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 #include <xnnpack/unaligned.h>
22
23
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)])24 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__xop_ld64(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 size_t ks,
29 const uint8_t** restrict a,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 size_t a_offset,
35 const uint8_t* zero,
36 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 2);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (2 * sizeof(void*)) == 0);
44 assert(a_offset % sizeof(uint8_t) == 0);
45 assert(a != NULL);
46 assert(w != NULL);
47 assert(c != NULL);
48
49 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
50 uint8_t* c0 = c;
51 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr != 2) {
53 c1 = c0;
54 }
55
56 do {
57 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
58 __m128i vacc1x0123 = vacc0x0123;
59 w = (const void*) ((const int32_t*) w + 4);
60
61 size_t p = ks;
62 do {
63 const uint8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const uint8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 a += 2;
72
73 size_t k = kc;
74 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
75 while (k >= 8 * sizeof(uint8_t)) {
76 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
77 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
78 a0 += 8;
79 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
80 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
81 a1 += 8;
82
83 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
84 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
85
86 vacc0x0123 = _mm_maddd_epi16(
87 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
88 vacc1x0123 = _mm_maddd_epi16(
89 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
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 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
98 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
99
100 vacc0x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
102 vacc1x0123 = _mm_maddd_epi16(
103 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
104 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
105 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
106
107 vacc0x0123 = _mm_maddd_epi16(
108 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
109 vacc1x0123 = _mm_maddd_epi16(
110 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
111
112 w = (const void*) ((const uint8_t*) w + 32);
113 k -= 8 * sizeof(uint8_t);
114 }
115 if (k != 0) {
116 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
117 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
118 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
119 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
120 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
121 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
122
123 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
124 w = (const void*) ((const uint8_t*) w + 8);
125 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
126
127 vacc0x0123 = _mm_maddd_epi16(
128 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
129 vacc1x0123 = _mm_maddd_epi16(
130 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
131
132 if (k > 2 * sizeof(uint8_t)) {
133 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
134 w = (const void*) ((const uint8_t*) w + 8);
135 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
136
137 vacc0x0123 = _mm_maddd_epi16(
138 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
139 vacc1x0123 = _mm_maddd_epi16(
140 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
141
142 if (k > 4 * sizeof(uint8_t)) {
143 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
144 w = (const void*) ((const uint8_t*) w + 8);
145 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
146
147 vacc0x0123 = _mm_maddd_epi16(
148 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
149 vacc1x0123 = _mm_maddd_epi16(
150 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
151 }
152 }
153 }
154 p -= 2 * sizeof(void*);
155 } while (p != 0);
156
157 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
158 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
159
160 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
161 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
162 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
163
164 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
165 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
166 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
167
168 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
169 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
170
171 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
172 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
173
174 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
175
176 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
177
178 if (nc >= 4) {
179 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
180 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
181 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
182 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
183
184 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
185
186 nc -= 4;
187 } else {
188 if (nc & 2) {
189 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
190 c1 += 2;
191 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
192 c0 += 2;
193 vout = _mm_srli_epi32(vout, 16);
194 }
195 if (nc & 1) {
196 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
197 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
198 }
199
200 nc = 0;
201 }
202 } while (nc != 0);
203 }
204