1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16
17
xnn_f32_qs8_vcvt_ukernel__sse2_x8(size_t n,const float * x,int8_t * y,const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_qs8_vcvt_ukernel__sse2_x8(
19 size_t n,
20 const float* x,
21 int8_t* y,
22 const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m128 vscale = _mm_load_ps(params->sse2.scale);
30 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
33
34 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
35 __m128 vx_lo = _mm_loadu_ps(x);
36 __m128 vx_hi = _mm_loadu_ps(x + 4);
37 x += 8;
38
39 vx_lo = _mm_mul_ps(vx_lo, vscale);
40 vx_hi = _mm_mul_ps(vx_hi, vscale);
41
42 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
43 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
44
45 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
46 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
47
48 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
49 vy = _mm_adds_epi16(vy, voutput_zero_point);
50 vy = _mm_max_epi16(vy, voutput_min);
51 vy = _mm_packs_epi16(vy, vy);
52
53 _mm_storel_epi64((__m128i*) y, vy);
54 y += 8;
55 }
56 if XNN_UNLIKELY(n != 0) {
57 __m128 vx_lo = _mm_loadu_ps(x);
58 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
59 __m128 vx_hi = _mm_loadu_ps(x_hi);
60
61 vx_lo = _mm_mul_ps(vx_lo, vscale);
62 vx_hi = _mm_mul_ps(vx_hi, vscale);
63
64 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
65 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
66
67 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
68 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
69
70 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
71 vy = _mm_adds_epi16(vy, voutput_zero_point);
72 vy = _mm_max_epi16(vy, voutput_min);
73 vy = _mm_packs_epi16(vy, vy);
74
75 if (n & (4 * sizeof(float))) {
76 *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vy);
77 y += 4;
78 vy = _mm_srli_epi64(vy, 32);
79 }
80 {
81 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
82 if (n & (2 * sizeof(float))) {
83 *((uint16_t*) y) = (uint16_t) vy_lo;
84 y += 2;
85 vy_lo >>= 16;
86 }
87 if (n & (1 * sizeof(float))) {
88 *y = (int8_t) vy_lo;
89 }
90 }
91 }
92 }
93