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_16x1__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_16x1__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 - 16 * sizeof(float);
35 while XNN_LIKELY(mc >= 16 * 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 float32x4_t vi4567 = vld1q_f32(input + 4);
43 float32x4_t vi89AB = vld1q_f32(input + 8);
44 float32x4_t viCDEF = vld1q_f32(input + 12);
45 size_t n = nc;
46 do {
47 uint32_t nnz = *nnzmap++;
48 float32x4_t vacc0123 = vw;
49 float32x4_t vacc4567 = vw;
50 float32x4_t vacc89AB = vw;
51 float32x4_t vaccCDEF = vw;
52 vw = vld1q_dup_f32(w); w += 1;
53 if XNN_LIKELY(nnz != 0) {
54 do {
55 vacc0123 = vmlaq_f32(vacc0123, vi0123, vw);
56 vacc4567 = vmlaq_f32(vacc4567, vi4567, vw);
57 vacc89AB = vmlaq_f32(vacc89AB, vi89AB, vw);
58 vaccCDEF = vmlaq_f32(vaccCDEF, viCDEF, vw);
59 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
60 __builtin_prefetch(input + 16);
61 diff = *dmap++;
62 vw = vld1q_dup_f32(w); w += 1;
63 __builtin_prefetch(w + 32);
64 vi0123 = vld1q_f32(input);
65 vi4567 = vld1q_f32(input + 4);
66 vi89AB = vld1q_f32(input + 8);
67 viCDEF = vld1q_f32(input + 12);
68 } while (--nnz != 0);
69 }
70 float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
71 float32x4_t vout4567 = vminq_f32(vacc4567, vmax);
72 float32x4_t vout89AB = vminq_f32(vacc89AB, vmax);
73 float32x4_t voutCDEF = vminq_f32(vaccCDEF, vmax);
74 vout0123 = vmaxq_f32(vout0123, vmin);
75 vout4567 = vmaxq_f32(vout4567, vmin);
76 vout89AB = vmaxq_f32(vout89AB, vmin);
77 voutCDEF = vmaxq_f32(voutCDEF, vmin);
78 vst1q_f32(output, vout0123);
79 vst1q_f32(output + 4, vout4567);
80 vst1q_f32(output + 8, vout89AB);
81 vst1q_f32(output + 12, voutCDEF);
82 output = (float*restrict) ((uintptr_t) output + output_stride);
83 } while (--n != 0);
84 output = (float*restrict) ((uintptr_t) output - output_decrement);
85 input += 16;
86 mc -= 16 * sizeof(float);
87 }
88 if XNN_UNLIKELY(mc != 0) {
89 output_decrement += 8 * sizeof(float);
90 if (mc & (8 * sizeof(float))) {
91 const float*restrict w = weights;
92 const int32_t* dmap = widx_dmap;
93 const uint32_t* nnzmap = nidx_nnzmap;
94 size_t n = nc;
95 do {
96 uint32_t nnz = *nnzmap++;
97 float32x4_t vacc0123 = vld1q_dup_f32(w); w += 1;
98 float32x4_t vacc4567 = vacc0123;
99 if XNN_LIKELY(nnz != 0) {
100 do {
101 const intptr_t diff = *dmap++;
102 const float32x4_t vi0123 = vld1q_f32(input);
103 const float32x4_t vi4567 = vld1q_f32(input + 4);
104 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
105 __builtin_prefetch(input + 16);
106 const float32x4_t vb = vld1q_dup_f32(w); w += 1;
107 __builtin_prefetch(w + 32);
108 vacc0123 = vmlaq_f32(vacc0123, vi0123, vb);
109 vacc4567 = vmlaq_f32(vacc4567, vi4567, vb);
110 } while (--nnz != 0);
111 }
112 float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
113 float32x4_t vout4567 = vminq_f32(vacc4567, vmax);
114 vout0123 = vmaxq_f32(vout0123, vmin);
115 vout4567 = vmaxq_f32(vout4567, vmin);
116 vst1q_f32(output, vout0123);
117 vst1q_f32(output + 4, vout4567);
118 output = (float*restrict) ((uintptr_t) output + output_stride);
119 } while (--n != 0);
120 output = (float*restrict) ((uintptr_t) output - output_decrement);
121 input += 8;
122 }
123 output_decrement += 4 * sizeof(float);
124 if (mc & (4 * sizeof(float))) {
125 const float*restrict w = weights;
126 const int32_t* dmap = widx_dmap;
127 const uint32_t* nnzmap = nidx_nnzmap;
128 size_t n = nc;
129 do {
130 uint32_t nnz = *nnzmap++;
131 float32x4_t vacc0123 = vld1q_dup_f32(w); w += 1;
132 if XNN_LIKELY(nnz != 0) {
133 do {
134 const intptr_t diff = *dmap++;
135 const float32x4_t vi0123 = vld1q_f32(input);
136 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
137 __builtin_prefetch(input + 16);
138 const float32x4_t vb = vld1q_dup_f32(w); w += 1;
139 __builtin_prefetch(w + 32);
140 vacc0123 = vmlaq_f32(vacc0123, vi0123, vb);
141 } while (--nnz != 0);
142 }
143 float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
144 vout0123 = vmaxq_f32(vout0123, vmin);
145 vst1q_f32(output, vout0123);
146 output = (float*restrict) ((uintptr_t) output + output_stride);
147 } while (--n != 0);
148 output = (float*restrict) ((uintptr_t) output - output_decrement);
149 input += 4;
150 }
151 output_decrement += 2 * sizeof(float);
152 if (mc & (2 * sizeof(float))) {
153 const float*restrict w = weights;
154 const int32_t* dmap = widx_dmap;
155 const uint32_t* nnzmap = nidx_nnzmap;
156 size_t n = nc;
157 do {
158 uint32_t nnz = *nnzmap++;
159 float32x2_t vacc01 = vld1_dup_f32(w); w += 1;
160 if XNN_LIKELY(nnz != 0) {
161 do {
162 const intptr_t diff = *dmap++;
163 const float32x2_t vi01 = vld1_f32(input);
164 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
165 __builtin_prefetch(input + 16);
166 const float32x2_t vb = vld1_dup_f32(w); w += 1;
167 __builtin_prefetch(w + 32);
168 vacc01 = vmla_f32(vacc01, vi01, vb);
169 } while (--nnz != 0);
170 }
171 float32x2_t vout01 = vmin_f32(vacc01, vget_low_f32(vmax));
172 vout01 = vmax_f32(vout01, vget_low_f32(vmin));
173 vst1_f32(output, vout01);
174 output = (float*restrict) ((uintptr_t) output + output_stride);
175 } while (--n != 0);
176 output = (float*restrict) ((uintptr_t) output - output_decrement);
177 input += 2;
178 }
179 output_decrement += 1 * sizeof(float);
180 if (mc & (1 * sizeof(float))) {
181 const float*restrict w = weights;
182 const int32_t* dmap = widx_dmap;
183 const uint32_t* nnzmap = nidx_nnzmap;
184 size_t n = nc;
185 do {
186 uint32_t nnz = *nnzmap++;
187 float32x2_t vacc0 = vld1_dup_f32(w); w += 1;
188 if XNN_LIKELY(nnz != 0) {
189 do {
190 const intptr_t diff = *dmap++;
191 const float32x2_t vi0 = vld1_dup_f32(input);
192 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
193 __builtin_prefetch(input + 16);
194 const float32x2_t vb = vld1_dup_f32(w); w += 1;
195 __builtin_prefetch(w + 32);
196 vacc0 = vmla_f32(vacc0, vi0, vb);
197 } while (--nnz != 0);
198 }
199 float32x2_t vout0 = vmin_f32(vacc0, vget_low_f32(vmax));
200 vout0 = vmax_f32(vout0, vget_low_f32(vmin));
201 vst1_lane_f32(output, vout0, 0);
202 output = (float*restrict) ((uintptr_t) output + output_stride);
203 } while (--n != 0);
204 output = (float*restrict) ((uintptr_t) output - output_decrement);
205 input += 1;
206 }
207 }
208 }
209