Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add_sample_logits_op
	
		
	
				
					
				
			test=developrevert-15774-anakin_subgraph_engine
						commit
						d12252e6a6
					
				
											
												
													File diff suppressed because one or more lines are too long
												
											
										
									
								@ -0,0 +1,149 @@
 | 
				
			||||
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
 | 
				
			||||
 *
 | 
				
			||||
 * Licensed under the Apache License, Version 2.0 (the "License");
 | 
				
			||||
 * you may not use this file except in compliance with the License.
 | 
				
			||||
 * You may obtain a copy of the License at
 | 
				
			||||
 *
 | 
				
			||||
 * http://www.apache.org/licenses/LICENSE-2.0
 | 
				
			||||
 *
 | 
				
			||||
 * Unless required by applicable law or agreed to in writing, software
 | 
				
			||||
 * distributed under the License is distributed on an "AS IS" BASIS,
 | 
				
			||||
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 | 
				
			||||
 * See the License for the specific language governing permissions and
 | 
				
			||||
 * limitations under the License. */
 | 
				
			||||
 | 
				
			||||
#include "paddle/fluid/operators/jit/gen/embseqpool.h"
 | 
				
			||||
#include <stddef.h>  // offsetof
 | 
				
			||||
#include <vector>
 | 
				
			||||
#include "paddle/fluid/operators/jit/gen/act.h"  // for exp_float_consts ones
 | 
				
			||||
#include "paddle/fluid/operators/jit/registry.h"
 | 
				
			||||
#include "paddle/fluid/platform/cpu_info.h"
 | 
				
			||||
 | 
				
			||||
namespace paddle {
 | 
				
			||||
namespace operators {
 | 
				
			||||
namespace jit {
 | 
				
			||||
namespace gen {
 | 
				
			||||
 | 
				
			||||
void EmbSeqPoolJitCode::genCode() {
 | 
				
			||||
  preCode();
 | 
				
			||||
  constexpr int block = YMM_FLOAT_BLOCK;
 | 
				
			||||
  constexpr int max_num_regs = 8;
 | 
				
			||||
  const int num_block = tbl_w_ / block;
 | 
				
			||||
  const int num_groups = num_block / max_num_regs;
 | 
				
			||||
  const size_t block_size = sizeof(float) * block;
 | 
				
			||||
  std::vector<int> groups(num_groups, max_num_regs);
 | 
				
			||||
  int rest_num_regs = num_block % max_num_regs;
 | 
				
			||||
  if (rest_num_regs > 0) {
 | 
				
			||||
    groups.push_back(rest_num_regs);
 | 
				
			||||
  }
 | 
				
			||||
 | 
				
			||||
  // protect param_dst
 | 
				
			||||
  mov(reg_ptr_param_dst, param_dst);
 | 
				
			||||
  mov(reg_idx_width_in_byte,
 | 
				
			||||
      qword[param_attr + offsetof(emb_seq_pool_attr_t, index_width)]);
 | 
				
			||||
  mov(reg_idx_height,
 | 
				
			||||
      qword[param_attr + offsetof(emb_seq_pool_attr_t, index_height)]);
 | 
				
			||||
  mov(rax, sizeof(int64_t));
 | 
				
			||||
  mul(reg_idx_width_in_byte);
 | 
				
			||||
  mov(reg_idx_width_in_byte, rax);
 | 
				
			||||
  const size_t tbl_width_in_byte = sizeof(float) * tbl_w_;
 | 
				
			||||
  int acc_num_regs = 0;
 | 
				
			||||
  for (int num_regs : groups) {
 | 
				
			||||
    Label l_next_idx_w, l_next_idx_h, l_save_now;
 | 
				
			||||
    xor_(reg_idx_w_i_in_byte, reg_idx_w_i_in_byte);
 | 
				
			||||
    mov(reg_ptr_dst_i, reg_ptr_param_dst);
 | 
				
			||||
    add(reg_ptr_dst_i, acc_num_regs * block_size);
 | 
				
			||||
 | 
				
			||||
    L(l_next_idx_w);
 | 
				
			||||
    {
 | 
				
			||||
      // h == 0
 | 
				
			||||
      mov(reg_ptr_idx_i, param_idx);
 | 
				
			||||
      add(reg_ptr_idx_i, reg_idx_w_i_in_byte);
 | 
				
			||||
      mov(reg_idx, qword[reg_ptr_idx_i]);
 | 
				
			||||
      mov(rax, tbl_width_in_byte);
 | 
				
			||||
      mul(reg_idx);
 | 
				
			||||
      mov(reg_ptr_tbl_i, rax);        // reg is offset now
 | 
				
			||||
      add(reg_ptr_tbl_i, param_tbl);  // reg is ptr_i now
 | 
				
			||||
      size_t w_offset = 0;
 | 
				
			||||
      for (int reg_i = 0; reg_i < num_regs; ++reg_i) {
 | 
				
			||||
        vmovups(ymm_t(reg_i + num_regs), ptr[reg_ptr_tbl_i + w_offset]);
 | 
				
			||||
        w_offset += block_size;
 | 
				
			||||
      }
 | 
				
			||||
      add(reg_ptr_idx_i, reg_idx_width_in_byte);
 | 
				
			||||
 | 
				
			||||
      // end condition of idx h
 | 
				
			||||
      mov(reg_idx_h_end, reg_idx_height);
 | 
				
			||||
      mov(rax, reg_idx_width_in_byte);
 | 
				
			||||
      mul(reg_idx_h_end);
 | 
				
			||||
      mov(reg_idx_h_end, rax);
 | 
				
			||||
      add(reg_idx_h_end, reg_idx_w_i_in_byte);
 | 
				
			||||
      add(reg_idx_h_end, param_idx);
 | 
				
			||||
 | 
				
			||||
      cmp(reg_ptr_idx_i, reg_idx_h_end);
 | 
				
			||||
      jge(l_save_now, T_NEAR);
 | 
				
			||||
      L(l_next_idx_h);
 | 
				
			||||
      {
 | 
				
			||||
        mov(reg_idx, qword[reg_ptr_idx_i]);
 | 
				
			||||
        mov(reg_ptr_tbl_i, reg_idx);
 | 
				
			||||
        mov(rax, tbl_width_in_byte);
 | 
				
			||||
        mul(reg_idx);
 | 
				
			||||
        mov(reg_ptr_tbl_i, rax);
 | 
				
			||||
        add(reg_ptr_tbl_i, param_tbl);
 | 
				
			||||
        size_t w_offset = 0;
 | 
				
			||||
        for (int reg_i = 0; reg_i < num_regs; ++reg_i) {
 | 
				
			||||
          vmovups(ymm_t(reg_i), ptr[reg_ptr_tbl_i + w_offset]);
 | 
				
			||||
          vaddps(ymm_t(reg_i + num_regs), ymm_t(reg_i + num_regs),
 | 
				
			||||
                 ymm_t(reg_i));
 | 
				
			||||
          w_offset += block_size;
 | 
				
			||||
        }
 | 
				
			||||
        add(reg_ptr_idx_i, reg_idx_width_in_byte);
 | 
				
			||||
        cmp(reg_ptr_idx_i, reg_idx_h_end);
 | 
				
			||||
        jl(l_next_idx_h, T_NEAR);
 | 
				
			||||
      }  // end of idx h
 | 
				
			||||
      L(l_save_now);
 | 
				
			||||
      // avg or sqrt here, if needed
 | 
				
			||||
      w_offset = 0;
 | 
				
			||||
      for (int reg_i = 0; reg_i < num_regs; ++reg_i) {
 | 
				
			||||
        vmovups(ptr[reg_ptr_dst_i + w_offset], ymm_t(reg_i + num_regs));
 | 
				
			||||
        w_offset += block_size;
 | 
				
			||||
      }
 | 
				
			||||
      add(reg_ptr_dst_i, tbl_width_in_byte);
 | 
				
			||||
      add(reg_idx_w_i_in_byte, sizeof(int64_t));
 | 
				
			||||
      cmp(reg_idx_w_i_in_byte, reg_idx_width_in_byte);
 | 
				
			||||
      jl(l_next_idx_w, T_NEAR);
 | 
				
			||||
    }  // end of idx w
 | 
				
			||||
 | 
				
			||||
    acc_num_regs += num_regs;
 | 
				
			||||
    add(param_tbl, num_regs * block_size);  // do not use acc_num_regs
 | 
				
			||||
  }                                         // end of groups
 | 
				
			||||
  postCode();
 | 
				
			||||
}
 | 
				
			||||
 | 
				
			||||
class EmbSeqPoolCreator : public JitCodeCreator<emb_seq_pool_attr_t> {
 | 
				
			||||
 public:
 | 
				
			||||
  bool UseMe(const emb_seq_pool_attr_t& attr) const override {
 | 
				
			||||
    return platform::MayIUse(platform::avx) &&
 | 
				
			||||
           attr.table_width % YMM_FLOAT_BLOCK == 0;
 | 
				
			||||
  }
 | 
				
			||||
  size_t CodeSize(const emb_seq_pool_attr_t& attr) const override {
 | 
				
			||||
    return 96 + (attr.table_width / YMM_FLOAT_BLOCK) * 96 * 8;
 | 
				
			||||
  }
 | 
				
			||||
  std::unique_ptr<GenBase> CreateJitCode(
 | 
				
			||||
      const emb_seq_pool_attr_t& attr) const override {
 | 
				
			||||
    PADDLE_ENFORCE_GT(attr.table_height, 0);
 | 
				
			||||
    PADDLE_ENFORCE_GT(attr.table_width, 0);
 | 
				
			||||
    PADDLE_ENFORCE_GT(attr.index_height, 0);
 | 
				
			||||
    PADDLE_ENFORCE_GT(attr.index_width, 0);
 | 
				
			||||
    PADDLE_ENFORCE_GT(attr.out_width, 0);
 | 
				
			||||
    return make_unique<EmbSeqPoolJitCode>(attr, CodeSize(attr));
 | 
				
			||||
  }
 | 
				
			||||
};
 | 
				
			||||
 | 
				
			||||
}  // namespace gen
 | 
				
			||||
}  // namespace jit
 | 
				
			||||
}  // namespace operators
 | 
				
			||||
}  // namespace paddle
 | 
				
			||||
 | 
				
			||||
namespace gen = paddle::operators::jit::gen;
 | 
				
			||||
 | 
				
			||||
REGISTER_JITKERNEL_GEN(kEmbSeqPool, gen::EmbSeqPoolCreator);
 | 
				
			||||
@ -0,0 +1,81 @@
 | 
				
			||||
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
 | 
				
			||||
 *
 | 
				
			||||
 * Licensed under the Apache License, Version 2.0 (the "License");
 | 
				
			||||
 * you may not use this file except in compliance with the License.
 | 
				
			||||
 * You may obtain a copy of the License at
 | 
				
			||||
 *
 | 
				
			||||
 * http://www.apache.org/licenses/LICENSE-2.0
 | 
				
			||||
 *
 | 
				
			||||
 * Unless required by applicable law or agreed to in writing, software
 | 
				
			||||
 * distributed under the License is distributed on an "AS IS" BASIS,
 | 
				
			||||
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 | 
				
			||||
 * See the License for the specific language governing permissions and
 | 
				
			||||
 * limitations under the License. */
 | 
				
			||||
 | 
				
			||||
#pragma once
 | 
				
			||||
 | 
				
			||||
#include <string>
 | 
				
			||||
#include "glog/logging.h"
 | 
				
			||||
#include "paddle/fluid/operators/jit/gen/jitcode.h"
 | 
				
			||||
#include "paddle/fluid/platform/enforce.h"
 | 
				
			||||
 | 
				
			||||
namespace paddle {
 | 
				
			||||
namespace operators {
 | 
				
			||||
namespace jit {
 | 
				
			||||
namespace gen {
 | 
				
			||||
 | 
				
			||||
class EmbSeqPoolJitCode : public JitCode {
 | 
				
			||||
 public:
 | 
				
			||||
  explicit EmbSeqPoolJitCode(const emb_seq_pool_attr_t& attr,
 | 
				
			||||
                             size_t code_size = 256 * 1024,
 | 
				
			||||
                             void* code_ptr = nullptr)
 | 
				
			||||
      : JitCode(code_size, code_ptr),
 | 
				
			||||
        tbl_w_(attr.table_width),
 | 
				
			||||
        type_(attr.pool_type) {
 | 
				
			||||
    if (type_ != SeqPoolType::kSum) {
 | 
				
			||||
      LOG(FATAL) << "Only support sum pool yet ";
 | 
				
			||||
    }
 | 
				
			||||
    this->genCode();
 | 
				
			||||
  }
 | 
				
			||||
 | 
				
			||||
  std::string name() const override {
 | 
				
			||||
    std::string base = "EmbSeqPoolJitCode";
 | 
				
			||||
    if (type_ == SeqPoolType::kSum) {
 | 
				
			||||
      base += "_Sum";
 | 
				
			||||
    } else if (type_ == SeqPoolType::kAvg) {
 | 
				
			||||
      base += "_Avg";
 | 
				
			||||
    } else if (type_ == SeqPoolType::kSqrt) {
 | 
				
			||||
      base += "_Sqrt";
 | 
				
			||||
    }
 | 
				
			||||
    base += ("_W" + std::to_string(tbl_w_));
 | 
				
			||||
    return base;
 | 
				
			||||
  }
 | 
				
			||||
  void genCode() override;
 | 
				
			||||
 | 
				
			||||
 private:
 | 
				
			||||
  int tbl_w_;
 | 
				
			||||
  SeqPoolType type_;
 | 
				
			||||
  reg64_t param_tbl{abi_param1};
 | 
				
			||||
  reg64_t param_idx{abi_param2};
 | 
				
			||||
  reg64_t param_dst{abi_param3};
 | 
				
			||||
  reg64_t param_attr{abi_param4};
 | 
				
			||||
 | 
				
			||||
  reg64_t reg_tmp{rax};
 | 
				
			||||
 | 
				
			||||
  reg64_t reg_idx_width_in_byte{r8};
 | 
				
			||||
  reg64_t reg_idx_height{r9};
 | 
				
			||||
 | 
				
			||||
  reg64_t reg_ptr_tbl_i{r10};
 | 
				
			||||
  reg64_t reg_idx{r10};  // could use same of reg_ptr_tbl_i
 | 
				
			||||
  reg64_t reg_ptr_idx_i{r11};
 | 
				
			||||
  reg64_t reg_ptr_dst_i{r12};
 | 
				
			||||
  reg64_t reg_ptr_param_dst{r13};  // rdx is used in mul so protect param_dst
 | 
				
			||||
 | 
				
			||||
  reg64_t reg_idx_w_i_in_byte{r14};
 | 
				
			||||
  reg64_t reg_idx_h_end{r15};
 | 
				
			||||
};
 | 
				
			||||
 | 
				
			||||
}  // namespace gen
 | 
				
			||||
}  // namespace jit
 | 
				
			||||
}  // namespace operators
 | 
				
			||||
}  // namespace paddle
 | 
				
			||||
Some files were not shown because too many files have changed in this diff Show More
					Loading…
					
					
				
		Reference in new issue