• 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}s4__psimd(
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        $for M in range(MR):
79          psimd_f32 va${M} = psimd_load_f32(a${M});
80          a${M} += 4;
81
82        $for L in range(4):
83
84          $for N in range(0, NR, 4):
85            const psimd_f32 vb${ABC[N:N+4]}c${L} = psimd_load_f32(w + ${L * NR + N});
86
87          $for N in range(0, NR, 4):
88            $for M in range(MR):
89              vacc${M}x${ABC[N:N+4]} = psimd_qfma_f32(vacc${M}x${ABC[N:N+4]}, va${M}, vb${ABC[N:N+4]}c${L});
90
91          $if L + 1 != 4:
92            #ifdef __clang__
93            $for M in range(MR):
94              va${M} = __builtin_shufflevector(va${M}, va${M}, 1, 2, 3, 0);
95            #else
96            $for M in range(MR):
97              va${M} = __builtin_shuffle(va${M}, va${M}, (psimd_s32) { 1, 2, 3, 0 });
98            #endif
99
100        w += ${4 * NR};
101        k -= 4 * sizeof(float);
102      }
103      if XNN_UNLIKELY(k != 0) {
104        do {
105          const psimd_f32 vb${ABC[0:4]} = psimd_load_f32(w);
106          $for N in range(4, NR, 4):
107            const psimd_f32 vb${ABC[N:N+4]} = psimd_load_f32(w + ${N});
108          w += ${NR};
109
110          $for M in range(MR):
111            const psimd_f32 va${M} = psimd_load_splat_f32(a${M});
112            a${M} += 1;
113
114          $for M in range(MR):
115            $for N in range(0, NR, 4):
116              vacc${M}x${ABC[N:N+4]} = psimd_qfma_f32(vacc${M}x${ABC[N:N+4]}, va${M}, vb${ABC[N:N+4]});
117          k -= sizeof(float);
118        } while (k != 0);
119      }
120      p -= ${MR} * sizeof(void*);
121    } while (p != 0);
122
123    const psimd_f32 vmax = psimd_load_splat_f32(&params->scalar.max);
124    $for N in range(0, NR, 4):
125      $for M in range(MR):
126        vacc${M}x${ABC[N:N+4]} = psimd_min_f32(vacc${M}x${ABC[N:N+4]}, vmax);
127
128    const psimd_f32 vmin = psimd_load_splat_f32(&params->scalar.min);
129    $for N in range(0, NR, 4):
130      $for M in range(MR):
131        vacc${M}x${ABC[N:N+4]} = psimd_max_f32(vacc${M}x${ABC[N:N+4]}, vmin);
132
133    if XNN_LIKELY(nc >= ${NR}) {
134      $for M in reversed(range(MR)):
135        psimd_store_f32(c${M}, vacc${M}x${ABC[0:4]});
136        $for N in range(4, NR, 4):
137          psimd_store_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
138        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
139
140      a = (const float**restrict) ((uintptr_t) a - ks);
141      nc -= ${NR};
142    } else {
143      $for LOG2N in reversed(range(NR.bit_length())):
144        $if NR != 1 << LOG2N:
145          if (nc & ${1 << LOG2N}) {
146            $if LOG2N >= 2:
147              $for M in reversed(range(MR)):
148                psimd_store_f32(c${M}, vacc${M}x${ABC[0:4]});
149                $for N in range(4, 1 << LOG2N, 4):
150                  psimd_store_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
151
152              $for M in reversed(range(MR)):
153                $for N in range(0, 1 << (LOG2N - 1), 4):
154                  vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]};
155
156              $for M in reversed(range(MR)):
157                c${M} += ${1 << LOG2N};
158            $elif LOG2N == 1:
159              $for M in reversed(range(MR)):
160                psimd_store2_f32(c${M}, vacc${M}x${ABC[0:4]});
161
162              $for M in reversed(range(MR)):
163                vacc${M}x${ABC[0:4]} = psimd_concat_hi_f32(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]});
164
165              $for M in reversed(range(MR)):
166                c${M} += 2;
167            $elif LOG2N == 0:
168              $for M in reversed(range(MR)):
169                psimd_store1_f32(c${M}, vacc${M}x${ABC[0:4]});
170          }
171
172      nc = 0;
173    }
174  } while (nc != 0);
175}
176