1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-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_3x4c8__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_3x4c8__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 <= 3);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (3 * 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, 8);
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 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 c2 = c1;
58 }
59
60 do {
61 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
62 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
63 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
64 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
65 __m128i vacc1x0 = vacc0x0;
66 __m128i vacc1x1 = vacc0x1;
67 __m128i vacc1x2 = vacc0x2;
68 __m128i vacc1x3 = vacc0x3;
69 __m128i vacc2x0 = vacc0x0;
70 __m128i vacc2x1 = vacc0x1;
71 __m128i vacc2x2 = vacc0x2;
72 __m128i vacc2x3 = vacc0x3;
73 w = (const int32_t*) w + 4;
74
75 size_t p = ks;
76 do {
77 const uint8_t* restrict a0 = a[0];
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
80 }
81 const uint8_t* restrict a1 = a[1];
82 if XNN_UNPREDICTABLE(a1 != zero) {
83 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
84 }
85 const uint8_t* restrict a2 = a[2];
86 if XNN_UNPREDICTABLE(a2 != zero) {
87 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
88 }
89 a += 3;
90
91 size_t k = 0;
92 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
93 while (k < kc) {
94 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
95 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
96 a0 += 8;
97 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
98 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
99 a1 += 8;
100 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
101 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
102 a2 += 8;
103
104 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
105 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
106
107 vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
108 vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
109 vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
110 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
111 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
112
113 vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
114 vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
115 vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
116 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
117 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
118
119 vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
120 vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
121 vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
122 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
123 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
124
125 vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
126 vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
127 vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
128
129 w = (const void*) ((const uint8_t*) w + 32);
130 k += 8 * sizeof(uint8_t);
131 }
132 p -= 3 * sizeof(void*);
133 } while (p != 0);
134
135 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
136 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
137 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
138 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
139 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
140 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
141
142 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
143 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
144 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
145
146 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
147 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
148 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
149
150 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
151 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
152 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
153 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
154
155 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
156 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
157 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
158 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
159
160 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
161 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
162 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
163
164 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
165 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
166 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
167
168 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
169
170 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
171
172 if (nc >= 4) {
173 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
174 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
175 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
176 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
177 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
178 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
179
180 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
181
182 nc -= 4;
183 } else {
184 if (nc & 2) {
185 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
186 c2 += 2;
187 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
188 c1 += 2;
189 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
190 c0 += 2;
191 vout = _mm_srli_epi32(vout, 16);
192 }
193 if (nc & 1) {
194 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
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