• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 = "0123456789ABCDEFGHIJKLMN"
8#include <assert.h>
9
10#include <psimd.h>
11
12#include <xnnpack/igemm.h>
13
14
15void xnn_f32_igemm_ukernel_${MR}x${NR}__psimd_splat(
16    size_t mr,
17    size_t nc,
18    size_t kc,
19    size_t ks,
20    const float**restrict a,
21    const float*restrict w,
22    float*restrict c,
23    size_t cm_stride,
24    size_t cn_stride,
25    size_t a_offset,
26    const float* zero,
27    const union xnn_f32_output_params params[restrict static 1])
28{
29  assert(mr != 0);
30  assert(mr <= ${MR});
31  assert(nc != 0);
32  assert(kc != 0);
33  assert(kc % sizeof(float) == 0);
34  assert(ks != 0);
35  assert(ks % (${MR} * sizeof(void*)) == 0);
36  assert(a_offset % sizeof(float) == 0);
37  assert(a != NULL);
38  assert(w != NULL);
39  assert(c != NULL);
40
41  float* c0 = c;
42  $for M in range(1, MR):
43    float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
44    $if M % 2 == 0:
45      if XNN_UNPREDICTABLE(mr <= ${M}) {
46        c${M} = c${M-1};
47      }
48    $elif M + 1 == MR:
49      if XNN_UNPREDICTABLE(mr != ${M+1}) {
50        c${M} = c${M-1};
51      }
52    $else:
53      if XNN_UNPREDICTABLE(mr < ${M+1}) {
54        c${M} = c${M-1};
55      }
56
57  do {
58    psimd_f32 vacc0x${ABC[0:4]} = psimd_load_f32(w);
59    $for N in range(4, NR, 4):
60      psimd_f32 vacc0x${ABC[N:N+4]} = psimd_load_f32(w + ${N});
61    $for M in range(1, MR):
62      $for N in range(0, NR, 4):
63        psimd_f32 vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
64    w += ${NR};
65
66    size_t p = ks;
67    do {
68      $for M in range(MR):
69        const float* restrict a${M} = a[${M}];
70        assert(a${M} != NULL);
71        if XNN_UNPREDICTABLE(a${M} != zero) {
72          a${M} = (const float*) ((uintptr_t) a${M} + a_offset);
73        }
74      a += ${MR};
75
76      size_t k = kc;
77      while (k >= 4 * sizeof(float)) {
78        $for M in range(MR):
79          const psimd_f32 va${M} = psimd_load_f32(a${M});
80          a${M} += 4;
81
82        $for L in range(4):
83          $for M in range(MR):
84            const psimd_f32 va${M}c${L} = psimd_splat${L}_f32(va${M});
85
86          $for N in range(0, NR, 4):
87            const psimd_f32 vb${ABC[N:N+4]}c${L} = psimd_load_f32(w + ${L * NR + N});
88
89          $for N in range(0, NR, 4):
90            $for M in range(MR):
91              vacc${M}x${ABC[N:N+4]} = psimd_qfma_f32(vacc${M}x${ABC[N:N+4]}, va${M}c${L}, vb${ABC[N:N+4]}c${L});
92
93        w += ${4 * NR};
94        k -= 4 * sizeof(float);
95      }
96      if XNN_UNLIKELY(k != 0) {
97        do {
98          const psimd_f32 vb${ABC[0:4]} = psimd_load_f32(w);
99          $for N in range(4, NR, 4):
100            const psimd_f32 vb${ABC[N:N+4]} = psimd_load_f32(w + ${N});
101          w += ${NR};
102
103          $for M in range(MR):
104            const psimd_f32 va${M} = psimd_load_splat_f32(a${M});
105            a${M} += 1;
106
107          $for M in range(MR):
108            $for N in range(0, NR, 4):
109              vacc${M}x${ABC[N:N+4]} = psimd_qfma_f32(vacc${M}x${ABC[N:N+4]}, va${M}, vb${ABC[N:N+4]});
110          k -= sizeof(float);
111        } while (k != 0);
112      }
113      p -= ${MR} * sizeof(void*);
114    } while (p != 0);
115
116    const psimd_f32 vmax = psimd_load_splat_f32(&params->scalar.max);
117    $for N in range(0, NR, 4):
118      $for M in range(MR):
119        vacc${M}x${ABC[N:N+4]} = psimd_min_f32(vacc${M}x${ABC[N:N+4]}, vmax);
120
121    const psimd_f32 vmin = psimd_load_splat_f32(&params->scalar.min);
122    $for N in range(0, NR, 4):
123      $for M in range(MR):
124        vacc${M}x${ABC[N:N+4]} = psimd_max_f32(vacc${M}x${ABC[N:N+4]}, vmin);
125
126    if XNN_LIKELY(nc >= ${NR}) {
127      $for M in reversed(range(MR)):
128        psimd_store_f32(c${M}, vacc${M}x${ABC[0:4]});
129        $for N in range(4, NR, 4):
130          psimd_store_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
131        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
132
133      a = (const float**restrict) ((uintptr_t) a - ks);
134      nc -= ${NR};
135    } else {
136      $for LOG2N in reversed(range(NR.bit_length())):
137        $if NR != 1 << LOG2N:
138          if (nc & ${1 << LOG2N}) {
139            $if LOG2N >= 2:
140              $for M in reversed(range(MR)):
141                psimd_store_f32(c${M}, vacc${M}x${ABC[0:4]});
142                $for N in range(4, 1 << LOG2N, 4):
143                  psimd_store_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
144
145              $for M in reversed(range(MR)):
146                $for N in range(0, 1 << (LOG2N - 1), 4):
147                  vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]};
148
149              $for M in reversed(range(MR)):
150                c${M} += ${1 << LOG2N};
151            $elif LOG2N == 1:
152              $for M in reversed(range(MR)):
153                psimd_store2_f32(c${M}, vacc${M}x${ABC[0:4]});
154
155              $for M in reversed(range(MR)):
156                vacc${M}x${ABC[0:4]} = psimd_concat_hi_f32(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]});
157
158              $for M in reversed(range(MR)):
159                c${M} += 2;
160            $elif LOG2N == 0:
161              $for M in reversed(range(MR)):
162                psimd_store1_f32(c${M}, vacc${M}x${ABC[0:4]});
163          }
164
165      nc = 0;
166    }
167  } while (nc != 0);
168}
169