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 #ifdef __GNUC__
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
xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__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_gemm_xw_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__xop(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 const int8_t* restrict a,
28 size_t a_stride,
29 const void* restrict w,
30 int8_t* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 const union xnn_qs8_gemm_xw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
34 {
35 assert(mr != 0);
36 assert(mr <= 4);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2);
45 const int8_t* a0 = a;
46 int8_t* c0 = c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr != 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65
66 do {
67 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
68 __m128i vacc1x0123 = vacc0x0123;
69 __m128i vacc2x0123 = vacc0x0123;
70 __m128i vacc3x0123 = vacc0x0123;
71 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
72
73 size_t k = kc;
74 while (k >= 8 * sizeof(int8_t)) {
75 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
76 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
77 a0 += 8;
78 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
79 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
80 a1 += 8;
81 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
82 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
83 a2 += 8;
84 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
85 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
86 a3 += 8;
87
88 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
89
90 vacc0x0123 = _mm_maddd_epi16(
91 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
92 vacc1x0123 = _mm_maddd_epi16(
93 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
94 vacc2x0123 = _mm_maddd_epi16(
95 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
96 vacc3x0123 = _mm_maddd_epi16(
97 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
98 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 8 * sizeof(int16_t)));
99
100 vacc0x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
102 vacc1x0123 = _mm_maddd_epi16(
103 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
104 vacc2x0123 = _mm_maddd_epi16(
105 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
106 vacc3x0123 = _mm_maddd_epi16(
107 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
108 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int16_t)));
109
110 vacc0x0123 = _mm_maddd_epi16(
111 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
112 vacc1x0123 = _mm_maddd_epi16(
113 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
114 vacc2x0123 = _mm_maddd_epi16(
115 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
116 vacc3x0123 = _mm_maddd_epi16(
117 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
118 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 24 * sizeof(int16_t)));
119
120 vacc0x0123 = _mm_maddd_epi16(
121 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
122 vacc1x0123 = _mm_maddd_epi16(
123 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
124 vacc2x0123 = _mm_maddd_epi16(
125 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
126 vacc3x0123 = _mm_maddd_epi16(
127 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
128
129 w = (const void*) ((uintptr_t) w + 32 * sizeof(int16_t));
130 k -= 8 * sizeof(int8_t);
131 }
132 if (k != 0) {
133 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
134 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
135 a0 = (const int8_t*) ((uintptr_t) a0 + k);
136 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
137 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
138 a1 = (const int8_t*) ((uintptr_t) a1 + k);
139 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
140 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
141 a2 = (const int8_t*) ((uintptr_t) a2 + k);
142 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
143 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
144 a3 = (const int8_t*) ((uintptr_t) a3 + k);
145
146 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
147 w = (const void*) ((uintptr_t) w + 8 * sizeof(int16_t));
148
149 vacc0x0123 = _mm_maddd_epi16(
150 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
151 vacc1x0123 = _mm_maddd_epi16(
152 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
153 vacc2x0123 = _mm_maddd_epi16(
154 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
155 vacc3x0123 = _mm_maddd_epi16(
156 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
157
158 if (k > 2 * sizeof(int8_t)) {
159 const __m128i vxb1 = _mm_load_si128((const __m128i*) w);
160 w = (const void*) ((uintptr_t) w + 8 * sizeof(int16_t));
161
162 vacc0x0123 = _mm_maddd_epi16(
163 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
164 vacc1x0123 = _mm_maddd_epi16(
165 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
166 vacc2x0123 = _mm_maddd_epi16(
167 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
168 vacc3x0123 = _mm_maddd_epi16(
169 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
170
171 if (k > 4 * sizeof(int8_t)) {
172 const __m128i vxb2 = _mm_load_si128((const __m128i*) w);
173 w = (const void*) ((uintptr_t) w + 8 * sizeof(int16_t));
174
175 vacc0x0123 = _mm_maddd_epi16(
176 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
177 vacc1x0123 = _mm_maddd_epi16(
178 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
179 vacc2x0123 = _mm_maddd_epi16(
180 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
181 vacc3x0123 = _mm_maddd_epi16(
182 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
183 }
184 }
185 }
186
187 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
188 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
189
190 const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
191 const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
192 const __m128i vacc2x1133 = _mm_shuffle_epi32(vacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));
193 const __m128i vacc3x1133 = _mm_shuffle_epi32(vacc3x0123, _MM_SHUFFLE(3, 3, 1, 1));
194
195 const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
196 const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
197 const __m128i vprod2x02 = _mm_add_epi64(_mm_mul_epi32(vacc2x0123, vmultiplier), vrounding);
198 const __m128i vprod3x02 = _mm_add_epi64(_mm_mul_epi32(vacc3x0123, vmultiplier), vrounding);
199
200 const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
201 const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
202 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding);
203 const __m128i vprod3x13 = _mm_add_epi64(_mm_mul_epi32(vacc3x1133, vmultiplier), vrounding);
204
205 const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
206 const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
207 const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
208 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
209 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31);
210 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13);
211 const __m128i vq31prod3x02 = _mm_srli_epi64(vprod3x02, 31);
212 const __m128i vq31prod3x13 = _mm_add_epi64(vprod3x13, vprod3x13);
213
214 const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
215 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
216 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC);
217 const __m128i vq31prod3x0123 = _mm_blend_epi16(vq31prod3x02, vq31prod3x13, 0xCC);
218
219 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
220 const __m128i vrem0x0123 =
221 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
222 const __m128i vrem1x0123 =
223 _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
224 const __m128i vrem2x0123 =
225 _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
226 const __m128i vrem3x0123 =
227 _mm_add_epi32(_mm_and_si128(vq31prod3x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod3x0123));
228
229 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
230 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
231 vacc0x0123 =
232 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
233 vacc1x0123 =
234 _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
235 vacc2x0123 =
236 _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
237 vacc3x0123 =
238 _mm_sub_epi32(_mm_sra_epi32(vq31prod3x0123, vshift), _mm_cmpgt_epi32(vrem3x0123, vremainder_threshold));
239
240 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
241 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
242 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
243
244 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
245 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
246 vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
247 vacc23x0123 = _mm_min_epi16(_mm_max_epi16(vacc23x0123, voutput_min), voutput_max);
248
249 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
250
251 if (nc >= 4) {
252 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
253 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
254 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
255 *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
256
257 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
258 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
259 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
260 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
261
262 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
263 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
264 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
265 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
266
267 nc -= 4;
268 } else {
269 if (nc & 2) {
270 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
271 c0 += 2;
272 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
273 c1 += 2;
274 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
275 c2 += 2;
276 *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
277 c3 += 2;
278 vout = _mm_srli_epi32(vout, 16);
279 }
280 if (nc & 1) {
281 *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
282 *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
283 *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
284 *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
285 }
286
287 nc = 0;
288 }
289 } while (nc != 0);
290 }
291