1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
20 #include <xnnpack/math.h>
21
22
23
xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__xop(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__xop(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 const int8_t* restrict a,
29 size_t a_stride,
30 const void* restrict w,
31 int8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 2);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 2);
46 const int8_t* a0 = a;
47 int8_t* c0 = c;
48 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
49 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr != 2) {
51 a1 = a0;
52 c1 = c0;
53 }
54
55 do {
56 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
57 __m128i vacc1x0123 = vacc0x0123;
58 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t k = kc;
61 while (k >= 8 * sizeof(int8_t)) {
62 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
63 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
64 a0 += 8;
65 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
66 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
67 a1 += 8;
68
69 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
70
71 vacc0x0123 = _mm_maddd_epi16(
72 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
73 vacc1x0123 = _mm_maddd_epi16(
74 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
75 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
76
77 vacc0x0123 = _mm_maddd_epi16(
78 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
79 vacc1x0123 = _mm_maddd_epi16(
80 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
81 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
82
83 vacc0x0123 = _mm_maddd_epi16(
84 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
85 vacc1x0123 = _mm_maddd_epi16(
86 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
87 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
88
89 vacc0x0123 = _mm_maddd_epi16(
90 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
91 vacc1x0123 = _mm_maddd_epi16(
92 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
93
94 w = (const void*) ((const int16_t*) w + 32);
95 k -= 8 * sizeof(int8_t);
96 }
97 if (k != 0) {
98 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
99 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
100 a0 = (const int8_t*) ((uintptr_t) a0 + k);
101 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
102 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
103 a1 = (const int8_t*) ((uintptr_t) a1 + k);
104
105 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
106 w = (const void*) ((const int16_t*) w + 8);
107
108 vacc0x0123 = _mm_maddd_epi16(
109 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
110 vacc1x0123 = _mm_maddd_epi16(
111 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
112
113 if (k > 2 * sizeof(int8_t)) {
114 const __m128i vxb1 = _mm_load_si128((const __m128i*) w);
115 w = (const void*) ((const int16_t*) w + 8);
116
117 vacc0x0123 = _mm_maddd_epi16(
118 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
119 vacc1x0123 = _mm_maddd_epi16(
120 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
121
122 if (k > 4 * sizeof(int8_t)) {
123 const __m128i vxb2 = _mm_load_si128((const __m128i*) w);
124 w = (const void*) ((const int16_t*) w + 8);
125
126 vacc0x0123 = _mm_maddd_epi16(
127 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
128 vacc1x0123 = _mm_maddd_epi16(
129 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
130 }
131 }
132 }
133
134 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
135 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
136
137 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
138 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
139 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
140
141 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
142 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
143 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
144
145 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
146 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
147
148 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
149 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
150
151
152 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
153
154 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
155
156 if (nc >= 4) {
157 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
158 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
159
160 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
161 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
162
163 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
164 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
165
166 nc -= 4;
167 } else {
168 if (nc & 2) {
169 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
170 c0 += 2;
171 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
172 c1 += 2;
173 vout = _mm_srli_epi32(vout, 16);
174 }
175 if (nc & 1) {
176 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
177 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
178 }
179
180 nc = 0;
181 }
182 } while (nc != 0);
183 }
184