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 <emmintrin.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__sse2_ld64(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__sse2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2);
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
52 __m128i vacc1x0123 = vacc0x0123;
53 w = (const void*) ((const int32_t*) w + 4);
54
55 size_t p = ks;
56 do {
57 const int8_t* restrict a0 = a[0];
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60 }
61 const int8_t* restrict a1 = a[1];
62 if XNN_UNPREDICTABLE(a1 != zero) {
63 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
64 }
65 a += 2;
66
67 size_t k = kc;
68 while (k >= 8 * sizeof(int8_t)) {
69 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
70 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
71 a0 += 8;
72 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
73 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
74 a1 += 8;
75
76 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
77 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
78
79 vacc0x0123 = _mm_add_epi32(vacc0x0123,
80 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
81 vacc1x0123 = _mm_add_epi32(vacc1x0123,
82 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
83 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
84 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
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 vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
91 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
92
93 vacc0x0123 = _mm_add_epi32(vacc0x0123,
94 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
95 vacc1x0123 = _mm_add_epi32(vacc1x0123,
96 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
97 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
98 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
99
100 vacc0x0123 = _mm_add_epi32(vacc0x0123,
101 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
102 vacc1x0123 = _mm_add_epi32(vacc1x0123,
103 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
104
105 w = (const void*) ((const int8_t*) w + 32);
106 k -= 8 * sizeof(int8_t);
107 }
108 if (k != 0) {
109 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
110 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
111 a0 = (const int8_t*) ((uintptr_t) a0 + k);
112 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
113 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
114 a1 = (const int8_t*) ((uintptr_t) a1 + k);
115
116 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
117 w = (const void*) ((const int8_t*) w + 8);
118 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
119
120 vacc0x0123 = _mm_add_epi32(vacc0x0123,
121 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
122 vacc1x0123 = _mm_add_epi32(vacc1x0123,
123 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
124
125 if (k > 2 * sizeof(int8_t)) {
126 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
127 w = (const void*) ((const int8_t*) w + 8);
128 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
129
130 vacc0x0123 = _mm_add_epi32(vacc0x0123,
131 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
132 vacc1x0123 = _mm_add_epi32(vacc1x0123,
133 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
134
135 if (k > 4 * sizeof(int8_t)) {
136 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
137 w = (const void*) ((const int8_t*) w + 8);
138 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
139
140 vacc0x0123 = _mm_add_epi32(vacc0x0123,
141 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
142 vacc1x0123 = _mm_add_epi32(vacc1x0123,
143 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
144 }
145 }
146 }
147 p -= 2 * sizeof(void*);
148 } while (p != 0);
149
150 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
151 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
152
153 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
154 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
155 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
156
157 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.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_sse2.output_zero_point);
165 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
166
167 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
168 vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
169
170 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
171
172
173 if (nc >= 4) {
174 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1)));
175 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
176 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
177 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
178
179 a = (const int8_t**restrict) ((uintptr_t) a - ks);
180
181 nc -= 4;
182 } else {
183 if (nc & 2) {
184 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
185 c1 += 2;
186 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
187 c0 += 2;
188 vout = _mm_srli_epi32(vout, 16);
189 }
190 if (nc & 1) {
191 *c1 = (int8_t) _mm_extract_epi16(vout, 2);
192 *c0 = (int8_t) _mm_cvtsi128_si32(vout);
193 }
194
195 nc = 0;
196 }
197 } while (nc != 0);
198 }
199