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_2x4c2__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)])23 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__xop_ld64(
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 <= 2);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(ks != 0);
42 assert(ks % (2 * 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
55 do {
56 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
57 __m128i vacc1x0123 = vacc0x0123;
58 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t p = ks;
61 do {
62 const uint8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const uint8_t* restrict a1 = a[1];
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 a += 2;
71
72 size_t k = kc;
73 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
74 while (k >= 8 * sizeof(uint8_t)) {
75 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
76 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
77 a0 += 8;
78 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
79 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
80 a1 += 8;
81
82 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
83 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), 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 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 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
97 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
98
99 vacc0x0123 = _mm_maddd_epi16(
100 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
101 vacc1x0123 = _mm_maddd_epi16(
102 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
103 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
104 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
105
106 vacc0x0123 = _mm_maddd_epi16(
107 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
108 vacc1x0123 = _mm_maddd_epi16(
109 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
110
111 w = (const void*) ((const uint8_t*) w + 32);
112 k -= 8 * sizeof(uint8_t);
113 }
114 if (k != 0) {
115 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
116 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
117 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
118 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
119 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
120 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
121
122 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
123 w = (const void*) ((const uint8_t*) w + 8);
124 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
125
126 vacc0x0123 = _mm_maddd_epi16(
127 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
128 vacc1x0123 = _mm_maddd_epi16(
129 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
130
131 if (k > 2 * sizeof(uint8_t)) {
132 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
133 w = (const void*) ((const uint8_t*) w + 8);
134 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
135
136 vacc0x0123 = _mm_maddd_epi16(
137 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
138 vacc1x0123 = _mm_maddd_epi16(
139 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
140
141 if (k > 4 * sizeof(uint8_t)) {
142 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
143 w = (const void*) ((const uint8_t*) w + 8);
144 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
145
146 vacc0x0123 = _mm_maddd_epi16(
147 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
148 vacc1x0123 = _mm_maddd_epi16(
149 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
150 }
151 }
152 }
153 p -= 2 * sizeof(void*);
154 } while (p != 0);
155
156 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
157 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
158
159 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
160 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
161 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
162
163 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
164 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
165 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
166
167 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
168 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
169
170 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
171 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
172
173 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
174
175 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
176
177 if (nc >= 4) {
178 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
179 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
180 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
181 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
182
183 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
184
185 nc -= 4;
186 } else {
187 if (nc & 2) {
188 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
189 c1 += 2;
190 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
191 c0 += 2;
192 vout = _mm_srli_epi32(vout, 16);
193 }
194 if (nc & 1) {
195 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
196 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
197 }
198
199 nc = 0;
200 }
201 } while (nc != 0);
202 }
203