Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-4x4c8-minmax-sse41-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_4x4c8__sse41_ld128(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_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(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
58 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr != 4) {
60 a3 = a2;
61 c3 = c2;
62 }
63
64 const __m128 vmin = _mm_set1_ps(params->scalar.min);
65 const __m128 vmax = _mm_set1_ps(params->scalar.max);
66 XNN_FORCE_REALIZATION(vmin);
67 XNN_FORCE_REALIZATION(vmax);
68
69 const __m128i vmask = _mm_set1_epi8(0xF0);
70 XNN_FORCE_REALIZATION(vmask);
71
72 do {
73 const __m128i vksum = _mm_load_si128((const __m128i*) w);
74 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
75 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
76 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
77 const __m128i vinput_zero_point23 = _mm_loadu_si128((const __m128i*) &quantization_params[2]);
78 const __m128i vinput_zero_point2 = _mm_shuffle_epi32(vinput_zero_point23, _MM_SHUFFLE(0, 0, 0, 0));
79 const __m128i vinput_zero_point3 = _mm_shuffle_epi32(vinput_zero_point23, _MM_SHUFFLE(2, 2, 2, 2));
80 const __m128i vzero = _mm_setzero_si128();
81 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
82 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
83 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point2);
84 const __m128i vinit3 = _mm_mullo_epi32(vksum, vinput_zero_point3);
85 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
86 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
87 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
88 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
89 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
90 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
91 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
92 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
93 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
94 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
95 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
96 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
97 __m128i vacc3x0 = _mm_blend_epi16(vinit3, vzero, 0xFC);
98 __m128i vacc3x1 = _mm_blend_epi16(vinit3, vzero, 0xF3);
99 __m128i vacc3x2 = _mm_blend_epi16(vinit3, vzero, 0xCF);
100 __m128i vacc3x3 = _mm_blend_epi16(vinit3, vzero, 0x3F);
101 w = (const int32_t*) w + 4;
102
103 size_t k = kc;
104
105 while (k >= 16 * sizeof(int8_t)) {
106 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
107 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
108 a0 += 8;
109 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
110 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
111 a1 += 8;
112 const __m128i va2c0 = _mm_loadl_epi64((const __m128i*) a2);
113 const __m128i vxa2c0 = _mm_cvtepi8_epi16(va2c0);
114 a2 += 8;
115 const __m128i va3c0 = _mm_loadl_epi64((const __m128i*) a3);
116 const __m128i vxa3c0 = _mm_cvtepi8_epi16(va3c0);
117 a3 += 8;
118
119 const __m128i vb01c01 = _mm_load_si128((const __m128i*) w);
120 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
121 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
122 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
123 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
124 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
125
126 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
127 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
128 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
129 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
130 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c0, vxb0c0));
131 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c0, vxb1c0));
132 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c0, vxb0c0));
133 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c0, vxb1c0));
134 const __m128i vb23c01 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
135 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
136 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
137 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
138 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
139 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
140
141 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
142 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
143 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
144 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
145 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c0, vxb2c0));
146 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c0, vxb3c0));
147 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c0, vxb2c0));
148 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c0, vxb3c0));
149
150 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
151 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
152 a0 += 8;
153 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
154 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
155 a1 += 8;
156 const __m128i va2c1 = _mm_loadl_epi64((const __m128i*) a2);
157 const __m128i vxa2c1 = _mm_cvtepi8_epi16(va2c1);
158 a2 += 8;
159 const __m128i va3c1 = _mm_loadl_epi64((const __m128i*) a3);
160 const __m128i vxa3c1 = _mm_cvtepi8_epi16(va3c1);
161 a3 += 8;
162
163 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
164 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
165 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
166 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
167
168 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
169 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
170 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
171 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
172 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2c1, vxb0c1));
173 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2c1, vxb1c1));
174 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3c1, vxb0c1));
175 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3c1, vxb1c1));
176 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
177 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
178 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
179 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
180
181 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
182 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
183 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
184 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
185 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2c1, vxb2c1));
186 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2c1, vxb3c1));
187 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3c1, vxb2c1));
188 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3c1, vxb3c1));
189
190
191 w = (const int8_t*) w + 32;
192 k -= 16 * sizeof(int8_t);
193 }
194
195 while (k >= 8 * sizeof(int8_t)) {
196 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
197 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
198 a0 += 8;
199 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
200 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
201 a1 += 8;
202 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
203 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
204 a2 += 8;
205 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
206 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
207 a3 += 8;
208
209 __m128i vb01 = _mm_load_si128((const __m128i*) w);
210 vb01 = _mm_slli_epi32(vb01, 4);
211 vb01 = _mm_and_si128(vb01, vmask);
212
213 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
214 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
215 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
216
217 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
218 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
219 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
220 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
221 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
222 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
223 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3, vxb0));
224 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3, vxb1));
225 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
226 vb23 = _mm_slli_epi32(vb23, 4);
227 vb23 = _mm_and_si128(vb23, vmask);
228
229 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
230 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
231 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
232
233 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
234 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
235 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
236 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
237 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
238 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
239 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3, vxb2));
240 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3, vxb3));
241
242 w = (const int8_t*) w + 32;
243 k -= 8 * sizeof(int8_t);
244 }
245
246 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
247 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
248 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
249 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
250 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
251 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
252 const __m128i vacc3x01 = _mm_hadd_epi32(vacc3x0, vacc3x1);
253 const __m128i vacc3x23 = _mm_hadd_epi32(vacc3x2, vacc3x3);
254
255 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
256 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
257 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
258 __m128i vacc3x0123 = _mm_hadd_epi32(vacc3x01, vacc3x23);
259
260 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
261 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
262 vacc2x0123 = _mm_srai_epi32(vacc2x0123, 4);
263 vacc3x0123 = _mm_srai_epi32(vacc3x0123, 4);
264 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
265 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
266 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
267 __m128 vout3x0123 = _mm_cvtepi32_ps(vacc3x0123);
268
269 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
270 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
271 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
272 const __m128i vinput_scale23 = _mm_loadu_si128((const __m128i*) &quantization_params[2]);
273 const __m128 vinput_scale2 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale23, _MM_SHUFFLE(1, 1, 1, 1)));
274 const __m128 vinput_scale3 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale23, _MM_SHUFFLE(3, 3, 3, 3)));
275
276 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
277 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
278 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
279 vout3x0123 = _mm_mul_ps(vout3x0123, vinput_scale3);
280
281 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
282 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
283 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
284 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
285 vout3x0123 = _mm_mul_ps(vout3x0123, vfilter_output_scale0123);
286
287 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
288 w = (const float*) w + 8;
289 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
290 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
291 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
292 vout3x0123 = _mm_add_ps(vout3x0123, vbias0123);
293
294 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
295 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
296 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
297 vout3x0123 = _mm_max_ps(vout3x0123, vmin);
298
299 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
300 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
301 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
302 vout3x0123 = _mm_min_ps(vout3x0123, vmax);
303
304 if XNN_LIKELY(nc >= 4) {
305 _mm_storeu_ps(c0, vout0x0123);
306 _mm_storeu_ps(c1, vout1x0123);
307 _mm_storeu_ps(c2, vout2x0123);
308 _mm_storeu_ps(c3, vout3x0123);
309
310 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
311 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
312 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
313 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
314
315 c0 = (float*) ((uintptr_t) c0 + cn_stride);
316 c1 = (float*) ((uintptr_t) c1 + cn_stride);
317 c2 = (float*) ((uintptr_t) c2 + cn_stride);
318 c3 = (float*) ((uintptr_t) c3 + cn_stride);
319
320 nc -= 4;
321 } else {
322 if (nc & 2) {
323 _mm_storel_pi((__m64*) c0, vout0x0123);
324 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
325 c0 += 2;
326 _mm_storel_pi((__m64*) c1, vout1x0123);
327 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
328 c1 += 2;
329 _mm_storel_pi((__m64*) c2, vout2x0123);
330 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
331 c2 += 2;
332 _mm_storel_pi((__m64*) c3, vout3x0123);
333 vout3x0123 = _mm_unpackhi_ps(vout3x0123, vout3x0123);
334 c3 += 2;
335 }
336 if (nc & 1) {
337 _mm_store_ss(c0, vout0x0123);
338 _mm_store_ss(c1, vout1x0123);
339 _mm_store_ss(c2, vout2x0123);
340 _mm_store_ss(c3, vout3x0123);
341 }
342 nc = 0;
343 }
344 } while (nc != 0);
345 }