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_x16(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_x16(
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 __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx512.output_zero_point);
28 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx512.output_min);
29 const __m128i voutput_max = _mm_load_si128((const __m128i*) 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 >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
35 const __m512i va0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input_a));
36 input_a += 16;
37
38 __m512i vacc0123456789ABCDEF = _mm512_add_epi32(vbias, _mm512_mullo_epi32(va0123456789ABCDEF, va_multiplier));
39
40 vacc0123456789ABCDEF = _mm512_sra_epi32(vacc0123456789ABCDEF, vshift);
41
42 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
43
44 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
45
46 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
47
48 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
49
50 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
51 output += 16;
52 }
53 if XNN_UNLIKELY(n != 0) {
54 {
55 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << n) - UINT32_C(1)));
56 const __m512i va0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_maskz_loadu_epi8(vmask, input_a));
57
58 __m512i vacc0123456789ABCDEF = _mm512_add_epi32(vbias, _mm512_mullo_epi32(va0123456789ABCDEF, va_multiplier));
59
60 vacc0123456789ABCDEF = _mm512_sra_epi32(vacc0123456789ABCDEF, vshift);
61
62 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
63 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
64 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
65 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
66
67 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
68 }
69 }
70 }
71