xref: /aosp_15_r20/external/XNNPACK/src/f16-spmm/neonfp16arith.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$assert MR % 8 == 0
7$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
8#include <assert.h>
9
10#include <arm_neon.h>
11
12#include <xnnpack/spmm.h>
13
14
15void xnn_f16_spmm_minmax_ukernel_${MR}x${NR}__neonfp16arith${"_x%d" % UNROLL if UNROLL > 1 else ""}(
16    size_t mc,
17    size_t nc,
18    const void*restrict input,
19    const void*restrict weights,
20    const int32_t*restrict widx_dmap,
21    const uint32_t*restrict nidx_nnzmap,
22    void*restrict output,
23    size_t output_stride,
24    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
25{
26  assert(mc != 0);
27  assert(mc % sizeof(__fp16) == 0);
28  assert(nc != 0);
29
30  const __fp16*restrict i = (const __fp16*) input;
31  __fp16*restrict o = (__fp16*) output;
32
33  const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.max));
34  const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.min));
35
36  size_t output_decrement = output_stride * nc - ${MR} * sizeof(__fp16);
37  while XNN_LIKELY(mc >= ${MR} * sizeof(__fp16)) {
38    const __fp16*restrict w = (const __fp16*) weights;
39    const int32_t* dmap = widx_dmap;
40    const uint32_t* nnzmap = nidx_nnzmap;
41    size_t n = nc;
42    do {
43      uint32_t nnz = *nnzmap++;
44      $if UNROLL > 1:
45        float16x8_t vacc01234567x0 = vld1q_dup_f16(w); w += 1;
46        $for K in range(1, UNROLL):
47          float16x8_t vacc01234567x${K} = vmovq_n_f16(0.0f);
48        $for M in range(8, MR, 8):
49          float16x8_t vacc${ABC[M:M+8]}x0 = vacc01234567x0;
50          $for K in range(1, UNROLL):
51            float16x8_t vacc${ABC[M:M+8]}x${K} = vmovq_n_f16(0.0f);
52        for (; nnz >= ${UNROLL}; nnz -= ${UNROLL}) {
53          $for K in range(UNROLL):
54            const intptr_t diff${K} = dmap[${K}];
55          dmap += ${UNROLL};
56          $for K in range(UNROLL):
57            const float16x8_t va01234567x${K} = vld1q_f16(i);
58            $for M in range(8, MR, 8):
59              const float16x8_t va${ABC[M:M+8]}x${K} = vld1q_f16(i + ${M});
60            i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff${K});
61            const float16x8_t vb${K} = vld1q_dup_f16(w); w += 1;
62            $for M in range(0, MR, 8):
63              vacc${ABC[M:M+8]}x${K} = vfmaq_f16(vacc${ABC[M:M+8]}x${K}, va${ABC[M:M+8]}x${K}, vb${K});
64        }
65        $for M in range(0, MR, 8):
66          float16x8_t vacc${ABC[M:M+8]} = vacc${ABC[M:M+8]}x0;
67        $for K in range(1, UNROLL):
68          $for M in range(0, MR, 8):
69            vacc${ABC[M:M+8]} = vaddq_f16(vacc${ABC[M:M+8]}, vacc${ABC[M:M+8]}x${K});
70      $else:
71        float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
72        $for M in range(8, MR, 8):
73          float16x8_t vacc${ABC[M:M+8]} = vacc01234567;
74      if XNN_LIKELY(nnz != 0) {
75        do {
76          const intptr_t diff = *dmap++;
77          const float16x8_t va01234567 = vld1q_f16(i);
78          $for M in range(8, MR, 8):
79            const float16x8_t va${ABC[M:M+8]} = vld1q_f16(i + ${M});
80          i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
81          const float16x8_t vb = vld1q_dup_f16(w); w += 1;
82          $for M in range(0, MR, 8):
83            vacc${ABC[M:M+8]} = vfmaq_f16(vacc${ABC[M:M+8]}, va${ABC[M:M+8]}, vb);
84        } while (--nnz != 0);
85      }
86      $for M in range(0, MR, 8):
87        float16x8_t vout${ABC[M:M+8]} = vminq_f16(vacc${ABC[M:M+8]}, vmax);
88      $for M in range(0, MR, 8):
89        vout${ABC[M:M+8]} = vmaxq_f16(vout${ABC[M:M+8]}, vmin);
90      vst1q_f16(o, vout01234567);
91      $for M in range(8, MR, 8):
92        vst1q_f16(o + ${M}, vout${ABC[M:M+8]});
93      o = (__fp16*restrict) ((uintptr_t) o + output_stride);
94    } while (--n != 0);
95    o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
96    i += ${MR};
97    mc -= ${MR} * sizeof(__fp16);
98  }
99  if XNN_UNLIKELY(mc != 0) {
100    $for LOG2M in reversed(range((MR - 1).bit_length())):
101      $SUBMR = 1 << LOG2M
102      $if SUBMR * 2 >= MR:
103        output_decrement += ${MR - SUBMR} * sizeof(__fp16);
104      $else:
105        output_decrement += ${SUBMR} * sizeof(__fp16);
106      if (mc & (${SUBMR} * sizeof(__fp16))) {
107        const __fp16*restrict w = (const __fp16*) weights;
108        const int32_t* dmap = widx_dmap;
109        const uint32_t* nnzmap = nidx_nnzmap;
110        size_t n = nc;
111        do {
112          uint32_t nnz = *nnzmap++;
113          $if SUBMR <= 4:
114            float16x4_t vacc${ABC[0:SUBMR]} = vld1_dup_f16(w); w += 1;
115          $else:
116            float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
117          $for M in range(8, SUBMR, 8):
118            float16x8_t vacc${ABC[M:M+8]} = vacc01234567;
119          if XNN_LIKELY(nnz != 0) {
120            do {
121              const intptr_t diff = *dmap++;
122              $if SUBMR == 1:
123                const float16x4_t va0 = vld1_dup_f16(i);
124              $elif SUBMR == 2:
125                const float16x4_t va01 = vreinterpret_f16_f32(vld1_dup_f32((const void*) i));
126              $elif SUBMR == 4:
127                const float16x4_t va0123 = vld1_f16(i);
128              $else:
129                const float16x8_t va01234567 = vld1q_f16(i);
130              $for M in range(8, SUBMR, 8):
131                const float16x8_t va${ABC[M:M+8]} = vld1q_f16(i + ${M});
132              i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
133              $if SUBMR <= 4:
134                const float16x4_t vb = vld1_dup_f16(w); w += 1;
135              $else:
136                const float16x8_t vb = vld1q_dup_f16(w); w += 1;
137              $if SUBMR <= 4:
138                vacc${ABC[0:SUBMR]} = vfma_f16(vacc${ABC[0:SUBMR]}, va${ABC[0:SUBMR]}, vb);
139              $else:
140                $for M in range(0, SUBMR, 8):
141                  vacc${ABC[M:M+8]} = vfmaq_f16(vacc${ABC[M:M+8]}, va${ABC[M:M+8]}, vb);
142            } while (--nnz != 0);
143          }
144          $if SUBMR <= 4:
145            float16x4_t vout${ABC[0:SUBMR]} = vmin_f16(vacc${ABC[0:SUBMR]}, vget_low_f16(vmax));
146            vout${ABC[0:SUBMR]} = vmax_f16(vout${ABC[0:SUBMR]}, vget_low_f16(vmin));
147            $if SUBMR == 1:
148              vst1_lane_f16(o, vout${ABC[0]}, 0);
149            $elif SUBMR == 2:
150              vst1_lane_f32((void*) o, vreinterpret_f32_f16(vout${ABC[0:SUBMR]}), 0);
151            $else:
152              vst1_f16(o, vout${ABC[0:SUBMR]});
153          $else:
154            $for M in range(0, SUBMR, 8):
155              float16x8_t vout${ABC[M:M+8]} = vminq_f16(vacc${ABC[M:M+8]}, vmax);
156            $for M in range(0, SUBMR, 8):
157              vout${ABC[M:M+8]} = vmaxq_f16(vout${ABC[M:M+8]}, vmin);
158            vst1q_f16(o, vout01234567);
159            $for M in range(8, SUBMR, 8):
160              vst1q_f16(o + ${M}, vout${ABC[M:M+8]});
161          o = (__fp16*restrict) ((uintptr_t) o + output_stride);
162        } while (--n != 0);
163        o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
164        i += ${SUBMR};
165      }
166  }
167}
168