bench/DepthwiseBenchmark.cc (305 lines of code) (raw):
/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* All rights reserved.
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
*/
#include <algorithm>
#include <chrono>
#include <cmath>
#include <cstdio>
#include <iostream>
#include <vector>
#ifdef _OPENMP
#include <omp.h>
#endif
#include "./AlignedVec.h"
#include "./BenchUtils.h"
#include "fbgemm/FbgemmI8DepthwiseAvx2.h"
#include "fbgemm/Utils.h"
#include "src/RefImplementations.h"
using namespace std;
using namespace fbgemm;
int main() {
#ifdef _OPENMP
// Use 1 thread unless OMP_NUM_THREADS is explicit set.
const char* val = getenv("OMP_NUM_THREADS");
if (val == nullptr || !*val) {
omp_set_num_threads(1);
}
#endif
// From Xray OCR
// clang-format off
vector<vector<int>> shapes = {
// NOTE: clang-format wants to use a different formatting but the current
// formatting should be easier to read.
// N, G, K_per_G, H_in, W_in, stride, kernel
{ 1, 272, 1, 47, 125, 1, 3, },
{ 1, 272, 1, 47, 125, 1, 5, },
{ 1, 272, 1, 64, 125, 1, 3, },
{ 1, 272, 1, 66, 125, 1, 3, },
{ 1, 272, 1, 67, 100, 1, 3, },
{ 1, 272, 1, 71, 125, 1, 3, },
{ 1, 272, 1, 74, 125, 1, 3, },
{ 1, 272, 1, 75, 75, 1, 3, },
{ 1, 272, 1, 75, 76, 1, 3, },
{ 1, 272, 1, 75, 79, 1, 3, },
{ 1, 272, 1, 75, 85, 1, 3, },
{ 1, 272, 1, 75, 100, 1, 3, },
{ 1, 272, 1, 75, 103, 1, 3, },
{ 1, 272, 1, 75, 111, 1, 3, },
{ 1, 272, 1, 75, 113, 1, 3, },
{ 1, 272, 1, 94, 75, 1, 3, },
{ 1, 272, 1, 109, 75, 1, 3, },
{ 1, 272, 1, 113, 75, 1, 3, },
{ 1, 272, 1, 117, 75, 1, 3, },
{ 1, 544, 1, 24, 63, 1, 3, },
{ 1, 544, 1, 32, 63, 1, 3, },
{ 1, 544, 1, 33, 63, 1, 3, },
{ 1, 544, 1, 34, 50, 1, 3, },
{ 1, 544, 1, 36, 63, 1, 3, },
{ 1, 544, 1, 37, 63, 1, 3, },
{ 1, 544, 1, 38, 38, 1, 3, },
{ 1, 544, 1, 38, 40, 1, 3, },
{ 1, 544, 1, 38, 43, 1, 3, },
{ 1, 544, 1, 38, 50, 1, 3, },
{ 1, 544, 1, 38, 52, 1, 3, },
{ 1, 544, 1, 38, 56, 1, 3, },
{ 1, 544, 1, 38, 57, 1, 3, },
{ 1, 544, 1, 47, 38, 1, 3, },
{ 1, 544, 1, 55, 38, 1, 3, },
{ 1, 544, 1, 57, 38, 1, 3, },
{ 1, 544, 1, 59, 38, 1, 3, },
{ 1, 1088, 1, 7, 7, 1, 3, },
{ 51, 1088, 1, 7, 7, 1, 3, },
{ 59, 1088, 1, 7, 7, 1, 3, },
{ 70, 1088, 1, 7, 7, 1, 3, },
{ 71, 1088, 1, 7, 7, 1, 3, },
{ 77, 1088, 1, 7, 7, 1, 3, },
{ 79, 1088, 1, 7, 7, 1, 3, },
{ 84, 1088, 1, 7, 7, 1, 3, },
{ 85, 1088, 1, 7, 7, 1, 3, },
{ 89, 1088, 1, 7, 7, 1, 3, },
{ 93, 1088, 1, 7, 7, 1, 3, },
{ 96, 1088, 1, 7, 7, 1, 3, },
{ 100, 1088, 1, 7, 7, 1, 3, },
{ 1, 248, 1, 93, 250, 2, 3, },
{ 1, 248, 1, 128, 250, 2, 3, },
{ 1, 248, 1, 132, 250, 2, 3, },
{ 1, 248, 1, 131, 250, 2, 3, },
{ 1, 248, 1, 133, 200, 2, 3, },
{ 1, 248, 1, 141, 250, 2, 3, },
{ 1, 248, 1, 148, 250, 2, 3, },
{ 1, 248, 1, 150, 150, 2, 3, },
{ 1, 248, 1, 150, 151, 2, 3, },
{ 1, 248, 1, 150, 158, 2, 3, },
{ 1, 248, 1, 150, 169, 2, 3, },
{ 1, 248, 1, 150, 200, 2, 3, },
{ 1, 248, 1, 150, 205, 2, 3, },
{ 1, 248, 1, 150, 221, 2, 3, },
{ 1, 248, 1, 150, 225, 2, 3, },
{ 1, 248, 1, 188, 150, 2, 3, },
{ 1, 248, 1, 218, 150, 2, 3, },
{ 1, 248, 1, 225, 150, 2, 3, },
{ 1, 248, 1, 234, 150, 2, 3, },
{ 1, 272, 1, 47, 125, 2, 3, },
{ 1, 272, 1, 64, 125, 2, 3, },
{ 1, 272, 1, 66, 125, 2, 3, },
{ 1, 272, 1, 67, 100, 2, 3, },
{ 1, 272, 1, 71, 125, 2, 3, },
{ 1, 272, 1, 74, 125, 2, 3, },
{ 1, 272, 1, 75, 75, 2, 3, },
{ 1, 272, 1, 75, 76, 2, 3, },
{ 1, 272, 1, 75, 79, 2, 3, },
{ 1, 272, 1, 75, 85, 2, 3, },
{ 1, 272, 1, 75, 100, 2, 3, },
{ 1, 272, 1, 75, 103, 2, 3, },
{ 1, 272, 1, 75, 111, 2, 3, },
{ 1, 272, 1, 75, 113, 2, 3, },
{ 1, 272, 1, 94, 75, 2, 3, },
{ 1, 272, 1, 109, 75, 2, 3, },
{ 1, 272, 1, 113, 75, 2, 3, },
{ 1, 272, 1, 117, 75, 2, 3, },
{ 1, 544, 1, 14, 14, 2, 3, },
{ 51, 544, 1, 14, 14, 2, 3, },
{ 59, 544, 1, 14, 14, 2, 3, },
{ 70, 544, 1, 14, 14, 2, 3, },
{ 71, 544, 1, 14, 14, 2, 3, },
{ 77, 544, 1, 14, 14, 2, 3, },
{ 79, 544, 1, 14, 14, 2, 3, },
{ 84, 544, 1, 14, 14, 2, 3, },
{ 85, 544, 1, 14, 14, 2, 3, },
{ 89, 544, 1, 14, 14, 2, 3, },
{ 93, 544, 1, 14, 14, 2, 3, },
{ 96, 544, 1, 14, 14, 2, 3, },
{ 100, 544, 1, 14, 14, 2, 3, },
{ 1, 16, 1, 112, 112, 1, 3, },
{ 1, 24, 1, 56, 56, 1, 3, },
{ 1, 96, 1, 112, 112, 2, 3, },
{ 1, 192, 1, 28, 28, 1, 3, },
{ 1, 96, 1, 28, 28, 1, 5, },
{ 1, 144, 1, 56, 56, 2, 5, },
{ 1, 192, 1, 28, 28, 1, 5, },
{ 1, 192, 1, 28, 28, 2, 5, },
{ 1, 192, 1, 14, 14, 1, 5, },
{ 1, 336, 1, 14, 14, 1, 5, },
{ 1, 384, 1, 14, 14, 1, 5, },
{ 1, 672, 1, 14, 14, 1, 5, },
{ 1, 672, 1, 14, 14, 2, 5, },
{ 1, 1104, 1, 7, 7, 1, 5, },
{ 1, 32, 1, 112, 112, 1, 3, },
{ 1, 144, 1, 56, 56, 1, 3, },
{ 1, 240, 1, 28, 28, 2, 3, },
{ 1, 480, 1, 14, 14, 1, 3, },
{ 1, 1152, 1, 7, 7, 1, 3, },
{ 1, 240, 1, 28, 28, 1, 5, },
{ 1, 480, 1, 14, 14, 1, 5, },
{ 1, 576, 1, 14, 14, 1, 5, },
{ 1, 768, 1, 14, 14, 2, 5, },
{ 1, 1104, 1, 7, 7, 1, 3, },
{ 1, 1152, 1, 7, 7, 1, 5, },
{ 1, 32, 1, 400, 400, 1, 3, },
{ 1, 96, 1, 400, 400, 2, 3, },
{ 1, 144, 1, 200, 200, 1, 3, },
{ 1, 144, 1, 200, 200, 2, 3, },
{ 1, 192, 1, 100, 100, 1, 3, },
{ 1, 192, 1, 100, 100, 2, 3, },
{ 1, 384, 1, 50, 50, 1, 3, },
{ 1, 576, 1, 50, 50, 1, 3, },
// 2 output channels per group
{ 1, 128, 2, 32, 100, 1, 3, },
};
// clang-format on
// Depthwise is memory BW bound so we want to flush LLC.
bool flush = true;
std::vector<char> llc;
if (flush) {
llc.resize(128 * 1024 * 1024, 1.0);
}
constexpr int NWARMUP = 4;
constexpr int NITER = 16;
for (auto shape : shapes) {
int N = shape[0];
int G = shape[1];
int OC_PER_G = shape[2];
int H = shape[3];
int W = shape[4];
int stride_h = shape[5];
int stride_w = stride_h;
int R = shape[6];
int S = R;
int PAD_T = (R - 1) / 2, PAD_B = (R - 1) / 2, PAD_L = (S - 1) / 2,
PAD_R = (S - 1) / 2;
int OC = G * OC_PER_G;
conv_param_t<2> conv_p(
N,
G,
OC,
{H, W},
G,
{R, S},
{stride_h, stride_w},
{PAD_T, PAD_L, PAD_B, PAD_R});
int H_OUT = conv_p.OUT_DIM[0];
int W_OUT = conv_p.OUT_DIM[1];
int MDim = N * H_OUT * W_OUT;
int KDim = R * S * G;
int KDimPerGroup = KDim / conv_p.G;
aligned_vector<uint8_t> A(N * H * W * G);
aligned_vector<int8_t> B(KDim * OC_PER_G);
aligned_vector<int8_t> B_tr(B.size());
aligned_vector<int32_t> C_ref(MDim * OC), C(C_ref.size());
aligned_vector<uint8_t> C_uint8_ref(C_ref.size()), C_uint8(C_ref.size());
randFill<uint8_t>(A, 0, 86);
int32_t A_zero_point = 43;
randFill<int8_t>(B, -16, 16);
int32_t B_zero_point = 5;
aligned_vector<float> C_multiplier(1);
randFill(C_multiplier, 0.001234f / 2, 0.001234f * 3 / 2);
int32_t C_zero_point = 5;
vector<int32_t> row_offsets(MDim);
// im2col to compute row offset later
vector<uint8_t> A_im2col(MDim * KDim);
im2col_ref(conv_p, A.data(), A_zero_point, A_im2col.data());
aligned_vector<int32_t> col_offsets(OC);
aligned_vector<int32_t> bias(OC);
randFill(col_offsets, -100, 100);
randFill(bias, -40, 40);
// reference implementation conv_ref expects weights to be in G (R S C/G)
// K/G
transposeConvWeights(conv_p, B.data(), B_tr.data());
conv_ref(conv_p, A.data(), A_zero_point, B_tr.data(), C_ref.data());
for (int g = 0; g < conv_p.G; ++g) {
// Compute row offset
row_offsets_u8acc32_ref(
MDim,
KDimPerGroup,
KDim,
A_im2col.data() + g * KDimPerGroup,
row_offsets.data());
// Requantization
requantize_u8acc32_ref(
MDim,
OC_PER_G,
OC,
C_ref.data() + g * OC_PER_G,
C_uint8_ref.data() + g * OC_PER_G,
C_multiplier.data(),
C_zero_point,
A_zero_point,
&B_zero_point,
row_offsets.data(),
col_offsets.data() + g * OC_PER_G,
bias.data() + g * OC_PER_G,
OC);
}
PackedDepthWiseConvMatrix Bp(OC, R * S, B.data());
double bytes = G *
(N * (2. * sizeof(int32_t) * H_OUT * W_OUT * OC_PER_G + H * W) +
R * S * OC_PER_G);
double ops = 2.0 * N * H_OUT * W_OUT * OC * R * S;
double ttot = measureWithWarmup(
[&]() {
int num_threads = fbgemm_get_num_threads();
int tid = fbgemm_get_thread_num();
depthwise_2d_same_pad<QuantizationGranularity::TENSOR>(
N,
H,
W,
G,
OC,
stride_h,
stride_w,
A_zero_point,
A.data(),
&B_zero_point,
Bp,
C_multiplier.data(),
C_zero_point,
C_uint8.data(),
col_offsets.data(),
bias.data(),
false, /* fuse_relu */
nullptr, /* act_scale * w_scale */
tid,
num_threads);
},
NWARMUP,
NITER,
[&]() {
if (flush) {
llc_flush(llc);
}
},
true /*useOpenMP*/);
// correctness check
for (int n = 0; n < N; ++n) {
for (int h = 0; h < H_OUT; ++h) {
for (int w = 0; w < W_OUT; ++w) {
for (int k = 0; k < OC; ++k) {
uint8_t expected =
C_uint8_ref[((n * H_OUT + h) * W_OUT + w) * OC + k];
uint8_t actual = C_uint8[((n * H_OUT + h) * W_OUT + w) * OC + k];
if (expected != actual) {
cerr << "Depthwise 3x3 results differ at (" << n << ", " << h
<< ", " << w << ", " << k << "). expected " << (int)expected
<< " actual " << (int)actual << endl;
return -1;
}
assert(expected == actual);
}
}
}
}
// Report performance
printf(
"N = %d G = %d OC = %d H = %d W = %d stride = %d R = %d\n",
N,
G,
OC,
H,
W,
stride_h,
R);
printf("GB/s = %f Gops/s = %f\n", bytes / ttot / 1e9, ops / ttot / 1e9);
} // for each shape
return 0;
}