Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x4c2-minmax-fp32-sse2-ld128.c

raw
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 <emmintrin.h>
13
14 #include "xnnpack/igemm.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x4c2__sse2_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_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(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
59 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
60 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
61 const __m128i voutput_min = _mm_set1_epi16(params->fp32_scalar.output_min);
62 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
63 XNN_FORCE_REALIZATION(voutput_zero_point);
64 XNN_FORCE_REALIZATION(voutput_min);
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*) ((const int32_t*) w + 4);
72
73 size_t p = ks;
74 do {
75 const int8_t* restrict a0 = a[0];
76 if XNN_UNPREDICTABLE(a0 != zero) {
77 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
78 }
79 const int8_t* restrict a1 = a[1];
80 if XNN_UNPREDICTABLE(a1 != zero) {
81 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
82 }
83 const int8_t* restrict a2 = a[2];
84 if XNN_UNPREDICTABLE(a2 != zero) {
85 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
86 }
87 const int8_t* restrict a3 = a[3];
88 if XNN_UNPREDICTABLE(a3 != zero) {
89 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
90 }
91 a += 4;
92
93 size_t k = kc;
94 while (k >= 8 * sizeof(int8_t)) {
95 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
96 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
97 a0 += 8;
98 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
99 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
100 a1 += 8;
101 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
102 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
103 a2 += 8;
104 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
105 const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
106 a3 += 8;
107
108 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
109 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
110 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
111 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
112
113 vacc0x0123 = _mm_add_epi32(vacc0x0123,
114 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
115 vacc1x0123 = _mm_add_epi32(vacc1x0123,
116 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
117 vacc2x0123 = _mm_add_epi32(vacc2x0123,
118 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
119 vacc3x0123 = _mm_add_epi32(vacc3x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
121
122 vacc0x0123 = _mm_add_epi32(vacc0x0123,
123 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
124 vacc1x0123 = _mm_add_epi32(vacc1x0123,
125 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
126 vacc2x0123 = _mm_add_epi32(vacc2x0123,
127 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
128 vacc3x0123 = _mm_add_epi32(vacc3x0123,
129 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
130 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
131 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
132 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
133 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
134
135 vacc0x0123 = _mm_add_epi32(vacc0x0123,
136 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
137 vacc1x0123 = _mm_add_epi32(vacc1x0123,
138 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
139 vacc2x0123 = _mm_add_epi32(vacc2x0123,
140 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
141 vacc3x0123 = _mm_add_epi32(vacc3x0123,
142 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
143
144 vacc0x0123 = _mm_add_epi32(vacc0x0123,
145 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
146 vacc1x0123 = _mm_add_epi32(vacc1x0123,
147 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
148 vacc2x0123 = _mm_add_epi32(vacc2x0123,
149 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
150 vacc3x0123 = _mm_add_epi32(vacc3x0123,
151 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
152
153 w = (const void*) ((const int8_t*) w + 32);
154 k -= 8 * sizeof(int8_t);
155 }
156 if (k != 0) {
157 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
158 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
159 a0 = (const int8_t*) ((uintptr_t) a0 + k);
160 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
161 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
162 a1 = (const int8_t*) ((uintptr_t) a1 + k);
163 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
164 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
165 a2 = (const int8_t*) ((uintptr_t) a2 + k);
166 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
167 const __m128i vxa3 = _mm_srai_epi16(_mm_unpacklo_epi8(va3, va3), 8);
168 a3 = (const int8_t*) ((uintptr_t) a3 + k);
169
170 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
171 w = (const int8_t*) w + 8;
172 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
173
174 vacc0x0123 = _mm_add_epi32(vacc0x0123,
175 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
176 vacc1x0123 = _mm_add_epi32(vacc1x0123,
177 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
178 vacc2x0123 = _mm_add_epi32(vacc2x0123,
179 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
180 vacc3x0123 = _mm_add_epi32(vacc3x0123,
181 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
182
183 if (k > 2 * sizeof(int8_t)) {
184 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
185 w = (const int8_t*) w + 8;
186 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
187
188 vacc0x0123 = _mm_add_epi32(vacc0x0123,
189 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
190 vacc1x0123 = _mm_add_epi32(vacc1x0123,
191 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
192 vacc2x0123 = _mm_add_epi32(vacc2x0123,
193 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
194 vacc3x0123 = _mm_add_epi32(vacc3x0123,
195 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
196
197 if (k > 4 * sizeof(int8_t)) {
198 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
199 w = (const int8_t*) w + 8;
200 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
201
202 vacc0x0123 = _mm_add_epi32(vacc0x0123,
203 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
204 vacc1x0123 = _mm_add_epi32(vacc1x0123,
205 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
206 vacc2x0123 = _mm_add_epi32(vacc2x0123,
207 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
208 vacc3x0123 = _mm_add_epi32(vacc3x0123,
209 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
210 }
211 }
212 }
213 p -= 4 * sizeof(void*);
214 } while (p != 0);
215
216 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
217 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
218 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
219 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
220
221 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
222 w = (const float*) w + 4;
223 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
224 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
225 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
226 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
227
228 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
229 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
230 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
231 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
232
233 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
234 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
235 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
236 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
237
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 vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
242 vacc23x0123 = _mm_max_epi16(vacc23x0123, voutput_min);
243
244 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
245
246
247 if (nc >= 4) {
248 unaligned_store_u32(c3, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(3, 3, 3, 3))));
249 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
250 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2))));
251 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
252 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1))));
253 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
254 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
255 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
256
257 a = (const int8_t**restrict) ((uintptr_t) a - ks);
258
259 nc -= 4;
260 } else {
261 if (nc & 2) {
262 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
263 c3 += 2;
264 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
265 c2 += 2;
266 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
267 c1 += 2;
268 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
269 c0 += 2;
270 vout = _mm_srli_epi32(vout, 16);
271 }
272 if (nc & 1) {
273 *c3 = (int8_t) _mm_extract_epi16(vout, 6);
274 *c2 = (int8_t) _mm_extract_epi16(vout, 4);
275 *c1 = (int8_t) _mm_extract_epi16(vout, 2);
276 *c0 = (int8_t) _mm_cvtsi128_si32(vout);
277 }
278
279 nc = 0;
280 }
281 } while (nc != 0);
282 }