Debian Codegraph @

src/qu8-vadd/gen/qu8-vadd-minmax-avx512skx-mul32-ld128-u32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/avx512skx-mul32-ld128.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/vbinary.h"
17
18
19 void xnn_qu8_vadd_minmax_ukernel__avx512skx_mul32_ld128_u32(
20 size_t batch,
21 const uint8_t* input_a,
22 const uint8_t* input_b,
23 uint8_t* output,
24 const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(batch != 0);
27 assert(batch % sizeof(uint8_t) == 0);
28 assert(input_a != NULL);
29 assert(input_b != NULL);
30 assert(output != NULL);
31
32 const __m512i vbias = _mm512_set1_epi32(params->scalar.bias);
33 const __m512i va_multiplier = _mm512_set1_epi32(params->scalar.a_multiplier);
34 const __m512i vb_multiplier = _mm512_set1_epi32(params->scalar.b_multiplier);
35 const __m128i vshift = _mm_set1_epi64x(params->scalar.shift);
36 const __m512i voutput_zero_point = _mm512_set1_epi16(params->scalar.output_zero_point);
37 const __m256i voutput_min = _mm256_set1_epi8(params->scalar.output_min);
38 const __m256i voutput_max = _mm256_set1_epi8(params->scalar.output_max);
39
40 XNN_FORCE_REALIZATION(vbias);
41 XNN_FORCE_REALIZATION(va_multiplier);
42 XNN_FORCE_REALIZATION(vb_multiplier);
43 XNN_FORCE_REALIZATION(vshift);
44 XNN_FORCE_REALIZATION(voutput_zero_point);
45 XNN_FORCE_REALIZATION(voutput_min);
46 XNN_FORCE_REALIZATION(voutput_max);
47
48 for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) {
49 const __m512i va0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input_a));
50 const __m512i vb0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input_b));
51 const __m512i vaGHIJKLMNOPQRSTUV = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) (input_a + 16)));
52 const __m512i vbGHIJKLMNOPQRSTUV = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) (input_b + 16)));
53 input_a += 32;
54 input_b += 32;
55
56 __m512i vacc0123456789ABCDEF = _mm512_add_epi32(vbias, _mm512_mullo_epi32(va0123456789ABCDEF, va_multiplier));
57 __m512i vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vbias, _mm512_mullo_epi32(vaGHIJKLMNOPQRSTUV, va_multiplier));
58
59 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vb0123456789ABCDEF, vb_multiplier));
60 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vbGHIJKLMNOPQRSTUV, vb_multiplier));
61
62 vacc0123456789ABCDEF = _mm512_sra_epi32(vacc0123456789ABCDEF, vshift);
63 vaccGHIJKLMNOPQRSTUV = _mm512_sra_epi32(vaccGHIJKLMNOPQRSTUV, vshift);
64
65 __m512i vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV = _mm512_adds_epi16(_mm512_packs_epi32(vacc0123456789ABCDEF, vaccGHIJKLMNOPQRSTUV), voutput_zero_point);
66
67 __m256i vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_permutevar8x32_epi32(_mm256_packus_epi16(_mm512_castsi512_si256(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV), _mm512_extracti32x8_epi32(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV, 1)), _mm256_set_epi32(7, 3, 5, 1, 6, 2, 4, 0));
68
69 vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_max_epu8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_min);
70
71 vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_min_epu8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_max);
72
73 _mm256_storeu_si256((__m256i*) output, vout0123456789ABCDEFGHIJKLMNOPQRSTUV);
74 output += 32;
75 }
76 if XNN_UNLIKELY(batch != 0) {
77 do {
78 const __m512i va0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input_a));
79 const __m512i vb0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input_b));
80 input_a += 16;
81 input_b += 16;
82
83 __m512i vacc0123456789ABCDEF = _mm512_add_epi32(vbias, _mm512_mullo_epi32(va0123456789ABCDEF, va_multiplier));
84
85 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vb0123456789ABCDEF, vb_multiplier));
86
87 vacc0123456789ABCDEF = _mm512_sra_epi32(vacc0123456789ABCDEF, vshift);
88
89 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), _mm512_castsi512_si256(voutput_zero_point));
90 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
91 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_min));
92 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_max));
93
94 if XNN_LIKELY(batch >= (16 * sizeof(uint8_t))) {
95 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
96 output += 16;
97 batch -= 16 * sizeof(uint8_t);
98 } else {
99 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
100 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
101 batch = 0;
102 }
103 } while (batch != 0);
104 }
105 }