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 #include <xnnpack/unaligned.h>
22
23
24
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)])25 void xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64(
26 size_t mr,
27 size_t nc,
28 size_t kc,
29 const uint8_t* restrict a,
30 size_t a_stride,
31 const void* restrict w,
32 uint8_t* restrict c,
33 size_t cm_stride,
34 size_t cn_stride,
35 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 2);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(uint8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
47 const uint8_t* a0 = a;
48 uint8_t* c0 = c;
49 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
50 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr != 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55
56 do {
57 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
58 __m128i vacc1x0123 = vacc0x0123;
59 w = (const void*) ((const int32_t*) w + 4);
60
61 size_t k = kc;
62 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
63 while (k >= 8 * sizeof(uint8_t)) {
64 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
65 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
66 a0 += 8;
67 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
68 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
69 a1 += 8;
70
71 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
72 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
73
74 vacc0x0123 = _mm_maddd_epi16(
75 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
76 vacc1x0123 = _mm_maddd_epi16(
77 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
78 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
79 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
80
81 vacc0x0123 = _mm_maddd_epi16(
82 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
83 vacc1x0123 = _mm_maddd_epi16(
84 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
85 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
86 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
87
88 vacc0x0123 = _mm_maddd_epi16(
89 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
90 vacc1x0123 = _mm_maddd_epi16(
91 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
92 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
93 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
94
95 vacc0x0123 = _mm_maddd_epi16(
96 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
97 vacc1x0123 = _mm_maddd_epi16(
98 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
99
100 w = (const void*) ((const uint8_t*) w + 32);
101 k -= 8 * sizeof(uint8_t);
102 }
103 if (k != 0) {
104 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
105 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
106 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
107 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
108 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
109 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
110
111 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
112 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
113 w = (const void*) ((const uint8_t*) w + 8);
114
115 vacc0x0123 = _mm_maddd_epi16(
116 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
117 vacc1x0123 = _mm_maddd_epi16(
118 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
119
120 if (k > 2 * sizeof(uint8_t)) {
121 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
122 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
123 w = (const void*) ((const uint8_t*) w + 8);
124
125 vacc0x0123 = _mm_maddd_epi16(
126 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
127 vacc1x0123 = _mm_maddd_epi16(
128 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
129
130 if (k > 4 * sizeof(uint8_t)) {
131 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
132 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
133 w = (const void*) ((const uint8_t*) w + 8);
134
135 vacc0x0123 = _mm_maddd_epi16(
136 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
137 vacc1x0123 = _mm_maddd_epi16(
138 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
139 }
140 }
141 }
142
143 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
144 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
145
146 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
147 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
148 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
149
150 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
151 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
152 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
153
154 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
155 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
156
157 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
158 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
159
160 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
161
162 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
163
164 if (nc >= 4) {
165 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
166 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
167
168 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
169 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
170
171 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
172 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
173
174 nc -= 4;
175 } else {
176 if (nc & 2) {
177 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
178 c0 += 2;
179 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
180 c1 += 2;
181 vout = _mm_srli_epi32(vout, 16);
182 }
183 if (nc & 1) {
184 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
185 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
186 }
187
188 nc = 0;
189 }
190 } while (nc != 0);
191 }
192