nihui
/
ncnn
Not watched
1
0
Fork 0
Code
Releases 46 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Merge 386eb61687 into d68ea6b4d5

pull/4068/merge
Xavier Hsinyuan GitHub 9 months ago
parent
d68ea6b4d5 386eb61687
commit
d0bfe60522
No known key found for this signature in database GPG Key ID: B5690EEEBB952194
3 changed files with 906 additions and 0 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +585
    -0
      src/layer/riscv/layernorm_riscv.cpp
  2. +37
    -0
      src/layer/riscv/layernorm_riscv.h
  3. +284
    -0
      src/layer/riscv/layernorm_rvv_fp16.h

+ 585
- 0
src/layer/riscv/layernorm_riscv.cpp View File

@@ -0,0 +1,585 @@
// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2022 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.
#include "layernorm_riscv.h"
#include <math.h>

#if __riscv_vector
#include <riscv_vector.h>
#endif // __riscv_vector

#include "riscv_usability.h"

namespace ncnn {

#if __riscv_vector && __riscv_zfh
#include "layernorm_rvv_fp16.h"
#endif

LayerNorm_riscv::LayerNorm_riscv()
{
#if __riscv_vector
support_packing = true;
#if __riscv_zfh
support_fp16_storage = true;
#endif
#endif // __riscv_vector
}

#if __riscv_vector
static inline int layernorm_rvv_pack1_procedure(int size, float* ptr, const float* gamma_data, const float* beta_data, float eps, int affine)
{
float sum = 0.f;
float sqsum = 0.f;
vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(), 0.f, vsetvlmax_e32m1());
vfloat32m1_t _sqsum = vfmv_s_f_f32m1(vundefined_f32m1(), 0.f, vsetvlmax_e32m1());
{
int n = size;
float* ptr_sum = ptr;
while (n > 0)
{
size_t vl = vsetvl_e32m8(n);
vfloat32m8_t _p = vle32_v_f32m8(ptr_sum, vl);
_sum = vfredusum_vs_f32m8_f32m1(_sum, _p, /* scalar */ _sum, vl);
// _sqsum = vfredusum_vs_f32m8_f32m1(_sqsum, vfmul_vv_f32m8(_p, _p, vl), /* scalar */ _sqsum, vl);
ptr_sum += vl;
n -= vl;
}
}
sum = vfmv_f_s_f32m1_f32(_sum);
float mean = sum / size;

{
int n = size;
float* ptr_sqsum = ptr;
while (n > 0)
{
size_t vl = vsetvl_e32m8(n);
vfloat32m8_t _p = vle32_v_f32m8(ptr_sqsum, vl);
_p = vfsub_vf_f32m8(_p, mean, vl);
_sqsum = vfredusum_vs_f32m8_f32m1(_sqsum, vfmul_vv_f32m8(_p, _p, vl), /* scalar */ _sqsum, vl);
n -= vl;
ptr_sqsum += vl;
}
}
sqsum = vfmv_f_s_f32m1_f32(_sqsum);
float var = sqsum / size;
// the var maybe minus due to accuracy
//float var = sqsum / size - mean * mean;
float a = static_cast<float>(1.f / (sqrt(var + eps)));
float b = -mean * a;

{
int n = size;
float* ptr_store = ptr;
const float* ptr_gamma = gamma_data;
const float* ptr_beta = beta_data;
if (affine)
{
while (n > 0)
{
size_t vl = vsetvl_e32m8(n);
vfloat32m8_t _p = vle32_v_f32m8(ptr_store, vl);
_p = vfmul_vf_f32m8(_p, a, vl);
vfloat32m8_t _gamma = vle32_v_f32m8(ptr_gamma, vl);
_p = vfadd_vf_f32m8(_p, b, vl);
vfloat32m8_t _beta = vle32_v_f32m8(ptr_beta, vl);
_p = vfmadd_vv_f32m8(_p, _gamma, _beta, vl);
vse32_v_f32m8(ptr_store, _p, vl);

n -= vl;
ptr_store += vl;
ptr_gamma += vl;
ptr_beta += vl;
}
}
else
{
while (n > 0)
{
size_t vl = vsetvl_e32m8(n);
vfloat32m8_t _p = vle32_v_f32m8(ptr_store, vl);
_p = vfmul_vf_f32m8(_p, a, vl);
_p = vfadd_vf_f32m8(_p, b, vl);
vse32_v_f32m8(ptr_store, _p, vl);
n -= vl;
ptr_store += vl;
}
}
}
return 0;
}

static inline int layernorm_rvv_packn_procedure(int size, float* ptr, const float* gamma_data, const float* beta_data, float eps, int affine, const size_t vl)
{
// mean and var
vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
vfloat32m1_t _sqsum = vfmv_v_f_f32m1(0.f, vl);
for (int i = 0; i < size; i++)
{
vfloat32m1_t _p = vle32_v_f32m1(ptr + vl * i, vl);
_sum = vfadd_vv_f32m1(_p, _sum, vl);
// _sqsum = vfmadd_vv_f32m1(_p,_p,_sqsum,vl);
}
vfloat32m1_t _mean = vfdiv_vf_f32m1(_sum, size, vl);
for (int i = 0; i < size; i++)
{
vfloat32m1_t _p = vle32_v_f32m1(ptr + vl * i, vl);
_p = vfsub_vv_f32m1(_p, _mean, vl);
_sqsum = vfmacc_vv_f32m1(_sqsum, _p, _p, vl);
}
vfloat32m1_t _var = vfdiv_vf_f32m1(_sqsum, size, vl);

// the var maybe minus due to accuracy
//float var = sqsum / size - mean * mean;
vfloat32m1_t _a = vfrdiv_vf_f32m1(vfsqrt_v_f32m1(vfadd_vf_f32m1(_var, eps, vl), vl), 1.f, vl);
// how about vfrsqrt7.v?
vfloat32m1_t _b = vfmul_vv_f32m1(vfsgnjn_vv_f32m1(_mean, _mean, vl), _a, vl);
if (affine)
{
for (int i = 0; i < size; i++)
{
const int offset = vl * i;
vfloat32m1_t _p = vle32_v_f32m1(ptr + offset, vl);
_p = vfmadd_vv_f32m1(_p, _a, _b, vl);
_p = vfmul_vf_f32m1(_p, gamma_data[i], vl);
_p = vfadd_vf_f32m1(_p, beta_data[i], vl);
vse32_v_f32m1(ptr + offset, _p, vl);
}
}
else
{
for (int i = 0; i < size; i++)
{
const int offset = vl * i;
vfloat32m1_t _p = vle32_v_f32m1(ptr + offset, vl);
_p = vfmadd_vv_f32m1(_p, _a, _b, vl);
vse32_v_f32m1(ptr + offset, _p, vl);
}
}

return 0;
}
#else
static inline int layernorm_scalar_procedure(int size, float* ptr, const float* gamma_data, const float* beta_data, float eps, int affine)
{
// mean and var
float sum = 0.f;
float sqsum = 0.f;
for (int i = 0; i < size; i++)
{
sum += ptr[i];
//sqsum += ptr[i] * ptr[i];
}
float mean = sum / size;
float tmp = 0.f;
for (int i = 0; i < size; i++)
{
tmp = ptr[i] - mean;
sqsum += tmp * tmp;
}
float var = sqsum / size;
// the var maybe minus due to accuracy
//float var = sqsum / size - mean * mean;

float a = static_cast<float>(1.f / (sqrt(var + eps)));
float b = -mean * a;

if (affine)
{
for (int i = 0; i < size; i++)
{
ptr[i] = (ptr[i] * a + b) * gamma_data[i] + beta_data[i];
}
}
else
{
for (int i = 0; i < size; i++)
{
ptr[i] = ptr[i] * a + b;
}
}
return 0;
}
#endif // __riscv_vector

int LayerNorm_riscv::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
{
// x = (x - mean) / sqrt(var + eps) * gamma + beta
#if __riscv_vector
int elembits = bottom_top_blob.elembits();
#if __riscv_zfh
if (opt.use_fp16_storage && elembits == 16)
{
if (opt.use_fp16_arithmetic)
return forward_inplace_fp16sa(bottom_top_blob, opt);
else
return forward_inplace_fp16s(bottom_top_blob, opt);
}
#endif // __riscv_zfh

int elempack = bottom_top_blob.elempack;
const int packn = csrr_vlenb() / 4;
#endif // __riscv_vector
int dims = bottom_top_blob.dims;
int w = bottom_top_blob.w;

if (dims == 1)
{
float* ptr = bottom_top_blob;
#if __riscv_vector
return layernorm_rvv_pack1_procedure(w * elempack, ptr, gamma_data, beta_data, eps, affine);
#else
return layernorm_scalar_procedure(w, ptr, gamma_data, beta_data, eps, affine);
#endif // __riscv_vector
}
#if __riscv_vector
if (elempack == 1)
#endif
{
if (dims == 2)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
// assert affine_size == w
#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < h; i++)
{
float* ptr = bottom_top_blob.row(i);
#if __riscv_vector
layernorm_rvv_pack1_procedure(w, ptr, gamma_data, beta_data, eps, affine);
#else
layernorm_scalar_procedure(w, ptr, gamma_data, beta_data, eps, affine);
#endif // __riscv_vector
}
}

if (dims == 3)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
int channels = bottom_top_blob.c;
int size = w * h;
if (affine_size == w)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
for (int i = 0; i < h; i++)
{
float* ptr = bottom_top_blob.channel(q).row(i);
#if __riscv_vector
layernorm_rvv_pack1_procedure(w, ptr, gamma_data, beta_data, eps, affine);
#else
layernorm_scalar_procedure(w, ptr, gamma_data, beta_data, eps, affine);
#endif // __riscv_vector
}
}
}
else // if (affine_size == size)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
float* ptr = bottom_top_blob.channel(q);
#if __riscv_vector
layernorm_rvv_pack1_procedure(size, ptr, gamma_data, beta_data, eps, affine);
#else
layernorm_scalar_procedure(size, ptr, gamma_data, beta_data, eps, affine);
#endif // __riscv_vector
}
}
}
}

#if __riscv_vector
if (elempack == packn)
{
const size_t vl = vsetvl_e32m1(packn);
if (dims == 2)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
// assert affine_size == w

#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < h; i++)
{
float* ptr = bottom_top_blob.row(i);
layernorm_rvv_packn_procedure(w, ptr, gamma_data, beta_data, eps, affine, vl);
}
}
if (dims == 3)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
int channels = bottom_top_blob.c;
int size = w * h;

if (affine_size == w)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
for (int i = 0; i < h; i++)
{
float* ptr = bottom_top_blob.channel(q).row(i);

layernorm_rvv_packn_procedure(w, ptr, gamma_data, beta_data, eps, affine, vl);
}
}
}
else // if (affine_size == size)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
float* ptr = bottom_top_blob.channel(q);
layernorm_rvv_packn_procedure(size, ptr, gamma_data, beta_data, eps, affine, vl);
}
}
}
}
#endif // __riscv_vector
return 0;
}

#if __riscv_vector && __riscv_zfh
int LayerNorm_riscv::forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const
{
// x = (x - mean) / sqrt(var + eps) * gamma + beta
int elempack = bottom_top_blob.elempack;
int dims = bottom_top_blob.dims;
int w = bottom_top_blob.w;

if (dims == 1)
{
__fp16* ptr = bottom_top_blob;

return layernorm_rvv_pack1_fp16s_procedure(w * elempack, ptr, gamma_data, beta_data, eps, affine);
}
if (elempack == 1)
{
if (dims == 2)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
// assert affine_size == w
#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < h; i++)
{
__fp16* ptr = bottom_top_blob.row<__fp16>(i);
layernorm_rvv_pack1_fp16s_procedure(w, ptr, gamma_data, beta_data, eps, affine);
}
}

if (dims == 3)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
int channels = bottom_top_blob.c;
int size = w * h;
if (affine_size == w)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
for (int i = 0; i < h; i++)
{
__fp16* ptr = bottom_top_blob.channel(q).row<__fp16>(i);

layernorm_rvv_pack1_fp16s_procedure(w, ptr, gamma_data, beta_data, eps, affine);
}
}
}
else // if (affine_size == size)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
__fp16* ptr = bottom_top_blob.channel(q);
layernorm_rvv_pack1_fp16s_procedure(size, ptr, gamma_data, beta_data, eps, affine);
}
}
}

return 0;
}

const int packn = csrr_vlenb() / 2; // fp16
if (elempack == packn)
{
const size_t vl = vsetvl_e16m1(packn);
if (dims == 2)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
// assert affine_size == w

#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < h; i++)
{
__fp16* ptr = bottom_top_blob.row<__fp16>(i);
layernorm_rvv_packn_fp16s_procedure(w, ptr, gamma_data, beta_data, eps, affine, vl);
}
}
if (dims == 3)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
int channels = bottom_top_blob.c;
int size = w * h;

if (affine_size == w)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
for (int i = 0; i < h; i++)
{
__fp16* ptr = bottom_top_blob.channel(q).row<__fp16>(i);

layernorm_rvv_packn_fp16s_procedure(w, ptr, gamma_data, beta_data, eps, affine, vl);
}
}
}
else // if (affine_size == size)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
__fp16* ptr = bottom_top_blob.channel(q);
layernorm_rvv_packn_fp16s_procedure(size, ptr, gamma_data, beta_data, eps, affine, vl);
}
}
}
}

return 0;
}

int LayerNorm_riscv::forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const
{
// x = (x - mean) / sqrt(var + eps) * gamma + beta
int elempack = bottom_top_blob.elempack;
int dims = bottom_top_blob.dims;
int w = bottom_top_blob.w;

if (dims == 1)
{
__fp16* ptr = bottom_top_blob;

return layernorm_rvv_pack1_fp16sa_procedure(w * elempack, ptr, gamma_data, beta_data, eps, affine);
}
if (elempack == 1)
{
if (dims == 2)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
// assert affine_size == w
#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < h; i++)
{
__fp16* ptr = bottom_top_blob.row<__fp16>(i);
layernorm_rvv_pack1_fp16sa_procedure(w, ptr, gamma_data, beta_data, eps, affine);
}
}

if (dims == 3)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
int channels = bottom_top_blob.c;
int size = w * h;
if (affine_size == w)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
for (int i = 0; i < h; i++)
{
__fp16* ptr = bottom_top_blob.channel(q).row<__fp16>(i);

layernorm_rvv_pack1_fp16sa_procedure(w, ptr, gamma_data, beta_data, eps, affine);
}
}
}
else // if (affine_size == size)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
__fp16* ptr = bottom_top_blob.channel(q);
layernorm_rvv_pack1_fp16sa_procedure(size, ptr, gamma_data, beta_data, eps, affine);
}
}
}

return 0;
}

const int packn = csrr_vlenb() / 2; // fp16
if (elempack == packn)
{
const size_t vl = vsetvl_e16m1(packn);
if (dims == 2)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
// assert affine_size == w

#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < h; i++)
{
__fp16* ptr = bottom_top_blob.row<__fp16>(i);
layernorm_rvv_packn_fp16sa_procedure(w, ptr, gamma_data, beta_data, eps, affine, vl);
}
}
if (dims == 3)
{
int w = bottom_top_blob.w;
int h = bottom_top_blob.h;
int channels = bottom_top_blob.c;
int size = w * h;

if (affine_size == w)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
for (int i = 0; i < h; i++)
{
__fp16* ptr = bottom_top_blob.channel(q).row<__fp16>(i);

layernorm_rvv_packn_fp16sa_procedure(w, ptr, gamma_data, beta_data, eps, affine, vl);
}
}
}
else // if (affine_size == size)
{
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
{
__fp16* ptr = bottom_top_blob.channel(q);
layernorm_rvv_packn_fp16sa_procedure(size, ptr, gamma_data, beta_data, eps, affine, vl);
}
}
}
}

return 0;
}
#endif

} // namespace ncnn

+ 37
- 0
src/layer/riscv/layernorm_riscv.h View File

@@ -0,0 +1,37 @@
// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2022 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.

#ifndef LAYER_LAYERNORM_RISCV_H
#define LAYER_LAYERNORM_RISCV_H

#include "layernorm.h"

namespace ncnn {
class LayerNorm_riscv : virtual public LayerNorm
{
public:
LayerNorm_riscv();

virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;

protected:
#if __riscv_vector && __riscv_zfh
int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
#endif
};

} // namespace ncnn

#endif // LAYER_LAYERNORM_RISCV_H

+ 284
- 0
src/layer/riscv/layernorm_rvv_fp16.h View File

@@ -0,0 +1,284 @@
// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2022 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.

// fp16s
static inline int layernorm_rvv_pack1_fp16s_procedure(int size, __fp16* ptr, const float* gamma_data, const float* beta_data, float eps, int affine)
{
float sum = 0.f;
float sqsum = 0.f;
vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(), 0.f, vsetvlmax_e32m1());
vfloat32m1_t _sqsum = vfmv_s_f_f32m1(vundefined_f32m1(), 0.f, vsetvlmax_e32m1());
{
int n = size;
__fp16* ptr_sum = ptr;
while (n > 0)
{
size_t vl = vsetvl_e16m4(n);
vfloat32m8_t _p = vfwcvt_f_f_v_f32m8(vle16_v_f16m4(ptr_sum, vl), vl);
_sum = vfredusum_vs_f32m8_f32m1(_sum, _p, /* scalar */ _sum, vl);
// _sqsum = vfredusum_vs_f32m8_f32m1(_sqsum, vfmul_vv_f32m8(_p, _p, vl), /* scalar */ _sqsum, vl);
ptr_sum += vl;
n -= vl;
}
}
sum = vfmv_f_s_f32m1_f32(_sum);
float mean = sum / size;

{
int n = size;
__fp16* ptr_sqsum = ptr;
while (n > 0)
{
size_t vl = vsetvl_e16m4(n);
vfloat32m8_t _p = vfwcvt_f_f_v_f32m8(vle16_v_f16m4(ptr_sqsum, vl), vl);
_p = vfsub_vf_f32m8(_p, mean, vl);
_sqsum = vfredusum_vs_f32m8_f32m1(_sqsum, vfmul_vv_f32m8(_p, _p, vl), /* scalar */ _sqsum, vl);
n -= vl;
ptr_sqsum += vl;
}
}
sqsum = vfmv_f_s_f32m1_f32(_sqsum);
float var = sqsum / size;
// the var maybe minus due to accuracy
//float var = sqsum / size - mean * mean;
float a = static_cast<float>(1.f / (sqrt(var + eps)));
float b = -mean * a;

{
int n = size;
__fp16* ptr_store = ptr;
const float* ptr_gamma = gamma_data;
const float* ptr_beta = beta_data;
if (affine)
{
while (n > 0)
{
size_t vl = vsetvl_e16m4(n);
vfloat32m8_t _p = vfwcvt_f_f_v_f32m8(vle16_v_f16m4(ptr_store, vl), vl);
_p = vfmul_vf_f32m8(_p, a, vl);
vfloat32m8_t _gamma = vle32_v_f32m8(ptr_gamma, vl);
_p = vfadd_vf_f32m8(_p, b, vl);
vfloat32m8_t _beta = vle32_v_f32m8(ptr_beta, vl);
_p = vfmadd_vv_f32m8(_p, _gamma, _beta, vl);
vse16_v_f16m4(ptr_store, vfncvt_f_f_w_f16m4(_p, vl), vl);

n -= vl;
ptr_store += vl;
ptr_gamma += vl;
ptr_beta += vl;
}
}
else
{
while (n > 0)
{
size_t vl = vsetvl_e16m4(n);
vfloat32m8_t _p = vfwcvt_f_f_v_f32m8(vle16_v_f16m4(ptr_store, vl), vl);
_p = vfmul_vf_f32m8(_p, a, vl);
_p = vfadd_vf_f32m8(_p, b, vl);
vse16_v_f16m4(ptr_store, vfncvt_f_f_w_f16m4(_p, vl), vl);
n -= vl;
ptr_store += vl;
}
}
}
return 0;
}

static inline int layernorm_rvv_packn_fp16s_procedure(int size, __fp16* ptr, const float* gamma_data, const float* beta_data, float eps, int affine, const size_t vl)
{
// mean and var
// f16m1 => f32m2
vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
vfloat32m2_t _sqsum = vfmv_v_f_f32m2(0.f, vl);
for (int i = 0; i < size; i++)
{
vfloat32m2_t _p = vfwcvt_f_f_v_f32m2(vle16_v_f16m1(ptr + vl * i, vl), vl);
_sum = vfadd_vv_f32m2(_p, _sum, vl);
}
vfloat32m2_t _mean = vfdiv_vf_f32m2(_sum, (float)size, vl);
for (int i = 0; i < size; i++)
{
vfloat32m2_t _p = vfwcvt_f_f_v_f32m2(vle16_v_f16m1(ptr + vl * i, vl), vl);
_p = vfsub_vv_f32m2(_p, _mean, vl);
_sqsum = vfmacc_vv_f32m2(_sqsum, _p, _p, vl);
}
vfloat32m2_t _var = vfdiv_vf_f32m2(_sqsum, (float)size, vl);

// the var maybe minus due to accuracy
//float var = sqsum / size - mean * mean;
vfloat32m2_t _a = vfrdiv_vf_f32m2(vfsqrt_v_f32m2(vfadd_vf_f32m2(_var, eps, vl), vl), 1.f, vl);
// how about vfrsqrt7.v?
vfloat32m2_t _b = vfmul_vv_f32m2(vfsgnjn_vv_f32m2(_mean, _mean, vl), _a, vl);
if (affine)
{
for (int i = 0; i < size; i++)
{
const int offset = vl * i;
vfloat32m2_t _p = vfwcvt_f_f_v_f32m2(vle16_v_f16m1(ptr + offset, vl), vl);
_p = vfmadd_vv_f32m2(_p, _a, _b, vl);
_p = vfmul_vf_f32m2(_p, gamma_data[i], vl);
_p = vfadd_vf_f32m2(_p, beta_data[i], vl);
vse16_v_f16m1(ptr + offset, vfncvt_f_f_w_f16m1(_p, vl), vl);
}
}
else
{
for (int i = 0; i < size; i++)
{
const int offset = vl * i;
vfloat32m2_t _p = vfwcvt_f_f_v_f32m2(vle16_v_f16m1(ptr + offset, vl), vl);
_p = vfmadd_vv_f32m2(_p, _a, _b, vl);
vse16_v_f16m1(ptr + offset, vfncvt_f_f_w_f16m1(_p, vl), vl);
}
}

return 0;
}

// fp16sa

static inline int layernorm_rvv_pack1_fp16sa_procedure(int size, __fp16* ptr, const float* gamma_data, const float* beta_data, float eps, int affine)
{
__fp16 sum = 0.f;
__fp16 sqsum = 0.f;
vfloat16m1_t _sum = vfmv_s_f_f16m1(vundefined_f16m1(), 0.f, vsetvlmax_e32m1());
vfloat16m1_t _sqsum = vfmv_s_f_f16m1(vundefined_f16m1(), 0.f, vsetvlmax_e32m1());
{
int n = size;
__fp16* ptr_sum = ptr;
while (n > 0)
{
size_t vl = vsetvl_e16m8(n);
vfloat16m8_t _p = vle16_v_f16m8(ptr_sum, vl);
_sum = vfredusum_vs_f16m8_f16m1(_sum, _p, /* scalar */ _sum, vl);
// _sqsum = vfredusum_vs_f32m8_f32m1(_sqsum, vfmul_vv_f32m8(_p, _p, vl), /* scalar */ _sqsum, vl);
ptr_sum += vl;
n -= vl;
}
}
sum = vfmv_f_s_f16m1_f16(_sum);
__fp16 mean = sum / size;

{
int n = size;
__fp16* ptr_sqsum = ptr;
while (n > 0)
{
size_t vl = vsetvl_e16m8(n);
vfloat16m8_t _p = vle16_v_f16m8(ptr_sqsum, vl);
_p = vfsub_vf_f16m8(_p, mean, vl);
_sqsum = vfredusum_vs_f16m8_f16m1(_sqsum, vfmul_vv_f16m8(_p, _p, vl), /* scalar */ _sqsum, vl);
n -= vl;
ptr_sqsum += vl;
}
}
sqsum = vfmv_f_s_f16m1_f16(_sqsum);
__fp16 var = sqsum / size;
// the var maybe minus due to accuracy
//float var = sqsum / size - mean * mean;
__fp16 a = static_cast<__fp16>(1.f / (sqrt(var + eps)));
__fp16 b = static_cast<__fp16>(-mean * a);

{
int n = size;
__fp16* ptr_store = ptr;
const float* ptr_gamma = gamma_data;
const float* ptr_beta = beta_data;
if (affine)
{
while (n > 0)
{
size_t vl = vsetvl_e16m4(n);
vfloat16m4_t _p = vle16_v_f16m4(ptr_store, vl);
_p = vfmul_vf_f16m4(_p, a, vl);
vfloat16m4_t _gamma = vfncvt_f_f_w_f16m4(vle32_v_f32m8(ptr_gamma, vl), vl);
_p = vfadd_vf_f16m4(_p, b, vl);
vfloat16m4_t _beta = vfncvt_f_f_w_f16m4(vle32_v_f32m8(ptr_beta, vl), vl);
_p = vfmadd_vv_f16m4(_p, _gamma, _beta, vl);
vse16_v_f16m4(ptr_store, _p, vl);

n -= vl;
ptr_store += vl;
ptr_gamma += vl;
ptr_beta += vl;
}
}
else
{
while (n > 0)
{
size_t vl = vsetvl_e16m8(n);
vfloat16m8_t _p = vle16_v_f16m8(ptr_store, vl);
_p = vfmul_vf_f16m8(_p, a, vl);
_p = vfadd_vf_f16m8(_p, b, vl);
vse16_v_f16m8(ptr_store, _p, vl);
n -= vl;
ptr_store += vl;
}
}
}
return 0;
}

static inline int layernorm_rvv_packn_fp16sa_procedure(int size, __fp16* ptr, const float* gamma_data, const float* beta_data, float eps, int affine, const size_t vl)
{
// mean and var
vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
vfloat16m1_t _sqsum = vfmv_v_f_f16m1(0.f, vl);
for (int i = 0; i < size; i++)
{
vfloat16m1_t _p = vle16_v_f16m1(ptr + vl * i, vl);
_sum = vfadd_vv_f16m1(_p, _sum, vl);
// _sqsum = vfmadd_vv_f16m1(_p,_p,_sqsum,vl);
}
vfloat16m1_t _mean = vfdiv_vf_f16m1(_sum, size, vl);
for (int i = 0; i < size; i++)
{
vfloat16m1_t _p = vle16_v_f16m1(ptr + vl * i, vl);
_p = vfsub_vv_f16m1(_p, _mean, vl);
_sqsum = vfmacc_vv_f16m1(_sqsum, _p, _p, vl);
}
vfloat16m1_t _var = vfdiv_vf_f16m1(_sqsum, size, vl);

// the var maybe minus due to accuracy
//float var = sqsum / size - mean * mean;
vfloat16m1_t _a = vfrdiv_vf_f16m1(vfsqrt_v_f16m1(vfadd_vf_f16m1(_var, eps, vl), vl), 1.f, vl);
// how about vfrsqrt7.v?
vfloat16m1_t _b = vfmul_vv_f16m1(vfsgnjn_vv_f16m1(_mean, _mean, vl), _a, vl);
if (affine)
{
for (int i = 0; i < size; i++)
{
const int offset = vl * i;
vfloat16m1_t _p = vle16_v_f16m1(ptr + offset, vl);
_p = vfmadd_vv_f16m1(_p, _a, _b, vl);
_p = vfmul_vf_f16m1(_p, gamma_data[i], vl);
_p = vfadd_vf_f16m1(_p, beta_data[i], vl);
vse16_v_f16m1(ptr + offset, _p, vl);
}
}
else
{
for (int i = 0; i < size; i++)
{
const int offset = vl * i;
vfloat16m1_t _p = vle16_v_f16m1(ptr + offset, vl);
_p = vfmadd_vv_f16m1(_p, _a, _b, vl);
vse16_v_f16m1(ptr + offset, _p, vl);
}
}

return 0;
}

Write Preview
Loading…
Cancel
Save