Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x4c2-minmax-fp32-sse2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-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/igemm.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_2x4c2__sse2_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 <= 2);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (2 * 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, 2 * 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
51 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);
52 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
53 const __m128i voutput_min = _mm_set1_epi16(params->fp32_scalar.output_min);
54 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
55 XNN_FORCE_REALIZATION(voutput_zero_point);
56 XNN_FORCE_REALIZATION(voutput_min);
57
58 do {
59 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
60 __m128i vacc1x0123 = vacc0x0123;
61 w = (const void*) ((const int32_t*) w + 4);
62
63 size_t p = ks;
64 do {
65 const int8_t* restrict a0 = a[0];
66 if XNN_UNPREDICTABLE(a0 != zero) {
67 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
68 }
69 const int8_t* restrict a1 = a[1];
70 if XNN_UNPREDICTABLE(a1 != zero) {
71 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
72 }
73 a += 2;
74
75 size_t k = kc;
76 while (k >= 8 * sizeof(int8_t)) {
77 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
78 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
79 a0 += 8;
80 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
81 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
82 a1 += 8;
83
84 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
85 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
86
87 vacc0x0123 = _mm_add_epi32(vacc0x0123,
88 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
89 vacc1x0123 = _mm_add_epi32(vacc1x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
91 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
92 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
93
94 vacc0x0123 = _mm_add_epi32(vacc0x0123,
95 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
96 vacc1x0123 = _mm_add_epi32(vacc1x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
98 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
99 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
100
101 vacc0x0123 = _mm_add_epi32(vacc0x0123,
102 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
103 vacc1x0123 = _mm_add_epi32(vacc1x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
105 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
106 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
107
108 vacc0x0123 = _mm_add_epi32(vacc0x0123,
109 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
110 vacc1x0123 = _mm_add_epi32(vacc1x0123,
111 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
112
113 w = (const void*) ((const int8_t*) w + 32);
114 k -= 8 * sizeof(int8_t);
115 }
116 if (k != 0) {
117 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
118 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
119 a0 = (const int8_t*) ((uintptr_t) a0 + k);
120 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
121 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
122 a1 = (const int8_t*) ((uintptr_t) a1 + k);
123
124 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
125 w = (const int8_t*) w + 8;
126 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
127
128 vacc0x0123 = _mm_add_epi32(vacc0x0123,
129 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
130 vacc1x0123 = _mm_add_epi32(vacc1x0123,
131 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
132
133 if (k > 2 * sizeof(int8_t)) {
134 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
135 w = (const int8_t*) w + 8;
136 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
137
138 vacc0x0123 = _mm_add_epi32(vacc0x0123,
139 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
140 vacc1x0123 = _mm_add_epi32(vacc1x0123,
141 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
142
143 if (k > 4 * sizeof(int8_t)) {
144 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
145 w = (const int8_t*) w + 8;
146 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
147
148 vacc0x0123 = _mm_add_epi32(vacc0x0123,
149 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
150 vacc1x0123 = _mm_add_epi32(vacc1x0123,
151 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
152 }
153 }
154 }
155 p -= 2 * sizeof(void*);
156 } while (p != 0);
157
158 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
159 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
160
161 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
162 w = (const float*) w + 4;
163 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
164 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
165
166 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
167 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
168
169 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
170 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
171
172 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
173
174 vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
175
176 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
177
178
179 if (nc >= 4) {
180 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1))));
181 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
182 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
183 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
184
185 a = (const int8_t**restrict) ((uintptr_t) a - ks);
186
187 nc -= 4;
188 } else {
189 if (nc & 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 *c1 = (int8_t) _mm_extract_epi16(vout, 2);
198 *c0 = (int8_t) _mm_cvtsi128_si32(vout);
199 }
200
201 nc = 0;
202 }
203 } while (nc != 0);
204 }