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
22
xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld128(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)])23 void xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld128(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 size_t ks,
28 const uint8_t** restrict a,
29 const void* restrict w,
30 uint8_t* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 size_t a_offset,
34 const uint8_t* zero,
35 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 3);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(ks != 0);
42 assert(ks % (3 * sizeof(void*)) == 0);
43 assert(a_offset % sizeof(uint8_t) == 0);
44 assert(a != NULL);
45 assert(w != NULL);
46 assert(c != NULL);
47
48 kc = round_up_po2(kc, 2);
49 uint8_t* c0 = c;
50 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 c1 = c0;
53 }
54 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 c2 = c1;
57 }
58
59 do {
60 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
61 __m128i vacc1x0123 = vacc0x0123;
62 __m128i vacc2x0123 = vacc0x0123;
63 w = (const void*) ((const int32_t*) w + 4);
64
65 size_t p = ks;
66 do {
67 const uint8_t* restrict a0 = a[0];
68 if XNN_UNPREDICTABLE(a0 != zero) {
69 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
70 }
71 const uint8_t* restrict a1 = a[1];
72 if XNN_UNPREDICTABLE(a1 != zero) {
73 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
74 }
75 const uint8_t* restrict a2 = a[2];
76 if XNN_UNPREDICTABLE(a2 != zero) {
77 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
78 }
79 a += 3;
80
81 size_t k = kc;
82 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
83 const __m128i vzero = _mm_setzero_si128();
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 vb01 = _mm_loadu_si128((const __m128i*) w);
96 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
97 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
98
99 vacc0x0123 = _mm_maddd_epi16(
100 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
101 vacc1x0123 = _mm_maddd_epi16(
102 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
103 vacc2x0123 = _mm_maddd_epi16(
104 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
105
106 vacc0x0123 = _mm_maddd_epi16(
107 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
108 vacc1x0123 = _mm_maddd_epi16(
109 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
110 vacc2x0123 = _mm_maddd_epi16(
111 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
112 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
113 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
114 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), 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
123 vacc0x0123 = _mm_maddd_epi16(
124 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
125 vacc1x0123 = _mm_maddd_epi16(
126 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
127 vacc2x0123 = _mm_maddd_epi16(
128 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
129
130 w = (const void*) ((const uint8_t*) w + 32);
131 k -= 8 * sizeof(uint8_t);
132 }
133 if (k != 0) {
134 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
135 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
136 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
137 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
138 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
139 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
140 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
141 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
142 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
143
144 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
145 w = (const void*) ((const uint8_t*) w + 8);
146 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
147
148 vacc0x0123 = _mm_maddd_epi16(
149 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
150 vacc1x0123 = _mm_maddd_epi16(
151 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
152 vacc2x0123 = _mm_maddd_epi16(
153 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
154
155 if (k > 2 * sizeof(uint8_t)) {
156 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
157 w = (const void*) ((const uint8_t*) w + 8);
158 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
159
160 vacc0x0123 = _mm_maddd_epi16(
161 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
162 vacc1x0123 = _mm_maddd_epi16(
163 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
164 vacc2x0123 = _mm_maddd_epi16(
165 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
166
167 if (k > 4 * sizeof(uint8_t)) {
168 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
169 w = (const void*) ((const uint8_t*) w + 8);
170 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
171
172 vacc0x0123 = _mm_maddd_epi16(
173 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
174 vacc1x0123 = _mm_maddd_epi16(
175 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
176 vacc2x0123 = _mm_maddd_epi16(
177 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
178 }
179 }
180 }
181 p -= 3 * sizeof(void*);
182 } while (p != 0);
183
184 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
185 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
186 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
187
188 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
189 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
190 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
191 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
192
193 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
194 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
195 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
196 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
197
198 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
199 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
200 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
201
202 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
203 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
204 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
205
206 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
207
208 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
209
210 if (nc >= 4) {
211 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
212 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
213 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
214 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
215 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
216 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
217
218 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
219
220 nc -= 4;
221 } else {
222 if (nc & 2) {
223 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
224 c2 += 2;
225 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
226 c1 += 2;
227 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
228 c0 += 2;
229 vout = _mm_srli_epi32(vout, 16);
230 }
231 if (nc & 1) {
232 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
233 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
234 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
235 }
236
237 nc = 0;
238 }
239 } while (nc != 0);
240 }
241