Implement base Webgpu device (#893)

* Implement Webgpu device.

All the tests pass, but it doesn't implement `Device` yet, because that
is a lot more work.

* Removed some of the more low level commands in favor of a wrapper struct

Also added tests for higher code coverage.

* AtomicPtr unsound fix

* Partial implementation of `Device<E>` for Webgpu

* Finished up with template tensor_ops

* Remove foolish Mutex

* Cargo fmt

* Adding f16 to webgpu

* Add Mutex back, since evidently it was causing issues.

Hopefully I can figure out a way to remove it again.

* Removed `num_traits::Num` requirement from Zeros.

Had to figure out a way to store zeros in place

* Fixing no-std stuff

dfdx-core/Cargo.toml

@@ -38,6 +38,9 @@ half = { version = "2.3.1", optional = true, features = ["num-traits", "rand_dis
 gemm = { version = "0.16.14", default-features = false, optional = true, features = ["rayon"] }
 rayon = { version = "1.7.0", optional = true }
 libm = { workspace = true }
+wgpu = { version = "0.18.0", optional = true }
+futures-lite = { version = "2.0.1", optional = true }
+thingbuf = { version = "0.1.4", optional = true }
 
 [dev-dependencies]
 tempfile = "3.3.0"
@@ -59,6 +62,7 @@ fast-alloc = ["std"]
 
 cuda = ["dep:cudarc", "dep:glob"]
 cudnn = ["cuda", "cudarc?/cudnn"]
+webgpu = ["dep:wgpu", "dep:futures-lite", "dep:thingbuf", "wgpu/expose-ids"]
 
 f16 = ["dep:half", "cudarc?/f16", "gemm?/f16"]
 

dfdx-core/src/tensor/error.rs

@@ -16,6 +16,12 @@ pub enum Error {
 
     #[cfg(feature = "cudnn")]
     CudnnError(cudarc::cudnn::CudnnError),
+
+    #[cfg(feature = "webgpu")]
+    WebgpuAdapterNotFound,
+
+    #[cfg(feature = "webgpu")]
+    WebgpuRequestDeviceError(wgpu::RequestDeviceError),
 }
 
 impl std::fmt::Display for Error {

dfdx-core/src/tensor/mod.rs

@@ -145,6 +145,8 @@ mod gradients;
 mod masks;
 #[cfg(feature = "numpy")]
 pub(crate) mod numpy;
+#[cfg(feature = "webgpu")]
+pub(crate) mod webgpu;
 #[cfg(feature = "numpy")]
 pub use numpy::NumpyDtype;
 mod error;
@@ -162,7 +164,7 @@ pub(crate) use storage_traits::{OneFillStorage, ZeroFillStorage};
 pub use tensorlike::Tensorlike;
 
 pub use cpu::Cpu;
-#[cfg(not(feature = "cuda"))]
+#[cfg(not(any(feature = "cuda", feature = "webgpu")))]
 pub type AutoDevice = Cpu;
 
 #[cfg(feature = "cuda")]
@@ -172,6 +174,11 @@ pub use cuda::Cuda;
 #[cfg(feature = "cuda")]
 pub type AutoDevice = Cuda;
 
+#[cfg(feature = "webgpu")]
+pub use webgpu::Webgpu;
+#[cfg(feature = "webgpu")]
+pub type AutoDevice = Webgpu;
+
 pub use storage_traits::{AsArray, CopySlice, TensorFrom, TensorFromVec, TensorToArray};
 pub use storage_traits::{Cache, RandomU64, Storage, Synchronize};
 pub use storage_traits::{OnesTensor, SampleTensor, TriangleTensor, ZerosTensor};

dfdx-core/src/tensor/webgpu/allocate.rs

@@ -0,0 +1,221 @@
+#![allow(clippy::needless_range_loop)]
+
+use crate::{
+    shapes::*,
+    tensor::{masks::triangle_mask, storage_traits::*, unique_id, Cpu, Error, NoneTape, Tensor},
+};
+
+use super::{device::CachableBuffer, Buffer, Webgpu};
+
+use core::marker::PhantomData;
+use rand::Rng;
+use std::{sync::Arc, vec::Vec};
+use wgpu::COPY_BUFFER_ALIGNMENT;
+
+pub(crate) fn round_to_buffer_alignment(size: u64) -> u64 {
+    (size + (COPY_BUFFER_ALIGNMENT - 1)) / COPY_BUFFER_ALIGNMENT * COPY_BUFFER_ALIGNMENT
+}
+
+impl Webgpu {
+    fn tensor_from_host_buf<S: Shape, E: Unit>(
+        &self,
+        shape: S,
+        buf: Vec<E>,
+    ) -> Result<Tensor<S, E, Self>, Error> {
+        let buffer = unsafe { self.alloc_empty::<E>(buf.len()) }?;
+        buffer.copy_to_device::<E>(&self.dev, &self.queue, &buf);
+
+        Ok(self.build_tensor(shape, shape.strides(), buffer))
+    }
+
+    pub(crate) fn build_tensor<S: Shape, E: Unit>(
+        &self,
+        shape: S,
+        strides: S::Concrete,
+        buffer: Buffer,
+    ) -> Tensor<S, E, Self> {
+        let data = CachableBuffer {
+            dev: self.dev.clone(),
+            queue: self.queue.clone(),
+            data: buffer,
+            cache: self.cache.clone(),
+            _phantom: PhantomData,
+        };
+        Tensor {
+            id: unique_id(),
+            data: Arc::new(data),
+            shape,
+            strides,
+            device: self.clone(),
+            tape: Default::default(),
+        }
+    }
+}
+
+impl<E: Unit + SafeZeros> ZerosTensor<E> for Webgpu {
+    fn try_zeros_like<S: HasShape>(&self, src: &S) -> Result<Tensor<S::Shape, E, Self>, Error> {
+        let shape = *src.shape();
+        let strides = shape.strides();
+        let data = unsafe { self.alloc_empty::<E>(shape.num_elements()) }?;
+        data.copy_to_device(&self.dev, &self.queue, &vec![0u8; data.size()]);
+
+        Ok(self.build_tensor(shape, strides, data))
+    }
+}
+
+impl<E: Unit + SafeZeros> ZeroFillStorage<E> for Webgpu {
+    fn try_fill_with_zeros(&self, storage: &mut Self::Vec) -> Result<(), Error> {
+        storage.copy_to_device(&self.dev, &self.queue, &vec![0u8; storage.size()]);
+
+        Ok(())
+    }
+}
+
+impl<E: Unit> OnesTensor<E> for Webgpu {
+    fn try_ones_like<S: HasShape>(&self, src: &S) -> Result<Tensor<S::Shape, E, Self>, Error> {
+        let shape = *src.shape();
+        let buf = vec![E::ONE; shape.num_elements()];
+        self.tensor_from_host_buf(shape, buf)
+    }
+}
+
+impl<E: Unit> TriangleTensor<E> for Webgpu
+where
+    Cpu: TriangleTensor<E>,
+{
+    fn try_upper_tri_like<S: HasShape>(
+        &self,
+        src: &S,
+        val: E,
+        diagonal: impl Into<Option<isize>>,
+    ) -> Result<Tensor<S::Shape, E, Self>, Error> {
+        let shape = *src.shape();
+        let mut data = vec![val; shape.num_elements()];
+        let offset = diagonal.into().unwrap_or(0);
+        triangle_mask(&mut data, &shape, true, offset);
+        self.tensor_from_host_buf(shape, data)
+    }
+
+    fn try_lower_tri_like<S: HasShape>(
+        &self,
+        src: &S,
+        val: E,
+        diagonal: impl Into<Option<isize>>,
+    ) -> Result<Tensor<S::Shape, E, Self>, Error> {
+        let shape = *src.shape();
+        let mut data = vec![val; shape.num_elements()];
+        let offset = diagonal.into().unwrap_or(0);
+        triangle_mask(&mut data, &shape, false, offset);
+        self.tensor_from_host_buf(shape, data)
+    }
+}
+
+impl<E: Unit> OneFillStorage<E> for Webgpu {
+    fn try_fill_with_ones(&self, storage: &mut Self::Vec) -> Result<(), Error> {
+        let len = storage.size() as usize / std::mem::size_of::<E>();
+        let buf = vec![E::ONE; len];
+        storage
+            .data
+            .copy_to_device::<E>(&self.dev, &self.queue, &buf);
+
+        Ok(())
+    }
+}
+
+impl<E: Unit> SampleTensor<E> for Webgpu
+where
+    Cpu: SampleTensor<E>,
+{
+    fn try_sample_like<S: HasShape, D: rand::prelude::Distribution<E>>(
+        &self,
+        src: &S,
+        distr: D,
+    ) -> Result<Tensor<S::Shape, E, Self>, Error> {
+        let shape = *src.shape();
+        let mut buf = Vec::with_capacity(shape.num_elements());
+        {
+            #[cfg(not(feature = "no-std"))]
+            let mut rng = self.cpu.rng.lock().unwrap();
+            #[cfg(feature = "no-std")]
+            let mut rng = self.cpu.rng.lock();
+            buf.resize_with(shape.num_elements(), || rng.sample(&distr));
+        }
+        self.tensor_from_host_buf::<S::Shape, E>(shape, buf)
+    }
+
+    fn try_fill_with_distr<D: rand::prelude::Distribution<E>>(
+        &self,
+        storage: &mut Self::Vec,
+        distr: D,
+    ) -> Result<(), Error> {
+        let len = storage.size() as usize / std::mem::size_of::<E>();
+        let mut buf = Vec::with_capacity(len);
+        {
+            #[cfg(not(feature = "no-std"))]
+            let mut rng = self.cpu.rng.lock().unwrap();
+            #[cfg(feature = "no-std")]
+            let mut rng = self.cpu.rng.lock();
+            buf.resize_with(len, || rng.sample(&distr));
+        }
+        unsafe {
+            std::ptr::copy_nonoverlapping(
+                buf.as_ptr(),
+                storage.data.slice(..).get_mapped_range_mut().as_mut_ptr() as *mut E,
+                len,
+            )
+        };
+        Ok(())
+    }
+}
+
+impl<E: Unit> CopySlice<E> for Webgpu {
+    fn copy_from<S: Shape, T>(dst: &mut Tensor<S, E, Self, T>, src: &[E]) {
+        assert_eq!(
+            dst.data.size() as usize,
+            src.len() * std::mem::size_of::<E>(),
+            "Slices must have same number of elements as *physical* Storage<E> of tensors."
+        );
+        dst.data
+            .data
+            .copy_to_device(&dst.device.dev, &dst.device.queue, src);
+    }
+
+    fn copy_into<S: Shape, T>(src: &Tensor<S, E, Self, T>, dst: &mut [E]) {
+        assert_eq!(
+            src.data.size() as usize,
+            dst.len() * std::mem::size_of::<E>(),
+            "Slices must have same number of elements as *physical* Storage<E> of tensors."
+        );
+        src.data
+            .data
+            .copy_to_host(&src.device.dev, &src.device.queue, dst);
+    }
+}
+
+impl<E: Unit> TensorFromVec<E> for Webgpu {
+    fn try_tensor_from_vec<S: Shape>(
+        &self,
+        src: Vec<E>,
+        shape: S,
+    ) -> Result<Tensor<S, E, Self>, Error> {
+        let num_elements = shape.num_elements();
+
+        if src.len() != num_elements {
+            Err(Error::WrongNumElements)
+        } else {
+            self.tensor_from_host_buf(shape, src)
+        }
+    }
+}
+
+impl<S: Shape, E: Unit> TensorToArray<S, E> for Webgpu
+where
+    Cpu: TensorToArray<S, E> + Storage<E>,
+{
+    type Array = <Cpu as TensorToArray<S, E>>::Array;
+    fn tensor_to_array<T>(&self, tensor: &Tensor<S, E, Self, T>) -> Self::Array {
+        let buf = tensor.as_vec();
+        let cpu_tensor = self.cpu.tensor_from_vec(buf, tensor.shape);
+        self.cpu.tensor_to_array::<NoneTape>(&cpu_tensor)
+    }
+}