ref: a1079c2ce33e198d582fabcfab79085726ff83cd
parent: 7d8b00f11d2ae57128ce8f235f341da36bc6f7bf
author: Jean-Marc Valin <jmvalin@amazon.com>
date: Thu Jul 8 09:20:15 EDT 2021
Again, same conversion as 3206cec, for NEON
--- a/dnn/vec_neon.h
+++ b/dnn/vec_neon.h
@@ -291,15 +291,15 @@
{int i, j;
signed char x[MAX_INPUTS];
- int out[MAX_OUTPUTS];
+ const float32x4_t scale = vdupq_n_f32(SCALE);
+ const float32x4_t scale_1 = vdupq_n_f32(SCALE_1);
(void)col_stride;
- for (i=0;i<rows;i++) out[i] = (int)floor(.5+SCALE*_out[i]);
for (i=0;i<cols;i++) x[i] = (int)floor(.5+127*_x[i]);
for (i=0;i<rows;i+=8)
{int32x4_t acc0, acc1;
- acc0 = vld1q_s32(&out[i]);
- acc1 = vld1q_s32(&out[i+4]);
+ acc0 = vcvtnq_s32_f32(vmulq_f32(scale, vld1q_f32(&_out[i])));
+ acc1 = vcvtnq_s32_f32(vmulq_f32(scale, vld1q_f32(&_out[i+4])));
for (j=0;j<cols;j+=4)
{int8x16_t vw0, vw1, vx;
@@ -310,10 +310,9 @@
acc1 = vdotprod(acc1, vw1, vx);
w += 32;
}
- vst1q_s32(&out[i], acc0);
- vst1q_s32(&out[i+4], acc1);
+ vst1q_f32(&_out[i], vmulq_f32(scale_1, vcvtq_f32_s32(acc0)));
+ vst1q_f32(&_out[i+4], vmulq_f32(scale_1, vcvtq_f32_s32(acc1)));
}
- for (i=0;i<rows;i++) _out[i] = SCALE_1*out[i];
}
static inline void sparse_sgemv_accum8x4(float *_out, const qweight *w, int rows, int cols, const int *idx, const float *_x)
@@ -320,15 +319,15 @@
{int i, j;
signed char x[MAX_INPUTS];
- int out[MAX_OUTPUTS];
- for (i=0;i<rows;i++) out[i] = (int)floor(.5+SCALE*_out[i]);
+ const float32x4_t scale = vdupq_n_f32(SCALE);
+ const float32x4_t scale_1 = vdupq_n_f32(SCALE_1);
for (i=0;i<cols;i++) x[i] = floor(.5+127*_x[i]);
for (i=0;i<rows;i+=8)
{int colblocks;
int32x4_t acc0, acc1;
- acc0 = vld1q_s32(&out[i]);
- acc1 = vld1q_s32(&out[i+4]);
+ acc0 = vcvtnq_s32_f32(vmulq_f32(scale, vld1q_f32(&_out[i])));
+ acc1 = vcvtnq_s32_f32(vmulq_f32(scale, vld1q_f32(&_out[i+4])));
colblocks = *idx++;
for (j=0;j<colblocks;j++)
{@@ -342,8 +341,7 @@
acc1 = vdotprod(acc1, vw1, vx);
w += 32;
}
- vst1q_s32(&out[i], acc0);
- vst1q_s32(&out[i+4], acc1);
+ vst1q_f32(&_out[i], vmulq_f32(scale_1, vcvtq_f32_s32(acc0)));
+ vst1q_f32(&_out[i+4], vmulq_f32(scale_1, vcvtq_f32_s32(acc1)));
}
- for (i=0;i<rows;i++) _out[i] = SCALE_1*out[i];
}
--
⑨