Bunch of tiny fixes and improvements

ptx/src/translate.rs

@@ -1020,12 +1020,8 @@ fn emit_function_header<'a>(
     kernel_info: &mut HashMap<String, KernelInfo>,
 ) -> Result<spirv::Word, TranslateError> {
     if let ast::MethodName::Kernel(name) = func_decl.name {
-        let input_args = if func_decl.shared_mem.is_none() {
+        let args_lens = func_decl
-            func_decl.input_arguments.as_slice()
+            .input_arguments
-        } else {
-            &func_decl.input_arguments[0..func_decl.input_arguments.len() - 1]
-        };
-        let args_lens = input_args
             .iter()
             .map(|param| param.v_type.size_of())
             .collect();

zluda/src/cuda.rs

@@ -3449,7 +3449,7 @@ pub extern "system" fn cuStreamQuery(hStream: CUstream) -> CUresult {
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "system" fn cuStreamSynchronize(hStream: CUstream) -> CUresult {
-    r#impl::unimplemented()
+    CUresult::CUDA_SUCCESS
 }
 
 #[cfg_attr(not(test), no_mangle)]

zluda/src/impl/function.rs

@@ -113,7 +113,7 @@ pub fn launch_kernel(
                         func.arg_size.iter().fold(0, |offset, size_of_arg| {
                             size_of_arg + round_up_to_multiple(offset, *size_of_arg)
                         });
-                    if buffer_size != sum_of_kernel_argument_sizes {
+                    if buffer_size < sum_of_kernel_argument_sizes {
                         return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
                     }
                     let mut offset = 0;

zluda_dump/src/cuda.rs

@@ -4399,3 +4399,20 @@ extern_redirect_with! {
 extern_redirect! {
     pub fn cuFuncGetModule(hmod: *mut CUmodule, hfunc: CUfunction) -> CUresult;
 }
+#[repr(transparent)]
+#[derive(Copy, Clone, Hash, PartialEq, Eq)]
+pub struct CUoutput_mode_enum(pub ::std::os::raw::c_uint);
+pub use self::CUoutput_mode_enum as CUoutput_mode;
+extern_redirect! {
+    pub fn cuProfilerInitialize(
+        configFile: *const ::std::os::raw::c_char,
+        outputFile: *const ::std::os::raw::c_char,
+        outputMode: CUoutput_mode,
+    ) -> CUresult;
+}
+extern_redirect! {
+    pub fn cuProfilerStart() -> CUresult;
+}
+extern_redirect! {
+    pub fn cuProfilerStop() -> CUresult;
+}

zluda_dump/src/lib.rs

@@ -576,7 +576,7 @@ fn dump_arguments(
                 let sum_of_kernel_argument_sizes = args.iter().fold(0, |offset, size_of_arg| {
                     size_of_arg + round_up_to_multiple(offset, *size_of_arg)
                 });
-                if buffer_size != sum_of_kernel_argument_sizes {
+                if buffer_size < sum_of_kernel_argument_sizes {
                     return Err("Malformed `extra` parameter to kernel launch")?;
                 }
                 let mut offset = 0;

zluda_dump/src/replay.py

@@ -51,9 +51,10 @@ def parse_arguments(dump_path, prefix):
     return [load_arguments(path.join(dir, f)) for f in sorted(arg_files)]
 
 
-def append_debug_buffer(args):
+def append_debug_buffer(args, grid, block):
     args = list(args)
-    debug_buff = np.zeros(1024 * 1024, np.single)
+    items = block[0] * block[1] * block[2] * block[0] * block[1] * block[2]
+    debug_buff = np.zeros(items, dtype=np.uint32)
     args.append((drv.InOut(debug_buff), debug_buff))
     return args
 
@@ -71,7 +72,7 @@ def verify_single_dump(input_path, max_block_threads):
         return
     module = drv.module_from_file(path.join(input_path, "module.ptx"))
     kernel = module.get_function(kernel_name)
-    pre_args = append_debug_buffer(parse_arguments(input_path, "pre"))
+    pre_args = append_debug_buffer(parse_arguments(input_path, "pre"), tuple(launch_lines[:3]), block)
     kernel_pre_args, host_pre_args = zip(*pre_args)
     kernel(*list(kernel_pre_args), grid=tuple(launch_lines[:3]), block=block, shared=launch_lines[6])
     post_args = parse_arguments(input_path, "post")
@@ -94,6 +95,7 @@ def main(argv):
         verify_single_dump(input_path, max_threads)
     else:
         for input_subdir in sorted([path.join(input_path, dir_name) for dir_name in os.listdir(input_path)]):
+            if os.path.isdir(input_subdir):
                 verify_single_dump(input_subdir, max_threads)