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 #include <smmintrin.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2__sse41_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)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2__sse41_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 2);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (2 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr != 2) {
48 c1 = c0;
49 }
50
51 do {
52 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
53 __m128i vacc1x0123 = vacc0x0123;
54 w = (const void*) ((const int32_t*) w + 4);
55
56 size_t p = ks;
57 do {
58 const int8_t* restrict a0 = a[0];
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 const int8_t* restrict a1 = a[1];
63 if XNN_UNPREDICTABLE(a1 != zero) {
64 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
65 }
66 a += 2;
67
68 size_t k = kc;
69 while (k >= 8 * sizeof(int8_t)) {
70 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
71 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
72 a0 += 8;
73 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
74 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
75 a1 += 8;
76
77 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
78 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
79 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
80
81 vacc0x0123 = _mm_add_epi32(vacc0x0123,
82 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
83 vacc1x0123 = _mm_add_epi32(vacc1x0123,
84 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
85
86 vacc0x0123 = _mm_add_epi32(vacc0x0123,
87 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
88 vacc1x0123 = _mm_add_epi32(vacc1x0123,
89 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
90 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
91 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
92 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
93
94 vacc0x0123 = _mm_add_epi32(vacc0x0123,
95 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
96 vacc1x0123 = _mm_add_epi32(vacc1x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
98
99 vacc0x0123 = _mm_add_epi32(vacc0x0123,
100 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
101 vacc1x0123 = _mm_add_epi32(vacc1x0123,
102 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
103
104 w = (const void*) ((const int8_t*) w + 32);
105 k -= 8 * sizeof(int8_t);
106 }
107 if (k != 0) {
108 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
109 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
110 a0 = (const int8_t*) ((uintptr_t) a0 + k);
111 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
112 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
113 a1 = (const int8_t*) ((uintptr_t) a1 + k);
114
115 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
116 w = (const void*) ((const int8_t*) w + 8);
117 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
118
119 vacc0x0123 = _mm_add_epi32(vacc0x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
121 vacc1x0123 = _mm_add_epi32(vacc1x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
123
124 if (k > 2 * sizeof(int8_t)) {
125 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
126 w = (const void*) ((const int8_t*) w + 8);
127 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
128
129 vacc0x0123 = _mm_add_epi32(vacc0x0123,
130 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
131 vacc1x0123 = _mm_add_epi32(vacc1x0123,
132 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
133
134 if (k > 4 * sizeof(int8_t)) {
135 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
136 w = (const void*) ((const int8_t*) w + 8);
137 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
138
139 vacc0x0123 = _mm_add_epi32(vacc0x0123,
140 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
141 vacc1x0123 = _mm_add_epi32(vacc1x0123,
142 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
143 }
144 }
145 }
146 p -= 2 * sizeof(void*);
147 } while (p != 0);
148
149 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
150 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
151
152 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
153 w = (const void*) ((const float*) w + 4);
154 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
155 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
156
157 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
158 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
159 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
160
161 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
162 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
163
164 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
165 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
166
167
168 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
169
170 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
171
172 if (nc >= 4) {
173 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
174 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
175 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
176 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
177
178 a = (const int8_t**restrict) ((uintptr_t) a - ks);
179
180 nc -= 4;
181 } else {
182 if (nc & 2) {
183 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
184 c1 += 2;
185 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
186 c0 += 2;
187 vout = _mm_srli_epi32(vout, 16);
188 }
189 if (nc & 1) {
190 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
191 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
192 }
193
194 nc = 0;
195 }
196 } while (nc != 0);
197 }
198