xref: /aosp_15_r20/external/XNNPACK/src/f32-vmulcaddc/neon.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1// Copyright 2019 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$assert CHANNEL_TILE % 4 == 0
7$assert CHANNEL_TILE >= 4
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_f32_vmulcaddc_minmax_ukernel_c${CHANNEL_TILE}__${"neonfma" if FMA else "neon"}_${ROW_TILE}x(
19    size_t rows,
20    size_t channels,
21    const float*restrict input,
22    size_t input_stride,
23    const float*restrict weights,
24    float*restrict output,
25    size_t output_stride,
26    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27{
28  assert(rows != 0);
29  assert(channels != 0);
30  assert(channels % sizeof(float) == 0);
31
32  const float* i0 = input;
33  float* o0 = output;
34  $for M in range(1, ROW_TILE):
35    const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_stride);
36    float* o${M} = (float*) ((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 float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
42  const float32x4_t vmax = vld1q_dup_f32(&params->scalar.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 float* w = weights;
57    size_t c = channels;
58    for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
59      $for C in range(0, CHANNEL_TILE, 4):
60        const float32x4_t vscale${ABC[C:C+4]} = vld1q_f32(w); w += 4;
61
62      $for M in range(ROW_TILE):
63        $for C in range(0, CHANNEL_TILE, 4):
64          float32x4_t vacc${M}x${ABC[C:C+4]} = vld1q_f32(i${M}); i${M} += 4;
65
66      $if not FMA:
67        $for M in range(ROW_TILE):
68          $for C in range(0, CHANNEL_TILE, 4):
69            vacc${M}x${ABC[C:C+4]} = vmulq_f32(vacc${M}x${ABC[C:C+4]}, vscale${ABC[C:C+4]});
70
71      $for C in range(0, CHANNEL_TILE, 4):
72        const float32x4_t vbias${ABC[C:C+4]} = vld1q_f32(w); w += 4;
73
74      $if not FMA:
75        $for M in range(ROW_TILE):
76          $for C in range(0, CHANNEL_TILE, 4):
77            vacc${M}x${ABC[C:C+4]} = vaddq_f32(vacc${M}x${ABC[C:C+4]}, vbias${ABC[C:C+4]});
78      $else:
79        $for M in range(ROW_TILE):
80          $for C in range(0, CHANNEL_TILE, 4):
81            vacc${M}x${ABC[C:C+4]} = vfmaq_f32(vbias${ABC[C:C+4]}, vscale${ABC[C:C+4]}, vacc${M}x${ABC[C:C+4]});
82
83      $for M in range(ROW_TILE):
84        $for C in range(0, CHANNEL_TILE, 4):
85          vacc${M}x${ABC[C:C+4]} = vmaxq_f32(vacc${M}x${ABC[C:C+4]}, vmin);
86
87      $for M in range(ROW_TILE):
88        $for C in range(0, CHANNEL_TILE, 4):
89          vacc${M}x${ABC[C:C+4]} = vminq_f32(vacc${M}x${ABC[C:C+4]}, vmax);
90
91      $for M in range(ROW_TILE):
92        $for C in range(0, CHANNEL_TILE, 4):
93          vst1q_f32(o${M}, vacc${M}x${ABC[C:C+4]}); o${M} += 4;
94    }
95    $if CHANNEL_TILE > 4:
96      for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
97        const float32x4_t vscale0123 = vld1q_f32(w); w += 4;
98
99        $for M in range(ROW_TILE):
100          float32x4_t vacc${M}x0123 = vld1q_f32(i${M}); i${M} += 4;
101
102        $if not FMA:
103          $for M in range(ROW_TILE):
104            vacc${M}x0123 = vmulq_f32(vacc${M}x0123, vscale0123);
105
106        const float32x4_t vbias0123 = vld1q_f32(w + ${CHANNEL_TILE - 4});
107
108        $if not FMA:
109          $for M in range(ROW_TILE):
110            vacc${M}x0123 = vaddq_f32(vacc${M}x0123, vbias0123);
111        $else:
112          $for M in range(ROW_TILE):
113            vacc${M}x0123 = vfmaq_f32(vbias0123, vscale0123, vacc${M}x0123);
114
115        $for M in range(ROW_TILE):
116          vacc${M}x0123 = vmaxq_f32(vacc${M}x0123, vmin);
117
118        $for M in range(ROW_TILE):
119          vacc${M}x0123 = vminq_f32(vacc${M}x0123, vmax);
120
121        $for M in range(ROW_TILE):
122          vst1q_f32(o${M}, vacc${M}x0123); o${M} += 4;
123      }
124    if XNN_UNLIKELY(c != 0) {
125      const float32x4_t vscale0123 = vld1q_f32(w);
126
127      $for M in range(ROW_TILE):
128        float32x4_t vacc${M}x0123 = vld1q_f32(i${M}); i${M} = (const float*) ((uintptr_t) i${M} + c);
129
130      $if not FMA:
131        $for M in range(ROW_TILE):
132          vacc${M}x0123 = vmulq_f32(vacc${M}x0123, vscale0123);
133
134      const float32x4_t vbias0123 = vld1q_f32(w + ${CHANNEL_TILE});
135
136      $if not FMA:
137        $for M in range(ROW_TILE):
138          vacc${M}x0123 = vaddq_f32(vacc${M}x0123, vbias0123);
139      $else:
140        $for M in range(ROW_TILE):
141          vacc${M}x0123 = vfmaq_f32(vbias0123, vscale0123, vacc${M}x0123);
142
143      $for M in range(ROW_TILE):
144        vacc${M}x0123 = vmaxq_f32(vacc${M}x0123, vmin);
145
146      $for M in range(ROW_TILE):
147        vacc${M}x0123 = vminq_f32(vacc${M}x0123, vmax);
148
149      $for M in range(ROW_TILE):
150        float32x2_t vacc${M}x01 = vget_low_f32(vacc${M}x0123);
151      if (c & (2 * sizeof(float))) {
152        $for M in range(ROW_TILE):
153          vst1_f32(o${M}, vacc${M}x01); o${M} += 2;
154
155        $for M in range(ROW_TILE):
156          vacc${M}x01 = vget_high_f32(vacc${M}x0123);
157      }
158      if (c & (1 * sizeof(float))) {
159        $for M in range(ROW_TILE):
160          vst1_lane_f32(o${M}, vacc${M}x01, 0); o${M} += 1;
161      }
162    }
163    $for M in range(ROW_TILE):
164      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
165      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
166    rows = doz(rows, ${ROW_TILE});
167  } while (rows != 0);
168}
169