bump cudarc

.github/workflows/cuda.yml

@@ -23,6 +23,7 @@ jobs:
     - run: |
          . .travis/ci-system-setup.sh
          echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"
+         echo "/usr/local/cuda/bin" >> "$GITHUB_PATH"
       env:
        RUNNER_ENVIRONMENT: ${{ runner.environment }}
 
@@ -45,6 +46,7 @@ jobs:
       - run: |
            . .travis/ci-system-setup.sh
            echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"
+           echo "/usr/local/cuda/bin" >> "$GITHUB_PATH"
         env:
          RUNNER_ENVIRONMENT: ${{ runner.environment }}
 

.travis/ci-system-setup.sh

@@ -39,6 +39,12 @@ then
     touch /tmp/ci-setup-done
 fi
 
+if [ -e /usr/local/cuda ]
+then
+    PATH=$PATH:/usr/local/cuda/bin
+    nvcc --version
+fi
+
 S3=https://s3.amazonaws.com/tract-ci-builds/tests
 
 if  [ "$GITHUB_WORKFLOW" = "Metal tests" -o "$GITHUB_WORKFLOW" = "CUDA tests" ] 

Cargo.toml

@@ -138,7 +138,7 @@ clap = { version = "~3.1", features = [ "cargo" ] }
 colorous = "1.0.5"
 core_affinity = "0.8.0"
 criterion = "0.6"
-cudarc = { version = "0.16.4", features = ["dynamic-loading", "cuda-12060", "f16"] }
+cudarc = { version = "0.17", features = ["dynamic-loading", "cuda-version-from-build-system", "f16"] }
 derive-new = "0.5.9"
 dinghy-test = "0.6"
 downcast-rs = "1.2.0"

cuda/src/kernels/mod.rs

@@ -9,6 +9,8 @@ mod unary;
 mod utils;
 
 use crate::ops::GgmlQuantQ81Fact;
+use std::mem::transmute;
+
 use crate::tensor::{CudaBuffer, CudaTensor};
 use anyhow::{bail, ensure};
 pub use binary::BinOps;
@@ -159,7 +161,9 @@ pub fn get_sliced_cuda_view_mut(
     len: usize,
 ) -> TractResult<CudaViewMut<'_, u8>> {
     ensure!(offset + len <= t.len() * t.datum_type().size_of());
-    let mut buffer = t.device_buffer().downcast_ref::<CudaBuffer>().unwrap();
+    let buffer: &CudaBuffer = t.device_buffer().downcast_ref::<CudaBuffer>().unwrap();
     let offset = t.buffer_offset::<usize>() + offset;
-    Ok(buffer.as_view_mut().slice_mut(offset..(offset + len)))
+    let ptr: *const CudaBuffer = buffer;
+    let mut_buffer: &mut CudaBuffer = unsafe { (ptr as *mut CudaBuffer).as_mut().unwrap() };
+    Ok(mut_buffer.as_view_mut().slice_mut(offset..(offset + len)))
 }

cuda/src/tensor.rs

@@ -1,4 +1,4 @@
-use std::ops::Deref;
+use std::ops::{Deref, DerefMut};
 
 use cudarc::driver::{CudaSlice, DevicePtr};
 use tract_core::internal::tract_smallvec::ToSmallVec;
@@ -30,6 +30,12 @@ impl Deref for CudaBuffer {
     }
 }
 
+impl DerefMut for CudaBuffer {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        &mut self.inner
+    }
+}
+
 #[derive(Clone)]
 pub struct CudaTensor {
     buffer: Arc<CudaBuffer>,