lib: various fixes to cmdbuf ndrange kernel launches

pocl · Mar 11, 2024 · d09cb7c · d09cb7c
1 parent f81985d
commit d09cb7c
Show file tree

Hide file tree

Showing 5 changed files with 4 additions and 6 deletions.
diff --git a/lib/CL/clCommandFillImageKHR.c b/lib/CL/clCommandFillImageKHR.c
@@ -24,7 +24,6 @@
 #include <CL/cl_ext.h>
 
 #include "pocl_cl.h"
-#include "pocl_image_util.h"
 #include "pocl_mem_management.h"
 #include "pocl_shared.h"
 #include "pocl_util.h"

diff --git a/lib/CL/clEnqueueCommandBufferKHR.c b/lib/CL/clEnqueueCommandBufferKHR.c
@@ -25,7 +25,6 @@
 
 #include "pocl_cl.h"
 #include "pocl_mem_management.h"
-#include "pocl_shared.h"
 #include "pocl_util.h"
 
 static void CL_CALLBACK

diff --git a/lib/CL/clRemapCommandBufferKHR.c b/lib/CL/clRemapCommandBufferKHR.c
@@ -147,8 +147,8 @@ POname (clRemapCommandBufferKHR) (cl_command_buffer_khr command_buffer,
           ulong *groups = cmd->command.run.pc.num_groups;
           size_t global_size[3]
             = { local_size[0] * groups[0],
-                work_dim > 1 ? (local_size[1] * groups[1]) : 0,
-                work_dim > 2 ? (local_size[2] * groups[2]) : 0 };
+                work_dim > 1 ? (local_size[1] * groups[1]) : 1,
+                work_dim > 2 ? (local_size[2] * groups[2]) : 1 };
 
           /* Re-record cmd using the original command's kernel arguments.
            *

diff --git a/lib/CL/pocl_ndrange_kernel.c b/lib/CL/pocl_ndrange_kernel.c
@@ -498,8 +498,7 @@ pocl_ndrange_kernel_common (
   if (errcode != CL_SUCCESS)
     goto ERROR;
 
-  errcode
-    = pocl_kernel_copy_args (kernel, src_arguments, &(*cmd)->command.run);
+  errcode = pocl_kernel_copy_args (kernel, src_arguments, &c->command.run);
   if (errcode != CL_SUCCESS)
     goto ERROR;
 

diff --git a/lib/CL/pocl_util.c b/lib/CL/pocl_util.c
@@ -2344,6 +2344,7 @@ pocl_copy_event_node (_cl_command_node *dst_node, _cl_command_node *src_node)
                                            &dst_node->command.run);
       if (errcode != CL_SUCCESS)
         return CL_OUT_OF_HOST_MEMORY;
+      dst_node->program_device_i = src_node->program_device_i;
       break;
 
     case CL_COMMAND_FILL_BUFFER: