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 #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/igemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22
23
xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld128(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)])24 void xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld128(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 size_t ks,
29 const uint8_t** restrict a,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 size_t a_offset,
35 const uint8_t* zero,
36 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 1);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (1 * sizeof(void*)) == 0);
44 assert(a_offset % sizeof(uint8_t) == 0);
45 assert(a != NULL);
46 assert(w != NULL);
47 assert(c != NULL);
48
49 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
50 uint8_t* c0 = c;
51
52 do {
53 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
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 a += 1;
63
64 size_t k = kc;
65 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
66 const __m128i vzero = _mm_setzero_si128();
67 while (k >= 8 * sizeof(uint8_t)) {
68 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
69 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
70 a0 += 8;
71
72 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
73 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
74 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
75
76 vacc0x0123 = _mm_maddd_epi16(
77 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
78
79 vacc0x0123 = _mm_maddd_epi16(
80 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
81 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
82 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
83 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
84
85 vacc0x0123 = _mm_maddd_epi16(
86 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
87
88 vacc0x0123 = _mm_maddd_epi16(
89 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
90
91 w = (const void*) ((const uint8_t*) w + 32);
92 k -= 8 * sizeof(uint8_t);
93 }
94 if (k != 0) {
95 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
96 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
97 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
98
99 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
100 w = (const void*) ((const uint8_t*) w + 8);
101 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
102
103 vacc0x0123 = _mm_maddd_epi16(
104 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
105
106 if (k > 2 * sizeof(uint8_t)) {
107 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
108 w = (const void*) ((const uint8_t*) w + 8);
109 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
110
111 vacc0x0123 = _mm_maddd_epi16(
112 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
113
114 if (k > 4 * sizeof(uint8_t)) {
115 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
116 w = (const void*) ((const uint8_t*) w + 8);
117 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
118
119 vacc0x0123 = _mm_maddd_epi16(
120 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
121 }
122 }
123 }
124 p -= 1 * sizeof(void*);
125 } while (p != 0);
126
127 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
128
129 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
130 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
131
132 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
133 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
134
135 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
136
137 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
138 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
139
140 __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123);
141
142 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
143
144 if (nc >= 4) {
145 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
146 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
147
148 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
149
150 nc -= 4;
151 } else {
152 if (nc & 2) {
153 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
154 c0 += 2;
155 vout = _mm_srli_epi32(vout, 16);
156 }
157 if (nc & 1) {
158 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
159 }
160
161 nc = 0;
162 }
163 } while (nc != 0);
164 }
165