xref: /aosp_15_r20/external/XNNPACK/src/f32-prelu/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$assert ROW_TILE >= 1
9$assert SSE in [1, 2, 4]
10$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
11$SSE_HEADER = {1: "xmmintrin.h", 2: "emmintrin.h", 4: "smmintrin.h"}[SSE]
12#include <assert.h>
13
14#include <${SSE_HEADER}>
15
16#include <xnnpack/math.h>
17#include <xnnpack/prelu.h>
18
19
20$ISA = {1: "sse", 2: "sse2", 4: "sse41"}[SSE]
21void xnn_f32_prelu_ukernel__${ISA}_${ROW_TILE}x${CHANNEL_TILE}(
22    size_t rows,
23    size_t channels,
24    const float*restrict input,
25    size_t input_stride,
26    const float*restrict weights,
27    float*restrict output,
28    size_t output_stride) XNN_OOB_READS
29{
30  assert(rows != 0);
31  assert(channels != 0);
32  assert(channels % sizeof(float) == 0);
33
34  const float* i0 = input;
35  float* o0 = output;
36  $for M in range(1, ROW_TILE):
37    const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_stride);
38    float* o${M} = (float*) ((uintptr_t) o${M-1} + output_stride);
39
40  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
41  const size_t output_increment = output_stride * ${ROW_TILE} - channels;
42
43  $if SSE == 1:
44    const __m128 vzero = _mm_setzero_ps();
45  do {
46    $for M in range(1, ROW_TILE):
47      $if M % 2 == 0:
48        if XNN_UNPREDICTABLE(rows <= ${M}) {
49          i${M} = i${M-1};
50          o${M} = o${M-1};
51        }
52      $else:
53        if XNN_UNPREDICTABLE(rows < ${M+1}) {
54          i${M} = i${M-1};
55          o${M} = o${M-1};
56        }
57
58    const float* w = weights;
59    size_t c = channels;
60    for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
61      const __m128 vw${ABC[0:4]} = _mm_load_ps(w);
62      $for C in range(4, CHANNEL_TILE, 4):
63        const __m128 vw${ABC[C:C+4]} = _mm_load_ps(w + ${C});
64      w += ${CHANNEL_TILE};
65
66      $for M in range(ROW_TILE):
67        $if SSE == 1:
68          __m128 vi${M}x${ABC[0:4]} = _mm_loadu_ps(i${M});
69          $for C in range(4, CHANNEL_TILE, 4):
70            __m128 vi${M}x${ABC[C:C+4]} = _mm_loadu_ps(i${M} + ${C});
71        $else:
72          const __m128 vi${M}x${ABC[0:4]} = _mm_loadu_ps(i${M});
73          $for C in range(4, CHANNEL_TILE, 4):
74            const __m128 vi${M}x${ABC[C:C+4]} = _mm_loadu_ps(i${M} + ${C});
75        i${M} += ${CHANNEL_TILE};
76
77      $for M in range(ROW_TILE):
78        $for C in range(0, CHANNEL_TILE, 4):
79          $if SSE == 1:
80            __m128 vacc${M}x${ABC[C:C+4]} = _mm_max_ps(_mm_setzero_ps(), vi${M}x${ABC[C:C+4]});
81            vi${M}x${ABC[C:C+4]} = _mm_min_ps(vi${M}x${ABC[C:C+4]}, vzero);
82          $else:
83            const __m128 vprod${M}x${ABC[C:C+4]} = _mm_mul_ps(vi${M}x${ABC[C:C+4]}, vw${ABC[C:C+4]});
84            $if SSE == 2:
85              const __m128 vmask${M}x${ABC[C:C+4]} = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi${M}x${ABC[C:C+4]})));
86
87      $for M in range(ROW_TILE):
88        $for C in range(0, CHANNEL_TILE, 4):
89          $if SSE == 1:
90            vacc${M}x${ABC[C:C+4]} = _mm_add_ps(vacc${M}x${ABC[C:C+4]}, _mm_mul_ps(vi${M}x${ABC[C:C+4]}, vw${ABC[C:C+4]}));
91          $elif SSE == 2:
92            const __m128 vacc${M}x${ABC[C:C+4]} = _mm_or_ps(_mm_and_ps(vprod${M}x${ABC[C:C+4]}, vmask${M}x${ABC[C:C+4]}), _mm_andnot_ps(vmask${M}x${ABC[C:C+4]}, vi${M}x${ABC[C:C+4]}));
93          $elif SSE == 4:
94            const __m128 vacc${M}x${ABC[C:C+4]} = _mm_blendv_ps(vi${M}x${ABC[C:C+4]}, vprod${M}x${ABC[C:C+4]}, vi${M}x${ABC[C:C+4]});
95
96      $for M in range(ROW_TILE):
97        _mm_storeu_ps(o${M}, vacc${M}x${ABC[0:4]});
98        $for C in range(4, CHANNEL_TILE, 4):
99          _mm_storeu_ps(o${M} + ${C}, vacc${M}x${ABC[C:C+4]});
100        o${M} += ${CHANNEL_TILE};
101    }
102    $if CHANNEL_TILE > 4:
103      for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
104        const __m128 vw0123 = _mm_load_ps(w);
105        w += 4;
106
107        $for M in range(ROW_TILE):
108          $if SSE == 1:
109            __m128 vi${M}x0123 = _mm_loadu_ps(i${M});
110          $else:
111            const __m128 vi${M}x0123 = _mm_loadu_ps(i${M});
112          i${M} += 4;
113
114        $for M in range(ROW_TILE):
115          $if SSE == 1:
116            __m128 vacc${M}x0123 = _mm_max_ps(_mm_setzero_ps(), vi${M}x0123);
117            vi${M}x0123 = _mm_min_ps(vi${M}x0123, vzero);
118          $else:
119            const __m128 vprod${M}x0123 = _mm_mul_ps(vi${M}x0123, vw0123);
120            $if SSE == 2:
121              const __m128 vmask${M}x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi${M}x0123)));
122
123        $for M in range(ROW_TILE):
124          $if SSE == 1:
125            vacc${M}x0123 = _mm_add_ps(vacc${M}x0123, _mm_mul_ps(vi${M}x0123, vw0123));
126          $elif SSE == 2:
127            __m128 vacc${M}x0123 = _mm_or_ps(_mm_and_ps(vprod${M}x0123, vmask${M}x0123), _mm_andnot_ps(vmask${M}x0123, vi${M}x0123));
128          $elif SSE == 4:
129            __m128 vacc${M}x0123 = _mm_blendv_ps(vi${M}x0123, vprod${M}x0123, vi${M}x0123);
130
131        $for M in range(ROW_TILE):
132          _mm_storeu_ps(o${M}, vacc${M}x0123);
133          o${M} += 4;
134      }
135    if XNN_UNLIKELY(c != 0) {
136      const __m128 vw0123 = _mm_load_ps(w);
137      w = (const float*) ((uintptr_t) w + c);
138
139      $for M in range(ROW_TILE):
140        $if SSE == 1:
141          __m128 vi${M}x0123 = _mm_loadu_ps(i${M});
142        $else:
143          const __m128 vi${M}x0123 = _mm_loadu_ps(i${M});
144        i${M} = (const float*) ((uintptr_t) i${M} + c);
145
146      $for M in range(ROW_TILE):
147        $if SSE == 1:
148          __m128 vacc${M}x0123 = _mm_max_ps(_mm_setzero_ps(), vi${M}x0123);
149          vi${M}x0123 = _mm_min_ps(vi${M}x0123, vzero);
150        $else:
151          const __m128 vprod${M}x0123 = _mm_mul_ps(vi${M}x0123, vw0123);
152          $if SSE == 2:
153            const __m128 vmask${M}x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi${M}x0123)));
154
155      $for M in range(ROW_TILE):
156        $if SSE == 1:
157          vacc${M}x0123 = _mm_add_ps(vacc${M}x0123, _mm_mul_ps(vi${M}x0123, vw0123));
158        $elif SSE == 2:
159          __m128 vacc${M}x0123 = _mm_or_ps(_mm_and_ps(vprod${M}x0123, vmask${M}x0123), _mm_andnot_ps(vmask${M}x0123, vi${M}x0123));
160        $elif SSE == 4:
161          __m128 vacc${M}x0123 = _mm_blendv_ps(vi${M}x0123, vprod${M}x0123, vi${M}x0123);
162
163      if (c & (2 * sizeof(float))) {
164        $for M in range(ROW_TILE):
165          _mm_storel_pi((__m64*) o${M}, vacc${M}x0123);
166
167        $for M in range(ROW_TILE):
168          vacc${M}x0123 = _mm_movehl_ps(vacc${M}x0123, vacc${M}x0123);
169
170        $for M in range(ROW_TILE):
171          o${M} += 2;
172      }
173      if (c & (1 * sizeof(float))) {
174        $for M in range(ROW_TILE):
175          _mm_store_ss(o${M}, vacc${M}x0123);
176
177        $for M in range(ROW_TILE):
178          o${M} += 1;
179      }
180    }
181    $for M in range(ROW_TILE):
182      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
183      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
184    rows = doz(rows, ${ROW_TILE});
185  } while (rows != 0);
186}
187