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 #include <xnnpack/unaligned.h>
22
23
xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__xop_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__xop_ld64(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 size_t ks,
29 const uint8_t** restrict a,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 size_t a_offset,
35 const uint8_t* zero,
36 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 4);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (4 * sizeof(void*)) == 0);
44 assert(a_offset % sizeof(uint8_t) == 0);
45 assert(a != NULL);
46 assert(w != NULL);
47 assert(c != NULL);
48
49 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
50 uint8_t* c0 = c;
51 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr < 2) {
53 c1 = c0;
54 }
55 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 c2 = c1;
58 }
59 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr != 4) {
61 c3 = c2;
62 }
63
64 do {
65 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
66 __m128i vacc1x0123 = vacc0x0123;
67 __m128i vacc2x0123 = vacc0x0123;
68 __m128i vacc3x0123 = vacc0x0123;
69 w = (const void*) ((const int32_t*) w + 4);
70
71 size_t p = ks;
72 do {
73 const uint8_t* restrict a0 = a[0];
74 if XNN_UNPREDICTABLE(a0 != zero) {
75 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
76 }
77 const uint8_t* restrict a1 = a[1];
78 if XNN_UNPREDICTABLE(a1 != zero) {
79 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
80 }
81 const uint8_t* restrict a2 = a[2];
82 if XNN_UNPREDICTABLE(a2 != zero) {
83 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
84 }
85 const uint8_t* restrict a3 = a[3];
86 if XNN_UNPREDICTABLE(a3 != zero) {
87 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
88 }
89 a += 4;
90
91 size_t k = kc;
92 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
93 while (k >= 8 * sizeof(uint8_t)) {
94 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
95 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
96 a0 += 8;
97 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
98 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
99 a1 += 8;
100 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
101 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
102 a2 += 8;
103 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
104 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
105 a3 += 8;
106
107 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
108 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
109
110 vacc0x0123 = _mm_maddd_epi16(
111 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
112 vacc1x0123 = _mm_maddd_epi16(
113 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
114 vacc2x0123 = _mm_maddd_epi16(
115 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
116 vacc3x0123 = _mm_maddd_epi16(
117 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
118 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
119 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
120
121 vacc0x0123 = _mm_maddd_epi16(
122 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
123 vacc1x0123 = _mm_maddd_epi16(
124 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
125 vacc2x0123 = _mm_maddd_epi16(
126 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
127 vacc3x0123 = _mm_maddd_epi16(
128 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
129 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
130 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
131
132 vacc0x0123 = _mm_maddd_epi16(
133 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
134 vacc1x0123 = _mm_maddd_epi16(
135 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
136 vacc2x0123 = _mm_maddd_epi16(
137 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
138 vacc3x0123 = _mm_maddd_epi16(
139 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
140 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
141 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
142
143 vacc0x0123 = _mm_maddd_epi16(
144 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
145 vacc1x0123 = _mm_maddd_epi16(
146 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
147 vacc2x0123 = _mm_maddd_epi16(
148 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
149 vacc3x0123 = _mm_maddd_epi16(
150 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
151
152 w = (const void*) ((const uint8_t*) w + 32);
153 k -= 8 * sizeof(uint8_t);
154 }
155 if (k != 0) {
156 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
157 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
158 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
159 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
160 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
161 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
162 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
163 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
164 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
165 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
166 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
167 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
168
169 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
170 w = (const void*) ((const uint8_t*) w + 8);
171 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
172
173 vacc0x0123 = _mm_maddd_epi16(
174 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
175 vacc1x0123 = _mm_maddd_epi16(
176 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
177 vacc2x0123 = _mm_maddd_epi16(
178 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
179 vacc3x0123 = _mm_maddd_epi16(
180 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
181
182 if (k > 2 * sizeof(uint8_t)) {
183 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
184 w = (const void*) ((const uint8_t*) w + 8);
185 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
186
187 vacc0x0123 = _mm_maddd_epi16(
188 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
189 vacc1x0123 = _mm_maddd_epi16(
190 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
191 vacc2x0123 = _mm_maddd_epi16(
192 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
193 vacc3x0123 = _mm_maddd_epi16(
194 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
195
196 if (k > 4 * sizeof(uint8_t)) {
197 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
198 w = (const void*) ((const uint8_t*) w + 8);
199 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
200
201 vacc0x0123 = _mm_maddd_epi16(
202 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
203 vacc1x0123 = _mm_maddd_epi16(
204 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
205 vacc2x0123 = _mm_maddd_epi16(
206 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
207 vacc3x0123 = _mm_maddd_epi16(
208 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
209 }
210 }
211 }
212 p -= 4 * sizeof(void*);
213 } while (p != 0);
214
215 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
216 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
217 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
218 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
219
220 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
221 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
222 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
223 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
224 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
225
226 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
227 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
228 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
229 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
230 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
231
232 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
233 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
234 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
235 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
236
237 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
238 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
239 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
240
241 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
242
243 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
244
245 if (nc >= 4) {
246 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
247 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
248 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
249 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
250 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
251 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
252 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
253 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
254
255 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
256
257 nc -= 4;
258 } else {
259 if (nc & 2) {
260 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
261 c3 += 2;
262 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
263 c2 += 2;
264 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
265 c1 += 2;
266 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
267 c0 += 2;
268 vout = _mm_srli_epi32(vout, 16);
269 }
270 if (nc & 1) {
271 *c3 = (uint8_t) _mm_extract_epi8(vout, 12);
272 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
273 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
274 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
275 }
276
277 nc = 0;
278 }
279 } while (nc != 0);
280 }
281