Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x4c8-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_2x4c8__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 <= 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 vb01c01 = _mm_load_si128((const __m128i*) w);
89 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
90 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
91 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
92 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
93 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
94
95 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
96 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
97 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
98 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
99 const __m128i vb23c01 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
100 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
101 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
102 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
103 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
104 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
105
106 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
107 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
108 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
109 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
110
111 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
112 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
113 a0 += 8;
114 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
115 const __m128i vxa1c1 = _mm_cvtepi8_epi16(va1c1);
116 a1 += 8;
117
118 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
119 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
120 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
121 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
122
123 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
124 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
125 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
126 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
127 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
128 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
129 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
130 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
131
132 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
133 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
134 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
135 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
136
137
138 w = (const int8_t*) w + 32;
139 k -= 16 * sizeof(int8_t);
140 }
141
142 while (k >= 8 * sizeof(int8_t)) {
143 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
144 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
145 a0 += 8;
146 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
147 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
148 a1 += 8;
149
150 __m128i vb01 = _mm_load_si128((const __m128i*) w);
151 vb01 = _mm_slli_epi32(vb01, 4);
152 vb01 = _mm_and_si128(vb01, vmask);
153
154 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
155 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
156 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
157
158 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
159 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
160 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
161 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
162 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
163 vb23 = _mm_slli_epi32(vb23, 4);
164 vb23 = _mm_and_si128(vb23, vmask);
165
166 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
167 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
168 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
169
170 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
171 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
172 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
173 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
174
175 w = (const int8_t*) w + 32;
176 k -= 8 * sizeof(int8_t);
177 }
178
179 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
180 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
181 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
182 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
183
184 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
185 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
186
187 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
188 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
189 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
190 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
191
192 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
193 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
194 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
195
196 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
197 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
198
199 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
200 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
201 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
202
203 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
204 w = (const float*) w + 8;
205 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
206 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
207
208 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
209 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
210
211 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
212 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
213
214 if XNN_LIKELY(nc >= 4) {
215 _mm_storeu_ps(c0, vout0x0123);
216 _mm_storeu_ps(c1, vout1x0123);
217
218 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
219 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
220
221 c0 = (float*) ((uintptr_t) c0 + cn_stride);
222 c1 = (float*) ((uintptr_t) c1 + cn_stride);
223
224 nc -= 4;
225 } else {
226 if (nc & 2) {
227 _mm_storel_pi((__m64*) c0, vout0x0123);
228 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
229 c0 += 2;
230 _mm_storel_pi((__m64*) c1, vout1x0123);
231 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
232 c1 += 2;
233 }
234 if (nc & 1) {
235 _mm_store_ss(c0, vout0x0123);
236 _mm_store_ss(c1, vout1x0123);
237 }
238 nc = 0;
239 }
240 } while (nc != 0);
241 }