Merge pull request filecoin-project#99 from filecoin-project/use-gpu-…

…tools

refact(gpu): big refactoring and use rust-gpu-tools

Cargo.toml

@@ -24,14 +24,14 @@ rand_core = "0.5"
 byteorder = "1"
 log = "0.4.8"
 lazy_static = "1.4.0"
-ocl = { version = "0.19.4", package = "fil-ocl", optional = true }
 ff-cl-gen = { version = "0.1.1", optional = true }
 fs2 = { version = "0.4.3", optional = true }
 rand = "0.7"
 rayon = "1.3.0"
 memmap = "0.7.0"
 thiserror = "1.0.10"
 ahash = "0.3.4"
+rust-gpu-tools = { version = "0.1.0", optional = true }
 
 [dev-dependencies]
 hex-literal = "0.2"
@@ -42,7 +42,7 @@ criterion = "0.3.2"
 
 [features]
 default = ["groth16", "multicore"]
-gpu = ["ocl", "ff-cl-gen", "fs2"]
+gpu = ["rust-gpu-tools", "ff-cl-gen", "fs2"]
 groth16 = ["paired"]
 multicore = ["futures-cpupool", "crossbeam", "num_cpus"]
 

README.md

@@ -11,7 +11,7 @@ booleans and number abstractions.
 This fork contains GPU parallel acceleration to the FFT and Multiexponentation algorithms in the groth16 prover codebase under a conditional compilation feature `#[cfg(feature = "gpu")]` and `gpu-test` for testing.
 
 ### Requirements
-- NVIDIA or AMD GPU Graphics Driver 
+- NVIDIA or AMD GPU Graphics Driver
 - OpenCL
 
 ( For AMD devices we recommend [ROCm](https://rocm-documentation.readthedocs.io/en/latest/Installation_Guide/Installation-Guide.html) )
@@ -29,31 +29,15 @@ The gpu extension contains some env vars that may be set externally to this libr
     env::set_var("BELLMAN_NO_GPU", "1");
     ```
 
-- `BELLMAN_PLATFORM`
+- `BELLMAN_VERIFIER`
 
-    Can be used to select the default OpenCL platform:
+    Chooses the device in which the batched verifier is going to run. Can be `cpu`, `gpu` or `auto`.
 
     ```rust
-    // Example
-    env::set_var("BELLMAN_PLATFORM", "AMD Accelerated Parallel Processing");
+    Example
+    env::set_var("BELLMAN_VERIFIER", "gpu");
     ```
 
-    Some possible values:
-
-`BELLMAN_VERIFIER`
-
-Chooses the device in which the batched verifier is going to run. Can be `cpu`, `gpu` or `auto`.
-
-```
-Example
-env::set_var("BELLMAN_VERIFIER", "gpu");
-```
-
-  - NVIDIA CUDA
-  - AMD Accelerated Parallel Processing
-
-    If not set, and the code does not select any platform, "NVIDIA CUDA" will be selected.
-
 - `BELLMAN_CUSTOM_GPU`
 
     Will allow for adding a GPU not in the tested list. This requires researching the name of the GPU device and the number of cores in the format `["name:cores"]`.
@@ -106,12 +90,11 @@ To run the multiexp_consistency test you can use:
 RUST_LOG=info cargo test --features gpu -- --exact multiexp::gpu_multiexp_consistency --nocapture
 ```
 
-to run on some specific platform you can do
+### Considerations
 
-```bash
-export BELLMAN_PLATFORM="AMD Accelerated Parallel Processing" 
-RUST_LOG=info cargo test --features gpu -- --exact multiexp::gpu_multiexp_consistency --nocapture
-```
+Bellperson uses `rust-gpu-tools` as its OpenCL backend, therefore you may see a
+directory named `~/.rust-gpu-tools` in your home folder, which contains the
+compiled binaries of OpenCL kernels used in this repository.
 
 ## License
 

src/domain.rs

@@ -559,11 +559,11 @@ fn parallel_fft_consistency() {
     test_consistency::<Bls12, _>(rng);
 }
 
-pub fn create_fft_kernel<E>(log_d: usize, priority: bool) -> Option<gpu::FFTKernel<E>>
+pub fn create_fft_kernel<E>(_log_d: usize, priority: bool) -> Option<gpu::FFTKernel<E>>
 where
     E: Engine,
 {
-    match gpu::FFTKernel::create(1 << log_d, priority) {
+    match gpu::FFTKernel::create(priority) {
         Ok(k) => {
             info!("GPU FFT kernel instantiated!");
             Some(k)
@@ -594,7 +594,7 @@ mod tests {
 
         let worker = Worker::new();
         let log_cpus = worker.log_num_cpus();
-        let mut kern = gpu::FFTKernel::create(1 << 24, false).expect("Cannot initialize kernel!");
+        let mut kern = gpu::FFTKernel::create(false).expect("Cannot initialize kernel!");
 
         for log_d in 1..25 {
             let d = 1 << log_d;

src/gpu/error.rs

@@ -1,27 +1,25 @@
+#[cfg(feature = "gpu")]
+use rust_gpu_tools::opencl;
+
 #[derive(thiserror::Error, Debug)]
 pub enum GPUError {
     #[error("GPUError: {0}")]
     Simple(&'static str),
     #[cfg(feature = "gpu")]
-    #[error("Ocl Error: {0}")]
-    Ocl(ocl::Error),
+    #[error("OpenCL Error: {0}")]
+    OpenCL(#[from] opencl::GPUError),
     #[cfg(feature = "gpu")]
     #[error("GPU taken by a high priority process!")]
     GPUTaken,
     #[cfg(feature = "gpu")]
     #[error("No kernel is initialized!")]
     KernelUninitialized,
+    #[error("GPU accelerator is disabled!")]
+    GPUDisabled,
 }
 
 pub type GPUResult<T> = std::result::Result<T, GPUError>;
 
-#[cfg(feature = "gpu")]
-impl From<ocl::Error> for GPUError {
-    fn from(error: ocl::Error) -> Self {
-        GPUError::Ocl(error)
-    }
-}
-
 #[cfg(feature = "gpu")]
 impl From<std::boxed::Box<dyn std::any::Any + std::marker::Send>> for GPUError {
     fn from(e: std::boxed::Box<dyn std::any::Any + std::marker::Send>) -> Self {