derek.holloway
7502bf64a4
CodeQL / Analyze (actions) (push) Canceled after 0s
CodeQL / Analyze (python) (push) Canceled after 0s
pre-commit / pre-commit (push) Canceled after 0s
Multi-Arch CI / setup (push) Canceled after 0s
Publish build_manylinux_x86_64 images / publish_build_manylinux_x86_64 (push) Canceled after 0s
Setup (Multi-Arch) / setup (push) Canceled after 0s
Multi-Arch CI / Linux::${{ fromJSON(needs.setup.outputs.linux_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubi10 image / publish_no_rocm_image_ubi10 (push) Canceled after 0s
Multi-Arch CI / Windows::${{ fromJSON(needs.setup.outputs.windows_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Multi-Arch CI / Manifest Diff (push) Canceled after 0s
TheRock Manifest Diff Report / Generate Manifest Diff Report (push) Canceled after 0s
Multi-Arch CI / CI Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_ocl_rt images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_rocgdb images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Unit Tests / Unit Tests :: ubuntu-24.04 (push) Canceled after 0s
Unit Tests / Unit Tests :: windows-2022 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_media (push) Canceled after 0s
Unit Tests / Unit Tests Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_openmpi (push) Canceled after 0s
511 lines
21 KiB
C++
511 lines
21 KiB
C++
/*******************************************************************************
|
|
*
|
|
* MIT License
|
|
*
|
|
* Copyright (c) 2024 Advanced Micro Devices, Inc.
|
|
*
|
|
* Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
* of this software and associated documentation files (the "Software"), to deal
|
|
* in the Software without restriction, including without limitation the rights
|
|
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
* copies of the Software, and to permit persons to whom the Software is
|
|
* furnished to do so, subject to the following conditions:
|
|
*
|
|
* The above copyright notice and this permission notice shall be included in all
|
|
* copies or substantial portions of the Software.
|
|
*
|
|
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
* SOFTWARE.
|
|
*
|
|
*******************************************************************************/
|
|
|
|
#include "../driver/tensor_driver.hpp"
|
|
#include "get_handle.hpp"
|
|
#include "random.hpp"
|
|
#include "tensor_holder.hpp"
|
|
#include "verify.hpp"
|
|
#include <gtest/gtest.h>
|
|
#include <miopen/t5layernorm.hpp>
|
|
#include <miopen/miopen.h>
|
|
|
|
template <class T>
|
|
void cpu_t5layernorm_forward(tensor<T> x,
|
|
tensor<T> weight,
|
|
tensor<T>& ref_y,
|
|
tensor<T>& ref_rstd,
|
|
float eps,
|
|
miopenNormMode_t mode)
|
|
{
|
|
auto dims = x.desc.GetLengths();
|
|
size_t outer_size = 1;
|
|
size_t inner_size = dims[dims.size() - 1];
|
|
|
|
for(size_t i = 0ULL; i < dims.size() - 1; ++i)
|
|
{
|
|
outer_size *= dims[i];
|
|
}
|
|
|
|
miopen::par_ford(outer_size)([&](int32_t o) {
|
|
float pvar = 0;
|
|
|
|
miopen::ford(inner_size)([&](int32_t i) {
|
|
float tmp = static_cast<float>(x[o * inner_size + i]);
|
|
pvar += tmp * tmp;
|
|
});
|
|
|
|
pvar = pvar / inner_size;
|
|
float prstd = 1 / sqrt(pvar + eps);
|
|
|
|
ref_rstd[o] = static_cast<T>(prstd);
|
|
|
|
miopen::ford(inner_size)([&](int32_t i) {
|
|
float pweight = mode ? static_cast<float>(weight[i]) : 1;
|
|
ref_y[o * inner_size + i] =
|
|
static_cast<T>(static_cast<float>(x[o * inner_size + i]) * prstd * pweight);
|
|
});
|
|
});
|
|
}
|
|
|
|
template <class T>
|
|
void cpu_t5layernorm_backward(tensor<T> dy,
|
|
tensor<T> x,
|
|
tensor<T> weight,
|
|
tensor<T> rstd,
|
|
tensor<T>& ref_dx,
|
|
miopenNormMode_t mode)
|
|
{
|
|
auto dims = dy.desc.GetLengths();
|
|
size_t outer_size = 1;
|
|
size_t inner_size = dims[dims.size() - 1];
|
|
|
|
for(size_t i = 0ULL; i < dims.size() - 1; ++i)
|
|
{
|
|
outer_size *= dims[i];
|
|
}
|
|
|
|
miopen::par_ford(outer_size)([&](int32_t o) {
|
|
float sum = 0;
|
|
|
|
miopen::ford(inner_size)([&](int32_t i) {
|
|
float pweight = mode ? static_cast<float>(weight[i]) : 1;
|
|
float pdy = (dy.GetSize() != 0) ? static_cast<float>(dy[o * inner_size + i]) : 0;
|
|
float px = static_cast<float>(x[o * inner_size + i]);
|
|
sum += pdy * px * pweight;
|
|
});
|
|
|
|
float s = 1 / static_cast<float>(inner_size);
|
|
float prstd = static_cast<float>(rstd[o]);
|
|
float a = sum * prstd * prstd * prstd * s;
|
|
|
|
miopen::ford(inner_size)([&](int32_t i) {
|
|
float pweight = mode ? static_cast<float>(weight[i]) : 1;
|
|
float pdy = (dy.GetSize() != 0) ? static_cast<float>(dy[o * inner_size + i]) : 0;
|
|
|
|
float val = prstd * pdy * pweight - a * static_cast<float>(x[o * inner_size + i]);
|
|
ref_dx[o * inner_size + i] = static_cast<T>(val);
|
|
});
|
|
});
|
|
}
|
|
|
|
template <class T>
|
|
void cpu_t5layernorm_backward_weight(
|
|
tensor<T> dy, tensor<T> x, tensor<T> rstd, tensor<T>& ref_dw, miopenNormMode_t /*mode*/)
|
|
{
|
|
auto dims = dy.desc.GetLengths();
|
|
size_t outer_size = 1;
|
|
size_t inner_size = dims[dims.size() - 1];
|
|
|
|
for(size_t i = 0ULL; i < dims.size() - 1; ++i)
|
|
{
|
|
outer_size *= dims[i];
|
|
}
|
|
|
|
miopen::par_ford(inner_size)([&](int32_t o) {
|
|
float sum = 0;
|
|
|
|
miopen::ford(outer_size)([&](int32_t i) {
|
|
float prstd = static_cast<float>(rstd[i]);
|
|
float pdy = (dy.GetSize() != 0) ? static_cast<float>(dy[i * inner_size + o]) : 0;
|
|
float px = static_cast<float>(x[i * inner_size + o]);
|
|
|
|
sum += pdy * px * prstd;
|
|
});
|
|
|
|
ref_dw[o] = sum;
|
|
});
|
|
}
|
|
|
|
struct T5LayerNormTestCase
|
|
{
|
|
size_t N;
|
|
size_t C;
|
|
size_t D;
|
|
size_t H;
|
|
size_t W;
|
|
float eps;
|
|
miopenNormMode_t ln_mode;
|
|
friend std::ostream& operator<<(std::ostream& os, const T5LayerNormTestCase& tc)
|
|
{
|
|
return os << " N:" << tc.N << " C:" << tc.C << " D:" << tc.D << " H:" << tc.H
|
|
<< " W:" << tc.W << " eps:" << tc.eps << " LayerNorm_mode:" << tc.ln_mode;
|
|
}
|
|
|
|
std::vector<size_t> GetInput()
|
|
{
|
|
if((N != 0) && (C != 0) && (D != 0) && (H != 0) && (W != 0))
|
|
{
|
|
return std::vector<size_t>({N, C, D, H, W});
|
|
}
|
|
else if((N != 0) && (C != 0) && (H != 0) && (W != 0))
|
|
{
|
|
return std::vector<size_t>({N, C, H, W});
|
|
}
|
|
else if((N != 0) && (C != 0) && (W != 0))
|
|
{
|
|
return std::vector<size_t>({N, C, W});
|
|
}
|
|
else if((N != 0) && (W != 0))
|
|
{
|
|
return std::vector<size_t>({N, W});
|
|
}
|
|
else
|
|
{
|
|
std::cout << "Error Input Tensor Lengths\n" << std::endl;
|
|
return std::vector<size_t>({0});
|
|
}
|
|
}
|
|
};
|
|
|
|
std::vector<T5LayerNormTestCase> T5LayerNormTestConfigs()
|
|
{ // n c d h w eps ln_mode
|
|
// clang-format off
|
|
return {
|
|
{ 32, 1, 32, 32, 32, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // 32x32x32 based on VoxNet arch
|
|
{ 32, 1, 14, 14, 14, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 32, 14, 14, 14, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 32, 12, 12, 12, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 32, 6, 6, 6, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 256, 1, 32, 32, 32, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // 32x32x32 based on VoxNet arch
|
|
{ 256, 32, 14, 14, 14, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 256, 32, 12, 12, 12, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 256, 32, 6, 6, 6, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 512, 1, 32, 32, 32, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // 32x32x32 based on VoxNet arch
|
|
{ 512, 32, 14, 14, 14, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 512, 32, 12, 12, 12, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 512, 32, 6, 6, 6, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 2, 32, 57, 125, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // Hand-gesture recognition CVPR 2015 paper High Res Net Path
|
|
{ 32, 32, 14, 25, 59, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 32, 6, 10, 27, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 32, 4, 6, 11, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 32, 2, 2, 3, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 32, 32, 28, 62, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // Hand-gesture recognition CVPR 2015 paper Low Res Net Path
|
|
{ 32, 32, 14, 12, 29, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 32, 6, 4, 12, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 32, 32, 4, 2, 2, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{ 16, 32, 6, 50, 50, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // Multi-view 3D convnet
|
|
{ 1, 3, 8, 240, 320, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // 3D convet on video
|
|
{ 1, 3, 16, 240, 320, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // 3D convet on video
|
|
{ 1, 3, 8, 128, 171, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // 3D convet on video
|
|
{ 1, 3, 16, 128, 171, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // 3D convet on video
|
|
{ 1, 3, 8, 112, 112, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // 3D convet on video
|
|
{ 1, 3, 16, 112, 112, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5}, // 3D convet on video
|
|
{ 32, 1, 32, 32, 32, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // 32x32x32 based on VoxNet arch
|
|
{ 32, 1, 14, 14, 14, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 32, 14, 14, 14, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 32, 12, 12, 12, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 32, 6, 6, 6, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 256, 1, 32, 32, 32, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // 32x32x32 based on VoxNet arch
|
|
{ 256, 32, 14, 14, 14, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 256, 32, 12, 12, 12, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 256, 32, 6, 6, 6, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 512, 1, 32, 32, 32, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // 32x32x32 based on VoxNet arch
|
|
{ 512, 32, 14, 14, 14, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 512, 32, 12, 12, 12, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 512, 32, 6, 6, 6, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 2, 32, 57, 125, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // Hand-gesture recognition CVPR 2015 paper High Res Net Path
|
|
{ 32, 32, 14, 25, 59, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 32, 6, 10, 27, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 32, 4, 6, 11, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 32, 2, 2, 3, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 32, 32, 28, 62, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // Hand-gesture recognition CVPR 2015 paper Low Res Net Path
|
|
{ 32, 32, 14, 12, 29, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 32, 6, 4, 12, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 32, 32, 4, 2, 2, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{ 16, 32, 6, 50, 50, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // Multi-view 3D convnet
|
|
{ 1, 3, 8, 240, 320, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // 3D convet on video
|
|
{ 1, 3, 16, 240, 320, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // 3D convet on video
|
|
{ 1, 3, 8, 128, 171, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // 3D convet on video
|
|
{ 1, 3, 16, 128, 171, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // 3D convet on video
|
|
{ 1, 3, 8, 112, 112, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // 3D convet on video
|
|
{ 1, 3, 16, 112, 112, 1e-5, MIOPEN_WEIGHT_BIAS_T5}, // 3D convet on video
|
|
{32, 4, 0, 4, 256, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{64, 4, 0, 4, 256, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{32, 4, 0, 4, 256, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{64, 4, 0, 4, 256, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{32, 0, 0, 0, 256, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{64, 0, 0, 0, 256, 1e-5, MIOPEN_ELEMENTWISE_AFFINE_T5},
|
|
{32, 0, 0, 0, 256, 1e-5, MIOPEN_WEIGHT_BIAS_T5},
|
|
{64, 0, 0, 0, 256, 1e-5, MIOPEN_WEIGHT_BIAS_T5}
|
|
};
|
|
// clang-format on
|
|
}
|
|
|
|
template <typename T = float>
|
|
struct T5LayerNormFwdTest : public ::testing::TestWithParam<T5LayerNormTestCase>
|
|
{
|
|
protected:
|
|
void SetUp() override
|
|
{
|
|
auto&& handle = get_handle();
|
|
t5layernorm_config = GetParam();
|
|
auto gen_value = [](auto...) { return prng::gen_descreet_uniform_sign<T>(1e-2, 100); };
|
|
|
|
eps = t5layernorm_config.eps;
|
|
ln_mode = t5layernorm_config.ln_mode;
|
|
|
|
auto in_dim = t5layernorm_config.GetInput();
|
|
x = tensor<T>{in_dim}.generate(gen_value);
|
|
|
|
std::vector<size_t> inner_dim = {in_dim[in_dim.size() - 1]};
|
|
|
|
if(ln_mode == MIOPEN_ELEMENTWISE_AFFINE_T5)
|
|
{
|
|
auto gen_one = [&](auto...) { return 1; };
|
|
weight = tensor<T>{inner_dim}.generate(gen_one);
|
|
}
|
|
else
|
|
{
|
|
weight = tensor<T>{inner_dim}.generate(gen_value);
|
|
}
|
|
|
|
std::vector<size_t> outer_dim;
|
|
|
|
outer_dim = {in_dim.begin(), in_dim.end() - 1};
|
|
|
|
y = tensor<T>{in_dim};
|
|
rstd = tensor<T>{outer_dim};
|
|
std::fill(y.begin(), y.end(), std::numeric_limits<T>::quiet_NaN());
|
|
std::fill(rstd.begin(), rstd.end(), std::numeric_limits<T>::quiet_NaN());
|
|
|
|
ref_y = tensor<T>{in_dim};
|
|
ref_rstd = tensor<T>{outer_dim};
|
|
std::fill(ref_y.begin(), ref_y.end(), std::numeric_limits<T>::quiet_NaN());
|
|
std::fill(ref_rstd.begin(), ref_rstd.end(), std::numeric_limits<T>::quiet_NaN());
|
|
|
|
x_dev = handle.Write(x.data);
|
|
weight_dev = handle.Write(weight.data);
|
|
y_dev = handle.Write(y.data);
|
|
rstd_dev = handle.Write(rstd.data);
|
|
}
|
|
void RunTest()
|
|
{
|
|
auto&& handle = get_handle();
|
|
|
|
cpu_t5layernorm_forward<T>(x, weight, ref_y, ref_rstd, eps, ln_mode);
|
|
|
|
miopenStatus_t status;
|
|
status = miopen::T5LayerNormForward(handle,
|
|
x.desc,
|
|
x_dev.get(),
|
|
weight.desc,
|
|
weight_dev.get(),
|
|
y.desc,
|
|
y_dev.get(),
|
|
rstd.desc,
|
|
rstd_dev.get(),
|
|
ln_mode,
|
|
eps);
|
|
EXPECT_EQ(status, miopenStatusSuccess);
|
|
|
|
y.data = handle.Read<T>(y_dev, y.data.size());
|
|
rstd.data = handle.Read<T>(rstd_dev, rstd.data.size());
|
|
}
|
|
|
|
void Verify()
|
|
{
|
|
// Computation error of fp16 is ~2^13 (=8192) bigger than
|
|
// the one of fp32 because mantissa is shorter by 13 bits.
|
|
// In the case of layernorm, there is a cumulative sum operation, and in the case of
|
|
// floating point operation, the result value can change if the order of the summed values
|
|
// is changed. So apply a threshold that is 10 times larger than other operations.
|
|
auto threshold = std::is_same<T, float>::value ? 1.5e-5 : 8.2e-2;
|
|
|
|
// bf16 mantissa has 7 bits, by 3 bits shorter than fp16.
|
|
if(std::is_same<T, bfloat16>::value)
|
|
threshold *= 8.0;
|
|
auto error = miopen::rms_range(ref_y, y);
|
|
EXPECT_TRUE(miopen::range_distance(ref_y) == miopen::range_distance(y));
|
|
EXPECT_TRUE(error < threshold)
|
|
<< "Error y beyond tolerance Error:" << error << ", Threshold: " << threshold;
|
|
|
|
error = miopen::rms_range(ref_rstd, rstd);
|
|
EXPECT_TRUE(miopen::range_distance(ref_rstd) == miopen::range_distance(rstd));
|
|
EXPECT_TRUE(error < threshold * 4) << "Error rstd beyond tolerance Error:" << error
|
|
<< ", Threshold x 4: " << threshold * 4;
|
|
}
|
|
T5LayerNormTestCase t5layernorm_config;
|
|
|
|
tensor<T> x;
|
|
tensor<T> weight;
|
|
tensor<T> y;
|
|
tensor<T> rstd;
|
|
|
|
tensor<T> ref_y;
|
|
tensor<T> ref_rstd;
|
|
|
|
miopen::Allocator::ManageDataPtr x_dev;
|
|
miopen::Allocator::ManageDataPtr weight_dev;
|
|
miopen::Allocator::ManageDataPtr y_dev;
|
|
miopen::Allocator::ManageDataPtr rstd_dev;
|
|
|
|
float eps;
|
|
miopenNormMode_t ln_mode;
|
|
};
|
|
|
|
template <typename T = float>
|
|
struct T5LayerNormBwdTest : public ::testing::TestWithParam<T5LayerNormTestCase>
|
|
{
|
|
protected:
|
|
void SetUp() override
|
|
{
|
|
auto&& handle = get_handle();
|
|
t5layernorm_config = GetParam();
|
|
auto gen_value = [](auto...) { return prng::gen_descreet_uniform_sign<T>(1e-2, 100); };
|
|
|
|
ln_mode = t5layernorm_config.ln_mode;
|
|
|
|
auto in_dim = t5layernorm_config.GetInput();
|
|
std::vector<size_t> outer_dim = {in_dim.begin(), in_dim.end() - 1};
|
|
|
|
x = tensor<T>{in_dim}.generate(gen_value);
|
|
dy = tensor<T>{in_dim}.generate(gen_value);
|
|
rstd = tensor<T>{outer_dim}.generate(gen_value);
|
|
|
|
std::vector<size_t> inner_dim = {in_dim[in_dim.size() - 1]};
|
|
|
|
if(ln_mode == MIOPEN_ELEMENTWISE_AFFINE_T5)
|
|
{
|
|
auto gen_one = [&](auto...) { return 1; };
|
|
weight = tensor<T>{inner_dim}.generate(gen_one);
|
|
}
|
|
else
|
|
{
|
|
weight = tensor<T>{inner_dim}.generate(gen_value);
|
|
}
|
|
|
|
dx = tensor<T>{in_dim};
|
|
dw = tensor<T>{inner_dim};
|
|
std::fill(dx.begin(), dx.end(), std::numeric_limits<T>::quiet_NaN());
|
|
std::fill(dw.begin(), dw.end(), std::numeric_limits<T>::quiet_NaN());
|
|
|
|
ref_dx = tensor<T>{in_dim};
|
|
ref_dw = tensor<T>{inner_dim};
|
|
std::fill(ref_dx.begin(), ref_dx.end(), std::numeric_limits<T>::quiet_NaN());
|
|
std::fill(ref_dw.begin(), ref_dw.end(), std::numeric_limits<T>::quiet_NaN());
|
|
|
|
std::vector<size_t> workspace_dims;
|
|
|
|
ws_sizeInBytes = miopen::GetT5LayerNormBackwardWorkspaceSize(
|
|
handle, dy.desc, x.desc, weight.desc, rstd.desc, dx.desc, dw.desc, ln_mode);
|
|
if(ws_sizeInBytes == static_cast<size_t>(-1))
|
|
GTEST_SKIP();
|
|
|
|
workspace_dims.push_back(ws_sizeInBytes / sizeof(T));
|
|
if(ws_sizeInBytes != 0)
|
|
{
|
|
workspace = tensor<T>{workspace_dims};
|
|
std::fill(workspace.begin(), workspace.end(), std::numeric_limits<T>::quiet_NaN());
|
|
workspace_dev = handle.Write(workspace.data);
|
|
}
|
|
|
|
x_dev = handle.Write(x.data);
|
|
weight_dev = handle.Write(weight.data);
|
|
rstd_dev = handle.Write(rstd.data);
|
|
dy_dev = handle.Write(dy.data);
|
|
dx_dev = handle.Write(dx.data);
|
|
dw_dev = handle.Write(dw.data);
|
|
}
|
|
void RunTest()
|
|
{
|
|
auto&& handle = get_handle();
|
|
cpu_t5layernorm_backward<T>(dy, x, weight, rstd, ref_dx, ln_mode);
|
|
cpu_t5layernorm_backward_weight<T>(dy, x, rstd, ref_dw, ln_mode);
|
|
|
|
miopenStatus_t status;
|
|
|
|
status = miopen::T5LayerNormBackward(handle,
|
|
workspace_dev.get(),
|
|
ws_sizeInBytes,
|
|
dy.desc,
|
|
dy_dev.get(),
|
|
x.desc,
|
|
x_dev.get(),
|
|
weight.desc,
|
|
weight_dev.get(),
|
|
rstd.desc,
|
|
rstd_dev.get(),
|
|
dx.desc,
|
|
dx_dev.get(),
|
|
dw.desc,
|
|
dw_dev.get(),
|
|
ln_mode);
|
|
|
|
EXPECT_EQ(status, miopenStatusSuccess);
|
|
|
|
dx.data = handle.Read<T>(dx_dev, dx.data.size());
|
|
dw.data = handle.Read<T>(dw_dev, dw.data.size());
|
|
}
|
|
|
|
void Verify()
|
|
{
|
|
// Computation error of fp16 is ~2^13 (=8192) bigger than
|
|
// the one of fp32 because mantissa is shorter by 13 bits.
|
|
// In the case of layernorm, there is a cumulative sum operation, and in the case of
|
|
// floating point operation, the result value can change if the order of the summed values
|
|
// is changed. So apply a threshold that is 10 times larger than other operations.
|
|
auto threshold = std::is_same<T, float>::value ? 1.5e-5 : 8.2e-2;
|
|
|
|
// bf16 mantissa has 7 bits, by 3 bits shorter than fp16.
|
|
if(std::is_same<T, bfloat16>::value)
|
|
threshold *= 80.0;
|
|
|
|
auto error = miopen::rms_range(ref_dx, dx);
|
|
EXPECT_TRUE(miopen::range_distance(ref_dx) == miopen::range_distance(dx));
|
|
EXPECT_TRUE(error < threshold)
|
|
<< "Error dx beyond tolerance Error:" << error << ", Threshold: " << threshold;
|
|
error = miopen::rms_range(ref_dw, dw);
|
|
EXPECT_TRUE(miopen::range_distance(ref_dw) == miopen::range_distance(dw));
|
|
EXPECT_TRUE(error < threshold * 2)
|
|
<< "Error dw beyond tolerance Error:" << error << ", Threshold x 2: " << threshold * 2;
|
|
}
|
|
T5LayerNormTestCase t5layernorm_config;
|
|
|
|
tensor<T> x;
|
|
tensor<T> weight;
|
|
tensor<T> rstd;
|
|
tensor<T> dy;
|
|
tensor<T> dx;
|
|
tensor<T> dw;
|
|
tensor<T> workspace;
|
|
|
|
tensor<T> ref_dx;
|
|
tensor<T> ref_dw;
|
|
|
|
miopen::Allocator::ManageDataPtr x_dev;
|
|
miopen::Allocator::ManageDataPtr weight_dev;
|
|
miopen::Allocator::ManageDataPtr rstd_dev;
|
|
miopen::Allocator::ManageDataPtr dy_dev;
|
|
miopen::Allocator::ManageDataPtr dx_dev;
|
|
miopen::Allocator::ManageDataPtr dw_dev;
|
|
miopen::Allocator::ManageDataPtr workspace_dev;
|
|
|
|
size_t ws_sizeInBytes;
|
|
|
|
miopenNormMode_t ln_mode;
|
|
};
|