|
| 1 | +// Copyright (c) 2024 PaddlePaddle Authors 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 | +#include <immintrin.h> |
| 16 | +#include <math.h> |
| 17 | +#include <omp.h> |
| 18 | +#include <stdio.h> |
| 19 | +#include <string.h> |
| 20 | + |
| 21 | +#include "paddle/phi/backends/cpu/cpu_context.h" |
| 22 | +#include "paddle/phi/core/kernel_registry.h" |
| 23 | +#include "paddle/phi/core/tensor_utils.h" |
| 24 | + |
| 25 | +namespace phi { |
| 26 | +namespace fusion { |
| 27 | + |
| 28 | +template <typename T> |
| 29 | +void ResidualBiasSumFunc(const T* x_data, |
| 30 | + const T* residual_data, |
| 31 | + const T* bias_data, |
| 32 | + const float residual_alpha, |
| 33 | + const int rows, |
| 34 | + const int cols, |
| 35 | + const int iStride, |
| 36 | + const int oStride, |
| 37 | + T* out_data) { |
| 38 | + __m512 vresidual_alpha = _mm512_set1_ps(residual_alpha); |
| 39 | + const T* pb = bias_data; |
| 40 | +#ifdef PADDLE_WITH_MKLML |
| 41 | +#pragma omp parallel for |
| 42 | +#endif |
| 43 | + for (int r = 0; r < rows; ++r) { |
| 44 | + const T* px = x_data + r * iStride; |
| 45 | + const T* pr = residual_data ? residual_data + r * iStride : nullptr; |
| 46 | + T* py = out_data + r * oStride; |
| 47 | + for (int col = 0; col < cols; col += 16) { |
| 48 | + int remain = cols - col; |
| 49 | + __mmask16 mask = (remain >= 16 ? 0xffff : (1 << remain) - 1); |
| 50 | + |
| 51 | + // residual*alpha + bias + x |
| 52 | + __m512 vx = _mm512_maskz_loadu_ps(mask, px + col); |
| 53 | + if (residual_data) { |
| 54 | + __m512 residual_vx = _mm512_maskz_loadu_ps(mask, pr + col); |
| 55 | + residual_vx = _mm512_mul_ps(residual_vx, vresidual_alpha); |
| 56 | + vx = _mm512_mask_add_ps(vx, mask, vx, residual_vx); |
| 57 | + } |
| 58 | + if (bias_data) { |
| 59 | + __m512 vb = _mm512_maskz_loadu_ps(mask, pb + col); |
| 60 | + vx = _mm512_mask_add_ps(vx, mask, vx, vb); |
| 61 | + } |
| 62 | + _mm512_mask_storeu_ps(py + col, mask, vx); |
| 63 | + } |
| 64 | + } |
| 65 | +} |
| 66 | + |
| 67 | +template <typename T> |
| 68 | +void LayerNormFunc(const T* x_data, |
| 69 | + const T* residual_data, |
| 70 | + const T* bias_data, |
| 71 | + const T* norm_weight_data, |
| 72 | + const T* norm_bias_data, |
| 73 | + const float epsilon, |
| 74 | + const float residual_alpha, |
| 75 | + const int rows, |
| 76 | + const int cols, |
| 77 | + const int iStride, |
| 78 | + const int oStride, |
| 79 | + T* out_data, |
| 80 | + T* residual_out_data, |
| 81 | + T* mean_out, |
| 82 | + T* var_out) { |
| 83 | + auto size = cols; |
| 84 | + __m512 vresidual_alpha = _mm512_set1_ps(residual_alpha); |
| 85 | + __m512 vgamma = _mm512_set1_ps(1); |
| 86 | + __m512 vbeta = _mm512_set1_ps(0); |
| 87 | + const T* pb = bias_data; |
| 88 | +#ifdef PADDLE_WITH_MKLML |
| 89 | +#pragma omp parallel for |
| 90 | +#endif |
| 91 | + for (int r = 0; r < rows; ++r) { |
| 92 | + const T* px = x_data + r * iStride; |
| 93 | + const T* pr = residual_data ? residual_data + r * iStride : nullptr; |
| 94 | + T* pr_out = residual_out_data ? residual_out_data + r * oStride : nullptr; |
| 95 | + T* py = out_data + r * oStride; |
| 96 | + |
| 97 | + T sum = 0; |
| 98 | + T squareSum = 0; |
| 99 | + |
| 100 | + __m512 vsum = _mm512_set1_ps(0); |
| 101 | + __m512 vsqare = _mm512_set1_ps(0); |
| 102 | + for (int col = 0; col < size; col += 16) { |
| 103 | + int remain = size - col; |
| 104 | + __mmask16 mask = (remain >= 16 ? 0xffff : (1 << remain) - 1); |
| 105 | + |
| 106 | + // SUM(x) |
| 107 | + __m512 vx = _mm512_maskz_loadu_ps(mask, px + col); |
| 108 | + if (residual_data) { |
| 109 | + __m512 residual_vx = _mm512_maskz_loadu_ps(mask, pr + col); |
| 110 | + residual_vx = _mm512_mul_ps(residual_vx, vresidual_alpha); |
| 111 | + vx = _mm512_mask_add_ps(vx, mask, vx, residual_vx); |
| 112 | + if (bias_data) { |
| 113 | + __m512 vb = _mm512_maskz_loadu_ps(mask, pb + col); |
| 114 | + vx = _mm512_mask_add_ps(vx, mask, vx, vb); |
| 115 | + } |
| 116 | + _mm512_mask_storeu_ps(pr_out + col, mask, vx); |
| 117 | + } |
| 118 | + vsum = _mm512_add_ps(vsum, vx); |
| 119 | + |
| 120 | + // SUM(x*x) |
| 121 | + __m512 tmp = _mm512_mul_ps(vx, vx); |
| 122 | + vsqare = _mm512_add_ps(vsqare, tmp); |
| 123 | + } |
| 124 | + |
| 125 | + sum = _mm512_reduce_add_ps(vsum); |
| 126 | + squareSum = _mm512_reduce_add_ps(vsqare); |
| 127 | + |
| 128 | + // Mean |
| 129 | + T mean = sum / size; |
| 130 | + mean_out[r] = mean; |
| 131 | + __m512 vmean = _mm512_set1_ps(mean); |
| 132 | + |
| 133 | + // Variance |
| 134 | + T var = 1 / sqrt(squareSum / size - mean * mean + epsilon); |
| 135 | + var_out[r] = var; |
| 136 | + __m512 vvar = _mm512_set1_ps(var); |
| 137 | + |
| 138 | + for (int col = 0; col < size; col += 16) { |
| 139 | + int remain = size - col; |
| 140 | + __mmask16 mask = (remain >= 16 ? 0xffff : (1 << remain) - 1); |
| 141 | + |
| 142 | + __m512 vx = _mm512_maskz_loadu_ps(mask, px + col); |
| 143 | + if (residual_data) { |
| 144 | + __m512 residual_vx = _mm512_maskz_loadu_ps(mask, pr + col); |
| 145 | + residual_vx = _mm512_mul_ps(residual_vx, vresidual_alpha); |
| 146 | + vx = _mm512_mask_add_ps(vx, mask, vx, residual_vx); |
| 147 | + if (bias_data) { |
| 148 | + __m512 vb = _mm512_maskz_loadu_ps(mask, pb + col); |
| 149 | + vx = _mm512_mask_add_ps(vx, mask, vx, vb); |
| 150 | + } |
| 151 | + } |
| 152 | + if (norm_weight_data) { |
| 153 | + vgamma = _mm512_maskz_loadu_ps(mask, norm_weight_data + col); |
| 154 | + } |
| 155 | + if (norm_bias_data) { |
| 156 | + vbeta = _mm512_maskz_loadu_ps(mask, norm_bias_data + col); |
| 157 | + } |
| 158 | + // (vx - vmean) * vgamma * vvar + vbeta |
| 159 | + __m512 vy; |
| 160 | + vx = _mm512_mask_sub_ps(vx, mask, vx, vmean); |
| 161 | + vx = _mm512_mask_mul_ps(vx, mask, vx, vgamma); |
| 162 | + vx = _mm512_mask_mul_ps(vx, mask, vx, vvar); |
| 163 | + vy = _mm512_mask_add_ps(vy, mask, vx, vbeta); |
| 164 | + _mm512_mask_storeu_ps(py + col, mask, vy); |
| 165 | + } |
| 166 | + } |
| 167 | +} |
| 168 | + |
| 169 | +template <typename T, typename Context> |
| 170 | +void FusedLayerNormAvxKernel(const Context& dev_ctx, |
| 171 | + const DenseTensor& x, |
| 172 | + const paddle::optional<DenseTensor>& bias, |
| 173 | + const paddle::optional<DenseTensor>& residual, |
| 174 | + const paddle::optional<DenseTensor>& norm_weight, |
| 175 | + const paddle::optional<DenseTensor>& norm_bias, |
| 176 | + const float epsilon, |
| 177 | + const float residual_alpha, |
| 178 | + const int begin_norm_axis, |
| 179 | + const float quant_scale, |
| 180 | + const int quant_round_type, |
| 181 | + const float quant_max_bound, |
| 182 | + const float quant_min_bound, |
| 183 | + DenseTensor* out, |
| 184 | + DenseTensor* residual_out, |
| 185 | + DenseTensor* mean, |
| 186 | + DenseTensor* variance) { |
| 187 | + if (quant_scale > 0.0f) { |
| 188 | + PD_THROW("NOT supported quant int8. "); |
| 189 | + } |
| 190 | + const auto x_dims = x.dims(); |
| 191 | + auto matrix_dim = common::flatten_to_2d(x_dims, begin_norm_axis); |
| 192 | + T* out_data = dev_ctx.template Alloc<T>(out); |
| 193 | + T* mean_out = dev_ctx.template Alloc<T>(mean); |
| 194 | + T* var_out = dev_ctx.template Alloc<T>(variance); |
| 195 | + |
| 196 | + const T* x_data = x.data<T>(); |
| 197 | + const T* bias_data = bias ? bias.get().data<T>() : nullptr; |
| 198 | + const T* residual_data = residual ? residual.get().data<T>() : nullptr; |
| 199 | + const T* norm_weight_data = |
| 200 | + norm_weight ? norm_weight.get().data<T>() : nullptr; |
| 201 | + const T* norm_bias_data = norm_bias ? norm_bias.get().data<T>() : nullptr; |
| 202 | + T* residual_out_data = |
| 203 | + residual ? dev_ctx.template Alloc<T>(residual_out) : nullptr; |
| 204 | + |
| 205 | + int32_t rows = static_cast<int32_t>(matrix_dim[0]); |
| 206 | + int32_t cols = static_cast<int32_t>(matrix_dim[1]); |
| 207 | + |
| 208 | + auto iStride = cols; |
| 209 | + auto oStride = cols; |
| 210 | + if (!norm_weight && !norm_bias_data) { |
| 211 | + ResidualBiasSumFunc(x_data, |
| 212 | + residual_data, |
| 213 | + bias_data, |
| 214 | + residual_alpha, |
| 215 | + rows, |
| 216 | + cols, |
| 217 | + iStride, |
| 218 | + oStride, |
| 219 | + out_data); |
| 220 | + } else { |
| 221 | + LayerNormFunc(x_data, |
| 222 | + residual_data, |
| 223 | + bias_data, |
| 224 | + norm_weight_data, |
| 225 | + norm_bias_data, |
| 226 | + epsilon, |
| 227 | + residual_alpha, |
| 228 | + rows, |
| 229 | + cols, |
| 230 | + iStride, |
| 231 | + oStride, |
| 232 | + out_data, |
| 233 | + residual_out_data, |
| 234 | + mean_out, |
| 235 | + var_out); |
| 236 | + } |
| 237 | +} |
| 238 | +} // namespace fusion |
| 239 | +} // namespace phi |
| 240 | + |
| 241 | +PD_REGISTER_KERNEL(fused_bias_residual_layernorm, |
| 242 | + CPU, |
| 243 | + ALL_LAYOUT, |
| 244 | + phi::fusion::FusedLayerNormAvxKernel, |
| 245 | + float) {} |
0 commit comments