Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x4c8-minmax-fp32-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 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (3 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54
55
56 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
57 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
58 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
59 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
60 XNN_FORCE_REALIZATION(voutput_zero_point);
61 XNN_FORCE_REALIZATION(voutput_min);
62
63
64 do {
65 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
66 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
67 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
68 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
69 __m128i vacc1x0 = vacc0x0;
70 __m128i vacc1x1 = vacc0x1;
71 __m128i vacc1x2 = vacc0x2;
72 __m128i vacc1x3 = vacc0x3;
73 __m128i vacc2x0 = vacc0x0;
74 __m128i vacc2x1 = vacc0x1;
75 __m128i vacc2x2 = vacc0x2;
76 __m128i vacc2x3 = vacc0x3;
77 w = (const int32_t*) w + 4;
78
79 size_t p = ks;
80 do {
81 const int8_t* restrict a0 = a[0];
82 if XNN_UNPREDICTABLE(a0 != zero) {
83 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
84 }
85 const int8_t* restrict a1 = a[1];
86 if XNN_UNPREDICTABLE(a1 != zero) {
87 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
88 }
89 const int8_t* restrict a2 = a[2];
90 if XNN_UNPREDICTABLE(a2 != zero) {
91 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
92 }
93 a += 3;
94
95 size_t k = 0;
96 while (k < kc) {
97 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
98 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
99 a0 += 8;
100 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
101 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
102 a1 += 8;
103 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
104 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
105 a2 += 8;
106
107 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
108 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
109
110 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
111 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
112 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
113 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
114 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
115
116 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
117 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
118 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
119 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
120 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
121
122 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
123 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
124 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
125 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
126 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
127
128 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
129 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
130 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
131
132 w = (const void*) ((const int8_t*) w + 32);
133 k += 8 * sizeof(int8_t);
134 }
135 p -= 3 * sizeof(void*);
136 } while (p != 0);
137
138 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
139 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
140 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
141 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
142 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
143 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
144
145 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
146 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
147 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
148
149 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
150 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
151 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
152
153 const __m128 vscale0123 = _mm_load_ps((const float*) w);
154 w = (const float*) w + 4;
155 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
156 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
157 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
158
159 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
160 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
161 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
162
163 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
164 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
165 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
166
167 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
168 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
169
170
171 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
172
173 vout = _mm_max_epi8(vout, voutput_min);
174
175 if (nc >= 4) {
176 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
177 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
178 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
179 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
180 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
181 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
182
183 a = (const int8_t**restrict) ((uintptr_t) a - ks);
184
185 nc -= 4;
186 } else {
187 if (nc & 2) {
188 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
189 c2 += 2;
190 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
191 c1 += 2;
192 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
193 c0 += 2;
194 vout = _mm_srli_epi32(vout, 16);
195 }
196 if (nc & 1) {
197 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
198 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
199 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
200 }
201
202 nc = 0;
203 }
204 } while (nc != 0);
205 }