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 #include <smmintrin.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__sse41_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)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__sse41_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const uint8_t** restrict a,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const uint8_t* zero,
31 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 4);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (4 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(uint8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
45 uint8_t* c0 = c;
46 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
59 do {
60 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
61 __m128i vacc1x0123 = vacc0x0123;
62 __m128i vacc2x0123 = vacc0x0123;
63 __m128i vacc3x0123 = vacc0x0123;
64 w = (const void*) ((const int32_t*) w + 4);
65
66 size_t p = ks;
67 do {
68 const uint8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
71 }
72 const uint8_t* restrict a1 = a[1];
73 if XNN_UNPREDICTABLE(a1 != zero) {
74 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
75 }
76 const uint8_t* restrict a2 = a[2];
77 if XNN_UNPREDICTABLE(a2 != zero) {
78 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
79 }
80 const uint8_t* restrict a3 = a[3];
81 if XNN_UNPREDICTABLE(a3 != zero) {
82 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
83 }
84 a += 4;
85
86 size_t k = kc;
87 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
88 while (k >= 8 * sizeof(uint8_t)) {
89 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
90 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
91 a0 += 8;
92 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
93 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
94 a1 += 8;
95 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
96 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
97 a2 += 8;
98 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
99 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
100 a3 += 8;
101
102 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
103 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
104
105 vacc0x0123 = _mm_add_epi32(vacc0x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
107 vacc1x0123 = _mm_add_epi32(vacc1x0123,
108 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
109 vacc2x0123 = _mm_add_epi32(vacc2x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
111 vacc3x0123 = _mm_add_epi32(vacc3x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
113 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
114 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
115
116 vacc0x0123 = _mm_add_epi32(vacc0x0123,
117 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
118 vacc1x0123 = _mm_add_epi32(vacc1x0123,
119 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
120 vacc2x0123 = _mm_add_epi32(vacc2x0123,
121 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
122 vacc3x0123 = _mm_add_epi32(vacc3x0123,
123 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
124 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
125 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
126
127 vacc0x0123 = _mm_add_epi32(vacc0x0123,
128 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
129 vacc1x0123 = _mm_add_epi32(vacc1x0123,
130 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
131 vacc2x0123 = _mm_add_epi32(vacc2x0123,
132 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
133 vacc3x0123 = _mm_add_epi32(vacc3x0123,
134 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
135 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
136 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
137
138 vacc0x0123 = _mm_add_epi32(vacc0x0123,
139 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
140 vacc1x0123 = _mm_add_epi32(vacc1x0123,
141 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
142 vacc2x0123 = _mm_add_epi32(vacc2x0123,
143 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
144 vacc3x0123 = _mm_add_epi32(vacc3x0123,
145 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
146
147 w = (const void*) ((const uint8_t*) w + 32);
148 k -= 8 * sizeof(uint8_t);
149 }
150 if (k != 0) {
151 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
152 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
153 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
154 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
155 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
156 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
157 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
158 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
159 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
160 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
161 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
162 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
163
164 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
165 w = (const void*) ((const uint8_t*) w + 8);
166 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
167
168 vacc0x0123 = _mm_add_epi32(vacc0x0123,
169 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
170 vacc1x0123 = _mm_add_epi32(vacc1x0123,
171 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
172 vacc2x0123 = _mm_add_epi32(vacc2x0123,
173 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
174 vacc3x0123 = _mm_add_epi32(vacc3x0123,
175 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
176
177 if (k > 2 * sizeof(uint8_t)) {
178 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
179 w = (const void*) ((const uint8_t*) w + 8);
180 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
181
182 vacc0x0123 = _mm_add_epi32(vacc0x0123,
183 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
184 vacc1x0123 = _mm_add_epi32(vacc1x0123,
185 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
186 vacc2x0123 = _mm_add_epi32(vacc2x0123,
187 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
188 vacc3x0123 = _mm_add_epi32(vacc3x0123,
189 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
190
191 if (k > 4 * sizeof(uint8_t)) {
192 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
193 w = (const void*) ((const uint8_t*) w + 8);
194 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
195
196 vacc0x0123 = _mm_add_epi32(vacc0x0123,
197 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
198 vacc1x0123 = _mm_add_epi32(vacc1x0123,
199 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
200 vacc2x0123 = _mm_add_epi32(vacc2x0123,
201 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
202 vacc3x0123 = _mm_add_epi32(vacc3x0123,
203 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
204 }
205 }
206 }
207 p -= 4 * sizeof(void*);
208 } while (p != 0);
209
210 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
211 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
212 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
213 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
214
215 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
216 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
217 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
218 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
219 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
220
221 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
222 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
223 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
224 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
225 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
226
227 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
228 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
229 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
230 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
231
232 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
233 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
234 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
235
236 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
237
238 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
239
240 if (nc >= 4) {
241 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
242 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
243 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
244 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
245 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
246 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
247 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
248 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
249
250 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
251
252 nc -= 4;
253 } else {
254 if (nc & 2) {
255 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
256 c3 += 2;
257 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
258 c2 += 2;
259 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
260 c1 += 2;
261 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
262 c0 += 2;
263 vout = _mm_srli_epi32(vout, 16);
264 }
265 if (nc & 1) {
266 *c3 = (uint8_t) _mm_extract_epi8(vout, 12);
267 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
268 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
269 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
270 }
271
272 nc = 0;
273 }
274 } while (nc != 0);
275 }
276