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_qc8_igemm_minmax_fp32_ukernel_3x4c8__xop_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qc8_igemm_minmax_fp32_ukernel_3x4c8__xop_ld128(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 size_t ks,
29 const int8_t** restrict a,
30 const void* restrict w,
31 int8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 size_t a_offset,
35 const int8_t* zero,
36 const union xnn_qc8_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(int8_t) == 0);
45 assert(a != NULL);
46 assert(w != NULL);
47 assert(c != NULL);
48
49 kc = round_up_po2(kc, 8);
50 int8_t* c0 = c;
51 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr < 2) {
53 c1 = c0;
54 }
55 int8_t* c2 = (int8_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 int8_t* restrict a0 = a[0];
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
80 }
81 const int8_t* restrict a1 = a[1];
82 if XNN_UNPREDICTABLE(a1 != zero) {
83 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
84 }
85 const int8_t* restrict a2 = a[2];
86 if XNN_UNPREDICTABLE(a2 != zero) {
87 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
88 }
89 a += 3;
90
91 size_t k = 0;
92 while (k < kc) {
93 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
94 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
95 a0 += 8;
96 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
97 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
98 a1 += 8;
99 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
100 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
101 a2 += 8;
102
103 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
104 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
105 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
106
107 vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
108 vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
109 vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
110 vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
111 vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
112 vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
113 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
114 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
115 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
116
117 vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
118 vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
119 vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
120 vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
121 vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
122 vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
123
124 w = (const void*) ((const int8_t*) w + 32);
125 k += 8 * sizeof(int8_t);
126 }
127 p -= 3 * sizeof(void*);
128 } while (p != 0);
129
130 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
131 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
132 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
133 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
134 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
135 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
136
137 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
138 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
139 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
140
141 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
142 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
143 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
144
145 const __m128 vscale0123 = _mm_load_ps((const float*) w);
146 w = (const void*) ((const float*) w + 4);
147 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
148 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
149 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
150
151 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
152 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
153 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
154 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
155
156 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
157 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
158 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
159
160 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
161 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
162 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
163
164
165 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
166
167 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
168
169 if (nc >= 4) {
170 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
171 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
172 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
173 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
174 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
175 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
176
177 a = (const int8_t**restrict) ((uintptr_t) a - ks);
178
179 nc -= 4;
180 } else {
181 if (nc & 2) {
182 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
183 c2 += 2;
184 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
185 c1 += 2;
186 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
187 c0 += 2;
188 vout = _mm_srli_epi32(vout, 16);
189 }
190 if (nc & 1) {
191 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
192 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
193 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
194 }
195
196 nc = 0;
197 }
198 } while (nc != 0);
199 }
200