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/intrinsics-polyfill.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_f32_qu8_vcvt_ukernel__avx_x32(size_t n,const float * x,uint8_t * y,const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_qu8_vcvt_ukernel__avx_x32(
20 size_t n,
21 const float* x,
22 uint8_t* y,
23 const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m256 vscale = _mm256_load_ps(params->avx.scale);
31 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->avx.output_max_less_zero_point);
32 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx.output_zero_point);
33 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx.output_min);
34
35 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
36 __m256 vx01234567 = _mm256_loadu_ps(x);
37 __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8);
38 __m256 vxGHIJKLMN = _mm256_loadu_ps(x + 16);
39 __m256 vxOPQRSTUV = _mm256_loadu_ps(x + 24);
40 x += 32;
41
42 vx01234567 = _mm256_mul_ps(vx01234567, vscale);
43 vx89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vscale);
44 vxGHIJKLMN = _mm256_mul_ps(vxGHIJKLMN, vscale);
45 vxOPQRSTUV = _mm256_mul_ps(vxOPQRSTUV, vscale);
46
47 vx01234567 = _mm256_min_ps(vx01234567, voutput_max_less_zero_point);
48 vx89ABCDEF = _mm256_min_ps(vx89ABCDEF, voutput_max_less_zero_point);
49 vxGHIJKLMN = _mm256_min_ps(vxGHIJKLMN, voutput_max_less_zero_point);
50 vxOPQRSTUV = _mm256_min_ps(vxOPQRSTUV, voutput_max_less_zero_point);
51
52 const __m256i vacc01234567 = _mm256_cvtps_epi32(vx01234567);
53 const __m256i vacc89ABCDEF = _mm256_cvtps_epi32(vx89ABCDEF);
54 const __m256i vaccGHIJKLMN = _mm256_cvtps_epi32(vxGHIJKLMN);
55 const __m256i vaccOPQRSTUV = _mm256_cvtps_epi32(vxOPQRSTUV);
56
57 __m128i vy01234567 = _mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extractf128_si256(vacc01234567, 1));
58 __m128i vy89ABCDEF = _mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extractf128_si256(vacc89ABCDEF, 1));
59 __m128i vyGHIJKLMN = _mm_packs_epi32(_mm256_castsi256_si128(vaccGHIJKLMN), _mm256_extractf128_si256(vaccGHIJKLMN, 1));
60 __m128i vyOPQRSTUV = _mm_packs_epi32(_mm256_castsi256_si128(vaccOPQRSTUV), _mm256_extractf128_si256(vaccOPQRSTUV, 1));
61
62 vy01234567 = _mm_adds_epi16(vy01234567, voutput_zero_point);
63 vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point);
64 vyGHIJKLMN = _mm_adds_epi16(vyGHIJKLMN, voutput_zero_point);
65 vyOPQRSTUV = _mm_adds_epi16(vyOPQRSTUV, voutput_zero_point);
66
67 __m128i vy0123456789ABCDEF = _mm_packus_epi16(vy01234567, vy89ABCDEF);
68 __m128i vyGHIJKLMNOPQRSTUV = _mm_packus_epi16(vyGHIJKLMN, vyOPQRSTUV);
69
70 vy0123456789ABCDEF = _mm_max_epu8(vy0123456789ABCDEF, voutput_min);
71 vyGHIJKLMNOPQRSTUV = _mm_max_epu8(vyGHIJKLMNOPQRSTUV, voutput_min);
72
73 _mm_storeu_si128((__m128i*) y, vy0123456789ABCDEF);
74 _mm_storeu_si128((__m128i*) (y + 16), vyGHIJKLMNOPQRSTUV);
75 y += 32;
76 }
77 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
78 __m256 vx = _mm256_loadu_ps(x);
79 vx = _mm256_mul_ps(vx, vscale);
80 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
81 x += 8;
82
83 const __m256i vacc = _mm256_cvtps_epi32(vx);
84
85 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extractf128_si256(vacc, 1));
86 vy = _mm_adds_epi16(vy, voutput_zero_point);
87 vy = _mm_packus_epi16(vy, vy);
88 vy = _mm_max_epu8(vy, voutput_min);
89
90 _mm_storel_epi64((__m128i*) y, vy);
91 y += 8;
92 }
93 if XNN_UNLIKELY(n != 0) {
94 assert(n >= 1 * sizeof(float));
95 assert(n <= 7 * sizeof(float));
96 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
97
98 __m256 vx = _mm256_maskload_ps(x, vmask);
99 vx = _mm256_mul_ps(vx, vscale);
100 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
101
102 const __m256i vacc = _mm256_cvtps_epi32(vx);
103
104 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extractf128_si256(vacc, 1));
105 vy = _mm_adds_epi16(vy, voutput_zero_point);
106 vy = _mm_packus_epi16(vy, vy);
107 vy = _mm_max_epu8(vy, voutput_min);
108
109 if (n & (4 * sizeof(float))) {
110 _mm_storeu_si32(y, vy);
111 y += 4;
112 vy = _mm_srli_epi64(vy, 32);
113 }
114 if (n & (2 * sizeof(float))) {
115 _mm_storeu_si16(y, vy);
116 y += 2;
117 vy = _mm_srli_epi32(vy, 16);
118 }
119 if (n & (1 * sizeof(float))) {
120 *y = (uint8_t) _mm_extract_epi8(vy, 0);
121 }
122 }
123 }
124