diff options
Diffstat (limited to 'src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwMatMul.cpp')
| -rw-r--r-- | src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwMatMul.cpp | 287 |
1 files changed, 287 insertions, 0 deletions
diff --git a/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwMatMul.cpp b/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwMatMul.cpp new file mode 100644 index 0000000000..14ad3847fc --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwMatMul.cpp @@ -0,0 +1,287 @@ +/* + * Copyright (c) 2023-2024 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/GpuCkwMatMul.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/utils/helpers/AdjustVecSize.h" +#include "arm_compute/core/Validate.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/CkwHelper.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/type_converter/Common.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwComponentArgument.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwVariableTable.h" +#include "support/StringSupport.h" + +#include "compute_kernel_writer/include/ckw/KernelWriter.h" +#include <cstdint> + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ + +GpuCkwMatMul::GpuCkwMatMul(ComponentId id, + const ArgumentPack<ITensorInfo> &tensors, + const Attributes &attributes, + const Settings &settings) + : IGpuCkwComponentDriver{id, tensors}, _lhs{}, _rhs{}, _dst{}, _attributes{attributes}, _settings{settings} +{ + _lhs = this->tensors().get_const_tensor(TensorType::ACL_SRC_0); + _rhs = this->tensors().get_const_tensor(TensorType::ACL_SRC_1); + _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0); + ARM_COMPUTE_ERROR_ON_NULLPTR(_lhs, _rhs, _dst); +} + +void GpuCkwMatMul::write_component_code(const ComponentGroup &comp_group, + GpuCkwVariableTable &vtable, + GpuCkwScopedKernelWriter writer) const +{ + /******************************************************************************** + * 1 - Define tensors + ********************************************************************************/ + GpuCkwComponentArgument *lhs = vtable.declare_variable(comp_group, writer, _lhs, "lhs"); + GpuCkwComponentArgument *rhs = vtable.declare_variable(comp_group, writer, _rhs, "rhs"); + GpuCkwComponentArgument *dst = vtable.declare_variable(comp_group, writer, _dst, "dst"); + + /******************************************************************************** + * 2 - Define CKW constants + ********************************************************************************/ + const auto k = + _attributes.adj_lhs() ? static_cast<int32_t>(_lhs->dimension(1)) : static_cast<int32_t>(_lhs->dimension(0)); + const auto k0 = static_cast<int32_t>(adjust_vec_size(_settings.k0(), k)); + const auto dst_dt = to_ckw(_dst->data_type()); + + // CKW constants + auto const_k_i32 = writer->declare_constant_tile(ckw::ConstantData({{k}}, ckw::DataType::Int32)); + auto const_k0_i32 = writer->declare_constant_tile(ckw::ConstantData({{k0}}, ckw::DataType::Int32)); + auto const_0_i32 = writer->declare_constant_tile(ckw::ConstantData({{0}}, ckw::DataType::Int32)); + auto const_pos_1_i32 = writer->declare_constant_tile(ckw::ConstantData({{1}}, ckw::DataType::Int32)); + auto const_0_fp = writer->declare_constant_tile(ckw::ConstantData({{0.0f}}, dst_dt)); + auto const_k_minus_k0_i32 = writer->declare_constant_tile(ckw::ConstantData({{k - k0}}, ckw::DataType::Int32)); + + /******************************************************************************** + * 3 - Define the compute block parameters and destination tile (if not root component) + * Bind the tile to the tensor to share it among different components and + * initialize the compute block parameters + ********************************************************************************/ + // The n0 and m0 parameters from root_window only refers to the output + const auto root_window = comp_group.get_root_component()->ckw_component_driver()->get_window(); + + // Destination compute block size + const int32_t dst_n0 = root_window.x().step(); + const int32_t dst_m0 = root_window.y().step(); + + // Destination compute block size left-over + const int32_t dst_n0_partial = _dst->dimension(0) % dst_n0; + const int32_t dst_m0_partial = _dst->dimension(1) % dst_m0; + + // Shift-back for the overlapping-min strategy + const int32_t dst_shift_back = (dst_n0 - dst_n0_partial) % dst_n0; + + ckw::TensorSampler sampler_dst; + sampler_dst.format(ckw::TensorSamplerFormat::Dim0_Dim1_Dim2); + if (dst_n0_partial == 0) + { + sampler_dst.address_mode_x(ckw::TensorSamplerAddressModeX::None); + } + else + { + sampler_dst.address_mode_x(ckw::TensorSamplerAddressModeX::OverlappingMin); + } + + if (dst_m0_partial == 0) + { + sampler_dst.address_mode_y(ckw::TensorSamplerAddressModeY::None); + } + else + { + sampler_dst.address_mode_y(ckw::TensorSamplerAddressModeY::ClampToBorderMaxOnly); + } + + sampler_dst.address_mode_z(ckw::TensorSamplerAddressModeZ::None); + sampler_dst.storage(ckw::TensorStorageType::BufferUint8Ptr); + + // Declare destination tile + auto tile_dst = writer->declare_tile("dst", ckw::TileInfo(dst_dt, dst_m0, dst_n0)); + + // Initialize destination tile + writer->op_assign(tile_dst, const_0_fp); + + // Bind tile to the tensor + dst->init_virtual_tensor(tile_dst, sampler_dst); + + /******************************************************************************** + * 4 - Define the compute block parameters CKW constants + ********************************************************************************/ + // Only now we can declare the N0 and M0 as constant + auto const_dst_n0_i32 = writer->declare_constant_tile(ckw::ConstantData({{dst_n0}}, ckw::DataType::Int32)); + auto const_dst_m0_i32 = writer->declare_constant_tile(ckw::ConstantData({{dst_m0}}, ckw::DataType::Int32)); + auto const_shift_back_dst_n0_i32 = + writer->declare_constant_tile(ckw::ConstantData({{dst_shift_back}}, ckw::DataType::Int32)); + + /******************************************************************************** + * 5 - Define the samplers for the input tensors + ********************************************************************************/ + // LHS SAMPLER + // The assumption here is that M is multiple of M0. This limitation will be removed once + // we have the support for OverlappingMin as address mode for the Y direction + ckw::TensorSampler sampler_lhs; + sampler_lhs.format(ckw::TensorSamplerFormat::Dim0_Dim1_Dim2); + sampler_lhs.address_mode_x(ckw::TensorSamplerAddressModeX::None); + sampler_lhs.address_mode_y(ckw::TensorSamplerAddressModeY::None); + sampler_lhs.address_mode_z(ckw::TensorSamplerAddressModeZ::None); + sampler_lhs.storage(ckw::TensorStorageType::BufferUint8Ptr); + + // RHS SAMPLER + ckw::TensorSampler sampler_rhs; + sampler_rhs.format(ckw::TensorSamplerFormat::Dim0_Dim1_Dim2); + sampler_rhs.address_mode_x(ckw::TensorSamplerAddressModeX::None); + sampler_rhs.address_mode_y(ckw::TensorSamplerAddressModeY::None); + sampler_rhs.address_mode_z(ckw::TensorSamplerAddressModeZ::None); + sampler_rhs.storage(ckw::TensorStorageType::BufferUint8Ptr); + + /******************************************************************************** + * 6 - Extra operations required before writing the main code (optional) + ********************************************************************************/ + + // Not required + + /******************************************************************************** + * 7 - Get the coordinates of the destination tile + ********************************************************************************/ + auto tile_gid_0 = writer->declare_tile("gid_0", ckw::TileInfo(ckw::DataType::Int32)); + auto tile_gid_1 = writer->declare_tile("gid_1", ckw::TileInfo(ckw::DataType::Int32)); + auto tile_gid_2 = writer->declare_tile("gid_2", ckw::TileInfo(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_idx_n = writer->declare_tile("idx_n", ckw::TileInfo(ckw::DataType::Int32)); // N index + auto tile_idx_m = writer->declare_tile("idx_m", ckw::TileInfo(ckw::DataType::Int32)); // M index + auto tile_idx_b = writer->declare_tile("idx_b", ckw::TileInfo(ckw::DataType::Int32)); // BATCH index + + // Calculate coordinates + get_coordinate_from_gws_overlapping_min(writer, tile_idx_n, tile_gid_0, const_dst_n0_i32, + const_shift_back_dst_n0_i32, const_0_i32); + get_coordinate_from_gws(writer, tile_idx_m, tile_gid_1, const_dst_m0_i32); + get_coordinate_from_gws(writer, tile_idx_b, tile_gid_2, const_pos_1_i32); + + /******************************************************************************** + * 8 - Write the rest of the code + ********************************************************************************/ + auto tile_idx_k = writer->declare_tile("idx_k", ckw::TileInfo(ckw::DataType::Int32)); // K index + + writer->op_assign(tile_idx_k, const_0_i32); + + // clang-format off + writer->op_for_loop(tile_idx_k, ckw::BinaryOp::LessEqual, const_k_minus_k0_i32, tile_idx_k, ckw::AssignmentOp::Increment, const_k0_i32, + [&]() + { + auto tile_lhs = writer->declare_tile("lhs", ckw::TileInfo(to_ckw(_lhs->data_type()), dst_m0, k0)); + auto tile_rhs = writer->declare_tile("rhs", ckw::TileInfo(to_ckw(_rhs->data_type()), dst_n0, k0)); + writer->op_assign(tile_lhs, const_0_fp); + writer->op_assign(tile_rhs, const_0_fp); + + writer->op_load(tile_lhs, lhs->tensor(), sampler_lhs, tile_idx_k, tile_idx_m, tile_idx_b, const_0_i32); + writer->op_load(tile_rhs, rhs->tensor(), sampler_rhs, tile_idx_k, tile_idx_n, tile_idx_b, const_0_i32); + + writer->op_binary(tile_dst, ckw::BinaryOp::MatMul_Nt_T, tile_lhs, tile_rhs); + + }); + + // Left-over accumulations for when K is not a multiple of k0 + if(((k % k0) != 0)) + { + writer->op_for_loop(tile_idx_k, ckw::BinaryOp::Less, const_k_i32, tile_idx_k, ckw::AssignmentOp::Increment, const_pos_1_i32, [&]() + { + auto tile_lhs = writer->declare_tile("lhs", ckw::TileInfo(to_ckw(_lhs->data_type()), dst_m0, 1)); + auto tile_rhs = writer->declare_tile("rhs", ckw::TileInfo(to_ckw(_rhs->data_type()), dst_n0, 1)); + writer->op_assign(tile_lhs, const_0_fp); + writer->op_assign(tile_rhs, const_0_fp); + + writer->op_load(tile_lhs, lhs->tensor(), sampler_lhs, tile_idx_k, tile_idx_m, tile_idx_b, const_0_i32); + writer->op_load(tile_rhs, rhs->tensor(), sampler_rhs, tile_idx_k, tile_idx_n, tile_idx_b, const_0_i32); + + writer->op_binary(tile_dst, ckw::BinaryOp::MatMul_Nt_T, tile_lhs, tile_rhs); + }); + } + // clang-format on +} + +Window GpuCkwMatMul::get_window() const +{ + ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized"); + + const int32_t m = _dst->dimension(1); + const int32_t n = _dst->dimension(0); + const bool adj_lhs = _attributes.adj_lhs(); + + const int32_t m0 = adj_lhs ? adjust_vec_size(_settings.m0(), m) : std::min(_settings.m0(), m); + const int32_t n0 = adjust_vec_size(_settings.n0(), n); + + // Configure kernel window + Window win = calculate_max_window(_dst->tensor_shape(), Steps(n0, m0)); + win = win.collapse(win, Window::DimZ); + + return win; +} + +std::string GpuCkwMatMul::get_name(const ComponentGroup &comp_group) const +{ + ARM_COMPUTE_UNUSED(comp_group); + + std::string kernel_name("mat_mul_native"); + + const int32_t m = _dst->dimension(1); + const int32_t n = _dst->dimension(0); + const int32_t k = _attributes.adj_lhs() ? _lhs->tensor_shape().y() : _lhs->tensor_shape().x(); + + kernel_name += _attributes.adj_lhs() ? "_t" : "_nt"; + kernel_name += _attributes.adj_rhs() ? "_t" : "_nt"; + kernel_name += "_"; + kernel_name += support::cpp11::to_string(m); + kernel_name += "_"; + kernel_name += support::cpp11::to_string(n); + kernel_name += "_"; + kernel_name += support::cpp11::to_string(k); + kernel_name += "_"; + kernel_name += support::cpp11::to_string(_dst->dimension(2)); + kernel_name += "_"; + kernel_name += support::cpp11::to_string(_settings.m0()); + kernel_name += "_"; + kernel_name += support::cpp11::to_string(_settings.n0()); + kernel_name += "_"; + kernel_name += support::cpp11::to_string(_settings.k0()); + + return kernel_name; +} + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute |