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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/igemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22
23
xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld64(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 size_t ks,
29 const uint8_t** restrict a,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 size_t a_offset,
35 const uint8_t* zero,
36 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 3);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (3 * sizeof(void*)) == 0);
44 assert(a_offset % sizeof(uint8_t) == 0);
45 assert(a != NULL);
46 assert(w != NULL);
47 assert(c != NULL);
48
49 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
50 uint8_t* c0 = c;
51 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr < 2) {
53 c1 = c0;
54 }
55 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 c2 = c1;
58 }
59
60 do {
61 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
62 __m128i vacc1x0123 = vacc0x0123;
63 __m128i vacc2x0123 = vacc0x0123;
64 w = (const void*) ((const int32_t*) w + 4);
65
66 size_t p = ks;
67 do {
68 const uint8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
71 }
72 const uint8_t* restrict a1 = a[1];
73 if XNN_UNPREDICTABLE(a1 != zero) {
74 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
75 }
76 const uint8_t* restrict a2 = a[2];
77 if XNN_UNPREDICTABLE(a2 != zero) {
78 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
79 }
80 a += 3;
81
82 size_t k = kc;
83 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
84 while (k >= 8 * sizeof(uint8_t)) {
85 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
86 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
87 a0 += 8;
88 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
89 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
90 a1 += 8;
91 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
92 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
93 a2 += 8;
94
95 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
96 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
97
98 vacc0x0123 = _mm_maddd_epi16(
99 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
100 vacc1x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
102 vacc2x0123 = _mm_maddd_epi16(
103 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
104 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
105 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
106
107 vacc0x0123 = _mm_maddd_epi16(
108 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
109 vacc1x0123 = _mm_maddd_epi16(
110 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
111 vacc2x0123 = _mm_maddd_epi16(
112 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
113 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
114 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
115
116 vacc0x0123 = _mm_maddd_epi16(
117 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
118 vacc1x0123 = _mm_maddd_epi16(
119 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
120 vacc2x0123 = _mm_maddd_epi16(
121 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
122 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
123 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
124
125 vacc0x0123 = _mm_maddd_epi16(
126 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
127 vacc1x0123 = _mm_maddd_epi16(
128 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
129 vacc2x0123 = _mm_maddd_epi16(
130 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
131
132 w = (const void*) ((const uint8_t*) w + 32);
133 k -= 8 * sizeof(uint8_t);
134 }
135 if (k != 0) {
136 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
137 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
138 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
139 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
140 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
141 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
142 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
143 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
144 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
145
146 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
147 w = (const void*) ((const uint8_t*) w + 8);
148 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
149
150 vacc0x0123 = _mm_maddd_epi16(
151 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
152 vacc1x0123 = _mm_maddd_epi16(
153 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
154 vacc2x0123 = _mm_maddd_epi16(
155 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
156
157 if (k > 2 * sizeof(uint8_t)) {
158 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
159 w = (const void*) ((const uint8_t*) w + 8);
160 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
161
162 vacc0x0123 = _mm_maddd_epi16(
163 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
164 vacc1x0123 = _mm_maddd_epi16(
165 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
166 vacc2x0123 = _mm_maddd_epi16(
167 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
168
169 if (k > 4 * sizeof(uint8_t)) {
170 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
171 w = (const void*) ((const uint8_t*) w + 8);
172 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
173
174 vacc0x0123 = _mm_maddd_epi16(
175 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
176 vacc1x0123 = _mm_maddd_epi16(
177 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
178 vacc2x0123 = _mm_maddd_epi16(
179 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
180 }
181 }
182 }
183 p -= 3 * sizeof(void*);
184 } while (p != 0);
185
186 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
187 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
188 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
189
190 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
191 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
192 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
193 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
194
195 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
196 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
197 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
198 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
199
200 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
201 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
202 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
203
204 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
205 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
206 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
207
208 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
209
210 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
211
212 if (nc >= 4) {
213 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
214 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
215 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
216 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
217 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
218 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
219
220 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
221
222 nc -= 4;
223 } else {
224 if (nc & 2) {
225 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
226 c2 += 2;
227 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
228 c1 += 2;
229 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
230 c0 += 2;
231 vout = _mm_srli_epi32(vout, 16);
232 }
233 if (nc & 1) {
234 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
235 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
236 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
237 }
238
239 nc = 0;
240 }
241 } while (nc != 0);
242 }
243