update submod

LuisaGroup · Dec 29, 2023 · 3638d4a · 3638d4a
1 parent 4db529c
commit 3638d4a
Show file tree

Hide file tree

Showing 3 changed files with 60 additions and 2 deletions.
diff --git a/luisa_compute/src/resource.rs b/luisa_compute/src/resource.rs
@@ -1358,6 +1358,9 @@ impl<T: IoTexel> Tex2d<T> {
     pub fn format(&self) -> PixelFormat {
         self.handle.format
     }
+    pub fn storage(&self) -> PixelStorage {
+        self.handle.storage
+    }
     // pub fn read(&self, uv: impl AsExpr<Value = Uint2>) -> Expr<T> {
     //     self.var().read(uv)
     // }
@@ -1384,6 +1387,9 @@ impl<T: IoTexel> Tex3d<T> {
     pub fn format(&self) -> PixelFormat {
         self.handle.format
     }
+    pub fn storage(&self) -> PixelStorage {
+        self.handle.storage
+    }
     // pub fn read(&self, uv: impl AsExpr<Value = Uint3>) -> Expr<T> {
     //     self.var().read(uv)
     // }

diff --git a/luisa_compute/tests/autodiff.rs b/luisa_compute/tests/autodiff.rs
@@ -956,7 +956,59 @@ fn autodiff_if_phi() {
         }
     }
 }
-
+#[test]
+fn autodiff_if_phi_outer_no_else() {
+    let device = get_device();
+    let x: Buffer<f32> = device.create_buffer(1024);
+    let y: Buffer<f32> = device.create_buffer(1024);
+    let dx: Buffer<f32> = device.create_buffer(1024);
+    let dy: Buffer<f32> = device.create_buffer(1024);
+    let mut rng = rand::thread_rng();
+    x.view(..).fill_fn(|_| rng.gen());
+    y.view(..).fill_fn(|_| rng.gen());
+    let kernel = Kernel::<fn()>::new(
+        &device,
+        &track!(|| {
+            let buf_x = x.var();
+            let buf_y = y.var();
+            let buf_dx = dx.var();
+            let buf_dy = dy.var();
+            let tid = dispatch_id().x;
+            let x = buf_x.read(tid);
+            let y = buf_y.read(tid);
+            let z = 0.0f32.var();
+            if true.expr() {
+                autodiff(|| {
+                    requires_grad(x);
+                    requires_grad(y);
+                    if x > y {
+                        let tmp = 0.0f32.var();
+                        *tmp = x * 4.0;
+                        *z = tmp;
+                    };
+                    backward(**z);
+                    buf_dx.write(tid, gradient(x));
+                    buf_dy.write(tid, gradient(y));
+                });
+            }
+        }),
+    );
+    kernel.dispatch([1024, 1, 1]);
+    let dx = dx.view(..).copy_to_vec();
+    let dy = dy.view(..).copy_to_vec();
+    let x = x.view(..).copy_to_vec();
+    let y = y.view(..).copy_to_vec();
+    let cache_dir = kernel.cache_dir();
+    for i in 0..1024 {
+        if x[i] > y[i] {
+            assert_eq!(dx[i], 4.0, "{} cache_dir: {:?}", dx[i], cache_dir);
+            assert_eq!(dy[i], 0.0, "{} cache_dir: {:?}", dy[i], cache_dir);
+        } else {
+            assert_eq!(dx[i], 0.0, "{} cache_dir: {:?}", dx[i], cache_dir);
+            assert_eq!(dy[i], 0.0, "{} cache_dir: {:?}", dy[i], cache_dir);
+        }
+    }
+}
 #[test]
 fn autodiff_if_phi2() {
     let device = get_device();

diff --git a/luisa_compute_sys/LuisaCompute b/luisa_compute_sys/LuisaCompute
+4 −1		include/luisa/core/basic_traits.h
+7 −4		include/luisa/core/basic_types.h
+27 −5		include/luisa/runtime/rhi/device_interface.h
+2 −2		include/luisa/vstl/functional.h
+13 −5		include/luisa/vstl/meta_lib.h
+285 −467		include/luisa/vstl/ranges.h
+0 −3		scripts/xmake_func.lua
+5 −1		src/ast/type.cpp
+1 −0		src/backends/CMakeLists.txt
+17 −4		src/backends/common/hlsl/hlsl_codegen.h
+5 −5		src/backends/common/hlsl/hlsl_codegen_util.cpp
+3 −4		src/backends/common/hlsl/struct_generator.cpp
+1 −0		src/backends/common/hlsl/struct_generator.h
+1 −0		src/backends/cuda/CMakeLists.txt
+1 −1		src/backends/cuda/cuda_bindless_array.cpp
+35 −24		src/backends/cuda/cuda_buffer.cpp
+15 −5		src/backends/cuda/cuda_buffer.h
+16 −6		src/backends/cuda/cuda_builtin/cuda_device_math.h
+14 −14		src/backends/cuda/cuda_builtin/cuda_device_resource.h
+6,481 −6,453		src/backends/cuda/cuda_builtin_embedded.cpp
+2 −2		src/backends/cuda/cuda_builtin_embedded.h
+2 −1		src/backends/cuda/cuda_codegen_ast.cpp
+8 −8		src/backends/cuda/cuda_command_encoder.cpp
+1 −0		src/backends/cuda/cuda_compiler.cpp
+4 −4		src/backends/cuda/cuda_curve.cpp
+6 −4		src/backends/cuda/cuda_device.cpp
+2 −0		src/backends/cuda/cuda_device.h
+4 −4		src/backends/cuda/cuda_mesh.cpp
+2 −3		src/backends/cuda/cuda_procedural_primitive.cpp
+3 −3		src/backends/cuda/cuda_shader_optix.cpp
+13 −12		src/backends/cuda/extensions/cuda_denoiser.cpp
+87 −0		src/backends/cuda/extensions/cuda_pinned_memory.cpp
+31 −0		src/backends/cuda/extensions/cuda_pinned_memory.h
+1 −1		src/backends/cuda/extensions/cuda_texture_compression.cpp
+10 −2		src/backends/cuda/generate_device_library.py
+3 −1		src/backends/cuda/xmake.lua
+4 −2		src/backends/dx/DXApi/LCDevice.cpp
+6 −4		src/backends/dx/DXRuntime/Device.cpp
+1 −0		src/backends/dx/DXRuntime/Device.h
+2 −2		src/backends/dx/Resource/GpuAllocator.cpp
+2 −2		src/backends/dx/Resource/GpuAllocator.h
+3 −0		src/backends/dx/Shader/BuiltinKernel.cpp
+30 −1		src/backends/dx/Shader/ComputeShader.cpp
+6 −3		src/backends/dx/Shader/ComputeShader.h
+12 −9		src/backends/dx/Shader/RasterShader.cpp
+38 −2		src/backends/dx/Shader/ShaderSerializer.cpp
+1 −0		src/backends/dx/Shader/ShaderSerializer.h
+6 −3		src/backends/validation/CMakeLists.txt
+1 −1		src/ext/magic_enum
+1 −1		src/ext/spdlog
+6 −0		src/ir/ir2ast.cpp
+16 −6		src/rust/luisa_compute_backend_impl/src/cpu/codegen/device_math.h
+1 −1		src/tests/test_pinned_mem.cpp
+6 −1		src/vstl/vstl.cpp