xref: /aosp_15_r20/external/XNNPACK/src/f16-vmulcaddc/neonfp16arith.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1// Copyright 2020 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6$assert CHANNEL_TILE % 8 == 0
7$assert CHANNEL_TILE >= 8
8$assert ROW_TILE >= 1
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11
12#include <arm_neon.h>
13
14#include <xnnpack/math.h>
15#include <xnnpack/vmulcaddc.h>
16
17
18void xnn_f16_vmulcaddc_minmax_ukernel_c${CHANNEL_TILE}__neonfp16arith_${ROW_TILE}x(
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(__fp16) == 0);
31
32  const __fp16* i0 = (const __fp16*) input;
33  __fp16* o0 = (__fp16*) output;
34  $for M in range(1, ROW_TILE):
35    const __fp16* i${M} = (const __fp16*) ((uintptr_t) i${M-1} + input_stride);
36    __fp16* o${M} = (__fp16*) ((uintptr_t) o${M-1} + output_stride);
37
38  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
39  const size_t output_increment = output_stride * ${ROW_TILE} - channels;
40
41  const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.min));
42  const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.max));
43  do {
44    $for M in range(1, ROW_TILE):
45      $if M % 2 == 0:
46        if XNN_UNPREDICTABLE(rows <= ${M}) {
47          i${M} = i${M-1};
48          o${M} = o${M-1};
49        }
50      $else:
51        if XNN_UNPREDICTABLE(rows < ${M+1}) {
52          i${M} = i${M-1};
53          o${M} = o${M-1};
54        }
55
56    const __fp16* w = (const __fp16*) weights;
57    size_t c = channels;
58    $if CHANNEL_TILE > 8:
59      for (; c >= ${CHANNEL_TILE} * sizeof(__fp16); c -= ${CHANNEL_TILE} * sizeof(__fp16)) {
60        $for C in range(0, CHANNEL_TILE, 8):
61          const float16x8_t vscale${ABC[C:C+8]} = vld1q_f16(w); w += 8;
62
63        $for M in range(ROW_TILE):
64          $for C in range(0, CHANNEL_TILE, 8):
65            float16x8_t vacc${M}x${ABC[C:C+8]} = vld1q_f16(i${M}); i${M} += 8;
66
67        $for C in range(0, CHANNEL_TILE, 8):
68          const float16x8_t vbias${ABC[C:C+8]} = vld1q_f16(w); w += 8;
69
70        $for M in range(ROW_TILE):
71          $for C in range(0, CHANNEL_TILE, 8):
72            vacc${M}x${ABC[C:C+8]} = vfmaq_f16(vbias${ABC[C:C+8]}, vscale${ABC[C:C+8]}, vacc${M}x${ABC[C:C+8]});
73
74        $for M in range(ROW_TILE):
75          $for C in range(0, CHANNEL_TILE, 8):
76            vacc${M}x${ABC[C:C+8]} = vmaxq_f16(vacc${M}x${ABC[C:C+8]}, vmin);
77
78        $for M in range(ROW_TILE):
79          $for C in range(0, CHANNEL_TILE, 8):
80            vacc${M}x${ABC[C:C+8]} = vminq_f16(vacc${M}x${ABC[C:C+8]}, vmax);
81
82        $for M in range(ROW_TILE):
83          $for C in range(0, CHANNEL_TILE, 8):
84            vst1q_f16(o${M}, vacc${M}x${ABC[C:C+8]}); o${M} += 8;
85      }
86    for (; c >= 8 * sizeof(__fp16); c -= 8 * sizeof(__fp16)) {
87      const float16x8_t vscale01234567 = vld1q_f16(w);
88
89      $for M in range(ROW_TILE):
90        float16x8_t vacc${M}x01234567 = vld1q_f16(i${M}); i${M} += 8;
91
92      const float16x8_t vbias01234567 = vld1q_f16(w + ${CHANNEL_TILE});
93      w += ${8 if CHANNEL_TILE > 8 else CHANNEL_TILE * 2};
94
95      $for M in range(ROW_TILE):
96        vacc${M}x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc${M}x01234567);
97
98      $for M in range(ROW_TILE):
99        vacc${M}x01234567 = vmaxq_f16(vacc${M}x01234567, vmin);
100
101      $for M in range(ROW_TILE):
102        vacc${M}x01234567 = vminq_f16(vacc${M}x01234567, vmax);
103
104      $for M in range(ROW_TILE):
105        vst1q_f16(o${M}, vacc${M}x01234567); o${M} += 8;
106    }
107    if XNN_UNLIKELY(c != 0) {
108      const float16x8_t vscale01234567 = vld1q_f16(w);
109
110      $for M in range(ROW_TILE):
111        float16x8_t vacc${M}x01234567 = vld1q_f16(i${M}); i${M} = (const __fp16*) ((uintptr_t) i${M} + c);
112
113      const float16x8_t vbias01234567 = vld1q_f16(w + ${CHANNEL_TILE});
114
115      $for M in range(ROW_TILE):
116        vacc${M}x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc${M}x01234567);
117
118      $for M in range(ROW_TILE):
119        vacc${M}x01234567 = vmaxq_f16(vacc${M}x01234567, vmin);
120
121      $for M in range(ROW_TILE):
122        vacc${M}x01234567 = vminq_f16(vacc${M}x01234567, vmax);
123
124      $for M in range(ROW_TILE):
125        float16x4_t vacc${M}x0123 = vget_low_f16(vacc${M}x01234567);
126      if (c & (4 * sizeof(__fp16))) {
127        $for M in range(ROW_TILE):
128          vst1_f16(o${M}, vacc${M}x0123); o${M} += 4;
129
130        $for M in range(ROW_TILE):
131          vacc${M}x0123 = vget_high_f16(vacc${M}x01234567);
132      }
133      if (c & (2 * sizeof(__fp16))) {
134        $for M in range(ROW_TILE):
135          vst1_lane_u32((void*) o${M}, vreinterpret_u32_f16(vacc${M}x0123), 0); o${M} += 2;
136
137        $for M in range(ROW_TILE):
138          vacc${M}x0123 = vext_f16(vacc${M}x0123, vacc${M}x0123, 2);
139      }
140      if (c & (1 * sizeof(__fp16))) {
141        $for M in range(ROW_TILE):
142          vst1_lane_f16(o${M}, vacc${M}x0123, 0); o${M} += 1;
143      }
144    }
145    $for M in range(ROW_TILE):
146      i${M} = (const __fp16*) ((uintptr_t) i${M} + input_increment);
147      o${M} = (__fp16*) ((uintptr_t) o${M} + output_increment);
148    rows = doz(rows, ${ROW_TILE});
149  } while (rows != 0);
150}
151