From 3ecf9fefa6f6299a0736599f150d4791cc8345d9 Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Wed, 28 Apr 2021 16:11:51 +0100 Subject: Remove OpenCL padding: CLReductionOperationKernel Change the parallel implementation across the X, now every thread computes one row Add missing test for MEAN_SUM Make reduction on any axis != 0 work with num_channels > 1 Resolve COMPMID-3917 Signed-off-by: Giorgio Arena Change-Id: Ib0f99540104e3c253bcd1ea637833db533f5e76e Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5522 Comments-Addressed: Arm Jenkins Reviewed-by: Manuel Bottini Reviewed-by: Gian Marco Iodice Tested-by: Arm Jenkins --- src/core/CL/kernels/CLReductionOperationKernel.cpp | 147 ++++++--------------- 1 file changed, 43 insertions(+), 104 deletions(-) (limited to 'src/core/CL/kernels/CLReductionOperationKernel.cpp') diff --git a/src/core/CL/kernels/CLReductionOperationKernel.cpp b/src/core/CL/kernels/CLReductionOperationKernel.cpp index a5734d43de..5c80f33802 100644 --- a/src/core/CL/kernels/CLReductionOperationKernel.cpp +++ b/src/core/CL/kernels/CLReductionOperationKernel.cpp @@ -42,10 +42,7 @@ namespace arm_compute { namespace { -// OpenCL kernel requires input width to be a power of 2 for x-axis. -constexpr unsigned int border_val = 64; - -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op, unsigned int width) +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); @@ -56,11 +53,13 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, u else { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 2, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(axis == 0); } ARM_COMPUTE_RETURN_ERROR_ON_MSG(op == ReductionOperation::SUM_SQUARE && input->data_type() == DataType::QASYMM8, "Not supported reduction operation for QASYMM8"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis"); - ARM_COMPUTE_RETURN_ERROR_ON((op == ReductionOperation::MEAN_SUM) && (axis == 0) && (width == 0) && (input->data_type() != DataType::QASYMM8) && (input->data_type() != DataType::QASYMM8_SIGNED)); + ARM_COMPUTE_RETURN_ERROR_ON((op == ReductionOperation::MEAN_SUM) && (axis == 0) && (input->dimension(0) == 0) && (input->data_type() != DataType::QASYMM8) + && (input->data_type() != DataType::QASYMM8_SIGNED)); ARM_COMPUTE_RETURN_ERROR_ON_MSG((op == ReductionOperation::ARG_IDX_MAX) || (op == ReductionOperation::ARG_IDX_MIN), "Not supported reduction operation, use CLArgMinMaxLayer"); if(output->total_size() != 0) @@ -71,77 +70,34 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, u return Status{}; } - -std::tuple validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, unsigned int axis, ReductionOperation op) -{ - // Output tensor auto initialization if not yet initialized - const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis, true); - DataType output_data_type = input->data_type(); - auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape).set_data_type(output_data_type).reset_padding().set_is_resizable(true)); - - const unsigned int num_elems_processed_per_iteration = (is_data_type_quantized(input->data_type()) && (axis == 0)) ? 1 : 16; - Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); - bool window_changed = false; - const bool is_serial_op = needs_serialized_reduction(op, input->data_type(), axis); - - switch(axis) - { - case 0: - { - if(!is_serial_op) - { - const unsigned int border_width = ((input->dimension(0) % border_val) != 0) ? border_val - input->dimension(0) % border_val : 0; - AccessWindowStatic input_access(input, 0, 0, input->dimension(0) + border_width, 1); - AccessWindowHorizontal output_access(output, 0, 1); - window_changed = update_window_and_padding(win, input_access, output_access); - } - } - break; - case 1: - case 2: - case 3: - { - AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); - AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); - window_changed = update_window_and_padding(win, input_access, output_access); - } - break; - default: - ARM_COMPUTE_ERROR("Not supported"); - } - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - - return std::make_tuple(err, win); -} } // namespace CLReductionOperationKernel::CLReductionOperationKernel() - : _input(nullptr), _output(nullptr), _reduction_axis(0), _op(ReductionOperation::SUM_SQUARE), _border_size() -{ -} - -BorderSize CLReductionOperationKernel::border_size() const + : _input(nullptr), _output(nullptr), _reduction_axis(0), _op(ReductionOperation::SUM_SQUARE) { - return _border_size; } -void CLReductionOperationKernel::configure(const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op, unsigned int width) +void CLReductionOperationKernel::configure(const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op) { - configure(CLKernelLibrary::get().get_compile_context(), input, output, axis, op, width); + configure(CLKernelLibrary::get().get_compile_context(), input, output, axis, op); } -void CLReductionOperationKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op, unsigned int width) +void CLReductionOperationKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis, op, width)); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis, op)); + + auto padding_info = get_padding_info({ input, output }); _input = input; _output = output; _reduction_axis = axis; _op = op; + const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, true); + auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).reset_padding().set_is_resizable(true)); + // Set build options CLBuildOptions build_opts; DataType data_type = input->info()->data_type(); @@ -156,8 +112,15 @@ void CLReductionOperationKernel::configure(const CLCompileContext &compile_conte data_type_promoted = get_cl_type_from_data_type(data_type); } + const unsigned int width = input->info()->dimension(0) * input->info()->num_channels(); + unsigned int vec_size = (is_data_type_quantized(input->info()->data_type()) && (axis == 0)) ? 1 : 16; + vec_size = adjust_vec_size(vec_size, width); + const unsigned int vec_size_leftover = width % vec_size; + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); build_opts.add_option("-DDATA_TYPE_PROMOTED=" + data_type_promoted); + build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size)); + build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_leftover)); build_opts.add_option_if(is_data_type_float(data_type), "-DFLOAT_DATA_TYPE"); build_opts.add_option_if(op == ReductionOperation::SUM_SQUARE, "-DSUM_SQUARE"); build_opts.add_option_if(op == ReductionOperation::MEAN_SUM, "-DMEAN"); @@ -165,7 +128,6 @@ void CLReductionOperationKernel::configure(const CLCompileContext &compile_conte build_opts.add_option_if(op == ReductionOperation::PROD, "-DPROD"); build_opts.add_option_if(op == ReductionOperation::MIN, "-DMIN"); build_opts.add_option_if(op == ReductionOperation::MAX, "-DMAX"); - build_opts.add_option_if(input->info()->num_channels() == 2, "-DCOMPLEX"); build_opts.add_option_if(is_data_type_quantized(data_type), "-DOFFSET=" + support::cpp11::to_string(input->info()->quantization_info().uniform().offset)); build_opts.add_option_if(is_data_type_quantized(data_type), "-DSCALE=" + float_to_string_with_full_precision(input->info()->quantization_info().uniform().scale)); @@ -189,7 +151,6 @@ void CLReductionOperationKernel::configure(const CLCompileContext &compile_conte } // Create kernel - cl::NDRange lws_hint = CLKernelLibrary::get().default_ndrange(); std::string kernel_axis_name; const bool is_serial_op = needs_serialized_reduction(_op, _input->info()->data_type(), _reduction_axis); @@ -197,22 +158,14 @@ void CLReductionOperationKernel::configure(const CLCompileContext &compile_conte { case 0: { + build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(width)); + kernel_axis_name = "x"; + if(is_serial_op) { - build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(input->info()->dimension(0))); build_opts.add_option_if_else(_input->info()->data_type() == DataType::F16, "-DCOND_DATA_TYPE=short", "-DCOND_DATA_TYPE=int"); kernel_axis_name = "non_parallel_x"; } - else - { - build_opts.add_option_if(op == ReductionOperation::MEAN_SUM, "-DWIDTH=" + support::cpp11::to_string(width)); - const unsigned int width_leftover = input->info()->dimension(0) % border_val; - const unsigned int border_width = (width_leftover != 0) ? border_val - width_leftover : 0; - kernel_axis_name = "x"; - - lws_hint = create_lws_hint_parallel_implementations(input->info()->dimension(0), border_val); - _border_size = BorderSize(0, border_width, 0, 0); - } } break; case 1: @@ -234,18 +187,16 @@ void CLReductionOperationKernel::configure(const CLCompileContext &compile_conte _kernel = create_kernel(compile_context, "reduction_operation_" + kernel_axis_name, build_opts.options()); // Configure kernel window - auto win_config = validate_and_configure_window(_input->info(), _output->info(), axis, op); + Window win = calculate_max_window(*input->info(), Steps(vec_size)); + win.set(Window::DimX, Window::Dimension(win.x().start(), win.x().end() * _input->info()->num_channels(), win.x().step())); + ICLKernel::configure_internal(win); - ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); - - ICLKernel::configure_internal(std::get<1>(win_config), lws_hint); + ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); } -Status CLReductionOperationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op, unsigned int width) +Status CLReductionOperationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis, op, width)); - ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get(), axis, op))); - + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis, op)); return Status{}; } @@ -277,36 +228,24 @@ void CLReductionOperationKernel::run(const Window &window, cl::CommandQueue &que unsigned int idx = 0; add_1D_tensor_argument(idx, _input, in_slice); add_1D_tensor_argument(idx, _output, out_slice); - enqueue(queue, *this, in_slice, lws_hint()); + enqueue(queue, *this, in_slice); } while(window_in.slide_window_slice_1D(in_slice) && out_window.slide_window_slice_1D(out_slice)); } else { // Set out window - Window out_window(window); - out_window.set(Window::DimX, Window::Dimension(0, 0, 0)); - - // Get first input and output slices - Window in_slice = window.first_slice_window_2D(); - Window out_slice = out_window.first_slice_window_2D(); - - // Reshape window - const unsigned int border_width = ((in_slice.x().end() % border_val) != 0) ? border_val - in_slice.x().end() % border_val : 0; - in_slice.set(Window::DimX, Window::Dimension(in_slice.x().start(), in_slice.x().end() + border_width, in_slice.x().step())); + bool has_collapsed = true; + Window window_in = window.collapse_if_possible(window, 1, &has_collapsed); + ARM_COMPUTE_ERROR_ON(!has_collapsed); - // Set local sums buffer - unsigned int local_res_size = lws_hint()[0] * _input->info()->element_size(); - _kernel.setArg(num_arguments_per_2D_tensor() * 2, local_res_size, nullptr); + Window window_out = window_in; + window_out.set(0, Window::Dimension()); - do - { - unsigned int idx = 0; - add_2D_tensor_argument(idx, _input, in_slice); - add_2D_tensor_argument(idx, _output, out_slice); - enqueue(queue, *this, in_slice, lws_hint()); - } - while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice)); + unsigned int idx = 0; + add_2D_tensor_argument(idx, _input, window_in); + add_2D_tensor_argument(idx, _output, window_out); + enqueue(queue, *this, window_in); } } break; @@ -323,7 +262,7 @@ void CLReductionOperationKernel::run(const Window &window, cl::CommandQueue &que unsigned int idx = 0; add_2D_tensor_argument(idx, _input, in_slice); add_2D_tensor_argument(idx, _output, out_slice); - enqueue(queue, *this, in_slice, lws_hint()); + enqueue(queue, *this, in_slice); } while(window_in.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice)); } @@ -341,7 +280,7 @@ void CLReductionOperationKernel::run(const Window &window, cl::CommandQueue &que unsigned int idx = 0; add_3D_tensor_argument(idx, _input, in_slice); add_3D_tensor_argument(idx, _output, out_slice); - enqueue(queue, *this, in_slice, lws_hint()); + enqueue(queue, *this, in_slice); } while(window_in.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(out_slice)); } @@ -359,7 +298,7 @@ void CLReductionOperationKernel::run(const Window &window, cl::CommandQueue &que unsigned int idx = 0; add_4D_tensor_argument(idx, _input, in_slice); add_4D_tensor_argument(idx, _output, out_slice); - enqueue(queue, *this, in_slice, lws_hint()); + enqueue(queue, *this, in_slice); } while(window_in.slide_window_slice_4D(in_slice) && window.slide_window_slice_4D(out_slice)); } -- cgit v1.2.1