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_8x1__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_8x1__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 - 8 * sizeof(__fp16);
40 while XNN_LIKELY(mc >= 8 * 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 for (; nnz >= 2; nnz -= 2) {
50 const intptr_t diff0 = dmap[0];
51 const intptr_t diff1 = dmap[1];
52 dmap += 2;
53 const float16x8_t va01234567x0 = vld1q_f16(i);
54 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff0);
55 const float16x8_t vb0 = vld1q_dup_f16(w); w += 1;
56 vacc01234567x0 = vfmaq_f16(vacc01234567x0, va01234567x0, vb0);
57 const float16x8_t va01234567x1 = vld1q_f16(i);
58 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff1);
59 const float16x8_t vb1 = vld1q_dup_f16(w); w += 1;
60 vacc01234567x1 = vfmaq_f16(vacc01234567x1, va01234567x1, vb1);
61 }
62 float16x8_t vacc01234567 = vacc01234567x0;
63 vacc01234567 = vaddq_f16(vacc01234567, vacc01234567x1);
64 if XNN_LIKELY(nnz != 0) {
65 do {
66 const intptr_t diff = *dmap++;
67 const float16x8_t va01234567 = vld1q_f16(i);
68 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
69 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
70 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
71 } while (--nnz != 0);
72 }
73 float16x8_t vout01234567 = vmulq_f16(vacc01234567, vscale);
74 vout01234567 = vminq_f16(vout01234567, vmax);
75 vout01234567 = vmaxq_f16(vout01234567, vmin);
76 vst1q_f16(o, vout01234567);
77 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
78 } while (--n != 0);
79 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
80 i += 8;
81 mc -= 8 * sizeof(__fp16);
82 }
83 if XNN_UNLIKELY(mc != 0) {
84 output_decrement += 4 * sizeof(__fp16);
85 if (mc & (4 * sizeof(__fp16))) {
86 const __fp16*restrict w = (const __fp16*) weights;
87 const int32_t* dmap = widx_dmap;
88 const uint32_t* nnzmap = nidx_nnzmap;
89 size_t n = nc;
90 do {
91 uint32_t nnz = *nnzmap++;
92 float16x4_t vacc0123 = vld1_dup_f16(w); w += 1;
93 if XNN_LIKELY(nnz != 0) {
94 do {
95 const intptr_t diff = *dmap++;
96 const float16x4_t va0123 = vld1_f16(i);
97 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
98 const float16x4_t vb = vld1_dup_f16(w); w += 1;
99 vacc0123 = vfma_f16(vacc0123, va0123, vb);
100 } while (--nnz != 0);
101 }
102 float16x4_t vout0123 = vmin_f16(vacc0123, vget_low_f16(vmax));
103 vout0123 = vmax_f16(vout0123, vget_low_f16(vmin));
104 vst1_f16(o, vout0123);
105 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
106 } while (--n != 0);
107 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
108 i += 4;
109 }
110 output_decrement += 2 * sizeof(__fp16);
111 if (mc & (2 * sizeof(__fp16))) {
112 const __fp16*restrict w = (const __fp16*) weights;
113 const int32_t* dmap = widx_dmap;
114 const uint32_t* nnzmap = nidx_nnzmap;
115 size_t n = nc;
116 do {
117 uint32_t nnz = *nnzmap++;
118 float16x4_t vacc01 = vld1_dup_f16(w); w += 1;
119 if XNN_LIKELY(nnz != 0) {
120 do {
121 const intptr_t diff = *dmap++;
122 const float16x4_t va01 = vreinterpret_f16_f32(vld1_dup_f32(__builtin_assume_aligned(i, 1)));
123 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
124 const float16x4_t vb = vld1_dup_f16(w); w += 1;
125 vacc01 = vfma_f16(vacc01, va01, vb);
126 } while (--nnz != 0);
127 }
128 float16x4_t vout01 = vmin_f16(vacc01, vget_low_f16(vmax));
129 vout01 = vmax_f16(vout01, vget_low_f16(vmin));
130 vst1_lane_f32(__builtin_assume_aligned(o, 1), vreinterpret_f32_f16(vout01), 0);
131 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
132 } while (--n != 0);
133 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
134 i += 2;
135 }
136 output_decrement += 1 * sizeof(__fp16);
137 if (mc & (1 * sizeof(__fp16))) {
138 const __fp16*restrict w = (const __fp16*) weights;
139 const int32_t* dmap = widx_dmap;
140 const uint32_t* nnzmap = nidx_nnzmap;
141 size_t n = nc;
142 do {
143 uint32_t nnz = *nnzmap++;
144 float16x4_t vacc0 = vld1_dup_f16(w); w += 1;
145 if XNN_LIKELY(nnz != 0) {
146 do {
147 const intptr_t diff = *dmap++;
148 const float16x4_t va0 = vld1_dup_f16(i);
149 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
150 const float16x4_t vb = vld1_dup_f16(w); w += 1;
151 vacc0 = vfma_f16(vacc0, va0, vb);
152 } while (--nnz != 0);
153 }
154 float16x4_t vout0 = vmin_f16(vacc0, vget_low_f16(vmax));
155 vout0 = vmax_f16(vout0, vget_low_f16(vmin));
156 vst1_lane_f16(o, vout0, 0);
157 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
158 } while (--n != 0);
159 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
160 i += 1;
161 }
162 }
163 }
164