Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x4c8-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_3x4c8__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 <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (3 * 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
56 const __m128 vmin = _mm_set1_ps(params->scalar.min);
57 const __m128 vmax = _mm_set1_ps(params->scalar.max);
58 XNN_FORCE_REALIZATION(vmin);
59 XNN_FORCE_REALIZATION(vmax);
60 __m128i vinput_zero_point = _mm_cvtsi32_si128(*((const int*) &quantization_params->zero_point));
61 vinput_zero_point = _mm_shuffle_epi32(vinput_zero_point, _MM_SHUFFLE(0, 0, 0, 0));
62 const __m128 vinput_scale = _mm_load1_ps(&quantization_params->inv_scale);
63
64
65 do {
66 const __m128i vksum = _mm_load_si128((const __m128i*) w);
67 const __m128i vzero = _mm_setzero_si128();
68 const __m128i vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point);
69 const __m128i vinit1 = _mm_mullo_epi32(vksum, vinput_zero_point);
70 const __m128i vinit2 = _mm_mullo_epi32(vksum, vinput_zero_point);
71 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
72 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
73 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
74 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
75 __m128i vacc1x0 = _mm_blend_epi16(vinit1, vzero, 0xFC);
76 __m128i vacc1x1 = _mm_blend_epi16(vinit1, vzero, 0xF3);
77 __m128i vacc1x2 = _mm_blend_epi16(vinit1, vzero, 0xCF);
78 __m128i vacc1x3 = _mm_blend_epi16(vinit1, vzero, 0x3F);
79 __m128i vacc2x0 = _mm_blend_epi16(vinit2, vzero, 0xFC);
80 __m128i vacc2x1 = _mm_blend_epi16(vinit2, vzero, 0xF3);
81 __m128i vacc2x2 = _mm_blend_epi16(vinit2, vzero, 0xCF);
82 __m128i vacc2x3 = _mm_blend_epi16(vinit2, vzero, 0x3F);
83 w = (const int32_t*) w + 4;
84
85 size_t p = ks;
86 do {
87 const int8_t* restrict a0 = a[0];
88 if XNN_UNPREDICTABLE(a0 != zero) {
89 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
90 } else {
91 a0 = zero_data;
92 }
93 const int8_t* restrict a1 = a[1];
94 if XNN_UNPREDICTABLE(a1 != zero) {
95 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
96 } else {
97 a1 = zero_data;
98 }
99 const int8_t* restrict a2 = a[2];
100 if XNN_UNPREDICTABLE(a2 != zero) {
101 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
102 } else {
103 a2 = zero_data;
104 }
105 a += 3;
106
107 size_t k = 0;
108 while (k < kc) {
109 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
110 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
111 a0 += 8;
112 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
113 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
114 a1 += 8;
115 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
116 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
117 a2 += 8;
118
119 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
120 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
121
122 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
123 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
124 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
125 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
126 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
127
128 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
129 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
130 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
131 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
132 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
133
134 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
135 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
136 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
137 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
138 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
139
140 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
141 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
142 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
143
144 w = (const void*) ((const int8_t*) w + 32);
145 k += 8 * sizeof(int8_t);
146 }
147 p -= 3 * sizeof(void*);
148 } while (p != 0);
149
150 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
151 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
152 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
153 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
154 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
155 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
156
157 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
158 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
159 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
160
161 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
162 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
163 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
164
165 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale);
166 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale);
167 vout2x0123 = _mm_mul_ps(vout2x0123, vinput_scale);
168
169 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
170 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
171 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
172 vout2x0123 = _mm_mul_ps(vout2x0123, vfilter_output_scale0123);
173
174 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
175 w = (const float*) w + 8;
176 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
177 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
178 vout2x0123 = _mm_add_ps(vout2x0123, vbias0123);
179
180 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
181 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
182 vout2x0123 = _mm_max_ps(vout2x0123, vmin);
183
184 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
185 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
186 vout2x0123 = _mm_min_ps(vout2x0123, vmax);
187
188 if XNN_LIKELY(nc >= 4) {
189 _mm_storeu_ps(c2, vout2x0123);
190 _mm_storeu_ps(c1, vout1x0123);
191 _mm_storeu_ps(c0, vout0x0123);
192
193 a = (const int8_t**restrict) ((uintptr_t) a - ks);
194
195 c0 = (float*) ((uintptr_t) c0 + cn_stride);
196 c1 = (float*) ((uintptr_t) c1 + cn_stride);
197 c2 = (float*) ((uintptr_t) c2 + cn_stride);
198
199 nc -= 4;
200 } else {
201 if (nc & 2) {
202 _mm_storel_pi((__m64*) c2, vout2x0123);
203 vout2x0123 = _mm_unpackhi_ps(vout2x0123, vout2x0123);
204 c2 += 2;
205 _mm_storel_pi((__m64*) c1, vout1x0123);
206 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
207 c1 += 2;
208 _mm_storel_pi((__m64*) c0, vout0x0123);
209 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
210 c0 += 2;
211 }
212 if (nc & 1) {
213 _mm_store_ss(c2, vout2x0123);
214 _mm_store_ss(c1, vout1x0123);
215 _mm_store_ss(c0, vout0x0123);
216 }
217 nc = 0;
218 }
219 } while (nc != 0);
220 }