+
Skip to content

Upgrade to cudarc==0.7.0 and use alloc_async instead of alloc_zeros_async #440

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

Merged
merged 1 commit into from
Feb 9, 2023
Merged

Upgrade to cudarc==0.7.0 and use alloc_async instead of alloc_zeros_async #440

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,7 @@ matrixmultiply = { version = "0.3.2", default-features = false }
zip = { version = "0.6.2", default-features = false, optional = true }
cblas-sys = { version = "0.1.4", default-features = false, optional = true }
libc = { version = "0.2", default-features = false, optional = true }
cudarc = { version = "0.6.1", default-features = false, optional = true }
cudarc = { version = "0.7.0", default-features = false, optional = true }

[features]
default = ["std", "numpy"]
Expand Down
1 change: 0 additions & 1 deletion build.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,6 @@ fn main() {
#[cfg(feature = "cuda")]
mod cuda {
pub fn build_ptx() {
// TODO build ptx file in source tree and don't call nvcc if so
let out_dir = std::env::var("OUT_DIR").unwrap();
let kernel_paths: Vec<std::path::PathBuf> = glob::glob("src/**/*.cu")
.unwrap()
Expand Down
10 changes: 3 additions & 7 deletions src/tensor_ops/boolean/cuda_kernels.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
use super::BooleanKernel;
use crate::prelude::{cuda::CudaArray, *};
use cudarc::prelude::*;
use cudarc::driver::*;

use std::sync::Arc;

Expand All @@ -24,10 +24,8 @@ impl Cuda {
let strides = lhs.shape.strides();
let numel = shape.num_elements();

let mut storage = self.dev.take_async(std::vec![false; numel])?;
let mut storage = unsafe { self.dev.alloc_async(numel) }?;

// TODO: modify this to be `self.dev.alloc_zeros_async(numel)?` once cudarc implements
// ValidAsZeroBits for bool
let dims: CudaSlice<usize> = self.dev.take_async(shape.concrete().into())?;
let lhs_strides: CudaSlice<usize> = self.dev.take_async(lhs.strides.into())?;
let rhs_strides: CudaSlice<usize> = self.dev.take_async(rhs.strides.into())?;
Expand Down Expand Up @@ -64,9 +62,7 @@ impl BooleanKernel for Cuda {
}

let numel = inp.data.len();
// TODO: modify this to be `self.dev.alloc_zeros_async(numel)?` once cudarc implements
// ValidAsZeroBits for bool
let mut storage = self.dev.take_async(std::vec![false; numel])?;
let mut storage = unsafe { self.dev.alloc_async(numel) }?;

let fwd_fn = self.dev.get_func(MODULE_NAME, "boolean_not").unwrap();
let cfg = LaunchConfig::for_num_elems(numel as u32);
Expand Down
2 changes: 1 addition & 1 deletion src/tensor_ops/choose/cuda_kernel.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@ impl ChooseKernel<f32> for Cuda {
let strides = lhs.shape.strides();
let numel = shape.num_elements();

let mut storage = self.dev.alloc_zeros_async::<f32>(numel)?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(numel) }?;

let dims: CudaSlice<usize> = self.dev.take_async(shape.concrete().into())?;
let cond_strides: CudaSlice<usize> = self.dev.take_async(cond.strides.into())?;
Expand Down
2 changes: 1 addition & 1 deletion src/tensor_ops/dropout/cuda_kernel.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,7 +36,7 @@ impl UnaryKernel<super::DropoutKernelOp, f32> for Cuda {
}

let numel = inp.data.len();
let mut storage = self.dev.alloc_zeros_async::<f32>(numel)?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(numel) }?;

let fwd_fn = self.dev.get_func(MODULE_NAME, FWD_FN_NAME).unwrap();
let cfg = LaunchConfig::for_num_elems(numel as u32);
Expand Down
12 changes: 6 additions & 6 deletions src/tensor_ops/matmul/cuda_kernel.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -164,7 +164,7 @@ impl super::VecVecKernel<f32> for Cuda {
let k = Const::<1>;
let shape = (m, n);
let strides = shape.strides();
let mut storage = self.dev.alloc_zeros_async::<f32>(shape.num_elements())?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(shape.num_elements()) }?;

unsafe {
sgemm(
Expand Down Expand Up @@ -239,7 +239,7 @@ impl super::VecMatKernel<f32> for Cuda {
let (k, n) = rhs.shape;
let shape = (n,);
let strides = shape.strides();
let mut storage = self.dev.alloc_zeros_async::<f32>(shape.num_elements())?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(shape.num_elements()) }?;

unsafe {
sgemm(
Expand Down Expand Up @@ -312,7 +312,7 @@ impl super::MatMatKernel<f32> for Cuda {
let (k, n) = rhs.shape;
let shape = (m, n);
let strides = shape.strides();
let mut storage = self.dev.alloc_zeros_async::<f32>(shape.num_elements())?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(shape.num_elements()) }?;

unsafe {
sgemm(
Expand Down Expand Up @@ -386,7 +386,7 @@ impl super::MatMatBrKernel<f32> for Cuda {
let (k, n) = rhs.shape;
let shape = (batch, m, n);
let strides = shape.strides();
let mut storage = self.dev.alloc_zeros_async::<f32>(shape.num_elements())?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(shape.num_elements()) }?;

unsafe {
sgemm_batch(
Expand Down Expand Up @@ -465,7 +465,7 @@ impl super::MatMatBatch3Kernel<f32> for Cuda {
let (_, k, n) = rhs.shape;
let shape = (batch, m, n);
let strides = shape.strides();
let mut storage = self.dev.alloc_zeros_async::<f32>(shape.num_elements())?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(shape.num_elements()) }?;

unsafe {
sgemm_batch(
Expand Down Expand Up @@ -539,7 +539,7 @@ impl super::MatMatBatch4Kernel<f32> for Cuda {
let (_, _, k, n) = rhs.shape;
let shape = (batch, seq, m, n);
let strides = shape.strides();
let mut storage = self.dev.alloc_zeros_async::<f32>(shape.num_elements())?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(shape.num_elements()) }?;

for b in 0..batch.size() {
// TODO: use separate streams
Expand Down
9 changes: 5 additions & 4 deletions src/tensor_ops/max_to/cuda_kernel.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,14 +28,15 @@ impl super::MaxReduceKernel<f32> for Cuda {
.load_ptx(PTX_SRC.into(), MODULE_NAME, &ALL_FN_NAMES)?;
}

let mut storage = self.dev.alloc_zeros_async::<f32>(dst.num_elements())?;
let fill_fn = self.dev.get_func(MODULE_NAME, "fill_with").unwrap();
unsafe {
let mut storage = unsafe {
let mut storage = self.dev.alloc_async::<f32>(dst.num_elements())?;
fill_fn.launch_async(
LaunchConfig::for_num_elems(dst.num_elements() as u32),
(&mut storage, f32::NEG_INFINITY, dst.num_elements()),
)
}?;
)?;
storage
};

let fwd_fn = self.dev.get_func(MODULE_NAME, FWD_FN_NAME).unwrap();

Expand Down
9 changes: 5 additions & 4 deletions src/tensor_ops/min_to/cuda_kernel.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,14 +28,15 @@ impl super::MinReduceKernel<f32> for Cuda {
.load_ptx(PTX_SRC.into(), MODULE_NAME, &ALL_FN_NAMES)?;
}

let mut storage = self.dev.alloc_zeros_async::<f32>(dst.num_elements())?;
let fill_fn = self.dev.get_func(MODULE_NAME, "fill_with").unwrap();
unsafe {
let mut storage = unsafe {
let mut storage = self.dev.alloc_async::<f32>(dst.num_elements())?;
fill_fn.launch_async(
LaunchConfig::for_num_elems(dst.num_elements() as u32),
(&mut storage, f32::INFINITY, dst.num_elements()),
)
}?;
)?;
storage
};

let fwd_fn = self.dev.get_func(MODULE_NAME, FWD_FN_NAME).unwrap();

Expand Down
2 changes: 1 addition & 1 deletion src/tensor_ops/reshape_to/cuda_kernel.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,7 @@ impl super::ReshapeKernel<f32> for Cuda {
}

let numel = inp.data.len();
let mut storage = self.dev.alloc_zeros_async::<f32>(numel)?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(numel) }?;

let inp_dims: CudaSlice<usize> = self.dev.take_async(inp.shape.concrete().into())?;
let dst_dims: CudaSlice<usize> = self.dev.take_async(dst.concrete().into())?;
Expand Down
4 changes: 2 additions & 2 deletions src/tensor_ops/utilities/cuda_kernels.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ impl<K: UnaryOpCudaKernel + AsKernelParam> UnaryKernel<K, f32> for Cuda {
}

let numel = inp.data.len();
let mut storage = self.dev.alloc_zeros_async::<f32>(numel)?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(numel) }?;

let fwd_fn = self.dev.get_func(K::MODULE_NAME, K::FWD_FN_NAME).unwrap();
let cfg = LaunchConfig::for_num_elems(numel as u32);
Expand Down Expand Up @@ -107,7 +107,7 @@ impl<K: BinaryOpCudaKernel + AsKernelParam> BinaryKernel<K, f32> for Cuda {
let strides = lhs.shape.strides();
let numel = shape.num_elements();

let mut storage = self.dev.alloc_zeros_async::<f32>(numel)?;
let mut storage = unsafe { self.dev.alloc_async::<f32>(numel) }?;

let dims: CudaSlice<usize> = self.dev.take_async(shape.concrete().into())?;
let lhs_strides: CudaSlice<usize> = self.dev.take_async(lhs.strides.into())?;
Expand Down
点击 这是indexloc提供的php浏览器服务，不要输入任何密码和下载