1 // Auto-generated file. Do not edit!
2 // Template: src/f32-spmm/neon.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/spmm.h>
15
16
xnn_f32_spmm_ukernel_4x1__neonfma(uint32_t m,uint32_t n,const float * restrict a,const float * restrict weights,const int32_t * restrict widx_dmap,const uint32_t * restrict nidx_nnzmap,float * restrict c,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_spmm_ukernel_4x1__neonfma(
18 uint32_t m,
19 uint32_t n,
20 const float*restrict a,
21 const float*restrict weights,
22 const int32_t*restrict widx_dmap,
23 const uint32_t*restrict nidx_nnzmap,
24 float*restrict c,
25 const union xnn_f32_output_params params[restrict static 1])
26 {
27 assert(m != 0);
28
29 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
30 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
31 size_t i = m;
32 while XNN_LIKELY(i >= 4) {
33 const float*restrict w = weights;
34 const int32_t* dmap = widx_dmap;
35 const uint32_t* nnzmap = nidx_nnzmap;
36 size_t j = n;
37 do {
38 uint32_t nnz = *nnzmap++;
39 float32x4_t vacc0123 = vld1q_dup_f32(w); w += 1;
40 if XNN_LIKELY(nnz != 0) {
41 do {
42 const intptr_t diff = *dmap++;
43 const float32x4_t va0123 = vld1q_f32(a);
44 a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
45 const float32x4_t vb = vld1q_dup_f32(w); w += 1;
46 vacc0123 = vfmaq_f32(vacc0123, va0123, vb);
47 } while (--nnz != 0);
48 }
49 float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
50 vout0123 = vmaxq_f32(vout0123, vmin);
51 vst1q_f32(c, vout0123);
52 c += m;
53 } while (--j != 0);
54 c -= m * n;
55 c += 4;
56 a += 4;
57 i -= 4;
58 }
59 if XNN_UNLIKELY(i != 0) {
60 if (i & 2) {
61 const float*restrict w = weights;
62 const int32_t* dmap = widx_dmap;
63 const uint32_t* nnzmap = nidx_nnzmap;
64 size_t j = n;
65 do {
66 uint32_t nnz = *nnzmap++;
67 float32x2_t vacc01 = vld1_dup_f32(w); w += 1;
68 if XNN_LIKELY(nnz != 0) {
69 do {
70 const intptr_t diff = *dmap++;
71 const float32x2_t va01 = vld1_f32(a);
72 a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
73 const float32x2_t vb = vld1_dup_f32(w); w += 1;
74 vacc01 = vfma_f32(vacc01, va01, vb);
75 } while (--nnz != 0);
76 }
77 float32x2_t vout01 = vmin_f32(vacc01, vget_low_f32(vmax));
78 vout01 = vmax_f32(vout01, vget_low_f32(vmin));
79 vst1_f32(c, vout01);
80 c += m;
81 } while (--j != 0);
82 c -= m * n;
83 c += 2;
84 a += 2;
85 }
86 if (i & 1) {
87 const float*restrict w = weights;
88 const int32_t* dmap = widx_dmap;
89 const uint32_t* nnzmap = nidx_nnzmap;
90 size_t j = n;
91 do {
92 uint32_t nnz = *nnzmap++;
93 float32x2_t vacc0 = vld1_dup_f32(w); w += 1;
94 if XNN_LIKELY(nnz != 0) {
95 do {
96 const intptr_t diff = *dmap++;
97 const float32x2_t va0 = vld1_dup_f32(a);
98 a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
99 const float32x2_t vb = vld1_dup_f32(w); w += 1;
100 vacc0 = vfma_f32(vacc0, va0, vb);
101 } while (--nnz != 0);
102 }
103 float32x2_t vout0 = vmin_f32(vacc0, vget_low_f32(vmax));
104 vout0 = vmax_f32(vout0, vget_low_f32(vmin));
105 vst1_lane_f32(c, vout0, 0);
106 c += m;
107 } while (--j != 0);
108 c -= m * n;
109 c += 1;
110 a += 1;
111 }
112 }
113 }
114