Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x4c8-minmax-sse2-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 <emmintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_3x4c8__sse2_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 float* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56
57 const __m128 vmin = _mm_set1_ps(params->scalar.min);
58 const __m128 vmax = _mm_set1_ps(params->scalar.max);
59 XNN_FORCE_REALIZATION(vmin);
60 XNN_FORCE_REALIZATION(vmax);
61
62
63 do {
64 const __m128i vksum = _mm_load_si128((const __m128i*) w);
65 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
66 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
67 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
68 __m128i vinput_zero_point2 = _mm_cvtsi32_si128(*((const int*) &quantization_params[2].zero_point));
69 vinput_zero_point2 = _mm_shuffle_epi32(vinput_zero_point2, _MM_SHUFFLE(0, 0, 0, 0));
70 const __m128i vzero = _mm_setzero_si128();
71 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
72 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
73
74 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point0, vksum_lo);
75 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point0, vksum_lo);
76 __m128i vzpprodksumhi1 = _mm_mulhi_epu16(vinput_zero_point1, vksum_lo);
77 const __m128i vzpprodksumlo1 = _mm_mullo_epi16(vinput_zero_point1, vksum_lo);
78 __m128i vzpprodksumhi2 = _mm_mulhi_epu16(vinput_zero_point2, vksum_lo);
79 const __m128i vzpprodksumlo2 = _mm_mullo_epi16(vinput_zero_point2, vksum_lo);
80
81 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point0, vksum_hi));
82 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point0, 15), vksum_lo));
83 vzpprodksumhi1 = _mm_add_epi16(vzpprodksumhi1, _mm_mullo_epi16(vinput_zero_point1, vksum_hi));
84 vzpprodksumhi1 = _mm_sub_epi16(vzpprodksumhi1, _mm_and_si128(_mm_srai_epi16(vinput_zero_point1, 15), vksum_lo));
85 vzpprodksumhi2 = _mm_add_epi16(vzpprodksumhi2, _mm_mullo_epi16(vinput_zero_point2, vksum_hi));
86 vzpprodksumhi2 = _mm_sub_epi16(vzpprodksumhi2, _mm_and_si128(_mm_srai_epi16(vinput_zero_point2, 15), vksum_lo));
87
88 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
89 vzpprodksumhi1 = _mm_slli_si128(vzpprodksumhi1, 2);
90 vzpprodksumhi2 = _mm_slli_si128(vzpprodksumhi2, 2);
91
92 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
93 const __m128i vksumzp1 = _mm_or_si128(vzpprodksumhi1, vzpprodksumlo1);
94 const __m128i vksumzp2 = _mm_or_si128(vzpprodksumhi2, vzpprodksumlo2);
95
96 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
97 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
98 const __m128i vksum011 = _mm_unpacklo_epi32(vksumzp1, vzero);
99 const __m128i vksum231 = _mm_unpackhi_epi32(vksumzp1, vzero);
100 const __m128i vksum012 = _mm_unpacklo_epi32(vksumzp2, vzero);
101 const __m128i vksum232 = _mm_unpackhi_epi32(vksumzp2, vzero);
102
103 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
104 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
105 __m128i vacc1x0 = _mm_unpacklo_epi64(vksum011, vzero);
106 __m128i vacc1x1 = _mm_unpackhi_epi64(vksum011, vzero);
107 __m128i vacc2x0 = _mm_unpacklo_epi64(vksum012, vzero);
108 __m128i vacc2x1 = _mm_unpackhi_epi64(vksum012, vzero);
109
110 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
111 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
112 __m128i vacc1x2 = _mm_unpacklo_epi64(vksum231, vzero);
113 __m128i vacc1x3 = _mm_unpackhi_epi64(vksum231, vzero);
114 __m128i vacc2x2 = _mm_unpacklo_epi64(vksum232, vzero);
115 __m128i vacc2x3 = _mm_unpackhi_epi64(vksum232, vzero);
116 w = (const int32_t*) w + 4;
117
118 size_t k = kc;
119
120
121 while (k >= 8 * sizeof(int8_t)) {
122 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
123 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
124 a0 += 8;
125 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
126 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
127 a1 += 8;
128 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
129 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
130 a2 += 8;
131
132 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
133
134 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
135 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
136 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
137
138 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
139 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
140 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
141 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
142 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
143 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
144 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
145
146 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
147 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
148 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
149
150 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
151 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
152 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
153 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
154 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
155 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
156
157 w = (const int8_t*) w + 32;
158 k -= 8 * sizeof(int8_t);
159 }
160
161 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
162 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
163 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
164 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
165 const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
166 const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
167
168 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
169 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
170 __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
171
172 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
173 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
174 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
175
176 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
177 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
178 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
179 const __m128 vinput_scale2 = _mm_load1_ps(&quantization_params[2].inv_scale);
180
181 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
182 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
183 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale2);
184
185 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
186 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
187 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
188 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
189
190 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
191 w = (const float*) w + 8;
192 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
193 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
194 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
195
196 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
197 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
198 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
199
200 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
201 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
202 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
203
204 if XNN_LIKELY(nc >= 4) {
205 _mm_storeu_ps(c0, vout0x0123);
206 _mm_storeu_ps(c1, vout1x0123);
207 _mm_storeu_ps(c2, vout2x0123);
208
209 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
210 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
211 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
212
213 c0 = (float*) ((uintptr_t) c0 + cn_stride);
214 c1 = (float*) ((uintptr_t) c1 + cn_stride);
215 c2 = (float*) ((uintptr_t) c2 + cn_stride);
216
217 nc -= 4;
218 } else {
219 if (nc & 2) {
220 _mm_storel_pi((__m64*) c0, vout0x0123);
221 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
222 c0 += 2;
223 _mm_storel_pi((__m64*) c1, vout1x0123);
224 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
225 c1 += 2;
226 _mm_storel_pi((__m64*) c2, vout2x0123);
227 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
228 c2 += 2;
229 }
230 if (nc & 1) {
231 _mm_store_ss(c0, vout0x0123);
232 _mm_store_ss(c1, vout1x0123);
233 _mm_store_ss(c2, vout2x0123);
234 }
235 nc = 0;
236 }
237 } while (nc != 0);
238 }