1 // Auto-generated file. Do not edit!
2 // Template: src/f16-vmulcaddc/fma3.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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 #include <immintrin.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/vmulcaddc.h>
16
17
xnn_f16_vmulcaddc_minmax_ukernel_c16__fma3_2x(size_t rows,size_t channels,const void * restrict input,size_t input_stride,const void * restrict weights,void * restrict output,size_t output_stride,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vmulcaddc_minmax_ukernel_c16__fma3_2x(
19 size_t rows,
20 size_t channels,
21 const void*restrict input,
22 size_t input_stride,
23 const void*restrict weights,
24 void*restrict output,
25 size_t output_stride,
26 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28 assert(rows != 0);
29 assert(channels != 0);
30 assert(channels % sizeof(uint16_t) == 0);
31
32 const uint16_t* i0 = (const uint16_t*) input;
33 uint16_t* o0 = (uint16_t*) output;
34 const uint16_t* i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
35 uint16_t* o1 = (uint16_t*) ((uintptr_t) o0 + output_stride);
36
37 const size_t input_increment = input_stride * 2 - channels;
38 const size_t output_increment = output_stride * 2 - channels;
39
40 const __m256 vmin = _mm256_load_ps(params->avx.min);
41 const __m256 vmax = _mm256_load_ps(params->avx.max);
42 do {
43 if XNN_UNPREDICTABLE(rows < 2) {
44 i1 = i0;
45 o1 = o0;
46 }
47
48 const uint16_t* w = (const uint16_t*) weights;
49 size_t c = channels;
50 for (; c >= 16 * sizeof(uint16_t); c -= 16 * sizeof(uint16_t)) {
51 const __m256 vscale01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) w));
52 const __m256 vscale89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 8)));
53
54 __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
55 __m256 vacc0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 8)));
56 i0 += 16;
57 __m256 vacc1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
58 __m256 vacc1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 8)));
59 i1 += 16;
60
61 const __m256 vbias01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 16)));
62 const __m256 vbias89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 24)));
63 w += 32;
64
65 vacc0x01234567 = _mm256_fmadd_ps(vacc0x01234567, vscale01234567, vbias01234567);
66 vacc0x89ABCDEF = _mm256_fmadd_ps(vacc0x89ABCDEF, vscale89ABCDEF, vbias89ABCDEF);
67 vacc1x01234567 = _mm256_fmadd_ps(vacc1x01234567, vscale01234567, vbias01234567);
68 vacc1x89ABCDEF = _mm256_fmadd_ps(vacc1x89ABCDEF, vscale89ABCDEF, vbias89ABCDEF);
69
70 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
71 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
72 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
73 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
74
75 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
76 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
77 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
78 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
79
80 _mm_storeu_si128((__m128i*) o0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
81 _mm_storeu_si128((__m128i*) (o0 + 8), _mm256_cvtps_ph(vacc0x89ABCDEF, _MM_FROUND_NO_EXC));
82 o0 += 16;
83 _mm_storeu_si128((__m128i*) o1, _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC));
84 _mm_storeu_si128((__m128i*) (o1 + 8), _mm256_cvtps_ph(vacc1x89ABCDEF, _MM_FROUND_NO_EXC));
85 o1 += 16;
86 }
87 for (; c >= 8 * sizeof(uint16_t); c -= 8 * sizeof(uint16_t)) {
88 const __m256 vscale = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) w));
89
90 __m256 vacc0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
91 i0 += 8;
92 __m256 vacc1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
93 i1 += 8;
94
95 const __m256 vbias = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 16)));
96 w += 8;
97
98 vacc0 = _mm256_fmadd_ps(vacc0, vscale, vbias);
99 vacc1 = _mm256_fmadd_ps(vacc1, vscale, vbias);
100
101 vacc0 = _mm256_max_ps(vacc0, vmin);
102 vacc1 = _mm256_max_ps(vacc1, vmin);
103
104 vacc0 = _mm256_min_ps(vacc0, vmax);
105 vacc1 = _mm256_min_ps(vacc1, vmax);
106
107 _mm_storeu_si128((__m128i*) o0, _mm256_cvtps_ph(vacc0, _MM_FROUND_NO_EXC));
108 o0 += 8;
109 _mm_storeu_si128((__m128i*) o1, _mm256_cvtps_ph(vacc1, _MM_FROUND_NO_EXC));
110 o1 += 8;
111 }
112 if XNN_UNLIKELY(c != 0) {
113 const __m256 vscale = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) w));
114
115 __m256 vacc0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
116 i0 = (const uint16_t*) ((uintptr_t) i0 + c);
117 __m256 vacc1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
118 i1 = (const uint16_t*) ((uintptr_t) i1 + c);
119
120 const __m256 vbias = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 16)));
121
122 vacc0 = _mm256_fmadd_ps(vacc0, vscale, vbias);
123 vacc1 = _mm256_fmadd_ps(vacc1, vscale, vbias);
124
125 vacc0 = _mm256_max_ps(vacc0, vmin);
126 vacc1 = _mm256_max_ps(vacc1, vmin);
127
128 vacc0 = _mm256_min_ps(vacc0, vmax);
129 vacc1 = _mm256_min_ps(vacc1, vmax);
130
131 __m128i vh0 = _mm256_cvtps_ph(vacc0, _MM_FROUND_NO_EXC);
132 __m128i vh1 = _mm256_cvtps_ph(vacc1, _MM_FROUND_NO_EXC);
133
134 if (c & (4 * sizeof(uint16_t))) {
135 _mm_storel_epi64((__m128i*) o0, vh0);
136 _mm_storel_epi64((__m128i*) o1, vh1);
137
138 vh0 = _mm_unpackhi_epi64(vh0, vh0);
139 vh1 = _mm_unpackhi_epi64(vh1, vh1);
140
141 o0 += 4;
142 o1 += 4;
143 }
144 if (c & (2 * sizeof(uint16_t))) {
145 *((uint32_t*) o0) = (uint32_t) _mm_cvtsi128_si32(vh0);
146 *((uint32_t*) o1) = (uint32_t) _mm_cvtsi128_si32(vh1);
147
148 vh0 = _mm_srli_epi64(vh0, 32);
149 vh1 = _mm_srli_epi64(vh1, 32);
150
151 o0 += 2;
152 o1 += 2;
153 }
154 if (c & (1 * sizeof(uint16_t))) {
155 *o0 = (uint16_t) _mm_extract_epi16(vh0, 0);
156 *o1 = (uint16_t) _mm_extract_epi16(vh1, 0);
157
158 o0 += 1;
159 o1 += 1;
160 }
161 }
162 i0 = (const uint16_t*) ((uintptr_t) i0 + input_increment);
163 o0 = (uint16_t*) ((uintptr_t) o0 + output_increment);
164 i1 = (const uint16_t*) ((uintptr_t) i1 + input_increment);
165 o1 = (uint16_t*) ((uintptr_t) o1 + output_increment);
166 rows = doz(rows, 2);
167 } while (rows != 0);
168 }
169