xref: /aosp_15_r20/external/XNNPACK/src/f32-vmulcaddc/sse.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 <xmmintrin.h>
13
14#include <xnnpack/math.h>
15#include <xnnpack/vmulcaddc.h>
16
17
18void xnn_f32_vmulcaddc_minmax_ukernel_c${CHANNEL_TILE}__sse_${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 __m128 vmin = _mm_load_ps(params->sse.min);
42  const __m128 vmax = _mm_load_ps(params->sse.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 -= ${CHANNEL_TILE} * sizeof(float)) {
59      const __m128 vscale${ABC[0:4]} = _mm_load_ps(w);
60      $for C in range(4, CHANNEL_TILE, 4):
61        const __m128 vscale${ABC[C:C+4]} = _mm_load_ps(w + ${C});
62
63      $for M in range(ROW_TILE):
64        __m128 vacc${M}x${ABC[0:4]} = _mm_loadu_ps(i${M});
65        $for C in range(4, CHANNEL_TILE, 4):
66          __m128 vacc${M}x${ABC[C:C+4]} = _mm_loadu_ps(i${M} + ${C});
67        i${M} += ${CHANNEL_TILE};
68
69      $for M in range(ROW_TILE):
70        $for C in range(0, CHANNEL_TILE, 4):
71          vacc${M}x${ABC[C:C+4]} = _mm_mul_ps(vacc${M}x${ABC[C:C+4]}, vscale${ABC[C:C+4]});
72
73      $for C in range(0, CHANNEL_TILE, 4):
74        const __m128 vbias${ABC[C:C+4]} = _mm_load_ps(w + ${C + CHANNEL_TILE});
75
76      $for M in range(ROW_TILE):
77        $for C in range(0, CHANNEL_TILE, 4):
78          vacc${M}x${ABC[C:C+4]} = _mm_add_ps(vacc${M}x${ABC[C:C+4]}, vbias${ABC[C:C+4]});
79
80      $for M in range(ROW_TILE):
81        $for C in range(0, CHANNEL_TILE, 4):
82          vacc${M}x${ABC[C:C+4]} = _mm_max_ps(vacc${M}x${ABC[C:C+4]}, vmin);
83
84      $for M in range(ROW_TILE):
85        $for C in range(0, CHANNEL_TILE, 4):
86          vacc${M}x${ABC[C:C+4]} = _mm_min_ps(vacc${M}x${ABC[C:C+4]}, vmax);
87
88      $for M in range(ROW_TILE):
89        _mm_storeu_ps(o${M}, vacc${M}x${ABC[0:4]});
90        $for C in range(4, CHANNEL_TILE, 4):
91          _mm_storeu_ps(o${M} + ${C}, vacc${M}x${ABC[C:C+4]});
92        o${M} += ${CHANNEL_TILE};
93
94      w += ${CHANNEL_TILE * 2};
95    }
96    $if CHANNEL_TILE > 4:
97      for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
98        const __m128 vscale0123 = _mm_load_ps(w);
99
100        $for M in range(ROW_TILE):
101          __m128 vacc${M}x0123 = _mm_loadu_ps(i${M});
102          i${M} += 4;
103
104        $for M in range(ROW_TILE):
105          vacc${M}x0123 = _mm_mul_ps(vacc${M}x0123, vscale0123);
106
107        const __m128 vbias0123 = _mm_load_ps(w + ${CHANNEL_TILE});
108
109        $for M in range(ROW_TILE):
110          vacc${M}x0123 = _mm_add_ps(vacc${M}x0123, vbias0123);
111
112        $for M in range(ROW_TILE):
113          vacc${M}x0123 = _mm_max_ps(vacc${M}x0123, vmin);
114
115        $for M in range(ROW_TILE):
116          vacc${M}x0123 = _mm_min_ps(vacc${M}x0123, vmax);
117
118        $for M in range(ROW_TILE):
119          _mm_storeu_ps(o${M}, vacc${M}x0123);
120          o${M} += 4;
121
122        w += 4;
123      }
124    if XNN_UNLIKELY(c != 0) {
125      const __m128 vscale0123 = _mm_load_ps(w);
126
127      $for M in range(ROW_TILE):
128        __m128 vacc${M}x0123 = _mm_loadu_ps(i${M});
129        i${M} = (const float*) ((uintptr_t) i${M} + c);
130
131      $for M in range(ROW_TILE):
132        vacc${M}x0123 = _mm_mul_ps(vacc${M}x0123, vscale0123);
133
134      const __m128 vbias0123 = _mm_load_ps(w + ${CHANNEL_TILE});
135
136      $for M in range(ROW_TILE):
137        vacc${M}x0123 = _mm_add_ps(vacc${M}x0123, vbias0123);
138
139      $for M in range(ROW_TILE):
140        vacc${M}x0123 = _mm_max_ps(vacc${M}x0123, vmin);
141
142      $for M in range(ROW_TILE):
143        vacc${M}x0123 = _mm_min_ps(vacc${M}x0123, vmax);
144
145      if (c & (2 * sizeof(float))) {
146        $for M in range(ROW_TILE):
147          _mm_storel_pi((__m64*) o${M}, vacc${M}x0123);
148
149        $for M in range(ROW_TILE):
150          vacc${M}x0123 = _mm_movehl_ps(vacc${M}x0123, vacc${M}x0123);
151
152        $for M in range(ROW_TILE):
153          o${M} += 2;
154      }
155      if (c & (1 * sizeof(float))) {
156        $for M in range(ROW_TILE):
157          _mm_store_ss(o${M}, vacc${M}x0123);
158
159        $for M in range(ROW_TILE):
160          o${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