Skip to content

Commit

Permalink
convolution sgemm pack4to16
Browse files Browse the repository at this point in the history
  • Loading branch information
@nihui
nihui committed Apr 11, 2022
1 parent 88970f4 commit 4f0eacb
Show file tree
Hide file tree
Showing 3 changed files with 426 additions and 0 deletions.
66 changes: 66 additions & 0 deletions src/layer/x86/convolution_1x1_pack4to16.h
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,66 @@
// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// 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.

static void conv1x1s1_sgemm_pack4to16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
{
int w = bottom_blob.w;
int h = bottom_blob.h;
const int size = w * h;

Mat bottom_im2col = bottom_blob;
bottom_im2col.w = size;
bottom_im2col.h = 1;

im2col_sgemm_pack4to16_avx512(bottom_im2col, top_blob, kernel, _bias, opt);
}

static void conv1x1s2_sgemm_pack4to16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
{
int w = bottom_blob.w;
int channels = bottom_blob.c;
size_t elemsize = bottom_blob.elemsize;
int elempack = bottom_blob.elempack;

int outw = top_blob.w;
int outh = top_blob.h;

const int tailstep = (w - 2 * outw + w) * 4;

Mat bottom_blob_shrinked;
bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);

#pragma omp parallel for num_threads(opt.num_threads)
for (int p = 0; p < channels; p++)
{
const float* r0 = bottom_blob.channel(p);
float* outptr = bottom_blob_shrinked.channel(p);

for (int i = 0; i < outh; i++)
{
int j = 0;
for (; j < outw; j++)
{
__m128 _v = _mm_load_ps(r0);
_mm_store_ps(outptr, _v);

r0 += 8;
outptr += 4;
}

r0 += tailstep;
}
}

conv1x1s1_sgemm_pack4to16_avx512(bottom_blob_shrinked, top_blob, kernel, _bias, opt);
}
317 changes: 317 additions & 0 deletions src/layer/x86/convolution_sgemm_pack4to16.h
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,317 @@
// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// 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.

static void im2col_sgemm_pack4to16_avx512(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
{
// Mat bottom_im2col(size, maxk, inch, 16u, 4, opt.workspace_allocator);

const int size = bottom_im2col.w;
const int maxk = bottom_im2col.h;
const int inch = bottom_im2col.c;

const int outch = top_blob.c;

const float* bias = _bias;

// permute
Mat tmp;
if (size >= 16)
tmp.create(16 * maxk, inch, size / 16 + size % 16, 16u, 4, opt.workspace_allocator);
else
tmp.create(maxk, inch, size, 16u, 4, opt.workspace_allocator);
{
int nn_size = size >> 4;
int remain_size_start = 0;

#pragma omp parallel for num_threads(opt.num_threads)
for (int ii = 0; ii < nn_size; ii++)
{
int i = remain_size_start + ii * 16;

float* tmpptr = tmp.channel(i / 16);

for (int q = 0; q < inch; q++)
{
const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;

for (int k = 0; k < maxk; k++)
{
// transpose 4x16
__m128 _r0 = _mm_load_ps(img0);
__m128 _r1 = _mm_load_ps(img0 + 4);
__m128 _r2 = _mm_load_ps(img0 + 4 * 2);
__m128 _r3 = _mm_load_ps(img0 + 4 * 3);
__m128 _r4 = _mm_load_ps(img0 + 4 * 4);
__m128 _r5 = _mm_load_ps(img0 + 4 * 5);
__m128 _r6 = _mm_load_ps(img0 + 4 * 6);
__m128 _r7 = _mm_load_ps(img0 + 4 * 7);
__m128 _r8 = _mm_load_ps(img0 + 4 * 8);
__m128 _r9 = _mm_load_ps(img0 + 4 * 9);
__m128 _ra = _mm_load_ps(img0 + 4 * 10);
__m128 _rb = _mm_load_ps(img0 + 4 * 11);
__m128 _rc = _mm_load_ps(img0 + 4 * 12);
__m128 _rd = _mm_load_ps(img0 + 4 * 13);
__m128 _re = _mm_load_ps(img0 + 4 * 14);
__m128 _rf = _mm_load_ps(img0 + 4 * 15);

_MM_TRANSPOSE4_PS(_r0, _r1, _r2, _r3);
_MM_TRANSPOSE4_PS(_r4, _r5, _r6, _r7);
_MM_TRANSPOSE4_PS(_r8, _r9, _ra, _rb);
_MM_TRANSPOSE4_PS(_rc, _rd, _re, _rf);

_mm_store_ps(tmpptr, _r0);
_mm_store_ps(tmpptr + 4, _r4);
_mm_store_ps(tmpptr + 4 * 2, _r8);
_mm_store_ps(tmpptr + 4 * 3, _rc);
_mm_store_ps(tmpptr + 4 * 4, _r1);
_mm_store_ps(tmpptr + 4 * 5, _r5);
_mm_store_ps(tmpptr + 4 * 6, _r9);
_mm_store_ps(tmpptr + 4 * 7, _rd);
_mm_store_ps(tmpptr + 4 * 8, _r2);
_mm_store_ps(tmpptr + 4 * 9, _r6);
_mm_store_ps(tmpptr + 4 * 10, _ra);
_mm_store_ps(tmpptr + 4 * 11, _re);
_mm_store_ps(tmpptr + 4 * 12, _r3);
_mm_store_ps(tmpptr + 4 * 13, _r7);
_mm_store_ps(tmpptr + 4 * 14, _rb);
_mm_store_ps(tmpptr + 4 * 15, _rf);

img0 += size * 4;
tmpptr += 64;
}
}
}

remain_size_start += nn_size << 4;

#pragma omp parallel for num_threads(opt.num_threads)
for (int i = remain_size_start; i < size; i++)
{
float* tmpptr = tmp.channel(i / 16 + i % 16);

for (int q = 0; q < inch; q++)
{
const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;

for (int k = 0; k < maxk; k++)
{
__m128 _val = _mm_load_ps(img0);
_mm_store_ps(tmpptr, _val);

img0 += size * 4;
tmpptr += 4;
}
}
}
}

#pragma omp parallel for num_threads(opt.num_threads)
for (int p = 0; p < outch; p++)
{
float* outptr0 = top_blob.channel(p);

const float zeros[16] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
const float* biasptr = bias ? bias + p * 16 : zeros;

int i = 0;
for (; i + 15 < size; i += 16)
{
float* tmpptr = tmp.channel(i / 16);
const float* kptr = kernel.channel(p);

int nn = inch * maxk * 4; // inch always > 0

__m512 _sum0 = _mm512_loadu_ps(biasptr);
__m512 _sum1 = _sum0;
__m512 _sum2 = _sum0;
__m512 _sum3 = _sum0;
__m512 _sum4 = _sum0;
__m512 _sum5 = _sum0;
__m512 _sum6 = _sum0;
__m512 _sum7 = _sum0;
__m512 _sum8 = _sum0;
__m512 _sum9 = _sum0;
__m512 _suma = _sum0;
__m512 _sumb = _sum0;
__m512 _sumc = _sum0;
__m512 _sumd = _sum0;
__m512 _sume = _sum0;
__m512 _sumf = _sum0;

for (int j = 0; j < nn; j++)
{
__m512 _w0 = _mm512_load_ps(kptr);

__m512 _val0 = _mm512_set1_ps(tmpptr[0]);
__m512 _val1 = _mm512_set1_ps(tmpptr[1]);
_sum0 = _mm512_fmadd_ps(_val0, _w0, _sum0);
_sum1 = _mm512_fmadd_ps(_val1, _w0, _sum1);
__m512 _val2 = _mm512_set1_ps(tmpptr[2]);
__m512 _val3 = _mm512_set1_ps(tmpptr[3]);
_sum2 = _mm512_fmadd_ps(_val2, _w0, _sum2);
_sum3 = _mm512_fmadd_ps(_val3, _w0, _sum3);
__m512 _val4 = _mm512_set1_ps(tmpptr[4]);
__m512 _val5 = _mm512_set1_ps(tmpptr[5]);
_sum4 = _mm512_fmadd_ps(_val4, _w0, _sum4);
_sum5 = _mm512_fmadd_ps(_val5, _w0, _sum5);
__m512 _val6 = _mm512_set1_ps(tmpptr[6]);
__m512 _val7 = _mm512_set1_ps(tmpptr[7]);
_sum6 = _mm512_fmadd_ps(_val6, _w0, _sum6);
_sum7 = _mm512_fmadd_ps(_val7, _w0, _sum7);
__m512 _val8 = _mm512_set1_ps(tmpptr[8]);
__m512 _val9 = _mm512_set1_ps(tmpptr[9]);
_sum8 = _mm512_fmadd_ps(_val8, _w0, _sum8);
_sum9 = _mm512_fmadd_ps(_val9, _w0, _sum9);
__m512 _vala = _mm512_set1_ps(tmpptr[10]);
__m512 _valb = _mm512_set1_ps(tmpptr[11]);
_suma = _mm512_fmadd_ps(_vala, _w0, _suma);
_sumb = _mm512_fmadd_ps(_valb, _w0, _sumb);
__m512 _valc = _mm512_set1_ps(tmpptr[12]);
__m512 _vald = _mm512_set1_ps(tmpptr[13]);
_sumc = _mm512_fmadd_ps(_valc, _w0, _sumc);
_sumd = _mm512_fmadd_ps(_vald, _w0, _sumd);
__m512 _vale = _mm512_set1_ps(tmpptr[14]);
__m512 _valf = _mm512_set1_ps(tmpptr[15]);
_sume = _mm512_fmadd_ps(_vale, _w0, _sume);
_sumf = _mm512_fmadd_ps(_valf, _w0, _sumf);

kptr += 16;
tmpptr += 16;
}

_mm512_store_ps(outptr0, _sum0);
_mm512_store_ps(outptr0 + 16, _sum1);
_mm512_store_ps(outptr0 + 16 * 2, _sum2);
_mm512_store_ps(outptr0 + 16 * 3, _sum3);
_mm512_store_ps(outptr0 + 16 * 4, _sum4);
_mm512_store_ps(outptr0 + 16 * 5, _sum5);
_mm512_store_ps(outptr0 + 16 * 6, _sum6);
_mm512_store_ps(outptr0 + 16 * 7, _sum7);
_mm512_store_ps(outptr0 + 16 * 8, _sum8);
_mm512_store_ps(outptr0 + 16 * 9, _sum9);
_mm512_store_ps(outptr0 + 16 * 10, _suma);
_mm512_store_ps(outptr0 + 16 * 11, _sumb);
_mm512_store_ps(outptr0 + 16 * 12, _sumc);
_mm512_store_ps(outptr0 + 16 * 13, _sumd);
_mm512_store_ps(outptr0 + 16 * 14, _sume);
_mm512_store_ps(outptr0 + 16 * 15, _sumf);

outptr0 += 16 * 16;
}
for (; i < size; i++)
{
float* tmpptr = tmp.channel(i / 16 + i % 16);
const float* kptr = kernel.channel(p);

int nn = inch * maxk * 4; // inch always > 0

__m512 _sum0 = _mm512_loadu_ps(biasptr);

for (int j = 0; j < nn; j++)
{
__m512 _w0 = _mm512_load_ps(kptr);
__m512 _val0 = _mm512_set1_ps(tmpptr[0]);
_sum0 = _mm512_fmadd_ps(_val0, _w0, _sum0);

kptr += 16;
tmpptr += 1;
}

_mm512_store_ps(outptr0, _sum0);
outptr0 += 16;
}
}
}

static void convolution_im2col_sgemm_transform_kernel_pack4to16_avx512(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
{
const int maxk = kernel_w * kernel_h;

// interleave
// src = maxk-inch-outch
// dst = 16b-4a-maxk-inch/4a-outch/16b
Mat kernel = _kernel.reshape(maxk, inch, outch);
kernel_tm.create(16 * 4 * maxk, inch / 4, outch / 16, (size_t)4u);

for (int q = 0; q + 15 < outch; q += 16)
{
float* g00 = kernel_tm.channel(q / 16);

for (int p = 0; p + 3 < inch; p += 4)
{
for (int k = 0; k < maxk; k++)
{
for (int i = 0; i < 4; i++)
{
for (int j = 0; j < 16; j++)
{
const float* k00 = kernel.channel(q + j).row(p + i);
g00[0] = k00[k];
g00++;
}
}
}
}
}
}

static void convolution_im2col_sgemm_pack4to16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
{
int w = bottom_blob.w;
int inch = bottom_blob.c;

int outw = top_blob.w;
int outh = top_blob.h;
const int size = outw * outh;

const int maxk = kernel_w * kernel_h;

// im2col
Mat bottom_im2col(size, maxk, inch, 16u, 4, opt.workspace_allocator);
{
const int gap = (w * stride_h - outw * stride_w) * 4;

#pragma omp parallel for num_threads(opt.num_threads)
for (int p = 0; p < inch; p++)
{
const Mat img = bottom_blob.channel(p);
float* ptr = bottom_im2col.channel(p);

for (int u = 0; u < kernel_h; u++)
{
for (int v = 0; v < kernel_w; v++)
{
const float* sptr = img.row(dilation_h * u) + dilation_w * v * 4;

for (int i = 0; i < outh; i++)
{
int j = 0;
for (; j < outw; j++)
{
__m128 _val = _mm_load_ps(sptr);
_mm_store_ps(ptr, _val);

sptr += stride_w * 4;
ptr += 4;
}

sptr += gap;
}
}
}
}
}

im2col_sgemm_pack4to16_avx512(bottom_im2col, top_blob, kernel, _bias, opt);
}
Loading

0 comments on commit 4f0eacb

Please sign in to comment.