Debian Codegraph @

src/f32-avgpool/gen/f32-avgpool-9p8x-minmax-rvv-c2v.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f32-avgpool/rvv_9p8x.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 Imagination Technologies, inc.
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 #include "xnnpack/avgpool.h"
12 #include <riscv_vector.h>
13
14 void xnn_f32_avgpool_minmax_ukernel_9p8x__rvv_c2v(
15 size_t output_pixels,
16 size_t kernel_elements,
17 size_t channels,
18 const float** input,
19 size_t input_offset,
20 const float* zero,
21 float* buffer,
22 float* output,
23 size_t input_increment,
24 size_t output_increment,
25 const struct xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
26 {
27 assert(output_pixels != 0);
28 assert(kernel_elements > 9);
29 assert(channels != 0);
30 assert((input_offset & 3) == 0);
31
32 input_offset >>= XNN_LOG2_SIZEOF_FLOAT;
33
34 do {
35 {
36 const float *i[9];
37 for (size_t kk = 0; kk < 9; ++kk) {
38 assert(input[kk] != NULL);
39 i[kk] = (input[kk] != zero ? input[kk] + input_offset : zero) ;
40 }
41 input += 9;
42
43 float* b = buffer;
44 for (size_t c = channels; c != 0; ) {
45 int32_t n = __riscv_vsetvl_e32m2(c);
46 c -= n;
47
48 vfloat32m2_t i0_f32v = __riscv_vle32_v_f32m2(i[0], n); i[0] += n;
49 vfloat32m2_t i1_f32v = __riscv_vle32_v_f32m2(i[1], n); i[1] += n;
50 vfloat32m2_t i2_f32v = __riscv_vle32_v_f32m2(i[2], n); i[2] += n;
51 vfloat32m2_t i3_f32v = __riscv_vle32_v_f32m2(i[3], n); i[3] += n;
52 vfloat32m2_t i4_f32v = __riscv_vle32_v_f32m2(i[4], n); i[4] += n;
53 vfloat32m2_t i5_f32v = __riscv_vle32_v_f32m2(i[5], n); i[5] += n;
54 vfloat32m2_t i6_f32v = __riscv_vle32_v_f32m2(i[6], n); i[6] += n;
55 vfloat32m2_t i7_f32v = __riscv_vle32_v_f32m2(i[7], n); i[7] += n;
56 vfloat32m2_t i8_f32v = __riscv_vle32_v_f32m2(i[8], n); i[8] += n;
57
58 vfloat32m2_t sum01_f32v = __riscv_vfadd_vv_f32m2(i0_f32v, i1_f32v, n);
59 vfloat32m2_t sum23_f32v = __riscv_vfadd_vv_f32m2(i2_f32v, i3_f32v, n);
60 vfloat32m2_t sum45_f32v = __riscv_vfadd_vv_f32m2(i4_f32v, i5_f32v, n);
61 vfloat32m2_t sum67_f32v = __riscv_vfadd_vv_f32m2(i6_f32v, i7_f32v, n);
62 vfloat32m2_t sum018_f32v = __riscv_vfadd_vv_f32m2(sum01_f32v, i8_f32v, n);
63 vfloat32m2_t sum2345_f32v = __riscv_vfadd_vv_f32m2(sum23_f32v, sum45_f32v, n);
64 vfloat32m2_t sum01678_f32v = __riscv_vfadd_vv_f32m2(sum018_f32v, sum67_f32v, n);
65 vfloat32m2_t sum_f32v = __riscv_vfadd_vv_f32m2(sum2345_f32v, sum01678_f32v, n);
66 __riscv_vse32_v_f32m2(b, sum_f32v, n); b += n;
67 }
68 }
69
70 size_t k = kernel_elements;
71 for (k -= 9; k > 8; k -= 8) {
72 const float *i[8];
73 for (size_t kk = 0; kk < 8; ++kk) {
74 assert(input[kk] != NULL);
75 i[kk] = (input[kk] != zero ? input[kk] + input_offset : zero) ;
76 }
77 input += 8;
78
79 float* b = buffer;
80 for (size_t c = channels; c != 0; ) {
81 int32_t n = __riscv_vsetvl_e32m2(c);
82 c -= n;
83
84 vfloat32m2_t i0_f32v = __riscv_vle32_v_f32m2(i[0], n); i[0] += n;
85 vfloat32m2_t i1_f32v = __riscv_vle32_v_f32m2(i[1], n); i[1] += n;
86 vfloat32m2_t i2_f32v = __riscv_vle32_v_f32m2(i[2], n); i[2] += n;
87 vfloat32m2_t i3_f32v = __riscv_vle32_v_f32m2(i[3], n); i[3] += n;
88 vfloat32m2_t i4_f32v = __riscv_vle32_v_f32m2(i[4], n); i[4] += n;
89 vfloat32m2_t i5_f32v = __riscv_vle32_v_f32m2(i[5], n); i[5] += n;
90 vfloat32m2_t i6_f32v = __riscv_vle32_v_f32m2(i[6], n); i[6] += n;
91 vfloat32m2_t i7_f32v = __riscv_vle32_v_f32m2(i[7], n); i[7] += n;
92 vfloat32m2_t vacc_f32v = __riscv_vle32_v_f32m2(b, n);
93
94 vfloat32m2_t sum01_f32v = __riscv_vfadd_vv_f32m2(i0_f32v, i1_f32v, n);
95 vfloat32m2_t sum23_f32v = __riscv_vfadd_vv_f32m2(i2_f32v, i3_f32v, n);
96 vfloat32m2_t sum45_f32v = __riscv_vfadd_vv_f32m2(i4_f32v, i5_f32v, n);
97 vfloat32m2_t sum67_f32v = __riscv_vfadd_vv_f32m2(i6_f32v, i7_f32v, n);
98 vfloat32m2_t sum01a_f32v = __riscv_vfadd_vv_f32m2(sum01_f32v, vacc_f32v, n);
99 vfloat32m2_t sum2345_f32v = __riscv_vfadd_vv_f32m2(sum23_f32v, sum45_f32v, n);
100 vfloat32m2_t sum0167a_f32v = __riscv_vfadd_vv_f32m2(sum01a_f32v, sum67_f32v, n);
101 vfloat32m2_t sum_f32v = __riscv_vfadd_vv_f32m2(sum2345_f32v, sum0167a_f32v, n);
102 __riscv_vse32_v_f32m2(b, sum_f32v, n); b += n;
103 }
104 }
105
106 {
107 const float *i[8];
108 for (size_t kk = 0; kk < k; ++kk) {
109 assert(input[kk] != NULL);
110 i[kk] = (input[kk] != zero ? input[kk] + input_offset : zero) ;
111 }
112 for (size_t tail = k; tail < 8; ++tail) {
113 i[tail] = zero;
114 }
115 input = (const float**) ((uintptr_t) input + input_increment);
116
117 const float scale = params->scalar.scale;
118 const float min = params->scalar.min;
119 const float max = params->scalar.max;
120 float* b = buffer;
121
122 for (size_t c = channels; c != 0; ) {
123 int32_t n = __riscv_vsetvl_e32m2(c);
124 c -= n;
125
126 vfloat32m2_t i0_f32v = __riscv_vle32_v_f32m2(i[0], n); i[0] += n;
127 vfloat32m2_t i1_f32v = __riscv_vle32_v_f32m2(i[1], n); i[1] += n;
128 vfloat32m2_t i2_f32v = __riscv_vle32_v_f32m2(i[2], n); i[2] += n;
129 vfloat32m2_t i3_f32v = __riscv_vle32_v_f32m2(i[3], n); i[3] += n;
130 vfloat32m2_t i4_f32v = __riscv_vle32_v_f32m2(i[4], n); i[4] += n;
131 vfloat32m2_t i5_f32v = __riscv_vle32_v_f32m2(i[5], n); i[5] += n;
132 vfloat32m2_t i6_f32v = __riscv_vle32_v_f32m2(i[6], n); i[6] += n;
133 vfloat32m2_t i7_f32v = __riscv_vle32_v_f32m2(i[7], n); i[7] += n;
134 vfloat32m2_t vacc_f32v = __riscv_vle32_v_f32m2(b, n); b += n;
135
136 vfloat32m2_t sum01_f32v = __riscv_vfadd_vv_f32m2(i0_f32v, i1_f32v, n);
137 vfloat32m2_t sum23_f32v = __riscv_vfadd_vv_f32m2(i2_f32v, i3_f32v, n);
138 vfloat32m2_t sum45_f32v = __riscv_vfadd_vv_f32m2(i4_f32v, i5_f32v, n);
139 vfloat32m2_t sum67_f32v = __riscv_vfadd_vv_f32m2(i6_f32v, i7_f32v, n);
140 vfloat32m2_t sum01a_f32v = __riscv_vfadd_vv_f32m2(sum01_f32v, vacc_f32v, n);
141 vfloat32m2_t sum2345_f32v = __riscv_vfadd_vv_f32m2(sum23_f32v, sum45_f32v, n);
142 vfloat32m2_t sum0167a_f32v = __riscv_vfadd_vv_f32m2(sum01a_f32v, sum67_f32v, n);
143 vfloat32m2_t sum_f32v = __riscv_vfadd_vv_f32m2(sum2345_f32v, sum0167a_f32v, n);
144 vfloat32m2_t out_f32v = __riscv_vfmul_vf_f32m2(sum_f32v, scale, n);
145 out_f32v = __riscv_vfmin_vf_f32m2(__riscv_vfmax_vf_f32m2(out_f32v, min, n), max, n);
146 __riscv_vse32_v_f32m2(output, out_f32v, n); output += n;
147 }
148 }
149 output = (float*) ((uintptr_t) output + output_increment);
150 } while (--output_pixels != 0);
151 }