1 // Auto-generated file. Do not edit!
2 // Template: src/f16-spmm/neonfp16arith.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_f16_spmm_minmax_ukernel_16x1__neonfp16arith_x2(size_t mc,size_t nc,const void * restrict input,const void * restrict weights,const int32_t * restrict widx_dmap,const uint32_t * restrict nidx_nnzmap,void * restrict output,size_t output_stride,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f16_spmm_minmax_ukernel_16x1__neonfp16arith_x2(
18 size_t mc,
19 size_t nc,
20 const void*restrict input,
21 const void*restrict weights,
22 const int32_t*restrict widx_dmap,
23 const uint32_t*restrict nidx_nnzmap,
24 void*restrict output,
25 size_t output_stride,
26 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mc != 0);
29 assert(mc % sizeof(__fp16) == 0);
30 assert(nc != 0);
31
32 const __fp16*restrict i = (const __fp16*) input;
33 __fp16*restrict o = (__fp16*) output;
34
35 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
36 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
37
38 size_t output_decrement = output_stride * nc - 16 * sizeof(__fp16);
39 while XNN_LIKELY(mc >= 16 * sizeof(__fp16)) {
40 const __fp16*restrict w = (const __fp16*) weights;
41 const int32_t* dmap = widx_dmap;
42 const uint32_t* nnzmap = nidx_nnzmap;
43 size_t n = nc;
44 do {
45 uint32_t nnz = *nnzmap++;
46 float16x8_t vacc01234567x0 = vld1q_dup_f16(w); w += 1;
47 float16x8_t vacc01234567x1 = vmovq_n_f16(0.0f);
48 float16x8_t vacc89ABCDEFx0 = vacc01234567x0;
49 float16x8_t vacc89ABCDEFx1 = vmovq_n_f16(0.0f);
50 for (; nnz >= 2; nnz -= 2) {
51 const intptr_t diff0 = dmap[0];
52 const intptr_t diff1 = dmap[1];
53 dmap += 2;
54 const float16x8_t va01234567x0 = vld1q_f16(i);
55 const float16x8_t va89ABCDEFx0 = vld1q_f16(i + 8);
56 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff0);
57 const float16x8_t vb0 = vld1q_dup_f16(w); w += 1;
58 vacc01234567x0 = vfmaq_f16(vacc01234567x0, va01234567x0, vb0);
59 vacc89ABCDEFx0 = vfmaq_f16(vacc89ABCDEFx0, va89ABCDEFx0, vb0);
60 const float16x8_t va01234567x1 = vld1q_f16(i);
61 const float16x8_t va89ABCDEFx1 = vld1q_f16(i + 8);
62 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff1);
63 const float16x8_t vb1 = vld1q_dup_f16(w); w += 1;
64 vacc01234567x1 = vfmaq_f16(vacc01234567x1, va01234567x1, vb1);
65 vacc89ABCDEFx1 = vfmaq_f16(vacc89ABCDEFx1, va89ABCDEFx1, vb1);
66 }
67 float16x8_t vacc01234567 = vacc01234567x0;
68 float16x8_t vacc89ABCDEF = vacc89ABCDEFx0;
69 vacc01234567 = vaddq_f16(vacc01234567, vacc01234567x1);
70 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vacc89ABCDEFx1);
71 if XNN_LIKELY(nnz != 0) {
72 do {
73 const intptr_t diff = *dmap++;
74 const float16x8_t va01234567 = vld1q_f16(i);
75 const float16x8_t va89ABCDEF = vld1q_f16(i + 8);
76 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
77 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
78 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
79 vacc89ABCDEF = vfmaq_f16(vacc89ABCDEF, va89ABCDEF, vb);
80 } while (--nnz != 0);
81 }
82 float16x8_t vout01234567 = vminq_f16(vacc01234567, vmax);
83 float16x8_t vout89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
84 vout01234567 = vmaxq_f16(vout01234567, vmin);
85 vout89ABCDEF = vmaxq_f16(vout89ABCDEF, vmin);
86 vst1q_f16(o, vout01234567);
87 vst1q_f16(o + 8, vout89ABCDEF);
88 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
89 } while (--n != 0);
90 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
91 i += 16;
92 mc -= 16 * sizeof(__fp16);
93 }
94 if XNN_UNLIKELY(mc != 0) {
95 output_decrement += 8 * sizeof(__fp16);
96 if (mc & (8 * sizeof(__fp16))) {
97 const __fp16*restrict w = (const __fp16*) weights;
98 const int32_t* dmap = widx_dmap;
99 const uint32_t* nnzmap = nidx_nnzmap;
100 size_t n = nc;
101 do {
102 uint32_t nnz = *nnzmap++;
103 float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
104 if XNN_LIKELY(nnz != 0) {
105 do {
106 const intptr_t diff = *dmap++;
107 const float16x8_t va01234567 = vld1q_f16(i);
108 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
109 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
110 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
111 } while (--nnz != 0);
112 }
113 float16x8_t vout01234567 = vminq_f16(vacc01234567, vmax);
114 vout01234567 = vmaxq_f16(vout01234567, vmin);
115 vst1q_f16(o, vout01234567);
116 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
117 } while (--n != 0);
118 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
119 i += 8;
120 }
121 output_decrement += 4 * sizeof(__fp16);
122 if (mc & (4 * sizeof(__fp16))) {
123 const __fp16*restrict w = (const __fp16*) weights;
124 const int32_t* dmap = widx_dmap;
125 const uint32_t* nnzmap = nidx_nnzmap;
126 size_t n = nc;
127 do {
128 uint32_t nnz = *nnzmap++;
129 float16x4_t vacc0123 = vld1_dup_f16(w); w += 1;
130 if XNN_LIKELY(nnz != 0) {
131 do {
132 const intptr_t diff = *dmap++;
133 const float16x4_t va0123 = vld1_f16(i);
134 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
135 const float16x4_t vb = vld1_dup_f16(w); w += 1;
136 vacc0123 = vfma_f16(vacc0123, va0123, vb);
137 } while (--nnz != 0);
138 }
139 float16x4_t vout0123 = vmin_f16(vacc0123, vget_low_f16(vmax));
140 vout0123 = vmax_f16(vout0123, vget_low_f16(vmin));
141 vst1_f16(o, vout0123);
142 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
143 } while (--n != 0);
144 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
145 i += 4;
146 }
147 output_decrement += 2 * sizeof(__fp16);
148 if (mc & (2 * sizeof(__fp16))) {
149 const __fp16*restrict w = (const __fp16*) weights;
150 const int32_t* dmap = widx_dmap;
151 const uint32_t* nnzmap = nidx_nnzmap;
152 size_t n = nc;
153 do {
154 uint32_t nnz = *nnzmap++;
155 float16x4_t vacc01 = vld1_dup_f16(w); w += 1;
156 if XNN_LIKELY(nnz != 0) {
157 do {
158 const intptr_t diff = *dmap++;
159 const float16x4_t va01 = vreinterpret_f16_f32(vld1_dup_f32((const void*) i));
160 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
161 const float16x4_t vb = vld1_dup_f16(w); w += 1;
162 vacc01 = vfma_f16(vacc01, va01, vb);
163 } while (--nnz != 0);
164 }
165 float16x4_t vout01 = vmin_f16(vacc01, vget_low_f16(vmax));
166 vout01 = vmax_f16(vout01, vget_low_f16(vmin));
167 vst1_lane_f32((void*) o, vreinterpret_f32_f16(vout01), 0);
168 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
169 } while (--n != 0);
170 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
171 i += 2;
172 }
173 output_decrement += 1 * sizeof(__fp16);
174 if (mc & (1 * sizeof(__fp16))) {
175 const __fp16*restrict w = (const __fp16*) weights;
176 const int32_t* dmap = widx_dmap;
177 const uint32_t* nnzmap = nidx_nnzmap;
178 size_t n = nc;
179 do {
180 uint32_t nnz = *nnzmap++;
181 float16x4_t vacc0 = vld1_dup_f16(w); w += 1;
182 if XNN_LIKELY(nnz != 0) {
183 do {
184 const intptr_t diff = *dmap++;
185 const float16x4_t va0 = vld1_dup_f16(i);
186 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
187 const float16x4_t vb = vld1_dup_f16(w); w += 1;
188 vacc0 = vfma_f16(vacc0, va0, vb);
189 } while (--nnz != 0);
190 }
191 float16x4_t vout0 = vmin_f16(vacc0, vget_low_f16(vmax));
192 vout0 = vmax_f16(vout0, vget_low_f16(vmin));
193 vst1_lane_f16(o, vout0, 0);
194 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
195 } while (--n != 0);
196 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
197 i += 1;
198 }
199 }
200 }
201