1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/avx512skx.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__avx512skx_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__avx512skx_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 __m512 vscale = _mm512_load_ps(params->avx2.scale);
31 const __m512 voutput_max_less_zero_point = _mm512_load_ps(params->avx512.output_max_less_zero_point);
32 const __m512i voutput_zero_point = _mm512_load_si512(params->avx512.output_zero_point);
33 const __m256i vshuffle256_mask = _mm256_load_si256((const __m256i*) params->avx512.shuffle256_mask);
34 const __m256i voutput_min = _mm256_load_si256((const __m256i*) params->avx512.output_min);
35 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
36 __m512 vx0123 = _mm512_loadu_ps(x);
37 __m512 vx4567 = _mm512_loadu_ps(x + 16);
38 x += 32;
39
40 vx0123 = _mm512_mul_ps(vx0123, vscale);
41 vx4567 = _mm512_mul_ps(vx4567, vscale);
42
43 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point);
44 vx4567 = _mm512_min_ps(vx4567, voutput_max_less_zero_point);
45
46 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123);
47 const __m512i vacc4567 = _mm512_cvtps_epi32(vx4567);
48
49 __m512i vacc04152637 = _mm512_packs_epi32(vacc0123, vacc4567);
50
51 vacc04152637 = _mm512_adds_epi16(vacc04152637, voutput_zero_point);
52
53 __m256i vy04261537 = _mm256_packus_epi16(_mm512_castsi512_si256(vacc04152637), _mm512_extracti32x8_epi32(vacc04152637, 1));
54
55 vy04261537 = _mm256_max_epu8(vy04261537, voutput_min);
56
57 const __m256i vy01234567 = _mm256_permutevar8x32_epi32(vy04261537, vshuffle256_mask);
58
59 _mm256_storeu_si256((__m256i*) y, vy01234567);
60 y += 32;
61 }
62 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
63 __m512 vx0123 = _mm512_loadu_ps(x);
64 vx0123 = _mm512_mul_ps(vx0123, vscale);
65 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point);
66 x += 16;
67
68 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123);
69
70 __m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(vacc0123, 1));
71 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point));
72 const __m128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213, 1));
73 __m128i vy0123 = _mm_shuffle_epi32(vy0213, _MM_SHUFFLE(3, 1, 2, 0));
74 vy0123 = _mm_max_epu8(vy0123, _mm256_castsi256_si128(voutput_min));
75
76 _mm_storeu_si128((__m128i*) y, vy0123);
77 y += 16;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 assert(n >= 1 * sizeof(float));
81 assert(n <= 15 * sizeof(float));
82
83 // Prepare mask for valid elements (depends on n).
84 n >>= 2 /* log2(sizeof(float)) */;
85 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
86
87 __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x);
88 vx0123 = _mm512_mul_ps(vx0123, vscale);
89 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point);
90
91 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123);
92
93 __m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(vacc0123, 1));
94 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point));
95 const __m128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213, 1));
96 __m128i vy0123 = _mm_shuffle_epi32(vy0213, _MM_SHUFFLE(3, 1, 2, 0));
97 vy0123 = _mm_max_epu8(vy0123, _mm256_castsi256_si128(voutput_min));
98
99 _mm_mask_storeu_epi8(y, vmask, vy0123);
100 }
101 }
102