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 struct xnn_f16_scaleminmax_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 vscale = vld1q_dup_f16((const __fp16*) &params->scale);
34  const float16x8_t vmax = vld1q_dup_f16((const __fp16*) &params->max);
35  const float16x8_t vmin = vld1q_dup_f16((const __fp16*) &params->min);
36
37  size_t output_decrement = output_stride * nc - ${MR} * sizeof(__fp16);
38  while XNN_LIKELY(mc >= ${MR} * sizeof(__fp16)) {
39    const __fp16*restrict w = (const __fp16*) weights;
40    const int32_t* dmap = widx_dmap;
41    const uint32_t* nnzmap = nidx_nnzmap;
42    size_t n = nc;
43    do {
44      uint32_t nnz = *nnzmap++;
45      $if UNROLL > 1:
46        float16x8_t vacc01234567x0 = vld1q_dup_f16(w); w += 1;
47        $for K in range(1, UNROLL):
48          float16x8_t vacc01234567x${K} = vmovq_n_f16(0.0f);
49        $for M in range(8, MR, 8):
50          float16x8_t vacc${ABC[M:M+8]}x0 = vacc01234567x0;
51          $for K in range(1, UNROLL):
52            float16x8_t vacc${ABC[M:M+8]}x${K} = vmovq_n_f16(0.0f);
53        for (; nnz >= ${UNROLL}; nnz -= ${UNROLL}) {
54          $for K in range(UNROLL):
55            const intptr_t diff${K} = dmap[${K}];
56          dmap += ${UNROLL};
57          $for K in range(UNROLL):
58            const float16x8_t va01234567x${K} = vld1q_f16(i);
59            $for M in range(8, MR, 8):
60              const float16x8_t va${ABC[M:M+8]}x${K} = vld1q_f16(i + ${M});
61            i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff${K});
62            const float16x8_t vb${K} = vld1q_dup_f16(w); w += 1;
63            $for M in range(0, MR, 8):
64              vacc${ABC[M:M+8]}x${K} = vfmaq_f16(vacc${ABC[M:M+8]}x${K}, va${ABC[M:M+8]}x${K}, vb${K});
65        }
66        $for M in range(0, MR, 8):
67          float16x8_t vacc${ABC[M:M+8]} = vacc${ABC[M:M+8]}x0;
68        $for K in range(1, UNROLL):
69          $for M in range(0, MR, 8):
70            vacc${ABC[M:M+8]} = vaddq_f16(vacc${ABC[M:M+8]}, vacc${ABC[M:M+8]}x${K});
71      $else:
72        float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
73        $for M in range(8, MR, 8):
74          float16x8_t vacc${ABC[M:M+8]} = vacc01234567;
75      if XNN_LIKELY(nnz != 0) {
76        do {
77          const intptr_t diff = *dmap++;
78          const float16x8_t va01234567 = vld1q_f16(i);
79          $for M in range(8, MR, 8):
80            const float16x8_t va${ABC[M:M+8]} = vld1q_f16(i + ${M});
81          i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
82          const float16x8_t vb = vld1q_dup_f16(w); w += 1;
83          $for M in range(0, MR, 8):
84            vacc${ABC[M:M+8]} = vfmaq_f16(vacc${ABC[M:M+8]}, va${ABC[M:M+8]}, vb);
85        } while (--nnz != 0);
86      }
87      $for M in range(0, MR, 8):
88        float16x8_t vout${ABC[M:M+8]} = vmulq_f16(vacc${ABC[M:M+8]}, vscale);
89      $for M in range(0, MR, 8):
90        vout${ABC[M:M+8]} = vminq_f16(vout${ABC[M:M+8]}, vmax);
91      $for M in range(0, MR, 8):
92        vout${ABC[M:M+8]} = vmaxq_f16(vout${ABC[M:M+8]}, vmin);
93      vst1q_f16(o, vout01234567);
94      $for M in range(8, MR, 8):
95        vst1q_f16(o + ${M}, vout${ABC[M:M+8]});
96      o = (__fp16*restrict) ((uintptr_t) o + output_stride);
97    } while (--n != 0);
98    o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
99    i += ${MR};
100    mc -= ${MR} * sizeof(__fp16);
101  }
102  if XNN_UNLIKELY(mc != 0) {
103    $for LOG2M in reversed(range((MR - 1).bit_length())):
104      $SUBMR = 1 << LOG2M
105      $if SUBMR * 2 >= MR:
106        output_decrement += ${MR - SUBMR} * sizeof(__fp16);
107      $else:
108        output_decrement += ${SUBMR} * sizeof(__fp16);
109      if (mc & (${SUBMR} * sizeof(__fp16))) {
110        const __fp16*restrict w = (const __fp16*) weights;
111        const int32_t* dmap = widx_dmap;
112        const uint32_t* nnzmap = nidx_nnzmap;
113        size_t n = nc;
114        do {
115          uint32_t nnz = *nnzmap++;
116          $if SUBMR <= 4:
117            float16x4_t vacc${ABC[0:SUBMR]} = vld1_dup_f16(w); w += 1;
118          $else:
119            float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
120          $for M in range(8, SUBMR, 8):
121            float16x8_t vacc${ABC[M:M+8]} = vacc01234567;
122          if XNN_LIKELY(nnz != 0) {
123            do {
124              const intptr_t diff = *dmap++;
125              $if SUBMR == 1:
126                const float16x4_t va0 = vld1_dup_f16(i);
127              $elif SUBMR == 2:
128                const float16x4_t va01 = vreinterpret_f16_f32(vld1_dup_f32(__builtin_assume_aligned(i, 1)));
129              $elif SUBMR == 4:
130                const float16x4_t va0123 = vld1_f16(i);
131              $else:
132                const float16x8_t va01234567 = vld1q_f16(i);
133              $for M in range(8, SUBMR, 8):
134                const float16x8_t va${ABC[M:M+8]} = vld1q_f16(i + ${M});
135              i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
136              $if SUBMR <= 4:
137                const float16x4_t vb = vld1_dup_f16(w); w += 1;
138              $else:
139                const float16x8_t vb = vld1q_dup_f16(w); w += 1;
140              $if SUBMR <= 4:
141                vacc${ABC[0:SUBMR]} = vfma_f16(vacc${ABC[0:SUBMR]}, va${ABC[0:SUBMR]}, vb);
142              $else:
143                $for M in range(0, SUBMR, 8):
144                  vacc${ABC[M:M+8]} = vfmaq_f16(vacc${ABC[M:M+8]}, va${ABC[M:M+8]}, vb);
145            } while (--nnz != 0);
146          }
147          $if SUBMR <= 4:
148            float16x4_t vout${ABC[0:SUBMR]} = vmin_f16(vacc${ABC[0:SUBMR]}, vget_low_f16(vmax));
149            vout${ABC[0:SUBMR]} = vmax_f16(vout${ABC[0:SUBMR]}, vget_low_f16(vmin));
150            $if SUBMR == 1:
151              vst1_lane_f16(o, vout${ABC[0]}, 0);
152            $elif SUBMR == 2:
153              vst1_lane_f32(__builtin_assume_aligned(o, 1), vreinterpret_f32_f16(vout${ABC[0:SUBMR]}), 0);
154            $else:
155              vst1_f16(o, vout${ABC[0:SUBMR]});
156          $else:
157            $for M in range(0, SUBMR, 8):
158              float16x8_t vout${ABC[M:M+8]} = vminq_f16(vacc${ABC[M:M+8]}, vmax);
159            $for M in range(0, SUBMR, 8):
160              vout${ABC[M:M+8]} = vmaxq_f16(vout${ABC[M:M+8]}, vmin);
161            vst1q_f16(o, vout01234567);
162            $for M in range(8, SUBMR, 8):
163              vst1q_f16(o + ${M}, vout${ABC[M:M+8]});
164          o = (__fp16*restrict) ((uintptr_t) o + output_stride);
165        } while (--n != 0);
166        o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
167        i += ${SUBMR};
168      }
169  }
170}
171