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_qu8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])25 void xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64(
26 size_t mr,
27 size_t nc,
28 size_t kc,
29 const uint8_t* restrict a,
30 size_t a_stride,
31 const void* restrict w,
32 uint8_t* restrict c,
33 size_t cm_stride,
34 size_t cn_stride,
35 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 4);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(uint8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
47 const uint8_t* a0 = a;
48 uint8_t* c0 = c;
49 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
50 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
56 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 2) {
58 a2 = a1;
59 c2 = c1;
60 }
61 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
62 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
63 if XNN_UNPREDICTABLE(mr != 4) {
64 a3 = a2;
65 c3 = c2;
66 }
67
68 do {
69 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
70 __m128i vacc1x0123 = vacc0x0123;
71 __m128i vacc2x0123 = vacc0x0123;
72 __m128i vacc3x0123 = vacc0x0123;
73 w = (const void*) ((const int32_t*) w + 4);
74
75 size_t k = kc;
76 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
77 while (k >= 8 * sizeof(uint8_t)) {
78 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
79 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
80 a0 += 8;
81 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
82 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
83 a1 += 8;
84 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
85 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
86 a2 += 8;
87 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
88 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
89 a3 += 8;
90
91 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
92 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
93
94 vacc0x0123 = _mm_maddd_epi16(
95 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
96 vacc1x0123 = _mm_maddd_epi16(
97 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
98 vacc2x0123 = _mm_maddd_epi16(
99 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
100 vacc3x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
102 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
103 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
104
105 vacc0x0123 = _mm_maddd_epi16(
106 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
107 vacc1x0123 = _mm_maddd_epi16(
108 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
109 vacc2x0123 = _mm_maddd_epi16(
110 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
111 vacc3x0123 = _mm_maddd_epi16(
112 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
113 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
114 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
115
116 vacc0x0123 = _mm_maddd_epi16(
117 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
118 vacc1x0123 = _mm_maddd_epi16(
119 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
120 vacc2x0123 = _mm_maddd_epi16(
121 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
122 vacc3x0123 = _mm_maddd_epi16(
123 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
124 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
125 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
126
127 vacc0x0123 = _mm_maddd_epi16(
128 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
129 vacc1x0123 = _mm_maddd_epi16(
130 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
131 vacc2x0123 = _mm_maddd_epi16(
132 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
133 vacc3x0123 = _mm_maddd_epi16(
134 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
135
136 w = (const void*) ((const uint8_t*) w + 32);
137 k -= 8 * sizeof(uint8_t);
138 }
139 if (k != 0) {
140 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
141 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
142 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
143 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
144 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
145 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
146 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
147 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
148 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
149 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
150 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
151 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
152
153 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
154 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
155 w = (const void*) ((const uint8_t*) w + 8);
156
157 vacc0x0123 = _mm_maddd_epi16(
158 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
159 vacc1x0123 = _mm_maddd_epi16(
160 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
161 vacc2x0123 = _mm_maddd_epi16(
162 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
163 vacc3x0123 = _mm_maddd_epi16(
164 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
165
166 if (k > 2 * sizeof(uint8_t)) {
167 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
168 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
169 w = (const void*) ((const uint8_t*) w + 8);
170
171 vacc0x0123 = _mm_maddd_epi16(
172 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
173 vacc1x0123 = _mm_maddd_epi16(
174 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
175 vacc2x0123 = _mm_maddd_epi16(
176 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
177 vacc3x0123 = _mm_maddd_epi16(
178 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
179
180 if (k > 4 * sizeof(uint8_t)) {
181 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
182 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
183 w = (const void*) ((const uint8_t*) w + 8);
184
185 vacc0x0123 = _mm_maddd_epi16(
186 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
187 vacc1x0123 = _mm_maddd_epi16(
188 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
189 vacc2x0123 = _mm_maddd_epi16(
190 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
191 vacc3x0123 = _mm_maddd_epi16(
192 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
193 }
194 }
195 }
196
197 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
198 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
199 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
200 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
201
202 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
203 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
204 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
205 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
206 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
207
208 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
209 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
210 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
211 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
212 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
213
214 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
215 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
216 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
217 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
218
219 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
220 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
221 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
222
223 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
224
225 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
226
227 if (nc >= 4) {
228 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
229 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
230 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
231 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
232
233 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
234 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
235 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
236 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
237
238 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
239 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
240 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
241 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
242
243 nc -= 4;
244 } else {
245 if (nc & 2) {
246 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
247 c0 += 2;
248 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
249 c1 += 2;
250 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
251 c2 += 2;
252 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
253 c3 += 2;
254 vout = _mm_srli_epi32(vout, 16);
255 }
256 if (nc & 1) {
257 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
258 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
259 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
260 *c3 = (uint8_t) _mm_extract_epi8(vout, 12);
261 }
262
263 nc = 0;
264 }
265 } while (nc != 0);
266 }
267