-
Notifications
You must be signed in to change notification settings - Fork 2.4k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[Snippets][CPU] Implemented BrgemmCopyB specific buffers
- Loading branch information
1 parent
7b790a6
commit 4d607ea
Showing
10 changed files
with
237 additions
and
131 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
140 changes: 140 additions & 0 deletions
140
.../intel_cpu/src/transformations/snippets/x64/pass/lowered/insert_brgemm_copy_b_buffers.cpp
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,140 @@ | ||
| // Copyright (C) 2018-2024 Intel Corporation | ||
| // SPDX-License-Identifier: Apache-2.0 | ||
| // | ||
|
|
||
| #include "insert_brgemm_copy_b_buffers.hpp" | ||
|
|
||
| #include "snippets/lowered/loop_manager.hpp" | ||
| #include "snippets/snippets_isa.hpp" | ||
| #include "snippets/utils/utils.hpp" | ||
| #include "snippets/itt.hpp" | ||
|
|
||
| #include "transformations/snippets/x64/op/brgemm_copy_b.hpp" | ||
| #include "utils/general_utils.h" | ||
|
|
||
|
|
||
| using namespace ov::intel_cpu::brgemm_utils::repacking; | ||
| using namespace ov::snippets::lowered; | ||
|
|
||
| namespace ov { | ||
| namespace intel_cpu { | ||
| namespace pass { | ||
|
|
||
| bool InsertBrgemmCopyBBuffers::run(LinearIR& linear_ir, LinearIR::constExprIt begin, LinearIR::constExprIt end) { | ||
| OV_ITT_SCOPED_TASK(ov::pass::itt::domains::SnippetsTransform, "Snippets::InsertBrgemmCopyBBuffers") | ||
|
|
||
| const auto& factory = linear_ir.get_expr_factory(); | ||
|
|
||
| auto insert_buffer = [&](const ExpressionPtr& copy_b_expr, size_t out_port, LinearIR::constExprIt insertion_pos) { | ||
| const auto& copy_b = ov::as_type_ptr<ov::intel_cpu::BrgemmCopyB>(copy_b_expr->get_node()); | ||
| const auto& copy_b_out = copy_b_expr->get_output_port_connector(out_port); | ||
| const auto copy_b_consumers = copy_b_out->get_consumers(); | ||
| OPENVINO_ASSERT(copy_b_consumers.size() == 1, "BufferCopyB must have only one consumer on each out port - Brgemm"); | ||
| const auto& buffer_op = std::make_shared<ov::snippets::op::Buffer>(copy_b->output(out_port)); | ||
| BufferExpressionPtr buffer_expr = nullptr; | ||
| if (out_port == 0) { | ||
| buffer_expr = factory->build<RepackedWeightsBufferExpression>(buffer_op, {copy_b_out}); | ||
| } else if (out_port == 1 && with_compensations(copy_b->get_type())) { | ||
| buffer_expr = factory->build<CompensationsBufferExpression>(buffer_op, {copy_b_out}); | ||
| } else { | ||
| OPENVINO_THROW("BrgemmCopyB has incorrect output ports"); | ||
| } | ||
| return linear_ir.insert_expr(buffer_expr, LoopManager::get_common_outer_loops(copy_b_expr, copy_b_consumers.begin()->get_expr()), | ||
| true, insertion_pos, {copy_b_consumers}); | ||
| }; | ||
|
|
||
| bool modified = false; | ||
| for (auto expr_it = begin; expr_it != end; ++expr_it) { | ||
| const auto expr = *expr_it; | ||
| if (auto copy_b = ov::as_type_ptr<ov::intel_cpu::BrgemmCopyB>(expr->get_node())) { | ||
| for (size_t i = 0; i < expr->get_output_count(); ++i) { | ||
| expr_it = insert_buffer(expr, i, std::next(expr_it)); // we insert after expression | ||
| } | ||
| modified = true; | ||
| } | ||
| } | ||
| return modified; | ||
| } | ||
|
|
||
| InsertBrgemmCopyBBuffers::RepackedWeightsBufferExpression::RepackedWeightsBufferExpression(const snippets::lowered::BufferExpression& other) | ||
| : BufferExpression(other) {} | ||
|
|
||
| InsertBrgemmCopyBBuffers::RepackedWeightsBufferExpression::RepackedWeightsBufferExpression(const std::shared_ptr<ov::Node>& n, | ||
| const std::shared_ptr<snippets::IShapeInferSnippetsFactory>& factory) : BufferExpression(n, factory) {} | ||
|
|
||
| snippets::lowered::ExpressionPtr InsertBrgemmCopyBBuffers::RepackedWeightsBufferExpression::clone() const { | ||
| return std::shared_ptr<RepackedWeightsBufferExpression>(new RepackedWeightsBufferExpression(*this)); | ||
| } | ||
|
|
||
| void InsertBrgemmCopyBBuffers::RepackedWeightsBufferExpression::validate() const { | ||
| BufferExpression::validate(); | ||
| OPENVINO_ASSERT(get_input_count() == 1, "RepackedWeightsBufferExpression must have only one input"); | ||
| const auto& parent_out = get_input_port_connector(0)->get_source(); | ||
| OPENVINO_ASSERT(ov::is_type<ov::intel_cpu::BrgemmCopyB>(parent_out.get_expr()->get_node()) && parent_out.get_index() == 0, | ||
| "RepackedWeightsBufferExpression expects BrgemmCopyB as parent expression"); | ||
| } | ||
|
|
||
| void InsertBrgemmCopyBBuffers::RepackedWeightsBufferExpression::init_allocation_size(const std::shared_ptr<snippets::lowered::LoopManager>& loop_manager, | ||
| size_t allocation_rank) { | ||
| const auto& parent_expr = get_input_port_connector(0)->get_source().get_expr(); | ||
| const auto& in_layout = parent_expr->get_input_port_descriptor(0)->get_layout(); | ||
| const auto& in_subtensor = ov::snippets::utils::get_projected_subtensor(parent_expr->get_input_port(0)); | ||
|
|
||
| const size_t n_blk = *in_subtensor.rbegin(); | ||
| const size_t k_blk = *++in_subtensor.rbegin(); | ||
| OPENVINO_ASSERT(!ov::snippets::utils::is_dynamic_value(n_blk) && !ov::snippets::utils::is_dynamic_value(k_blk), | ||
| "RepackedWeightsBufferExpression supports only static subtensor values"); | ||
|
|
||
| const auto& precision = get_node()->get_input_element_type(0); | ||
| // Repacking buffer shape is set in accordance to OneDNN requirements | ||
| const size_t N_dim = std::max(n_blk, compute_inner_n_block(precision)); | ||
| if (!in_layout.empty() && in_layout.back() != in_layout.size() - 1) { | ||
| // In case of transpose, K dimension must be rounded-up to number of elems in vector register | ||
| // For the details, please see 'transpose16x8' and 'fixup16x16' implementations and usage in onednn/src/cpu/x64/matmul/brgemm_matmul_copy_utils.cpp | ||
| const auto elems_in_vec = brgemm_utils::get_elems_in_vec(precision); | ||
| m_allocation_size = N_dim * rnd_up(k_blk, elems_in_vec); | ||
| } else { | ||
| // Low precision repacking writes the result by m_brgemmVNNIFactor * m_inner_n_block blocks | ||
| // despite the actual size of the input data. Because of that we have to round-up the allocation shape to always have enough memory allocated. | ||
| // For the details, please see 'copy_4x64' and 'copy_2x32' implementations and usage in onednn/src/cpu/x64/matmul/brgemm_matmul_copy_utils.cpp | ||
| const auto brgemmVNNIFactor = brgemm_utils::compute_vnni_factor(precision); | ||
| OPENVINO_ASSERT(brgemmVNNIFactor > 0, "brgemmVNNIFactor value must be positive."); | ||
| m_allocation_size = N_dim * rnd_up(k_blk, brgemmVNNIFactor); | ||
| } | ||
| } | ||
|
|
||
| InsertBrgemmCopyBBuffers::CompensationsBufferExpression::CompensationsBufferExpression(const snippets::lowered::BufferExpression& other) | ||
| : BufferExpression(other) {} | ||
|
|
||
| InsertBrgemmCopyBBuffers::CompensationsBufferExpression::CompensationsBufferExpression(const std::shared_ptr<ov::Node>& n, | ||
| const std::shared_ptr<snippets::IShapeInferSnippetsFactory>& factory) : BufferExpression(n, factory) {} | ||
|
|
||
| snippets::lowered::ExpressionPtr InsertBrgemmCopyBBuffers::CompensationsBufferExpression::clone() const { | ||
| return std::shared_ptr<CompensationsBufferExpression>(new CompensationsBufferExpression(*this)); | ||
| } | ||
|
|
||
| void InsertBrgemmCopyBBuffers::CompensationsBufferExpression::validate() const { | ||
| BufferExpression::validate(); | ||
| OPENVINO_ASSERT(get_input_count() == 1, "CompensationsBufferExpression must have only one input"); | ||
| const auto& parent_out = get_input_port_connector(0)->get_source(); | ||
| OPENVINO_ASSERT(ov::is_type<ov::intel_cpu::BrgemmCopyB>(parent_out.get_expr()->get_node()) && parent_out.get_index() == 1, | ||
| "CompensationsBufferExpression expects BrgemmCopyB as parent expression"); | ||
| } | ||
|
|
||
| void InsertBrgemmCopyBBuffers::CompensationsBufferExpression::init_allocation_size(const std::shared_ptr<snippets::lowered::LoopManager>& loop_manager, | ||
| size_t allocation_rank) { | ||
| const auto& parent_expr = get_input_port_connector(0)->get_source().get_expr(); | ||
| const auto& in_subtensor = ov::snippets::utils::get_projected_subtensor(parent_expr->get_input_port(0)); | ||
| const size_t n_blk = *in_subtensor.rbegin(); | ||
| OPENVINO_ASSERT(!ov::snippets::utils::is_dynamic_value(n_blk), "CompensationsBufferExpression supports only static subtensor values"); | ||
| const auto& precision = parent_expr->get_node()->get_input_element_type(0); | ||
| // Compensations are computed during repacking, so we need to round-up allocation shape according to m_inner_n_block | ||
| // because of OneDNN implementation nuances (as in get_repacking_buffer_size). | ||
| // However, the compensations are computed by N dimension, so K dimension doesn't affect the compensations buffer | ||
| m_allocation_size = std::max(n_blk, compute_inner_n_block(precision)); | ||
| } | ||
|
|
||
| } // namespace pass | ||
| } // namespace intel_cpu | ||
| } // namespace ov | ||
|
|
65 changes: 65 additions & 0 deletions
65
.../intel_cpu/src/transformations/snippets/x64/pass/lowered/insert_brgemm_copy_b_buffers.hpp
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,65 @@ | ||
| // Copyright (C) 2018-2024 Intel Corporation | ||
| // SPDX-License-Identifier: Apache-2.0 | ||
| // | ||
|
|
||
| #pragma once | ||
|
|
||
| #include "snippets/lowered/pass/pass.hpp" | ||
|
|
||
| #include "snippets/lowered/expressions/buffer_expression.hpp" | ||
|
|
||
| namespace ov { | ||
| namespace intel_cpu { | ||
| namespace pass { | ||
|
|
||
| /** | ||
| * @interface InsertBrgemmCopyBBuffers | ||
| * @brief Insert Buffers after BrgemmCopyB with algorithm of allocation size calculation which | ||
| * distinguishes with common algorithm | ||
| * @ingroup snippets | ||
| */ | ||
| class InsertBrgemmCopyBBuffers: public snippets::lowered::pass::RangedPass { | ||
| public: | ||
| InsertBrgemmCopyBBuffers() = default; | ||
| OPENVINO_RTTI("InsertBrgemmCopyBBuffers", "Pass"); | ||
| bool run(snippets::lowered::LinearIR& linear_ir, snippets::lowered::LinearIR::constExprIt begin, snippets::lowered::LinearIR::constExprIt end) override; | ||
|
|
||
| private: | ||
| class RepackedWeightsBufferExpression : public snippets::lowered::BufferExpression { | ||
| friend class snippets::lowered::ExpressionFactory; | ||
| public: | ||
| OPENVINO_RTTI("RepackedWeightsBufferExpression", "0", BufferExpression) | ||
| RepackedWeightsBufferExpression() = default; | ||
|
|
||
| void validate() const override; | ||
|
|
||
| void init_allocation_size(const std::shared_ptr<snippets::lowered::LoopManager>& loop_manager, size_t allocation_rank) override; | ||
|
|
||
| private: | ||
| RepackedWeightsBufferExpression(const snippets::lowered::BufferExpression& other); | ||
| RepackedWeightsBufferExpression(const std::shared_ptr<ov::Node>& n, const std::shared_ptr<snippets::IShapeInferSnippetsFactory>& factory); | ||
|
|
||
| snippets::lowered::ExpressionPtr clone() const override; | ||
| }; | ||
|
|
||
| class CompensationsBufferExpression : public snippets::lowered::BufferExpression { | ||
| friend class snippets::lowered::ExpressionFactory; | ||
| public: | ||
| OPENVINO_RTTI("CompensationsBufferExpression", "0", BufferExpression) | ||
| CompensationsBufferExpression() = default; | ||
|
|
||
| void validate() const override; | ||
|
|
||
| void init_allocation_size(const std::shared_ptr<snippets::lowered::LoopManager>& loop_manager, size_t allocation_rank) override; | ||
|
|
||
| private: | ||
| CompensationsBufferExpression(const snippets::lowered::BufferExpression& other); | ||
| CompensationsBufferExpression(const std::shared_ptr<ov::Node>& n, const std::shared_ptr<snippets::IShapeInferSnippetsFactory>& factory); | ||
|
|
||
| snippets::lowered::ExpressionPtr clone() const override; | ||
| }; | ||
| }; | ||
|
|
||
| } // namespace pass | ||
| } // namespace intel_cpu | ||
| } // namespace ov |
Oops, something went wrong.