1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/avx.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 <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16
17
xnn_f32_qs8_vcvt_ukernel__avx_x32(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__avx_x32(
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)])
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m256 vscale = _mm256_load_ps(params->avx.scale);
30 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->avx.output_max_less_zero_point);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx.output_min);
33
34 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
35 __m256 vx01234567 = _mm256_loadu_ps(x);
36 __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8);
37 __m256 vxGHIJKLMN = _mm256_loadu_ps(x + 16);
38 __m256 vxOPQRSTUV = _mm256_loadu_ps(x + 24);
39 x += 32;
40
41 vx01234567 = _mm256_mul_ps(vx01234567, vscale);
42 vx89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vscale);
43 vxGHIJKLMN = _mm256_mul_ps(vxGHIJKLMN, vscale);
44 vxOPQRSTUV = _mm256_mul_ps(vxOPQRSTUV, vscale);
45
46 vx01234567 = _mm256_min_ps(vx01234567, voutput_max_less_zero_point);
47 vx89ABCDEF = _mm256_min_ps(vx89ABCDEF, voutput_max_less_zero_point);
48 vxGHIJKLMN = _mm256_min_ps(vxGHIJKLMN, voutput_max_less_zero_point);
49 vxOPQRSTUV = _mm256_min_ps(vxOPQRSTUV, voutput_max_less_zero_point);
50
51 const __m256i vacc01234567 = _mm256_cvtps_epi32(vx01234567);
52 const __m256i vacc89ABCDEF = _mm256_cvtps_epi32(vx89ABCDEF);
53 const __m256i vaccGHIJKLMN = _mm256_cvtps_epi32(vxGHIJKLMN);
54 const __m256i vaccOPQRSTUV = _mm256_cvtps_epi32(vxOPQRSTUV);
55
56 __m128i vy01234567 = _mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extractf128_si256(vacc01234567, 1));
57 __m128i vy89ABCDEF = _mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extractf128_si256(vacc89ABCDEF, 1));
58 __m128i vyGHIJKLMN = _mm_packs_epi32(_mm256_castsi256_si128(vaccGHIJKLMN), _mm256_extractf128_si256(vaccGHIJKLMN, 1));
59 __m128i vyOPQRSTUV = _mm_packs_epi32(_mm256_castsi256_si128(vaccOPQRSTUV), _mm256_extractf128_si256(vaccOPQRSTUV, 1));
60
61 vy01234567 = _mm_adds_epi16(vy01234567, voutput_zero_point);
62 vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point);
63 vyGHIJKLMN = _mm_adds_epi16(vyGHIJKLMN, voutput_zero_point);
64 vyOPQRSTUV = _mm_adds_epi16(vyOPQRSTUV, voutput_zero_point);
65
66 __m128i vy0123456789ABCDEF = _mm_packs_epi16(vy01234567, vy89ABCDEF);
67 __m128i vyGHIJKLMNOPQRSTUV = _mm_packs_epi16(vyGHIJKLMN, vyOPQRSTUV);
68
69 vy0123456789ABCDEF = _mm_max_epi8(vy0123456789ABCDEF, voutput_min);
70 vyGHIJKLMNOPQRSTUV = _mm_max_epi8(vyGHIJKLMNOPQRSTUV, voutput_min);
71
72 _mm_storeu_si128((__m128i*) y, vy0123456789ABCDEF);
73 _mm_storeu_si128((__m128i*) (y + 16), vyGHIJKLMNOPQRSTUV);
74 y += 32;
75 }
76 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
77 __m256 vx = _mm256_loadu_ps(x);
78 vx = _mm256_mul_ps(vx, vscale);
79 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
80 x += 8;
81
82 const __m256i vacc = _mm256_cvtps_epi32(vx);
83
84 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extractf128_si256(vacc, 1));
85 vy = _mm_adds_epi16(vy, voutput_zero_point);
86 vy = _mm_packs_epi16(vy, vy);
87 vy = _mm_max_epi8(vy, voutput_min);
88
89 _mm_storel_epi64((__m128i*) y, vy);
90 y += 8;
91 }
92 if XNN_UNLIKELY(n != 0) {
93 assert(n >= 1 * sizeof(float));
94 assert(n <= 7 * sizeof(float));
95 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
96
97 __m256 vx = _mm256_maskload_ps(x, vmask);
98 vx = _mm256_mul_ps(vx, vscale);
99 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
100
101 const __m256i vacc = _mm256_cvtps_epi32(vx);
102
103 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extractf128_si256(vacc, 1));
104 vy = _mm_adds_epi16(vy, voutput_zero_point);
105 vy = _mm_packs_epi16(vy, vy);
106 vy = _mm_max_epi8(vy, voutput_min);
107
108 if (n & (4 * sizeof(float))) {
109 *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vy);
110 y += 4;
111 vy = _mm_srli_epi64(vy, 32);
112 }
113 if (n & (2 * sizeof(float))) {
114 *((uint16_t*) y) = (uint16_t) _mm_extract_epi16(vy, 0);
115 y += 2;
116 vy = _mm_srli_epi32(vy, 16);
117 }
118 if (n & (1 * sizeof(float))) {
119 *y = (int8_t) _mm_extract_epi8(vy, 0);
120 }
121 }
122 }
123