xref: /aosp_15_r20/external/XNNPACK/src/f16-gemm/neonfp16arith-ld64.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1// Copyright 2020 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$assert NR % 8 == 0
7$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
8
9#include <assert.h>
10
11#include <arm_neon.h>
12
13#include <xnnpack/common.h>
14
15#include <xnnpack/gemm.h>
16
17
18void xnn_f16_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__neonfp16arith_ld64(
19    size_t mr,
20    size_t nc,
21    size_t kc,
22    const void* restrict a,
23    size_t a_stride,
24    const void* restrict w,
25    void* restrict c,
26    size_t cm_stride,
27    size_t cn_stride,
28    $if INC:
29      const void*restrict acc,
30    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31{
32  assert(mr != 0);
33  assert(mr <= ${MR});
34  assert(nc != 0);
35  assert(kc != 0);
36  assert(kc % sizeof(__fp16) == 0);
37  assert(a != NULL);
38  assert(w != NULL);
39  assert(c != NULL);
40  $if INC:
41    assert(acc != NULL);
42
43  const __fp16* a0 = (const __fp16*) a;
44  __fp16* c0 = (__fp16*) c;
45  $for M in range(1, MR):
46    const __fp16* a${M} = (const __fp16*) ((uintptr_t) a${M-1} + a_stride);
47    __fp16* c${M} = (__fp16*) ((uintptr_t) c${M-1} + cm_stride);
48    $if M % 2 == 0:
49      if XNN_UNPREDICTABLE(mr <= ${M}) {
50        a${M} = a${M-1};
51        c${M} = c${M-1};
52      }
53    $elif M + 1 == MR:
54      if XNN_UNPREDICTABLE(mr != ${M+1}) {
55        a${M} = a${M-1};
56        c${M} = c${M-1};
57      }
58    $else:
59      if XNN_UNPREDICTABLE(mr < ${M+1}) {
60        a${M} = a${M-1};
61        c${M} = c${M-1};
62      }
63
64  do {
65    $if INC:
66      $for M in range(MR):
67        $for N in range(0, NR, 8):
68          float16x8_t vacc${M}x${ABC[N:N+8]} = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
69    $else:
70      $for N in range(0, NR, 8):
71        float16x8_t vacc0x${ABC[N:N+8]} = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
72      $for M in range(1, MR):
73        $for N in range(0, NR, 8):
74          float16x8_t vacc${M}x${ABC[N:N+8]} = vacc0x${ABC[N:N+8]};
75
76    size_t k = kc;
77    while (k >= 4 * sizeof(__fp16)) {
78      $for M in range(MR):
79        const float16x4_t va${M} = vld1_f16(a${M}); a${M} += 4;
80
81      $for L in range(4):
82        $for N in range(0, NR, 8):
83          const float16x8_t vb${ABC[N:N+8]}c${L} = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
84
85        #if XNN_ARCH_ARM64
86          $for N in range(0, NR, 8):
87            $for M in range(MR):
88              vacc${M}x${ABC[N:N+8]} = vfmaq_lane_f16(vacc${M}x${ABC[N:N+8]}, vb${ABC[N:N+8]}c${L}, va${M}, ${L});
89        #else
90          $for M in range(MR):
91            const float16x8_t va${M}c${L} = vdupq_lane_f16(va${M}, ${L});
92
93          $for N in range(0, NR, 8):
94            $for M in range(MR):
95              vacc${M}x${ABC[N:N+8]} = vfmaq_f16(vacc${M}x${ABC[N:N+8]}, va${M}c${L}, vb${ABC[N:N+8]}c${L});
96        #endif
97
98      k -= 4 * sizeof(__fp16);
99    }
100    if XNN_UNLIKELY(k != 0) {
101      do {
102        $for M in range(MR):
103          const float16x8_t va${M} = vld1q_dup_f16(a${M}); a${M} += 1;
104
105        $for N in range(0, NR, 8):
106          const float16x8_t vb${ABC[N:N+8]} = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
107
108        $for N in range(0, NR, 8):
109          $for M in range(MR):
110            vacc${M}x${ABC[N:N+8]} = vfmaq_f16(vacc${M}x${ABC[N:N+8]}, va${M}, vb${ABC[N:N+8]});
111
112        k -= sizeof(__fp16);
113      } while (k != 0);
114    }
115
116
117    const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.max));
118    $for N in range(0, NR, 8):
119      $for M in range(MR):
120        vacc${M}x${ABC[N:N+8]} = vminq_f16(vacc${M}x${ABC[N:N+8]}, vmax);
121
122    const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.min));
123    $for N in range(0, NR, 8):
124      $for M in range(MR):
125        vacc${M}x${ABC[N:N+8]} = vmaxq_f16(vacc${M}x${ABC[N:N+8]}, vmin);
126
127    if XNN_LIKELY(nc >= ${NR}) {
128      $for M in range(MR):
129        vst1q_f16(c${M}, vacc${M}x${ABC[0:8]});
130        $for N in range(8, NR, 8):
131          vst1q_f16(c${M} + ${N}, vacc${M}x${ABC[N:N+8]});
132        c${M} = (__fp16*) ((uintptr_t) c${M} + cn_stride);
133
134      $for M in range(MR):
135        a${M} = (const __fp16*) ((uintptr_t) a${M} - kc);
136
137      nc -= ${NR};
138    } else {
139      $for LOG2N in reversed(range(NR.bit_length())):
140        $if NR != 1 << LOG2N:
141          if (nc & ${1 << LOG2N}) {
142            $if LOG2N >= 3:
143              $for N in range(0, 1 << LOG2N, 8):
144                $for M in range(MR):
145                  vst1q_f16(c${M}, vacc${M}x${ABC[N:N+8]}); c${M} += 8;
146
147              $for M in range(MR):
148                $for N in range(0, 1 << (LOG2N - 1), 8):
149                  vacc${M}x${ABC[N:N+8]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+8]};
150            $elif LOG2N == 2:
151              $for M in range(MR):
152                vst1_f16(c${M}, vacc${M}x${ABC[0:4]}); c${M} += 4;
153
154              $for M in range(MR):
155                vacc${M}x${ABC[0:4]} = vget_high_f16(vacc${M}x${ABC[0:8]});
156            $elif LOG2N == 1:
157              $for M in range(MR):
158                vst1_lane_u32((void*) c${M}, vreinterpret_u32_f16(vacc${M}x${ABC[0:4]}), 0); c${M} += 2;
159
160              $for M in range(MR):
161                vacc${M}x${ABC[0:4]} = vext_f16(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}, 2);
162            $elif LOG2N == 0:
163              $for M in range(MR):
164                vst1_lane_f16(c${M}, vacc${M}x${ABC[0:4]}, 0);
165          }
166        $if LOG2N == 3:
167          $for M in range(MR):
168            float16x4_t vacc${M}x${ABC[0:4]} = vget_low_f16(vacc${M}x${ABC[0:8]});
169
170      nc = 0;
171    }
172  } while (nc != 0);
173}
174