diff options
Diffstat (limited to 'src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwDirectConv2d.cpp')
| -rw-r--r-- | src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwDirectConv2d.cpp | 333 |
1 files changed, 333 insertions, 0 deletions
diff --git a/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwDirectConv2d.cpp b/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwDirectConv2d.cpp new file mode 100644 index 0000000000..3c906646a6 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwDirectConv2d.cpp @@ -0,0 +1,333 @@ +/* + * Copyright (c) 2023 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/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwDirectConv2d.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/helpers/AdjustVecSize.h" +#include "arm_compute/core/utils/StringUtils.h" + +#include "ckw/TensorTileSampler.h" +#include "ckw/TileInfo.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.h" +#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwKernelWriter.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwScopedKernelWriter.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwVariableTable.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/WriterHelper.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/type_converter/Common.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ + +using TileContainer = std::vector<std::vector<std::string>>; + +GpuCkwDirectConv2d::GpuCkwDirectConv2d(ComponentId id, + const ArgumentPack<ITensorInfo> &tensors, + const Attributes &attributes, + const Settings &settings) + : IGpuCkwComponentDriver{ id, tensors }, + _src{}, + _wei{}, + _bia{}, + _dst{}, + _attributes{ attributes }, + _settings{ settings } +{ + _src = this->tensors().get_const_tensor(TensorType::ACL_SRC_0); + _wei = this->tensors().get_const_tensor(TensorType::ACL_SRC_1); + _bia = this->tensors().get_const_tensor(TensorType::ACL_SRC_2); + _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0); + ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _wei, _dst); // Bias can be null +} + +void GpuCkwDirectConv2d::write_component_code(const ComponentGroup &comp_group, GpuCkwVariableTable &vtable, GpuCkwScopedKernelWriter writer) const +{ + const auto desc = _settings.direct_conv_descriptor(); + ARM_COMPUTE_ERROR_ON_MSG(desc.export_input_to_cl_image || desc.export_output_to_cl_image, + "Only the weights tensor can be exported to cl_image"); + + const unsigned int channel_idx = get_data_layout_dimension_index(_src->data_layout(), DataLayoutDimension::CHANNEL); + const unsigned int width_idx = get_data_layout_dimension_index(_wei->data_layout(), DataLayoutDimension::WIDTH); + const unsigned int height_idx = get_data_layout_dimension_index(_wei->data_layout(), DataLayoutDimension::HEIGHT); + + const auto root_window = comp_group.get_root_component()->ckw_component_driver()->get_window(); + + // Tunable parameters + const int32_t m0 = root_window.y().step(); + const int32_t n0 = root_window.x().step(); + const int32_t k0 = adjust_vec_size(_settings.direct_conv_descriptor().k0, _src->dimension(channel_idx)); + const int32_t partial_n0 = _dst->dimension(0) % n0; + + const int32_t K = _src->dimension(channel_idx); + + // Exporting the weights tensor to an OpenCL image object is currently only supported when: + // a) k0 is equal to 4 + // The current implementation expects to read a vector of 4 float values into the OpenCL image object. + // b) K is a multiple of 4 + // This is a limitation in the current interface due to the variable table being responsible for maintaining + // information about the TensorStorageType rather than the TensorTileSampler. As a result, TensorStorageType cannot + // be reassigned, and we cannot use a texture object for the weights tensor in cases where we expect to have an + // extra loop to compute the left-over elements. + const bool use_cl_image_for_weights = desc.export_weights_to_cl_image && (k0 == 4) && (K % 4 == 0); + + GpuCkwComponentArgument *src = vtable.declare_variable(comp_group, writer, _src, TensorStorageType::ClBufferUint8Ptr, "src"); + GpuCkwComponentArgument *wei = vtable.declare_variable( + comp_group, writer, _wei, use_cl_image_for_weights ? TensorStorageType::ClImage2dReadOnly : TensorStorageType::ClBufferUint8Ptr, "wei"); + GpuCkwComponentArgument *dst = vtable.declare_variable(comp_group, writer, _dst, TensorStorageType::ClBufferUint8Ptr, "dst"); + GpuCkwComponentArgument *bia = nullptr; + + const bool using_bias = _bia != nullptr; + + if(using_bias) + { + bia = vtable.declare_variable(comp_group, writer, _bia, TensorStorageType::ClBufferUint8Ptr, "bia"); + } + + // Constants + const auto kernel_height = static_cast<int32_t>(_wei->dimension(height_idx)); + const auto kernel_width = static_cast<int32_t>(_wei->dimension(width_idx)); + const auto src_channels = static_cast<int32_t>(_src->dimension(channel_idx)); + auto &tile_kernel_w = writer->declare_tile("kernel_w", kernel_width); + auto &tile_kernel_size = writer->declare_tile("kernel_size", kernel_width * kernel_height); + auto &tile_src_c = writer->declare_tile("src_c", static_cast<int32_t>(_src->dimension(channel_idx))); + auto &tile_dst_w = writer->declare_tile("dst_w", static_cast<int32_t>(_dst->dimension(width_idx))); + auto &tile_stride_x = writer->declare_tile("stride_x", static_cast<int32_t>(_attributes.stride().x())); + auto &tile_stride_y = writer->declare_tile("stride_y", static_cast<int32_t>(_attributes.stride().y())); + auto &tile_pad_x = writer->declare_tile("pad_x", static_cast<int32_t>(_attributes.pad().left)); + auto &tile_pad_y = writer->declare_tile("pad_y", static_cast<int32_t>(_attributes.pad().top)); + auto &tile_k0 = writer->declare_tile("k0", k0); + auto &tile_0 = writer->declare_tile("0", 0); + auto &tile_1 = writer->declare_tile("1", 1); + + auto &tile_gid_0 = writer->declare_tile("gid_0", ckw::DataType::Int32); + auto &tile_gid_1 = writer->declare_tile("gid_1", ckw::DataType::Int32); + auto &tile_gid_2 = writer->declare_tile("gid_2", ckw::DataType::Int32); + + writer->op_get_global_id(tile_gid_0, 0); + writer->op_get_global_id(tile_gid_1, 1); + writer->op_get_global_id(tile_gid_2, 2); + + auto &tile_cout = writer->declare_tile("cout", ckw::DataType::Int32); // OFM + auto &tile_mout = writer->declare_tile("mout", ckw::DataType::Int32); // WIDTH x HEIGHT + auto &tile_bout = writer->declare_tile("bout", ckw::DataType::Int32); // BATCH SIZE IDX + + // Get the boundary aware coordinates at each global dimension index + get_coord(writer, tile_cout, tile_gid_0, n0, partial_n0, tile_cout.name() + "_dim0_", tile_0); + get_coord(writer, tile_mout, tile_gid_1, m0, 0, tile_mout.name() + "_dim1_", tile_0); + get_coord(writer, tile_bout, tile_gid_2, 1, 0, tile_bout.name() + "_dim2_", tile_0); + + TensorTileSampler src_sampler; + src_sampler.width(k0); + src_sampler.height(m0); + src_sampler.format(TensorSamplerFormat::C_WH_1); + // We cannot have out-of-bounds reads in the X dimension (mapped to the IFMs) as we have an extra loop to + // compute left-over elements + src_sampler.address_mode_x(TensorSamplerAddressModeX::None); + // We cannot have out-of-bounds reads when the kernel height is equal to 1. Otherwise, we need to ensure the + // indirection buffer mi does not contain negative values representing out-of-bounds reads. + src_sampler.address_mode_y(kernel_height == 1 ? TensorSamplerAddressModeY::None : TensorSamplerAddressModeY::SkipMinEdgeOnly); + src_sampler.address_mode_z(TensorSamplerAddressModeZ::None); + + TensorTileSampler wei_sampler; + wei_sampler.width(k0); + wei_sampler.height(n0); + wei_sampler.format(TensorSamplerFormat::C_WH_1); + // We cannot have out-of-bounds accesses for the weights + wei_sampler.address_mode_x(TensorSamplerAddressModeX::None); + wei_sampler.address_mode_y(TensorSamplerAddressModeY::None); + wei_sampler.address_mode_z(TensorSamplerAddressModeZ::None); + + TensorTileSampler dst_sampler; + dst_sampler.width(n0); + dst_sampler.height(m0); + dst_sampler.format(TensorSamplerFormat::C_WH_1); + dst_sampler.address_mode_x(TensorSamplerAddressModeX::OverlappingMin); + dst_sampler.address_mode_y(TensorSamplerAddressModeY::ClampToMaxEdgeOnly); + dst_sampler.address_mode_z(TensorSamplerAddressModeZ::None); + dst_sampler.x(tile_cout); + dst_sampler.y(tile_mout); + dst_sampler.z(tile_0); + dst_sampler.b(tile_bout); + + if(!dst->has_tile()) + { + auto &tile = writer->declare_tile("dst", TileInfo(to_ckw(_dst->data_type()), m0, n0)); + dst->init_virtual_tensor(tile, dst_sampler); + } + auto &tile_dst = dst->tile(); + + writer->op_assign(tile_dst, tile_0); + + // We create a 2d container of size (M0, 1) to store the indices for iteration + TileContainer it; + for(int m = 0; m < m0; ++m) + { + std::vector<std::string> idx { std::to_string(m) }; + it.push_back({ idx }); + } + const auto &tile_it = writer->declare_tile("it", it, ckw::DataType::Int32); + + auto &tile_xi = writer->declare_tile("xi", TileInfo(ckw::DataType::Int32, m0, 1)); + auto &tile_yi = writer->declare_tile("yi", TileInfo(ckw::DataType::Int32, m0, 1)); + + // Convert the linear index to coordinate + // xi = ((mout + i) % dst_w) * stride_x - pad_x + // yi = ((mout + i) / dst_w) * stride_y - pad_y + writer->op_binary_expression(tile_xi, tile_mout, BinaryOp::Add, tile_it); + writer->op_binary_expression(tile_yi, tile_mout, BinaryOp::Add, tile_it); + writer->op_binary_expression(tile_xi, tile_xi, BinaryOp::Mod, tile_dst_w); + writer->op_binary_expression(tile_yi, tile_yi, BinaryOp::Div, tile_dst_w); + writer->op_binary_expression(tile_xi, tile_xi, BinaryOp::Mul, tile_stride_x); + writer->op_binary_expression(tile_yi, tile_yi, BinaryOp::Mul, tile_stride_y); + writer->op_binary_expression(tile_xi, tile_xi, BinaryOp::Sub, tile_pad_x); + writer->op_binary_expression(tile_yi, tile_yi, BinaryOp::Sub, tile_pad_y); + + auto &tile_y_b = writer->declare_tile("y_b", ckw::DataType::Int32); + writer->op_binary_expression(tile_y_b, tile_cout, BinaryOp::Mul, tile_kernel_size); + + auto &tile_i = writer->declare_tile("i", ckw::DataType::Int32); + writer->op_assign(tile_i, tile_0); + + // clang-format off + writer->op_for_loop(tile_i, BinaryOp::Less, tile_kernel_size, tile_i, AssignmentOp::Increment, tile_1, [&]() + { + auto &tile_x_k = writer->declare_tile("x_k", ckw::DataType::Int32); + auto &tile_y_k = writer->declare_tile("y_k", ckw::DataType::Int32); + + writer->op_binary_expression(tile_x_k, tile_i, BinaryOp::Mod, tile_kernel_w); + writer->op_binary_expression(tile_y_k, tile_i, BinaryOp::Div, tile_kernel_w); + + auto &tile_ck = writer->declare_tile("ck", ckw::DataType::Int32); + writer->op_assign(tile_ck, tile_0); + + auto &tile_mi = writer->declare_tile("mi", TileInfo(ckw::DataType::Int32, m0, 1)); + // Construct an indirection buffer containing the precalculated addresses of elements in the source tensor + // x_s = xi + x_k + // y_s = yi + y_k + // mi = x_s + y_s * width; + // mi = select(-1, mi, x_s >= 0); + // mi = select(-1, mi, x_s < width); + // mi = select(-1, mi, y_s >= 0); + // mi = select(-1, mi, y_s < height); + writer->util_get_indirect_buffer(tile_mi, src->tensor(), src_sampler, tile_xi, tile_yi, tile_x_k, tile_y_k); + + src_sampler.x(tile_ck); + src_sampler.y(tile_mi); + src_sampler.z(tile_0); + src_sampler.b(tile_bout); + + wei_sampler.x(tile_ck); + wei_sampler.y(tile_y_b); + wei_sampler.z(tile_0); + wei_sampler.b(tile_0); + + auto &tile_src_c_minus_k0 = writer->declare_tile("src_c_minus_k0", src_channels - k0); + + writer->op_for_loop(tile_ck, BinaryOp::LessEqual, tile_src_c_minus_k0, tile_ck, AssignmentOp::Increment, tile_k0, [&]() + { + auto &tile_lhs = writer->declare_tile("lhs", TileInfo(to_ckw(_src->data_type()), m0, k0)); + auto &tile_rhs = writer->declare_tile("rhs", TileInfo(to_ckw(_wei->data_type()), n0, k0)); + writer->op_assign(tile_lhs, tile_0); + writer->op_assign(tile_rhs, tile_0); + + writer->op_load_indirect(tile_lhs, src->tensor(), src_sampler); + writer->op_load(tile_rhs, wei->tensor(), wei_sampler, tile_kernel_size); + + writer->op_binary_expression(tile_dst, tile_lhs, BinaryOp::MatMul_Nt_T, tile_rhs); + }); + + // Left-over accumulations for when K is not a multiple of k0 + if(!(K % k0 == 0)) + { + writer->op_for_loop(tile_ck, BinaryOp::Less, tile_src_c, tile_ck, AssignmentOp::Increment, tile_1, [&]() + { + auto &tile_lhs = writer->declare_tile("lhs_leftover", TileInfo(to_ckw(_src->data_type()), m0, 1)); + auto &tile_rhs = writer->declare_tile("rhs_leftover", TileInfo(to_ckw(_wei->data_type()), n0, 1)); + writer->op_assign(tile_lhs, tile_0); + writer->op_assign(tile_rhs, tile_0); + + writer->op_load_indirect(tile_lhs, src->tensor(), src_sampler); + writer->op_load(tile_rhs, wei->tensor(), wei_sampler, tile_kernel_size); + + writer->op_binary_expression(tile_dst, tile_lhs, BinaryOp::MatMul_Nt_T, tile_rhs); + }); + } + + writer->op_binary_expression(tile_y_b, tile_y_b, BinaryOp::Add, tile_1); + }); + // clang-format on + + // Bias addition + // NOTE: This operation will be removed from this kernel as the interface is standardized. The intended way of + // performing bias addition is to fuse this convolution kernel with a following elementwise addition kernel. + if(using_bias) + { + if(!bia->has_tile()) + { + // Reuse the destination sampler for the bias + writer->op_load_once(bia, dst_sampler); + } + auto &tile_bia = bia->tile(); + + writer->op_binary_expression(tile_dst, tile_dst, BinaryOp::Add, tile_bia); + } +} + +Window GpuCkwDirectConv2d::get_window() const +{ + ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized"); + + const auto dst_shape = _dst->tensor_shape(); + const auto desc = _settings.direct_conv_descriptor(); + + const unsigned int n0 = adjust_vec_size(desc.n0, dst_shape[0]); + const unsigned int m0 = adjust_vec_size(desc.m0, dst_shape[1] * dst_shape[2]); + + Window win = calculate_max_window(dst_shape, Steps(n0, m0)); + + const size_t dim_y_collapsed = ceil_to_multiple(dst_shape[1] * dst_shape[2], m0); + win.set(Window::DimY, Window::Dimension(0, dim_y_collapsed, m0)); + win.set(Window::DimZ, Window::Dimension(0, dst_shape.total_size_upper(3), 1)); + + return win; +} + +std::string GpuCkwDirectConv2d::get_name(const ComponentGroup &comp_group) const +{ + ARM_COMPUTE_UNUSED(comp_group); + + return "direct_conv2d"; +} + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute |