ARM_COMPUTE_UNUSED(num_groups, weights_info); - ARM_COMPUTE_ERROR_THROW_ON(CpuGemmConvolution::validate(src, - weights, - biases, - dst, - conv_info, - weights_info, - dilation, - act_info, - enable_fast_math, - num_groups)); - - const DataType data_type = src->data_type(); - const DataLayout data_layout = src->data_layout(); - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); - - const unsigned int kernel_width = weights->dimension(idx_width); - const unsigned int kernel_height = weights->dimension(idx_height); - - _is_prepared = weights_info.retain_internal_weights(); - _is_quantized = is_data_type_quantized_asymmetric(src->data_type()); - _data_layout = data_layout; - _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1); - - const ITensorInfo *gemm_input_to_use = src; - ITensorInfo *gemm_output_to_use = dst; - - // Get convolved dimensions - unsigned int conv_w = 0; - unsigned int conv_h = 0; - std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), - src->dimension(idx_height), - kernel_width, - kernel_height, - conv_info, - dilation); - ARM_COMPUTE_ERROR_ON_MSG((dst->dimension(idx_width) != conv_w) || (dst->dimension(idx_height) != conv_h), - "Output shape does not match the expected one"); - - // Check if GEMM3D is supported - if(data_layout == DataLayout::NHWC) - { - _skip_col2im = bool(validate_gemm3d(src, weights, act_info, conv_h, true)); - // If not supported, we need to perform im2col and col2im (or reshape layer) - if(!_skip_col2im) - { - _skip_im2col = false; - } - } - else - { - _skip_col2im = false; - } - - // Get parameters from conv_info - unsigned int stride_x = 0; - unsigned int stride_y = 0; - std::tie(stride_x, stride_y) = conv_info.stride(); - - unsigned int mat_weights_cols = weights->dimension(idx_kernels); - - // _weights_reshaped will be auto configured in the kernel. - // Just append biases and do not transpose 1xW as it will be reshaped in CpuGemm - _weights_reshape_kernel = std::make_unique(); - _weights_reshape_kernel->configure(weights, nullptr, &_weights_reshaped); - _weights_reshaped.set_quantization_info(weights->quantization_info()); - - // Create tensor to store im2col reshaped inputs - if(!_skip_im2col) - { - // Configure - _im2col_kernel = std::make_unique(); - _im2col_kernel->configure(src, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, false, dilation); - - // Update GEMM input - gemm_input_to_use = &_im2col_output; - } - - // Create temporary GEMM output tensor in case we cannot skip col2im - const DataType output_data_type = data_type == DataType::BFLOAT16 ? DataType::F32 : data_type; - if(!_skip_col2im) - { - TensorShape shape_gemm; - - // Calculate GEMM output shape - shape_gemm = _im2col_output.tensor_shape(); - shape_gemm.set(0, mat_weights_cols); - shape_gemm.set(1, conv_w * conv_h); - - _gemm_output = TensorInfo(shape_gemm, 1, output_data_type); - _gemm_output.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); - _gemm_output_3d = TensorInfo(_gemm_output); - - // Update GEMM output - gemm_output_to_use = &_gemm_output; - } - else - { - _gemm_output_3d = TensorInfo(*dst); - _gemm_output_3d.set_data_type(output_data_type).set_data_layout(src->data_layout()).set_is_resizable(true); - _gemm_output = TensorInfo(_gemm_output_3d); - - // Update GEMM output - gemm_output_to_use = &_gemm_output_3d; - } - - // Configure GEMM - // In case we need to skip col2im, GEMM3D (gemm_3d_depth != 0) must be called in order to avoid reshaping the output matrix - const unsigned int gemm_3d_depth = _skip_col2im ? conv_h : 0; - configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, act_info, enable_fast_math, gemm_3d_depth); - - if(!_skip_col2im && _data_layout == DataLayout::NCHW) - { - // Configure col2im - _col2im_kernel = std::make_unique(); - _col2im_kernel->configure(gemm_output_to_use, dst, Size2D(conv_w, conv_h)); - } - else - { - // Configure reshape layer - _reshape_kernel = std::make_unique(); - _reshape_kernel->configure(gemm_output_to_use, dst); - } - - // Check if GEMM transforms weights - // Modernise through COMPMID-4535 - bool gemm_trans_wei = _aux_mem[1].size > 0; // Asm Pretranspose - gemm_trans_wei = _mm_gemm != nullptr ? _aux_mem[3].size > 0 : gemm_trans_wei; // Tranpose RHS - gemm_trans_wei = _mm_gemmlowp != nullptr ? _aux_mem[5].size > 0 : gemm_trans_wei; // Transpose RHS - - // Check lifetime - _aux_mem[Im2ColOutput] = MemoryInfo(offset_int_vec(Im2ColOutput), MemoryLifetime::Temporary, _im2col_output.total_size()); - _aux_mem[WeightsReshaped] = MemoryInfo(offset_int_vec(WeightsReshaped), gemm_trans_wei ? MemoryLifetime::Prepare : MemoryLifetime::Persistent, _weights_reshaped.total_size()); - _aux_mem[GemmOutput] = MemoryInfo(offset_int_vec(GemmOutput), MemoryLifetime::Temporary, _gemm_output.total_size()); -} - -Status CpuGemmConvolution::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, - const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!"); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::BFLOAT16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::BFLOAT16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, weights); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups > 1, "Grouping (num_groups != 1) is not supported"); - - const DataLayout data_layout = src->data_layout(); - const DataType data_type = src->data_type(); - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); - const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); - - const unsigned int kernel_width = weights->dimension(idx_width); - const unsigned int kernel_height = weights->dimension(idx_height); - - TensorInfo im2col_reshaped_info{}; - TensorInfo info_gemm{}; - TensorInfo tmp_info{}; - TensorInfo weights_reshaped_info{}; - const ITensorInfo *gemm_input_to_use = src; - const ITensorInfo *gemm_output_to_use = dst; - const ITensorInfo *weights_to_use = weights; - - const bool append_bias = false; - const bool is_quantized = is_data_type_quantized_asymmetric(data_type); - const bool is_bf16 = data_type == DataType::BFLOAT16; - bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1); - - // Get convolved dimensions - unsigned int conv_w = 0; - unsigned int conv_h = 0; - - std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), - src->dimension(idx_height), - kernel_width, - kernel_height, - conv_info, - dilation); - - // Check if GEMM3D is supported - bool skip_col2im = false; - if(data_layout == DataLayout::NHWC) - { - skip_col2im = bool(validate_gemm3d(src, weights, act_info, conv_h, true)); - // If not supported, we need to perform im2col and col2im (or reshape layer) - if(!skip_col2im) - { - skip_im2col = false; - } - } - - if(skip_col2im) - { - // If not supported, we need to perform im2col and col2im (or reshape layer) - if(!bool(validate_gemm3d(src, weights, act_info, conv_h, skip_im2col))) - { - skip_im2col = false; - skip_col2im = false; - } - } - - ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_channel) != src->dimension(idx_channel)); - ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); - - // Validate biases - if(biases != nullptr) - { - if(is_quantized) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); - } - else if(is_bf16) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32); - } - else - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases); - } - ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(idx_kernels)); - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - } - - unsigned int mat_weights_cols = weights->dimension(idx_kernels); - unsigned int mat_weights_rows = weights->dimension(idx_width) * weights->dimension(idx_height) * weights->dimension(idx_channel); - - weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, append_bias), 1, data_type); - weights_reshaped_info.set_quantization_info(weights->quantization_info()); - weights_to_use = &weights_reshaped_info; - - if(!skip_im2col) - { - // Create tensor info for im2col reshaped inputs - // For CPU, the batch size is on the fourth dimension - TensorShape shape_im2col = src->tensor_shape(); - shape_im2col.set(0, mat_weights_rows); - shape_im2col.set(1, conv_w * conv_h); - shape_im2col.set(2, 1); - - im2col_reshaped_info = TensorInfo(shape_im2col, 1, data_type); - im2col_reshaped_info.set_quantization_info(src->quantization_info()); - ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuIm2ColKernel::validate(src, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation)); - gemm_input_to_use = &im2col_reshaped_info; - } - - // Create temporary GEMM output tensor in case we cannot skip col2im - const DataType output_data_type = data_type == DataType::BFLOAT16 ? DataType::F32 : data_type; - if(!skip_col2im) - { - TensorShape shape_gemm = gemm_input_to_use->tensor_shape(); - shape_gemm.set(0, mat_weights_cols); - shape_gemm.set(1, conv_w * conv_h); - info_gemm = TensorInfo(shape_gemm, 1, output_data_type); - } - else - { - info_gemm = TensorInfo(dst->tensor_shape(), 1, output_data_type); - } - info_gemm.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); - gemm_output_to_use = &info_gemm; - ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases, gemm_output_to_use, act_info, enable_fast_math, skip_col2im ? conv_h : 0, skip_im2col)); - - // Validate Col2Im/ReshapeLayer - if(!skip_col2im && (data_layout == DataLayout::NCHW)) - { - ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuCol2ImKernel::validate(gemm_output_to_use, dst, Size2D(conv_w, conv_h))); - } - - return Status{}; -} - -void CpuGemmConvolution::run(ITensorPack &tensors) -{ - prepare(tensors); - - auto src = tensors.get_const_tensor(ACL_SRC_0); - auto dst = tensors.get_tensor(ACL_DST); - auto gemm_input_to_use = src; - - CpuAuxTensorHandler im2col_output(offset_int_vec(Im2ColOutput), _im2col_output, tensors, false); - CpuAuxTensorHandler gemm_output(offset_int_vec(GemmOutput), _gemm_output, tensors, false); - CpuAuxTensorHandler reshaped_wei(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors, false); - - bool out_has_padding = _skip_col2im && (dst->info()->padding().bottom != 0 || dst->info()->padding().top != 0); - if(!_skip_im2col) - { - // Run input reshaping - unsigned int y_dim = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); - ITensorPack pack = - { - { TensorType::ACL_SRC, src }, - { TensorType::ACL_DST, im2col_output.get() } - }; - NEScheduler::get().schedule_op(_im2col_kernel.get(), y_dim, _im2col_kernel->window(), pack); - gemm_input_to_use = im2col_output.get(); - } - - // Handle the case where output has top/bottom padding - const ITensor *out_to_use = out_has_padding ? gemm_output.get() : dst; - Tensor gemm3d; - _gemm_output_3d.extend_padding(out_to_use->info()->padding()); - gemm3d.allocator()->soft_init(_gemm_output_3d); - gemm3d.allocator()->import_memory(out_to_use->buffer()); - auto gemm_output_to_use = gemm_output.get(); - - if(_skip_im2col) - { - gemm_output_to_use = &gemm3d; - } - if(_skip_col2im && !out_has_padding) - { - gemm_output_to_use = dst; - } - - // Runs CpuGemm or CpuGemmLowpMatrixMultiplyCore functions - ITensorPack pack_mm = tensors; - pack_mm.add_const_tensor(TensorType::ACL_SRC_0, gemm_input_to_use); - pack_mm.add_const_tensor(TensorType::ACL_SRC_1, reshaped_wei.get()); - pack_mm.add_tensor(TensorType::ACL_DST, gemm_output_to_use); - if(_is_quantized) - { - // Run gemmlowp - _mm_gemmlowp->run(pack_mm); - } - else - { - // Run gemm - _mm_gemm->run(pack_mm); - } - - // Reshape output matrix - if(!_skip_col2im) - { - if(_data_layout == DataLayout::NCHW) - { - ITensorPack pack = - { - { TensorType::ACL_SRC, gemm_output.get() }, - { TensorType::ACL_DST, dst } - }; - NEScheduler::get().schedule_op(_col2im_kernel.get(), Window::DimY, _col2im_kernel->window(), pack); - } - else - { - ITensorPack pack = - { - { TensorType::ACL_SRC, gemm_output_to_use }, - { TensorType::ACL_DST, dst } - }; - NEScheduler::get().schedule_op(_reshape_kernel.get(), Window::DimY, _reshape_kernel->window(), pack); - } - } - else if(out_has_padding) - { - ITensorPack pack = - { - { TensorType::ACL_SRC, gemm_output_to_use }, - { TensorType::ACL_DST, dst } - }; - NEScheduler::get().schedule_op(_reshape_kernel.get(), Window::DimY, _reshape_kernel->window(), pack); - } -} - -void CpuGemmConvolution::prepare(ITensorPack &tensors) -{ - if(!_is_prepared) - { - // Run weights reshaping and mark original weights tensor as unused - CpuAuxTensorHandler weights_reshaped(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors); - auto weights = tensors.get_const_tensor(TensorType::ACL_SRC_1); - ITensorPack pack = - { - { TensorType::ACL_SRC, weights }, - { TensorType::ACL_DST, weights_reshaped.get() } - }; - NEScheduler::get().schedule_op(_weights_reshape_kernel.get(), 3, _weights_reshape_kernel->window(), pack); - weights->mark_as_unused(); - - // Prepare GEMM - ITensorPack gemm_pack = tensors; - gemm_pack.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get()); - _is_quantized ? _mm_gemmlowp->prepare(gemm_pack) : _mm_gemm->prepare(gemm_pack); - - _is_prepared = true; - } -} -experimental::MemoryRequirements CpuGemmConvolution::workspace() const -{ - return _aux_mem; -} -} // namespace cpu -} // namespace arm_compute \ No newline at end of file diff --git a/src/runtime/cpu/operators/CpuGemmConvolution.h b/src/runtime/cpu/operators/CpuGemmConvolution.h deleted file mode 100644 index 7755bbe2a2..0000000000 --- a/src/runtime/cpu/operators/CpuGemmConvolution.h +++ /dev/null @@ -1,203 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CPU_GEMMCONVOLUTION_H -#define ARM_COMPUTE_CPU_GEMMCONVOLUTION_H - -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Types.h" -#include "src/runtime/cpu/ICpuOperator.h" - -#include - -namespace arm_compute -{ -namespace cpu -{ -class CpuGemm; -class CpuGemmLowpMatrixMultiplyCore; -class CpuGemmLowpOutputStage; -namespace kernels -{ -class CpuWeightsReshapeKernel; -class CpuIm2ColKernel; -class CpuCol2ImKernel; -class CpuReshapeKernel; -} // namespace kernels - -/** Basic function to compute the convolution layer. This function calls the following kernels/functions: - * - * -# @ref cpu::kernels::CpuIm2ColKernel - * -# @ref CpuGemm (if the data type is BFLOAT16/FP16/FP32) - * -# @ref CpuGemmLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED) - * -# @ref CpuGemmLowpOutputStage (if the data type is QASYMM8/QASYMM8_SIGNED) - * -# @ref cpu::kernels::CpuCol2ImKernel (if NCHW data layout) - * -# @ref kernels::CpuWeightsReshapeKernel - * - */ -class CpuGemmConvolution : public ICpuOperator -{ -public: - /** Constructor */ - CpuGemmConvolution(); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - CpuGemmConvolution(const CpuGemmConvolution &) = delete; - /** Prevent instances of this class from being moved (As this class contains non movable objects) */ - CpuGemmConvolution(CpuGemmConvolution &&) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - CpuGemmConvolution &operator=(const CpuGemmConvolution &) = delete; - /** Prevent instances of this class from being moved (As this class contains non movable objects) */ - CpuGemmConvolution &operator=(CpuGemmConvolution &&) = delete; - /** Destructor */ - ~CpuGemmConvolution(); - /** Set the input and output tensors. - * - * Valid data layouts: - * - NHWC - * - NCHW - * - * Valid data type configurations: - * |src0 |src1 |src2 |dst | - * |:--------------|:------------------|:--------|:--------------| - * |F16 |F16 |F16 |F16 | - * |F32 |F32 |F32 |F32 | - * |BFLOAT16 |BFLOAT16 |BFLOAT16 |BFLOAT16 | - * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | - * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | - * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | - * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | - * - * @param[in] src Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], - * while every optional dimension from 4 and above represent a batch of inputs. - * Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. - * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. - * Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32. - * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. - * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. - * @param[out] dst Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. - * Data types supported: Same as @p input. - * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. - * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights - * tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input. - * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). - * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. - * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation - * available which may introduce a drop of accuracy as well. Default is false - * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported - */ - void configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(), - const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false, unsigned int num_groups = 1); - /** Static function to check if given info will lead to a valid configuration - * - * Similar to CpuGemmConvolution::configure() - * - * @return a status - */ - static Status validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, - const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), - bool enable_fast_math = false, unsigned int num_groups = 1); - - // Inherited methods overridden: - void run(ITensorPack &tensors) override; - void prepare(ITensorPack &tensors) override; - experimental::MemoryRequirements workspace() const override; - -private: - /** Configures the appropriate matrix multiply routine - * - * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. - * @param[in] weights Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32. - * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. - * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. - * @param[out] dst Output tensor info. Data types supported: Same as @p input, - * except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type. - * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. - * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation - * available which may introduce a drop of accuracy as well. Default is false - * @param[in] gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1) - */ - void configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo(), - bool enable_fast_math = false, int gemm_3d_depth = 1); - /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer matrix multiply routines - * - * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. - * @param[in] weights Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32. - * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. - * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. - * @param[in] dst Output tensor info. Data types supported: Same as @p input, - * except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type. - * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. - * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation - * available which may introduce a drop of accuracy as well. Default is false - * @param[in] gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1) - * @param[in] skip_im2col (Optional) Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout. (Default to false) - * - * @return a status - */ - static Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo(), - bool enable_fast_math = false, int gemm_3d_depth = 1, bool skip_im2col = false); - /** Static function to check if GEMM3D is supported in @ref NEGEMM or in @ref CpuGemmMLowpMatrixMultiplyCore - * - * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. - * @param[in] weights Weights tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. - * @param[in] act_info Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. - * @param[in] gemm_3d_depth Depth of GEMM 3D - * @param[in] skip_im2col Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout - * - * @return a status - */ - static Status validate_gemm3d(const ITensorInfo *src, const ITensorInfo *weights, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col); - - enum AuxTensorIdx - { - // CpuGemmLowpMatrixMultiplyCore has up to 8 internal tensors - Im2ColOutput = 9, - WeightsReshaped, - GemmOutput, - Count - }; - - std::unique_ptr _weights_reshape_kernel; - std::unique_ptr _im2col_kernel; - std::unique_ptr _mm_gemm; - std::unique_ptr _mm_gemmlowp; - std::unique_ptr _col2im_kernel; - std::unique_ptr _reshape_kernel; - - TensorInfo _im2col_output; - TensorInfo _weights_reshaped; - TensorInfo _gemm_output; - TensorInfo _gemm_output_3d; - - DataLayout _data_layout; - - bool _skip_im2col; - bool _skip_col2im; - bool _is_quantized; - bool _is_prepared; - - experimental::MemoryRequirements _aux_mem{ Count }; -}; -} // namespace cpu -} // namespace arm_compute -#endif /* ARM_COMPUTE_CPU_GEMMCONVOLUTION_H */ diff --git a/src/runtime/gpu/cl/operators/ClConv2d.cpp b/src/runtime/gpu/cl/operators/ClConv2d.cpp index 2f4d673d9c..0cb3a968e6 100644 --- a/src/runtime/gpu/cl/operators/ClConv2d.cpp +++ b/src/runtime/gpu/cl/operators/ClConv2d.cpp @@ -31,7 +31,7 @@ #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h" #include "src/runtime/gpu/cl/operators/ClDirectConv2d.h" -#include "src/runtime/gpu/cl/operators/ClGemmConvolution.h" +#include "src/runtime/gpu/cl/operators/ClGemmConv2d.h" #include "src/runtime/gpu/cl/operators/ClWinogradConv2d.h" #include @@ -104,7 +104,7 @@ void ClConv2d::configure(const CLCompileContext &compile_context, ITensorInfo *s } case ConvolutionMethod::GEMM: { - auto f = std::make_unique(); + auto f = std::make_unique(); f->configure(compile_context, src, weights, biases, dst, conv2d_info, weights_info); _operator = std::move(f); break; @@ -143,7 +143,7 @@ Status ClConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, co case ConvolutionMethod::GEMM: { // Validate gemm-based convolution layer - ARM_COMPUTE_RETURN_ON_ERROR(ClGemmConvolution::validate(src, weights, biases, dst, conv2d_info, weights_info)); + ARM_COMPUTE_RETURN_ON_ERROR(ClGemmConv2d::validate(src, weights, biases, dst, conv2d_info, weights_info)); break; } default: diff --git a/src/runtime/gpu/cl/operators/ClConv2d.h b/src/runtime/gpu/cl/operators/ClConv2d.h index 0888c2f47b..cdf3b7df32 100644 --- a/src/runtime/gpu/cl/operators/ClConv2d.h +++ b/src/runtime/gpu/cl/operators/ClConv2d.h @@ -36,7 +36,7 @@ namespace opencl { /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions: * - * -# @ref opencl::ClGemmConvolution + * -# @ref opencl::ClGemmConv2d * -# @ref opencl::ClWinogradConv2d * -# @ref opencl::ClDirectConv2d * -# @ref CLFFTConvolutionLayer diff --git a/src/runtime/gpu/cl/operators/ClGemmConv2d.cpp b/src/runtime/gpu/cl/operators/ClGemmConv2d.cpp new file mode 100644 index 0000000000..8c796e0712 --- /dev/null +++ b/src/runtime/gpu/cl/operators/ClGemmConv2d.cpp @@ -0,0 +1,628 @@ +/* + * Copyright (c) 2017-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "src/runtime/gpu/cl/operators/ClGemmConv2d.h" + +#include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/PixelValue.h" +#include "arm_compute/core/Size2D.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "arm_compute/runtime/CL/CLScheduler.h" +#include "src/core/gpu/cl/kernels/ClActivationKernel.h" +#include "src/core/gpu/cl/kernels/ClCol2ImKernel.h" +#include "src/core/gpu/cl/kernels/ClIm2ColKernel.h" +#include "src/core/gpu/cl/kernels/ClWeightsReshapeKernel.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/runtime/gpu/cl/operators/ClGemm.h" +#include "src/runtime/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.h" +#include "src/runtime/gpu/cl/utils/ClAuxTensorHandler.h" +#include "support/Cast.h" + +namespace arm_compute +{ +using namespace experimental; +using namespace misc::shape_calculator; +using namespace utils::cast; +namespace opencl +{ +ClGemmConv2d::ClGemmConv2d() + : _weights_reshape_kernel(nullptr), _im2col_kernel(nullptr), _mm_gemm(nullptr), _mm_gemmlowp(nullptr), _col2im_kernel(nullptr), _activation_kernel(nullptr), _im2col_output(), _weights_reshaped(), + _gemm_output(), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _fuse_activation(true), _append_bias(false), _is_prepared(false), _aux_mem(AuxTensorIdx::Count) +{ +} +ClGemmConv2d::~ClGemmConv2d() = default; + +void ClGemmConv2d::configure_mm(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, + const GEMMLowpOutputStageInfo &gemmlowp_output_stage, + int gemm_3d_depth, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights); + ARM_COMPUTE_ERROR_THROW_ON(validate_mm(src, weights, biases, dst, gemmlowp_output_stage, gemm_3d_depth, _skip_im2col, act_info)); + + const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped + false, // is_b_reshaped + true, // reshape_b_only_on_first_run + gemm_3d_depth, // depth_output_gemm3d + _skip_im2col, // reinterpret_input_as_3d + false, // retain_internal_weights + gemmlowp_output_stage, // gemmlowp_output_stage + false, // fast_math + false, // fp_mixed_precision + true, // broadcast_bias + act_info); // activation_info + + TensorInfo tmp_src{ *src }; + if(_is_quantized) + { + // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() + // Extract and negate input and weights offset + const QuantizationInfo input_quantization_info = src->quantization_info(); + const QuantizationInfo weights_quantization_info = weights->quantization_info(); + + tmp_src.set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset)); + weights->set_quantization_info(QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset)); + + _mm_gemmlowp = std::make_unique(); + _mm_gemmlowp->configure(compile_context, &tmp_src, weights, biases, dst, gemm_info); + + // Revert back QuantizatioInfo as weights could be used in other convolution layers + weights->set_quantization_info(weights_quantization_info); + + auto mm_mem_req = _mm_gemmlowp->workspace(); + for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) + { + _aux_mem[cont] = mm_mem_req[cont]; + } + } + else + { + // Configure matrix multiply function + _mm_gemm = std::make_unique(); + _mm_gemm->configure(compile_context, &tmp_src, weights, biases, dst, 1.0f, 1.0f, gemm_info); + auto mm_mem_req = _mm_gemm->workspace(); + for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) + { + _aux_mem[cont] = mm_mem_req[cont]; + } + } +} + +Status ClGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, + const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth, bool skip_im2col, const ActivationLayerInfo &act_info) +{ + const bool is_quantized = is_data_type_quantized_asymmetric(src->data_type()); + + const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped + false, // is_b_reshaped + true, // reshape_b_only_on_first_run + gemm_3d_depth, // depth_output_gemm3d + skip_im2col, // reinterpret_input_as_3d + false, // retain_internal_weights + gemmlowp_output_stage, // gemmlowp_output_stage + false, // fast_math + false, // fp_mixed_precision + true, // broadcast_bias + act_info); // activation_info + + if(is_quantized) + { + // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() + // Extract and negate input and weights offset + const QuantizationInfo input_quantization_info = src->quantization_info(); + const QuantizationInfo weights_quantization_info = weights->quantization_info(); + + std::unique_ptr src_qa = src->clone(); + std::unique_ptr weights_qa = weights->clone(); + src_qa->set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset)); + weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset)); + + // Perform validation step on GEMMLowp + return ClGemmLowpMatrixMultiplyCore::validate(src_qa.get(), weights_qa.get(), biases, dst, gemm_info); + } + else + { + // Perform validation step on Matrix multiply function + return ClGemm::validate(src, weights, biases, dst, 1.0f, 1.0f, gemm_info); + } +} + +void ClGemmConv2d::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, + const Conv2dInfo &conv2d_info, const WeightsInfo &weights_info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); + + ARM_COMPUTE_ERROR_THROW_ON(ClGemmConv2d::validate(src, weights, biases, dst, + conv2d_info, + weights_info)); + + const DataType data_type = src->data_type(); + const DataLayout data_layout = src->data_layout(); + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); + + const unsigned int kernel_width = weights->dimension(idx_width); + const unsigned int kernel_height = weights->dimension(idx_height); + const unsigned int num_kernels = weights->dimension(idx_kernels); + + const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); + + _is_prepared = weights_info.retain_internal_weights(); + _is_quantized = is_data_type_quantized_asymmetric(src->data_type()); + _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv2d_info.conv_info.stride().first == 1 && conv2d_info.conv_info.stride().second == 1); + _skip_col2im = data_layout == DataLayout::NHWC; + + // Only for quantize there are few cases where we cannot fuse the activation function in GEMM + _fuse_activation = true; + + const ITensorInfo *gemm_input_to_use = src; + ITensorInfo *gemm_output_to_use = dst; + + // Get parameters from conv_info + unsigned int stride_x = 0; + unsigned int stride_y = 0; + std::tie(stride_x, stride_y) = conv2d_info.conv_info.stride(); + + // Get convolved dimensions + unsigned int conv_w = 0; + unsigned int conv_h = 0; + std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), + src->dimension(idx_height), + kernel_width, + kernel_height, + conv2d_info.conv_info, + conv2d_info.dilation); + + unsigned int mat_weights_cols = num_kernels / conv2d_info.num_groups; + + ITensorInfo *biases_to_use = biases; + _append_bias = false; + + _weights_reshape_kernel = std::make_unique(); + if(conv2d_info.num_groups != 1 && biases != nullptr) + { + // num_groups != 1 can only be for NCHW + // Since it is missing an utility function to reshape the biases, we append the biases into the weights tensor + biases_to_use = nullptr; + _append_bias = true; + _weights_reshape_kernel->configure(compile_context, weights, biases, &_weights_reshaped, conv2d_info.num_groups); + } + else + { + _weights_reshape_kernel->configure(compile_context, weights, nullptr, &_weights_reshaped, conv2d_info.num_groups); + } + + // Create tensor to store im2col reshaped inputs + if(!_skip_im2col) + { + // Configure and tune im2col. im2col output shape is auto-initialized + _im2col_kernel = std::make_unique(); + + // Set the GPU target for im2col + _im2col_kernel->set_target(CLScheduler::get().target()); + _im2col_kernel->configure(compile_context, src, &_im2col_output, Size2D(kernel_width, kernel_height), conv2d_info.conv_info, _append_bias, conv2d_info.dilation, conv2d_info.num_groups); + + // Set quantization info + _im2col_output.set_quantization_info(src->quantization_info()); + CLScheduler::get().tune_kernel_static(*_im2col_kernel); + + // Update GEMM input + gemm_input_to_use = &_im2col_output; + } + + // Create GEMM output tensor + if(!_skip_col2im) + { + TensorShape shape_gemm; + + // If we cannot skip col2im it means we run im2col as well + shape_gemm = _im2col_output.tensor_shape(); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, conv_w * conv_h); + + _gemm_output = TensorInfo(shape_gemm, 1, data_type); + _gemm_output.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); + + // Update GEMM output + gemm_output_to_use = &_gemm_output; + } + + GEMMLowpOutputStageInfo gemmlowp_output_stage; + gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; + gemmlowp_output_stage.gemmlowp_offset = 0; + + // Configure output stage for quantized case + if(_is_quantized) + { + const auto output_quant_info = (dst->total_size() == 0) ? iq_info : oq_info; + const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->data_type()); + const unsigned int num_filters = (is_quantized_per_channel) ? num_kernels : 1; + + gemmlowp_output_stage.is_quantized_per_channel = is_quantized_per_channel; + + gemmlowp_output_stage.gemmlowp_multipliers.resize(num_filters); + gemmlowp_output_stage.gemmlowp_shifts.resize(num_filters); + quantization::compute_quantized_multipliers_and_shifts(src, weights, dst, + gemmlowp_output_stage.gemmlowp_multipliers.data(), + gemmlowp_output_stage.gemmlowp_shifts.data()); + gemmlowp_output_stage.gemmlowp_multiplier = gemmlowp_output_stage.gemmlowp_multipliers[0]; + gemmlowp_output_stage.gemmlowp_shift = gemmlowp_output_stage.gemmlowp_shifts[0]; + + PixelValue min_val{}; + PixelValue max_val{}; + std::tie(min_val, max_val) = get_min_max(dst->data_type()); + + auto min_activation = min_val.get(); + auto max_activation = max_val.get(); + + const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, + ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, + ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU + }; + + if(conv2d_info.act_info.enabled()) + { + if(supported_acts.count(conv2d_info.act_info.activation()) != 0) + { + std::tie(min_activation, max_activation) = get_quantized_activation_min_max(conv2d_info.act_info, data_type, output_quant_info); + } + else + { + _fuse_activation = false; + } + } + + // Set the GEMMLowp output stage info + gemmlowp_output_stage.gemmlowp_offset = output_quant_info.offset; + gemmlowp_output_stage.gemmlowp_min_bound = min_activation; + gemmlowp_output_stage.gemmlowp_max_bound = max_activation; + } + + // Configure and tune GEMM + // In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix + const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0; + + configure_mm(compile_context, gemm_input_to_use, &_weights_reshaped, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, conv2d_info.act_info); + + if(!_skip_col2im) + { + // Set the GPU target for col2im + _col2im_kernel = std::make_unique(); + _col2im_kernel->set_target(CLScheduler::get().target()); + // Configure and tune Col2Im + _col2im_kernel->configure(compile_context, gemm_output_to_use, dst, Size2D(conv_w, conv_h), conv2d_info.num_groups); + CLScheduler::get().tune_kernel_static(*_col2im_kernel.get()); + } + + ARM_COMPUTE_ERROR_ON_MSG((dst->dimension(idx_width) != conv_w) || (dst->dimension(idx_height) != conv_h), + "Output shape does not match the expected one"); + + if(!_fuse_activation) + { + _activation_kernel = std::make_unique(); + _activation_kernel->configure(compile_context, dst, nullptr, conv2d_info.act_info); + } + + _aux_mem[Im2ColOutput] = MemoryInfo(offset_int_vec(Im2ColOutput), MemoryLifetime::Temporary, _im2col_output.total_size()); + _aux_mem[WeightsReshaped] = MemoryInfo(offset_int_vec(WeightsReshaped), MemoryLifetime::Persistent, _weights_reshaped.total_size()); + _aux_mem[GemmOutput] = MemoryInfo(offset_int_vec(GemmOutput), MemoryLifetime::Temporary, _gemm_output.total_size()); +} + +Status ClGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const Conv2dInfo &conv2d_info, + const WeightsInfo &weights_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!"); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); + const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->data_type()); + + if(!is_quantized_per_channel) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights); + } + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, weights); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((conv2d_info.num_groups != 1) && (src->data_layout() != DataLayout::NCHW), "Grouping (num_groups != 1) with NHWC data layout is not supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((conv2d_info.num_groups != 1) && (src->data_type() == DataType::QASYMM8), "Grouping (num_groups != 1) is not supported with QASYMM8"); + ARM_COMPUTE_RETURN_ERROR_ON(((src->dimension(2) / weights->dimension(2)) != conv2d_info.num_groups) && (src->data_layout() == DataLayout::NCHW)); + + const DataLayout data_layout = src->data_layout(); + const DataType data_type = src->data_type(); + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); + const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); + + const unsigned int kernel_width = weights->dimension(idx_width); + const unsigned int kernel_height = weights->dimension(idx_height); + const unsigned int num_kernels = weights->dimension(idx_kernels); + + TensorInfo im2col_reshaped_info{}; + TensorInfo info_gemm{}; + TensorInfo weights_reshaped_info{}; + const ITensorInfo *gemm_input_to_use = src; + const ITensorInfo *gemm_output_to_use = dst; + const ITensorInfo *weights_to_use = weights; + const bool is_quantized = is_data_type_quantized_asymmetric(data_type); + const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv2d_info.conv_info.stride().first == 1 + && conv2d_info.conv_info.stride().second == 1); + const bool skip_col2im = data_layout == DataLayout::NHWC; + bool fuse_activation = true; + + ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(idx_channel) * conv2d_info.num_groups) != src->dimension(idx_channel)); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); + + // Validate biases + if(biases != nullptr) + { + if(is_quantized) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases); + } + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(idx_kernels)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + if(conv2d_info.act_info.enabled()) + { + ARM_COMPUTE_ERROR_ON(conv2d_info.act_info.b() > conv2d_info.act_info.a()); + } + + // Get convolved dimensions + unsigned int conv_w = 0; + unsigned int conv_h = 0; + + std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), + src->dimension(idx_height), + kernel_width, + kernel_height, + conv2d_info.conv_info, + conv2d_info.dilation); + + unsigned int mat_weights_cols = num_kernels / conv2d_info.num_groups; + + const ITensorInfo *biases_to_use = biases; + bool append_bias = false; + + if(conv2d_info.num_groups != 1 && biases != nullptr) + { + // num_groups != 1 can only be for NCHW + // Since it is missing an utility function to reshape the biases, we append the biases into the weights tensor + biases_to_use = nullptr; + append_bias = true; + weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, true, conv2d_info.num_groups), 1, data_type); + } + else + { + weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, false, conv2d_info.num_groups), 1, data_type); + } + + weights_to_use = &weights_reshaped_info; + + if(!skip_im2col) + { + const Size2D kernel_dims(kernel_width, kernel_height); + + // Output tensor auto initialization if not yet initialized + TensorShape expected_output_shape = compute_im2col_conv_shape(src, kernel_dims, conv2d_info.conv_info, append_bias, conv2d_info.dilation, conv2d_info.num_groups == 1, conv2d_info.num_groups); + + auto_init_if_empty(im2col_reshaped_info, src->clone()->set_tensor_shape(expected_output_shape)); + + ARM_COMPUTE_RETURN_ON_ERROR(opencl::kernels::ClIm2ColKernel::validate(src, &im2col_reshaped_info, kernel_dims, conv2d_info.conv_info, append_bias, conv2d_info.dilation, conv2d_info.num_groups)); + gemm_input_to_use = &im2col_reshaped_info; + } + + // Create GEMM output tensor + if(!skip_col2im) + { + TensorShape shape_gemm; + + shape_gemm = gemm_input_to_use->tensor_shape(); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, conv_w * conv_h); + + info_gemm = TensorInfo(shape_gemm, 1, data_type); + info_gemm.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); + gemm_output_to_use = &info_gemm; + } + + GEMMLowpOutputStageInfo gemmlowp_output_stage; + gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; + gemmlowp_output_stage.gemmlowp_offset = 0; + gemmlowp_output_stage.is_quantized_per_channel = is_quantized_per_channel; + + if(is_quantized) + { + const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); + const auto output_quant_info = (dst->total_size() == 0) ? iq_info : oq_info; + const unsigned int num_filters = (is_quantized_per_channel) ? num_kernels : 1; + + gemmlowp_output_stage.gemmlowp_multipliers.resize(num_filters); + gemmlowp_output_stage.gemmlowp_shifts.resize(num_filters); + quantization::compute_quantized_multipliers_and_shifts(src, weights, dst, + gemmlowp_output_stage.gemmlowp_multipliers.data(), + gemmlowp_output_stage.gemmlowp_shifts.data()); + gemmlowp_output_stage.gemmlowp_multiplier = gemmlowp_output_stage.gemmlowp_multipliers[0]; + gemmlowp_output_stage.gemmlowp_shift = gemmlowp_output_stage.gemmlowp_shifts[0]; + + int min_activation = 0; + int max_activation = 0; + + const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, + ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, + ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU + }; + + if(conv2d_info.act_info.enabled()) + { + if(supported_acts.count(conv2d_info.act_info.activation()) != 0) + { + std::tie(min_activation, max_activation) = get_quantized_activation_min_max(conv2d_info.act_info, data_type, output_quant_info); + } + else + { + fuse_activation = false; + } + } + + // Set the GEMMLowp output stage info + gemmlowp_output_stage.gemmlowp_offset = output_quant_info.offset; + gemmlowp_output_stage.gemmlowp_min_bound = min_activation; + gemmlowp_output_stage.gemmlowp_max_bound = max_activation; + } + + // In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix + const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0; + + ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, skip_im2col, conv2d_info.act_info)); + + // Validate Col2Im + if(!skip_col2im) + { + ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClCol2ImKernel::validate(gemm_output_to_use, dst, Size2D(conv_w, conv_h), conv2d_info.num_groups)); + } + + //Validate Activation Layer + if(!fuse_activation) + { + ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClActivationKernel::validate(dst, nullptr, conv2d_info.act_info)); + } + + return Status{}; +} + +void ClGemmConv2d::run(ITensorPack &tensors) +{ + prepare(tensors); + + auto src = tensors.get_const_tensor(ACL_SRC_0); + auto biases = tensors.get_const_tensor(ACL_SRC_2); + auto dst = tensors.get_tensor(ACL_DST); + auto gemm_input_to_use = src; + auto gemm_output_to_use = dst; + + CLAuxTensorHandler im2col_output(offset_int_vec(Im2ColOutput), _im2col_output, tensors, false); + CLAuxTensorHandler gemm_output(offset_int_vec(GemmOutput), _gemm_output, tensors, false); + CLAuxTensorHandler weights_reshaped(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors, false); + + // Run im2col + if(!_skip_im2col) + { + ITensorPack pack = + { + { TensorType::ACL_SRC, src }, + { TensorType::ACL_DST, im2col_output.get() } + }; + CLScheduler::get().enqueue_op(*_im2col_kernel, pack, false); + gemm_input_to_use = im2col_output.get(); + } + if(!_skip_col2im) + { + gemm_output_to_use = gemm_output.get(); + } + ITensorPack pack_mm = tensors; + pack_mm.add_const_tensor(TensorType::ACL_SRC_0, gemm_input_to_use); + pack_mm.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get()); + if(!_append_bias) + { + pack_mm.add_const_tensor(TensorType::ACL_SRC_2, biases); + } + pack_mm.add_tensor(TensorType::ACL_DST, gemm_output_to_use); + // Runs ClGemm or ClGemmLowpMatrixMultiplyCore functions + if(_is_quantized) + { + // Run gemmlowp + _mm_gemmlowp->run(pack_mm); + } + else + { + // Run gemm + _mm_gemm->run(pack_mm); + } + + // Reshape output matrix + if(!_skip_col2im) + { + ITensorPack pack = + { + { TensorType::ACL_SRC, gemm_output_to_use }, + { TensorType::ACL_DST, dst } + }; + CLScheduler::get().enqueue_op(*_col2im_kernel.get(), pack, false); + } + + //Run Activation Layer if we cannot fuse in GEMM + if(!_fuse_activation) + { + ITensorPack pack = + { + { TensorType::ACL_SRC, dst }, + { TensorType::ACL_DST, dst } + }; + CLScheduler::get().enqueue_op(*_activation_kernel.get(), pack, false); + } +} + +void ClGemmConv2d::prepare(ITensorPack &tensors) +{ + if(!_is_prepared) + { + // Run weights reshaping and mark original weights tensor as unused + ICLTensor *weights_reshaped_p = utils::cast::polymorphic_downcast(tensors.get_tensor(offset_int_vec(WeightsReshaped))); + CLAuxTensorHandler weights_reshaped(_weights_reshaped, *weights_reshaped_p); + auto weights = tensors.get_const_tensor(TensorType::ACL_SRC_1); + ITensorPack pack = + { + { TensorType::ACL_SRC, weights }, + { TensorType::ACL_DST, weights_reshaped.get() } + }; + + if(_append_bias) + { + const auto biases = tensors.get_const_tensor(TensorType::ACL_SRC_2); + pack.add_const_tensor(TensorType::ACL_BIAS, biases); + } + CLScheduler::get().enqueue_op(*_weights_reshape_kernel.get(), pack, true); + tensors.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get()); + + // Prepare GEMM + _is_quantized ? _mm_gemmlowp->prepare(tensors) : _mm_gemm->prepare(tensors); + _is_prepared = true; + } +} +experimental::MemoryRequirements ClGemmConv2d::workspace() const +{ + return _aux_mem; +} +} // namespace opencl +} // namespace arm_compute diff --git a/src/runtime/gpu/cl/operators/ClGemmConv2d.h b/src/runtime/gpu/cl/operators/ClGemmConv2d.h new file mode 100644 index 0000000000..e16d029e71 --- /dev/null +++ b/src/runtime/gpu/cl/operators/ClGemmConv2d.h @@ -0,0 +1,185 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_CL_GEMM_CONV2D_H +#define ARM_COMPUTE_CL_GEMM_CONV2D_H + +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/FunctionDescriptors.h" +#include "src/core/gpu/cl/ClCompileContext.h" +#include "src/runtime/gpu/cl/IClOperator.h" + +#include + +namespace arm_compute +{ +namespace opencl +{ +class ClGemm; +class ClGemmLowpMatrixMultiplyCore; +namespace kernels +{ +class ClIm2ColKernel; +class ClCol2ImKernel; +class ClWeightsReshapeKernel; +class ClActivationKernel; +} // namespace kernels + +/** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions: + * + * -# @ref opencl::kernels::ClIm2ColKernel + * -# @ref ClGemm (if the data type is FP32 or FP16) + * -# @ref CLGEMMLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED) + * -# @ref ClGemmLowpOutputStage with QUANTIZE_DOWN_FIXEDPOINT type of quantization (if the data type is QASYMM8/QASYMM8_SIGNED) + * -# @ref opencl::kernels::ClCol2ImKernel (if NCHW data layout) + * -# @ref opencl::kernels::ClActivationKernel + */ +class ClGemmConv2d : public IClOperator +{ +public: + /** Constructor */ + ClGemmConv2d(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + ClGemmConv2d(const ClGemmConv2d &) = delete; + /** Default move constructor */ + ClGemmConv2d(ClGemmConv2d &&) = default; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + ClGemmConv2d &operator=(const ClGemmConv2d &) = delete; + /** Default move assignment operator */ + ClGemmConv2d &operator=(ClGemmConv2d &&) = default; + /**Default destructor */ + ~ClGemmConv2d(); + /** Set the input and output tensors. + * + * Valid data layouts: + * - NHWC + * - NCHW + * + * Valid data type configurations: + * |src0 |src1 |src2 |dst | + * |:--------------|:------------------|:--------|:--------------| + * |F16 |F16 |F16 |F16 | + * |F32 |F32 |F32 |F32 | + * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | + * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | + * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | + * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | + * + * @param[in] compile_context The compile context to be used. + * @param[in] src Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. + * Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8 or QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8_SIGNED. + * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. + * Data type supported: Should match @p input data type, except for input of quantized type where biases should be of S32 type. + * @param[out] dst Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * Data types supported: Same as @p input. + * @param[in] conv2d_info Contains convolution 2d info described in @ref Conv2dInfo. + * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. If this is not part of the fully connected layer the weights + * tensor has also been transposed with CLGEMMReshapeRHSMatrixKernel. Data type supported: Same as @p input. + */ + void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, const Conv2dInfo &conv2d_info, + const WeightsInfo &weights_info = WeightsInfo()); + /** Static function to check if given info will lead to a valid configuration + * + * Similar to ClGemmConvolution::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const Conv2dInfo &conv2d_info, + const WeightsInfo &weights_info = WeightsInfo()); + + // Inherited methods overridden: + void run(ITensorPack &tensors) override; + void prepare(ITensorPack &constants) override; + experimental::MemoryRequirements workspace() const override; + +private: + /** Configures the appropriate matrix multiply routine + * + * @param[in] compile_context The compile context to be used. + * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[in] weights Weights tensor info. Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8 or + * QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8_SIGNED. + * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. + * Data type supported: Should match @p input data type, except for input of quantized type where biases should be of S32 type. + * @param[in, out] dst Output tensor info. Data types supported: same as @p input. + * @param[in] gemmlowp_output_stage GEMMLowp output stage info + * @param[in] gemm_3d_depth Depth of GEMM 3D + * @param[in] act_info Activation to apply after the matrix multiplication + */ + void configure_mm(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, + const GEMMLowpOutputStageInfo &gemmlowp_output_stage, + int gemm_3d_depth, const ActivationLayerInfo &act_info); + /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMConvolutionLayer matrix multiply routines + * + * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[in] weights Weights tensor info. Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8 or + * QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8_SIGNED. + * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. + * Data type supported: Should match @p input data type, except for input of quantized type where biases should be of S32 type. + * @param[in] dst Output tensor info. Data types supported: same as @p input. + * @param[in] gemmlowp_output_stage GEMMLowp output stage info + * @param[in] gemm_3d_depth Depth of GEMM 3D + * @param[in] skip_im2col Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout. + * @param[in] act_info Activation to apply after the matrix multiplication + * + * @return a status + */ + static Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const GEMMLowpOutputStageInfo &gemmlowp_output_stage, + int gemm_3d_depth, bool skip_im2col, const ActivationLayerInfo &act_info); + + enum AuxTensorIdx + { + // ClGemmLowpMatrixMultiplyCore has up to 7 internal tensors + Im2ColOutput = 8, + WeightsReshaped, + GemmOutput, + Count + }; + + std::unique_ptr _weights_reshape_kernel; + std::unique_ptr _im2col_kernel; + std::unique_ptr _mm_gemm; + std::unique_ptr _mm_gemmlowp; + std::unique_ptr _col2im_kernel; + std::unique_ptr _activation_kernel; + + TensorInfo _im2col_output; + TensorInfo _weights_reshaped; + TensorInfo _gemm_output; + + bool _skip_im2col; + bool _skip_col2im; + bool _is_quantized; + bool _fuse_activation; + bool _append_bias; + bool _is_prepared; + + experimental::MemoryRequirements _aux_mem; +}; +} // namespace opencl +} // namespace arm_compute +#endif /* ARM_COMPUTE_CL_GEMM_CONV2D_H */ diff --git a/src/runtime/gpu/cl/operators/ClGemmConvolution.cpp b/src/runtime/gpu/cl/operators/ClGemmConvolution.cpp deleted file mode 100644 index 1926cbbe4d..0000000000 --- a/src/runtime/gpu/cl/operators/ClGemmConvolution.cpp +++ /dev/null @@ -1,628 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/runtime/gpu/cl/operators/ClGemmConvolution.h" - -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/PixelValue.h" -#include "arm_compute/core/Size2D.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/core/utils/quantization/AsymmHelpers.h" -#include "arm_compute/runtime/CL/CLScheduler.h" -#include "src/core/gpu/cl/kernels/ClActivationKernel.h" -#include "src/core/gpu/cl/kernels/ClCol2ImKernel.h" -#include "src/core/gpu/cl/kernels/ClIm2ColKernel.h" -#include "src/core/gpu/cl/kernels/ClWeightsReshapeKernel.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/MemoryHelpers.h" -#include "src/runtime/gpu/cl/operators/ClGemm.h" -#include "src/runtime/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.h" -#include "src/runtime/gpu/cl/utils/ClAuxTensorHandler.h" -#include "support/Cast.h" - -namespace arm_compute -{ -using namespace experimental; -using namespace misc::shape_calculator; -using namespace utils::cast; -namespace opencl -{ -ClGemmConvolution::ClGemmConvolution() - : _weights_reshape_kernel(nullptr), _im2col_kernel(nullptr), _mm_gemm(nullptr), _mm_gemmlowp(nullptr), _col2im_kernel(nullptr), _activation_kernel(nullptr), _im2col_output(), _weights_reshaped(), - _gemm_output(), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _fuse_activation(true), _append_bias(false), _is_prepared(false), _aux_mem(AuxTensorIdx::Count) -{ -} -ClGemmConvolution::~ClGemmConvolution() = default; - -void ClGemmConvolution::configure_mm(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, - const GEMMLowpOutputStageInfo &gemmlowp_output_stage, - int gemm_3d_depth, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights); - ARM_COMPUTE_ERROR_THROW_ON(validate_mm(src, weights, biases, dst, gemmlowp_output_stage, gemm_3d_depth, _skip_im2col, act_info)); - - const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped - false, // is_b_reshaped - true, // reshape_b_only_on_first_run - gemm_3d_depth, // depth_output_gemm3d - _skip_im2col, // reinterpret_input_as_3d - false, // retain_internal_weights - gemmlowp_output_stage, // gemmlowp_output_stage - false, // fast_math - false, // fp_mixed_precision - true, // broadcast_bias - act_info); // activation_info - - TensorInfo tmp_src{ *src }; - if(_is_quantized) - { - // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() - // Extract and negate input and weights offset - const QuantizationInfo input_quantization_info = src->quantization_info(); - const QuantizationInfo weights_quantization_info = weights->quantization_info(); - - tmp_src.set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset)); - weights->set_quantization_info(QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset)); - - _mm_gemmlowp = std::make_unique(); - _mm_gemmlowp->configure(compile_context, &tmp_src, weights, biases, dst, gemm_info); - - // Revert back QuantizatioInfo as weights could be used in other convolution layers - weights->set_quantization_info(weights_quantization_info); - - auto mm_mem_req = _mm_gemmlowp->workspace(); - for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) - { - _aux_mem[cont] = mm_mem_req[cont]; - } - } - else - { - // Configure matrix multiply function - _mm_gemm = std::make_unique(); - _mm_gemm->configure(compile_context, &tmp_src, weights, biases, dst, 1.0f, 1.0f, gemm_info); - auto mm_mem_req = _mm_gemm->workspace(); - for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) - { - _aux_mem[cont] = mm_mem_req[cont]; - } - } -} - -Status ClGemmConvolution::validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, - const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth, bool skip_im2col, const ActivationLayerInfo &act_info) -{ - const bool is_quantized = is_data_type_quantized_asymmetric(src->data_type()); - - const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped - false, // is_b_reshaped - true, // reshape_b_only_on_first_run - gemm_3d_depth, // depth_output_gemm3d - skip_im2col, // reinterpret_input_as_3d - false, // retain_internal_weights - gemmlowp_output_stage, // gemmlowp_output_stage - false, // fast_math - false, // fp_mixed_precision - true, // broadcast_bias - act_info); // activation_info - - if(is_quantized) - { - // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() - // Extract and negate input and weights offset - const QuantizationInfo input_quantization_info = src->quantization_info(); - const QuantizationInfo weights_quantization_info = weights->quantization_info(); - - std::unique_ptr src_qa = src->clone(); - std::unique_ptr weights_qa = weights->clone(); - src_qa->set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset)); - weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset)); - - // Perform validation step on GEMMLowp - return ClGemmLowpMatrixMultiplyCore::validate(src_qa.get(), weights_qa.get(), biases, dst, gemm_info); - } - else - { - // Perform validation step on Matrix multiply function - return ClGemm::validate(src, weights, biases, dst, 1.0f, 1.0f, gemm_info); - } -} - -void ClGemmConvolution::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, - const Conv2dInfo &conv2d_info, const WeightsInfo &weights_info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); - - ARM_COMPUTE_ERROR_THROW_ON(ClGemmConvolution::validate(src, weights, biases, dst, - conv2d_info, - weights_info)); - - const DataType data_type = src->data_type(); - const DataLayout data_layout = src->data_layout(); - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); - - const unsigned int kernel_width = weights->dimension(idx_width); - const unsigned int kernel_height = weights->dimension(idx_height); - const unsigned int num_kernels = weights->dimension(idx_kernels); - - const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); - - _is_prepared = weights_info.retain_internal_weights(); - _is_quantized = is_data_type_quantized_asymmetric(src->data_type()); - _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv2d_info.conv_info.stride().first == 1 && conv2d_info.conv_info.stride().second == 1); - _skip_col2im = data_layout == DataLayout::NHWC; - - // Only for quantize there are few cases where we cannot fuse the activation function in GEMM - _fuse_activation = true; - - const ITensorInfo *gemm_input_to_use = src; - ITensorInfo *gemm_output_to_use = dst; - - // Get parameters from conv_info - unsigned int stride_x = 0; - unsigned int stride_y = 0; - std::tie(stride_x, stride_y) = conv2d_info.conv_info.stride(); - - // Get convolved dimensions - unsigned int conv_w = 0; - unsigned int conv_h = 0; - std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), - src->dimension(idx_height), - kernel_width, - kernel_height, - conv2d_info.conv_info, - conv2d_info.dilation); - - unsigned int mat_weights_cols = num_kernels / conv2d_info.num_groups; - - ITensorInfo *biases_to_use = biases; - _append_bias = false; - - _weights_reshape_kernel = std::make_unique(); - if(conv2d_info.num_groups != 1 && biases != nullptr) - { - // num_groups != 1 can only be for NCHW - // Since it is missing an utility function to reshape the biases, we append the biases into the weights tensor - biases_to_use = nullptr; - _append_bias = true; - _weights_reshape_kernel->configure(compile_context, weights, biases, &_weights_reshaped, conv2d_info.num_groups); - } - else - { - _weights_reshape_kernel->configure(compile_context, weights, nullptr, &_weights_reshaped, conv2d_info.num_groups); - } - - // Create tensor to store im2col reshaped inputs - if(!_skip_im2col) - { - // Configure and tune im2col. im2col output shape is auto-initialized - _im2col_kernel = std::make_unique(); - - // Set the GPU target for im2col - _im2col_kernel->set_target(CLScheduler::get().target()); - _im2col_kernel->configure(compile_context, src, &_im2col_output, Size2D(kernel_width, kernel_height), conv2d_info.conv_info, _append_bias, conv2d_info.dilation, conv2d_info.num_groups); - - // Set quantization info - _im2col_output.set_quantization_info(src->quantization_info()); - CLScheduler::get().tune_kernel_static(*_im2col_kernel); - - // Update GEMM input - gemm_input_to_use = &_im2col_output; - } - - // Create GEMM output tensor - if(!_skip_col2im) - { - TensorShape shape_gemm; - - // If we cannot skip col2im it means we run im2col as well - shape_gemm = _im2col_output.tensor_shape(); - shape_gemm.set(0, mat_weights_cols); - shape_gemm.set(1, conv_w * conv_h); - - _gemm_output = TensorInfo(shape_gemm, 1, data_type); - _gemm_output.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); - - // Update GEMM output - gemm_output_to_use = &_gemm_output; - } - - GEMMLowpOutputStageInfo gemmlowp_output_stage; - gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; - gemmlowp_output_stage.gemmlowp_offset = 0; - - // Configure output stage for quantized case - if(_is_quantized) - { - const auto output_quant_info = (dst->total_size() == 0) ? iq_info : oq_info; - const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->data_type()); - const unsigned int num_filters = (is_quantized_per_channel) ? num_kernels : 1; - - gemmlowp_output_stage.is_quantized_per_channel = is_quantized_per_channel; - - gemmlowp_output_stage.gemmlowp_multipliers.resize(num_filters); - gemmlowp_output_stage.gemmlowp_shifts.resize(num_filters); - quantization::compute_quantized_multipliers_and_shifts(src, weights, dst, - gemmlowp_output_stage.gemmlowp_multipliers.data(), - gemmlowp_output_stage.gemmlowp_shifts.data()); - gemmlowp_output_stage.gemmlowp_multiplier = gemmlowp_output_stage.gemmlowp_multipliers[0]; - gemmlowp_output_stage.gemmlowp_shift = gemmlowp_output_stage.gemmlowp_shifts[0]; - - PixelValue min_val{}; - PixelValue max_val{}; - std::tie(min_val, max_val) = get_min_max(dst->data_type()); - - auto min_activation = min_val.get(); - auto max_activation = max_val.get(); - - const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, - ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, - ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU - }; - - if(conv2d_info.act_info.enabled()) - { - if(supported_acts.count(conv2d_info.act_info.activation()) != 0) - { - std::tie(min_activation, max_activation) = get_quantized_activation_min_max(conv2d_info.act_info, data_type, output_quant_info); - } - else - { - _fuse_activation = false; - } - } - - // Set the GEMMLowp output stage info - gemmlowp_output_stage.gemmlowp_offset = output_quant_info.offset; - gemmlowp_output_stage.gemmlowp_min_bound = min_activation; - gemmlowp_output_stage.gemmlowp_max_bound = max_activation; - } - - // Configure and tune GEMM - // In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix - const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0; - - configure_mm(compile_context, gemm_input_to_use, &_weights_reshaped, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, conv2d_info.act_info); - - if(!_skip_col2im) - { - // Set the GPU target for col2im - _col2im_kernel = std::make_unique(); - _col2im_kernel->set_target(CLScheduler::get().target()); - // Configure and tune Col2Im - _col2im_kernel->configure(compile_context, gemm_output_to_use, dst, Size2D(conv_w, conv_h), conv2d_info.num_groups); - CLScheduler::get().tune_kernel_static(*_col2im_kernel.get()); - } - - ARM_COMPUTE_ERROR_ON_MSG((dst->dimension(idx_width) != conv_w) || (dst->dimension(idx_height) != conv_h), - "Output shape does not match the expected one"); - - if(!_fuse_activation) - { - _activation_kernel = std::make_unique(); - _activation_kernel->configure(compile_context, dst, nullptr, conv2d_info.act_info); - } - - _aux_mem[Im2ColOutput] = MemoryInfo(offset_int_vec(Im2ColOutput), MemoryLifetime::Temporary, _im2col_output.total_size()); - _aux_mem[WeightsReshaped] = MemoryInfo(offset_int_vec(WeightsReshaped), MemoryLifetime::Persistent, _weights_reshaped.total_size()); - _aux_mem[GemmOutput] = MemoryInfo(offset_int_vec(GemmOutput), MemoryLifetime::Temporary, _gemm_output.total_size()); -} - -Status ClGemmConvolution::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const Conv2dInfo &conv2d_info, - const WeightsInfo &weights_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!"); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); - const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->data_type()); - - if(!is_quantized_per_channel) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights); - } - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, weights); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((conv2d_info.num_groups != 1) && (src->data_layout() != DataLayout::NCHW), "Grouping (num_groups != 1) with NHWC data layout is not supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((conv2d_info.num_groups != 1) && (src->data_type() == DataType::QASYMM8), "Grouping (num_groups != 1) is not supported with QASYMM8"); - ARM_COMPUTE_RETURN_ERROR_ON(((src->dimension(2) / weights->dimension(2)) != conv2d_info.num_groups) && (src->data_layout() == DataLayout::NCHW)); - - const DataLayout data_layout = src->data_layout(); - const DataType data_type = src->data_type(); - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); - const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); - - const unsigned int kernel_width = weights->dimension(idx_width); - const unsigned int kernel_height = weights->dimension(idx_height); - const unsigned int num_kernels = weights->dimension(idx_kernels); - - TensorInfo im2col_reshaped_info{}; - TensorInfo info_gemm{}; - TensorInfo weights_reshaped_info{}; - const ITensorInfo *gemm_input_to_use = src; - const ITensorInfo *gemm_output_to_use = dst; - const ITensorInfo *weights_to_use = weights; - const bool is_quantized = is_data_type_quantized_asymmetric(data_type); - const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv2d_info.conv_info.stride().first == 1 - && conv2d_info.conv_info.stride().second == 1); - const bool skip_col2im = data_layout == DataLayout::NHWC; - bool fuse_activation = true; - - ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(idx_channel) * conv2d_info.num_groups) != src->dimension(idx_channel)); - ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); - - // Validate biases - if(biases != nullptr) - { - if(is_quantized) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); - } - else - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases); - } - ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(idx_kernels)); - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - } - - if(conv2d_info.act_info.enabled()) - { - ARM_COMPUTE_ERROR_ON(conv2d_info.act_info.b() > conv2d_info.act_info.a()); - } - - // Get convolved dimensions - unsigned int conv_w = 0; - unsigned int conv_h = 0; - - std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), - src->dimension(idx_height), - kernel_width, - kernel_height, - conv2d_info.conv_info, - conv2d_info.dilation); - - unsigned int mat_weights_cols = num_kernels / conv2d_info.num_groups; - - const ITensorInfo *biases_to_use = biases; - bool append_bias = false; - - if(conv2d_info.num_groups != 1 && biases != nullptr) - { - // num_groups != 1 can only be for NCHW - // Since it is missing an utility function to reshape the biases, we append the biases into the weights tensor - biases_to_use = nullptr; - append_bias = true; - weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, true, conv2d_info.num_groups), 1, data_type); - } - else - { - weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, false, conv2d_info.num_groups), 1, data_type); - } - - weights_to_use = &weights_reshaped_info; - - if(!skip_im2col) - { - const Size2D kernel_dims(kernel_width, kernel_height); - - // Output tensor auto initialization if not yet initialized - TensorShape expected_output_shape = compute_im2col_conv_shape(src, kernel_dims, conv2d_info.conv_info, append_bias, conv2d_info.dilation, conv2d_info.num_groups == 1, conv2d_info.num_groups); - - auto_init_if_empty(im2col_reshaped_info, src->clone()->set_tensor_shape(expected_output_shape)); - - ARM_COMPUTE_RETURN_ON_ERROR(opencl::kernels::ClIm2ColKernel::validate(src, &im2col_reshaped_info, kernel_dims, conv2d_info.conv_info, append_bias, conv2d_info.dilation, conv2d_info.num_groups)); - gemm_input_to_use = &im2col_reshaped_info; - } - - // Create GEMM output tensor - if(!skip_col2im) - { - TensorShape shape_gemm; - - shape_gemm = gemm_input_to_use->tensor_shape(); - shape_gemm.set(0, mat_weights_cols); - shape_gemm.set(1, conv_w * conv_h); - - info_gemm = TensorInfo(shape_gemm, 1, data_type); - info_gemm.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout()); - gemm_output_to_use = &info_gemm; - } - - GEMMLowpOutputStageInfo gemmlowp_output_stage; - gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; - gemmlowp_output_stage.gemmlowp_offset = 0; - gemmlowp_output_stage.is_quantized_per_channel = is_quantized_per_channel; - - if(is_quantized) - { - const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); - const auto output_quant_info = (dst->total_size() == 0) ? iq_info : oq_info; - const unsigned int num_filters = (is_quantized_per_channel) ? num_kernels : 1; - - gemmlowp_output_stage.gemmlowp_multipliers.resize(num_filters); - gemmlowp_output_stage.gemmlowp_shifts.resize(num_filters); - quantization::compute_quantized_multipliers_and_shifts(src, weights, dst, - gemmlowp_output_stage.gemmlowp_multipliers.data(), - gemmlowp_output_stage.gemmlowp_shifts.data()); - gemmlowp_output_stage.gemmlowp_multiplier = gemmlowp_output_stage.gemmlowp_multipliers[0]; - gemmlowp_output_stage.gemmlowp_shift = gemmlowp_output_stage.gemmlowp_shifts[0]; - - int min_activation = 0; - int max_activation = 0; - - const std::set supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, - ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, - ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU - }; - - if(conv2d_info.act_info.enabled()) - { - if(supported_acts.count(conv2d_info.act_info.activation()) != 0) - { - std::tie(min_activation, max_activation) = get_quantized_activation_min_max(conv2d_info.act_info, data_type, output_quant_info); - } - else - { - fuse_activation = false; - } - } - - // Set the GEMMLowp output stage info - gemmlowp_output_stage.gemmlowp_offset = output_quant_info.offset; - gemmlowp_output_stage.gemmlowp_min_bound = min_activation; - gemmlowp_output_stage.gemmlowp_max_bound = max_activation; - } - - // In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix - const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0; - - ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, skip_im2col, conv2d_info.act_info)); - - // Validate Col2Im - if(!skip_col2im) - { - ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClCol2ImKernel::validate(gemm_output_to_use, dst, Size2D(conv_w, conv_h), conv2d_info.num_groups)); - } - - //Validate Activation Layer - if(!fuse_activation) - { - ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClActivationKernel::validate(dst, nullptr, conv2d_info.act_info)); - } - - return Status{}; -} - -void ClGemmConvolution::run(ITensorPack &tensors) -{ - prepare(tensors); - - auto src = tensors.get_const_tensor(ACL_SRC_0); - auto biases = tensors.get_const_tensor(ACL_SRC_2); - auto dst = tensors.get_tensor(ACL_DST); - auto gemm_input_to_use = src; - auto gemm_output_to_use = dst; - - CLAuxTensorHandler im2col_output(offset_int_vec(Im2ColOutput), _im2col_output, tensors, false); - CLAuxTensorHandler gemm_output(offset_int_vec(GemmOutput), _gemm_output, tensors, false); - CLAuxTensorHandler weights_reshaped(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors, false); - - // Run im2col - if(!_skip_im2col) - { - ITensorPack pack = - { - { TensorType::ACL_SRC, src }, - { TensorType::ACL_DST, im2col_output.get() } - }; - CLScheduler::get().enqueue_op(*_im2col_kernel, pack, false); - gemm_input_to_use = im2col_output.get(); - } - if(!_skip_col2im) - { - gemm_output_to_use = gemm_output.get(); - } - ITensorPack pack_mm = tensors; - pack_mm.add_const_tensor(TensorType::ACL_SRC_0, gemm_input_to_use); - pack_mm.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get()); - if(!_append_bias) - { - pack_mm.add_const_tensor(TensorType::ACL_SRC_2, biases); - } - pack_mm.add_tensor(TensorType::ACL_DST, gemm_output_to_use); - // Runs ClGemm or ClGemmLowpMatrixMultiplyCore functions - if(_is_quantized) - { - // Run gemmlowp - _mm_gemmlowp->run(pack_mm); - } - else - { - // Run gemm - _mm_gemm->run(pack_mm); - } - - // Reshape output matrix - if(!_skip_col2im) - { - ITensorPack pack = - { - { TensorType::ACL_SRC, gemm_output_to_use }, - { TensorType::ACL_DST, dst } - }; - CLScheduler::get().enqueue_op(*_col2im_kernel.get(), pack, false); - } - - //Run Activation Layer if we cannot fuse in GEMM - if(!_fuse_activation) - { - ITensorPack pack = - { - { TensorType::ACL_SRC, dst }, - { TensorType::ACL_DST, dst } - }; - CLScheduler::get().enqueue_op(*_activation_kernel.get(), pack, false); - } -} - -void ClGemmConvolution::prepare(ITensorPack &tensors) -{ - if(!_is_prepared) - { - // Run weights reshaping and mark original weights tensor as unused - ICLTensor *weights_reshaped_p = utils::cast::polymorphic_downcast(tensors.get_tensor(offset_int_vec(WeightsReshaped))); - CLAuxTensorHandler weights_reshaped(_weights_reshaped, *weights_reshaped_p); - auto weights = tensors.get_const_tensor(TensorType::ACL_SRC_1); - ITensorPack pack = - { - { TensorType::ACL_SRC, weights }, - { TensorType::ACL_DST, weights_reshaped.get() } - }; - - if(_append_bias) - { - const auto biases = tensors.get_const_tensor(TensorType::ACL_SRC_2); - pack.add_const_tensor(TensorType::ACL_BIAS, biases); - } - CLScheduler::get().enqueue_op(*_weights_reshape_kernel.get(), pack, true); - tensors.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get()); - - // Prepare GEMM - _is_quantized ? _mm_gemmlowp->prepare(tensors) : _mm_gemm->prepare(tensors); - _is_prepared = true; - } -} -experimental::MemoryRequirements ClGemmConvolution::workspace() const -{ - return _aux_mem; -} -} // namespace opencl -} // namespace arm_compute diff --git a/src/runtime/gpu/cl/operators/ClGemmConvolution.h b/src/runtime/gpu/cl/operators/ClGemmConvolution.h deleted file mode 100644 index 444516eaaa..0000000000 --- a/src/runtime/gpu/cl/operators/ClGemmConvolution.h +++ /dev/null @@ -1,185 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_GEMMCONVOLUTION_H -#define ARM_COMPUTE_CL_GEMMCONVOLUTION_H - -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/runtime/FunctionDescriptors.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/runtime/gpu/cl/IClOperator.h" - -#include - -namespace arm_compute -{ -namespace opencl -{ -class ClGemm; -class ClGemmLowpMatrixMultiplyCore; -namespace kernels -{ -class ClIm2ColKernel; -class ClCol2ImKernel; -class ClWeightsReshapeKernel; -class ClActivationKernel; -} // namespace kernels - -/** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions: - * - * -# @ref opencl::kernels::ClIm2ColKernel - * -# @ref ClGemm (if the data type is FP32 or FP16) - * -# @ref CLGEMMLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED) - * -# @ref ClGemmLowpOutputStage with QUANTIZE_DOWN_FIXEDPOINT type of quantization (if the data type is QASYMM8/QASYMM8_SIGNED) - * -# @ref opencl::kernels::ClCol2ImKernel (if NCHW data layout) - * -# @ref opencl::kernels::ClActivationKernel - */ -class ClGemmConvolution : public IClOperator -{ -public: - /** Constructor */ - ClGemmConvolution(); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - ClGemmConvolution(const ClGemmConvolution &) = delete; - /** Default move constructor */ - ClGemmConvolution(ClGemmConvolution &&) = default; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - ClGemmConvolution &operator=(const ClGemmConvolution &) = delete; - /** Default move assignment operator */ - ClGemmConvolution &operator=(ClGemmConvolution &&) = default; - /**Default destructor */ - ~ClGemmConvolution(); - /** Set the input and output tensors. - * - * Valid data layouts: - * - NHWC - * - NCHW - * - * Valid data type configurations: - * |src0 |src1 |src2 |dst | - * |:--------------|:------------------|:--------|:--------------| - * |F16 |F16 |F16 |F16 | - * |F32 |F32 |F32 |F32 | - * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | - * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | - * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | - * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], - * while every optional dimension from 4 and above represent a batch of inputs. - * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. - * Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8 or QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8_SIGNED. - * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. - * Data type supported: Should match @p input data type, except for input of quantized type where biases should be of S32 type. - * @param[out] dst Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. - * Data types supported: Same as @p input. - * @param[in] conv2d_info Contains convolution 2d info described in @ref Conv2dInfo. - * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. If this is not part of the fully connected layer the weights - * tensor has also been transposed with CLGEMMReshapeRHSMatrixKernel. Data type supported: Same as @p input. - */ - void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, const Conv2dInfo &conv2d_info, - const WeightsInfo &weights_info = WeightsInfo()); - /** Static function to check if given info will lead to a valid configuration - * - * Similar to ClGemmConvolution::configure() - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const Conv2dInfo &conv2d_info, - const WeightsInfo &weights_info = WeightsInfo()); - - // Inherited methods overridden: - void run(ITensorPack &tensors) override; - void prepare(ITensorPack &constants) override; - experimental::MemoryRequirements workspace() const override; - -private: - /** Configures the appropriate matrix multiply routine - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[in] weights Weights tensor info. Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8 or - * QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8_SIGNED. - * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. - * Data type supported: Should match @p input data type, except for input of quantized type where biases should be of S32 type. - * @param[in, out] dst Output tensor info. Data types supported: same as @p input. - * @param[in] gemmlowp_output_stage GEMMLowp output stage info - * @param[in] gemm_3d_depth Depth of GEMM 3D - * @param[in] act_info Activation to apply after the matrix multiplication - */ - void configure_mm(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, - const GEMMLowpOutputStageInfo &gemmlowp_output_stage, - int gemm_3d_depth, const ActivationLayerInfo &act_info); - /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMConvolutionLayer matrix multiply routines - * - * @param[in] src Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[in] weights Weights tensor info. Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8 or - * QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8_SIGNED. - * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. - * Data type supported: Should match @p input data type, except for input of quantized type where biases should be of S32 type. - * @param[in] dst Output tensor info. Data types supported: same as @p input. - * @param[in] gemmlowp_output_stage GEMMLowp output stage info - * @param[in] gemm_3d_depth Depth of GEMM 3D - * @param[in] skip_im2col Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout. - * @param[in] act_info Activation to apply after the matrix multiplication - * - * @return a status - */ - static Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const GEMMLowpOutputStageInfo &gemmlowp_output_stage, - int gemm_3d_depth, bool skip_im2col, const ActivationLayerInfo &act_info); - - enum AuxTensorIdx - { - // ClGemmLowpMatrixMultiplyCore has up to 7 internal tensors - Im2ColOutput = 8, - WeightsReshaped, - GemmOutput, - Count - }; - - std::unique_ptr _weights_reshape_kernel; - std::unique_ptr _im2col_kernel; - std::unique_ptr _mm_gemm; - std::unique_ptr _mm_gemmlowp; - std::unique_ptr _col2im_kernel; - std::unique_ptr _activation_kernel; - - TensorInfo _im2col_output; - TensorInfo _weights_reshaped; - TensorInfo _gemm_output; - - bool _skip_im2col; - bool _skip_col2im; - bool _is_quantized; - bool _fuse_activation; - bool _append_bias; - bool _is_prepared; - - experimental::MemoryRequirements _aux_mem; -}; -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_GEMMCONVOLUTION_H */ diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp index 4332db605d..2178b9b209 100644 --- a/tests/validation/NEON/ConvolutionLayer.cpp +++ b/tests/validation/NEON/ConvolutionLayer.cpp @@ -29,7 +29,7 @@ #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" #include "src/core/helpers/MemoryHelpers.h" -#include "src/runtime/cpu/operators/CpuGemmConvolution.h" +#include "src/runtime/cpu/operators/CpuGemmConv2d.h" #include "src/runtime/cpu/operators/CpuGemmDirectConv2d.h" #include "src/runtime/cpu/operators/CpuWinogradConv2d.h" #include "tests/NEON/Accessor.h" @@ -510,7 +510,7 @@ using NEGEMMConvolutionLayerFixture = ConvolutionValidationFixture using NEGEMMConvolutionLayerMixedDataLayoutFixture = ConvolutionValidationFixture; -/** Test case for memory injection in @ref cpu::CpuGemmConvolution. +/** Test case for memory injection in @ref cpu::CpuGemmConv2d. * * Configure the operator once and inject memory at run-time in multiple executions. * @@ -519,7 +519,7 @@ using NEGEMMConvolutionLayerMixedDataLayoutFixture = ConvolutionValidationFixtur */ TEST_CASE(MemoryInjection, framework::DatasetMode::ALL) { - auto conv = std::make_unique(); + auto conv = std::make_unique(); const auto src_info = TensorInfo(TensorShape(1U, 5U, 2U), 1, DataType::F32, DataLayout::NCHW); const auto weight_info = TensorInfo(TensorShape(1U, 3U, 2U, 3U), 1, DataType::F32, DataLayout::NCHW); const auto bias_info = TensorInfo(TensorShape(3U), 1, DataType::F32, DataLayout::NCHW); -- cgit v1.2.1