ref: 54abdb6f5d671a1fc9756e475012c365aae71f59
parent: 2681822c1825d06a626284f7d6e3228e6109b9ec
author: Jean-Marc Valin <jmvalin@amazon.com>
date: Tue Jul 6 13:05:07 EDT 2021
Sparse matrix indexing optimization The 4* is now stored in the table to avoid computing it in the loop
--- a/dnn/training_tf2/dump_lpcnet.py
+++ b/dnn/training_tf2/dump_lpcnet.py
@@ -80,7 +80,7 @@
qblock = AQ[j*4:(j+1)*4, i*8:(i+1)*8]
if np.sum(np.abs(block)) > 1e-10:
nb_nonzero = nb_nonzero + 1
- idx = np.append(idx, j)
+ idx = np.append(idx, j*4)
vblock = qblock.transpose((1,0)).reshape((-1,))
W0 = np.concatenate([W0, block.reshape((-1,))])
W = np.concatenate([W, vblock])
--- a/dnn/vec.h
+++ b/dnn/vec.h
@@ -250,7 +250,7 @@
int pos;
float * restrict y;
int xj0, xj1, xj2, xj3;
- pos = 4 * (*idx++);
+ pos = (*idx++);
xj0 = x[pos+0];
xj1 = x[pos+1];
xj2 = x[pos+2];
@@ -318,7 +318,7 @@
int pos;
float * restrict y;
int xj0, xj1, xj2, xj3;
- pos = 4 * (*idx++);
+ pos = (*idx++);
xj0 = x[pos+0];
xj1 = x[pos+1];
xj2 = x[pos+2];
@@ -357,7 +357,7 @@
int pos;
float * restrict y;
float xj0, xj1, xj2, xj3;
- pos = 4 * (*idx++);
+ pos = (*idx++);
xj0 = x[pos+0];
xj1 = x[pos+1];
xj2 = x[pos+2];
--- a/dnn/vec_avx.h
+++ b/dnn/vec_avx.h
@@ -508,7 +508,7 @@
__m256i vxj;
__m256i vw;
int pos;
- pos = 4 * (*idx++);
+ pos = (*idx++);
vxj = _mm256_set1_epi32(*(int*)&x[pos]);
vw = _mm256_loadu_si256((const __m256i *)w); //_mm256_lddqu_si256?
tmp = _mm256_maddubs_epi16(vxj, vw); //swap?
@@ -544,19 +544,19 @@
__m256 vxj;
__m256 vw;
id = *idx++;
- vxj = _mm256_broadcast_ss(&x[4*id]);
+ vxj = _mm256_broadcast_ss(&x[id]);
vw = _mm256_loadu_ps(&weights[0]);
vy0 = _mm256_fmadd_ps(vw, vxj, vy0);
- vxj = _mm256_broadcast_ss(&x[4*id+1]);
+ vxj = _mm256_broadcast_ss(&x[id+1]);
vw = _mm256_loadu_ps(&weights[8]);
vy0 = _mm256_fmadd_ps(vw, vxj, vy0);
- vxj = _mm256_broadcast_ss(&x[4*id+2]);
+ vxj = _mm256_broadcast_ss(&x[id+2]);
vw = _mm256_loadu_ps(&weights[16]);
vy0 = _mm256_fmadd_ps(vw, vxj, vy0);
- vxj = _mm256_broadcast_ss(&x[4*id+3]);
+ vxj = _mm256_broadcast_ss(&x[id+3]);
vw = _mm256_loadu_ps(&weights[24]);
vy0 = _mm256_fmadd_ps(vw, vxj, vy0);
--- a/dnn/vec_neon.h
+++ b/dnn/vec_neon.h
@@ -333,7 +333,7 @@
for (j=0;j<colblocks;j++)
{int pos;
- pos = 4 * (*idx++);
+ pos = (*idx++);
int8x16_t vw0, vw1, vx;
vx = (int8x16_t)vld1q_dup_s32((int*)&x[pos]);
vw0 = vld1q_s8(w);
--
⑨