mirror of
https://github.com/Tencent/ncnn.git
synced 2026-10-18 21:54:39 +08:00
nihui
231 lines
12 KiB
C++
231 lines
12 KiB
C++
// Copyright 2026 Tencent
|
|
// SPDX-License-Identifier: BSD-3-Clause
|
|
|
|
#include "testutil.h"
|
|
|
|
static int test_convolution(int w, int h, int c, int outch, int kernel, int dilation, int stride, int pad, int bias, bool use_sgemm, bool use_winograd)
|
|
{
|
|
ncnn::Mat a = RandomMat(w, h, c);
|
|
|
|
ncnn::ParamDict pd;
|
|
pd.set(0, outch);
|
|
pd.set(1, kernel);
|
|
pd.set(2, dilation);
|
|
pd.set(3, stride);
|
|
pd.set(4, pad);
|
|
pd.set(5, bias);
|
|
pd.set(6, outch * c * kernel * kernel);
|
|
|
|
int activation_type = RAND() % 7; // 0 1 2 3 4 5 6
|
|
ncnn::Mat activation_params(2);
|
|
activation_params[0] = (activation_type == 6) ? RandomFloat(0, 1) : RandomFloat(-1, 0); // alpha
|
|
activation_params[1] = RandomFloat(0, 1); // beta
|
|
pd.set(9, activation_type);
|
|
pd.set(10, activation_params);
|
|
|
|
std::vector<ncnn::Mat> weights(bias ? 2 : 1);
|
|
weights[0] = RandomMat(outch * c * kernel * kernel);
|
|
if (bias)
|
|
weights[1] = RandomMat(outch);
|
|
|
|
float epsilon = 0.001;
|
|
|
|
// fp32 path (no bf16)
|
|
{
|
|
ncnn::Option opt;
|
|
opt.num_threads = 1;
|
|
opt.use_packing_layout = true;
|
|
opt.use_fp16_packed = false;
|
|
opt.use_fp16_storage = false;
|
|
opt.use_fp16_arithmetic = false;
|
|
opt.use_bf16_packed = false;
|
|
opt.use_bf16_storage = false;
|
|
opt.use_sgemm_convolution = use_sgemm;
|
|
opt.use_winograd_convolution = use_winograd;
|
|
|
|
int ret = test_layer_opt("Convolution", pd, weights, opt, a, epsilon);
|
|
if (ret != 0)
|
|
{
|
|
fprintf(stderr, "test_convolution failed w=%d h=%d c=%d outch=%d kernel=%d dilation=%d stride=%d pad=%d bias=%d sgemm=%d winograd=%d act=%d actparams=[%f,%f]\n", w, h, c, outch, kernel, dilation, stride, pad, bias, use_sgemm, use_winograd, activation_type, activation_params[0], activation_params[1]);
|
|
return ret;
|
|
}
|
|
}
|
|
|
|
// bf16 path
|
|
{
|
|
ncnn::Option opt;
|
|
opt.num_threads = 1;
|
|
opt.use_packing_layout = true;
|
|
opt.use_fp16_packed = false;
|
|
opt.use_fp16_storage = false;
|
|
opt.use_fp16_arithmetic = false;
|
|
opt.use_bf16_packed = true;
|
|
opt.use_bf16_storage = true;
|
|
opt.use_sgemm_convolution = use_sgemm;
|
|
opt.use_winograd_convolution = use_winograd;
|
|
|
|
int ret = test_layer_opt("Convolution", pd, weights, opt, a, epsilon);
|
|
if (ret != 0)
|
|
{
|
|
fprintf(stderr, "test_convolution failed w=%d h=%d c=%d outch=%d kernel=%d dilation=%d stride=%d pad=%d bias=%d sgemm=%d winograd=%d act=%d actparams=[%f,%f]\n", w, h, c, outch, kernel, dilation, stride, pad, bias, use_sgemm, use_winograd, activation_type, activation_params[0], activation_params[1]);
|
|
return ret;
|
|
}
|
|
}
|
|
|
|
return 0;
|
|
}
|
|
|
|
// BF16 winograd path: kernel=3, dilation=1, stride=1, winograd=true, sgemm=false
|
|
// Need num_input>8 || num_output>8 for prefer_winograd
|
|
// winograd43 is the default variant (when neither 63 nor 23 is preferred)
|
|
// winograd63: num_input<64 and specific size ranges (e.g. c=16,outch=16,minwh~25-44)
|
|
// winograd23: large channels with small spatial (e.g. c=512,outch=512,minwh=3-14)
|
|
static int test_convolution_winograd()
|
|
{
|
|
// winograd43: default path for moderate sizes
|
|
// c=16,outch=16 => neither prefer_winograd63 nor prefer_winograd23 for w=11,h=10
|
|
int ret = 0
|
|
// Various elempack/out_elempack combos for winograd43
|
|
|| test_convolution(11, 10, 16, 16, 3, 1, 1, 1, 1, false, true)
|
|
|| test_convolution(11, 10, 16, 24, 3, 1, 1, 1, 0, false, true)
|
|
|| test_convolution(11, 10, 24, 16, 3, 1, 1, 1, 1, false, true)
|
|
|| test_convolution(11, 10, 16, 3, 3, 1, 1, 1, 0, false, true)
|
|
|| test_convolution(11, 10, 3, 16, 3, 1, 1, 1, 1, false, true)
|
|
|| test_convolution(11, 10, 16, 1, 3, 1, 1, 1, 0, false, true)
|
|
|| test_convolution(11, 10, 1, 16, 3, 1, 1, 1, 1, false, true)
|
|
|| test_convolution(11, 10, 16, 2, 3, 1, 1, 1, 0, false, true)
|
|
|| test_convolution(11, 10, 2, 16, 3, 1, 1, 1, 1, false, true)
|
|
|| test_convolution(11, 10, 9, 9, 3, 1, 1, 1, 0, false, true)
|
|
// out_elempack=8 (outch%16!=0 && outch%8==0)
|
|
|| test_convolution(11, 10, 16, 8, 3, 1, 1, 1, 1, false, true)
|
|
// out_elempack=4 (outch%8!=0 && outch%4==0)
|
|
|| test_convolution(11, 10, 16, 4, 3, 1, 1, 1, 0, false, true)
|
|
|| test_convolution(11, 10, 16, 12, 3, 1, 1, 1, 1, false, true)
|
|
|
|
// winograd63: c=16,outch=16, larger spatial => minwh in [23..44]
|
|
|| test_convolution(30, 30, 16, 16, 3, 1, 1, 1, 1, false, true)
|
|
|| test_convolution(30, 30, 16, 24, 3, 1, 1, 1, 0, false, true)
|
|
|| test_convolution(30, 30, 24, 16, 3, 1, 1, 1, 1, false, true)
|
|
|| test_convolution(30, 30, 16, 3, 3, 1, 1, 1, 0, false, true)
|
|
|| test_convolution(30, 30, 3, 16, 3, 1, 1, 1, 1, false, true)
|
|
// winograd63 with out_elempack=8
|
|
|| test_convolution(30, 30, 16, 8, 3, 1, 1, 1, 0, false, true)
|
|
// winograd63 with out_elempack=4 (covers the uncovered winograd63 output transform)
|
|
// Need larger spatial for outch<16 to prefer winograd63 (minwh in [47..128] for num_output>=8)
|
|
|| test_convolution(50, 50, 16, 12, 3, 1, 1, 1, 1, false, true)
|
|
// outch=20 (out_elempack=4, outch>=16 for 16-wide output transform)
|
|
|| test_convolution(50, 50, 16, 20, 3, 1, 1, 1, 0, false, true)
|
|
|| test_convolution(30, 30, 16, 4, 3, 1, 1, 1, 1, false, true)
|
|
|| test_convolution(30, 30, 16, 12, 3, 1, 1, 1, 0, false, true)
|
|
|
|
// winograd23: large channels, small spatial
|
|
|| test_convolution(5, 5, 64, 64, 3, 1, 1, 1, 1, false, true)
|
|
|| test_convolution(5, 5, 64, 32, 3, 1, 1, 1, 0, false, true)
|
|
|| test_convolution(5, 5, 32, 64, 3, 1, 1, 1, 1, false, true)
|
|
// winograd23 with out_elempack=4
|
|
|| test_convolution(5, 5, 64, 4, 3, 1, 1, 1, 0, false, true);
|
|
|
|
if (ret != 0)
|
|
return -1;
|
|
|
|
return 0;
|
|
}
|
|
|
|
// BF16 im2col_gemm (sgemm) path: sgemm=true, winograd=false
|
|
// Triggered when prefer_sgemm=true (large matrix) or kernel=1x1
|
|
// prefer_sgemm: num_input * num_output * k*k * dilation^2 * stride^2 * sizeof(bf16) * 2 > L2
|
|
// or num_input > 16 || num_output > 16
|
|
static int test_convolution_sgemm()
|
|
{
|
|
// 1x1 convolution always goes to sgemm path
|
|
int ret = 0
|
|
// exact 16 spatial locations for jj+15 bf16 deshuffle paths
|
|
|| test_convolution(4, 4, 16, 24, 1, 1, 1, 0, 1, true, false)
|
|
|| test_convolution(4, 4, 16, 20, 1, 1, 1, 0, 0, true, false)
|
|
|| test_convolution(4, 4, 16, 8, 1, 1, 1, 0, 1, true, false)
|
|
|| test_convolution(4, 4, 16, 12, 1, 1, 1, 0, 0, true, false)
|
|
|
|
|| test_convolution(11, 10, 16, 16, 1, 1, 1, 0, 1, true, false)
|
|
|| test_convolution(11, 10, 16, 24, 1, 1, 1, 0, 0, true, false)
|
|
|| test_convolution(11, 10, 24, 16, 1, 1, 1, 0, 1, true, false)
|
|
|| test_convolution(11, 10, 16, 3, 1, 1, 1, 0, 0, true, false)
|
|
|| test_convolution(11, 10, 3, 16, 1, 1, 1, 0, 1, true, false)
|
|
|| test_convolution(11, 10, 16, 1, 1, 1, 1, 0, 0, true, false)
|
|
|| test_convolution(11, 10, 1, 16, 1, 1, 1, 0, 1, true, false)
|
|
|| test_convolution(11, 10, 16, 2, 1, 1, 1, 0, 0, true, false)
|
|
|| test_convolution(11, 10, 2, 16, 1, 1, 1, 0, 1, true, false)
|
|
|| test_convolution(11, 10, 9, 9, 1, 1, 1, 0, 0, true, false)
|
|
|| test_convolution(11, 10, 1, 1, 1, 1, 1, 0, 1, true, false)
|
|
// out_elempack=8 (1x1 with outch%8==0, outch%16!=0)
|
|
|| test_convolution(11, 10, 16, 8, 1, 1, 1, 0, 1, true, false)
|
|
// out_elempack=4 (1x1 with outch%4==0, outch%8!=0)
|
|
|| test_convolution(11, 10, 16, 4, 1, 1, 1, 0, 0, true, false)
|
|
|| test_convolution(11, 10, 16, 12, 1, 1, 1, 0, 1, true, false)
|
|
|
|
// non-1x1 with prefer_sgemm (num_output > 16)
|
|
|| test_convolution(11, 10, 16, 20, 3, 1, 1, 1, 1, true, false)
|
|
|| test_convolution(11, 10, 20, 16, 3, 1, 2, 1, 0, true, false)
|
|
|| test_convolution(11, 10, 16, 20, 5, 1, 1, 2, 1, true, false)
|
|
|| test_convolution(11, 10, 20, 16, 3, 2, 1, 2, 0, true, false)
|
|
|| test_convolution(11, 10, 3, 20, 3, 1, 1, 1, 1, true, false)
|
|
|| test_convolution(11, 10, 20, 3, 3, 1, 1, 1, 0, true, false)
|
|
// non-1x1 with out_elempack=8
|
|
|| test_convolution(11, 10, 16, 8, 3, 1, 1, 1, 1, true, false)
|
|
// non-1x1 with out_elempack=4
|
|
|| test_convolution(11, 10, 16, 4, 3, 1, 1, 1, 0, true, false);
|
|
|
|
if (ret != 0)
|
|
return -1;
|
|
|
|
return 0;
|
|
}
|
|
|
|
// BF16 packed path: sgemm=false, winograd=false (or conditions not met for winograd)
|
|
// Falls through to convolution_packed_bf16s
|
|
// Kernel configs: A={3,1,2,1} B={5,1,1,-234} C={3,2,1,-234} are spread across test items
|
|
static int test_convolution_packed()
|
|
{
|
|
return 0
|
|
// out_elempack=16 k d s p
|
|
|| test_convolution(11, 10, 16, 16, 3, 1, 2, 1, 1, false, false) // ep=16
|
|
|| test_convolution(11, 10, 24, 16, 5, 1, 1, -234, 0, false, false) // ep=8
|
|
|| test_convolution(11, 10, 1, 16, 3, 2, 1, -234, 1, false, false) // ep=1
|
|
|| test_convolution(11, 10, 3, 16, 3, 1, 2, 1, 0, false, false) // ep=1
|
|
|| test_convolution(11, 10, 4, 16, 5, 1, 1, -234, 1, false, false) // ep=4, kernel_tm inch=4
|
|
|
|
// out_elempack=8
|
|
|| test_convolution(11, 10, 16, 8, 3, 2, 1, -234, 0, false, false) // ep=16
|
|
|| test_convolution(11, 10, 24, 8, 3, 1, 2, 1, 1, false, false) // ep=8, q+15 ep==8
|
|
|| test_convolution(11, 10, 20, 8, 5, 1, 1, -234, 0, false, false) // ep=4, q+15 ep==4
|
|
|| test_convolution(11, 10, 17, 8, 3, 2, 1, -234, 1, false, false) // ep=1, q+15 ep==1
|
|
|| test_convolution(11, 10, 1, 8, 3, 1, 2, 1, 0, false, false) // ep=1, kernel_tm inch=1
|
|
|
|
// out_elempack=4
|
|
|| test_convolution(11, 10, 16, 4, 5, 1, 1, -234, 0, false, false) // ep=16
|
|
|| test_convolution(11, 10, 24, 4, 3, 2, 1, -234, 1, false, false) // ep=8, q+15 ep==8
|
|
|| test_convolution(11, 10, 20, 4, 3, 1, 2, 1, 0, false, false) // ep=4, q+15 ep==4
|
|
|| test_convolution(11, 10, 17, 4, 5, 1, 1, -234, 1, false, false) // ep=1, q+15 ep==1
|
|
|| test_convolution(11, 10, 8, 4, 3, 2, 1, -234, 0, false, false) // ep=8, q+7 ep==8
|
|
|| test_convolution(11, 10, 12, 4, 3, 1, 2, 1, 1, false, false) // ep=4, q+7/q+3 ep==4
|
|
|| test_convolution(11, 10, 9, 4, 5, 1, 1, -234, 0, false, false) // ep=1, q+7 ep==1
|
|
|| test_convolution(11, 10, 5, 4, 3, 2, 1, -234, 1, false, false) // ep=1, q+3 ep==1
|
|
|| test_convolution(11, 10, 2, 4, 3, 1, 2, 1, 0, false, false) // kernel_tm inch=2
|
|
|| test_convolution(11, 10, 16, 12, 5, 1, 1, -234, 1, false, false) // outch=12
|
|
|
|
// out_elempack=1
|
|
|| test_convolution(11, 10, 16, 3, 3, 2, 1, -234, 1, false, false) // ep=16
|
|
|| test_convolution(11, 10, 24, 1, 3, 1, 2, 1, 0, false, false) // ep=8, q+15 ep==8
|
|
|| test_convolution(11, 10, 24, 3, 5, 1, 1, -234, 1, false, false) // ep=8, outch=2+1
|
|
|| test_convolution(11, 10, 20, 1, 5, 1, 1, -234, 0, false, false) // ep=4, q+15 ep==4
|
|
|| test_convolution(11, 10, 17, 1, 3, 2, 1, -234, 1, false, false) // ep=1, q+15 ep==1
|
|
|| test_convolution(11, 10, 12, 1, 3, 1, 2, 1, 0, false, false) // ep=4, q+7 ep==4
|
|
|| test_convolution(11, 10, 9, 1, 5, 1, 1, -234, 1, false, false); // ep=1, q+7 ep==1
|
|
}
|
|
|
|
int main()
|
|
{
|
|
SRAND(7767517);
|
|
|
|
return test_convolution_winograd() || test_convolution_sgemm() || test_convolution_packed();
|
|
}
|