fixed some bugs

Author: zibai.wang

include/acl_types.h

@@ -922,7 +922,7 @@ typedef struct _cl_kernel {
   // Eventually this should be an array of ACLDeviceBinaries similar
   // to how cl_program contains an array of dev_prog.
   const acl_device_binary_t *dev_bin;
-  int processed_buffer_size;
+  size_t processed_buffer_size;
 } _cl_kernel;
 
 ACL_DECLARE_CL_OBJECT_ALLOC_FUNCTIONS(cl_kernel);
@@ -1347,7 +1347,7 @@ typedef struct {
   // it sees the CL_COMPLETE status for the first time, but it won't
   // change after that.
   cl_uint num_printf_bytes_pending;
-  int debug_dump; // TODO, initialize to 0, somewhere
+  int debug_dump = 0; // TODO, initialize to 0, somewhere
 } acl_device_op_info_t;
 
 // An operation to be performed on a device.

src/acl_printf.cpp

@@ -38,7 +38,7 @@
 #pragma GCC visibility push(protected)
 #endif
 
-static void l_dump_printf_buffer(cl_event event, cl_kernel kernel,
+static size_t l_dump_printf_buffer(cl_event event, cl_kernel kernel,
                                  unsigned size);
 static void decode_string(std::string &print_data);
 
@@ -719,12 +719,13 @@ static std::string::const_iterator get_data_elem_at_offset(
   return end_of_string;
 }
 
-static void l_dump_printf_buffer(cl_event event, cl_kernel kernel,
+static size_t l_dump_printf_buffer(cl_event event, cl_kernel kernel,
                                  unsigned size) {
   unsigned global_offset;        // the location in the printf buffer
   unsigned single_printf_offset; // the offset of a single printf
   void (*hal_dma_fn)(cl_event, const void *, void *, size_t) = 0;
   int src_on_host = 1;
+  size_t dumped_buffer_size = 0;
 #ifdef _WIN32
   __declspec(align(64)) char
       buffer[ACL_PRINTF_BUFFER_TOTAL_SIZE]; // Aligned to 64, for dma transfers
@@ -743,11 +744,11 @@ static void l_dump_printf_buffer(cl_event event, cl_kernel kernel,
   if (!verify_types()) {
     printf("Host data types are incompatible with ACL compiler, ignoring "
            "printfs...\n");
-    return;
+    return dumped_buffer_size;
   }
 
   if (printf_infos.empty())
-    return;
+    return dumped_buffer_size;
 
   // Memory is on the device if all of these are true:
   //   The memory is not SVM or the device does not support SVM.
@@ -783,9 +784,10 @@ static void l_dump_printf_buffer(cl_event event, cl_kernel kernel,
   if (size > (unsigned int) kernel->processed_buffer_size){
     printf("ZIBAI has something to print! \n");
     hal_dma_fn(NULL, unprocessed_begin, buffer, size - kernel->processed_buffer_size);
+    dumped_buffer_size = size - kernel->processed_buffer_size;
   } else{
     printf("Zibai Added, nothing to print \n");
-    return;
+    return dumped_buffer_size;
   }
 
 
@@ -839,7 +841,7 @@ static void l_dump_printf_buffer(cl_event event, cl_kernel kernel,
     if (!success) {
       acl_print_debug_msg(
           "corrupt printf data, ignoring remaining printfs...\n");
-      return;
+      return dumped_buffer_size;
     }
 
 #ifdef DEBUG
@@ -902,7 +904,7 @@ static void l_dump_printf_buffer(cl_event event, cl_kernel kernel,
       if (vector_size == -1) {
         acl_print_debug_msg("wrong vector specifier in printf call, ignoring "
                             "remaining printfs...\n");
-        return;
+        return dumped_buffer_size;
       }
 
       // get the length specifier
@@ -923,7 +925,7 @@ static void l_dump_printf_buffer(cl_event event, cl_kernel kernel,
       if (size_of_data == 0) {
         acl_print_debug_msg("wrong length modifier in printf call, ignoring "
                             "remaining printfs...\n");
-        return;
+        return dumped_buffer_size;
       }
 
       for (i = 0; i < vector_size; i++) {
@@ -979,6 +981,7 @@ static void l_dump_printf_buffer(cl_event event, cl_kernel kernel,
 #ifdef DEBUG
   printf("exiting acl_dump_buffer...\n");
 #endif
+  return dumped_buffer_size;
 }
 
 //
@@ -1030,14 +1033,14 @@ void acl_process_printf_buffer(void *user_data, acl_device_op_t *op) {
 
     // Grab the printf data and emit it.
     cl_uint num_bytes = op->info.num_printf_bytes_pending;
-    l_dump_printf_buffer(event, kernel, num_bytes);
+    size_t dumped_buffer_size = l_dump_printf_buffer(event, kernel, num_bytes);
 
     // Mark this printf work as done.  Must do this *before* unstalling
     // the kernel, to avoid a race against the kernel filling up the
     // buffer again.
     // Zibai Testing
     if (op->info.debug_dump == 1){
-      kernel->processed_buffer_size += num_bytes; // TODO, Fix this, should be += processed bytes
+      kernel->processed_buffer_size += dumped_buffer_size;
     } else{
       // Full dump, clear this global var
       kernel->processed_buffer_size = 0;
@@ -1049,9 +1052,11 @@ void acl_process_printf_buffer(void *user_data, acl_device_op_t *op) {
     acl_memory_barrier();
 
     // Allow the kernel to continue running.
-    // Zibai Testing!
-    // acl_get_hal()->unstall_kernel(
-    //     event->cmd.info.ndrange_kernel.device->def.physical_device_id, op->id);
+    // We don't need to unstall the kernel during the early flushing during debug.
+    if (op->info.debug_dump == 0){
+      acl_get_hal()->unstall_kernel(
+          event->cmd.info.ndrange_kernel.device->def.physical_device_id, op->id);
+    }
   }
 }