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 #include <immintrin.h>
13 
14 #include <xnnpack/intrinsics-polyfill.h>
15 #include <xnnpack/vadd.h>
16 
17 
xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16(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)])18 void xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16(
19     size_t n,
20     const int8_t* input_x,
21     const int8_t* input_y,
22     int8_t* output,
23     const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
24 {
25   const __m128i vx_multiplier = _mm_load_si128((const __m128i*) params->sse2.x_multiplier);
26   const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
27   const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
28   const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
29   const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
30   const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
31   const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
32 
33   __m128i vzero_point_product = _mm_cvtsi32_si128(params->sse2.y_multiplier[0] * (int32_t) *input_y);
34   vzero_point_product = _mm_shuffle_epi32(vzero_point_product, _MM_SHUFFLE(0, 0, 0, 0));
35   vzero_point_product = _mm_add_epi32(vzero_point_product, _mm_load_si128((const __m128i*) params->sse2.zero_point_product));
36   for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
37     const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
38     const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
39     const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8));
40     const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12));
41     input_x += 16;
42     input_y += 16;
43 
44     __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx0123, vx_multiplier));
45     __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx4567, vx_multiplier));
46     __m128i vacc89AB = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx89AB, vx_multiplier));
47     __m128i vaccCDEF = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vxCDEF, vx_multiplier));
48 
49     const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
50     const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
51     const __m128i vrem89AB = _mm_add_epi32(_mm_and_si128(vacc89AB, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc89AB));
52     const __m128i vremCDEF = _mm_add_epi32(_mm_and_si128(vaccCDEF, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccCDEF));
53 
54     vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
55     vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
56     vacc89AB = _mm_sub_epi32(_mm_sra_epi32(vacc89AB, vshift), _mm_cmpgt_epi32(vrem89AB, vremainder_threshold));
57     vaccCDEF = _mm_sub_epi32(_mm_sra_epi32(vaccCDEF, vshift), _mm_cmpgt_epi32(vremCDEF, vremainder_threshold));
58 
59     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
60     __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
61 
62     vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
63     vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
64 
65     vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
66     vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
67 
68     const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
69 
70     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
71     output += 16;
72   }
73   if XNN_UNLIKELY(n != 0) {
74     do {
75       const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
76       const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
77       input_x += 8;
78 
79       __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx0123, vx_multiplier));
80       __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx4567, vx_multiplier));
81 
82       const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
83       const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
84 
85       vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
86       vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
87 
88       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
89       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
90       vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
91 
92       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
93 
94       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
95         _mm_storel_epi64((__m128i*) output, vout0123456701234567);
96         output += 8;
97         n -= 8 * sizeof(int8_t);
98       } else {
99         if (n & (4 * sizeof(int8_t))) {
100           *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
101           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
102           output += 4;
103         }
104         if (n & (2 * sizeof(int8_t))) {
105           *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
106           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
107           output += 2;
108         }
109         if (n & (1 * sizeof(int8_t))) {
110           *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
111         }
112         n = 0;
113       }
114     } while (n != 0);
115   }
116 }
117