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_1x4c2__xop_ld128(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_1x4c2__xop_ld128(
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 <= 1);
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
50 do {
51 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
52 w = (const void*) ((const int32_t*) w + 4);
53
54 size_t k = kc;
55 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
56 const __m128i vzero = _mm_setzero_si128();
57 while (k >= 8 * sizeof(uint8_t)) {
58 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
59 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
60 a0 += 8;
61
62 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
63 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
64 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
65
66 vacc0x0123 = _mm_maddd_epi16(
67 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
68
69 vacc0x0123 = _mm_maddd_epi16(
70 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
71 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
72 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
73 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
74
75 vacc0x0123 = _mm_maddd_epi16(
76 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
77
78 vacc0x0123 = _mm_maddd_epi16(
79 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
80
81 w = (const void*) ((const uint8_t*) w + 32);
82 k -= 8 * sizeof(uint8_t);
83 }
84 if (k != 0) {
85 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
86 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
87 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
88
89 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
90 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
91 w = (const void*) ((const uint8_t*) w + 8);
92
93 vacc0x0123 = _mm_maddd_epi16(
94 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
95
96 if (k > 2 * sizeof(uint8_t)) {
97 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
98 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
99 w = (const void*) ((const uint8_t*) w + 8);
100
101 vacc0x0123 = _mm_maddd_epi16(
102 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
103
104 if (k > 4 * sizeof(uint8_t)) {
105 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
106 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
107 w = (const void*) ((const uint8_t*) w + 8);
108
109 vacc0x0123 = _mm_maddd_epi16(
110 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
111 }
112 }
113 }
114
115 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
116
117 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
118 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
119
120 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
121 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
122
123 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
124
125 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
126 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
127
128 __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123);
129
130 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
131
132 if (nc >= 4) {
133 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
134
135 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
136
137 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
138
139 nc -= 4;
140 } else {
141 if (nc & 2) {
142 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
143 c0 += 2;
144 vout = _mm_srli_epi32(vout, 16);
145 }
146 if (nc & 1) {
147 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
148 }
149
150 nc = 0;
151 }
152 } while (nc != 0);
153 }
154