diff options
Diffstat (limited to 'src/core/CL/kernels')
6 files changed, 68 insertions, 37 deletions
diff --git a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp index 0b663e8498..f2119728c9 100644 --- a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp +++ b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp @@ -48,16 +48,17 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, C ARM_COMPUTE_RETURN_ERROR_ON(input == output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, - DataType::U8, DataType::S8, DataType::S16, + DataType::U8, DataType::S8, DataType::QSYMM8_PER_CHANNEL, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, - DataType::U8, DataType::S8, DataType::S16, + DataType::U8, DataType::S8, DataType::QASYMM8, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == output->data_type(), "Input and output data types must be different"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_float(input->data_type()) && shift != 0, "Shift is used only with integer inputs"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_float(input->data_type()) && shift != 0, "Shift is used only with integer non-quantized inputs"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized(input->data_type()) && shift != 0, "Shift is used only with integer non-quantized inputs"); ARM_COMPUTE_RETURN_ERROR_ON(shift >= 8); // Validate in case of configured output @@ -94,13 +95,14 @@ void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *out // Conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined build_opts.add_option_if(is_data_type_float(input->info()->data_type()) || policy == ConvertPolicy::SATURATE, "-DSATURATE"); build_opts.add_option_if(is_data_type_float(input->info()->data_type()) || is_data_type_float(output->info()->data_type()), "-DIS_DATA_TYPE_FLOAT"); + build_opts.add_option_if(is_data_type_quantized(input->info()->data_type()), "-DIS_DATA_TYPE_QUANTIZED"); // Create kernel const std::string kernel_name = (input_size >= output_size) ? "convert_depth_down" : "convert_depth_up"; _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); // Set shift arg - unsigned int idx = 2 * num_arguments_per_3D_tensor(); //Skip the input and output parameters + unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters _kernel.setArg(idx++, shift); // Configure kernel diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.cpp index 4bcfa82ca7..09caeeea55 100644 --- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.cpp @@ -46,8 +46,6 @@ namespace arm_compute { using namespace misc::shape_calculator; -class Coordinates; - namespace { using ElementsProcessed = Steps; @@ -56,7 +54,6 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const GEMMReshapeInfo &gemm_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3"); diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp index 27d5b28943..779f96e7cf 100644 --- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp @@ -54,7 +54,6 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const GEMMReshapeInfo &gemm_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3"); diff --git a/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp b/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp index 1852262337..2ebd76e1bf 100644 --- a/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp @@ -37,17 +37,12 @@ #include <cstddef> #include <cstdint> -using namespace arm_compute; - namespace arm_compute { -class Coordinates; -} // namespace arm_compute - namespace { Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, - int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage) + int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32); ARM_COMPUTE_RETURN_ERROR_ON(output_stage.type == GEMMLowpOutputStageType::NONE); @@ -61,6 +56,16 @@ Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vecto ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != bias->dimension(0)); } + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1); + if(output_stage.is_quantized_per_channel) + { + ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != output_shifts->dimension(0)); + ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != output_multipliers->dimension(0)); + } + // If a_offset == 0, vector_sum_col can be a nullptr if(a_offset != 0) { @@ -109,11 +114,14 @@ Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vecto ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mm_result, output); } + ARM_COMPUTE_RETURN_ERROR_ON_MSG(output_stage.gemmlowp_multipliers.size() != output_stage.gemmlowp_shifts.size(), + "per channel quantization info is incorrect"); + return Status{}; } std::pair<Status, Window> validate_and_configure_window(ITensorInfo *mm_result, ITensorInfo *vector_sum_col, ITensorInfo *vector_sum_row, ITensorInfo *bias, ITensorInfo *output, - int32_t a_offset, int32_t b_offset) + int32_t a_offset, int32_t b_offset, ITensorInfo *output_multipliers, ITensorInfo *output_shifts) { constexpr unsigned int num_elems_processed_per_iteration = 4; bool window_changed = false; @@ -147,36 +155,55 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *mm_result, window_changed = window_changed || update_window_and_padding(win, bias_access); } + if(output_multipliers->dimension(0) > 1) + { + AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_processed_per_iteration); + AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_processed_per_iteration); + window_changed = window_changed || update_window_and_padding(win, output_multipliers_access, output_shifts_access); + } + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; return std::make_pair(err, win); } } // namespace CLGEMMLowpOffsetContributionOutputStageKernel::CLGEMMLowpOffsetContributionOutputStageKernel() - : _mm_result(nullptr), _vector_sum_col(nullptr), _vector_sum_row(nullptr), _bias(nullptr), _output(nullptr) + : _mm_result(nullptr), + _vector_sum_col(nullptr), + _vector_sum_row(nullptr), + _bias(nullptr), + _output(nullptr), + _output_multipliers(nullptr), + _output_shifts(nullptr), + _is_quantized_per_channel(false) { } void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const ICLTensor *mm_result, const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias, ICLTensor *output, - int32_t k, int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage) + int32_t k, int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage, + const ICLTensor *output_multipliers, const ICLTensor *output_shifts) { // Perform validate step - ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result, output); + ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result, output, output_multipliers, output_shifts); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(mm_result->info(), vector_sum_col != nullptr ? vector_sum_col->info() : nullptr, vector_sum_row != nullptr ? vector_sum_row->info() : nullptr, bias != nullptr ? bias->info() : nullptr, output->info(), - a_offset, b_offset, output_stage)); // NOLINT + a_offset, b_offset, output_stage, + output_multipliers->info(), output_shifts->info())); // NOLINT const int min = output_stage.gemmlowp_min_bound; const int max = output_stage.gemmlowp_max_bound; - _vector_sum_col = vector_sum_col; - _vector_sum_row = vector_sum_row; - _mm_result = mm_result; - _bias = bias; - _output = output; + _vector_sum_col = vector_sum_col; + _vector_sum_row = vector_sum_row; + _mm_result = mm_result; + _bias = bias; + _output = output; + _output_multipliers = output_multipliers; + _output_shifts = output_shifts; + _is_quantized_per_channel = output_stage.is_quantized_per_channel; // Check if input is a 3D reinterpretation const bool reinterpret_as_3d = vector_sum_row != nullptr @@ -199,8 +226,9 @@ void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const ICLTensor *m build_opts.add_option_if(reinterpret_as_3d, "-DDEPTH_INPUT3D=" + support::cpp11::to_string(mm_result->info()->dimension(2))); build_opts.add_option_if(bias != nullptr, "-DADD_BIAS"); build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(output_stage.gemmlowp_offset)); - build_opts.add_option("-DRESULT_MULTIPLIER=" + support::cpp11::to_string(output_stage.gemmlowp_multiplier)); - build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(output_stage.gemmlowp_shift)); + build_opts.add_option("-DRESULT_MULTIPLIER=" + support::cpp11::to_string(output_stage.gemmlowp_multipliers[0])); + build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(output_stage.gemmlowp_shifts[0])); + build_opts.add_option_if(_is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION"); build_opts.add_option_if((min != 0) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min)); build_opts.add_option_if((max != 255) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max)); @@ -225,7 +253,8 @@ void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const ICLTensor *m vector_sum_row != nullptr ? vector_sum_row->info() : nullptr, bias != nullptr ? bias->info() : nullptr, output->info(), - a_offset, b_offset); // NOLINT + a_offset, b_offset, + output_multipliers->info(), output_shifts->info()); // NOLINT ARM_COMPUTE_ERROR_THROW_ON(win_config.first); ICLKernel::configure_internal(win_config.second); @@ -239,16 +268,17 @@ void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const ICLTensor *m } Status CLGEMMLowpOffsetContributionOutputStageKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, - const ITensorInfo *output, - int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage) + const ITensorInfo *output, int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage, + const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, output, a_offset, b_offset, output_stage)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, output, a_offset, b_offset, output_stage, output_multipliers, output_shifts)); ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(mm_result->clone().get(), vector_sum_col != nullptr ? vector_sum_col->clone().get() : nullptr, vector_sum_row != nullptr ? vector_sum_row->clone().get() : nullptr, bias != nullptr ? bias->clone().get() : nullptr, output->clone().get(), - a_offset, b_offset) + a_offset, b_offset, + output_multipliers->clone().get(), output_shifts->clone().get()) .first); // NOLINT return Status{}; @@ -285,7 +315,10 @@ void CLGEMMLowpOffsetContributionOutputStageKernel::run(const Window &window, cl add_2D_tensor_argument_if((_vector_sum_row != nullptr), idx, _vector_sum_row, win_vector_sum_row); add_1D_tensor_argument_if((_bias != nullptr), idx, _bias, biases_slice); add_3D_tensor_argument(idx, _output, slice); + add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_multipliers, biases_slice); + add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_shifts, biases_slice); enqueue(queue, *this, slice, lws_hint()); } while(collapsed.slide_window_slice_3D(slice)); } +} // namespace arm_compute diff --git a/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp b/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp index 6f6019d26a..3d681dd13e 100644 --- a/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp +++ b/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp @@ -55,9 +55,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c ARM_COMPUTE_RETURN_ERROR_ON((rhs_info.k0 == 1) && (rhs_info.transpose)); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::U8, DataType::S8, - DataType::U16, DataType::S16, DataType::U32, DataType::S32, - DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); if(output->total_size() != 0) { diff --git a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp index 9330b3b8a1..e325feac1f 100644 --- a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp +++ b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp @@ -33,7 +33,8 @@ #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" -using namespace arm_compute; +namespace arm_compute +{ using namespace arm_compute::misc::shape_calculator; namespace @@ -42,7 +43,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, c { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); ARM_COMPUTE_RETURN_ERROR_ON(num_groups == 0); ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::NHWC && num_groups > 1); ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4 && num_groups > 1); @@ -50,7 +51,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, c if(biases != nullptr) { - ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type())); + ARM_COMPUTE_RETURN_ERROR_ON(!is_data_type_float(input->data_type())); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->num_dimensions() != 1)); ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->num_dimensions() != 2)); @@ -160,3 +161,4 @@ void CLWeightsReshapeKernel::run(const Window &window, cl::CommandQueue &queue) } while(window.slide_window_slice_4D(in_slice) && out_window.slide_window_slice_2D(out_slice)); } +} // namespace arm_compute |