Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x4c8-minmax-sse41-ld64.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_2x4c8__sse41_ld64(
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 <= 2);
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
52 const __m128 vmin = _mm_set1_ps(params->scalar.min);
53 const __m128 vmax = _mm_set1_ps(params->scalar.max);
54 XNN_FORCE_REALIZATION(vmin);
55 XNN_FORCE_REALIZATION(vmax);
56
57 const __m128i vmask = _mm_set1_epi8(0xF0);
58 XNN_FORCE_REALIZATION(vmask);
59
60 do {
61 const __m128i vksum = _mm_load_si128((const __m128i*) w);
62 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
63 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
64 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
65 const __m128i vzero = _mm_setzero_si128();
66 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
67 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point1);
68 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
69 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
70 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
71 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
72 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
73 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
74 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
75 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
76 w = (const int32_t*) w + 4;
77
78 size_t k = kc;
79
80 while (k >= 16 * sizeof(int8_t)) {
81 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
82 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
83 a0 += 8;
84 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
85 const __m128i vxa1c0 = _mm_cvtepi8_epi16(va1c0);
86 a1 += 8;
87
88 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
89 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
90 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
91 const __m128i vxb0c0 = _mm_cvtepi8_epi16(vb0c0);
92
93 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
94 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
95 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
96 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
97 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
98 const __m128i vxb1c0 = _mm_cvtepi8_epi16(vb1c0);
99
100 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
101 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
102 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
103 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
104 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
105 const __m128i vxb2c0 = _mm_cvtepi8_epi16(vb2c0);
106
107 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
108 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
109 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
110 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
111 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
112 const __m128i vxb3c0 = _mm_cvtepi8_epi16(vb3c0);
113
114 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
115 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
116
117 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
118 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
119 a0 += 8;
120 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
121 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
122 a1 += 8;
123
124 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
125 const __m128i vxb0c1 = _mm_cvtepi8_epi16(vb0c1);
126
127 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
128 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
129 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
130 const __m128i vxb1c1 = _mm_cvtepi8_epi16(vb1c1);
131
132 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
133 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
134 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
135 const __m128i vxb2c1 = _mm_cvtepi8_epi16(vb2c1);
136
137 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
138 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
139 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
140 const __m128i vxb3c1 = _mm_cvtepi8_epi16(vb3c1);
141
142 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
143 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
144
145
146 w = (const int8_t*) w + 32;
147 k -= 16 * sizeof(int8_t);
148 }
149
150 while (k >= 8 * sizeof(int8_t)) {
151 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
152 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
153 a0 += 8;
154 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
155 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
156 a1 += 8;
157
158 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
159 vb0 = _mm_slli_epi32(vb0, 4);
160 vb0 = _mm_and_si128(vb0, vmask);
161
162 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
163
164 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
165 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
166 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
167 vb1 = _mm_slli_epi32(vb1, 4);
168 vb1 = _mm_and_si128(vb1, vmask);
169
170 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
171
172 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
173 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
174 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
175 vb2 = _mm_slli_epi32(vb2, 4);
176 vb2 = _mm_and_si128(vb2, vmask);
177
178 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
179
180 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
181 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
182 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
183 vb3 = _mm_slli_epi32(vb3, 4);
184 vb3 = _mm_and_si128(vb3, vmask);
185
186 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
187
188 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
189 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
190
191 w = (const int8_t*) w + 32;
192 k -= 8 * sizeof(int8_t);
193 }
194
195 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
196 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
197 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
198 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
199
200 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
201 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
202
203 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
204 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
205 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
206 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
207
208 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
209 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
210 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
211
212 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
213 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
214
215 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
216 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
217 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
218
219 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
220 w = (const float*) w + 8;
221 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
222 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
223
224 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
225 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
226
227 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
228 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
229
230 if XNN_LIKELY(nc >= 4) {
231 _mm_storeu_ps(c0, vout0x0123);
232 _mm_storeu_ps(c1, vout1x0123);
233
234 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
235 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
236
237 c0 = (float*) ((uintptr_t) c0 + cn_stride);
238 c1 = (float*) ((uintptr_t) c1 + cn_stride);
239
240 nc -= 4;
241 } else {
242 if (nc & 2) {
243 _mm_storel_pi((__m64*) c0, vout0x0123);
244 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
245 c0 += 2;
246 _mm_storel_pi((__m64*) c1, vout1x0123);
247 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
248 c1 += 2;
249 }
250 if (nc & 1) {
251 _mm_store_ss(c0, vout0x0123);
252 _mm_store_ss(c1, vout1x0123);
253 }
254 nc = 0;
255 }
256 } while (nc != 0);
257 }