ConvTranspose1d cuda support. (#1697)

* ConvTranspose1d cuda support.

* Add the conv-transpose1d kernel.

* Remove some unused variables.

candle-core/src/cuda_backend.rs

@@ -1149,6 +1149,55 @@ impl<'a> Map2 for Conv2D<'a> {
     }
 }
 
+struct ConvTranspose1D<'a>(&'a crate::conv::ParamsConvTranspose1D);
+impl<'a> Map2 for ConvTranspose1D<'a> {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        inp: &CudaSlice<T>,
+        inp_l: &Layout,
+        k: &CudaSlice<T>,
+        k_l: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<CudaSlice<T>> {
+        // Kernel shape: (c_in_k, c_out, l_k)
+        // Input shape: (b_size, c_in, l_in)
+        let p = &self.0;
+        let l_out = p.l_out();
+        let dst_el = p.c_out * l_out * p.b_size;
+        let inp = &inp.slice(inp_l.start_offset()..);
+        let k = &k.slice(k_l.start_offset()..);
+        let shape = inp_l.shape();
+        let dims = shape.dims();
+        let el = shape.elem_count();
+
+        // SAFETY: Set later by running the kernel.
+        let out = unsafe { dev.alloc::<T>(dst_el) }.w()?;
+        let cfg = LaunchConfig::for_num_elems(dst_el as u32);
+        let func = dev.get_or_load_func(&kernel_name::<T>("conv_transpose1d"), kernels::CONV)?;
+        let ds = if dims.len() == 3 {
+            [dims, inp_l.stride(), k_l.dims(), k_l.stride()].concat()
+        } else {
+            crate::bail!("unexpected input shape for conv_transpose1d {dims:?}")
+        };
+        let ds = dev.htod_copy(ds).w()?;
+        let params = (
+            el,
+            l_out,
+            p.stride,
+            p.padding,
+            p.output_padding,
+            p.dilation,
+            &ds,
+            inp,
+            k,
+            &out,
+        );
+        // SAFETY: ffi.
+        unsafe { func.launch(cfg, params) }.w()?;
+        Ok(out)
+    }
+}
+
 struct ConvTranspose2D<'a>(&'a crate::conv::ParamsConvTranspose2D);
 impl<'a> Map2 for ConvTranspose2D<'a> {
     fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
@@ -1810,12 +1859,15 @@ impl BackendStorage for CudaStorage {
 
     fn conv_transpose1d(
         &self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: &crate::conv::ParamsConvTranspose1D,
+        l: &Layout,
+        kernel: &Self,
+        kernel_l: &Layout,
+        params: &crate::conv::ParamsConvTranspose1D,
     ) -> Result<Self> {
-        todo!()
+        let device = self.device().clone();
+        let slice =
+            ConvTranspose1D(params).map(&self.slice, l, &kernel.slice, kernel_l, &device)?;
+        Ok(Self { slice, device })
     }
 
     #[cfg(not(feature = "cudnn"))]

candle-core/tests/conv_tests.rs

@@ -50,17 +50,15 @@ fn conv1d(dev: &Device) -> Result<()> {
         test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
         [2.4509, 2.6357, -1.3336, 4.1393, 0.5657, 1.8091, -1.1784, 3.5675, 0.5069, 3.3352]
     );
-    if dev.is_cpu() {
-        let res = t.conv_transpose1d(&w.transpose(0, 1)?, 0, 0, 1, 1)?;
-        assert_eq!(res.dims(), [1, 2, 7]);
-        assert_eq!(
-            test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
-            [
-                0.0699, -1.2899, 8.3018, 5.5873, 2.4572, -2.6143, -0.0706, 1.8765, 4.8318, 1.1538,
-                4.7076, -5.9745, -0.8276, 1.621
-            ],
-        );
-    }
+    let res = t.conv_transpose1d(&w.transpose(0, 1)?, 0, 0, 1, 1)?;
+    assert_eq!(res.dims(), [1, 2, 7]);
+    assert_eq!(
+        test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
+        [
+            0.0699, -1.2899, 8.3018, 5.5873, 2.4572, -2.6143, -0.0706, 1.8765, 4.8318, 1.1538,
+            4.7076, -5.9745, -0.8276, 1.621
+        ],
+    );
     Ok(())
 }