Debian Codegraph @

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

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/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/intrinsics-polyfill.h"
15 #include "xnnpack/vbinary.h"
16
17
18 void xnn_qu8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_u32(
19 size_t batch,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input_a != NULL);
28 assert(input_b != NULL);
29 assert(output != NULL);
30
31 const __m512i vbias = _mm512_set1_epi32((params->scalar.b_multiplier * (int32_t) *input_b) + params->scalar.bias);
32 const __m512i va_multiplier = _mm512_set1_epi32(params->scalar.a_multiplier);
33 const __m128i vshift = _mm_set1_epi64x(params->scalar.shift);
34 const __m512i voutput_zero_point = _mm512_set1_epi16(params->scalar.output_zero_point);
35 const __m256i voutput_min = _mm256_set1_epi8(params->scalar.output_min);
36 const __m256i voutput_max = _mm256_set1_epi8(params->scalar.output_max);
37
38 XNN_FORCE_REALIZATION(vbias);
39 XNN_FORCE_REALIZATION(va_multiplier);
40 XNN_FORCE_REALIZATION(voutput_zero_point);
41 XNN_FORCE_REALIZATION(voutput_min);
42 XNN_FORCE_REALIZATION(voutput_max);
43
44 for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) {
45 const __m512i va0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input_a));
46 const __m512i vaGHIJKLMNOPQRSTUV = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) (input_a + 16)));
47 input_a += 32;
48
49 __m512i vacc0123456789ABCDEF = _mm512_add_epi32(vbias, _mm512_mullo_epi32(va0123456789ABCDEF, va_multiplier));
50 __m512i vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vbias, _mm512_mullo_epi32(vaGHIJKLMNOPQRSTUV, va_multiplier));
51
52 vacc0123456789ABCDEF = _mm512_sra_epi32(vacc0123456789ABCDEF, vshift);
53 vaccGHIJKLMNOPQRSTUV = _mm512_sra_epi32(vaccGHIJKLMNOPQRSTUV, vshift);
54
55 __m512i vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV = _mm512_adds_epi16(_mm512_packs_epi32(vacc0123456789ABCDEF, vaccGHIJKLMNOPQRSTUV), voutput_zero_point);
56
57 __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));
58
59 vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_max_epu8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_min);
60
61 vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_min_epu8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_max);
62
63 _mm256_storeu_si256((__m256i*) output, vout0123456789ABCDEFGHIJKLMNOPQRSTUV);
64 output += 32;
65 }
66 if XNN_UNLIKELY(batch != 0) {
67 do {
68 const __m512i va0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input_a));
69 input_a += 16;
70
71 __m512i vacc0123456789ABCDEF = _mm512_add_epi32(vbias, _mm512_mullo_epi32(va0123456789ABCDEF, va_multiplier));
72
73 vacc0123456789ABCDEF = _mm512_sra_epi32(vacc0123456789ABCDEF, vshift);
74
75 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), _mm512_castsi512_si256(voutput_zero_point));
76 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
77 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_min));
78 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_max));
79
80 if XNN_LIKELY(batch >= (16 * sizeof(uint8_t))) {
81 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
82 output += 16;
83 batch -= 16 * sizeof(uint8_t);
84 } else {
85 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
86 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
87 batch = 0;
88 }
89 } while (batch != 0);
90 }
91 }