| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/f32-avgpool/rvv_9x.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_9x__rvv_c1v( |
| 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* output, |
| 22 |
size_t input_increment, |
| 23 |
size_t output_increment, |
| 24 |
const struct xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 25 |
{ |
| 26 |
assert(output_pixels != 0); |
| 27 |
assert(kernel_elements != 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 |
const float scale = params->scalar.scale; |
| 35 |
const float min = params->scalar.min; |
| 36 |
const float max = params->scalar.max; |
| 37 |
|
| 38 |
do { |
| 39 |
const float *i[9]; |
| 40 |
for (size_t kk = 0; kk < kernel_elements; ++kk) { |
| 41 |
assert(input[kk] != NULL); |
| 42 |
i[kk] = (input[kk] != zero ? input[kk] + input_offset : zero) ; |
| 43 |
} |
| 44 |
for (size_t tail = kernel_elements; tail < 9; ++tail) { |
| 45 |
i[tail] = zero; |
| 46 |
} |
| 47 |
input = (const float**) ((uintptr_t) input + input_increment); |
| 48 |
|
| 49 |
for (size_t c = channels; c != 0; ) { |
| 50 |
int32_t n = __riscv_vsetvl_e32m1(c); |
| 51 |
|
| 52 |
vfloat32m1_t i0_f32v = __riscv_vle32_v_f32m1(i[0], n); i[0] += n; |
| 53 |
vfloat32m1_t i1_f32v = __riscv_vle32_v_f32m1(i[1], n); i[1] += n; |
| 54 |
vfloat32m1_t i2_f32v = __riscv_vle32_v_f32m1(i[2], n); i[2] += n; |
| 55 |
vfloat32m1_t i3_f32v = __riscv_vle32_v_f32m1(i[3], n); i[3] += n; |
| 56 |
vfloat32m1_t i4_f32v = __riscv_vle32_v_f32m1(i[4], n); i[4] += n; |
| 57 |
vfloat32m1_t i5_f32v = __riscv_vle32_v_f32m1(i[5], n); i[5] += n; |
| 58 |
vfloat32m1_t i6_f32v = __riscv_vle32_v_f32m1(i[6], n); i[6] += n; |
| 59 |
vfloat32m1_t i7_f32v = __riscv_vle32_v_f32m1(i[7], n); i[7] += n; |
| 60 |
vfloat32m1_t i8_f32v = __riscv_vle32_v_f32m1(i[8], n); i[8] += n; |
| 61 |
|
| 62 |
vfloat32m1_t sum01_f32v = __riscv_vfadd_vv_f32m1(i0_f32v, i1_f32v, n); |
| 63 |
vfloat32m1_t sum23_f32v = __riscv_vfadd_vv_f32m1(i2_f32v, i3_f32v, n); |
| 64 |
vfloat32m1_t sum45_f32v = __riscv_vfadd_vv_f32m1(i4_f32v, i5_f32v, n); |
| 65 |
vfloat32m1_t sum67_f32v = __riscv_vfadd_vv_f32m1(i6_f32v, i7_f32v, n); |
| 66 |
vfloat32m1_t sum018_f32v = __riscv_vfadd_vv_f32m1(sum01_f32v, i8_f32v, n); |
| 67 |
vfloat32m1_t sum2345_f32v = __riscv_vfadd_vv_f32m1(sum23_f32v, sum45_f32v, n); |
| 68 |
vfloat32m1_t sum01678_f32v = __riscv_vfadd_vv_f32m1(sum018_f32v, sum67_f32v, n); |
| 69 |
vfloat32m1_t sum_f32v = __riscv_vfadd_vv_f32m1(sum2345_f32v, sum01678_f32v, n); |
| 70 |
vfloat32m1_t out_f32v = __riscv_vfmul_vf_f32m1(sum_f32v, scale, n); |
| 71 |
out_f32v = __riscv_vfmin_vf_f32m1(__riscv_vfmax_vf_f32m1(out_f32v, min, n), max, n); |
| 72 |
__riscv_vse32_v_f32m1(output, out_f32v, n); output += n; |
| 73 |
|
| 74 |
c -= n; |
| 75 |
} |
| 76 |
output = (float*) ((uintptr_t) output + output_increment); |
| 77 |
} while (--output_pixels != 0); |
| 78 |
} |