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
17
xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2__sse41_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2__sse41_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2);
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 do {
59 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
60 __m128i vacc1x0123 = vacc0x0123;
61 __m128i vacc2x0123 = vacc0x0123;
62 __m128i vacc3x0123 = vacc0x0123;
63 w = (const void*) ((const int32_t*) w + 4);
64
65 size_t p = ks;
66 do {
67 const int8_t* restrict a0 = a[0];
68 if XNN_UNPREDICTABLE(a0 != zero) {
69 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
70 }
71 const int8_t* restrict a1 = a[1];
72 if XNN_UNPREDICTABLE(a1 != zero) {
73 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
74 }
75 const int8_t* restrict a2 = a[2];
76 if XNN_UNPREDICTABLE(a2 != zero) {
77 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
78 }
79 const int8_t* restrict a3 = a[3];
80 if XNN_UNPREDICTABLE(a3 != zero) {
81 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
82 }
83 a += 4;
84
85 size_t k = kc;
86 while (k >= 8 * sizeof(int8_t)) {
87 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
88 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
89 a0 += 8;
90 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
91 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
92 a1 += 8;
93 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
94 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
95 a2 += 8;
96 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
97 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
98 a3 += 8;
99
100 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
101 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
102 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
103
104 vacc0x0123 = _mm_add_epi32(vacc0x0123,
105 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
106 vacc1x0123 = _mm_add_epi32(vacc1x0123,
107 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
108 vacc2x0123 = _mm_add_epi32(vacc2x0123,
109 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
110 vacc3x0123 = _mm_add_epi32(vacc3x0123,
111 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
112
113 vacc0x0123 = _mm_add_epi32(vacc0x0123,
114 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
115 vacc1x0123 = _mm_add_epi32(vacc1x0123,
116 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
117 vacc2x0123 = _mm_add_epi32(vacc2x0123,
118 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
119 vacc3x0123 = _mm_add_epi32(vacc3x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
121 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
122 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
123 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
124
125 vacc0x0123 = _mm_add_epi32(vacc0x0123,
126 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
127 vacc1x0123 = _mm_add_epi32(vacc1x0123,
128 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
129 vacc2x0123 = _mm_add_epi32(vacc2x0123,
130 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
131 vacc3x0123 = _mm_add_epi32(vacc3x0123,
132 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
133
134 vacc0x0123 = _mm_add_epi32(vacc0x0123,
135 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
136 vacc1x0123 = _mm_add_epi32(vacc1x0123,
137 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
138 vacc2x0123 = _mm_add_epi32(vacc2x0123,
139 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
140 vacc3x0123 = _mm_add_epi32(vacc3x0123,
141 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
142
143 w = (const void*) ((const int8_t*) w + 32);
144 k -= 8 * sizeof(int8_t);
145 }
146 if (k != 0) {
147 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
148 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
149 a0 = (const int8_t*) ((uintptr_t) a0 + k);
150 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
151 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
152 a1 = (const int8_t*) ((uintptr_t) a1 + k);
153 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
154 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
155 a2 = (const int8_t*) ((uintptr_t) a2 + k);
156 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
157 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
158 a3 = (const int8_t*) ((uintptr_t) a3 + k);
159
160 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
161 w = (const void*) ((const int8_t*) w + 8);
162 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
163
164 vacc0x0123 = _mm_add_epi32(vacc0x0123,
165 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
166 vacc1x0123 = _mm_add_epi32(vacc1x0123,
167 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
168 vacc2x0123 = _mm_add_epi32(vacc2x0123,
169 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
170 vacc3x0123 = _mm_add_epi32(vacc3x0123,
171 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
172
173 if (k > 2 * sizeof(int8_t)) {
174 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
175 w = (const void*) ((const int8_t*) w + 8);
176 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
177
178 vacc0x0123 = _mm_add_epi32(vacc0x0123,
179 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
180 vacc1x0123 = _mm_add_epi32(vacc1x0123,
181 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
182 vacc2x0123 = _mm_add_epi32(vacc2x0123,
183 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
184 vacc3x0123 = _mm_add_epi32(vacc3x0123,
185 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
186
187 if (k > 4 * sizeof(int8_t)) {
188 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
189 w = (const void*) ((const int8_t*) w + 8);
190 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
191
192 vacc0x0123 = _mm_add_epi32(vacc0x0123,
193 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
194 vacc1x0123 = _mm_add_epi32(vacc1x0123,
195 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
196 vacc2x0123 = _mm_add_epi32(vacc2x0123,
197 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
198 vacc3x0123 = _mm_add_epi32(vacc3x0123,
199 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
200 }
201 }
202 }
203 p -= 4 * sizeof(void*);
204 } while (p != 0);
205
206 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
207 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
208 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
209 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
210
211 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
212 w = (const void*) ((const float*) w + 4);
213 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
214 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
215 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
216 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
217
218 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
219 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
220 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
221 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
222 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
223
224 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
225 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
226 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
227 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
228
229 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
230 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
231 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
232
233
234 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
235
236 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
237
238 if (nc >= 4) {
239 *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
240 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
241 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
242 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
243 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
244 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
245 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
246 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
247
248 a = (const int8_t**restrict) ((uintptr_t) a - ks);
249
250 nc -= 4;
251 } else {
252 if (nc & 2) {
253 *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
254 c3 += 2;
255 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
256 c2 += 2;
257 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
258 c1 += 2;
259 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
260 c0 += 2;
261 vout = _mm_srli_epi32(vout, 16);
262 }
263 if (nc & 1) {
264 *c3 = (int8_t) _mm_extract_epi8(vout, 12);
265 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
266 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
267 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
268 }
269
270 nc = 0;
271 }
272 } while (nc != 0);
273 }
274