1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-sse.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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/igemm.h>
20 #include <xnnpack/math.h>
21
22
xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld128(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 size_t ks,
28 const int8_t** restrict a,
29 const void* restrict w,
30 int8_t* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 size_t a_offset,
34 const int8_t* zero,
35 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 1);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(ks != 0);
42 assert(ks % (1 * sizeof(void*)) == 0);
43 assert(a_offset % sizeof(int8_t) == 0);
44 assert(a != NULL);
45 assert(w != NULL);
46 assert(c != NULL);
47
48 kc = round_up_po2(kc, 2);
49 int8_t* c0 = c;
50
51 do {
52 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
53 w = (const void*) ((const int32_t*) w + 4);
54
55 size_t p = ks;
56 do {
57 const int8_t* restrict a0 = a[0];
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60 }
61 a += 1;
62
63 size_t k = kc;
64 while (k >= 8 * sizeof(int8_t)) {
65 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
66 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
67 a0 += 8;
68
69 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
70 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
71 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
72
73 vacc0x0123 = _mm_maddd_epi16(
74 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
75
76 vacc0x0123 = _mm_maddd_epi16(
77 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
78 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
79 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
80 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
81
82 vacc0x0123 = _mm_maddd_epi16(
83 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
84
85 vacc0x0123 = _mm_maddd_epi16(
86 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
87
88 w = (const void*) ((const int8_t*) w + 32);
89 k -= 8 * sizeof(int8_t);
90 }
91 if (k != 0) {
92 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
93 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
94 a0 = (const int8_t*) ((uintptr_t) a0 + k);
95
96 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
97 w = (const void*) ((const int8_t*) w + 8);
98 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
99
100 vacc0x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
102
103 if (k > 2 * sizeof(int8_t)) {
104 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
105 w = (const void*) ((const int8_t*) w + 8);
106 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
107
108 vacc0x0123 = _mm_maddd_epi16(
109 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
110
111 if (k > 4 * sizeof(int8_t)) {
112 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
113 w = (const void*) ((const int8_t*) w + 8);
114 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
115
116 vacc0x0123 = _mm_maddd_epi16(
117 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
118 }
119 }
120 }
121 p -= 1 * sizeof(void*);
122 } while (p != 0);
123
124 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
125
126 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
127 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
128
129 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
130 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
131
132 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
133
134 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
135 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
136
137
138 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
139
140 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
141
142 if (nc >= 4) {
143 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
144 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
145
146 a = (const int8_t**restrict) ((uintptr_t) a - ks);
147
148 nc -= 4;
149 } else {
150 if (nc & 2) {
151 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
152 c0 += 2;
153 vout = _mm_srli_epi32(vout, 16);
154 }
155 if (nc & 1) {
156 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
157 }
158
159 nc = 0;
160 }
161 } while (nc != 0);
162 }
163