xref: /aosp_15_r20/external/XNNPACK/src/f32-gemm/sse-shuffle.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$assert NR % 4 == 0
7$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
8#include <assert.h>
9
10#include <xmmintrin.h>
11
12#include <xnnpack/gemm.h>
13
14
15void xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}s4__sse(
16    size_t mr,
17    size_t nc,
18    size_t kc,
19    const float*restrict a,
20    size_t a_stride,
21    const float*restrict w,
22    float*restrict c,
23    size_t cm_stride,
24    size_t cn_stride,
25    $if INC:
26      const float*restrict acc,
27    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28{
29  assert(mr != 0);
30  assert(mr <= ${MR});
31  assert(nc != 0);
32  assert(kc != 0);
33  assert(kc % sizeof(float) == 0);
34  assert(a != NULL);
35  assert(w != NULL);
36  assert(c != NULL);
37  $if INC:
38    assert(acc != NULL);
39
40  const float* a0 = a;
41  float* c0 = c;
42  $for M in range(1, MR):
43    const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride);
44    float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
45    $if M % 2 == 0:
46      if XNN_UNPREDICTABLE(mr <= ${M}) {
47        a${M} = a${M-1};
48        c${M} = c${M-1};
49      }
50    $elif M + 1 == MR:
51      if XNN_UNPREDICTABLE(mr != ${M+1}) {
52        a${M} = a${M-1};
53        c${M} = c${M-1};
54      }
55    $else:
56      if XNN_UNPREDICTABLE(mr < ${M+1}) {
57        a${M} = a${M-1};
58        c${M} = c${M-1};
59      }
60
61  do {
62    $if INC:
63      $for M in range(MR):
64        $for N in range(0, NR, 4):
65          __m128 vacc${M}x${ABC[N:N+4]} = _mm_load_ps(acc + ${M*NR+N});
66      acc += ${MR*NR};
67    $else:
68      $for N in range(0, NR, 4):
69        __m128 vacc0x${ABC[N:N+4]} = _mm_load_ps(w + ${N});
70      $for M in range(1, MR):
71        $for N in range(0, NR, 4):
72          __m128 vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
73      w += ${NR};
74
75    size_t k = kc;
76    while (k >= 4 * sizeof(float)) {
77      $for M in range(MR):
78        __m128 va${M} = _mm_loadu_ps(a${M});
79        a${M} += 4;
80
81      $for L in range(4):
82
83        $for N in range(0, NR, 4):
84          const __m128 vb${ABC[N:N+4]}c${L} = _mm_load_ps(w + ${L * NR + N});
85
86        $for N in range(0, NR, 4):
87          $for M in range(MR):
88            vacc${M}x${ABC[N:N+4]} = _mm_add_ps(vacc${M}x${ABC[N:N+4]}, _mm_mul_ps(va${M}, vb${ABC[N:N+4]}c${L}));
89
90        $if L + 1 != 4:
91          $for M in range(MR):
92            va${M} = _mm_shuffle_ps(va${M}, va${M}, _MM_SHUFFLE(0, 3, 2, 1));
93
94      w += ${4 * NR};
95      k -= 4 * sizeof(float);
96    }
97    if XNN_UNLIKELY(k != 0) {
98      $for M in range(MR):
99        __m128 va${M} = _mm_loadu_ps(a${M});
100        a${M} = (const float*) ((uintptr_t) a${M} + k);
101
102      $for L in range(4):
103
104        $for N in range(0, NR, 4):
105          const __m128 vb${ABC[N:N+4]}c${L} = _mm_load_ps(w + ${L * NR + N});
106
107        $for N in range(0, NR, 4):
108          $for M in range(MR):
109            vacc${M}x${ABC[N:N+4]} = _mm_add_ps(vacc${M}x${ABC[N:N+4]}, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb${ABC[N:N+4]}c${L}), va${M}), vb${ABC[N:N+4]}c${L}));
110
111        $if L + 1 != 4:
112          $for M in range(MR):
113            va${M} = _mm_shuffle_ps(va${M}, va${M}, _MM_SHUFFLE(0, 3, 2, 1));
114
115      w += ${4 * NR};
116    }
117
118    const __m128 vmax = _mm_load_ps(params->sse.max);
119    $for N in range(0, NR, 4):
120      $for M in range(MR):
121        vacc${M}x${ABC[N:N+4]} = _mm_min_ps(vacc${M}x${ABC[N:N+4]}, vmax);
122
123    const __m128 vmin = _mm_load_ps(params->sse.min);
124    $for N in range(0, NR, 4):
125      $for M in range(MR):
126        vacc${M}x${ABC[N:N+4]} = _mm_max_ps(vacc${M}x${ABC[N:N+4]}, vmin);
127
128    if XNN_LIKELY(nc >= ${NR}) {
129      $for M in reversed(range(MR)):
130        _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]});
131        $for N in range(4, NR, 4):
132          _mm_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
133        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
134
135      $for M in reversed(range(MR)):
136        a${M} = (const float*) ((uintptr_t) a${M} - kc);
137
138      nc -= ${NR};
139    } else {
140      $for LOG2N in reversed(range(NR.bit_length())):
141        $if NR != 1 << LOG2N:
142          if (nc & ${1 << LOG2N}) {
143            $if LOG2N >= 2:
144              $for M in reversed(range(MR)):
145                _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]});
146                $for N in range(4, 1 << LOG2N, 4):
147                  _mm_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
148
149              $for M in reversed(range(MR)):
150                $for N in range(0, 1 << (LOG2N - 1), 4):
151                  vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]};
152
153              $for M in reversed(range(MR)):
154                c${M} += ${1 << LOG2N};
155            $elif LOG2N == 1:
156              $for M in reversed(range(MR)):
157                _mm_storel_pi((__m64*) c${M}, vacc${M}x${ABC[0:4]});
158
159              $for M in reversed(range(MR)):
160                vacc${M}x${ABC[0:4]} = _mm_movehl_ps(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]});
161
162              $for M in reversed(range(MR)):
163                c${M} += 2;
164            $elif LOG2N == 0:
165              $for M in reversed(range(MR)):
166                _mm_store_ss(c${M}, vacc${M}x${ABC[0:4]});
167          }
168
169      nc = 0;
170    }
171  } while (nc != 0);
172}
173