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_qc8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])25 void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128(
26 size_t mr,
27 size_t nc,
28 size_t kc,
29 const int8_t* restrict a,
30 size_t a_stride,
31 const void* restrict w,
32 int8_t* restrict c,
33 size_t cm_stride,
34 size_t cn_stride,
35 const union xnn_qc8_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(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 2 * sizeof(int8_t));
47 const int8_t* a0 = a;
48 int8_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 while (k >= 8 * sizeof(int8_t)) {
56 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
57 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
58 a0 += 8;
59
60 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
61 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
62 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
63
64 vacc0x0123 = _mm_maddd_epi16(
65 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
66
67 vacc0x0123 = _mm_maddd_epi16(
68 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
69 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
70 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
71 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
72
73 vacc0x0123 = _mm_maddd_epi16(
74 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
75
76 vacc0x0123 = _mm_maddd_epi16(
77 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
78
79 w = (const void*) ((const int8_t*) w + 32);
80 k -= 8 * sizeof(int8_t);
81 }
82 if (k != 0) {
83 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
84 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
85 a0 = (const int8_t*) ((uintptr_t) a0 + k);
86
87 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
88 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
89 w = (const void*) ((const int8_t*) w + 8);
90
91 vacc0x0123 = _mm_maddd_epi16(
92 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
93
94 if (k > 2 * sizeof(int8_t)) {
95 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
96 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
97 w = (const void*) ((const int8_t*) w + 8);
98
99 vacc0x0123 = _mm_maddd_epi16(
100 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
101
102 if (k > 4 * sizeof(int8_t)) {
103 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
104 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
105 w = (const void*) ((const int8_t*) w + 8);
106
107 vacc0x0123 = _mm_maddd_epi16(
108 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
109 }
110 }
111 }
112
113 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
114
115 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
116 w = (const void*) ((const float*) w + 4);
117 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
118
119 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
120 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
121
122 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
123
124 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
125 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
126
127
128 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
129
130 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
131
132 if (nc >= 4) {
133 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
134
135 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
136
137 a0 = (const int8_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 = (int8_t) _mm_extract_epi8(vout, 0);
148 }
149
150 nc = 0;
151 }
152 } while (nc != 0);
153 }
154