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_qu8_igemm_minmax_fp32_ukernel_2x4c2__sse41_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__sse41_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 while (k >= 8 * sizeof(uint8_t)) {
71 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
72 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
73 a0 += 8;
74 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
75 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
76 a1 += 8;
77
78 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
79 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
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 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
86 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
87
88 vacc0x0123 = _mm_add_epi32(vacc0x0123,
89 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
90 vacc1x0123 = _mm_add_epi32(vacc1x0123,
91 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
92 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
93 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
94
95 vacc0x0123 = _mm_add_epi32(vacc0x0123,
96 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
97 vacc1x0123 = _mm_add_epi32(vacc1x0123,
98 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
99 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
100 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
101
102 vacc0x0123 = _mm_add_epi32(vacc0x0123,
103 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
104 vacc1x0123 = _mm_add_epi32(vacc1x0123,
105 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
106
107 w = (const void*) ((const uint8_t*) w + 32);
108 k -= 8 * sizeof(uint8_t);
109 }
110 if (k != 0) {
111 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
112 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
113 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
114 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
115 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
116 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
117
118 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
119 w = (const void*) ((const uint8_t*) w + 8);
120 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
121
122 vacc0x0123 = _mm_add_epi32(vacc0x0123,
123 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
124 vacc1x0123 = _mm_add_epi32(vacc1x0123,
125 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
126
127 if (k > 2 * sizeof(uint8_t)) {
128 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
129 w = (const void*) ((const uint8_t*) w + 8);
130 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
131
132 vacc0x0123 = _mm_add_epi32(vacc0x0123,
133 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
134 vacc1x0123 = _mm_add_epi32(vacc1x0123,
135 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
136
137 if (k > 4 * sizeof(uint8_t)) {
138 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
139 w = (const void*) ((const uint8_t*) w + 8);
140 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
141
142 vacc0x0123 = _mm_add_epi32(vacc0x0123,
143 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
144 vacc1x0123 = _mm_add_epi32(vacc1x0123,
145 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
146 }
147 }
148 }
149 p -= 2 * sizeof(void*);
150 } while (p != 0);
151
152 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
153 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
154
155 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
156 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
157 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
158
159 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
160 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
161 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
162
163 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
164 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
165
166 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
167 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
168
169 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
170
171 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
172
173 if (nc >= 4) {
174 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
175 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
176 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
177 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
178
179 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
180
181 nc -= 4;
182 } else {
183 if (nc & 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 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
192 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
193 }
194
195 nc = 0;
196 }
197 } while (nc != 0);
198 }
199