我的书签
添加书签
移除书签名词：
GEMM = General Matrix Multiplication
译:
通用矩阵乘
前言：
参考文献依然是放前面：
https://blog.csdn.net/caicaiatnbu/category_9096319.html
https://blog.csdn.net/caicaiatnbu/article/details/100708754 源码解析还带图解
darknet版本： https://github.com/AlexeyAB/darknet ，与原始的版本还是有一点区别的。
因为第一次读源码，我就直接按照参考文献的顺序来了，到时候再查漏补缺，加油！
今天看的代码是 gemm 主要完成矢量和矩阵的加速运算，是darknet卷积底层实现的核心，其实也是caffe卷积实现的核心。
进入代码：
GEMM 在深度学习中是十分重要的，全连接层以及卷积层基本上都是通过GEMM来实现的，而网络中大约90%的运算都是在这两层中。而一个良好的GEMM的实现可以充分利用系统的多级存储结构和程序执行的局部性来充分加速运算。
gemm.c中总结起来就完成一个矩阵乘法的运算：
【类卷积操作】
上述公式中，A、B、C为矩阵，其中A、B为输入矩阵，C矩阵保存运算结果。 、为系数。
接下来我们需要考虑矩阵A、B的行数和列数分别是多少，这里我们假设矩阵A为[M,K]，矩阵B为[K,N]，那么矩阵C为[M,N] 。我们都知道矩阵A、B、C在逻辑上是一个二维的结果，但要注意在这里实际的存储结构是一个一维数组，按行存储。
gemm_tn()
用于计算 这种类型；
gemm_nt()
用于计算 这种类型；
gemm_tt()
用于计算 这种类型；
gemm_cpu()
就是根据矩阵A和B的情况来实际调用 gemm_nn()、gemm_tn()、gemm_nt()、gemm_tt()；
gemm()
其实就是在 gemm_cpu() 上再封装一层，参数原封不动传递给gemm_cpu函数；
函数里完成了主要工作：矩阵乘，二维卷积操作
读代码时间长是因为还有一些加速，并行的知识存在
【可以看一下参考文献里的图解，代码是把整个矩阵运算拆分了】
1. gemm.h
#ifndef GEMM_H
#define GEMM_H
#include "activations.h"
#include <stdint.h>
#include <stddef.h>
#ifdef __cplusplus
extern "C" {
#endif
void convolution_2d(int w, int h, int ksize, int n, int c, int pad, int stride,
float *weights, float *input, float *output, float *mean);
static inline void set_bit(unsigned char *const dst, size_t index) {
size_t dst_i = index / 8;
int dst_shift = index % 8;
dst[dst_i] |= 1 << dst_shift;
//dst[dst_i] |= 1 << (8 - dst_shift);
}
static inline unsigned char get_bit(unsigned char const*const src, size_t index) {
size_t src_i = index / 8;
int src_shift = index % 8;
unsigned char val = (src[src_i] & (1 << src_shift)) > 0;
//unsigned char val = (src[src_i] & (1 << (8 - src_shift))) > 0;
return val;
}
int is_avx();
int is_fma_avx2();
void float_to_bit(float *src, unsigned char *dst, size_t size);
void transpose_block_SSE4x4(float *A, float *B, const int n, const int m,
const int lda, const int ldb, const int block_size);
void transpose_bin(uint32_t *A, uint32_t *B, const int n, const int m,
const int lda, const int ldb, const int block_size);
void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
unsigned char *A, int lda,
unsigned char *B, int ldb,
float *C, int ldc, float *mean_arr);
void im2col_cpu_custom(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col);
void im2col_cpu_custom_align(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col, int bit_align);
void im2col_cpu_custom_bin(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col, int bit_align);
void im2col_cpu_custom_transpose(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col, int ldb_align);
void activate_array_cpu_custom(float *x, const int n, const ACTIVATION a);
void transpose_32x32_bits_reversed_diagonale(uint32_t *A, uint32_t *B, int m, int n);
void gemm_bin(int M, int N, int K, float ALPHA,
char *A, int lda,
float *B, int ldb,
float *C, int ldc);
void repack_input(float *input, float *re_packed_input, int w, int h, int c);
void convolution_repacked(uint32_t *packed_input, uint32_t *packed_weights, float *output,
int w, int h, int c, int n, int size, int pad, int new_lda, float *mean_arr);
void gemm_nn_bin_32bit_packed(int M, int N, int K, float ALPHA,
uint32_t *A, int lda,
uint32_t *B, int ldb,
float *C, int ldc, float *mean_arr);
void transpose_uint32(uint32_t *src, uint32_t *dst, int src_h, int src_w, int src_align, int dst_align);
void gemm_nn_bin_transposed_32bit_packed(int M, int N, int K, float ALPHA,
uint32_t *A, int lda,
uint32_t *B, int ldb,
float *C, int ldc, float *mean_arr);
void forward_maxpool_layer_avx(float *src, float *dst, int *indexes, int size, int w, int h, int out_w, int out_h, int c,
int pad, int stride, int batch);
"""多维矩阵的点乘计算"""
void gemm(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc);
void gemm_cpu(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc);
#ifdef GPU
void gemm_ongpu(int TA, int TB, int M, int N, int K, float ALPHA,
float *A_gpu, int lda,
float *B_gpu, int ldb,
float BETA,
float *C_gpu, int ldc);
void gemm_gpu(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc);
#endif
#ifdef __cplusplus
}
#endif
#endif2. gemm.c
#include "gemm.h"
#include "utils.h"
#include "im2col.h"
#include "dark_cuda.h"
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <float.h>
#include <string.h>
#include <stdint.h>
#ifdef _WIN32
#include <intrin.h>
#endif
#if defined(_OPENMP)
#include <omp.h>
#endif
#define TILE_M 4 // 4 ops
#define TILE_N 16 // AVX2 = 2 ops * 8 floats
#define TILE_K 16 // loop
#ifdef __cplusplus
#define PUT_IN_REGISTER
#else
#define PUT_IN_REGISTER register
#endif
void gemm_bin(int M, int N, int K, float ALPHA,
char *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i,j,k;
for(i = 0; i < M; ++i){
for(k = 0; k < K; ++k){
char A_PART = A[i*lda+k];
if(A_PART){//判断A[i,j]的值是正还是负，做不同的操作
for(j = 0; j < N; ++j){
C[i*ldc+j] += B[k*ldb+j];
}
} else {
for(j = 0; j < N; ++j){
C[i*ldc+j] -= B[k*ldb+j];
}
}
}
}
}
//生成随机的数组，并返回
float *random_matrix(int rows, int cols)
{
int i;
float* m = (float*)xcalloc(rows * cols, sizeof(float));
for(i = 0; i < rows*cols; ++i){
m[i] = (float)rand()/RAND_MAX;
}
return m;
}
void time_random_matrix(int TA, int TB, int m, int k, int n)
{
float *a;
if(!TA) a = random_matrix(m,k);//如果A矩阵不需要转置
else a = random_matrix(k,m);//如果A矩阵需要转置
int lda = (!TA)?k:m;//判断lda应该为A的行数还是列数，因为有A[i,j]=a[i*lda+j]
float *b;//矩阵B做同样的判断
if(!TB) b = random_matrix(k,n);
else b = random_matrix(n,k);
int ldb = (!TB)?n:k;
float *c = random_matrix(m,n);//矩阵C不需要判断是否转置啦
int i;
clock_t start = clock(), end;//获取当前时间
for(i = 0; i<10; ++i){
gemm_cpu(TA,TB,m,n,k,1,a,lda,b,ldb,1,c,n);
}
end = clock();//获取处理完矩阵计算操作后的时间
//输出多维矩阵计算的时长
printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %lf ms\n",m,k,k,n, TA, TB, (float)(end-start)/CLOCKS_PER_SEC);
free(a);
free(b);
free(c);//释放空间
}
//gemm函数利用了gemm_cpu函数，进行多维矩阵计算
void gemm(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc)
{
gemm_cpu( TA, TB, M, N, K, ALPHA,A,lda, B, ldb,BETA,C,ldc);
}
//--------------------------------------------
// XNOR bitwise GEMM for binary neural network
// 二进制神经网络的按位异或GEMM【？】
//--------------------------------------------
static inline unsigned char xnor(unsigned char a, unsigned char b) {
//return a == b;
return !(a^b);//按位异或再取反
}
// INT-32
static inline uint32_t get_bit_int32(uint32_t const*const src, size_t index) {
size_t src_i = index / 32;
int src_shift = index % 32;
unsigned char val = (src[src_i] & (1 << src_shift)) > 0;
return val;
}
static inline uint32_t xnor_int32(uint32_t a, uint32_t b) {
return ~(a^b);//按位异或再a按位取反
}
static inline uint64_t xnor_int64(uint64_t a, uint64_t b) {
return ~(a^b);//按位异或再a按位取反
}
static inline uint32_t fill_bit_int32(char src) {
if (src == 0) return 0x00000000;
else return 0xFFFFFFFF;
}
static inline uint64_t fill_bit_int64(char src) {
if (src == 0) return 0x0000000000000000;
else return 0xFFFFFFFFFFFFFFFF;
}
void binary_int32_printf(uint32_t src) {
int i;
for (i = 0; i < 32; ++i) {
if (src & 1) printf("1");
else printf("0");
src = src >> 1;
}
printf("\n");
}
void binary_int64_printf(uint64_t src) {
int i;
for (i = 0; i < 64; ++i) {
if (src & 1) printf("1");
else printf("0");
src = src >> 1;
}
printf("\n");
}
//----------------------------
// is not used
/*
void transpose_32x32_bits_my(uint32_t *A, uint32_t *B, int lda, int ldb)
{
unsigned int x, y;
for (y = 0; y < 32; ++y) {
for (x = 0; x < 32; ++x) {
if (A[y * lda] & ((uint32_t)1 << x)) B[x * ldb] |= (uint32_t)1 << y;
}
}
}
*/
#ifndef GPU
uint8_t reverse_8_bit(uint8_t a) {//头尾反转一个字节里的bits
return ((a * 0x0802LU & 0x22110LU) | (a * 0x8020LU & 0x88440LU)) * 0x10101LU >> 16;
}
uint32_t reverse_32_bit(uint32_t a)
{
// unsigned int __rbit(unsigned int val) // for ARM //__asm__("rbit %0, %1\n" : "=r"(output) : "r"(input));
return (reverse_8_bit(a >> 24) << 0) |
(reverse_8_bit(a >> 16) << 8) |
(reverse_8_bit(a >> 8) << 16) |
(reverse_8_bit(a >> 0) << 24);
}
#define swap(a0, a1, j, m) t = (a0 ^ (a1 >>j)) & m; a0 = a0 ^ t; a1 = a1 ^ (t << j);
void transpose32_optimized(uint32_t A[32]) {
int j, k;
unsigned m, t;
//m = 0x0000FFFF;
//for (j = 16; j != 0; j = j >> 1, m = m ^ (m << j)) {
// for (k = 0; k < 32; k = (k + j + 1) & ~j) {
// t = (A[k] ^ (A[k + j] >> j)) & m;
// A[k] = A[k] ^ t;
// A[k + j] = A[k + j] ^ (t << j);
// }
//}
j = 16;
m = 0x0000FFFF;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
j = 8;
m = 0x00ff00ff;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
j = 4;
m = 0x0f0f0f0f;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
j = 2;
m = 0x33333333;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
j = 1;
m = 0x55555555;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
// reverse Y
for (j = 0; j < 16; ++j) {
uint32_t tmp = A[j];
A[j] = reverse_32_bit(A[31 - j]);
A[31 - j] = reverse_32_bit(tmp);
}
}
void transpose_32x32_bits_reversed_diagonale(uint32_t *A, uint32_t *B, int m, int n)
{
unsigned A_tmp[32];
int i;
#pragma unroll
for (i = 0; i < 32; ++i) A_tmp[i] = A[i * m];
transpose32_optimized(A_tmp);
#pragma unroll
for (i = 0; i < 32; ++i) B[i*n] = A_tmp[i];
}
void transpose_8x8_bits_my(unsigned char *A, unsigned char *B, int lda, int ldb)
{
unsigned x, y;
for (y = 0; y < 8; ++y) {
for (x = 0; x < 8; ++x) {
if (A[y * lda] & (1 << x)) B[x * ldb] |= 1 << y;
}
}
}
unsigned char reverse_byte_1(char a)
{
return ((a & 0x1) << 7) | ((a & 0x2) << 5) |
((a & 0x4) << 3) | ((a & 0x8) << 1) |
((a & 0x10) >> 1) | ((a & 0x20) >> 3) |
((a & 0x40) >> 5) | ((a & 0x80) >> 7);
}
unsigned char reverse_byte(unsigned char a)
{
return ((a * 0x0802LU & 0x22110LU) | (a * 0x8020LU & 0x88440LU)) * 0x10101LU >> 16;
}
static unsigned char lookup[16] = {
0x0, 0x8, 0x4, 0xc, 0x2, 0xa, 0x6, 0xe,
0x1, 0x9, 0x5, 0xd, 0x3, 0xb, 0x7, 0xf, };
unsigned char reverse_byte_3(unsigned char n) {
// Reverse the top and bottom nibble then swap them.
return (lookup[n & 0b1111] << 4) | lookup[n >> 4];
}
void transpose8rS32_reversed_diagonale(unsigned char* A, unsigned char* B, int m, int n)
{
unsigned x, y, t;
x = y = 0;
// Load the array and pack it into x and y.
//x = (A[0] << 24) | (A[m] << 16) | (A[2 * m] << 8) | A[3 * m];
//y = (A[4 * m] << 24) | (A[5 * m] << 16) | (A[6 * m] << 8) | A[7 * m];
t = (x ^ (x >> 7)) & 0x00AA00AA; x = x ^ t ^ (t << 7);
t = (y ^ (y >> 7)) & 0x00AA00AA; y = y ^ t ^ (t << 7);
t = (x ^ (x >> 14)) & 0x0000CCCC; x = x ^ t ^ (t << 14);
t = (y ^ (y >> 14)) & 0x0000CCCC; y = y ^ t ^ (t << 14);
t = (x & 0xF0F0F0F0) | ((y >> 4) & 0x0F0F0F0F);
y = ((x << 4) & 0xF0F0F0F0) | (y & 0x0F0F0F0F);
x = t;
B[7 * n] = reverse_byte(x >> 24); B[6 * n] = reverse_byte(x >> 16); B[5 * n] = reverse_byte(x >> 8); B[4 * n] = reverse_byte(x);
B[3 * n] = reverse_byte(y >> 24); B[2 * n] = reverse_byte(y >> 16); B[1 * n] = reverse_byte(y >> 8); B[0 * n] = reverse_byte(y);
}
/*
// transpose by 8-bit
void transpose_bin(char *A, char *B, const int n, const int m,
const int lda, const int ldb, const int block_size)
{
//printf("\n n = %d, ldb = %d \t\t m = %d, lda = %d \n", n, ldb, m, lda);
int i;
#pragma omp parallel for
for (i = 0; i < n; i += 8) {
int j;
for (j = 0; j < m; j += 8) {
int a_index = i*lda + j;
int b_index = j*ldb + i;
//transpose_8x8_bits_my(&A[a_index/8], &B[b_index/8], lda/8, ldb/8);
transpose8rS32_reversed_diagonale(&A[a_index / 8], &B[b_index / 8], lda / 8, ldb / 8);
}
for (; j < m; ++j) {
if (get_bit(A, i*lda + j)) set_bit(B, j*ldb + i);
}
}
}
*/
#endif
// transpose by 32-bit
void transpose_bin(uint32_t *A, uint32_t *B, const int n, const int m,
const int lda, const int ldb, const int block_size)
{
//printf("\n n = %d (n mod 32 = %d), m = %d (m mod 32 = %d) \n", n, n % 32, m, m % 32);
//printf("\n lda = %d (lda mod 32 = %d), ldb = %d (ldb mod 32 = %d) \n", lda, lda % 32, ldb, ldb % 32);
int i;
#pragma omp parallel for
for (i = 0; i < n; i += 32) {
int j;
for (j = 0; j < m; j += 32) {
int a_index = i*lda + j;
int b_index = j*ldb + i;
transpose_32x32_bits_reversed_diagonale(&A[a_index / 32], &B[b_index / 32], lda / 32, ldb / 32);
//transpose_32x32_bits_my(&A[a_index/32], &B[b_index/32], lda/32, ldb/32);
}
for (; j < m; ++j) {
if (get_bit((const unsigned char* const)A, i * lda + j)) set_bit((unsigned char* const)B, j * ldb + i);
}
}
}
static inline int popcnt_32(uint32_t val32) {
#ifdef WIN32 // Windows MSVS
int tmp_count = __popcnt(val32);
#else // Linux GCC
int tmp_count = __builtin_popcount(val32);
#endif
return tmp_count;
}
//----------------------------
#if (defined(__AVX__) && defined(__x86_64__)) || (defined(_WIN64) && !defined(__MINGW32__))
#ifdef _WIN64
#include <intrin.h>
#include <ammintrin.h>
#include <immintrin.h>
#include <smmintrin.h>
#if defined(_MSC_VER) && _MSC_VER <= 1900
static inline __int32 _mm256_extract_epi64(__m256i a, const int index) {
return a.m256i_i64[index];
}
static inline __int32 _mm256_extract_epi32(__m256i a, const int index) {
return a.m256i_i32[index];
}
#endif
static inline float _castu32_f32(uint32_t a) {
return *((float *)&a);
}
static inline float _mm256_extract_float32(__m256 a, const int index) {
return a.m256_f32[index];
}
#else // Linux GCC/Clang
#include <x86intrin.h>
#include <ammintrin.h>
#include <immintrin.h>
#include <smmintrin.h>
#include <cpuid.h>
static inline float _castu32_f32(uint32_t a) {
return *((float *)&a);
}
static inline float _mm256_extract_float32(__m256 a, const int index) {
switch(index) {
case 0:
return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 0));
case 1:
return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 1));
case 2:
return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 2));
case 3:
return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 3));
case 4:
return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 4));
case 5:
return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 5));
case 6:
return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 6));
case 7:
return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 7));
default:
return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 0));
}
}
void asm_cpuid(uint32_t* abcd, uint32_t eax)
{
uint32_t ebx = 0, edx = 0, ecx = 0;
// EBX is saved to EDI and later restored
__asm__("movl %%ebx, %%edi;"
"cpuid;"
"xchgl %%ebx, %%edi;"
: "=D"(ebx),
"+a"(eax), "+c"(ecx), "=d"(edx));
abcd[0] = eax;
abcd[1] = ebx;
abcd[2] = ecx;
abcd[3] = edx;
}
#endif
#ifdef _WIN32
// Windows
#define cpuid(info, x) __cpuidex(info, x, 0)
#else
// GCC Intrinsics
void cpuid(int info[4], int InfoType) {
__cpuid_count(InfoType, 0, info[0], info[1], info[2], info[3]);
}
#endif
// Misc.
static int HW_MMX, HW_x64, HW_RDRAND, HW_BMI1, HW_BMI2, HW_ADX, HW_PREFETCHWT1;
static int HW_ABM; // Advanced Bit Manipulation
// SIMD: 128-bit
static int HW_SSE, HW_SSE2, HW_SSE3, HW_SSSE3, HW_SSE41, HW_SSE42, HW_SSE4a, HW_AES, HW_SHA;
// SIMD: 256-bit
static int HW_AVX, HW_XOP, HW_FMA3, HW_FMA4, HW_AVX2;
// SIMD: 512-bit
static int HW_AVX512F; // AVX512 Foundation
static int HW_AVX512CD; // AVX512 Conflict Detection
static int HW_AVX512PF; // AVX512 Prefetch
static int HW_AVX512ER; // AVX512 Exponential + Reciprocal
static int HW_AVX512VL; // AVX512 Vector Length Extensions
static int HW_AVX512BW; // AVX512 Byte + Word
static int HW_AVX512DQ; // AVX512 Doubleword + Quadword
static int HW_AVX512IFMA; // AVX512 Integer 52-bit Fused Multiply-Add
static int HW_AVX512VBMI; // AVX512 Vector Byte Manipulation Instructions
// https://stackoverflow.com/questions/6121792/how-to-check-if-a-cpu-supports-the-sse3-instruction-set
void check_cpu_features(void) {
int info[4];
cpuid(info, 0);//#define cpuid(info,x) __cpuidex(info, x, 0)
//cpuid(info, 0)-->__cpuidex(info, 0, 0)
/*
关于cpuid
CPUID指令主要可用于判断Intel架构CPU的制造商,比如在一些细微的功能实现上,Intel和AMD有差别,要用到此功能的代码就需要先判断制造商.
CPUID指令也可用来检测某功能是否支持,从而选择实现方式,比如判断当前CPU是否支持SSE2指令集,从而决定是采用优化浮点计算的SSE2，还是X87浮点计算.
CPUID汇编指令接受输入的寄存器是EAX，存放需要的子功能号，从0x00开始；存放输出的寄存器是EAX，EBX，ECX，EDX。
CPUID的子功能集有两大类，基本和扩展。
基本功能：
【https://blog.csdn.net/razor87/article/details/8711712?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-4.nonecase&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-4.nonecase】
EAX=00H-->获得最大可用的基本功能号，以及CPU Vendor ID
EAX=01H-->获得CPU各项基本属性，包括Family, Model, Stepping ID, 和其他特性支持描述符
EAX=02H-->CPU缓存和TLB的基本描述
EAX=03H-->CPU序列号 （由于隐私考虑，在Pentium III以后处理器中该功能被禁止）
__cpuid / __cpuidex 函数是微软对CPUID汇编指令的C++封装，易于使用，不再需要操作寄存器，而是int32数组，方便高级程序访问CPU信息。
函数声明如下：
void __cpuid(
int CPUInfo[4],
int InfoType
);
void __cpuidex(
int CPUInfo[4],
int InfoType,
int ECXValue
);
InfoType就是输入的EAX值，CPUInfo就是输出的int数组，CPUInfo[0]，CPUInfo[1]，CPUInfo[2]，CPUInfo[3]分别对应EAX，EBX，ECX，EDX寄存器。
ECXValue就是输入的ECX寄存器值。
本函数中：__cpuidex(info, 0, 0)
info-->CPUInfo
0-->InfoType
所以有：
EAX 返回CPU基本信息时EAX能接受的最大值
EBX “Genu”
ECX “intel”
EDX “ineI”
*/
int nIds = info[0];//返回CPU基本信息时EAX能接受的最大值
cpuid(info, 0x80000000);
/*
扩展功能：
EAX = 80000000H 获得最大可用的扩展功能号
EAX = 80000001H 获得扩展CPU信息和支持的扩展CPU功能表
EAX = 80000002H,80000003H,80000004H 获得CPU描述字符串
EAX = 80000005H 获得CPU L1 Cache 和 TLB 属性描述
EAX = 80000006H 获得CPU L2 Cache 的属性描述
EAX = 80000007H 获得高级电源管理的属性描述 （APMI）
*/
unsigned nExIds = info[0];//获得最大可用的扩展功能号
//检查CPU是否支持SSE3指令集
//这里不太懂，网上能查到一模一样的代码，但是我不是很懂
//就是说这里检查寄存器中的指令集
/*
应用程序是通过三个步奏来检查AVX2指令集的——
1) 使用cpuid指令的功能1，检测OSXSAVE和AVX标志。
2) 使用cpuid指令的功能7，检测AVX2标志。
3) 使用XGETBV指令获取XCR0寄存器的值，判断操作系统是否支持XMM和YMM状态。
*/
// Detect Features
if (nIds >= 0x00000001) {
cpuid(info, 0x00000001);
HW_MMX = (info[3] & ((uint32_t)1 << 23)) != 0;
HW_SSE = (info[3] & ((uint32_t)1 << 25)) != 0;
HW_SSE2 = (info[3] & ((uint32_t)1 << 26)) != 0;
HW_SSE3 = (info[2] & ((uint32_t)1 << 0)) != 0;
HW_SSSE3 = (info[2] & ((uint32_t)1 << 9)) != 0;
HW_SSE41 = (info[2] & ((uint32_t)1 << 19)) != 0;
HW_SSE42 = (info[2] & ((uint32_t)1 << 20)) != 0;
HW_AES = (info[2] & ((uint32_t)1 << 25)) != 0;
HW_AVX = (info[2] & ((uint32_t)1 << 28)) != 0;
HW_FMA3 = (info[2] & ((uint32_t)1 << 12)) != 0;
HW_RDRAND = (info[2] & ((uint32_t)1 << 30)) != 0;
}
if (nIds >= 0x00000007) {
cpuid(info, 0x00000007);
HW_AVX2 = (info[1] & ((uint32_t)1 << 5)) != 0;
HW_BMI1 = (info[1] & ((uint32_t)1 << 3)) != 0;
HW_BMI2 = (info[1] & ((uint32_t)1 << 8)) != 0;
HW_ADX = (info[1] & ((uint32_t)1 << 19)) != 0;
HW_SHA = (info[1] & ((uint32_t)1 << 29)) != 0;
HW_PREFETCHWT1 = (info[2] & ((uint32_t)1 << 0)) != 0;
HW_AVX512F = (info[1] & ((uint32_t)1 << 16)) != 0;
HW_AVX512CD = (info[1] & ((uint32_t)1 << 28)) != 0;
HW_AVX512PF = (info[1] & ((uint32_t)1 << 26)) != 0;
HW_AVX512ER = (info[1] & ((uint32_t)1 << 27)) != 0;
HW_AVX512VL = (info[1] & ((uint32_t)1 << 31)) != 0;
HW_AVX512BW = (info[1] & ((uint32_t)1 << 30)) != 0;
HW_AVX512DQ = (info[1] & ((uint32_t)1 << 17)) != 0;
HW_AVX512IFMA = (info[1] & ((uint32_t)1 << 21)) != 0;
HW_AVX512VBMI = (info[2] & ((uint32_t)1 << 1)) != 0;
}
if (nExIds >= 0x80000001) {
cpuid(info, 0x80000001);
HW_x64 = (info[3] & ((uint32_t)1 << 29)) != 0;
HW_ABM = (info[2] & ((uint32_t)1 << 5)) != 0;
HW_SSE4a = (info[2] & ((uint32_t)1 << 6)) != 0;
HW_FMA4 = (info[2] & ((uint32_t)1 << 16)) != 0;
HW_XOP = (info[2] & ((uint32_t)1 << 11)) != 0;
}
}
//这里应该就是判断是不是使用了AVX指令集
//AVX对CPU的性能有提升，感兴趣的可以去看一下AVX指令集相关资料
int is_avx() {
static int result = -1;
if (result == -1) {
check_cpu_features();
result = HW_AVX;
if (result == 1) printf(" Used AVX \n");
else printf(" Not used AVX \n");
}
return result;
}
int is_fma_avx2() {
static int result = -1;
if (result == -1) {
check_cpu_features();
result = HW_FMA3 && HW_AVX2;
if (result == 1) printf(" Used FMA & AVX2 \n");
else printf(" Not used FMA & AVX2 \n");
}
return result;
}
// https://software.intel.com/sites/landingpage/IntrinsicsGuide
/**
* 被gemm_cpu函数调用，实际完成 C = ALPHA * A * B + C 矩阵运算，输出的C也是按行存储(所有行并成一行)
* @param M A，C的行数（不做转置）
* @param N B，C的列数（不做装置）
* @param K A的列数，B的行数（不做转置）
* @param ALPHA 系数
* @param A 输入矩阵(一维数组格式)
* @param lda A的列数（不做转置）
* @param B 输入矩阵(一维数组格式)
* @param ldb B的列数（不做转置）
* @param C 输入矩阵(一维数组格式)
* @param ldc C的列数（不做转置）
*
* 说明：此函数在gemm_cpu()函数中调用，是其中四中情况之一，A不进行转置，B不进行转置
* 函数名gemm_nn()中nt分别表示 not transpose, not transpose
*/
/*
举个栗子
A=[ [1,2,3,4], B=[ [1,0],
[2,3,4,5], [1,1],
[3,4,5,6]] [1,2],
[0,1] ]
a=[1,2,3,4,2,3,4,5,3,4,5,6]
b=[1,0,1,1,1,2,0,1]
D=A*B
D[0,0]=1*1+1*2+1*3+0*4=6
D[0,1]=1*0+2*1+3*2+4*1=12
D[0,:]=A[0,0]*B[0,:]+A[0,1]*B[1,:]+A[0,2]*B[2,:]+A[0,3]*B[3,:]
=[1,0]+[2,2]+[3,6]+[0,4]
=[6,12]
D[i,:]=A[i,0]*B[0,:]+A[i,1]*B[1,:]+A[i,2]*B[2,:]+A[i,3]*B[3,:]
其实归纳起来整个数组的计算过程是，C数组的第i行的值等于A数组的第i行分别去对B数组做乘法相加得到的。
*/
//可以看看博客学习一下 【https://blog.csdn.net/grafx/article/details/20001589?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.nonecase&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.nonecase】
void gemm_nn(int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i, j, k;
if (is_avx() == 1) { // AVX，Inter内部的数据操作指令
for (i = 0; i < M; ++i) {
for (k = 0; k < K; ++k) {
float A_PART = ALPHA*A[i*lda + k];//计算a*A矩阵的结果，常数乘数组
__m256 a256, b256, c256, result256; // AVX，_m256是AVX中的数据类型，包含8个float类型的数字向量
a256 = _mm256_set1_ps(A_PART);//AVX中的函数：_mm256_set1_ps--用一个float类型的数填充向量,
//a256={A_PART,A_PART,A_PART,A_PART,A_PART,A_PART,A_PART,A_PART}
for (j = 0; j < N - 8; j += 8) {
//从循环来看是一次加载8个位的数据
//查资料有：__m128，如果使用sizeof(__m128)计算该类型大小，结果为16，即等于四个浮点数长度。
//类推__m256应该是8个浮点数的长度。
b256 = _mm256_loadu_ps(&B[k*ldb + j]);//从未对齐的内存地址加载浮点向量,获得数组B的相应值
/*
所以这里举个栗子 假设j=0,那么应该是B[k*ldb+0]--B[k*ldb+7];C[k*ldb+0]--C[k*ldb+7]
*/
c256 = _mm256_loadu_ps(&C[i*ldc + j]);//从未对齐的内存地址加载浮点向量,获得数组C的相应值
// FMA - Intel Haswell (2013), AMD Piledriver (2012)
//result256 = _mm256_fmadd_ps(a256, b256, c256);
result256 = _mm256_mul_ps(a256, b256);//对两个float类型的向量进行相乘,对数组A,B相应的值进行点乘的操作
result256 = _mm256_add_ps(result256, c256);//对两个浮点向量做加法，加上C数组中相应的值
_mm256_storeu_ps(&C[i*ldc + j], result256);//将计算的值，反填充回C数组
}
int prev_end = (N % 8 == 0) ? (N - 8) : (N / 8) * 8;//确认B，C的行数是否为8的倍数,把没有计算填充数组序号求出来
for (j = prev_end; j < N; ++j)
C[i*ldc + j] += A_PART*B[k*ldb + j];//把没计算的数组值计算好并填充
}
}
}
else {// 如果不使用Inter内部的数据操作指令，就是很简单的点乘相加的过程
for (i = 0; i < M; ++i) {
// 大循环：遍历A的每一行，i表示A的第i行，也是C的第i行
for (k = 0; k < K; ++k) {
// 中循环：遍历每一行的所有列，k表示A的第k列，同时表示B的第k行
PUT_IN_REGISTER float A_PART = ALPHA * A[i * lda + k];// 先计算ALPHA * A (A中每一个元素乘以ALPHA)
for (j = 0; j < N; ++j) {
// 内循环：遍历B中所有列，每次大循环完毕，将计算得到A×B一行的结果
// j是B的第j列，也是C的第j列
C[i*ldc + j] += A_PART*B[k*ldb + j];
// A中第i行k列与B中第k行i列对应相乘，因为一个大循环要计算A×B一行的结果
// 因此，这里用一个内循环，并没有直接乘以B[k*ldb+i]
// 每个内循环完毕，将计算A×B整行的部分结果（A中第i行k列与B所有列第k行所有元素相乘的结果）
}
/* // SSE
__m128 a128, b128, c128, result128; // SSE
a128 = _mm_set1_ps(A_PART);
for (j = 0; j < N - 4; j += 4) {
b128 = _mm_loadu_ps(&B[k*ldb + j]);
c128 = _mm_loadu_ps(&C[i*ldc + j]);
//result128 = _mm_fmadd_ps(a128, b128, c128);
result128 = _mm_mul_ps(a128, b128);
result128 = _mm_add_ps(result128, c128);
_mm_storeu_ps(&C[i*ldc + j], result128);
}
int prev_end = (N % 4 == 0) ? (N - 4) : (N / 4) * 4;
for (j = prev_end; j < N; ++j){
C[i*ldc + j] += A_PART*B[k*ldb + j];
}
*/
}
}
}
}
//gemm_nn加速版，默认要计算的数组很大，不然都没有数据填充计算来着
void gemm_nn_fast(int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i;
#pragma omp parallel for
//是OpenMP中的一个指令，表示接下来的for循环将被多线程执行，另外每次循环之间不能有关系
for (i = 0; i < (M / TILE_M)*TILE_M; i += TILE_M)//#define TILE_M 4
//对数组A和C来说一下子要跑4行数据
//#define TILE_M 4 // 4 ops
//#define TILE_N 16 // AVX2 = 2 ops * 8 floats
//#define TILE_K 16 // loop
{
int j, k;
int i_d, k_d;
for (k = 0; k < (K / TILE_K)*TILE_K; k += TILE_K) //#define TILE_K 16
//对数组A来说一下子还要跑十六列数据,B数组是十六行
{
for (j = 0; j < (N / TILE_N)*TILE_N; j += TILE_N)//#define TILE_N 16
{
// L1 - 6 bits tag [11:6] - cache size 32 KB, conflict for each 4 KB
// L2 - 9 bits tag [14:6] - cache size 256 KB, conflict for each 32 KB
// L3 - 13 bits tag [18:6] - cache size 8 MB, conflict for each 512 KB
//跟gemm_nn函数一样，先定义需要的变量
__m256 result256;
__m256 a256_0, b256_0; // AVX
__m256 a256_1, b256_1; // AVX
__m256 a256_2;// , b256_2; // AVX
__m256 a256_3;// , b256_3; // AVX
__m256 c256_0, c256_1, c256_2, c256_3;
__m256 c256_4, c256_5, c256_6, c256_7;
//加载数据
c256_0 = _mm256_loadu_ps(&C[(0 + i)*ldc + (0 + j)]);
//假设j=0,C[(0 + i)*ldc+0]--C[(0 + i)*ldc+7]
c256_1 = _mm256_loadu_ps(&C[(1 + i)*ldc + (0 + j)]);
//假设j=0,C[(1 + i)*ldc+0]--C[(1 + i)*ldc+7]
c256_2 = _mm256_loadu_ps(&C[(0 + i)*ldc + (8 + j)]);
//假设j=0,C[(0 + i)*ldc+8]--C[(0 + i)*ldc+15]
c256_3 = _mm256_loadu_ps(&C[(1 + i)*ldc + (8 + j)]);
//假设j=0,C[(1 + i)*ldc+8]--C[(1 + i)*ldc+15]
//所以上面这个操作是一下子取了两行十六列的浮点数组数据
//同样的下面这个操作也是一下子取了两行十六列的浮点数组数据
//也是满足了循环的要求，i就是一次性跳4
c256_4 = _mm256_loadu_ps(&C[(2 + i)*ldc + (0 + j)]);
c256_5 = _mm256_loadu_ps(&C[(3 + i)*ldc + (0 + j)]);
c256_6 = _mm256_loadu_ps(&C[(2 + i)*ldc + (8 + j)]);
c256_7 = _mm256_loadu_ps(&C[(3 + i)*ldc + (8 + j)]);
for (k_d = 0; k_d < (TILE_K); ++k_d)//#define TILE_K 16
{
//对于数组A来说，按行分别计算，16列的数据（每列）乘以ALPHA的值，并填充到a256_0中，刚好是4行
a256_0 = _mm256_set1_ps(ALPHA*A[(0 + i)*lda + (k_d + k)]);
//A_PART_i_k_d=ALPHA*A[(0 + i)*lda + (k_d + k)]
//若i=0且k_d=0，a256_0={A_PART_0_0,A_PART_0_0,A_PART_0_0,A_PART_0_0,A_PART_0_0,A_PART_0_0,A_PART_0_0,A_PART_0_0}
//表示，A[0,0]*ALPHA,复制8份
//下面是一样的A[0,1]*ALPHA，A[0,2]*ALPHA，A[0,3]*ALPHA,分别复制8份
a256_1 = _mm256_set1_ps(ALPHA*A[(1 + i)*lda + (k_d + k)]);
a256_2 = _mm256_set1_ps(ALPHA*A[(2 + i)*lda + (k_d + k)]);
a256_3 = _mm256_set1_ps(ALPHA*A[(3 + i)*lda + (k_d + k)]);
b256_0 = _mm256_loadu_ps(&B[(k_d + k)*ldb + (0 + j)]);//加载数组B的数据，一次8列
b256_1 = _mm256_loadu_ps(&B[(k_d + k)*ldb + (8 + j)]);
// FMA - Intel Haswell (2013), AMD Piledriver (2012)
//c256_0 = _mm256_fmadd_ps(a256_0, b256_0, c256_0);
//c256_1 = _mm256_fmadd_ps(a256_1, b256_0, c256_1);
//c256_2 = _mm256_fmadd_ps(a256_0, b256_1, c256_2);
//c256_3 = _mm256_fmadd_ps(a256_1, b256_1, c256_3);
//c256_4 = _mm256_fmadd_ps(a256_2, b256_0, c256_4);
//c256_5 = _mm256_fmadd_ps(a256_3, b256_0, c256_5);
//c256_6 = _mm256_fmadd_ps(a256_2, b256_1, c256_6);
//c256_7 = _mm256_fmadd_ps(a256_3, b256_1, c256_7);
result256 = _mm256_mul_ps(a256_0, b256_0);//i=0,j=0,k_d=0时，A[0,0]*B[0,:]
c256_0 = _mm256_add_ps(result256, c256_0);//i=0,j=0,k_d=0时，c256_0=A[0,0]*B[0,:]+C[0,:]
result256 = _mm256_mul_ps(a256_1, b256_0);//i=0,j=0,k_d=0时，A[1,0]*B[0,:]
c256_1 = _mm256_add_ps(result256, c256_1);//i=0,j=0,k_d=0时，c256_1=A[1,0]*B[0,:]+C[1,:]
//后面就类推啦
result256 = _mm256_mul_ps(a256_0, b256_1);
c256_2 = _mm256_add_ps(result256, c256_2);
result256 = _mm256_mul_ps(a256_1, b256_1);
c256_3 = _mm256_add_ps(result256, c256_3);
result256 = _mm256_mul_ps(a256_2, b256_0);
c256_4 = _mm256_add_ps(result256, c256_4);
result256 = _mm256_mul_ps(a256_3, b256_0);
c256_5 = _mm256_add_ps(result256, c256_5);
result256 = _mm256_mul_ps(a256_2, b256_1);
c256_6 = _mm256_add_ps(result256, c256_6);
result256 = _mm256_mul_ps(a256_3, b256_1);
c256_7 = _mm256_add_ps(result256, c256_7);
}
//一样的反向把数据从寄存器推给数组内存
_mm256_storeu_ps(&C[(0 + i)*ldc + (0 + j)], c256_0);
_mm256_storeu_ps(&C[(1 + i)*ldc + (0 + j)], c256_1);
_mm256_storeu_ps(&C[(0 + i)*ldc + (8 + j)], c256_2);
_mm256_storeu_ps(&C[(1 + i)*ldc + (8 + j)], c256_3);
_mm256_storeu_ps(&C[(2 + i)*ldc + (0 + j)], c256_4);
_mm256_storeu_ps(&C[(3 + i)*ldc + (0 + j)], c256_5);
_mm256_storeu_ps(&C[(2 + i)*ldc + (8 + j)], c256_6);
_mm256_storeu_ps(&C[(3 + i)*ldc + (8 + j)], c256_7);
}
//查漏补缺，看有没有没有计算的
for (j = (N / TILE_N)*TILE_N; j < N; ++j) {
for (i_d = i; i_d < (i + TILE_M); ++i_d)
{
for (k_d = k; k_d < (k + TILE_K); ++k_d)
{
PUT_IN_REGISTER float A_PART = ALPHA*A[i_d*lda + k_d];
C[i_d*ldc + j] += A_PART*B[k_d*ldb + j];
}
}
}
}
//查漏补缺
for (k = (K / TILE_K)*TILE_K; k < K; ++k)
{
for (i_d = i; i_d < (i + TILE_M); ++i_d)
{
PUT_IN_REGISTER float A_PART = ALPHA*A[i_d*lda + k];
for (j = 0; j < N; ++j) {
C[i_d*ldc + j] += A_PART*B[k*ldb + j];
}
}
}
}
//查漏补缺
for (i = (M / TILE_M)*TILE_M; i < M; ++i) {
int j, k;
for (k = 0; k < K; ++k) {
PUT_IN_REGISTER float A_PART = ALPHA*A[i*lda + k];
for (j = 0; j < N; ++j) {
C[i*ldc + j] += A_PART*B[k*ldb + j];
}
}
}
}
void gemm_nn_bin_32bit_packed(int M, int N, int K, float ALPHA,
uint32_t *A, int lda,
uint32_t *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int i;
#pragma omp parallel for
for (i = 0; i < M; ++i) { // l.n
int j, s;
float mean_val = mean_arr[i];
//printf(" l.mean_arr[i] = %d \n ", l.mean_arr[i]);
for (s = 0; s < K; ++s) // l.size*l.size*l.c/32 or (l.size*l.size*l.c)
{
PUT_IN_REGISTER uint32_t A_PART = A[i*lda + s];
__m256i a256 = _mm256_set1_epi32(A_PART);
for (j = 0; j < N - 8; j += 8)
{
__m256i b256 = *((__m256i*)&B[s*ldb + j]);
__m256i xor256 = _mm256_xor_si256(a256, b256); // xnor = xor(a,b)
__m256i all_1 = _mm256_set1_epi8((char)255);
__m256i xnor256 = _mm256_andnot_si256(xor256, all_1); // xnor = not(xor(a,b))
// waiting for - CPUID Flags: AVX512VPOPCNTDQ: __m512i _mm512_popcnt_epi32(__m512i a)
__m256 count = _mm256_setr_ps(
popcnt_32(_mm256_extract_epi32(xnor256, 0)),
popcnt_32(_mm256_extract_epi32(xnor256, 1)),
popcnt_32(_mm256_extract_epi32(xnor256, 2)),
popcnt_32(_mm256_extract_epi32(xnor256, 3)),
popcnt_32(_mm256_extract_epi32(xnor256, 4)),
popcnt_32(_mm256_extract_epi32(xnor256, 5)),
popcnt_32(_mm256_extract_epi32(xnor256, 6)),
popcnt_32(_mm256_extract_epi32(xnor256, 7)));
__m256 val2 = _mm256_set1_ps(2);
count = _mm256_mul_ps(count, val2); // count * 2
__m256 val32 = _mm256_set1_ps(32);
count = _mm256_sub_ps(count, val32); // count - 32
__m256 mean256 = _mm256_set1_ps(mean_val);
count = _mm256_mul_ps(count, mean256); // count * mean_val
__m256 c256 = *((__m256*)&C[i*ldc + j]);
count = _mm256_add_ps(count, c256); // c = c + count
*((__m256*)&C[i*ldc + j]) = count;
}
for (; j < N; ++j) // out_h*out_w;
{
PUT_IN_REGISTER uint32_t B_PART = B[s*ldb + j];
uint32_t xnor_result = ~(A_PART ^ B_PART);
int32_t count = popcnt_32(xnor_result); // must be Signed int
C[i*ldc + j] += (2 * count - 32) * mean_val;
}
}
}
}
//二维卷积操作，给input,weights,得到outputs
void convolution_2d_old(int w, int h, int ksize, int n, int c, int pad, int stride,
float *weights, float *input, float *output)
{
//const int out_h = (h + 2 * pad - ksize) / stride + 1; // output_height=input_height for stride=1 and pad=1
//const int out_w = (w + 2 * pad - ksize) / stride + 1; // output_width=input_width for stride=1 and pad=1
int fil;
// filter index
//并行加速
#pragma omp parallel for // "omp parallel for" - automatic parallelization of loop by using OpenMP
for (fil = 0; fil < n; ++fil) {//n=输出的通道数
//int i, f, j;
int chan, y, x, f_y, f_x;
// channel index
for (chan = 0; chan < c; ++chan)
// input - y
for (y = 0; y < h; ++y)
// input - x
for (x = 0; x < w; ++x)
{
//计算输出特征的output index,第fil个输出通道数，第y行，第x列的位置
int const output_index = fil*w*h + y*w + x;
//fil*c*ksize*ksize-->第fil个输出通道数，第chan通道的卷积核的初始位置
int const weights_pre_index = fil*c*ksize*ksize + chan*ksize*ksize;
//计算输入特征的index,第几个通道的特征的初始位置
int const input_pre_index = chan*w*h;
float sum = 0;
// filter - y
//找到当前需要计算的卷积核参数和对应的特征层以后，进行小循环计算相应的输出
//卷积计算，循环卷积核的尺寸ksize
for (f_y = 0; f_y < ksize; ++f_y)
{
int input_y = y + f_y - pad;
// filter - x
for (f_x = 0; f_x < ksize; ++f_x)
{
int input_x = x + f_x - pad;
//确认即将要进行卷积的特征是否满足大小，不满足就说明无法进行卷积计算。
if (input_y < 0 || input_x < 0 || input_y >= h || input_x >= w) continue;
//对应卷积核的尺寸，去进行点乘和的操作
//这里是计算需要的index
int input_index = input_pre_index + input_y*w + input_x;
int weights_index = weights_pre_index + f_y*ksize + f_x;
//将这个通道特征和卷积核点乘和计算好的值存放好
sum += input[input_index] * weights[weights_index];
}
}
// l.output[filters][width][height] +=
// state.input[channels][width][height] *
// l.weights[filters][channels][filter_width][filter_height];
//最终得到该点的输出值：
//这里应该好懂吧--假设input feature size =3*5*5
// input feature size= 4*3*3[pad=0,stride=1]
// 那么： kernel szie =4*3*3
//sum的值就相当于一个卷积核3*3,与一层input feature做点乘和的值
//output[output_index]就是输出位置相应的3个这样的点乘和的值相加，即output feature的输出
output[output_index] += sum;
}
}
}
//加速版本的卷积操作
//稍后再看..
void convolution_2d(int w, int h, int ksize, int n, int c, int pad, int stride,
float *weights, float *input, float *output, float *mean)
{
//const int out_h = (h + 2 * pad - ksize) / stride + 1; // output_height=input_height for stride=1 and pad=1
//const int out_w = (w + 2 * pad - ksize) / stride + 1; // output_width=input_width for stride=1 and pad=1
int i;
#if defined(_OPENMP)
static int max_num_threads = 0;
if (max_num_threads == 0) {
max_num_threads = omp_get_max_threads();
//omp_set_num_threads( max_num_threads / 2);
}
#endif
//convolution_2d_old(w, h, ksize, n, c, pad, stride, weights, input, output);
//整型变量定义
__m256i all256_sing1 = _mm256_set_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000);
for (i = 0; i < ksize*ksize*n*c; i+=8) {
*((__m256*)&weights[i]) = _mm256_and_ps(*((__m256*)&weights[i]), _mm256_castsi256_ps(all256_sing1));
}
//for (i = 0; i < w*h*c; i += 8) {
//*((__m256*)&input[i]) = _mm256_and_ps(*((__m256*)&input[i]), _mm256_castsi256_ps(all256_sing1));
//}
//__m256i all256_last_zero = _mm256_set1_epi32(0xFFFFFFFF);
//all256_last_zero.m256i_i32[7] = 0;
__m256i all256_last_zero =
_mm256_set_epi32(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0x0);
__m256i idx256 = _mm256_set_epi32(0, 7, 6, 5, 4, 3, 2, 1);
//__m256 all256_sing1 = _mm256_set1_ps(0x80000000);
__m256 all256_one = _mm256_set1_ps(1);
__m256i all256i_one = _mm256_set1_epi32(1);
///__m256i src256 = _mm256_loadu_si256((__m256i *)(&src[i]));
///__m256i result256 = _mm256_and_si256(src256, all256_sing1); // check sign in 8 x 32-bit floats
int fil;
// filter index
#pragma omp parallel for // "omp parallel for" - automatic parallelization of loop by using OpenMP
for (fil = 0; fil < n; ++fil) {
int chan, y, x, f_y, f_x;
float cur_mean = fabs(mean[fil]);
__m256 mean256 = _mm256_set1_ps(cur_mean);
// channel index
//for (chan = 0; chan < c; ++chan)
// input - y
for (y = 0; y < h; ++y)
// input - x
for (x = 0; x < w-8; x+=8)
{
int const output_index = fil*w*h + y*w + x;
float sum = 0;
__m256 sum256 = _mm256_set1_ps(0);
for (chan = 0; chan < c; ++chan) {
int const weights_pre_index = fil*c*ksize*ksize + chan*ksize*ksize;
int const input_pre_index = chan*w*h;
// filter - y
for (f_y = 0; f_y < ksize; ++f_y)
{
int input_y = y + f_y - pad;
//__m256 in = *((__m256*)&input[input_pre_index + input_y*w]);
if (input_y < 0 || input_y >= h) continue;
//__m256 in = _mm256_loadu_ps(&input[input_pre_index + input_y*w + x - pad]);
// filter - x
for (f_x = 0; f_x < ksize; ++f_x)
{
int input_x = x + f_x - pad;
//if (input_y < 0 || input_x < 0 || input_y >= h || input_x >= w) continue;
int input_index = input_pre_index + input_y*w + input_x;
int weights_index = weights_pre_index + f_y*ksize + f_x;
//if (input_y < 0 || input_y >= h) continue;
//sum += input[input_index] * weights[weights_index];
__m256 in = *((__m256*)&input[input_index]);
__m256 w = _mm256_set1_ps(weights[weights_index]);
//__m256 w_sign = _mm256_and_ps(w, _mm256_castsi256_ps(all256_sing1)); // check sign in 8 x 32-bit floats
__m256 xor256 = _mm256_xor_ps(w, in);
//printf("\n xor256_1 = %f, xor256_2 = %f \n", xor256.m256_f32[0], xor256.m256_f32[1]);
//printf("\n in = %f, w = %f, xor256 = %f \n", in.m256_f32[0], w_sign.m256_f32[0], xor256.m256_f32[0]);
//__m256 pn1 = _mm256_and_ps(_mm256_castsi256_ps(all256i_one), xor256);
//sum256 = xor256;
sum256 = _mm256_add_ps(xor256, sum256);
//printf("\n --- \n");
//printf("\n 0 = %f, 1 = %f, 2 = %f, 3 = %f, 4 = %f, 5 = %f, 6 = %f, 7 = %f \n", in.m256_f32[0], in.m256_f32[1], in.m256_f32[2], in.m256_f32[3], in.m256_f32[4], in.m256_f32[5], in.m256_f32[6], in.m256_f32[7]);
if (f_x < ksize-1) {
//in = _mm256_permutevar8x32_ps(in, idx256);
//in = _mm256_and_ps(in, _mm256_castsi256_ps(all256_last_zero));
}
}
}
}
// l.output[filters][width][height] +=
// state.input[channels][width][height] *
// l.weights[filters][channels][filter_width][filter_height];
//output[output_index] += sum;
sum256 = _mm256_mul_ps(sum256, mean256);
//printf("\n cur_mean = %f, sum256 = %f, sum256 = %f, in = %f \n",
// cur_mean, sum256.m256_f32[0], sum256.m256_f32[1], input[input_pre_index]);
//__m256 out = *((__m256*)&output[output_index]);
//out = _mm256_add_ps(out, sum256);
//*((__m256*)&output[output_index]) = out;
*((__m256*)&output[output_index]) = sum256;
//_mm256_storeu_ps(&C[i*ldc + j], result256);
}
}
}
// http://graphics.stanford.edu/~seander/bithacks.html
// https://stackoverflow.com/questions/17354971/fast-counting-the-number-of-set-bits-in-m128i-register
// https://arxiv.org/pdf/1611.07612.pdf
static inline int popcnt128(__m128i n) {
const __m128i n_hi = _mm_unpackhi_epi64(n, n);
#if defined(_MSC_VER)
return __popcnt64(_mm_cvtsi128_si64(n)) + __popcnt64(_mm_cvtsi128_si64(n_hi));
#elif defined(__APPLE__) && defined(__clang__)
return _mm_popcnt_u64(_mm_cvtsi128_si64(n)) + _mm_popcnt_u64(_mm_cvtsi128_si64(n_hi));
#else
return __popcntq(_mm_cvtsi128_si64(n)) + __popcntq(_mm_cvtsi128_si64(n_hi));
#endif
}
static inline int popcnt256(__m256i n) {
return popcnt128(_mm256_extractf128_si256(n, 0)) + popcnt128(_mm256_extractf128_si256(n, 1));
}
static inline __m256i count256(__m256i v) {
__m256i lookup =
_mm256_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2,
2, 3, 2, 3, 3, 4, 0, 1, 1, 2, 1, 2, 2, 3,
1, 2, 2, 3, 2, 3, 3, 4);
__m256i low_mask = _mm256_set1_epi8(0x0f);
__m256i lo = _mm256_and_si256(v, low_mask);
__m256i hi = _mm256_and_si256(_mm256_srli_epi32(v, 4), low_mask);
__m256i popcnt1 = _mm256_shuffle_epi8(lookup, lo);
__m256i popcnt2 = _mm256_shuffle_epi8(lookup, hi);
__m256i total = _mm256_add_epi8(popcnt1, popcnt2);
return _mm256_sad_epu8(total, _mm256_setzero_si256());
}
static inline int popcnt256_custom(__m256i n) {
__m256i val = count256(n);
//return val.m256i_i64[0] +
//val.m256i_i64[1] +
//val.m256i_i64[2] +
//val.m256i_i64[3];
return _mm256_extract_epi64(val, 0)
+ _mm256_extract_epi64(val, 1)
+ _mm256_extract_epi64(val, 2)
+ _mm256_extract_epi64(val, 3);
}
static inline void xnor_avx2_popcnt(__m256i a_bit256, __m256i b_bit256, __m256i *count_sum) {
__m256i c_bit256 = _mm256_set1_epi8((char)255);
__m256i xor256 = _mm256_xor_si256(a_bit256, b_bit256); // xnor = not(xor(a,b))
c_bit256 = _mm256_andnot_si256(xor256, c_bit256); // can be optimized - we can do other NOT for wegihts once and do not do this NOT
*count_sum = _mm256_add_epi64(count256(c_bit256), *count_sum); // 1st part - popcnt Mula's algorithm
}
// 2nd part - popcnt Mula's algorithm
static inline int get_count_mula(__m256i count_sum) {
return _mm256_extract_epi64(count_sum, 0)
+ _mm256_extract_epi64(count_sum, 1)
+ _mm256_extract_epi64(count_sum, 2)
+ _mm256_extract_epi64(count_sum, 3);
}
// 5x times faster than gemm()-float32
// further optimizations: do mean-mult only for the last layer
void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
unsigned char *A, int lda,
unsigned char *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int i;
#if defined(_OPENMP)
static int max_num_threads = 0;
if (max_num_threads == 0) {
max_num_threads = omp_get_max_threads();
//omp_set_num_threads(max_num_threads / 2);
}
#endif
//#pragma omp parallel for
//for (i = 0; i < M; ++i)
#pragma omp parallel for
for (i = 0; i < (M/2)*2; i += 2)
{ // l.n - filters [16 - 55 - 1024]
float mean_val_0 = mean_arr[i + 0];
float mean_val_1 = mean_arr[i + 1];
int j, k;
//__m256i all_1 = _mm256_set1_epi8(255);
//for (j = 0; j < N; ++j)
for (j = 0; j < (N/2)*2; j += 2)
{ // out_h*out_w - one channel output size [169 - 173056]
//int count = 0;
const int bit_step = 256;
__m256i count_sum_0 = _mm256_set1_epi8(0);
__m256i count_sum_1 = _mm256_set1_epi8(0);
__m256i count_sum_2 = _mm256_set1_epi8(0);
__m256i count_sum_3 = _mm256_set1_epi8(0);
for (k = 0; k < K; k += bit_step) { // l.size*l.size*l.c - one filter size [27 - 9216]
__m256i a_bit256_0 = _mm256_loadu_si256((__m256i *)(A + ((i + 0)*lda + k) / 8));
__m256i b_bit256_0 = _mm256_loadu_si256((__m256i *)(B + ((j + 0)*ldb + k) / 8));
__m256i a_bit256_1 = _mm256_loadu_si256((__m256i *)(A + ((i + 1)*lda + k) / 8));
__m256i b_bit256_1 = _mm256_loadu_si256((__m256i *)(B + ((j + 1)*ldb + k) / 8));
xnor_avx2_popcnt(a_bit256_0, b_bit256_0, &count_sum_0);
xnor_avx2_popcnt(a_bit256_0, b_bit256_1, &count_sum_1);
xnor_avx2_popcnt(a_bit256_1, b_bit256_0, &count_sum_2);
xnor_avx2_popcnt(a_bit256_1, b_bit256_1, &count_sum_3);
//count += popcnt256(c_bit256);
//binary_int64_printf(c_bit64);
//printf(", count = %d \n\n", tmp_count);
}
int count_0 = get_count_mula(count_sum_0);
int count_1 = get_count_mula(count_sum_1);
int count_2 = get_count_mula(count_sum_2);
int count_3 = get_count_mula(count_sum_3);
const int f1 = (K % bit_step == 0) ? 0 : (bit_step - (K % bit_step));
count_0 = count_0 - f1; // remove extra bits (from empty space for align only)
count_1 = count_1 - f1;
count_2 = count_2 - f1;
count_3 = count_3 - f1;
C[i*ldc + (j + 0)] = (2 * count_0 - K) * mean_val_0;
C[i*ldc + (j + 1)] = (2 * count_1 - K) * mean_val_0;
C[(i + 1)*ldc + (j + 0)] = (2 * count_2 - K) * mean_val_1;
C[(i + 1)*ldc + (j + 1)] = (2 * count_3 - K) * mean_val_1;
}
int i_d;
for (i_d = 0; i_d < 2; ++i_d)
{
float mean_val = mean_arr[i + i_d];
for (j = (N / 2) * 2; j < N; j += 1)
{ // out_h*out_w - one channel output size [169 - 173056]
const int bit_step = 256;
__m256i count_sum = _mm256_set1_epi8(0);
for (k = 0; k < K; k += bit_step) { // l.size*l.size*l.c - one filter size [27 - 9216]
__m256i a_bit256_0 = _mm256_loadu_si256((__m256i *)(A + ((i + i_d + 0)*lda + k) / 8));
__m256i b_bit256_0 = _mm256_loadu_si256((__m256i *)(B + ((j + 0)*ldb + k) / 8));
xnor_avx2_popcnt(a_bit256_0, b_bit256_0, &count_sum);
}
int count = get_count_mula(count_sum);
const int f1 = (K % bit_step == 0) ? 0 : (bit_step - (K % bit_step));
count = count - f1; // remove extra bits (from empty space for align only)
C[(i + i_d)*ldc + j] = (2 * count - K) * mean_val;
}
}
}
for (i = (M / 2) * 2; i < M; i += 1)
{
float mean_val = mean_arr[i];
int j, k;
for (j = 0; j < N; j += 1)
{ // out_h*out_w - one channel output size [169 - 173056]
const int bit_step = 256;
__m256i count_sum = _mm256_set1_epi8(0);
for (k = 0; k < K; k += bit_step) { // l.size*l.size*l.c - one filter size [27 - 9216]
__m256i a_bit256_0 = _mm256_loadu_si256((__m256i *)(A + ((i + 0)*lda + k) / 8));
__m256i b_bit256_0 = _mm256_loadu_si256((__m256i *)(B + ((j + 0)*ldb + k) / 8));
xnor_avx2_popcnt(a_bit256_0, b_bit256_0, &count_sum);
}
int count = get_count_mula(count_sum);
const int f1 = (K % bit_step == 0) ? 0 : (bit_step - (K % bit_step));
count = count - f1; // remove extra bits (from empty space for align only)
C[i*ldc + j] = (2 * count - K) * mean_val;
}
}
}
//From Berkeley Vision's Caffe!
//https://github.com/BVLC/caffe/blob/master/LICENSE
void im2col_cpu_custom_transpose(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col, int ldb_align)
{
const int height_col = (height + 2 * pad - ksize) / stride + 1;
const int width_col = (width + 2 * pad - ksize) / stride + 1;
const int channels_col = channels * ksize * ksize;
int c;
// optimized version
if (height_col == height && width_col == width && stride == 1 && pad == 1)
{
#pragma omp parallel for
for (c = 0; c < channels_col; ++c) {
int h, w;
int w_offset = c % ksize;
int h_offset = (c / ksize) % ksize;
int c_im = c / ksize / ksize;
for (h = pad; h < height_col - pad; ++h) {
for (w = pad; w < width_col - pad - 4; w+=8) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = (h * width_col + w)*ldb_align + c; // transposed & aligned
//data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
__m256 src256 = _mm256_loadu_ps((float *)(&data_im[im_col + width*(im_row + height*c_im)]));
data_col[col_index + ldb_align * 0] = _mm256_extract_float32(src256, 0);// src256.m256_f32[0];
data_col[col_index + ldb_align * 1] = _mm256_extract_float32(src256, 1);// src256.m256_f32[1];
data_col[col_index + ldb_align * 2] = _mm256_extract_float32(src256, 2);// src256.m256_f32[2];
data_col[col_index + ldb_align * 3] = _mm256_extract_float32(src256, 3);// src256.m256_f32[3];
data_col[col_index + ldb_align * 4] = _mm256_extract_float32(src256, 4);// src256.m256_f32[4];
data_col[col_index + ldb_align * 5] = _mm256_extract_float32(src256, 5);// src256.m256_f32[5];
data_col[col_index + ldb_align * 6] = _mm256_extract_float32(src256, 6);// src256.m256_f32[6];
data_col[col_index + ldb_align * 7] = _mm256_extract_float32(src256, 7);// src256.m256_f32[7];
//_mm256_storeu_ps(&data_col[col_index], src256);
}
for (; w < width_col - pad; ++w) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
int col_index = (h * width_col + w)*ldb_align + c; // transposed & aligned
data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
}
}
{
w = 0;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (h * width_col + w)*ldb_align + c; // transposed & aligned
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
{
w = width_col - 1;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (h * width_col + w)*ldb_align + c; // transposed & aligned
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
{
h = 0;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (h * width_col + w)*ldb_align + c; // transposed & aligned
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
{
h = height_col - 1;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (h * width_col + w)*ldb_align + c; // transposed & aligned
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
}
}
else {
#pragma omp parallel for
for (c = 0; c < channels_col; ++c) {
int h, w;
int w_offset = c % ksize;
int h_offset = (c / ksize) % ksize;
int c_im = c / ksize / ksize;
for (h = 0; h < height_col; ++h) {
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h * stride;
int im_col = w_offset + w * stride;
int col_index = (h * width_col + w)*ldb_align + c; // transposed & aligned
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
}
}
}
//From Berkeley Vision's Caffe!
//https://github.com/BVLC/caffe/blob/master/LICENSE
void im2col_cpu_custom(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col)
{
int c;
const int height_col = (height + 2 * pad - ksize) / stride + 1;
const int width_col = (width + 2 * pad - ksize) / stride + 1;
const int channels_col = channels * ksize * ksize;
// optimized version
if (height_col == height && width_col == width && stride == 1 && pad == 1 && is_fma_avx2())
{
#pragma omp parallel for
for (c = 0; c < channels_col; ++c) {
int h, w;
int w_offset = c % ksize;
int h_offset = (c / ksize) % ksize;
int c_im = c / ksize / ksize;
for (h = pad; h < height_col-pad; ++h) {
for (w = pad; w < width_col-pad-8; w += 8) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
int col_index = (c * height_col + h) * width_col + w;
//data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
__m256 src256 = _mm256_loadu_ps((float *)(&data_im[im_col + width*(im_row + height*c_im)]));
_mm256_storeu_ps(&data_col[col_index], src256);
}
for (; w < width_col - pad; ++w) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
}
}
{
w = 0;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
{
w = width_col-1;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
{
h = 0;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
{
h = height_col-1;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
}
}
else {
//printf("\n Error: is no non-optimized version \n");
im2col_cpu(data_im, channels, height, width, ksize, stride, pad, data_col);
}
}
//From Berkeley Vision's Caffe!
//https://github.com/BVLC/caffe/blob/master/LICENSE
void im2col_cpu_custom_align(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col, int bit_align)
{
int c;
const int height_col = (height + 2 * pad - ksize) / stride + 1;
const int width_col = (width + 2 * pad - ksize) / stride + 1;
const int channels_col = channels * ksize * ksize;
// optimized version
if (height_col == height && width_col == width && stride == 1 && pad == 1 && is_fma_avx2())
{
int new_ldb = bit_align;
#pragma omp parallel for
for (c = 0; c < channels_col; ++c) {
int h, w;
int w_offset = c % ksize;
int h_offset = (c / ksize) % ksize;
int c_im = c / ksize / ksize;
for (h = pad; h < height_col - pad; ++h) {
for (w = pad; w < width_col - pad - 8; w += 8) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
__m256 src256 = _mm256_loadu_ps((float *)(&data_im[im_col + width*(im_row + height*c_im)]));
_mm256_storeu_ps(&data_col[col_index], src256);
}
for (; w < width_col - pad; ++w) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
}
}
{
w = 0;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
}
}
{
w = width_col - 1;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
}
}
{
h = 0;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
}
}
{
h = height_col - 1;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
}
}
}
}
else {
printf("\n Error: is no non-optimized version \n");
//im2col_cpu(data_im, channels, height, width, ksize, stride, pad, data_col); // must be aligned for transpose after float_to_bin
// float_to_bit(b, t_input, src_size);
// transpose_bin(t_input, *t_bit_input, k, n, bit_align, new_ldb, 8);
}
}
//From Berkeley Vision's Caffe!
//https://github.com/BVLC/caffe/blob/master/LICENSE
void im2col_cpu_custom_bin(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col, int bit_align)
{
int c;
const int height_col = (height + 2 * pad - ksize) / stride + 1;
const int width_col = (width + 2 * pad - ksize) / stride + 1;
const int channels_col = channels * ksize * ksize;
// optimized version
if (height_col == height && width_col == width && stride == 1 && pad == 1 && is_fma_avx2())
{
__m256i all256_sing1 = _mm256_set_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000);
__m256 float_zero256 = _mm256_set1_ps(0.00);
int new_ldb = bit_align;
#pragma omp parallel for
for (c = 0; c < channels_col; ++c) {
int h, w;
int w_offset = c % ksize;
int h_offset = (c / ksize) % ksize;
int c_im = c / ksize / ksize;
for (h = pad; h < height_col - pad; ++h) {
for (w = pad; w < width_col - pad - 8; w += 8) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//__m256i src256 = _mm256_loadu_si256((__m256i *)(&data_im[im_col + width*(im_row + height*c_im)]));
//__m256i result256 = _mm256_and_si256(src256, all256_sing1); // check sign in 8 x 32-bit floats
//uint16_t mask = _mm256_movemask_ps(_mm256_castsi256_ps(result256)); // (val >= 0) ? 0 : 1
//mask = ~mask; // inverse mask, (val >= 0) ? 1 : 0
__m256 src256 = _mm256_loadu_ps((float *)(&data_im[im_col + width*(im_row + height*c_im)]));
__m256 result256 = _mm256_cmp_ps(src256, float_zero256, _CMP_GT_OS);
uint16_t mask = _mm256_movemask_ps(result256); // (val > 0) ? 0 : 1
uint16_t* dst_ptr = (uint16_t*)&((uint8_t*)data_col)[col_index / 8];
*dst_ptr |= (mask << (col_index % 8));
}
for (; w < width_col - pad; ++w) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
float val = data_im[im_col + width*(im_row + height*c_im)];
if (val > 0) set_bit((unsigned char* const)data_col, col_index);
}
}
{
w = 0;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
float val = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
if (val > 0) set_bit((unsigned char* const)data_col, col_index);
}
}
{
w = width_col - 1;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
float val = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
if (val > 0) set_bit((unsigned char* const)data_col, col_index);
}
}
{
h = 0;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
float val = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
if (val > 0) set_bit((unsigned char* const)data_col, col_index);
}
}
{
h = height_col - 1;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
float val = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
if (val > 0) set_bit((unsigned char* const)data_col, col_index);
}
}
}
}
else {
printf("\n Error: is no non-optimized version \n");
//im2col_cpu(data_im, channels, height, width, ksize, stride, pad, data_col); // must be aligned for transpose after float_to_bin
// float_to_bit(b, t_input, src_size);
// transpose_bin(t_input, *t_bit_input, k, n, bit_align, new_ldb, 8);
}
}
void activate_array_cpu_custom(float *x, const int n, const ACTIVATION a)
{
int i = 0;
if (a == LINEAR)
{}
else if (a == LEAKY)
{
if (is_fma_avx2()) {
__m256i all256_sing1 = _mm256_set_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000);
__m256 all256_01 = _mm256_set1_ps(0.1F);
for (i = 0; i < n - 8; i += 8) {
//x[i] = (x[i]>0) ? x[i] : .1*x[i];
__m256 src256 = _mm256_loadu_ps(&x[i]);
__m256 mult256 = _mm256_mul_ps((src256), all256_01); // mult * 0.1
__m256i sign256 = _mm256_and_si256(_mm256_castps_si256(src256), all256_sing1); // check sign in 8 x 32-bit floats
__m256 result256 = _mm256_blendv_ps(src256, mult256, _mm256_castsi256_ps(sign256)); // (sign>0) ? src : mult;
_mm256_storeu_ps(&x[i], result256);
}
}
for (; i < n; ++i) {
x[i] = (x[i]>0) ? x[i] : .1*x[i];
}
}
else {
for (i = 0; i < n; ++i) {
x[i] = activate(x[i], a);
}
}
}
void float_to_bit(float *src, unsigned char *dst, size_t size)
{
size_t dst_size = size / 8 + 1;
memset(dst, 0, dst_size);
size_t i;
//__m256i all256_sing1 = _mm256_set_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000);
__m256 float_zero256 = _mm256_set1_ps(0.0);
for (i = 0; i < size; i+=8)
{
//__m256i src256 = _mm256_loadu_si256((__m256i *)(&src[i]));
//__m256i result256 = _mm256_and_si256(src256, all256_sing1); // check sign in 8 x 32-bit floats
//uint32_t mask = _mm256_movemask_ps(_mm256_castsi256_ps(result256)); // (val >= 0) ? 0 : 1
mask = ~mask; // inverse mask, (val >= 0) ? 1 : 0
__m256 src256 = _mm256_loadu_ps((float *)(&src[i]));
__m256 result256 = _mm256_cmp_ps(src256, float_zero256, _CMP_GT_OS);
uint32_t mask = _mm256_movemask_ps(result256); // (val > 0) ? 0 : 1
dst[i / 8] = mask;
}
}
static inline void transpose4x4_SSE(float *A, float *B, const int lda, const int ldb)
{
__m128 row1 = _mm_loadu_ps(&A[0 * lda]);
__m128 row2 = _mm_loadu_ps(&A[1 * lda]);
__m128 row3 = _mm_loadu_ps(&A[2 * lda]);
__m128 row4 = _mm_loadu_ps(&A[3 * lda]);
_MM_TRANSPOSE4_PS(row1, row2, row3, row4);
_mm_storeu_ps(&B[0 * ldb], row1);
_mm_storeu_ps(&B[1 * ldb], row2);
_mm_storeu_ps(&B[2 * ldb], row3);
_mm_storeu_ps(&B[3 * ldb], row4);
}
void transpose_block_SSE4x4(float *A, float *B, const int n, const int m,
const int lda, const int ldb, const int block_size)
{
int i;
#pragma omp parallel for
for (i = 0; i < n; i += block_size) {
int j, i2, j2;
//int max_i2 = (i + block_size < n) ? (i + block_size) : n;
if (i + block_size < n) {
int max_i2 = i + block_size;
for (j = 0; j < m; j += block_size) {
//int max_j2 = (j + block_size < m) ? (j + block_size) : m;
if (j + block_size < m) {
int max_j2 = j + block_size;
for (i2 = i; i2 < max_i2; i2 += 4) {
for (j2 = j; j2 < max_j2; j2 += 4) {
transpose4x4_SSE(&A[i2*lda + j2], &B[j2*ldb + i2], lda, ldb);
}
}
}
else {
for (i2 = i; i2 < max_i2; ++i2) {
for (j2 = j; j2 < m; ++j2) {
B[j2*ldb + i2] = A[i2*lda + j2];
}
}
}
}
}
else {
for (i2 = i; i2 < n; ++i2) {
for (j2 = 0; j2 < m; ++j2) {
B[j2*ldb + i2] = A[i2*lda + j2];
}
}
}
}
}
void forward_maxpool_layer_avx(float *src, float *dst, int *indexes, int size, int w, int h, int out_w, int out_h, int c,
int pad, int stride, int batch)
{
const int w_offset = -pad / 2;
const int h_offset = -pad / 2;
int b, k;
for (b = 0; b < batch; ++b) {
#pragma omp parallel for
for (k = 0; k < c; ++k) {
int i, j, m, n;
for (i = 0; i < out_h; ++i) {
//for (j = 0; j < out_w; ++j) {
j = 0;
if(stride == 1 && is_avx() == 1) {
for (j = 0; j < out_w - 8 - (size - 1); j += 8) {
int out_index = j + out_w*(i + out_h*(k + c*b));
__m256 max256 = _mm256_set1_ps(-FLT_MAX);
for (n = 0; n < size; ++n) {
for (m = 0; m < size; ++m) {
int cur_h = h_offset + i*stride + n;
int cur_w = w_offset + j*stride + m;
int index = cur_w + w*(cur_h + h*(k + b*c));
int valid = (cur_h >= 0 && cur_h < h &&
cur_w >= 0 && cur_w < w);
if (!valid) continue;
__m256 src256 = _mm256_loadu_ps(&src[index]);
max256 = _mm256_max_ps(src256, max256);
}
}
_mm256_storeu_ps(&dst[out_index], max256);
}
}
else if (size == 2 && stride == 2 && is_avx() == 1) {
for (j = 0; j < out_w - 4; j += 4) {
int out_index = j + out_w*(i + out_h*(k + c*b));
//float max = -FLT_MAX;
//int max_i = -1;
__m128 max128 = _mm_set1_ps(-FLT_MAX);
for (n = 0; n < size; ++n) {
//for (m = 0; m < size; ++m)
m = 0;
{
int cur_h = h_offset + i*stride + n;
int cur_w = w_offset + j*stride + m;
int index = cur_w + w*(cur_h + h*(k + b*c));
int valid = (cur_h >= 0 && cur_h < h &&
cur_w >= 0 && cur_w < w);
if (!valid) continue;
__m256 src256 = _mm256_loadu_ps(&src[index]);
__m256 src256_2 = _mm256_permute_ps(src256, (1 << 0) | (3 << 4));
__m256 max256 = _mm256_max_ps(src256, src256_2);
__m128 src128_0 = _mm256_extractf128_ps(max256, 0);
__m128 src128_1 = _mm256_extractf128_ps(max256, 1);
__m128 src128 = _mm_shuffle_ps(src128_0, src128_1, (2 << 2) | (2 << 6));
max128 = _mm_max_ps(src128, max128);
}
}
_mm_storeu_ps(&dst[out_index], max128);
}
}
for (; j < out_w; ++j) {
int out_index = j + out_w*(i + out_h*(k + c*b));
float max = -FLT_MAX;
int max_i = -1;
for (n = 0; n < size; ++n) {
for (m = 0; m < size; ++m) {
int cur_h = h_offset + i*stride + n;
int cur_w = w_offset + j*stride + m;
int index = cur_w + w*(cur_h + h*(k + b*c));
int valid = (cur_h >= 0 && cur_h < h &&
cur_w >= 0 && cur_w < w);
float val = (valid != 0) ? src[index] : -FLT_MAX;
max_i = (val > max) ? index : max_i;
max = (val > max) ? val : max;
}
}
dst[out_index] = max;
if (indexes) indexes[out_index] = max_i;
}
}
}
}
}
#else // AVX
int is_avx() {
return 0;
}
int is_fma_avx2() {
return 0;
}
void gemm_nn(int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i, j, k;
for (i = 0; i < M; ++i) {
for (k = 0; k < K; ++k) {
PUT_IN_REGISTER float A_PART = ALPHA * A[i * lda + k];
for (j = 0; j < N; ++j) {
C[i*ldc + j] += A_PART*B[k*ldb + j];
}
}
}
}
void gemm_nn_fast(int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i, j, k;
#pragma omp parallel for
for (i = 0; i < M; ++i) {
for (k = 0; k < K; ++k) {
PUT_IN_REGISTER float A_PART = ALPHA*A[i*lda + k];
for (j = 0; j < N; ++j) {
C[i*ldc + j] += A_PART*B[k*ldb + j];
}
}
}
}
void gemm_nn_bin_32bit_packed(int M, int N, int K, float ALPHA,
uint32_t *A, int lda,
uint32_t *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int i;
#pragma omp parallel for
for (i = 0; i < M; ++i) { // l.n
int j, s;
float mean_val = mean_arr[i];
//printf(" l.mean_arr[i] = %d \n ", l.mean_arr[i]);
for (s = 0; s < K; ++s) // l.size*l.size*l.c/32 or (l.size*l.size*l.c)
{
//PUT_IN_REGISTER float A_PART = 1*a[i*k + s];
PUT_IN_REGISTER uint32_t A_PART = A[i * lda + s];
for (j = 0; j < N; ++j) // out_h*out_w;
{
//c[i*n + j] += A_PART*b[s*n + j];
PUT_IN_REGISTER uint32_t B_PART = B[s * ldb + j];
uint32_t xnor_result = ~(A_PART ^ B_PART);
//printf(" xnor_result = %d, ", xnor_result);
int32_t count = popcnt_32(xnor_result); // must be Signed int
C[i*ldc + j] += (2 * count - 32) * mean_val;
//c[i*n + j] += count*mean;
}
}
}
}
void convolution_2d(int w, int h, int ksize, int n, int c, int pad, int stride,
float *weights, float *input, float *output, float *mean)
{
const int out_h = (h + 2 * pad - ksize) / stride + 1; // output_height=input_height for stride=1 and pad=1
const int out_w = (w + 2 * pad - ksize) / stride + 1; // output_width=input_width for stride=1 and pad=1
//int i, f, j;
int fil;
// filter index
#pragma omp parallel for // "omp parallel for" - automatic parallelization of loop by using OpenMP
for (fil = 0; fil < n; ++fil) {
int chan, y, x, f_y, f_x;
// channel index
for (chan = 0; chan < c; ++chan)
// input - y
for (y = 0; y < h; ++y)
// input - x
for (x = 0; x < w; ++x)
{
int const output_index = fil*w*h + y*w + x;
int const weights_pre_index = fil*c*ksize*ksize + chan*ksize*ksize;
int const input_pre_index = chan*w*h;
float sum = 0;
// filter - y
for (f_y = 0; f_y < ksize; ++f_y)
{
int input_y = y + f_y - pad;
// filter - x
for (f_x = 0; f_x < ksize; ++f_x)
{
int input_x = x + f_x - pad;
if (input_y < 0 || input_x < 0 || input_y >= h || input_x >= w) continue;
int input_index = input_pre_index + input_y*w + input_x;
int weights_index = weights_pre_index + f_y*ksize + f_x;
sum += input[input_index] * weights[weights_index];
}
}
// l.output[filters][width][height] +=
// state.input[channels][width][height] *
// l.weights[filters][channels][filter_width][filter_height];
output[output_index] += sum;
}
}
}
static inline int popcnt_64(uint64_t val64) {
#ifdef WIN32 // Windows
#ifdef _WIN64 // Windows 64-bit
int tmp_count = __popcnt64(val64);
#else // Windows 32-bit
int tmp_count = __popcnt(val64);
tmp_count += __popcnt(val64 >> 32);
#endif
#else // Linux
#if defined(__x86_64__) || defined(__aarch64__) // Linux 64-bit
int tmp_count = __builtin_popcountll(val64);
#else // Linux 32-bit
int tmp_count = __builtin_popcount(val64);
tmp_count += __builtin_popcount(val64 >> 32);
#endif
#endif
return tmp_count;
}
void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
unsigned char *A, int lda,
unsigned char *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int i;
#pragma omp parallel for
for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
int j, k;
float mean_val = mean_arr[i];
for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
int count = 0;
for (k = 0; k < K; k += 64) { // l.size*l.size*l.c - one filter size [27 - 9216]
uint64_t a_bit64 = *((uint64_t *)(A + (i*lda + k) / 8));
uint64_t b_bit64 = *((uint64_t *)(B + (j*ldb + k) / 8));
uint64_t c_bit64 = xnor_int64(a_bit64, b_bit64);
int tmp_count = popcnt_64(c_bit64);
if (K - k < 64) tmp_count = tmp_count - (64 - (K - k)); // remove extra bits
count += tmp_count;
//binary_int64_printf(c_bit64);
//printf(", count = %d \n\n", tmp_count);
}
C[i*ldc + j] = (2 * count - K) * mean_val;
}
}
}
void im2col_cpu_custom_transpose(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col, int ldb_align)
{
printf("\n im2col_cpu_custom_transpose() isn't implemented without AVX \n");
}
//From Berkeley Vision's Caffe!
//https://github.com/BVLC/caffe/blob/master/LICENSE
void im2col_cpu_custom(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col)
{
im2col_cpu(data_im, channels, height, width, ksize, stride, pad, data_col);
return;
int c;
const int height_col = (height + 2 * pad - ksize) / stride + 1;
const int width_col = (width + 2 * pad - ksize) / stride + 1;
const int channels_col = channels * ksize * ksize;
// optimized version
if (height_col == height && width_col == width && stride == 1 && pad == 1)
{
#pragma omp parallel for
for (c = 0; c < channels_col; ++c) {
int h, w;
int w_offset = c % ksize;
int h_offset = (c / ksize) % ksize;
int c_im = c / ksize / ksize;
for (h = pad; h < height_col - pad; ++h) {
for (w = pad; w < width_col - pad; ++w) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
}
for (; w < width_col - pad; ++w) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
}
}
{
w = 0;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
{
w = width_col - 1;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
{
h = 0;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
{
h = height_col - 1;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
int col_index = (c * height_col + h) * width_col + w;
data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
im_row, im_col, c_im, pad);
}
}
}
}
else {
//printf("\n Error: is no non-optimized version \n");
im2col_cpu(data_im, channels, height, width, ksize, stride, pad, data_col);
}
}
//From Berkeley Vision's Caffe!
//https://github.com/BVLC/caffe/blob/master/LICENSE
void im2col_cpu_custom_bin(float* data_im,
int channels, int height, int width,
int ksize, int stride, int pad, float* data_col, int bit_align)
{
int c;
const int height_col = (height + 2 * pad - ksize) / stride + 1;
const int width_col = (width + 2 * pad - ksize) / stride + 1;
const int channels_col = channels * ksize * ksize;
// optimized version
if (height_col == height && width_col == width && stride == 1 && pad == 1)
{
int new_ldb = bit_align;
#pragma omp parallel for
for (c = 0; c < channels_col; ++c) {
int h, w;
int w_offset = c % ksize;
int h_offset = (c / ksize) % ksize;
int c_im = c / ksize / ksize;
for (h = pad; h < height_col - pad; ++h) {
for (w = pad; w < width_col - pad - 8; w += 1) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
float val = data_im[im_col + width*(im_row + height*c_im)];
if (val > 0) set_bit((unsigned char*)data_col, col_index);
}
for (; w < width_col - pad; ++w) {
int im_row = h_offset + h - pad;
int im_col = w_offset + w - pad;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
float val = data_im[im_col + width*(im_row + height*c_im)];
if (val > 0) set_bit((unsigned char*)data_col, col_index);
}
}
{
w = 0;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
float val = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
if (val > 0) set_bit((unsigned char*)data_col, col_index);
}
}
{
w = width_col - 1;
for (h = 0; h < height_col; ++h) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
float val = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
if (val > 0) set_bit((unsigned char*)data_col, col_index);
}
}
{
h = 0;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
float val = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
if (val > 0) set_bit((unsigned char*)data_col, col_index);
}
}
{
h = height_col - 1;
for (w = 0; w < width_col; ++w) {
int im_row = h_offset + h;
int im_col = w_offset + w;
//int col_index = (c * height_col + h) * width_col + w;
int col_index = c * new_ldb + h * width_col + w;
//data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
float val = im2col_get_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad);
if (val > 0) set_bit((unsigned char*)data_col, col_index);
}
}
}
}
else {
printf("\n Error: is no non-optimized version \n");
//im2col_cpu(data_im, channels, height, width, ksize, stride, pad, data_col); // must be aligned for transpose after float_to_bin
// float_to_bit(b, t_input, src_size);
// transpose_bin(t_input, *t_bit_input, k, n, bit_align, new_ldb, 8);
}
}
void activate_array_cpu_custom(float *x, const int n, const ACTIVATION a)
{
int i;
if (a == LINEAR)
{
}
else if (a == LEAKY)
{
for (i = 0; i < n; ++i) {
x[i] = (x[i]>0) ? x[i] : .1*x[i];
}
}
else {
for (i = 0; i < n; ++i) {
x[i] = activate(x[i], a);
}
}
}
void float_to_bit(float *src, unsigned char *dst, size_t size)
{
size_t dst_size = size / 8 + 1;
memset(dst, 0, dst_size);
size_t i;
char* byte_arr = (char*)xcalloc(size, sizeof(char));
for (i = 0; i < size; ++i) {
if (src[i] > 0) byte_arr[i] = 1;
}
//for (i = 0; i < size; ++i) {
// dst[i / 8] |= byte_arr[i] << (i % 8);
//}
for (i = 0; i < size; i += 8) {
char dst_tmp = 0;
dst_tmp |= byte_arr[i + 0] << 0;
dst_tmp |= byte_arr[i + 1] << 1;
dst_tmp |= byte_arr[i + 2] << 2;
dst_tmp |= byte_arr[i + 3] << 3;
dst_tmp |= byte_arr[i + 4] << 4;
dst_tmp |= byte_arr[i + 5] << 5;
dst_tmp |= byte_arr[i + 6] << 6;
dst_tmp |= byte_arr[i + 7] << 7;
dst[i / 8] = dst_tmp;
}
free(byte_arr);
}
static inline void transpose_scalar_block(float *A, float *B, const int lda, const int ldb, const int block_size)
{
int i;
//#pragma omp parallel for
for (i = 0; i<block_size; i++) {
int j;
for (j = 0; j<block_size; j++) {
B[j*ldb + i] = A[i*lda + j];
}
}
}
void transpose_block_SSE4x4(float *A, float *B, const int n, const int m,
const int lda, const int ldb, const int block_size)
{
int i;
#pragma omp parallel for
for (i = 0; i < n; i += block_size) {
int j, i2, j2;
for (j = 0; j < m; j += block_size) {
int max_i2 = i + block_size < n ? i + block_size : n;
int max_j2 = j + block_size < m ? j + block_size : m;
for (i2 = i; i2 < max_i2; ++i2) {
for (j2 = j; j2 < max_j2; ++j2) {
B[j2*ldb + i2] = A[i2*lda + j2];
}
}
}
}
}
void forward_maxpool_layer_avx(float *src, float *dst, int *indexes, int size, int w, int h, int out_w, int out_h, int c,
int pad, int stride, int batch)
{
int b, k;
const int w_offset = -pad / 2;
const int h_offset = -pad / 2;
for (b = 0; b < batch; ++b) {
#pragma omp parallel for
for (k = 0; k < c; ++k) {
int i, j, m, n;
for (i = 0; i < out_h; ++i) {
for (j = 0; j < out_w; ++j) {
int out_index = j + out_w*(i + out_h*(k + c*b));
float max = -FLT_MAX;
int max_i = -1;
for (n = 0; n < size; ++n) {
for (m = 0; m < size; ++m) {
int cur_h = h_offset + i*stride + n;
int cur_w = w_offset + j*stride + m;
int index = cur_w + w*(cur_h + h*(k + b*c));
int valid = (cur_h >= 0 && cur_h < h &&
cur_w >= 0 && cur_w < w);
float val = (valid != 0) ? src[index] : -FLT_MAX;
max_i = (val > max) ? index : max_i;
max = (val > max) ? val : max;
}
}
dst[out_index] = max;
if (indexes) indexes[out_index] = max_i;
}
}
}
}
}
#endif // AVX
// 32 channels -> 1 channel (with 32 floats)
// 256 channels -> 8 channels (with 32 floats)
void repack_input(float *input, float *re_packed_input, int w, int h, int c)
{
const int items_per_channel = w * h;
int chan, i;
for (chan = 0; chan < c; chan += 32)
{
for (i = 0; i < items_per_channel; ++i)
{
int c_pack;
for (c_pack = 0; c_pack < 32; ++c_pack) {
float src = input[(chan + c_pack)*items_per_channel + i];
re_packed_input[chan*items_per_channel + i * 32 + c_pack] = src;
}
}
}
}
void transpose_uint32(uint32_t *src, uint32_t *dst, int src_h, int src_w, int src_align, int dst_align)
{
//l.bit_align - algined (n) by 32
//new_ldb - aligned (k) by 256
int i;
//#pragma omp parallel for
for (i = 0; i < src_h; i += 1) // l.size*l.size*l.c;
{
int j;
for (j = 0; j < src_w; j += 1) // out_h*out_w;
{
((uint32_t *)dst)[j*dst_align / 32 + i] = ((uint32_t *)src)[i*src_align + j];
}
}
}
void gemm_nn_bin_transposed_32bit_packed(int M, int N, int K, float ALPHA,
uint32_t *A, int lda,
uint32_t *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int i;
#pragma omp parallel for
for (i = 0; i < M; ++i) { // l.n
int j, s;
float mean_val = mean_arr[i];
for (j = 0; j < N; ++j) // out_h*out_w;
{
float val = 0;
for (s = 0; s < K; ++s) // l.size*l.size*l.c/32 or (l.size*l.size*l.c)
{
PUT_IN_REGISTER uint32_t A_PART = ((uint32_t*)A)[i*lda + s];
PUT_IN_REGISTER uint32_t B_PART = ((uint32_t*)B)[j * ldb + s];
uint32_t xnor_result = ~(A_PART ^ B_PART);
int32_t count = popcnt_32(xnor_result); // must be Signed int
val += (2 * count - 32) * mean_val;
}
C[i*ldc + j] += val;
}
}
}
void convolution_repacked(uint32_t *packed_input, uint32_t *packed_weights, float *output,
int w, int h, int c, int n, int size, int pad, int new_lda, float *mean_arr)
{
int fil;
// filter index
#pragma omp parallel for
for (fil = 0; fil < n; ++fil) {
float mean_val = mean_arr[fil];
int chan, y, x, f_y, f_x; // c_pack
// channel index
for (chan = 0; chan < c / 32; ++chan)
//for (chan = 0; chan < l.c; chan += 32)
//for (c_pack = 0; c_pack < 32; ++c_pack)
// input - y
for (y = 0; y < h; ++y)
// input - x
for (x = 0; x < w; ++x)
{
int const output_index = fil*w*h + y*w + x;
float sum = 0;
// filter - y
for (f_y = 0; f_y < size; ++f_y)
{
int input_y = y + f_y - pad;
// filter - x
for (f_x = 0; f_x < size; ++f_x)
{
int input_x = x + f_x - pad;
if (input_y < 0 || input_x < 0 || input_y >= h || input_x >= w) continue;
// normal
//float input = state.input[(chan + c_pack)*l.w*l.h + input_y*l.w + input_x];
//float weight = l.weights[fil*l.c*l.size*l.size + (chan + c_pack)*l.size*l.size + f_y*l.size + f_x];
// packed
//float input = re_packed_input[chan*l.w*l.h + (input_y*l.w + input_x) * 32 + c_pack];
//float weight = l.weights[fil*l.c*l.size*l.size + chan*l.size*l.size + (f_y*l.size + f_x) * 32 + c_pack];
//sum += input * weight;
//float input = re_packed_input[chan*l.w*l.h + (input_y*l.w + input_x) * 32 + c_pack];
//float weight = l.weights[fil*l.c*l.size*l.size + chan*l.size*l.size + (f_y*l.size + f_x) * 32 + c_pack];
//uint32_t bit1 = input > 0;
//uint32_t bit2 = weight > 0;
//uint32_t count = (~(bit1 ^ bit2)) & 1;
//float result = (2 * (float)count - 1) * mean_val;
//printf("\n mul = %f, bit1 = %d, bit2 = %d, count = %d, mean = %f, result = %f ", input*weight, bit1, bit2, count, mean_val, result);
//sum += result;
uint32_t input = ((uint32_t *)packed_input)[chan*w*h + input_y*w + input_x];
//uint32_t weight = ((uint32_t *)l.align_bit_weights)[fil*l.c*l.size*l.size/32 + chan*l.size*l.size + f_y*l.size + f_x];
uint32_t weight = ((uint32_t *)packed_weights)[fil*new_lda / 32 + chan*size*size + f_y*size + f_x];
uint32_t xnor_result = ~(input ^ weight);
int32_t count = popcnt_32(xnor_result); // mandatory Signed int
sum += (2 * count - 32) * mean_val;
}
}
// l.output[filters][width][height] +=
// state.input[channels][width][height] *
// l.weights[filters][channels][filter_width][filter_height];
output[output_index] += sum;
}
}
}
//计算A*B^T+C
/**
* 被gemm_cpu()函数调用,实际完成 C = ALPHA * A * B^T + C 矩阵计算
* @param M A,C的行数（不做转置）或者A^T的行数（做转置），此处A未转置，故为A的行数
* @param N B,C的列数（不做转置）或者B^T的列数（做转置），此处B转置，故为B^T的列数
* @param K A的列数（不做转置）或者A^T的列数（做转置）,B的行数（不做转置）或者B^T（做转置），此处A未转置，B转置，故为A的列数，B^T的行数
* @param ALPHA 系数
* @param A 输入矩阵
* @param lda A的列数（不做转置）或者A^T的行数（做转置），此处A未转置，故为A的列数
* @param B 输入矩阵
* @param ldb B的列数（不做转置）或者B^T的行数（做转置），此处B转置，故为B^T的行数
* @param C 输入矩阵
* @param ldc 矩阵C的列数
* 说明：此函数在gemm_cpu()函数中调用，是其中四中情况之一，A不进行转置，B转置
* 函数名gemm_nt()中nt分别表示 not transpose, tranpose
*/
void gemm_nt(int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i,j,k;
for(i = 0; i < M; ++i){
for(j = 0; j < N; ++j){
PUT_IN_REGISTER float sum = 0;
for(k = 0; k < K; ++k){
//内循环：每次内循环结束，将计算A中第i行与B中第j列相乘的结果
//也就是得到C[i][j],因为C也一维化，且按行存储，所以得到C[i*lda+j]
// k表示A的第几列，也表示
sum += ALPHA*A[i*lda+k]*B[j*ldb + k];
}
C[i*ldc+j] += sum;
}
}
}
/**
* 被gemm_cpu()函数调用,实际完成 C = ALPHA * A^T * B + C 矩阵计算
* @param M A,C的行数（不做转置）或者A^T的行数（做转置），此处A转置，故为A^T的行数
* @param N B,C的列数（不做转置）或者B^T的列数（做转置），此处B未转置，故为B的列数
* @param K A的列数（不做转置）或者A^T的列数（做转置）,B的行数（不做转置）或者B^T行数（做转置），此处A转置，B未转置，故为A^T的行数，B的列数
* @param ALPHA 系数
* @param A 输入矩阵
* @param lda A的列数（不做转置）或者A^T的行数（做转置），此处A转置，故为A^T的行数
* @param B 输入矩阵
* @param ldb B的列数（不做转置）或者B^T的行数（做转置），此处B未转置，故为B的列数
* @param C 输入矩阵
* @param ldc 矩阵C的列数
* 说明：此函数在gemm_cpu()函数中调用，是其中四中情况之一，A进行转置，B不进行转置
* 函数名gemm_tn()中tn分别表示 tranpose,not transpose
*/
void gemm_tn(int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i,j,k;
for(i = 0; i < M; ++i){
for(k = 0; k < K; ++k){
PUT_IN_REGISTER float A_PART = ALPHA * A[k * lda + i];
for(j = 0; j < N; ++j){
//内循环：每次内循环结束，将计算A中第i行与B中第j列相乘的结果
//也就是得到C[i][j],因为C也一维化，且按行存储，所以得到C[i*lda+j]
// k表示A的第几列，也表示
C[i*ldc+j] += A_PART*B[k*ldb+j];
}
}
}
}
/**
* 被gemm_cpu()函数调用,实际完成 C = ALPHA * A^T * B^T + C 矩阵计算
* @param M A,C的行数（不做转置）或者A^T的行数（做转置），此处A转置，故为A^T的行数
* @param N B,C的列数（不做转置）或者B^T的列数（做转置），此处B转置，故为B^T的列数
* @param K A的列数（不做转置）或者A^T的列数（做转置）,B的行数（不做转置）或者B^T（做转置），此处A转置，B转置，故为A^T的列数，B^T的行数
* @param ALPHA 系数
* @param A 输入矩阵
* @param lda A的列数（不做转置）或者A^T的行数（做转置），此处A转置，故为A^T的行数
* @param B 输入矩阵
* @param ldb B的列数（不做转置）或者B^T的行数（做转置），此处B转置，故为B^T的行数
* @param C 输入矩阵
* @param ldc 矩阵C的列数
* 说明：此函数在gemm_cpu()函数中调用，是其中四中情况之一，A进行转置，B进行转置
* 函数名gemm_tt()中tt分别表示 transpose, tranpose
*/
void gemm_tt(int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i,j,k;
for(i = 0; i < M; ++i){
for(j = 0; j < N; ++j){
PUT_IN_REGISTER float sum = 0;
for(k = 0; k < K; ++k){
sum += ALPHA*A[i+k*lda]*B[k+j*ldb];
}
C[i*ldc+j] += sum;
}
}
}
void gemm_cpu(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc)
{
//printf("cpu: %d %d %d %d %d %f %d %d %f %d\n",TA, TB, M, N, K, ALPHA, lda, ldb, BETA, ldc);
if (BETA != 1){
int i, j;
for(i = 0; i < M; ++i){
for(j = 0; j < N; ++j){
// 先行计算BETA * C,并把结果存入C中，得到C将为M行N列（按行存储在一维数组中）
C[i*ldc + j] *= BETA;
}
}
}
is_avx(); // initialize static variable
if (is_fma_avx2() && !TA && !TB) {
gemm_nn_fast(M, N, K, ALPHA, A, lda, B, ldb, C, ldc);
}
else {
int t;
#pragma omp parallel for
//根据AB矩阵的情况计算ALPHA*A*B+C
for (t = 0; t < M; ++t) {
if (!TA && !TB)
gemm_nn(1, N, K, ALPHA, A + t*lda, lda, B, ldb, C + t*ldc, ldc);
else if (TA && !TB)
gemm_tn(1, N, K, ALPHA, A + t, lda, B, ldb, C + t*ldc, ldc);
else if (!TA && TB)
gemm_nt(1, N, K, ALPHA, A + t*lda, lda, B, ldb, C + t*ldc, ldc);
else
gemm_tt(1, N, K, ALPHA, A + t, lda, B, ldb, C + t*ldc, ldc);
}
}
}
#ifdef GPU
#include <math.h>
void gemm_ongpu(int TA, int TB, int M, int N, int K, float ALPHA,
float *A_gpu, int lda,
float *B_gpu, int ldb,
float BETA,
float *C_gpu, int ldc)
{
cublasHandle_t handle = blas_handle();
cudaError_t stream_status = (cudaError_t)cublasSetStream(handle, get_cuda_stream());
CHECK_CUDA(stream_status);
cudaError_t status = (cudaError_t)cublasSgemm(handle, (TB ? CUBLAS_OP_T : CUBLAS_OP_N),
(TA ? CUBLAS_OP_T : CUBLAS_OP_N), N, M, K, &ALPHA, B_gpu, ldb, A_gpu, lda, &BETA, C_gpu, ldc);
CHECK_CUDA(status);
}
//矩阵计算GPU实现，调用CUDA中cublasSgemm()函数完成 C_gpu = ALPHA + A_gpu * B_gpu + BETA * C_gpu的线性矩阵运算
// 与gemm_cpu()基本类似，输入参数也基本相同，但是存在两点不同：
// 1. 此处是直接调用CUDA cuBLAS库中的cublasSgemm()函数进行矩阵运算，而无需gemm_cpu()那样，需要自己用循环挨个元素相乘实现；
// 2. 在GPU中,默认采用的矩阵存储格式是按列存储,而不是我们之前一度习惯的按行存储,此处调用的cublasSgemm()也不例外,
// 所以下面会有一些不同的操作（由于这个原因，相比于cpu版本的gemm_cpu(),又要复杂一些）
//
/*
* GPU使用cuBLAS库中cublasSgemm()函数进行矩阵乘法计算，参看：
* 这个网址是CUDA关于cuBLAS库的官方文档，此处cublasSgemm()函数在2.7.1节: cublas<t>gemm();
* 可以看出cublasSgem()函数完成C_gpu = ALPHA * A_gpu * B_gpu + BETA * C_gpu的线性矩阵计算
*
* @param TA 是否需要对A做转置操作，是为1，否为0（要不要转置取决于A，B之间的维度是否匹配，比如A:3*2, B:4*2, 则需要对B转置，才满足矩阵乘法）
* @param TB 同上
* @param M A,C 的行数（若A需要转置，则此出给出转置后A即A^T的行数，而不是转置前的）
* @param N B,C 的列数（若B需要转置，则此处给出转置后B即B^T的列数，而不是转置前的）
* @param K A的列数,B的行数（同样，若A与B中的二者或者其中一个需要转置，则不管怎么样，转置后的A,B必须行列能够匹配，符合矩阵乘法规则，K也是转置后的值，不是转置的）
* @param ALPHA 系数
* @param A_gpu 输入矩阵,且其内存在GPU设备内存中，不在主机内存中（由cudaMalloc分配，由cudaFree释放）
* @param lda A的列数(不做转置)或者行数(做转置，且给的是转置后A即A^T的行数)
* @param B_gpu 输入矩阵,且其内存在GPU设备内存中，不在主机内存中（由cudaMalloc分配，由cudaFree释放）
* @param ldb B的列数(不做转置)或者行数(做转置，且给的是转置后B即B^T的行数)
* @param BETA 系数
* @param C_gpu 输入矩阵,且其内存在GPU设备内存中，不在主机内存中（由cudaMalloc分配，由cudaFree释放）
* @param ldc C的列数
*
* 可以看出，如果不是因为存储方式的不同，cublasSgemm()函数的结构也与darknet自己实现的cpu版本的gemm_cpu一模一样；
* 因为二者存储格式不同，需要交换A_gpu, B_gpu的位置，对应M和N之间，TB与TA之间，ldb与lda之间都要相互交换；
*/
void gemm_gpu(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc)
{
float *A_gpu = cuda_make_array(A, (TA ? lda*K:lda*M));
float *B_gpu = cuda_make_array(B, (TB ? ldb*N : ldb*K));
float *C_gpu = cuda_make_array(C, ldc*M);
gemm_ongpu(TA, TB, M, N, K, ALPHA, A_gpu, lda, B_gpu, ldb, BETA, C_gpu, ldc);
cuda_pull_array(C_gpu, C, ldc*M);
cuda_free(A_gpu);
cuda_free(B_gpu);
cuda_free(C_gpu);
}
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
void time_gpu_random_matrix(int TA, int TB, int m, int k, int n)
{
float *a;
if(!TA) a = random_matrix(m,k);
else a = random_matrix(k,m);
int lda = (!TA)?k:m;
float *b;
if(!TB) b = random_matrix(k,n);
else b = random_matrix(n,k);
int ldb = (!TB)?n:k;
float *c = random_matrix(m,n);
int i;
clock_t start = clock(), end;
for(i = 0; i<32; ++i){
gemm_gpu(TA,TB,m,n,k,1,a,lda,b,ldb,1,c,n);
}
end = clock();
printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %lf s\n",m,k,k,n, TA, TB, (float)(end-start)/CLOCKS_PER_SEC);
free(a);
free(b);
free(c);
}
void time_ongpu(int TA, int TB, int m, int k, int n)
{
int iter = 10;
float *a = random_matrix(m,k);
float *b = random_matrix(k,n);
int lda = (!TA)?k:m;
int ldb = (!TB)?n:k;
float *c = random_matrix(m,n);
float *a_cl = cuda_make_array(a, m*k);
float *b_cl = cuda_make_array(b, k*n);
float *c_cl = cuda_make_array(c, m*n);
int i;
clock_t start = clock(), end;
for(i = 0; i<iter; ++i){
gemm_ongpu(TA,TB,m,n,k,1,a_cl,lda,b_cl,ldb,1,c_cl,n);
cudaDeviceSynchronize();
}
double flop = ((double)m)*n*(2.*k + 2.)*iter;
double gflop = flop/pow(10., 9);
end = clock();
double seconds = sec(end-start);
printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %lf s, %lf GFLOPS\n",m,k,k,n, TA, TB, seconds, gflop/seconds);
cuda_free(a_cl);
cuda_free(b_cl);
cuda_free(c_cl);
free(a);
free(b);
free(c);
}
void test_gpu_accuracy(int TA, int TB, int m, int k, int n)
{
srand(0);
float *a;
if(!TA) a = random_matrix(m,k);
else a = random_matrix(k,m);
int lda = (!TA)?k:m;
float *b;
if(!TB) b = random_matrix(k,n);
else b = random_matrix(n,k);
int ldb = (!TB)?n:k;
float *c = random_matrix(m,n);
float *c_gpu = random_matrix(m,n);
memset(c, 0, m*n*sizeof(float));
memset(c_gpu, 0, m*n*sizeof(float));
int i;
//pm(m,k,b);
gemm_gpu(TA,TB,m,n,k,1,a,lda,b,ldb,1,c_gpu,n);
//printf("GPU\n");
//pm(m, n, c_gpu);
gemm_cpu(TA,TB,m,n,k,1,a,lda,b,ldb,1,c,n);
//printf("\n\nCPU\n");
//pm(m, n, c);
double sse = 0;
for(i = 0; i < m*n; ++i) {
//printf("%f %f\n", c[i], c_gpu[i]);
sse += pow(c[i]-c_gpu[i], 2);
}
printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %g SSE\n",m,k,k,n, TA, TB, sse/(m*n));
free(a);
free(b);
free(c);
free(c_gpu);
}
int test_gpu_blas()
{
/*
test_gpu_accuracy(0,0,10,576,75);
test_gpu_accuracy(0,0,17,10,10);
test_gpu_accuracy(1,0,17,10,10);
test_gpu_accuracy(0,1,17,10,10);
test_gpu_accuracy(1,1,17,10,10);
test_gpu_accuracy(0,0,1000,10,100);
test_gpu_accuracy(1,0,1000,10,100);
test_gpu_accuracy(0,1,1000,10,100);
test_gpu_accuracy(1,1,1000,10,100);
test_gpu_accuracy(0,0,10,10,10);
time_ongpu(0,0,64,2916,363);
time_ongpu(0,0,64,2916,363);
time_ongpu(0,0,64,2916,363);
time_ongpu(0,0,192,729,1600);
time_ongpu(0,0,384,196,1728);
time_ongpu(0,0,256,196,3456);
time_ongpu(0,0,256,196,2304);
time_ongpu(0,0,128,4096,12544);
time_ongpu(0,0,128,4096,4096);
*/
time_ongpu(0,0,64,75,12544);
time_ongpu(0,0,64,75,12544);
time_ongpu(0,0,64,75,12544);
time_ongpu(0,0,64,576,12544);
time_ongpu(0,0,256,2304,784);
time_ongpu(1,1,2304,256,784);
time_ongpu(0,0,512,4608,196);
time_ongpu(1,1,4608,512,196);
return 0;
}
#endif
void init_cpu() {
is_avx();
is_fma_avx2();
}
