Bunch of tiny fixes and improvements

3 files changed, 24 insertions, 5 deletions

diff --git a/zluda_dump/src/cuda.rs b/zluda_dump/src/cuda.rs
index d715689..50082d1 100644
--- a/zluda_dump/src/cuda.rs
+++ b/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;
+}
diff --git a/zluda_dump/src/lib.rs b/zluda_dump/src/lib.rs
index cff6733..92238c6 100644
--- a/zluda_dump/src/lib.rs
+++ b/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;
diff --git a/zluda_dump/src/replay.py b/zluda_dump/src/replay.py
index 52f250c..723d954 100644
--- a/zluda_dump/src/replay.py
+++ b/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")

@@ -94,7 +95,8 @@ 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)]):

-            verify_single_dump(input_subdir, max_threads)

+            if os.path.isdir(input_subdir):

+                verify_single_dump(input_subdir, max_threads)

 

 

 if __name__ == "__main__":