1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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 #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/gemm.h>
20 #include <xnnpack/math.h>
21
22
23
xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 const uint8_t* restrict a,
29 size_t a_stride,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 2);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(uint8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 2);
46 const uint8_t* a0 = a;
47 uint8_t* c0 = c;
48 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
49 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr != 2) {
51 a1 = a0;
52 c1 = c0;
53 }
54
55 do {
56 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
57 __m128i vacc1x0123 = vacc0x0123;
58 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t k = kc;
61 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
62 while (k >= 8 * sizeof(uint8_t)) {
63 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
64 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
65 a0 += 8;
66 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
67 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
68 a1 += 8;
69
70 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
71 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
72
73 vacc0x0123 = _mm_maddd_epi16(
74 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
75 vacc1x0123 = _mm_maddd_epi16(
76 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
77 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
78 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
79
80 vacc0x0123 = _mm_maddd_epi16(
81 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
82 vacc1x0123 = _mm_maddd_epi16(
83 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
84 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
85 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
86
87 vacc0x0123 = _mm_maddd_epi16(
88 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
89 vacc1x0123 = _mm_maddd_epi16(
90 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
91 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
92 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
93
94 vacc0x0123 = _mm_maddd_epi16(
95 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
96 vacc1x0123 = _mm_maddd_epi16(
97 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
98
99 w = (const void*) ((const uint8_t*) w + 32);
100 k -= 8 * sizeof(uint8_t);
101 }
102 if (k != 0) {
103 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
104 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
105 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
106 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
107 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
108 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
109
110 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
111 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
112 w = (const void*) ((const uint8_t*) w + 8);
113
114 vacc0x0123 = _mm_maddd_epi16(
115 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
116 vacc1x0123 = _mm_maddd_epi16(
117 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
118
119 if (k > 2 * sizeof(uint8_t)) {
120 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
121 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
122 w = (const void*) ((const uint8_t*) w + 8);
123
124 vacc0x0123 = _mm_maddd_epi16(
125 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
126 vacc1x0123 = _mm_maddd_epi16(
127 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
128
129 if (k > 4 * sizeof(uint8_t)) {
130 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
131 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
132 w = (const void*) ((const uint8_t*) w + 8);
133
134 vacc0x0123 = _mm_maddd_epi16(
135 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
136 vacc1x0123 = _mm_maddd_epi16(
137 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
138 }
139 }
140 }
141
142 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
143 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
144
145 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
146 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
147 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
148
149 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
150 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
151 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
152
153 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
154 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
155
156 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
157 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
158
159 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
160
161 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
162
163 if (nc >= 4) {
164 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
165 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
166
167 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
168 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
169
170 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
171 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
172
173 nc -= 4;
174 } else {
175 if (nc & 2) {
176 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
177 c0 += 2;
178 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
179 c1 += 2;
180 vout = _mm_srli_epi32(vout, 16);
181 }
182 if (nc & 1) {
183 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
184 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
185 }
186
187 nc = 0;
188 }
189 } while (nc != 0);
190 }
191