blas.cc 5.5 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
/* 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. */
T
tensor-tang 已提交
14 15 16

#include "paddle/fluid/operators/jit/gen/blas.h"
#include "paddle/fluid/operators/jit/registry.h"
T
tensor-tang 已提交
17
#include "paddle/fluid/platform/cpu_info.h"
T
tensor-tang 已提交
18 19 20

namespace paddle {
namespace operators {
T
tensor-tang 已提交
21 22
namespace jit {
namespace gen {
T
tensor-tang 已提交
23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41

void VXXJitCode::genCode() {
  // do not need push stack, and do not need save avx512reg if do not use avx512
  int offset = 0;
  if (with_relu_) {
    vxorps(ymm_zero, ymm_zero, ymm_zero);
  }
  if (scalar_index_ == 1) {
    vbroadcastss(ymm_src1, ptr[param1]);
  } else if (scalar_index_ == 2) {
    vbroadcastss(ymm_src2, ptr[param2]);
  }
  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
    if (scalar_index_ != 1) {
      vmovups(ymm_src1, ptr[param1 + offset]);
    }
    if (scalar_index_ != 2) {
      vmovups(ymm_src2, ptr[param2 + offset]);
    }
42
    if (type_ == operand_type::MUL) {
T
tensor-tang 已提交
43
      vmulps(ymm_dst, ymm_src1, ymm_src2);
44
    } else if (type_ == operand_type::ADD) {
T
tensor-tang 已提交
45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81
      vaddps(ymm_dst, ymm_src1, ymm_src2);
    }
    if (with_relu_) {
      vmaxps(ymm_dst, ymm_zero, ymm_dst);
    }
    vmovups(ptr[param3 + offset], ymm_dst);
    offset += sizeof(float) * YMM_FLOAT_BLOCK;
  }
  int rest = num_ % YMM_FLOAT_BLOCK;
  while (rest > 0) {
    int block = XMM_FLOAT_BLOCK;
    if (rest >= 4) {
      block = 4;
      if (scalar_index_ != 1) {
        vmovups(xmm_src1, ptr[param1 + offset]);
      }
      if (scalar_index_ != 2) {
        vmovups(xmm_src2, ptr[param2 + offset]);
      }
    } else if (rest >= 2) {
      block = 2;
      if (scalar_index_ != 1) {
        vmovq(xmm_src1, ptr[param1 + offset]);
      }
      if (scalar_index_ != 2) {
        vmovq(xmm_src2, ptr[param2 + offset]);
      }
    } else {
      block = 1;
      if (scalar_index_ != 1) {
        vmovss(xmm_src1, ptr[param1 + offset]);
      }
      if (scalar_index_ != 2) {
        vmovss(xmm_src2, ptr[param2 + offset]);
      }
    }
    switch (type_) {
82
      case operand_type::MUL:
T
tensor-tang 已提交
83 84
        vmulps(xmm_dst, xmm_src1, xmm_src2);
        break;
85
      case operand_type::ADD:
T
tensor-tang 已提交
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106
        vaddps(xmm_dst, xmm_src1, xmm_src2);
        break;
      default:
        break;
    }
    if (with_relu_) {
      vmaxps(xmm_dst, xmm_zero, xmm_dst);
    }
    if (rest >= 4) {
      vmovups(ptr[param3 + offset], xmm_dst);
    } else if (rest >= 2) {
      vmovq(ptr[param3 + offset], xmm_dst);
    } else {
      vmovss(ptr[param3 + offset], xmm_dst);
    }
    offset += sizeof(float) * block;
    rest -= block;
  }
  ret();
}

T
tensor-tang 已提交
107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148
void NCHW16CMulNCJitCode::genCode() {
  // RDI is ptr x_input
  // RSI is ptr y_input
  // RDX is ptr output
  // RCX is height
  // r8 is width

  push(rbx);

  xor_(rax, rax);
  xor_(r10, r10);
  vmovups(zmm3, ptr[rsi]);

  L("h_loop");
  xor_(rbx, rbx);
  L("w_loop");
  vmovups(zmm2, ptr[rdi + rax]);
  vmulps(zmm1, zmm2, zmm3);
  vmovups(ptr[rdx + rax], zmm1);
  add(rax, 64);
  inc(rbx);
  cmp(r8, rbx);
  jnz("w_loop");
  inc(r10);
  cmp(r10, rcx);
  jnz("h_loop");

  pop(rbx);
  ret();
}

class NCHW16CMulNCCreator : public JitCodeCreator<int> {
 public:
  bool UseMe(const int& attr) const override {
    return platform::MayIUse(platform::avx512f);
  }
  size_t CodeSize(const int& d) const override { return 256 * 1024; }
  std::unique_ptr<GenBase> CreateJitCode(const int& attr) const override {
    return make_unique<NCHW16CMulNCJitCode>(attr, CodeSize(attr));
  }
};

149 150 151 152 153 154 155 156 157 158 159 160
#define DECLARE_BLAS_CREATOR(name)                                           \
  class name##Creator : public JitCodeCreator<int> {                         \
   public:                                                                   \
    bool UseMe(const int& attr) const override {                             \
      return platform::MayIUse(platform::avx);                               \
    }                                                                        \
    size_t CodeSize(const int& d) const override {                           \
      return 96 + d / YMM_FLOAT_BLOCK * 4 * 8;                               \
    }                                                                        \
    std::unique_ptr<GenBase> CreateJitCode(const int& attr) const override { \
      return make_unique<name##JitCode>(attr, CodeSize(attr));               \
    }                                                                        \
T
tensor-tang 已提交
161
  }
162 163 164 165 166 167 168 169 170

DECLARE_BLAS_CREATOR(VMul);
DECLARE_BLAS_CREATOR(VAdd);
DECLARE_BLAS_CREATOR(VSub);
DECLARE_BLAS_CREATOR(VAddRelu);
DECLARE_BLAS_CREATOR(VScal);
DECLARE_BLAS_CREATOR(VAddBias);

#undef DECLARE_BLAS_CREATOR
T
tensor-tang 已提交
171

T
tensor-tang 已提交
172
}  // namespace gen
T
tensor-tang 已提交
173
}  // namespace jit
T
tensor-tang 已提交
174 175
}  // namespace operators
}  // namespace paddle
T
tensor-tang 已提交
176 177 178 179

namespace gen = paddle::operators::jit::gen;

REGISTER_JITKERNEL_GEN(vmul, gen::VMulCreator);
180 181 182 183 184 185
REGISTER_JITKERNEL_GEN(vadd, gen::VAddCreator);
// TODO(TJ): enable sub
// REGISTER_JITKERNEL_GEN(vsub, gen::VSubCreator);
REGISTER_JITKERNEL_GEN(vaddrelu, gen::VAddReluCreator);
REGISTER_JITKERNEL_GEN(vscal, gen::VScalCreator);
REGISTER_JITKERNEL_GEN(vaddbias, gen::VAddBiasCreator);
T
tensor-tang 已提交
186
REGISTER_JITKERNEL_GEN(nchw16cmulnc, gen::NCHW16CMulNCCreator);