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_x24(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_x24(
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 >= 24 * sizeof(int8_t); n -= 24 * 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     const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8));
45     const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12));
46     const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16));
47     const __m128i vxKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 20));
48     input_x += 24;
49     input_y += 24;
50 
51     __m128i vacc0123 = _mm_macc_epi32(vx0123, vx_multiplier, vzero_point_product);
52     __m128i vacc4567 = _mm_macc_epi32(vx4567, vx_multiplier, vzero_point_product);
53     __m128i vacc89AB = _mm_macc_epi32(vx89AB, vx_multiplier, vzero_point_product);
54     __m128i vaccCDEF = _mm_macc_epi32(vxCDEF, vx_multiplier, vzero_point_product);
55     __m128i vaccGHIJ = _mm_macc_epi32(vxGHIJ, vx_multiplier, vzero_point_product);
56     __m128i vaccKLMN = _mm_macc_epi32(vxKLMN, vx_multiplier, vzero_point_product);
57 
58     const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
59     const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
60     const __m128i vrem89AB = _mm_add_epi32(_mm_and_si128(vacc89AB, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc89AB));
61     const __m128i vremCDEF = _mm_add_epi32(_mm_and_si128(vaccCDEF, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccCDEF));
62     const __m128i vremGHIJ = _mm_add_epi32(_mm_and_si128(vaccGHIJ, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccGHIJ));
63     const __m128i vremKLMN = _mm_add_epi32(_mm_and_si128(vaccKLMN, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccKLMN));
64 
65     vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
66     vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
67     vacc89AB = _mm_sub_epi32(_mm_sra_epi32(vacc89AB, vshift), _mm_cmpgt_epi32(vrem89AB, vremainder_threshold));
68     vaccCDEF = _mm_sub_epi32(_mm_sra_epi32(vaccCDEF, vshift), _mm_cmpgt_epi32(vremCDEF, vremainder_threshold));
69     vaccGHIJ = _mm_sub_epi32(_mm_sra_epi32(vaccGHIJ, vshift), _mm_cmpgt_epi32(vremGHIJ, vremainder_threshold));
70     vaccKLMN = _mm_sub_epi32(_mm_sra_epi32(vaccKLMN, vshift), _mm_cmpgt_epi32(vremKLMN, vremainder_threshold));
71 
72     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
73     __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
74     __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
75 
76     vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
77     vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
78     voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
79 
80     vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
81     vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
82     voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
83 
84     const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
85     const __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
86 
87     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
88     _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
89     output += 24;
90   }
91   if XNN_UNLIKELY(n != 0) {
92     do {
93       const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
94       const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
95       input_x += 8;
96 
97       __m128i vacc0123 = _mm_macc_epi32(vx0123, vx_multiplier, vzero_point_product);
98       __m128i vacc4567 = _mm_macc_epi32(vx4567, vx_multiplier, vzero_point_product);
99 
100       const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
101       const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
102 
103       vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
104       vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
105 
106       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
107       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
108       vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
109 
110       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
111 
112       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
113         _mm_storel_epi64((__m128i*) output, vout0123456701234567);
114         output += 8;
115         n -= 8 * sizeof(int8_t);
116       } else {
117         if (n & (4 * sizeof(int8_t))) {
118           *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
119           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
120           output += 4;
121         }
122         if (n & (2 * sizeof(int8_t))) {
123           *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
124           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
125           output += 2;
126         }
127         if (n & (1 * sizeof(int8_t))) {
128           *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
129         }
130         n = 0;
131       }
132     } while (n != 0);
133   }
134 }
135