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/vadd.h>
16
17
xnn_qu8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x32(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x32(
19 size_t n,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 const __m512i va_multiplier = _mm512_load_si512(params->avx512.a_multiplier);
26 const __m128i vshift = _mm_load_si128((const __m128i*) params->avx512.shift);
27 const __m512i voutput_zero_point = _mm512_load_si512(params->avx512.output_zero_point);
28 const __m256i voutput_min = _mm256_load_si256((const __m256i*) params->avx512.output_min);
29 const __m256i voutput_max = _mm256_load_si256((const __m256i*) params->avx512.output_max);
30
31 const __m512i vbias = _mm512_add_epi32(
32 _mm512_broadcastd_epi32(_mm_cvtsi32_si128(params->avx512.b_multiplier[0] * (int32_t) *input_b)),
33 _mm512_load_si512(params->avx512.bias));
34 for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
35 const __m512i va0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input_a));
36 const __m512i vaGHIJKLMNOPQRSTUV = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) (input_a + 16)));
37 input_a += 32;
38
39 __m512i vacc0123456789ABCDEF = _mm512_add_epi32(vbias, _mm512_mullo_epi32(va0123456789ABCDEF, va_multiplier));
40 __m512i vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vbias, _mm512_mullo_epi32(vaGHIJKLMNOPQRSTUV, va_multiplier));
41
42 vacc0123456789ABCDEF = _mm512_sra_epi32(vacc0123456789ABCDEF, vshift);
43 vaccGHIJKLMNOPQRSTUV = _mm512_sra_epi32(vaccGHIJKLMNOPQRSTUV, vshift);
44
45 __m512i vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV = _mm512_adds_epi16(_mm512_packs_epi32(vacc0123456789ABCDEF, vaccGHIJKLMNOPQRSTUV), voutput_zero_point);
46
47 __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));
48
49 vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_max_epu8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_min);
50
51 vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_min_epu8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_max);
52
53 _mm256_storeu_si256((__m256i*) output, vout0123456789ABCDEFGHIJKLMNOPQRSTUV);
54 output += 32;
55 }
56 if XNN_UNLIKELY(n != 0) {
57 do {
58 const __m512i va0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input_a));
59 input_a += 16;
60
61 __m512i vacc0123456789ABCDEF = _mm512_add_epi32(vbias, _mm512_mullo_epi32(va0123456789ABCDEF, va_multiplier));
62
63 vacc0123456789ABCDEF = _mm512_sra_epi32(vacc0123456789ABCDEF, vshift);
64
65 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), _mm512_castsi512_si256(voutput_zero_point));
66 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
67 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_min));
68 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_max));
69
70 if XNN_LIKELY(n >= (16 * sizeof(uint8_t))) {
71 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
72 output += 16;
73 n -= 16 * sizeof(uint8_t);
74 } else {
75 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << n) - UINT32_C(1)));
76 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
77 n = 0;
78 }
79 } while (n != 0);
80 }
81 }
82