mirror/darknet
1
0
Fork 0
mirror of https://github.com/pjreddie/darknet.git synced 2023-08-10 21:13:14 +03:00
Code Issues Projects Releases Wiki Activity

Some fixes for AVX support on CPU

Browse Source
This commit is contained in:
AlexeyAB
2018-08-14 01:51:31 +03:00
parent b1dddf02cc
commit f92b20580a
2 changed files with 20 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
src/convolutional_layer.c
View File

@ -621,7 +621,7 @@ void binary_align_weights(convolutional_layer *l)
free(align_weights);
}
// further optimizations: im2col_bin() for XNOR, and then transpose_aling_bin()
size_t binary_transpose_align_input(int k, int n, float *b, char **t_bit_input, size_t ldb_align)
{
size_t new_ldb = k + (ldb_align - k%ldb_align); // (k / 8 + 1) * 8;
@ -690,6 +690,7 @@ void forward_convolutional_layer(convolutional_layer l, network_state state)
//}
//if (l.xnor && l.size == 3 && l.stride == 1 && l.pad == 1) {}
//else
// further optimizations: im2col_bin() for XNOR, and then transpose_aling_bin()
im2col_cpu_custom(state.input, l.c, l.h, l.w, l.size, l.stride, l.pad, b);
@ -793,6 +794,8 @@ void forward_convolutional_layer(convolutional_layer l, network_state state)
//char *t_bit_input = calloc(new_ldb * n, sizeof(char)); // for im2col_cpu_custom_transpose() only
//float_to_bit(t_input, t_bit_input, new_ldb * n); // for im2col_cpu_custom_transpose() only
// 5x times faster than gemm()-float32
// further optimizations: accelerate maxpool-layer with OpenMP/AVX
gemm_nn_custom_bin_mean_transposed(m, n, k, 1, l.align_bit_weights, new_ldb, t_bit_input, new_ldb, c, n, l.mean_arr);
//gemm_nn_custom_bin_mean_transposed(m, n, k, 1, bit_weights, k, t_bit_input, new_ldb, c, n, mean_arr);

10
src/gemm.c
View File

@ -674,6 +674,8 @@ static inline int popcnt256_custom(__m256i n) {
+ _mm256_extract_epi64(val, 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,
@ -873,7 +875,7 @@ void im2col_cpu_custom(float* data_im,
int channels_col = channels * ksize * ksize;
// optimized version
if (height_col == height && width_col == width && stride == 1 && pad == 1)
if (height_col == height && width_col == width && stride == 1 && pad == 1 && is_fma_avx())
{
#pragma omp parallel for
for (c = 0; c < channels_col; ++c) {
@ -954,15 +956,16 @@ void im2col_cpu_custom(float* data_im,
void activate_array_cpu_custom(float *x, const int n, const ACTIVATION a)
{
int i;
int i = 0;
if (a == LINEAR)
{}
else if (a == LEAKY)
{
if (is_fma_avx()) {
__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) {
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]);
@ -973,6 +976,7 @@ void activate_array_cpu_custom(float *x, const int n, const ACTIVATION a)
__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];