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 struct xnn_f16_scaleminmax_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 struct xnn_f16_scaleminmax_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 vscale = vld1q_dup_f16((const __fp16*) ¶ms->scale);
36 const float16x8_t vmax = vld1q_dup_f16((const __fp16*) ¶ms->max);
37 const float16x8_t vmin = vld1q_dup_f16((const __fp16*) ¶ms->min);
38
39 size_t output_decrement = output_stride * nc - 16 * sizeof(__fp16);
40 while XNN_LIKELY(mc >= 16 * sizeof(__fp16)) {
41 const __fp16*restrict w = (const __fp16*) weights;
42 const int32_t* dmap = widx_dmap;
43 const uint32_t* nnzmap = nidx_nnzmap;
44 size_t n = nc;
45 do {
46 uint32_t nnz = *nnzmap++;
47 float16x8_t vacc01234567x0 = vld1q_dup_f16(w); w += 1;
48 float16x8_t vacc01234567x1 = vmovq_n_f16(0.0f);
49 float16x8_t vacc89ABCDEFx0 = vacc01234567x0;
50 float16x8_t vacc89ABCDEFx1 = vmovq_n_f16(0.0f);
51 for (; nnz >= 2; nnz -= 2) {
52 const intptr_t diff0 = dmap[0];
53 const intptr_t diff1 = dmap[1];
54 dmap += 2;
55 const float16x8_t va01234567x0 = vld1q_f16(i);
56 const float16x8_t va89ABCDEFx0 = vld1q_f16(i + 8);
57 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff0);
58 const float16x8_t vb0 = vld1q_dup_f16(w); w += 1;
59 vacc01234567x0 = vfmaq_f16(vacc01234567x0, va01234567x0, vb0);
60 vacc89ABCDEFx0 = vfmaq_f16(vacc89ABCDEFx0, va89ABCDEFx0, vb0);
61 const float16x8_t va01234567x1 = vld1q_f16(i);
62 const float16x8_t va89ABCDEFx1 = vld1q_f16(i + 8);
63 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff1);
64 const float16x8_t vb1 = vld1q_dup_f16(w); w += 1;
65 vacc01234567x1 = vfmaq_f16(vacc01234567x1, va01234567x1, vb1);
66 vacc89ABCDEFx1 = vfmaq_f16(vacc89ABCDEFx1, va89ABCDEFx1, vb1);
67 }
68 float16x8_t vacc01234567 = vacc01234567x0;
69 float16x8_t vacc89ABCDEF = vacc89ABCDEFx0;
70 vacc01234567 = vaddq_f16(vacc01234567, vacc01234567x1);
71 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vacc89ABCDEFx1);
72 if XNN_LIKELY(nnz != 0) {
73 do {
74 const intptr_t diff = *dmap++;
75 const float16x8_t va01234567 = vld1q_f16(i);
76 const float16x8_t va89ABCDEF = vld1q_f16(i + 8);
77 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
78 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
79 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
80 vacc89ABCDEF = vfmaq_f16(vacc89ABCDEF, va89ABCDEF, vb);
81 } while (--nnz != 0);
82 }
83 float16x8_t vout01234567 = vmulq_f16(vacc01234567, vscale);
84 float16x8_t vout89ABCDEF = vmulq_f16(vacc89ABCDEF, vscale);
85 vout01234567 = vminq_f16(vout01234567, vmax);
86 vout89ABCDEF = vminq_f16(vout89ABCDEF, vmax);
87 vout01234567 = vmaxq_f16(vout01234567, vmin);
88 vout89ABCDEF = vmaxq_f16(vout89ABCDEF, vmin);
89 vst1q_f16(o, vout01234567);
90 vst1q_f16(o + 8, vout89ABCDEF);
91 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
92 } while (--n != 0);
93 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
94 i += 16;
95 mc -= 16 * sizeof(__fp16);
96 }
97 if XNN_UNLIKELY(mc != 0) {
98 output_decrement += 8 * sizeof(__fp16);
99 if (mc & (8 * sizeof(__fp16))) {
100 const __fp16*restrict w = (const __fp16*) weights;
101 const int32_t* dmap = widx_dmap;
102 const uint32_t* nnzmap = nidx_nnzmap;
103 size_t n = nc;
104 do {
105 uint32_t nnz = *nnzmap++;
106 float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
107 if XNN_LIKELY(nnz != 0) {
108 do {
109 const intptr_t diff = *dmap++;
110 const float16x8_t va01234567 = vld1q_f16(i);
111 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
112 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
113 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
114 } while (--nnz != 0);
115 }
116 float16x8_t vout01234567 = vminq_f16(vacc01234567, vmax);
117 vout01234567 = vmaxq_f16(vout01234567, vmin);
118 vst1q_f16(o, vout01234567);
119 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
120 } while (--n != 0);
121 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
122 i += 8;
123 }
124 output_decrement += 4 * sizeof(__fp16);
125 if (mc & (4 * sizeof(__fp16))) {
126 const __fp16*restrict w = (const __fp16*) weights;
127 const int32_t* dmap = widx_dmap;
128 const uint32_t* nnzmap = nidx_nnzmap;
129 size_t n = nc;
130 do {
131 uint32_t nnz = *nnzmap++;
132 float16x4_t vacc0123 = vld1_dup_f16(w); w += 1;
133 if XNN_LIKELY(nnz != 0) {
134 do {
135 const intptr_t diff = *dmap++;
136 const float16x4_t va0123 = vld1_f16(i);
137 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
138 const float16x4_t vb = vld1_dup_f16(w); w += 1;
139 vacc0123 = vfma_f16(vacc0123, va0123, vb);
140 } while (--nnz != 0);
141 }
142 float16x4_t vout0123 = vmin_f16(vacc0123, vget_low_f16(vmax));
143 vout0123 = vmax_f16(vout0123, vget_low_f16(vmin));
144 vst1_f16(o, vout0123);
145 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
146 } while (--n != 0);
147 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
148 i += 4;
149 }
150 output_decrement += 2 * sizeof(__fp16);
151 if (mc & (2 * sizeof(__fp16))) {
152 const __fp16*restrict w = (const __fp16*) weights;
153 const int32_t* dmap = widx_dmap;
154 const uint32_t* nnzmap = nidx_nnzmap;
155 size_t n = nc;
156 do {
157 uint32_t nnz = *nnzmap++;
158 float16x4_t vacc01 = vld1_dup_f16(w); w += 1;
159 if XNN_LIKELY(nnz != 0) {
160 do {
161 const intptr_t diff = *dmap++;
162 const float16x4_t va01 = vreinterpret_f16_f32(vld1_dup_f32(__builtin_assume_aligned(i, 1)));
163 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
164 const float16x4_t vb = vld1_dup_f16(w); w += 1;
165 vacc01 = vfma_f16(vacc01, va01, vb);
166 } while (--nnz != 0);
167 }
168 float16x4_t vout01 = vmin_f16(vacc01, vget_low_f16(vmax));
169 vout01 = vmax_f16(vout01, vget_low_f16(vmin));
170 vst1_lane_f32(__builtin_assume_aligned(o, 1), vreinterpret_f32_f16(vout01), 0);
171 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
172 } while (--n != 0);
173 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
174 i += 2;
175 }
176 output_decrement += 1 * sizeof(__fp16);
177 if (mc & (1 * sizeof(__fp16))) {
178 const __fp16*restrict w = (const __fp16*) weights;
179 const int32_t* dmap = widx_dmap;
180 const uint32_t* nnzmap = nidx_nnzmap;
181 size_t n = nc;
182 do {
183 uint32_t nnz = *nnzmap++;
184 float16x4_t vacc0 = vld1_dup_f16(w); w += 1;
185 if XNN_LIKELY(nnz != 0) {
186 do {
187 const intptr_t diff = *dmap++;
188 const float16x4_t va0 = vld1_dup_f16(i);
189 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
190 const float16x4_t vb = vld1_dup_f16(w); w += 1;
191 vacc0 = vfma_f16(vacc0, va0, vb);
192 } while (--nnz != 0);
193 }
194 float16x4_t vout0 = vmin_f16(vacc0, vget_low_f16(vmax));
195 vout0 = vmax_f16(vout0, vget_low_f16(vmin));
196 vst1_lane_f16(o, vout0, 0);
197 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
198 } while (--n != 0);
199 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
200 i += 1;
201 }
202 }
203 }
204