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 #include <xnnpack/unaligned.h>
22
23
24
xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__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)])25 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__xop(
26 size_t mr,
27 size_t nc,
28 size_t kc,
29 const int8_t* restrict a,
30 size_t a_stride,
31 const void* restrict w,
32 int8_t* restrict c,
33 size_t cm_stride,
34 size_t cn_stride,
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 <= 3);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 2 * sizeof(int8_t));
47 const int8_t* a0 = a;
48 int8_t* c0 = c;
49 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
50 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
56 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 2) {
58 a2 = a1;
59 c2 = c1;
60 }
61
62 do {
63 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
64 __m128i vacc1x0123 = vacc0x0123;
65 __m128i vacc2x0123 = vacc0x0123;
66 w = (const void*) ((const int32_t*) w + 4);
67
68 size_t k = kc;
69 while (k >= 8 * sizeof(int8_t)) {
70 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
71 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
72 a0 += 8;
73 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
74 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
75 a1 += 8;
76 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
77 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
78 a2 += 8;
79
80 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
81
82 vacc0x0123 = _mm_maddd_epi16(
83 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
84 vacc1x0123 = _mm_maddd_epi16(
85 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
86 vacc2x0123 = _mm_maddd_epi16(
87 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
88 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
89
90 vacc0x0123 = _mm_maddd_epi16(
91 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
92 vacc1x0123 = _mm_maddd_epi16(
93 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
94 vacc2x0123 = _mm_maddd_epi16(
95 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
96 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
97
98 vacc0x0123 = _mm_maddd_epi16(
99 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
100 vacc1x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
102 vacc2x0123 = _mm_maddd_epi16(
103 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
104 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
105
106 vacc0x0123 = _mm_maddd_epi16(
107 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
108 vacc1x0123 = _mm_maddd_epi16(
109 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
110 vacc2x0123 = _mm_maddd_epi16(
111 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
112
113 w = (const void*) ((const int16_t*) w + 32);
114 k -= 8 * sizeof(int8_t);
115 }
116 if (k != 0) {
117 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
118 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
119 a0 = (const int8_t*) ((uintptr_t) a0 + k);
120 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
121 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
122 a1 = (const int8_t*) ((uintptr_t) a1 + k);
123 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
124 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
125 a2 = (const int8_t*) ((uintptr_t) a2 + k);
126
127 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
128 w = (const void*) ((const int16_t*) w + 8);
129
130 vacc0x0123 = _mm_maddd_epi16(
131 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
132 vacc1x0123 = _mm_maddd_epi16(
133 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
134 vacc2x0123 = _mm_maddd_epi16(
135 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
136
137 if (k > 2 * sizeof(int8_t)) {
138 const __m128i vxb1 = _mm_load_si128((const __m128i*) w);
139 w = (const void*) ((const int16_t*) w + 8);
140
141 vacc0x0123 = _mm_maddd_epi16(
142 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
143 vacc1x0123 = _mm_maddd_epi16(
144 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
145 vacc2x0123 = _mm_maddd_epi16(
146 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
147
148 if (k > 4 * sizeof(int8_t)) {
149 const __m128i vxb2 = _mm_load_si128((const __m128i*) w);
150 w = (const void*) ((const int16_t*) w + 8);
151
152 vacc0x0123 = _mm_maddd_epi16(
153 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
154 vacc1x0123 = _mm_maddd_epi16(
155 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
156 vacc2x0123 = _mm_maddd_epi16(
157 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
158 }
159 }
160 }
161
162 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
163 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
164 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
165
166 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
167 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
168 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
169 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
170
171 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
172 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
173 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
174 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
175
176 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
177 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
178 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
179
180 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
181 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
182 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
183
184
185 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
186
187 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
188
189 if (nc >= 4) {
190 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
191 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
192 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
193
194 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
195 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
196 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
197
198 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
199 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
200 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
201
202 nc -= 4;
203 } else {
204 if (nc & 2) {
205 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
206 c0 += 2;
207 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
208 c1 += 2;
209 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
210 c2 += 2;
211 vout = _mm_srli_epi32(vout, 16);
212 }
213 if (nc & 1) {
214 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
215 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
216 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
217 }
218
219 nc = 0;
220 }
221 } while (nc != 0);
222 }
223