1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22
23
24
xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])25 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64(
26 size_t mr,
27 size_t nc,
28 size_t kc,
29 const uint8_t* restrict a,
30 size_t a_stride,
31 const void* restrict w,
32 uint8_t* restrict c,
33 size_t cm_stride,
34 size_t cn_stride,
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(kc % sizeof(uint8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
47 const uint8_t* a0 = a;
48 uint8_t* c0 = c;
49 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
50 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
56 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 2) {
58 a2 = a1;
59 c2 = c1;
60 }
61
62 do {
63 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
64 __m128i vacc1x0123 = vacc0x0123;
65 __m128i vacc2x0123 = vacc0x0123;
66 w = (const void*) ((const int32_t*) w + 4);
67
68 size_t k = kc;
69 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
70 while (k >= 8 * sizeof(uint8_t)) {
71 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
72 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
73 a0 += 8;
74 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
75 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
76 a1 += 8;
77 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
78 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
79 a2 += 8;
80
81 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
82 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
83
84 vacc0x0123 = _mm_maddd_epi16(
85 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
86 vacc1x0123 = _mm_maddd_epi16(
87 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
88 vacc2x0123 = _mm_maddd_epi16(
89 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
90 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
91 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
92
93 vacc0x0123 = _mm_maddd_epi16(
94 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
95 vacc1x0123 = _mm_maddd_epi16(
96 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
97 vacc2x0123 = _mm_maddd_epi16(
98 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
99 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
100 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
101
102 vacc0x0123 = _mm_maddd_epi16(
103 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
104 vacc1x0123 = _mm_maddd_epi16(
105 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
106 vacc2x0123 = _mm_maddd_epi16(
107 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
108 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
109 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
110
111 vacc0x0123 = _mm_maddd_epi16(
112 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
113 vacc1x0123 = _mm_maddd_epi16(
114 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
115 vacc2x0123 = _mm_maddd_epi16(
116 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
117
118 w = (const void*) ((const uint8_t*) w + 32);
119 k -= 8 * sizeof(uint8_t);
120 }
121 if (k != 0) {
122 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
123 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
124 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
125 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
126 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
127 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
128 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
129 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
130 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
131
132 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
133 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
134 w = (const void*) ((const uint8_t*) w + 8);
135
136 vacc0x0123 = _mm_maddd_epi16(
137 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
138 vacc1x0123 = _mm_maddd_epi16(
139 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
140 vacc2x0123 = _mm_maddd_epi16(
141 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
142
143 if (k > 2 * sizeof(uint8_t)) {
144 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
145 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
146 w = (const void*) ((const uint8_t*) w + 8);
147
148 vacc0x0123 = _mm_maddd_epi16(
149 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
150 vacc1x0123 = _mm_maddd_epi16(
151 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
152 vacc2x0123 = _mm_maddd_epi16(
153 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
154
155 if (k > 4 * sizeof(uint8_t)) {
156 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
157 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
158 w = (const void*) ((const uint8_t*) w + 8);
159
160 vacc0x0123 = _mm_maddd_epi16(
161 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
162 vacc1x0123 = _mm_maddd_epi16(
163 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
164 vacc2x0123 = _mm_maddd_epi16(
165 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
166 }
167 }
168 }
169
170 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
171 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
172 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
173
174 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
175 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
176 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
177 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
178
179 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
180 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
181 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
182 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
183
184 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
185 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
186 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
187
188 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
189 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
190 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
191
192 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
193
194 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
195
196 if (nc >= 4) {
197 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
198 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
199 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
200
201 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
202 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
203 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
204
205 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
206 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
207 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
208
209 nc -= 4;
210 } else {
211 if (nc & 2) {
212 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
213 c0 += 2;
214 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
215 c1 += 2;
216 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
217 c2 += 2;
218 vout = _mm_srli_epi32(vout, 16);
219 }
220 if (nc & 1) {
221 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
222 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
223 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
224 }
225
226 nc = 0;
227 }
228 } while (nc != 0);
229 }
230