Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add bfloat16 support for several operators and apis. #52696

Merged
merged 6 commits into from
Apr 9, 2023
Merged

Add bfloat16 support for several operators and apis. #52696

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
da4ba01
Cherry-pick the register of bfloat16 for amp_kernel, pull request #45…
Xreki Apr 9, 2023
1387e55
Cherry-pick the master_grad support of adamw, pull request #51141.
Xreki Apr 9, 2023
6080005
add bf16 for some ops in static mode (#51582)
shaojiewang Apr 9, 2023
7c00a83
Add bfloat16 support for some api in static mode.
Xreki Apr 9, 2023
14da52e
Fix codestyle.
Xreki Apr 9, 2023
77edf2e
Revert the change of layer_function_generator.py.
Xreki Apr 9, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
154 changes: 108 additions & 46 deletions paddle/phi/kernels/gpu/adamw_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,7 @@
#include "paddle/phi/kernels/funcs/for_range.h"

namespace phi {
template <typename T, typename MT>
template <typename T, typename TG, typename MT>
__global__ void AdamWKernelREG(MT beta1,
MT beta2,
MT epsilon,
Expand All @@ -42,7 +42,7 @@ __global__ void AdamWKernelREG(MT beta1,
const MT* moment2,
MT* moment2_out,
const MT* lr_,
const T* grad,
const TG* grad,
const T* param,
T* param_out,
const MT* master_param,
Expand Down Expand Up @@ -78,7 +78,7 @@ __global__ void AdamWKernelREG(MT beta1,
}
}

template <typename T, typename MT>
template <typename T, typename TG, typename MT>
__global__ void AdamWKernelMEM(MT beta1,
MT beta2,
MT epsilon,
Expand All @@ -91,7 +91,7 @@ __global__ void AdamWKernelMEM(MT beta1,
const MT* moment2,
MT* moment2_out,
const MT* lr_,
const T* grad,
const TG* grad,
const T* param,
T* param_out,
const MT* master_param,
Expand Down Expand Up @@ -167,6 +167,8 @@ void AdamwDenseKernel(const Context& dev_ctx,
DenseTensor* master_param_outs) {
using MPDType = typename phi::dtype::MPTypeTrait<T>::Type;

const auto grad_type = grad.dtype();

VLOG(4) << "use_global_beta_pow:" << use_global_beta_pow;

MPDType coeff_ = static_cast<MPDType>(coeff);
Expand All @@ -191,8 +193,10 @@ void AdamwDenseKernel(const Context& dev_ctx,
phi::Copy(dev_ctx, param, dev_ctx.GetPlace(), false, param_out);
phi::Copy(dev_ctx, moment1, dev_ctx.GetPlace(), false, moment1_out);
phi::Copy(dev_ctx, moment2, dev_ctx.GetPlace(), false, moment2_out);
phi::Copy(dev_ctx, beta1_pow, beta1_pow.place(), false, beta1_pow_out);
phi::Copy(dev_ctx, beta2_pow, beta2_pow.place(), false, beta2_pow_out);
if (!use_global_beta_pow) {
phi::Copy(dev_ctx, beta1_pow, beta1_pow.place(), false, beta1_pow_out);
phi::Copy(dev_ctx, beta2_pow, beta2_pow.place(), false, beta2_pow_out);
}
return;
}

Expand Down Expand Up @@ -233,25 +237,49 @@ void AdamwDenseKernel(const Context& dev_ctx,

if (beta1_pow.place() == CPUPlace() && beta2_pow.place() == CPUPlace()) {
// Compute with betapow in REG
AdamWKernelREG<T, MPDType><<<blocks, threads, 0, dev_ctx.stream()>>>(
beta1_,
beta2_,
epsilon_,
coeff_,
lr_ratio_,
*beta1_pow.data<MPDType>(),
*beta2_pow.data<MPDType>(),
moment1.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment1_out),
moment2.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment2_out),
learning_rate.data<MPDType>(),
grad.data<T>(),
param.data<T>(),
dev_ctx.template Alloc<T>(param_out),
master_in_data,
master_out_data,
param.numel());
if (grad_type == phi::DataType::FLOAT32)
AdamWKernelREG<T, float, MPDType>
<<<blocks, threads, 0, dev_ctx.stream()>>>(
beta1_,
beta2_,
epsilon_,
coeff_,
lr_ratio_,
*beta1_pow.data<MPDType>(),
*beta2_pow.data<MPDType>(),
moment1.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment1_out),
moment2.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment2_out),
learning_rate.data<MPDType>(),
grad.data<float>(),
param.data<T>(),
dev_ctx.template Alloc<T>(param_out),
master_in_data,
master_out_data,
param.numel());

else

AdamWKernelREG<T, T, MPDType><<<blocks, threads, 0, dev_ctx.stream()>>>(
beta1_,
beta2_,
epsilon_,
coeff_,
lr_ratio_,
*beta1_pow.data<MPDType>(),
*beta2_pow.data<MPDType>(),
moment1.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment1_out),
moment2.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment2_out),
learning_rate.data<MPDType>(),
grad.data<T>(),
param.data<T>(),
dev_ctx.template Alloc<T>(param_out),
master_in_data,
master_out_data,
param.numel());
if (!use_global_beta_pow) {
// Cpu update
dev_ctx.template HostAlloc<MPDType>(beta1_pow_out)[0] =
Expand All @@ -260,28 +288,50 @@ void AdamwDenseKernel(const Context& dev_ctx,
beta2_ * beta2_pow.data<MPDType>()[0];
}
} else {
AdamWKernelMEM<T, MPDType><<<blocks, threads, 0, dev_ctx.stream()>>>(
beta1_,
beta2_,
epsilon_,
coeff_,
lr_ratio_,
beta1_pow.data<MPDType>(),
beta2_pow.data<MPDType>(),
moment1.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment1_out),
moment2.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment2_out),
learning_rate.data<MPDType>(),
grad.data<T>(),
param.data<T>(),
dev_ctx.template Alloc<T>(param_out),
master_in_data,
master_out_data,
param.numel());
if (grad_type == phi::DataType::FLOAT32)
AdamWKernelMEM<T, float, MPDType>
<<<blocks, threads, 0, dev_ctx.stream()>>>(
beta1_,
beta2_,
epsilon_,
coeff_,
lr_ratio_,
beta1_pow.data<MPDType>(),
beta2_pow.data<MPDType>(),
moment1.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment1_out),
moment2.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment2_out),
learning_rate.data<MPDType>(),
grad.data<float>(),
param.data<T>(),
dev_ctx.template Alloc<T>(param_out),
master_in_data,
master_out_data,
param.numel());
else
AdamWKernelMEM<T, T, MPDType><<<blocks, threads, 0, dev_ctx.stream()>>>(
beta1_,
beta2_,
epsilon_,
coeff_,
lr_ratio_,
beta1_pow.data<MPDType>(),
beta2_pow.data<MPDType>(),
moment1.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment1_out),
moment2.data<MPDType>(),
dev_ctx.template Alloc<MPDType>(moment2_out),
learning_rate.data<MPDType>(),
grad.data<T>(),
param.data<T>(),
dev_ctx.template Alloc<T>(param_out),
master_in_data,
master_out_data,
param.numel());
if (!use_global_beta_pow) {
// Update with gpu
UpdateAdamWBetaPow<MPDType><<<1, 32, 0, dev_ctx.stream()>>>(
UpdateAdamWBetaPow<MPDType><<<1, 1, 0, dev_ctx.stream()>>>(
beta1_,
beta2_,
beta1_pow.data<MPDType>(),
Expand All @@ -300,9 +350,21 @@ PD_REGISTER_KERNEL(adamw,
phi::AdamwDenseKernel,
float,
double,
phi::dtype::float16) {
phi::dtype::float16,
phi::dtype::bfloat16) {
// Skip beta1_pow, beta2_pow, skip_update data transform
kernel->InputAt(5).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(6).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(8).SetBackend(phi::Backend::ALL_BACKEND);

if (kernel_key.dtype() == phi::DataType::FLOAT16 ||
kernel_key.dtype() == phi::DataType::BFLOAT16) {
kernel->OutputAt(1).SetDataType(phi::DataType::FLOAT32);
kernel->OutputAt(2).SetDataType(phi::DataType::FLOAT32);
kernel->OutputAt(3).SetDataType(phi::DataType::FLOAT32);
kernel->OutputAt(4).SetDataType(phi::DataType::FLOAT32);
kernel->OutputAt(5).SetDataType(phi::DataType::FLOAT32);
}
kernel->OutputAt(3).SetBackend(phi::Backend::UNDEFINED);
kernel->OutputAt(4).SetBackend(phi::Backend::UNDEFINED);
}
6 changes: 4 additions & 2 deletions paddle/phi/kernels/gpu/amp_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -357,14 +357,16 @@ PD_REGISTER_KERNEL(check_finite_and_unscale,
phi::CheckFiniteAndUnscaleKernel,
float,
double,
phi::dtype::float16) {}
phi::dtype::float16,
phi::dtype::bfloat16) {}

PD_REGISTER_KERNEL(update_loss_scaling,
GPU,
ALL_LAYOUT,
phi::UpdateLossScalingKernel,
float,
double,
phi::dtype::float16) {
phi::dtype::float16,
phi::dtype::bfloat16) {
kernel->InputAt(1).SetBackend(phi::Backend::ALL_BACKEND);
}
1 change: 1 addition & 0 deletions paddle/phi/kernels/gpu/matmul_grad_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,6 +55,7 @@ PD_REGISTER_KERNEL(matmul_with_flatten_grad,
phi::MatmulWithFlattenGradKernel,
float,
double,
phi::dtype::bfloat16,
phi::dtype::float16) {}

PD_REGISTER_KERNEL(matmul_with_flatten_double_grad,
Expand Down
1 change: 1 addition & 0 deletions paddle/phi/kernels/gpu/matmul_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,4 +36,5 @@ PD_REGISTER_KERNEL(matmul_with_flatten,
phi::MatmulWithFlattenKernel,
float,
double,
phi::dtype::bfloat16,
phi::dtype::float16) {}
Loading