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_unroll2(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_unroll2(
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 vacc0123x0 = vld1q_dup_f32(w); w += 1;
40 float32x4_t vacc0123x1 = vmovq_n_f32(0.0f);
41 for (; nnz >= 2; nnz -= 2) {
42 const intptr_t diff0 = dmap[0];
43 const intptr_t diff1 = dmap[1];
44 dmap += 2;
45 const float32x4_t va0123x0 = vld1q_f32(a);
46 a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff0);
47 const float32x4_t vb0 = vld1q_dup_f32(w); w += 1;
48 vacc0123x0 = vfmaq_f32(vacc0123x0, va0123x0, vb0);
49 const float32x4_t va0123x1 = vld1q_f32(a);
50 a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff1);
51 const float32x4_t vb1 = vld1q_dup_f32(w); w += 1;
52 vacc0123x1 = vfmaq_f32(vacc0123x1, va0123x1, vb1);
53 }
54 float32x4_t vacc0123 = vacc0123x0;
55 vacc0123 = vaddq_f32(vacc0123, vacc0123x1);
56 if XNN_LIKELY(nnz != 0) {
57 do {
58 const intptr_t diff = *dmap++;
59 const float32x4_t va0123 = vld1q_f32(a);
60 a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
61 const float32x4_t vb = vld1q_dup_f32(w); w += 1;
62 vacc0123 = vfmaq_f32(vacc0123, va0123, vb);
63 } while (--nnz != 0);
64 }
65 float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
66 vout0123 = vmaxq_f32(vout0123, vmin);
67 vst1q_f32(c, vout0123);
68 c += m;
69 } while (--j != 0);
70 c -= m * n;
71 c += 4;
72 a += 4;
73 i -= 4;
74 }
75 if XNN_UNLIKELY(i != 0) {
76 if (i & 2) {
77 const float*restrict w = weights;
78 const int32_t* dmap = widx_dmap;
79 const uint32_t* nnzmap = nidx_nnzmap;
80 size_t j = n;
81 do {
82 uint32_t nnz = *nnzmap++;
83 float32x2_t vacc01 = vld1_dup_f32(w); w += 1;
84 if XNN_LIKELY(nnz != 0) {
85 do {
86 const intptr_t diff = *dmap++;
87 const float32x2_t va01 = vld1_f32(a);
88 a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
89 const float32x2_t vb = vld1_dup_f32(w); w += 1;
90 vacc01 = vfma_f32(vacc01, va01, vb);
91 } while (--nnz != 0);
92 }
93 float32x2_t vout01 = vmin_f32(vacc01, vget_low_f32(vmax));
94 vout01 = vmax_f32(vout01, vget_low_f32(vmin));
95 vst1_f32(c, vout01);
96 c += m;
97 } while (--j != 0);
98 c -= m * n;
99 c += 2;
100 a += 2;
101 }
102 if (i & 1) {
103 const float*restrict w = weights;
104 const int32_t* dmap = widx_dmap;
105 const uint32_t* nnzmap = nidx_nnzmap;
106 size_t j = n;
107 do {
108 uint32_t nnz = *nnzmap++;
109 float32x2_t vacc0 = vld1_dup_f32(w); w += 1;
110 if XNN_LIKELY(nnz != 0) {
111 do {
112 const intptr_t diff = *dmap++;
113 const float32x2_t va0 = vld1_dup_f32(a);
114 a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
115 const float32x2_t vb = vld1_dup_f32(w); w += 1;
116 vacc0 = vfma_f32(vacc0, va0, vb);
117 } while (--nnz != 0);
118 }
119 float32x2_t vout0 = vmin_f32(vacc0, vget_low_f32(vmax));
120 vout0 = vmax_f32(vout0, vget_low_f32(vmin));
121 vst1_lane_f32(c, vout0, 0);
122 c += m;
123 } while (--j != 0);
124 c -= m * n;
125 c += 1;
126 a += 1;
127 }
128 }
129 }
130