diff options
Diffstat (limited to 'src/cpu/operators/CpuGemmConv2d.cpp')
| -rw-r--r-- | src/cpu/operators/CpuGemmConv2d.cpp | 378 |
1 files changed, 212 insertions, 166 deletions
diff --git a/src/cpu/operators/CpuGemmConv2d.cpp b/src/cpu/operators/CpuGemmConv2d.cpp index 39b410d609..7c59d88c61 100644 --- a/src/cpu/operators/CpuGemmConv2d.cpp +++ b/src/cpu/operators/CpuGemmConv2d.cpp @@ -26,9 +26,9 @@ #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/core/Validate.h" #include "arm_compute/runtime/NEON/NEScheduler.h" #include "src/common/utils/Log.h" @@ -52,8 +52,11 @@ namespace arm_compute { namespace cpu { -CpuGemmConv2d::SkipInfo CpuGemmConv2d::skip_im_col_info(const ITensorInfo *src, const ITensorInfo *weights, const PadStrideInfo &conv_info, - const Size2D &dilation, const ActivationLayerInfo &act_info) +CpuGemmConv2d::SkipInfo CpuGemmConv2d::skip_im_col_info(const ITensorInfo *src, + const ITensorInfo *weights, + const PadStrideInfo &conv_info, + const Size2D &dilation, + const ActivationLayerInfo &act_info) { const DataLayout data_layout = src->data_layout(); const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); @@ -62,63 +65,86 @@ CpuGemmConv2d::SkipInfo CpuGemmConv2d::skip_im_col_info(const ITensorInfo *src, const unsigned int kernel_height = weights->dimension(idx_height); 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); - const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1); - - if(skip_im2col) + std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), src->dimension(idx_height), kernel_width, + kernel_height, conv_info, dilation); + const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && + conv_info.stride().first == 1 && conv_info.stride().second == 1); + + if (skip_im2col) { - const bool skip_col2im = (data_layout == DataLayout::NHWC && (bool(CpuGemmConv2d::validate_gemm3d(src, weights, act_info, conv_h, /*skip_im2col*/ true)))); - if(skip_col2im) + const bool skip_col2im = + (data_layout == DataLayout::NHWC && + (bool(CpuGemmConv2d::validate_gemm3d(src, weights, act_info, conv_h, /*skip_im2col*/ true)))); + if (skip_col2im) { - return { true, true }; + return {true, true}; } } else { - const bool skip_col2im = (data_layout == DataLayout::NHWC && (bool(CpuGemmConv2d::validate_gemm3d(src, weights, act_info, conv_h, /*skip_im2col*/ false)))); - if(skip_col2im) + const bool skip_col2im = + (data_layout == DataLayout::NHWC && + (bool(CpuGemmConv2d::validate_gemm3d(src, weights, act_info, conv_h, /*skip_im2col*/ false)))); + if (skip_col2im) { - return { false, true }; + return {false, true}; } } // Default case when we cannot reinterpret the input and output as 3D. - return { false, false }; + return {false, false}; } CpuGemmConv2d::CpuGemmConv2d() - : _weights_reshape_kernel(nullptr), _im2col_kernel(), _mm_gemm(), _mm_gemmlowp(), _col2im_kernel(), _reshape(), _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) + : _weights_reshape_kernel(nullptr), + _im2col_kernel(), + _mm_gemm(), + _mm_gemmlowp(), + _col2im_kernel(), + _reshape(), + _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, bool fixed_format, arm_compute::WeightFormat weight_format) +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, + bool fixed_format, + arm_compute::WeightFormat weight_format) { 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, fixed_format, weight_format)); + ARM_COMPUTE_ERROR_THROW_ON(validate_mm(src, weights, biases, dst, act_info, enable_fast_math, gemm_3d_depth, + _skip_im2col, fixed_format, weight_format)); // 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, fixed_format, weight_format); + 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, fixed_format, weight_format); // Supported activations in GEMM - const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, - ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, - ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU - }; + const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { + ActivationLayerInfo::ActivationFunction::RELU, ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, + ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU}; - if(_is_quantized) + if (_is_quantized) { - TensorInfo tmp_src{ *src }; - TensorInfo tmp_weights{ *weights }; + 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(); @@ -129,7 +155,7 @@ void CpuGemmConv2d::configure_mm(const ITensorInfo *src, const ITensorInfo *weig 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())) + 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)); @@ -142,7 +168,7 @@ void CpuGemmConv2d::configure_mm(const ITensorInfo *src, const ITensorInfo *weig int32_t min_activation = type_min.get<int32_t>(); int32_t max_activation = type_max.get<int32_t>(); - if(supported_acts.count(act_info.activation()) != 0) + if (supported_acts.count(act_info.activation()) != 0) { std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act_info, data_type, uoqinfo); } @@ -156,11 +182,12 @@ void CpuGemmConv2d::configure_mm(const ITensorInfo *src, const ITensorInfo *weig quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info); _mm_gemmlowp = std::make_unique<CpuGemmLowpMatrixMultiplyCore>(); - _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, fixed_format, - weight_format)); + _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, fixed_format, weight_format)); auto mm_mem_req = _mm_gemmlowp->workspace(); - for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) + for (unsigned int cont = 0; cont < mm_mem_req.size(); ++cont) { _aux_mem[cont] = mm_mem_req[cont]; } @@ -171,26 +198,35 @@ void CpuGemmConv2d::configure_mm(const ITensorInfo *src, const ITensorInfo *weig _mm_gemm = std::make_unique<CpuGemm>(); _mm_gemm->configure(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) + 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, bool fixed_format, arm_compute::WeightFormat weight_format) +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, + bool fixed_format, + arm_compute::WeightFormat weight_format) { 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, fixed_format, weight_format); + 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, fixed_format, weight_format); - if(is_quantized) + if (is_quantized) { // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() // Extract and negate input and weights offset @@ -206,11 +242,10 @@ Status CpuGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *wei int32_t min_activation = type_min.get<int32_t>(); int32_t max_activation = type_max.get<int32_t>(); - const std::set<ActivationLayerInfo::ActivationFunction> 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) + const std::set<ActivationLayerInfo::ActivationFunction> 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); } @@ -229,8 +264,9 @@ Status CpuGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *wei 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)); + 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 { @@ -239,36 +275,44 @@ Status CpuGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *wei } } -Status CpuGemmConv2d::validate_gemm3d(const ITensorInfo *input_info, const ITensorInfo *weights_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col) +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_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()); + 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); + 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) +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)); - ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, conv_info, weights_info, dilation, act_info, enable_fast_math, num_groups); + ARM_COMPUTE_ERROR_THROW_ON(CpuGemmConv2d::validate(src, weights, biases, dst, conv_info, weights_info, dilation, + act_info, enable_fast_math, num_groups)); + ARM_COMPUTE_LOG_PARAMS(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(); @@ -283,7 +327,8 @@ void CpuGemmConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights _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); + _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; @@ -291,20 +336,17 @@ void CpuGemmConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights // 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); + 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 - const CpuGemmConv2d::SkipInfo skip_info = CpuGemmConv2d::skip_im_col_info(src, weights, conv_info, dilation, act_info); - _skip_im2col = skip_info.skip_im2col; - _skip_col2im = skip_info.skip_col2im; + const CpuGemmConv2d::SkipInfo skip_info = + CpuGemmConv2d::skip_im_col_info(src, weights, conv_info, dilation, act_info); + _skip_im2col = skip_info.skip_im2col; + _skip_col2im = skip_info.skip_col2im; // Get parameters from conv_info unsigned int stride_x = 0; @@ -320,17 +362,19 @@ void CpuGemmConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights _weights_reshaped.set_quantization_info(weights->quantization_info()); // Create tensor to store im2col reshaped inputs - if(!_skip_im2col) + if (!_skip_im2col) { const int block_by = arm_compute::block_by(weights_info.weight_format()); unsigned int input_pad_right = 0; - if(block_by > 1) + if (block_by > 1) { - input_pad_right = (src->dimension(idx_channel) % block_by) == 0 ? 0 : block_by - (src->dimension(idx_channel) % block_by); + input_pad_right = + (src->dimension(idx_channel) % block_by) == 0 ? 0 : block_by - (src->dimension(idx_channel) % block_by); } // Configure _im2col_kernel = std::make_unique<kernels::CpuIm2ColKernel>(); - _im2col_kernel->configure(src, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, false, dilation, num_groups, input_pad_right); + _im2col_kernel->configure(src, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, false, dilation, + num_groups, input_pad_right); // Update GEMM input gemm_input_to_use = &_im2col_output; @@ -338,7 +382,7 @@ void CpuGemmConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights // 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) + if (!_skip_col2im) { TensorShape shape_gemm; @@ -368,9 +412,10 @@ void CpuGemmConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights // 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; const bool fixed_format = weights_info.weight_format() != arm_compute::WeightFormat::UNSPECIFIED; - configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, act_info, enable_fast_math, gemm_3d_depth, fixed_format, weights_info.weight_format()); + configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, act_info, enable_fast_math, + gemm_3d_depth, fixed_format, weights_info.weight_format()); - if(!_skip_col2im && _data_layout == DataLayout::NCHW) + if (!_skip_col2im && _data_layout == DataLayout::NCHW) { // Configure col2im _col2im_kernel = std::make_unique<kernels::CpuCol2ImKernel>(); @@ -390,14 +435,24 @@ void CpuGemmConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights 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()); + _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::has_opt_impl(arm_compute::WeightFormat &expected_weight_format, 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, const bool enable_fast_math) +Status CpuGemmConv2d::has_opt_impl(arm_compute::WeightFormat &expected_weight_format, + 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, + const bool enable_fast_math) { const DataLayout data_layout = src->data_layout(); const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); @@ -406,36 +461,44 @@ Status CpuGemmConv2d::has_opt_impl(arm_compute::WeightFormat &expected_weight_fo const unsigned int kernel_height = weights->dimension(idx_height); 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); + std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width), src->dimension(idx_height), kernel_width, + kernel_height, conv_info, dilation); - const CpuGemmConv2d::SkipInfo skip_info = CpuGemmConv2d::skip_im_col_info(src, weights, conv_info, - dilation, act_info); + const CpuGemmConv2d::SkipInfo skip_info = + CpuGemmConv2d::skip_im_col_info(src, weights, conv_info, dilation, act_info); const bool skip_im2col = skip_info.skip_im2col; const bool skip_col2im = skip_info.skip_col2im; const unsigned int gemm_3d_depth = skip_col2im ? conv_h : 0; const bool fixed_format = weights_info.weight_format() != arm_compute::WeightFormat::UNSPECIFIED; - 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, fixed_format, weights_info.weight_format()); + 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, fixed_format, weights_info.weight_format()); return CpuGemm::has_opt_impl(expected_weight_format, src, weights, biases, dst, gemm_info); } -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) +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_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); - if(!is_fixed_format(weights_info.weight_format())) + if (!is_fixed_format(weights_info.weight_format())) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, weights); } @@ -468,29 +531,25 @@ Status CpuGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weight 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); + 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 - const CpuGemmConv2d::SkipInfo skip_info = CpuGemmConv2d::skip_im_col_info(src, weights, conv_info, - dilation, act_info); - const bool skip_im2col = skip_info.skip_im2col, skip_col2im = skip_info.skip_col2im; + const CpuGemmConv2d::SkipInfo skip_info = + CpuGemmConv2d::skip_im_col_info(src, weights, conv_info, dilation, act_info); + const bool skip_im2col = skip_info.skip_im2col, skip_col2im = skip_info.skip_col2im; 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 (biases != nullptr) { - if(is_quantized) + if (is_quantized) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); } - else if(is_bf16) + else if (is_bf16) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32); } @@ -503,20 +562,23 @@ Status CpuGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weight } 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); + 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, weights->data_type()); weights_reshaped_info.set_quantization_info(weights->quantization_info()); weights_to_use = &weights_reshaped_info; - if(!skip_im2col) + if (!skip_im2col) { const int block_by = arm_compute::block_by(weights_info.weight_format()); int input_pad_right = 0; - if(block_by > 1) + if (block_by > 1) { - input_pad_right = (src->dimension(idx_channel) % block_by) == 0 ? 0 : block_by - (src->dimension(idx_channel) % block_by); - mat_weights_rows = weights->dimension(idx_width) * weights->dimension(idx_height) * (weights->dimension(idx_channel) + input_pad_right); + input_pad_right = + (src->dimension(idx_channel) % block_by) == 0 ? 0 : block_by - (src->dimension(idx_channel) % block_by); + mat_weights_rows = weights->dimension(idx_width) * weights->dimension(idx_height) * + (weights->dimension(idx_channel) + input_pad_right); } // Create tensor info for im2col reshaped inputs @@ -528,13 +590,15 @@ Status CpuGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weight 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, num_groups, input_pad_right)); + ARM_COMPUTE_RETURN_ON_ERROR( + kernels::CpuIm2ColKernel::validate(src, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), + conv_info, append_bias, dilation, num_groups, input_pad_right)); 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) + if (!skip_col2im) { TensorShape shape_gemm = gemm_input_to_use->tensor_shape(); shape_gemm.set(0, mat_weights_cols); @@ -549,13 +613,15 @@ Status CpuGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weight gemm_output_to_use = &info_gemm; const bool fixed_format = weights_info.weight_format() != arm_compute::WeightFormat::UNSPECIFIED; - 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, fixed_format, + 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, fixed_format, weights_info.weight_format())); // Validate Col2Im/ReshapeLayer - if(!skip_col2im && (data_layout == DataLayout::NCHW)) + 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))); + ARM_COMPUTE_RETURN_ON_ERROR( + kernels::CpuCol2ImKernel::validate(gemm_output_to_use, dst, Size2D(conv_w, conv_h))); } return Status{}; @@ -574,15 +640,11 @@ void CpuGemmConv2d::run(ITensorPack &tensors) 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) + 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() } - }; + 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(); } @@ -595,11 +657,11 @@ void CpuGemmConv2d::run(ITensorPack &tensors) gemm3d.allocator()->import_memory(out_to_use->buffer()); auto gemm_output_to_use = gemm_output.get(); - if(_skip_im2col) + if (_skip_im2col) { gemm_output_to_use = &gemm3d; } - if(_skip_col2im && !out_has_padding) + if (_skip_col2im && !out_has_padding) { gemm_output_to_use = dst; } @@ -607,12 +669,12 @@ void CpuGemmConv2d::run(ITensorPack &tensors) // Runs CpuGemm or CpuGemmLowpMatrixMultiplyCore functions ITensorPack pack_mm = tensors; pack_mm.add_const_tensor(TensorType::ACL_SRC_0, gemm_input_to_use); - if(!this->isVarWeightsKernel()) + if (!this->isVarWeightsKernel()) { 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) + if (_is_quantized) { // Run gemmlowp _mm_gemmlowp->run(pack_mm); @@ -624,45 +686,33 @@ void CpuGemmConv2d::run(ITensorPack &tensors) } // Reshape output matrix - if(!_skip_col2im) + if (!_skip_col2im) { - if(_data_layout == DataLayout::NCHW) + if (_data_layout == DataLayout::NCHW) { - ITensorPack pack = - { - { TensorType::ACL_SRC, gemm_output.get() }, - { TensorType::ACL_DST, dst } - }; + 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 } - }; + ITensorPack pack = {{TensorType::ACL_SRC, gemm_output_to_use}, {TensorType::ACL_DST, dst}}; _reshape->run(pack); } } - else if(out_has_padding) + else if (out_has_padding) { - ITensorPack pack = - { - { TensorType::ACL_SRC, gemm_output_to_use }, - { TensorType::ACL_DST, dst } - }; + ITensorPack pack = {{TensorType::ACL_SRC, gemm_output_to_use}, {TensorType::ACL_DST, dst}}; _reshape->run(pack); } } void CpuGemmConv2d::prepare(ITensorPack &tensors) { - if(!_is_prepared) + if (!_is_prepared) { // Variable weights executions that use fixed-format kernels // need no reshaping of the weights. - if(this->isVarWeightsKernel()) + if (this->isVarWeightsKernel()) { _is_quantized ? _mm_gemmlowp->prepare(tensors) : _mm_gemm->prepare(tensors); _is_prepared = true; @@ -672,11 +722,7 @@ void CpuGemmConv2d::prepare(ITensorPack &tensors) // 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() } - }; + 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(); ITensorPack gemm_pack = tensors; |