1 // Auto-generated file. Do not edit!
2 // Template: src/f32-spmm/neon-pipelined.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_minmax_ukernel_4x1__neon_pipelined(size_t mc,size_t nc,const float * restrict input,const float * restrict weights,const int32_t * restrict widx_dmap,const uint32_t * restrict nidx_nnzmap,float * restrict output,size_t output_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_spmm_minmax_ukernel_4x1__neon_pipelined(
18 size_t mc,
19 size_t nc,
20 const float*restrict input,
21 const float*restrict weights,
22 const int32_t*restrict widx_dmap,
23 const uint32_t*restrict nidx_nnzmap,
24 float*restrict output,
25 size_t output_stride,
26 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mc != 0);
29 assert(mc % sizeof(float) == 0);
30 assert(nc != 0);
31
32 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
33 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
34 size_t output_decrement = output_stride * nc - 4 * sizeof(float);
35 while XNN_LIKELY(mc >= 4 * sizeof(float)) {
36 const float*restrict w = weights;
37 const int32_t* dmap = widx_dmap;
38 const uint32_t* nnzmap = nidx_nnzmap;
39 float32x4_t vw = vld1q_dup_f32(w); w += 1;
40 intptr_t diff = *dmap++;
41 float32x4_t vi0123 = vld1q_f32(input);
42 size_t n = nc;
43 do {
44 uint32_t nnz = *nnzmap++;
45 float32x4_t vacc0123 = vw;
46 vw = vld1q_dup_f32(w); w += 1;
47 if XNN_LIKELY(nnz != 0) {
48 do {
49 vacc0123 = vmlaq_f32(vacc0123, vi0123, vw);
50 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
51 __builtin_prefetch(input + 16);
52 diff = *dmap++;
53 vw = vld1q_dup_f32(w); w += 1;
54 __builtin_prefetch(w + 32);
55 vi0123 = vld1q_f32(input);
56 } while (--nnz != 0);
57 }
58 float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
59 vout0123 = vmaxq_f32(vout0123, vmin);
60 vst1q_f32(output, vout0123);
61 output = (float*restrict) ((uintptr_t) output + output_stride);
62 } while (--n != 0);
63 output = (float*restrict) ((uintptr_t) output - output_decrement);
64 input += 4;
65 mc -= 4 * sizeof(float);
66 }
67 if XNN_UNLIKELY(mc != 0) {
68 output_decrement += 2 * sizeof(float);
69 if (mc & (2 * sizeof(float))) {
70 const float*restrict w = weights;
71 const int32_t* dmap = widx_dmap;
72 const uint32_t* nnzmap = nidx_nnzmap;
73 size_t n = nc;
74 do {
75 uint32_t nnz = *nnzmap++;
76 float32x2_t vacc01 = vld1_dup_f32(w); w += 1;
77 if XNN_LIKELY(nnz != 0) {
78 do {
79 const intptr_t diff = *dmap++;
80 const float32x2_t vi01 = vld1_f32(input);
81 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
82 __builtin_prefetch(input + 16);
83 const float32x2_t vb = vld1_dup_f32(w); w += 1;
84 __builtin_prefetch(w + 32);
85 vacc01 = vmla_f32(vacc01, vi01, vb);
86 } while (--nnz != 0);
87 }
88 float32x2_t vout01 = vmin_f32(vacc01, vget_low_f32(vmax));
89 vout01 = vmax_f32(vout01, vget_low_f32(vmin));
90 vst1_f32(output, vout01);
91 output = (float*restrict) ((uintptr_t) output + output_stride);
92 } while (--n != 0);
93 output = (float*restrict) ((uintptr_t) output - output_decrement);
94 input += 2;
95 }
96 output_decrement += 1 * sizeof(float);
97 if (mc & (1 * sizeof(float))) {
98 const float*restrict w = weights;
99 const int32_t* dmap = widx_dmap;
100 const uint32_t* nnzmap = nidx_nnzmap;
101 size_t n = nc;
102 do {
103 uint32_t nnz = *nnzmap++;
104 float32x2_t vacc0 = vld1_dup_f32(w); w += 1;
105 if XNN_LIKELY(nnz != 0) {
106 do {
107 const intptr_t diff = *dmap++;
108 const float32x2_t vi0 = vld1_dup_f32(input);
109 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
110 __builtin_prefetch(input + 16);
111 const float32x2_t vb = vld1_dup_f32(w); w += 1;
112 __builtin_prefetch(w + 32);
113 vacc0 = vmla_f32(vacc0, vi0, vb);
114 } while (--nnz != 0);
115 }
116 float32x2_t vout0 = vmin_f32(vacc0, vget_low_f32(vmax));
117 vout0 = vmax_f32(vout0, vget_low_f32(vmin));
118 vst1_lane_f32(output, vout0, 0);
119 output = (float*restrict) ((uintptr_t) output + output_stride);
120 } while (--n != 0);
121 output = (float*restrict) ((uintptr_t) output - output_decrement);
122 input += 1;
123 }
124 }
125 }
126