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
22
23
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)])24 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 const uint8_t* restrict a,
29 size_t a_stride,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 3);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(uint8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 2);
46 const uint8_t* a0 = a;
47 uint8_t* c0 = c;
48 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
49 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 a1 = a0;
52 c1 = c0;
53 }
54 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
55 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 a2 = a1;
58 c2 = c1;
59 }
60
61 do {
62 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
63 __m128i vacc1x0123 = vacc0x0123;
64 __m128i vacc2x0123 = vacc0x0123;
65 w = (const void*) ((const int32_t*) w + 4);
66
67 size_t k = kc;
68 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
69 while (k >= 8 * sizeof(uint8_t)) {
70 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
71 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
72 a0 += 8;
73 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
74 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
75 a1 += 8;
76 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
77 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
78 a2 += 8;
79
80 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
81 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
82
83 vacc0x0123 = _mm_maddd_epi16(
84 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
85 vacc1x0123 = _mm_maddd_epi16(
86 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
87 vacc2x0123 = _mm_maddd_epi16(
88 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
89 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
90 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
91
92 vacc0x0123 = _mm_maddd_epi16(
93 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
94 vacc1x0123 = _mm_maddd_epi16(
95 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
96 vacc2x0123 = _mm_maddd_epi16(
97 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
98 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
99 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
100
101 vacc0x0123 = _mm_maddd_epi16(
102 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
103 vacc1x0123 = _mm_maddd_epi16(
104 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
105 vacc2x0123 = _mm_maddd_epi16(
106 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
107 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
108 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
109
110 vacc0x0123 = _mm_maddd_epi16(
111 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
112 vacc1x0123 = _mm_maddd_epi16(
113 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
114 vacc2x0123 = _mm_maddd_epi16(
115 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
116
117 w = (const void*) ((const uint8_t*) w + 32);
118 k -= 8 * sizeof(uint8_t);
119 }
120 if (k != 0) {
121 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
122 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
123 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
124 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
125 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
126 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
127 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
128 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
129 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
130
131 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
132 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
133 w = (const void*) ((const uint8_t*) w + 8);
134
135 vacc0x0123 = _mm_maddd_epi16(
136 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
137 vacc1x0123 = _mm_maddd_epi16(
138 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
139 vacc2x0123 = _mm_maddd_epi16(
140 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
141
142 if (k > 2 * sizeof(uint8_t)) {
143 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
144 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
145 w = (const void*) ((const uint8_t*) w + 8);
146
147 vacc0x0123 = _mm_maddd_epi16(
148 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
149 vacc1x0123 = _mm_maddd_epi16(
150 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
151 vacc2x0123 = _mm_maddd_epi16(
152 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
153
154 if (k > 4 * sizeof(uint8_t)) {
155 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
156 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
157 w = (const void*) ((const uint8_t*) w + 8);
158
159 vacc0x0123 = _mm_maddd_epi16(
160 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
161 vacc1x0123 = _mm_maddd_epi16(
162 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
163 vacc2x0123 = _mm_maddd_epi16(
164 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
165 }
166 }
167 }
168
169 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
170 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
171 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
172
173 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
174 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
175 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
176 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
177
178 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
179 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
180 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
181 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
182
183 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
184 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
185 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
186
187 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
188 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
189 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
190
191 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
192
193 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
194
195 if (nc >= 4) {
196 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
197 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
198 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
199
200 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
201 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
202 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
203
204 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
205 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
206 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
207
208 nc -= 4;
209 } else {
210 if (nc & 2) {
211 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
212 c0 += 2;
213 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
214 c1 += 2;
215 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
216 c2 += 2;
217 vout = _mm_srli_epi32(vout, 16);
218 }
219 if (nc & 1) {
220 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
221 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
222 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
223 }
224
225 nc = 0;
226 }
227 } while (nc != 0);
228 }
229