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 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__sse2_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)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__sse2_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const uint8_t** restrict a,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const uint8_t* zero,
31 const union xnn_qu8_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(uint8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
45 uint8_t* c0 = c;
46 uint8_t* c1 = (uint8_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 uint8_t* restrict a0 = a[0];
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 const uint8_t* restrict a1 = a[1];
63 if XNN_UNPREDICTABLE(a1 != zero) {
64 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
65 }
66 a += 2;
67
68 size_t k = kc;
69 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
70 const __m128i vzero = _mm_setzero_si128();
71 while (k >= 8 * sizeof(uint8_t)) {
72 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
73 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
74 a0 += 8;
75 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
76 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
77 a1 += 8;
78
79 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
80 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
81
82 vacc0x0123 = _mm_add_epi32(vacc0x0123,
83 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
84 vacc1x0123 = _mm_add_epi32(vacc1x0123,
85 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
86 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
87 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
88
89 vacc0x0123 = _mm_add_epi32(vacc0x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
91 vacc1x0123 = _mm_add_epi32(vacc1x0123,
92 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
93 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
94 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
95
96 vacc0x0123 = _mm_add_epi32(vacc0x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
98 vacc1x0123 = _mm_add_epi32(vacc1x0123,
99 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
100 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
101 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
102
103 vacc0x0123 = _mm_add_epi32(vacc0x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
105 vacc1x0123 = _mm_add_epi32(vacc1x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
107
108 w = (const void*) ((const uint8_t*) w + 32);
109 k -= 8 * sizeof(uint8_t);
110 }
111 if (k != 0) {
112 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
113 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
114 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
115 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
116 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
117 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
118
119 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
120 w = (const void*) ((const uint8_t*) w + 8);
121 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
122
123 vacc0x0123 = _mm_add_epi32(vacc0x0123,
124 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
125 vacc1x0123 = _mm_add_epi32(vacc1x0123,
126 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
127
128 if (k > 2 * sizeof(uint8_t)) {
129 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
130 w = (const void*) ((const uint8_t*) w + 8);
131 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
132
133 vacc0x0123 = _mm_add_epi32(vacc0x0123,
134 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
135 vacc1x0123 = _mm_add_epi32(vacc1x0123,
136 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
137
138 if (k > 4 * sizeof(uint8_t)) {
139 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
140 w = (const void*) ((const uint8_t*) w + 8);
141 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
142
143 vacc0x0123 = _mm_add_epi32(vacc0x0123,
144 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
145 vacc1x0123 = _mm_add_epi32(vacc1x0123,
146 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
147 }
148 }
149 }
150 p -= 2 * sizeof(void*);
151 } while (p != 0);
152
153 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
154 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
155
156 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
157 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
158 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
159
160 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
161 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
162 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
163
164 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
165 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
166
167 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
168 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
169
170 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
171
172 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
173
174 if (nc >= 4) {
175 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 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(c1, (uint16_t) _mm_extract_epi16(vout, 2));
186 c1 += 2;
187 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
188 c0 += 2;
189 vout = _mm_srli_epi32(vout, 16);
190 }
191 if (nc & 1) {
192 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
193 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
194 }
195
196 nc = 0;
197 }
198 } while (nc != 0);
199 }
200