Make sure it is possible to log 32bit PhysX (#374)

dark_api/src/lib.rs

@@ -227,7 +227,7 @@ dark_api! {
         [1] = get_module_from_cubin(
             module: *mut cuda_types::cuda::CUmodule,
             fatbinc_wrapper: *const cuda_types::dark_api::FatbincWrapper
-        ) -> (),
+        ) -> cuda_types::cuda::CUresult,
         [2] = cudart_interface_fn2(
             pctx: *mut cuda_types::cuda::CUcontext,
             dev: cuda_types::cuda::CUdevice

format/src/lib.rs

@@ -428,8 +428,9 @@ impl CudaDisplay for CUmemcpy3DOperand_st {
                 CudaDisplay::write(unsafe { &self.op.ptr }, fn_name, index, writer)?;
             }
             _ => {
+                const CU_MEMCPY_3D_OP_SIZE: usize = mem::size_of::<CUmemcpy3DOperand_st__bindgen_ty_1>();
                 CudaDisplay::write(
-                    &unsafe { mem::transmute::<_, [u8; 32]>(self.op) },
+                    &unsafe { mem::transmute::<_, [u8; CU_MEMCPY_3D_OP_SIZE]>(self.op) },
                     fn_name,
                     index,
                     writer,

ptx/src/test/spirv_run/mod.rs

@@ -500,7 +500,7 @@ macro_rules! dynamic_fns {
         $(
             #[allow(dead_code)]
             unsafe fn $fn_name(&self, $($arg_id : $arg_type),*) -> Result<$ret_type, libloading::Error> {
-                let func = unsafe { self.lib.get::<unsafe extern "system" fn ($($arg_type),*) -> $ret_type>(concat!(stringify!($fn_name), "\0").as_bytes()) };
+                let func = unsafe { self.lib.get::<unsafe extern $abi fn ($($arg_type),*) -> $ret_type>(concat!(stringify!($fn_name), "\0").as_bytes()) };
                 func.map(|f| f($($arg_id),*) )
             }
         )*

zluda/src/impl/driver.rs

@@ -106,7 +106,7 @@ impl ::dark_api::cuda::CudaDarkApi for DarkApi {
     unsafe extern "system" fn get_module_from_cubin(
         module: *mut cuda_types::cuda::CUmodule,
         fatbinc_wrapper: *const cuda_types::dark_api::FatbincWrapper,
-    ) -> () {
+    ) -> cuda_types::cuda::CUresult {
         todo!()
     }
 

zluda_dump/src/lib.rs

@@ -286,7 +286,7 @@ impl DarkApiDump {
         fatbinc_wrapper: *const cuda_types::dark_api::FatbincWrapper,
         state: &mut trace::StateTracker,
         fn_logger: &mut FnCallLog,
-        _result: (),
+        _result: CUresult,
     ) {
         fn_logger.try_(|fn_logger| unsafe {
             trace::record_submodules_from_wrapped_fatbin(*module, fatbinc_wrapper, fn_logger, state)
@@ -391,7 +391,7 @@ impl ::dark_api::cuda::CudaDarkApi for DarkApiDump {
             [1] = get_module_from_cubin(
                 module: *mut cuda_types::cuda::CUmodule,
                 fatbinc_wrapper: *const cuda_types::dark_api::FatbincWrapper // FatbincWrapper
-            ) -> (),
+            ) -> cuda_types::cuda::CUresult,
             [6] = get_module_from_cubin_ext1(
                 result: *mut cuda_types::cuda::CUmodule,
                 fatbinc_wrapper: *const cuda_types::dark_api::FatbincWrapper, // FatbincWrapper

zluda_dump/src/log.rs

@@ -380,7 +380,7 @@ impl Display for ErrorEntry {
                                                         observed,
                                                     } => write!(
                                                         f,
-                                                        "Unexpected field {}. Expected one of: {{{}}}, observed: {}",
+                                                        "Unexpected field {}. Expected one of: [{}], observed: {}",
                                                         field_name,
                                                         expected
                                                             .iter()

zluda_dump_common/src/lib.rs

@@ -1,6 +1,6 @@
 use cuda_types::cuda::{CUerror, CUresult, CUresultConsts, CUuuid};
 use dark_api::ByteVecFfi;
-use std::{num::NonZero, ptr, sync::LazyLock};
+use std::{ffi::c_void, num::NonZero, ptr, sync::LazyLock};
 
 pub fn get_export_table() -> Option<::dark_api::zluda_dump::ZludaDumpInternal> {
     static CU_GET_EXPORT_TABLE: LazyLock<
@@ -117,7 +117,7 @@ impl<T> ReprUsize for *const T {
 
     extern "C" fn format_status(x: usize) -> ByteVecFfi {
         let mut writer = Vec::new();
-        format::CudaDisplay::write(&x, "", 0, &mut writer).ok();
+        format::CudaDisplay::write(&Self::from_usize(x).cast::<c_void>(), "", 0, &mut writer).ok();
         ByteVecFfi::new(writer)
     }
 }
@@ -169,15 +169,9 @@ impl ReprUsize for () {
         ()
     }
 
-    extern "C" fn format_status(x: usize) -> ByteVecFfi {
+    extern "C" fn format_status(_: usize) -> ByteVecFfi {
         let mut writer = Vec::new();
-        format::CudaDisplay::write(
+        format::CudaDisplay::write(&(), "", 0, &mut writer).ok();
-            &cuda_types::cublas::cublasStatus_t::from_usize(x),
-            "",
-            0,
-            &mut writer,
-        )
-        .ok();
         ByteVecFfi::new(writer)
     }
 }
@@ -193,13 +187,7 @@ impl ReprUsize for u32 {
 
     extern "C" fn format_status(x: usize) -> ByteVecFfi {
         let mut writer = Vec::new();
-        format::CudaDisplay::write(
+        format::CudaDisplay::write(&Self::from_usize(x), "", 0, &mut writer).ok();
-            &cuda_types::cublas::cublasStatus_t::from_usize(x),
-            "",
-            0,
-            &mut writer,
-        )
-        .ok();
         ByteVecFfi::new(writer)
     }
 }
@@ -215,13 +203,7 @@ impl ReprUsize for i32 {
 
     extern "C" fn format_status(x: usize) -> ByteVecFfi {
         let mut writer = Vec::new();
-        format::CudaDisplay::write(
+        format::CudaDisplay::write(&Self::from_usize(x), "", 0, &mut writer).ok();
-            &cuda_types::cublas::cublasStatus_t::from_usize(x),
-            "",
-            0,
-            &mut writer,
-        )
-        .ok();
         ByteVecFfi::new(writer)
     }
 }
@@ -237,13 +219,7 @@ impl ReprUsize for u64 {
 
     extern "C" fn format_status(x: usize) -> ByteVecFfi {
         let mut writer = Vec::new();
-        format::CudaDisplay::write(
+        format::CudaDisplay::write(&Self::from_usize(x), "", 0, &mut writer).ok();
-            &cuda_types::cublas::cublasStatus_t::from_usize(x),
-            "",
-            0,
-            &mut writer,
-        )
-        .ok();
         ByteVecFfi::new(writer)
     }
 }
@@ -259,13 +235,7 @@ impl ReprUsize for *mut std::ffi::c_void {
 
     extern "C" fn format_status(x: usize) -> ByteVecFfi {
         let mut writer = Vec::new();
-        format::CudaDisplay::write(
+        format::CudaDisplay::write(&Self::from_usize(x), "", 0, &mut writer).ok();
-            &cuda_types::cublas::cublasStatus_t::from_usize(x),
-            "",
-            0,
-            &mut writer,
-        )
-        .ok();
         ByteVecFfi::new(writer)
     }
 }

zluda_inject/build.rs

@@ -43,7 +43,6 @@ fn main() -> Result<(), VarError> {
             rustc_cmd.arg(r"-lucrt");
         }
         rustc_cmd
-            .arg("-ldylib=nvcuda")
             .arg("-C")
             .arg(format!("opt-level={}", opt_level))
             .arg("-L")

zluda_inject/tests/helpers/direct_cuinit.rs

@@ -1,5 +1,6 @@
 #![crate_type = "bin"]
 
+#[link(name = "nvcuda", kind = "raw-dylib")]
 extern "system" {
     fn cuInit(flags: u32) -> u32;
 }

zluda_inject/tests/helpers/do_cuinit.rs

@@ -1,5 +1,6 @@
 #![crate_type = "cdylib"]
 
+#[link(name = "nvcuda", kind = "raw-dylib")]
 extern "system" {
     fn cuInit(flags: u32) -> u32;
 }

zluda_redirect/src/lib.rs

@@ -720,12 +720,7 @@ fn get_payload(guid: &detours_sys::GUID) -> Option<&'static [u8]> {
     let mut size = 0;
     let payload_ptr = unsafe { detours_sys::DetourFindPayloadEx(guid, &mut size) };
     if payload_ptr != ptr::null_mut() {
-        Some(unsafe {
+        Some(unsafe { slice::from_raw_parts(payload_ptr as *const _, size as usize) })
-            slice::from_raw_parts(
-                payload_ptr as *const _,
-                (size as usize) / mem::size_of::<u16>(),
-            )
-        })
     } else {
         None
     }