From 5e549fa8ff058eb06cb74de43b9f89a08d0b4a9c Mon Sep 17 00:00:00 2001 From: Mohammed Suhail Munshi Date: Wed, 16 Mar 2022 11:14:06 +0000 Subject: Add CLPool3d Int8 Support - Adds Qasymm8 and Qasymm8_signed support to the 3d pool operator Resolves: COMPMID-4669 Signed-off-by: Mohammed Suhail Munshi Change-Id: I36038c2b7c4f36baf67f7aae801356890e104538 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/410496 Tested-by: bsgcomp Reviewed-by: Sheri Zhang Comments-Addressed: bsgcomp Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7391 Tested-by: Arm Jenkins Reviewed-by: Gunes Bayir Comments-Addressed: Arm Jenkins --- src/gpu/cl/ClKernelLibrary.cpp | 5 ++++ src/gpu/cl/kernels/ClPool3dKernel.cpp | 55 +++++++++++++++++++++++++++-------- src/gpu/cl/kernels/ClPool3dKernel.h | 2 +- 3 files changed, 49 insertions(+), 13 deletions(-) (limited to 'src/gpu') diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp index f675dbe6fd..1bf7f2b3ac 100644 --- a/src/gpu/cl/ClKernelLibrary.cpp +++ b/src/gpu/cl/ClKernelLibrary.cpp @@ -438,6 +438,7 @@ const std::map ClKernelLibrary::_kernel_program_map = { "pooling_layer_2x2_nhwc", "nhwc/pooling_layer.cl" }, { "pooling_layer_MxN_quantized_nhwc", "nhwc/pooling_layer_quantized.cl" }, { "pooling_3d_layer_MxN_ndhwc", "nhwc/pooling_3d_layer.cl" }, + { "pooling_3d_layer_MxN_ndhwc_quantized", "nhwc/pooling_3d_layer_quantized.cl" }, { "reorg_layer_nhwc", "nhwc/reorg_layer.cl" }, { "scale_nearest_neighbour_nhwc", "nhwc/scale.cl" }, { "scale_bilinear_nhwc", "nhwc/scale.cl" }, @@ -905,6 +906,10 @@ const std::map ClKernelLibrary::_program_source_map = { "nhwc/pooling_3d_layer.cl", #include "./cl_kernels/nhwc/pooling_3d_layer.clembed" + }, + { + "nhwc/pooling_3d_layer_quantized.cl", +#include "./cl_kernels/nhwc/pooling_3d_layer_quantized.clembed" }, { "nhwc/pooling_layer_quantized.cl", diff --git a/src/gpu/cl/kernels/ClPool3dKernel.cpp b/src/gpu/cl/kernels/ClPool3dKernel.cpp index 929ccf7cb6..a090ac5774 100644 --- a/src/gpu/cl/kernels/ClPool3dKernel.cpp +++ b/src/gpu/cl/kernels/ClPool3dKernel.cpp @@ -49,19 +49,22 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_info.stride.x() == 0 || pool_info.stride.y() == 0 || pool_info.stride.z() == 0), "Strides cannot be zero."); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32, DataType::QASYMM8_SIGNED, DataType::QASYMM8); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((!is_data_type_float(src->data_type())) && (!pool_info.exclude_padding + && (pool_info.pool_type == PoolingType::AVG)), + "Exclude padding is unsupported for non-float types for Avg op"); - const auto 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_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::DEPTH); - const bool is_global_pooling = pool_info.is_global_pooling; + const auto 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_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::DEPTH); + const bool is_global_pooling = pool_info.is_global_pooling; const unsigned int pool_size_x = is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width; const unsigned int pool_size_y = is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height; const unsigned int pool_size_z = is_global_pooling ? src->dimension(idx_depth) : pool_info.pool_size.depth; - int output_width = 0; - int output_height = 0; - int output_depth = 0; + int output_width = 0; + int output_height = 0; + int output_depth = 0; bool round_type_ceil_with_asymm_padding = (pool_info.round_type == DimensionRoundingType::CEIL) && (!is_symmetric(pool_info.padding)); ARM_COMPUTE_RETURN_ERROR_ON_MSG(round_type_ceil_with_asymm_padding, "Cannot use dimension round type CEIL when padding is asymmetric."); @@ -143,10 +146,31 @@ void ClPool3dKernel::configure(const ClCompileContext &compile_context, const IT build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height))); build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(idx_depth))); + // If datatype is quantized add relevant parameters + if(is_data_type_quantized_asymmetric(data_type) && src->quantization_info() != dst->quantization_info()) + { + const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); + + build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset)); + build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); + build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale)); + build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); + } + // Set the initial value for the pooling operation accordingly with the data type if(pool_type == PoolingType::MAX) { - build_opts.add_option("-DINITIAL_VALUE=" + float_to_string_with_full_precision(std::numeric_limits::lowest())); + if(is_data_type_quantized(data_type)) + { + PixelValue type_min{}; + std::tie(type_min, std::ignore) = get_min_max(data_type); + build_opts.add_option("-DINITIAL_VALUE=" + support::cpp11::to_string(type_min.get())); + } + else + { + build_opts.add_option("-DINITIAL_VALUE=" + float_to_string_with_full_precision(std::numeric_limits::lowest())); + } } else { @@ -164,6 +188,11 @@ void ClPool3dKernel::configure(const ClCompileContext &compile_context, const IT { acc_data_type = DataType::F32; } + else if(is_data_type_quantized(data_type) && pool_type != PoolingType::MAX) // Use S32 for avg pooling to allow for integer division + { + acc_data_type = DataType::S32; + } + build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(acc_data_type)); build_opts.add_option_if(use_fp_mixed_precision, "-DFP_MIXED_PRECISION"); build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING"); @@ -172,8 +201,10 @@ void ClPool3dKernel::configure(const ClCompileContext &compile_context, const IT build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst->dimension(idx_channel))); build_opts.add_option("-DDST_BATCH_SIZE=" + support::cpp11::to_string(dst->dimension(idx_batch_size))); build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % _num_elems_processed_per_iteration)); - std::string kernel_name = "pooling_3d_layer_MxN_ndhwc"; - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); + + // if datatype is quantized use quantized kernel function + std::string kernel_name = (is_data_type_quantized_asymmetric(data_type) ? "pooling_3d_layer_MxN_ndhwc_quantized" : "pooling_3d_layer_MxN_ndhwc"); + _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); // Configure kernel window Window win = calculate_max_window(*dst, Steps(_num_elems_processed_per_iteration)); diff --git a/src/gpu/cl/kernels/ClPool3dKernel.h b/src/gpu/cl/kernels/ClPool3dKernel.h index 30c76ed632..00852349e6 100644 --- a/src/gpu/cl/kernels/ClPool3dKernel.h +++ b/src/gpu/cl/kernels/ClPool3dKernel.h @@ -46,7 +46,7 @@ public: * @note Asymmetric padding is not supported when dimension rounding type == CEIL. * * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data types supported: F16/F32. + * @param[in] src Source tensor info. Data types supported: F16/F32/QASYMM8/QASYMM8_SIGNED * @param[out] dst Destination tensor info. Data types supported: same as @p src. * @param[in] pool_info Contains pooling operation information described in @ref Pooling3dLayerInfo. */ -- cgit v1.2.1