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_c8__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_c8__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 >= 8 * sizeof(uint16_t); c -= 8 * sizeof(uint16_t)) {
51 const __m256 vscale = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) w));
52
53 __m256 vacc0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
54 i0 += 8;
55 __m256 vacc1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
56 i1 += 8;
57
58 const __m256 vbias = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 8)));
59 w += 16;
60
61 vacc0 = _mm256_fmadd_ps(vacc0, vscale, vbias);
62 vacc1 = _mm256_fmadd_ps(vacc1, vscale, vbias);
63
64 vacc0 = _mm256_max_ps(vacc0, vmin);
65 vacc1 = _mm256_max_ps(vacc1, vmin);
66
67 vacc0 = _mm256_min_ps(vacc0, vmax);
68 vacc1 = _mm256_min_ps(vacc1, vmax);
69
70 _mm_storeu_si128((__m128i*) o0, _mm256_cvtps_ph(vacc0, _MM_FROUND_NO_EXC));
71 o0 += 8;
72 _mm_storeu_si128((__m128i*) o1, _mm256_cvtps_ph(vacc1, _MM_FROUND_NO_EXC));
73 o1 += 8;
74 }
75 if XNN_UNLIKELY(c != 0) {
76 const __m256 vscale = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) w));
77
78 __m256 vacc0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
79 i0 = (const uint16_t*) ((uintptr_t) i0 + c);
80 __m256 vacc1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
81 i1 = (const uint16_t*) ((uintptr_t) i1 + c);
82
83 const __m256 vbias = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 8)));
84
85 vacc0 = _mm256_fmadd_ps(vacc0, vscale, vbias);
86 vacc1 = _mm256_fmadd_ps(vacc1, vscale, vbias);
87
88 vacc0 = _mm256_max_ps(vacc0, vmin);
89 vacc1 = _mm256_max_ps(vacc1, vmin);
90
91 vacc0 = _mm256_min_ps(vacc0, vmax);
92 vacc1 = _mm256_min_ps(vacc1, vmax);
93
94 __m128i vh0 = _mm256_cvtps_ph(vacc0, _MM_FROUND_NO_EXC);
95 __m128i vh1 = _mm256_cvtps_ph(vacc1, _MM_FROUND_NO_EXC);
96
97 if (c & (4 * sizeof(uint16_t))) {
98 _mm_storel_epi64((__m128i*) o0, vh0);
99 _mm_storel_epi64((__m128i*) o1, vh1);
100
101 vh0 = _mm_unpackhi_epi64(vh0, vh0);
102 vh1 = _mm_unpackhi_epi64(vh1, vh1);
103
104 o0 += 4;
105 o1 += 4;
106 }
107 if (c & (2 * sizeof(uint16_t))) {
108 *((uint32_t*) o0) = (uint32_t) _mm_cvtsi128_si32(vh0);
109 *((uint32_t*) o1) = (uint32_t) _mm_cvtsi128_si32(vh1);
110
111 vh0 = _mm_srli_epi64(vh0, 32);
112 vh1 = _mm_srli_epi64(vh1, 32);
113
114 o0 += 2;
115 o1 += 2;
116 }
117 if (c & (1 * sizeof(uint16_t))) {
118 *o0 = (uint16_t) _mm_extract_epi16(vh0, 0);
119 *o1 = (uint16_t) _mm_extract_epi16(vh1, 0);
120
121 o0 += 1;
122 o1 += 1;
123 }
124 }
125 i0 = (const uint16_t*) ((uintptr_t) i0 + input_increment);
126 o0 = (uint16_t*) ((uintptr_t) o0 + output_increment);
127 i1 = (const uint16_t*) ((uintptr_t) i1 + input_increment);
128 o1 = (uint16_t*) ((uintptr_t) o1 + output_increment);
129 rows = doz(rows, 2);
130 } while (rows != 0);
131 }
132