• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-vaddc/sse-mul32-ld32.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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 #ifdef __GNUC__
13   #include <x86intrin.h>
14 #else
15   #include <immintrin.h>
16   #include <ammintrin.h>
17 #endif
18 
19 #include <xnnpack/intrinsics-polyfill.h>
20 #include <xnnpack/vadd.h>
21 
22 
xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8(size_t n,const int8_t * input_x,const int8_t * input_y,int8_t * output,const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8(
24     size_t n,
25     const int8_t* input_x,
26     const int8_t* input_y,
27     int8_t* output,
28     const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30   const __m128i vx_multiplier = _mm_load_si128((const __m128i*) params->sse2.x_multiplier);
31   const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
32   const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
33   const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
34   const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
35   const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
36   const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
37 
38   __m128i vzero_point_product = _mm_cvtsi32_si128(params->sse2.y_multiplier[0] * (int32_t) *input_y);
39   vzero_point_product = _mm_shuffle_epi32(vzero_point_product, _MM_SHUFFLE(0, 0, 0, 0));
40   vzero_point_product = _mm_add_epi32(vzero_point_product, _mm_load_si128((const __m128i*) params->sse2.zero_point_product));
41   for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
42     const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
43     const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
44     input_x += 8;
45     input_y += 8;
46 
47     __m128i vacc0123 = _mm_macc_epi32(vx0123, vx_multiplier, vzero_point_product);
48     __m128i vacc4567 = _mm_macc_epi32(vx4567, vx_multiplier, vzero_point_product);
49 
50     const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
51     const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
52 
53     vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
54     vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
55 
56     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
57 
58     vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
59 
60     vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
61 
62     const __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
63 
64     _mm_storel_epi64((__m128i*) output, vout0123456701234567);
65     output += 8;
66   }
67   if XNN_UNLIKELY(n != 0) {
68     {
69       const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
70       const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
71 
72       __m128i vacc0123 = _mm_macc_epi32(vx0123, vx_multiplier, vzero_point_product);
73       __m128i vacc4567 = _mm_macc_epi32(vx4567, vx_multiplier, vzero_point_product);
74 
75       const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
76       const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
77 
78       vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
79       vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
80 
81       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
82       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
83       vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
84 
85       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
86 
87       if (n & (4 * sizeof(int8_t))) {
88         *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
89         vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
90         output += 4;
91       }
92       if (n & (2 * sizeof(int8_t))) {
93         *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
94         vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
95         output += 2;
96       }
97       if (n & (1 * sizeof(int8_t))) {
98         *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
99       }
100     }
101   }
102 }
103