1 // Copyright 2015 Google Inc. All Rights Reserved. 2 // 3 // Licensed under the Apache License, Version 2.0 (the "License"); 4 // you may not use this file except in compliance with the License. 5 // You may obtain a copy of the License at 6 // 7 // http://www.apache.org/licenses/LICENSE-2.0 8 // 9 // Unless required by applicable law or agreed to in writing, software 10 // distributed under the License is distributed on an "AS IS" BASIS, 11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 // See the License for the specific language governing permissions and 13 // limitations under the License. 14 15 // compute.h: the central stage of the Gemm computation, operates 16 // on already-packed LHS and RHS blocks and calls the Gemm kernel 17 // to compute a block of the product. 18 19 #ifndef GEMMLOWP_INTERNAL_COMPUTE_H_ 20 #define GEMMLOWP_INTERNAL_COMPUTE_H_ 21 22 #include "block_params.h" 23 #include "kernel.h" 24 #include "pack.h" 25 26 namespace gemmlowp { 27 28 template <typename PackedLhs, typename PackedRhs, typename PackedResult> 29 class ComputeImpl { 30 typedef typename PackedLhs::KernelSideFormat KernelLhsFormat; 31 typedef typename PackedRhs::KernelSideFormat KernelRhsFormat; 32 typedef KernelFormat<KernelLhsFormat, KernelRhsFormat> Format; 33 34 const KernelBase& kernel_; 35 const BlockParams& block_params_; 36 37 PackedResult* const packed_result_; 38 const PackedLhs& packed_lhs_; 39 const PackedRhs& packed_rhs_; 40 41 public: 42 ComputeImpl(const KernelBase& _kernel, const BlockParams& _block_params, 43 PackedResult* _packed_result, const PackedLhs& _packed_lhs, 44 const PackedRhs& _packed_rhs) 45 : kernel_(_kernel), 46 block_params_(_block_params), 47 packed_result_(_packed_result), 48 packed_lhs_(_packed_lhs), 49 packed_rhs_(_packed_rhs) {} 50 51 void Compute() { 52 for (int d = 0; d < block_params_.l2_depth; d += block_params_.l1_depth) { 53 int ds = std::min(block_params_.l1_depth, block_params_.l2_depth - d); 54 55 for (int r = 0; r < block_params_.l2_rows; r += block_params_.l1_rows) { 56 int rs = std::min(block_params_.l1_rows, block_params_.l2_rows - r); 57 58 ComputeL1(r, rs, 0, block_params_.l2_cols, d, ds); 59 } 60 } 61 } 62 63 private: 64 void ComputeRun(int start_row, int start_col, int start_depth, 65 int depth) GEMMLOWP_NOINLINE { 66 packed_lhs_.seek_run(start_row, start_depth); 67 packed_rhs_.seek_run(start_col, start_depth); 68 auto packed_result_block = packed_result_->Map().block( 69 start_row, start_col, Format::kRows, Format::kCols); 70 kernel_.Run(packed_result_block.data(), packed_result_block.rows_stride(), 71 packed_result_block.cols_stride(), packed_lhs_.current_data(), 72 packed_rhs_.current_data(), start_depth, depth); 73 } 74 75 void ComputeL1(int start_row, int rows, int start_col, int cols, 76 int start_depth, int depth) { 77 assert(rows % Format::kRows == 0); 78 assert(cols % Format::kCols == 0); 79 assert(depth % Format::kDepth == 0); 80 81 for (int c = 0; c < cols; c += Format::kCols) { 82 for (int r = 0; r < rows; r += Format::kRows) { 83 ComputeRun(start_row + r, start_col + c, start_depth, depth); 84 } 85 } 86 } 87 }; 88 89 template <typename PackedLhs, typename PackedRhs, typename PackedResult> 90 void Compute(const KernelBase& kernel, const BlockParams& block_params, 91 PackedResult* packed_result, const PackedLhs& packed_lhs, 92 const PackedRhs& packed_rhs) { 93 ScopedProfilingLabel label("compute"); 94 ComputeImpl<PackedLhs, PackedRhs, PackedResult> impl( 95 kernel, block_params, packed_result, packed_lhs, packed_rhs); 96 97 impl.Compute(); 98 } 99 100 } // namespace gemmlowp 101 102 #endif // GEMMLOWP_INTERNAL_COMPUTE_H_ 103
