1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/avx2.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__avx2_x64(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__avx2_x64(
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->avx2.scale);
31 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->avx2.output_max_less_zero_point);
32 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
33 const __m256i vshuffle_mask = _mm256_load_si256((const __m256i*) params->avx2.shuffle_mask);
34 const __m256i voutput_min = _mm256_load_si256((const __m256i*) params->avx2.output_min);
35
36 for (; n >= 64 * sizeof(float); n -= 64 * sizeof(float)) {
37 __m256 vx01 = _mm256_loadu_ps(x);
38 __m256 vx23 = _mm256_loadu_ps(x + 8);
39 __m256 vx45 = _mm256_loadu_ps(x + 16);
40 __m256 vx67 = _mm256_loadu_ps(x + 24);
41 __m256 vx89 = _mm256_loadu_ps(x + 32);
42 __m256 vxAB = _mm256_loadu_ps(x + 40);
43 __m256 vxCD = _mm256_loadu_ps(x + 48);
44 __m256 vxEF = _mm256_loadu_ps(x + 56);
45 x += 64;
46
47 vx01 = _mm256_mul_ps(vx01, vscale);
48 vx23 = _mm256_mul_ps(vx23, vscale);
49 vx45 = _mm256_mul_ps(vx45, vscale);
50 vx67 = _mm256_mul_ps(vx67, vscale);
51 vx89 = _mm256_mul_ps(vx89, vscale);
52 vxAB = _mm256_mul_ps(vxAB, vscale);
53 vxCD = _mm256_mul_ps(vxCD, vscale);
54 vxEF = _mm256_mul_ps(vxEF, vscale);
55
56 vx01 = _mm256_min_ps(vx01, voutput_max_less_zero_point);
57 vx23 = _mm256_min_ps(vx23, voutput_max_less_zero_point);
58 vx45 = _mm256_min_ps(vx45, voutput_max_less_zero_point);
59 vx67 = _mm256_min_ps(vx67, voutput_max_less_zero_point);
60 vx89 = _mm256_min_ps(vx89, voutput_max_less_zero_point);
61 vxAB = _mm256_min_ps(vxAB, voutput_max_less_zero_point);
62 vxCD = _mm256_min_ps(vxCD, voutput_max_less_zero_point);
63 vxEF = _mm256_min_ps(vxEF, voutput_max_less_zero_point);
64
65 const __m256i vacc01 = _mm256_cvtps_epi32(vx01);
66 const __m256i vacc23 = _mm256_cvtps_epi32(vx23);
67 const __m256i vacc45 = _mm256_cvtps_epi32(vx45);
68 const __m256i vacc67 = _mm256_cvtps_epi32(vx67);
69 const __m256i vacc89 = _mm256_cvtps_epi32(vx89);
70 const __m256i vaccAB = _mm256_cvtps_epi32(vxAB);
71 const __m256i vaccCD = _mm256_cvtps_epi32(vxCD);
72 const __m256i vaccEF = _mm256_cvtps_epi32(vxEF);
73
74 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23);
75 __m256i vacc4657 = _mm256_packs_epi32(vacc45, vacc67);
76 __m256i vacc8A9B = _mm256_packs_epi32(vacc89, vaccAB);
77 __m256i vaccCEDF = _mm256_packs_epi32(vaccCD, vaccEF);
78
79 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point);
80 vacc4657 = _mm256_adds_epi16(vacc4657, voutput_zero_point);
81 vacc8A9B = _mm256_adds_epi16(vacc8A9B, voutput_zero_point);
82 vaccCEDF = _mm256_adds_epi16(vaccCEDF, voutput_zero_point);
83
84 const __m256i vy02461357 = _mm256_packus_epi16(vacc0213, vacc4657);
85 const __m256i vy8ACE9BDF = _mm256_packus_epi16(vacc8A9B, vaccCEDF);
86
87 __m256i vy01234567 = _mm256_permutevar8x32_epi32(vy02461357, vshuffle_mask);
88 __m256i vy89ABCDEF = _mm256_permutevar8x32_epi32(vy8ACE9BDF, vshuffle_mask);
89
90 vy01234567 = _mm256_max_epu8(vy01234567, voutput_min);
91 vy89ABCDEF = _mm256_max_epu8(vy89ABCDEF, voutput_min);
92
93 _mm256_storeu_si256((__m256i*) y, vy01234567);
94 _mm256_storeu_si256((__m256i*) (y + 32), vy89ABCDEF);
95 y += 64;
96 }
97 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
98 __m256 vx = _mm256_loadu_ps(x);
99 vx = _mm256_mul_ps(vx, vscale);
100 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
101 x += 8;
102
103 const __m256i vacc = _mm256_cvtps_epi32(vx);
104
105 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
106 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
107 vy = _mm_packus_epi16(vy, vy);
108 vy = _mm_max_epu8(vy, _mm256_castsi256_si128(voutput_min));
109
110 _mm_storel_epi64((__m128i*) y, vy);
111 y += 8;
112 }
113 if XNN_UNLIKELY(n != 0) {
114 assert(n >= 1 * sizeof(float));
115 assert(n <= 7 * sizeof(float));
116 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2.mask_table[7] - n));
117
118 __m256 vx = _mm256_maskload_ps(x, vmask);
119 vx = _mm256_mul_ps(vx, vscale);
120 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
121
122 const __m256i vacc = _mm256_cvtps_epi32(vx);
123
124 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
125 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
126 vy = _mm_packus_epi16(vy, vy);
127 vy = _mm_max_epu8(vy, _mm256_castsi256_si128(voutput_min));
128
129 if (n & (4 * sizeof(float))) {
130 _mm_storeu_si32(y, vy);
131 y += 4;
132 vy = _mm_srli_epi64(vy, 32);
133 }
134 if (n & (2 * sizeof(float))) {
135 _mm_storeu_si16(y, vy);
136 y += 2;
137 vy = _mm_srli_epi32(vy, 16);
138 }
139 if (n & (1 * sizeof(float))) {
140 *y = (uint8_t) _mm_extract_epi8(vy, 0);
141 }
142 }
143 }
144