From f007f7b90ec1a3d83fd638d81b9b209448f659a4 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Wed, 15 Aug 2018 20:47:56 +0000
Subject: [PATCH] Fixes for MSVS2017
---
src/convolutional_layer.c | 82 ++++++++++++++++++++++------------------
1 files changed, 45 insertions(+), 37 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index a820588..e7c8e3f 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -44,7 +44,7 @@
}
mean = mean / size;
for(i = 0; i < size; ++i){
- binary[f*size + i] = (weights[f*size + i] > 0) ? mean : -mean;
+ binary[f*size + i] = (weights[f*size + i] > 0) ? mean: -mean;
}
}
}
@@ -593,11 +593,11 @@
}
}
-void binary_align_weights(convolutional_layer *l, size_t lda_align)
+void binary_align_weights(convolutional_layer *l)
{
int m = l->n;
int k = l->size*l->size*l->c;
- size_t new_lda = k + (lda_align - k%lda_align); // (k / 8 + 1) * 8;
+ size_t new_lda = k + (l->lda_align - k % l->lda_align); // (k / 8 + 1) * 8;
binarize_weights(l->weights, m, k, l->binary_weights);
@@ -621,7 +621,7 @@
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;
@@ -637,32 +637,9 @@
//printf("\n align_weights_size = %d, k = %d, m = %d, lda = %d \n", align_weights_size, k, m, k);
//printf("\n align_bit_weights_size = %d, k = %d, m = %d, new_lda = %d \n", align_bit_weights_size, k, m, new_ldb);
- // transpose and align B
- int i, j;
- //#pragma omp parallel for
- /*
- for (i = 0; i < n; ++i) {
- for (j = 0; j < k; ++j) {
- t_input[i*new_ldb + j] = b[j*n + i];
- }
- }*/
- //transpose_block_SSE4x4(float *A, float *B, const int n, const int m, const int lda, const int ldb, const int block_size)
-
- //transpose_block(b, t_input, k, n, n, new_ldb, 16);
-
- int blocksize = 1;
- int mod_k = 1, mod_n = 1;
- for (i = 2; i < 256; i *= 2)
- if (k % i == 0) mod_k = i;
-
- for (i = 2; i < 256; i *= 2)
- if (n % i == 0) mod_n = i;
-
- blocksize = (mod_k < mod_n) ? mod_k : mod_n;
-
+ int blocksize = 64;
transpose_block_SSE4x4(b, t_input, k, n, n, new_ldb, blocksize);
- //transpose_block(b, t_input, k, n, n, new_ldb, blocksize);
//printf("\n blocksize = %d \n", blocksize);
float_to_bit(t_input, *t_bit_input, t_intput_size);
@@ -703,8 +680,20 @@
for(i = 0; i < l.batch; ++i){
//im2col_cpu(state.input, l.c, l.h, l.w, l.size, l.stride, l.pad, b);
+
+ //float *t_input = NULL;
+ //if (l.xnor) {
+ // size_t new_ldb = k + (l.lda_align - k%l.lda_align);
+ // size_t t_intput_size = new_ldb * n;
+ // t_input = calloc(t_intput_size, sizeof(float));
+ // im2col_cpu_custom_transpose(state.input, l.c, l.h, l.w, l.size, l.stride, l.pad, t_input, new_ldb);
+ //}
+ //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);
+
//gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
//gemm_nn_custom(m, n, k, 1, a, k, b, n, c, n);
if (l.xnor) {
@@ -783,19 +772,38 @@
free(align_weights);
}
*/
- size_t ldb_align = 256; // 256 bit for AVX2
- size_t new_ldb = k + (ldb_align - k%ldb_align);
- char *t_bit_input = NULL;
- size_t t_intput_size = binary_transpose_align_input(k, n, b, &t_bit_input, ldb_align);
- 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);
+ /*
+ if (l.size == 3 && l.stride == 1 && l.pad == 1)
+ {
+ //binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.binary_weights);
+ //printf("\n mean = %f \n", l.mean_arr[0]);
- //gemm_nn_custom_bin_mean_transposed(m, n, k, 1, bit_weights, k, t_bit_input, new_ldb, c, n, mean_arr);
+ convolution_2d(l.w, l.h, l.size, l.n, l.c, l.pad, l.stride,
+ //l.weights, state.input, l.output, l.mean_arr);
+ l.binary_weights, state.input, l.output, l.mean_arr);
+ }
+ else {
+ */
- //free(t_input);
- free(t_bit_input);
+ //size_t ldb_align = 256; // 256 bit for AVX2
+ int ldb_align = l.lda_align;
+ size_t new_ldb = k + (ldb_align - k%ldb_align);
+ char *t_bit_input = NULL;
+ size_t t_intput_size = binary_transpose_align_input(k, n, b, &t_bit_input, ldb_align);
+ //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
- //free(align_bit_weights);
+ // 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);
+
+ //free(t_input);
+ free(t_bit_input);
+ //}
+
}
// for bit_input: (k * n)
--
Gitblit v1.10.0