Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x4c8-minmax-sse41-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_4x4c8__sse41_ld64(
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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (4 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 const __m128 vmin = _mm_set1_ps(params->scalar.min);
61 const __m128 vmax = _mm_set1_ps(params->scalar.max);
62 XNN_FORCE_REALIZATION(vmin);
63 XNN_FORCE_REALIZATION(vmax);
64 __m128i vinput_zero_point = _mm_cvtsi32_si128(*((const int*) &quantization_params->zero_point));
65 vinput_zero_point = _mm_shuffle_epi32(vinput_zero_point, _MM_SHUFFLE(0, 0, 0, 0));
66 const __m128 vinput_scale = _mm_load1_ps(&quantization_params->inv_scale);
67
68
69 do {
70 const __m128i vksum = _mm_load_si128((const __m128i*) w);
71 const __m128i vzero = _mm_setzero_si128();
72 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point);
73 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point);
74 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point);
75 const __m128i vinit3 = _mm_mullo_epi32(vksum, vinput_zero_point);
76 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
77 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
78 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
79 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
80 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
81 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
82 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
83 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
84 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
85 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
86 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
87 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
88 __m128i vacc3x0 = _mm_blend_epi16(vinit3, vzero, 0xFC);
89 __m128i vacc3x1 = _mm_blend_epi16(vinit3, vzero, 0xF3);
90 __m128i vacc3x2 = _mm_blend_epi16(vinit3, vzero, 0xCF);
91 __m128i vacc3x3 = _mm_blend_epi16(vinit3, vzero, 0x3F);
92 w = (const int32_t*) w + 4;
93
94 size_t p = ks;
95 do {
96 const int8_t* restrict a0 = a[0];
97 if XNN_UNPREDICTABLE(a0 != zero) {
98 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
99 } else {
100 a0 = zero_data;
101 }
102 const int8_t* restrict a1 = a[1];
103 if XNN_UNPREDICTABLE(a1 != zero) {
104 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
105 } else {
106 a1 = zero_data;
107 }
108 const int8_t* restrict a2 = a[2];
109 if XNN_UNPREDICTABLE(a2 != zero) {
110 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
111 } else {
112 a2 = zero_data;
113 }
114 const int8_t* restrict a3 = a[3];
115 if XNN_UNPREDICTABLE(a3 != zero) {
116 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
117 } else {
118 a3 = zero_data;
119 }
120 a += 4;
121
122 size_t k = 0;
123 while (k < kc) {
124 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
125 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
126 a0 += 8;
127 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
128 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
129 a1 += 8;
130 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
131 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
132 a2 += 8;
133 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
134 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
135 a3 += 8;
136
137 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
138 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
139
140 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
141 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
142 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
143 vacc3x0 = _mm_add_epi32(vacc3x0, _mm_madd_epi16(vxa3, vxb0));
144 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
145 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
146
147 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
148 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
149 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
150 vacc3x1 = _mm_add_epi32(vacc3x1, _mm_madd_epi16(vxa3, vxb1));
151 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
152 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
153
154 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
155 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
156 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
157 vacc3x2 = _mm_add_epi32(vacc3x2, _mm_madd_epi16(vxa3, vxb2));
158 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
159 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
160
161 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
162 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
163 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
164 vacc3x3 = _mm_add_epi32(vacc3x3, _mm_madd_epi16(vxa3, vxb3));
165
166 w = (const void*) ((const int8_t*) w + 32);
167 k += 8 * sizeof(int8_t);
168 }
169 p -= 4 * sizeof(void*);
170 } while (p != 0);
171
172 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
173 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
174 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
175 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
176 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
177 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
178 const __m128i vacc3x01 = _mm_hadd_epi32(vacc3x0, vacc3x1);
179 const __m128i vacc3x23 = _mm_hadd_epi32(vacc3x2, vacc3x3);
180
181 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
182 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
183 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
184 __m128i vacc3x0123 = _mm_hadd_epi32(vacc3x01, vacc3x23);
185
186 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
187 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
188 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
189 __m128 vout3x0123 = _mm_cvtepi32_ps(vacc3x0123);
190
191 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale);
192 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale);
193 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale);
194 vout3x0123 = _mm_mul_ps(vout3x0123, vinput_scale);
195
196 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
197 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
198 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
199 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
200 vout3x0123 = _mm_mul_ps(vout3x0123, vfilter_output_scale0123);
201
202 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
203 w = (const float*) w + 8;
204 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
205 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
206 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
207 vout3x0123 = _mm_add_ps(vout3x0123, vbias0123);
208
209 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
210 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
211 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
212 vout3x0123 = _mm_max_ps(vout3x0123, vmin);
213
214 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
215 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
216 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
217 vout3x0123 = _mm_min_ps(vout3x0123, vmax);
218
219 if XNN_LIKELY(nc >= 4) {
220 _mm_storeu_ps(c3, vout3x0123);
221 _mm_storeu_ps(c2, vout2x0123);
222 _mm_storeu_ps(c1, vout1x0123);
223 _mm_storeu_ps(c0, vout0x0123);
224
225 a = (const int8_t**restrict) ((uintptr_t) a - ks);
226
227 c0 = (float*) ((uintptr_t) c0 + cn_stride);
228 c1 = (float*) ((uintptr_t) c1 + cn_stride);
229 c2 = (float*) ((uintptr_t) c2 + cn_stride);
230 c3 = (float*) ((uintptr_t) c3 + cn_stride);
231
232 nc -= 4;
233 } else {
234 if (nc & 2) {
235 _mm_storel_pi((__m64*) c3, vout3x0123);
236 vout3x0123 = _mm_unpackhi_ps(vout3x0123, vout3x0123);
237 c3 += 2;
238 _mm_storel_pi((__m64*) c2, vout2x0123);
239 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
240 c2 += 2;
241 _mm_storel_pi((__m64*) c1, vout1x0123);
242 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
243 c1 += 2;
244 _mm_storel_pi((__m64*) c0, vout0x0123);
245 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
246 c0 += 2;
247 }
248 if (nc & 1) {
249 _mm_store_ss(c3, vout3x0123);
250 _mm_store_ss(c2, vout2x0123);
251 _mm_store_ss(c1, vout1x0123);
252 _mm_store_ss(c0, vout0x0123);
253 }
254 nc = 0;
255 }
256 } while (nc != 0);
257 }