From e0976bcb30fa50e6e33c701fc057a4e93935bccf Mon Sep 17 00:00:00 2001
From: Edmond Yoo <hj3yoo@uwaterloo.ca>
Date: Sat, 13 Oct 2018 06:17:09 +0000
Subject: [PATCH] Update README
---
src/gemm.c | 694 ++++++++++++++++++++++++++++++++++++++++++++++++++++++---
1 files changed, 649 insertions(+), 45 deletions(-)
diff --git a/src/gemm.c b/src/gemm.c
index 2fd7f50..6ff57cd 100644
--- a/src/gemm.c
+++ b/src/gemm.c
@@ -5,6 +5,8 @@
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
+#include <float.h>
+#include <string.h>
#if defined(_OPENMP)
#include <omp.h>
@@ -304,13 +306,84 @@
//----------------------------
-#if (defined(__AVX__) && defined(__x86_64__)) || defined(_WIN64)
+void transpose_8x8_bits_my(unsigned char *A, unsigned char *B, int lda, int ldb)
+{
+ unsigned x, y, t;
+ for (y = 0; y < 8; ++y) {
+ for (x = 0; x < 8; ++x) {
+ if (A[y * lda] & (1 << x)) B[x * ldb] |= 1 << y;
+ }
+ }
+}
-#define OSXSAVEFlag (1UL<<27)
-#define AVXFlag ((1UL<<28)|OSXSAVEFlag)
-#define FMAFlag ((1UL<<12)|AVXFlag|OSXSAVEFlag)
-#define CLMULFlag ((1UL<< 1)|AVXFlag|OSXSAVEFlag)
-#define VAESFlag ((1UL<<25)|AVXFlag|OSXSAVEFlag)
+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_2(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(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, int m, int n, unsigned char* B)
+{
+ unsigned x, y, t;
+
+ // 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);
+}
+
+void transpose_bin(char *A, char *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 += 8) {
+ int j;
+ for (j = 0; j < m - 8; 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], lda / 8, ldb / 8, &B[b_index / 8]);
+ }
+ for (; j < m; ++j) {
+ if (get_bit(A, i*lda + j)) set_bit(B, j*ldb + i);
+ }
+ }
+}
+
+//----------------------------
+
+
+#if (defined(__AVX__) && defined(__x86_64__)) || defined(_WIN64)
#ifdef _WIN64
#include <intrin.h>
@@ -326,7 +399,6 @@
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) {
@@ -368,35 +440,121 @@
abcd[2] = ecx;
abcd[3] = edx;
}
-
#endif
-int simd_detect_x86(unsigned int idFeature)
-{
- uint32_t regs[4]; // EAX, EBX, ECX, EDX;
+
+
#ifdef _WIN32
- __cpuid(regs, 0);
- if (regs[0] > 1U) __cpuid(regs, 1);
+// Windows
+#define cpuid(info, x) __cpuidex(info, x, 0)
#else
- __get_cpuid(0, ®s[0], ®s[1], ®s[2], ®s[3]);
- if(regs[0] > 1U) __get_cpuid(1, ®s[0], ®s[1], ®s[2], ®s[3]);
+// GCC Intrinsics
+void cpuid(int info[4], int InfoType) {
+ __cpuid_count(InfoType, 0, info[0], info[1], info[2], info[3]);
+}
#endif
- if ((regs[2] & idFeature) != idFeature)
- return 0;
- return 1;
+
+// 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);
+ int nIds = info[0];
+
+ cpuid(info, 0x80000000);
+ unsigned nExIds = info[0];
+
+ // Detect Features
+ if (nIds >= 0x00000001) {
+ cpuid(info, 0x00000001);
+ HW_MMX = (info[3] & ((int)1 << 23)) != 0;
+ HW_SSE = (info[3] & ((int)1 << 25)) != 0;
+ HW_SSE2 = (info[3] & ((int)1 << 26)) != 0;
+ HW_SSE3 = (info[2] & ((int)1 << 0)) != 0;
+
+ HW_SSSE3 = (info[2] & ((int)1 << 9)) != 0;
+ HW_SSE41 = (info[2] & ((int)1 << 19)) != 0;
+ HW_SSE42 = (info[2] & ((int)1 << 20)) != 0;
+ HW_AES = (info[2] & ((int)1 << 25)) != 0;
+
+ HW_AVX = (info[2] & ((int)1 << 28)) != 0;
+ HW_FMA3 = (info[2] & ((int)1 << 12)) != 0;
+
+ HW_RDRAND = (info[2] & ((int)1 << 30)) != 0;
+ }
+ if (nIds >= 0x00000007) {
+ cpuid(info, 0x00000007);
+ HW_AVX2 = (info[1] & ((int)1 << 5)) != 0;
+
+ HW_BMI1 = (info[1] & ((int)1 << 3)) != 0;
+ HW_BMI2 = (info[1] & ((int)1 << 8)) != 0;
+ HW_ADX = (info[1] & ((int)1 << 19)) != 0;
+ HW_SHA = (info[1] & ((int)1 << 29)) != 0;
+ HW_PREFETCHWT1 = (info[2] & ((int)1 << 0)) != 0;
+
+ HW_AVX512F = (info[1] & ((int)1 << 16)) != 0;
+ HW_AVX512CD = (info[1] & ((int)1 << 28)) != 0;
+ HW_AVX512PF = (info[1] & ((int)1 << 26)) != 0;
+ HW_AVX512ER = (info[1] & ((int)1 << 27)) != 0;
+ HW_AVX512VL = (info[1] & ((int)1 << 31)) != 0;
+ HW_AVX512BW = (info[1] & ((int)1 << 30)) != 0;
+ HW_AVX512DQ = (info[1] & ((int)1 << 17)) != 0;
+ HW_AVX512IFMA = (info[1] & ((int)1 << 21)) != 0;
+ HW_AVX512VBMI = (info[2] & ((int)1 << 1)) != 0;
+ }
+ if (nExIds >= 0x80000001) {
+ cpuid(info, 0x80000001);
+ HW_x64 = (info[3] & ((int)1 << 29)) != 0;
+ HW_ABM = (info[2] & ((int)1 << 5)) != 0;
+ HW_SSE4a = (info[2] & ((int)1 << 6)) != 0;
+ HW_FMA4 = (info[2] & ((int)1 << 16)) != 0;
+ HW_XOP = (info[2] & ((int)1 << 11)) != 0;
+ }
}
-int is_fma_avx() {
+int is_avx() {
static int result = -1;
if (result == -1) {
- result = simd_detect_x86(AVXFlag);
+ 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
void gemm_nn(int M, int N, int K, float ALPHA,
float *A, int lda,
@@ -404,7 +562,7 @@
float *C, int ldc)
{
int i, j, k;
- if (is_fma_avx() == 1) { // AVX
+ if (is_avx() == 1) { // AVX
for (i = 0; i < M; ++i) {
for (k = 0; k < K; ++k) {
float A_PART = ALPHA*A[i*lda + k];
@@ -515,7 +673,7 @@
static int max_num_threads = 0;
if (max_num_threads == 0) {
max_num_threads = omp_get_max_threads();
- omp_set_num_threads(4);// max_num_threads / 2);
+ //omp_set_num_threads( max_num_threads / 2);
}
#endif
@@ -878,7 +1036,7 @@
int channels_col = channels * ksize * ksize;
// optimized version
- if (height_col == height && width_col == width && stride == 1 && pad == 1 && is_fma_avx())
+ 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) {
@@ -957,29 +1115,226 @@
}
}
-void transpose_8x8_bits(unsigned char A[8], unsigned char B[8], int m, int n)
+//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)
{
- unsigned x, y, t;
+ int c, h, w;
+ int height_col = (height + 2 * pad - ksize) / stride + 1;
+ int width_col = (width + 2 * pad - ksize) / stride + 1;
+ int channels_col = channels * ksize * ksize;
- // Load the array and pack it into x and y.
+ // optimized version
+ if (height_col == height && width_col == width && stride == 1 && pad == 1 && is_fma_avx2())
+ {
+ int new_ldb = bit_align;
- 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];
+ #pragma omp parallel for
+ for (c = 0; c < channels_col; ++c) {
+ 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;
- t = (x ^ (x >> 7)) & 0x00AA00AA; x = x ^ t ^ (t << 7);
- t = (y ^ (y >> 7)) & 0x00AA00AA; y = y ^ t ^ (t << 7);
+ //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);
+ }
- t = (x ^ (x >> 14)) & 0x0000CCCC; x = x ^ t ^ (t << 14);
- t = (y ^ (y >> 14)) & 0x0000CCCC; y = y ^ t ^ (t << 14);
+ 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)];
+ }
+ }
- t = (x & 0xF0F0F0F0) | ((y >> 4) & 0x0F0F0F0F);
- y = ((x << 4) & 0xF0F0F0F0) | (y & 0x0F0F0F0F);
- x = t;
+ {
+ 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);
+ }
+ }
- B[0] = x >> 24; B[n] = x >> 16; B[2 * n] = x >> 8; B[3 * n] = x;
- B[4 * n] = y >> 24; B[5 * n] = y >> 16; B[6 * n] = y >> 8; B[7 * n] = y;
+ {
+ 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, h, w;
+ int height_col = (height + 2 * pad - ksize) / stride + 1;
+ int width_col = (width + 2 * pad - ksize) / stride + 1;
+ 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 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 = &((unsigned char*)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(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(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(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(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(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;
@@ -987,7 +1342,7 @@
{}
else if (a == LEAKY)
{
- if (is_fma_avx()) {
+ 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);
@@ -1022,14 +1377,18 @@
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
+ //__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
- 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;
}
@@ -1088,6 +1447,100 @@
}
+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 w_offset = -pad / 2;
+ 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;
+ indexes[out_index] = max_i;
+ }
+ }
+ }
+ }
+}
+
#else
void gemm_nn(int M, int N, int K, float ALPHA,
@@ -1204,6 +1657,8 @@
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, h, w;
int height_col = (height + 2 * pad - ksize) / stride + 1;
@@ -1225,7 +1680,7 @@
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;
@@ -1234,7 +1689,7 @@
data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
}
-}
+ }
{
w = 0;
@@ -1288,6 +1743,118 @@
}
}
+
+//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, h, w;
+ int height_col = (height + 2 * pad - ksize) / stride + 1;
+ int width_col = (width + 2 * pad - ksize) / stride + 1;
+ 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 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(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(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(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(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(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(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;
@@ -1366,7 +1933,44 @@
}
}
}
-#endif // __x86_64
+
+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;
+ int w_offset = -pad / 2;
+ 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;
+ indexes[out_index] = max_i;
+ }
+ }
+ }
+ }
+}
+
+#endif // AVX
void gemm_nt(int M, int N, int K, float ALPHA,
float *A, int lda,
--
Gitblit v1.10.0