From 3e856ec04ebe3cf072939641ba6cd7871bfefc32 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Thu, 09 Aug 2018 22:27:20 +0000
Subject: [PATCH] Optimized: transpose
---
src/convolutional_layer.c | 25 ------------
src/gemm.c | 61 ++++++++++++++++--------------
2 files changed, 33 insertions(+), 53 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index a820588..3c9efdd 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -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);
diff --git a/src/gemm.c b/src/gemm.c
index ff5edfa..4a7dad7 100644
--- a/src/gemm.c
+++ b/src/gemm.c
@@ -686,47 +686,50 @@
static inline void transpose4x4_SSE(float *A, float *B, const int lda, const int ldb)
{
- __m128 row1 = _mm_load_ps(&A[0 * lda]);
- __m128 row2 = _mm_load_ps(&A[1 * lda]);
- __m128 row3 = _mm_load_ps(&A[2 * lda]);
- __m128 row4 = _mm_load_ps(&A[3 * lda]);
+ __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_store_ps(&B[0 * ldb], row1);
- _mm_store_ps(&B[1 * ldb], row2);
- _mm_store_ps(&B[2 * ldb], row3);
- _mm_store_ps(&B[3 * ldb], 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;
- if (block_size % 4 == 0) {
- #pragma omp parallel for
- for (i = 0; i < n; i += block_size) {
- int j, i2, j2;
+ #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_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 += 4) {
- for (j2 = j; j2 < max_j2; j2 += 4) {
- transpose4x4_SSE(&A[i2*lda + j2], &B[j2*ldb + i2], lda, ldb);
+ //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 {
- #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];
- }
+ else {
+ for (i2 = i; i2 < n; ++i2) {
+ for (j2 = 0; j2 < m; ++j2) {
+ B[j2*ldb + i2] = A[i2*lda + j2];
}
}
}
--
Gitblit v1.10.0