jit_code.cc 3.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
/* 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/math/jit_code.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
#include "paddle/fluid/platform/cpu_info.h"

namespace paddle {
namespace operators {
namespace math {
namespace jitkernel {
namespace gen {

using namespace platform::jit;  // NOLINT

bool VMulJitCode::init(int d) {
28 29
  // It's not necessary to use avx512 since it would slow down the frequency
  // and this kernel is not compute bound.
T
tensor-tang 已提交
30
  return MayIUse(avx);
31 32 33
}

void VMulJitCode::generate() {
T
tensor-tang 已提交
34
  // do not need push stack, and do not need save avx512reg if do not use avx512
35
  int offset = 0;
36
  for (int i = 0; i < num_ / AVX_FLOAT_BLOCK; ++i) {
37 38
    vmovups(ymm_src1, ptr[param1 + offset]);
    vmovups(ymm_src2, ptr[param2 + offset]);
39
    vmulps(ymm_dst, ymm_src1, ymm_src2);
40 41 42 43 44 45 46 47 48 49 50 51 52
    vmovups(ptr[param3 + offset], ymm_dst);
    offset += sizeof(float) * AVX_FLOAT_BLOCK;
  }
  int rest = num_ % AVX_FLOAT_BLOCK;
  if (rest >= 4) {
    vmovups(xmm_src1, ptr[param1 + offset]);
    vmovups(xmm_src2, ptr[param2 + offset]);
    vmulps(xmm_dst, xmm_src1, xmm_src2);
    vmovups(ptr[param3 + offset], xmm_dst);
    offset += sizeof(float) * 4;
    rest -= 4;
  }
  if (rest >= 2) {
T
tensor-tang 已提交
53 54
    vmovq(xmm_src1, ptr[param1 + offset]);
    vmovq(xmm_src2, ptr[param2 + offset]);
55
    vmulps(xmm_dst, xmm_src1, xmm_src2);
T
tensor-tang 已提交
56
    vmovq(ptr[param3 + offset], xmm_dst);
57 58
    offset += sizeof(float) * 2;
    rest -= 2;
59
  }
T
tensor-tang 已提交
60 61 62 63 64 65
  if (rest > 0) {
    vmovss(xmm_src1, ptr[param1 + offset]);
    vmovss(xmm_src2, ptr[param2 + offset]);
    vmulss(xmm_dst, xmm_src1, xmm_src2);
    vmovss(ptr[param3 + offset], xmm_dst);
  }
T
tensor-tang 已提交
66
  ret();
67 68
}

T
tensor-tang 已提交
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
bool VAddJitCode::init(int d) { return MayIUse(avx); }

void VAddJitCode::generate() {
  int offset = 0;
  for (int i = 0; i < num_ / AVX_FLOAT_BLOCK; ++i) {
    vmovups(ymm_src1, ptr[param1 + offset]);
    vmovups(ymm_src2, ptr[param2 + offset]);
    vaddps(ymm_dst, ymm_src1, ymm_src2);
    vmovups(ptr[param3 + offset], ymm_dst);
    offset += sizeof(float) * AVX_FLOAT_BLOCK;
  }
  int rest = num_ % AVX_FLOAT_BLOCK;
  if (rest >= 4) {
    vmovups(xmm_src1, ptr[param1 + offset]);
    vmovups(xmm_src2, ptr[param2 + offset]);
    vaddps(xmm_dst, xmm_src1, xmm_src2);
    vmovups(ptr[param3 + offset], xmm_dst);
    offset += sizeof(float) * 4;
    rest -= 4;
  }
  if (rest >= 2) {
    vmovq(xmm_src1, ptr[param1 + offset]);
    vmovq(xmm_src2, ptr[param2 + offset]);
    vaddps(xmm_dst, xmm_src1, xmm_src2);
    vmovq(ptr[param3 + offset], xmm_dst);
    offset += sizeof(float) * 2;
    rest -= 2;
  }
  if (rest > 0) {
    vmovss(xmm_src1, ptr[param1 + offset]);
    vmovss(xmm_src2, ptr[param2 + offset]);
    vaddss(xmm_dst, xmm_src1, xmm_src2);
    vmovss(ptr[param3 + offset], xmm_dst);
  }
  ret();
}
105 106 107 108 109
}  // namespace gen
}  // namespace jitkernel
}  // namespace math
}  // namespace operators
}  // namespace paddle