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_ld128(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_ld128(
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 const __m128i vzero = _mm_setzero_si128();
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 vb01 = _mm_loadu_si128((const __m128i*) w);
82 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
83 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
84
85 vacc0x0123 = _mm_maddd_epi16(
86 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
87 vacc1x0123 = _mm_maddd_epi16(
88 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
89 vacc2x0123 = _mm_maddd_epi16(
90 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
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 vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
99 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
100 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), 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
109 vacc0x0123 = _mm_maddd_epi16(
110 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
111 vacc1x0123 = _mm_maddd_epi16(
112 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
113 vacc2x0123 = _mm_maddd_epi16(
114 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
115
116 w = (const void*) ((const uint8_t*) w + 32);
117 k -= 8 * sizeof(uint8_t);
118 }
119 if (k != 0) {
120 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
121 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
122 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
123 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
124 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
125 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
126 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
127 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
128 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
129
130 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
131 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
132 w = (const void*) ((const uint8_t*) w + 8);
133
134 vacc0x0123 = _mm_maddd_epi16(
135 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
136 vacc1x0123 = _mm_maddd_epi16(
137 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
138 vacc2x0123 = _mm_maddd_epi16(
139 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
140
141 if (k > 2 * sizeof(uint8_t)) {
142 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
143 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
144 w = (const void*) ((const uint8_t*) w + 8);
145
146 vacc0x0123 = _mm_maddd_epi16(
147 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
148 vacc1x0123 = _mm_maddd_epi16(
149 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
150 vacc2x0123 = _mm_maddd_epi16(
151 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
152
153 if (k > 4 * sizeof(uint8_t)) {
154 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
155 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
156 w = (const void*) ((const uint8_t*) w + 8);
157
158 vacc0x0123 = _mm_maddd_epi16(
159 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
160 vacc1x0123 = _mm_maddd_epi16(
161 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
162 vacc2x0123 = _mm_maddd_epi16(
163 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
164 }
165 }
166 }
167
168 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
169 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
170 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
171
172 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
173 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
174 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
175 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
176
177 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
178 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
179 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
180 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
181
182 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
183 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
184 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
185
186 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
187 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
188 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
189
190 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
191
192 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
193
194 if (nc >= 4) {
195 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
196 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
197 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
198
199 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
200 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
201 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
202
203 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
204 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
205 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
206
207 nc -= 4;
208 } else {
209 if (nc & 2) {
210 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
211 c0 += 2;
212 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
213 c1 += 2;
214 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
215 c2 += 2;
216 vout = _mm_srli_epi32(vout, 16);
217 }
218 if (nc & 1) {
219 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
220 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
221 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
222 }
223
224 nc = 0;
225 }
226 } while (nc != 0);
227 }
228