Revert "Fix offset calculation in kernel launch"

This reverts commit d7d38256e0580c3cfb649a641d9ed62c6ff0fc20.

2 files changed, 16 insertions, 19 deletions

diff --git a/zluda/src/impl/function.rs b/zluda/src/impl/function.rs
index 638d08a..4f2006c 100644
--- a/zluda/src/impl/function.rs
+++ b/zluda/src/impl/function.rs
@@ -110,24 +110,23 @@ pub fn launch_kernel(
             match (buffer_size, buffer_ptr) {
                 (Some(buffer_size), Some(buffer_ptr)) => {
                     let sum_of_kernel_argument_sizes =
-                        func.arg_size
-                            .iter()
-                            .fold(0, |sum_of_arg_sizes, size_of_arg| {
-                                sum_of_arg_sizes + align_to_usize(*size_of_arg)
-                            });
+                        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 {
                         return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
                     }
                     let mut offset = 0;
                     for (i, arg_size) in func.arg_size.iter().enumerate() {
+                        let buffer_offset = round_up_to_multiple(offset, *arg_size);
                         unsafe {
                             func.base.set_arg_raw(
                                 i as u32,
                                 *arg_size,
-                                buffer_ptr.add(offset) as *const _,
+                                buffer_ptr.add(buffer_offset) as *const _,
                             )?
                         };
-                        offset += align_to_usize(*arg_size);
+                        offset = buffer_offset + *arg_size;
                     }
                 }
                 _ => return Err(CUresult::CUDA_ERROR_INVALID_VALUE),
@@ -160,9 +159,8 @@ pub fn launch_kernel(
     })?
 }
 
-fn align_to_usize(value: usize) -> usize {
-    let multiple = std::mem::size_of::<usize>();
-    ((value + multiple - 1) / multiple) * multiple
+fn round_up_to_multiple(x: usize, multiple: usize) -> usize {
+    ((x + multiple - 1) / multiple) * multiple
 }
 
 pub(crate) fn get_attribute(
diff --git a/zluda_dump/src/lib.rs b/zluda_dump/src/lib.rs
index 79aca8d..cff6733 100644
--- a/zluda_dump/src/lib.rs
+++ b/zluda_dump/src/lib.rs
@@ -573,24 +573,24 @@ fn dump_arguments(
         }
         match (buffer_size, buffer_ptr) {
             (Some(buffer_size), Some(buffer_ptr)) => {
-                let sum_of_kernel_argument_sizes =
-                    args.iter().fold(0, |sum_of_arg_sizes, size_of_arg| {
-                        sum_of_arg_sizes + align_to_usize(*size_of_arg)
-                    });
+                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 {
                     return Err("Malformed `extra` parameter to kernel launch")?;
                 }
                 let mut offset = 0;
                 for (i, arg_size) in args.iter().enumerate() {
+                    let buffer_offset = round_up_to_multiple(offset, *arg_size);
                     unsafe {
                         dump_argument_to_file(
                             &dump_dir,
                             i,
                             *arg_size,
-                            buffer_ptr.add(offset) as *const _,
+                            buffer_ptr.add(buffer_offset) as *const _,
                         )?
                     };
-                    offset += align_to_usize(*arg_size);
+                    offset = buffer_offset + *arg_size;
                 }
             }
             _ => return Err("Malformed `extra` parameter to kernel launch")?,
@@ -599,9 +599,8 @@ fn dump_arguments(
     Ok(())
 }
 
-fn align_to_usize(value: usize) -> usize {
-    let multiple = std::mem::size_of::<usize>();
-    ((value + multiple - 1) / multiple) * multiple
+fn round_up_to_multiple(x: usize, multiple: usize) -> usize {
+    ((x + multiple - 1) / multiple) * multiple
 }
 
 unsafe fn dump_argument_to_file(