GPU and Compiler backends have a different way to refresh memory.

Author: isazi

kernel_tuner/backends/backend.py

@@ -57,11 +57,10 @@ def memcpy_htod(self, dest, src):
         """This method must implement a host to device copy."""
         pass
 
-    def refresh_memory(self, arguments, should_sync):
-        """Copy the original content of the output memory to device memory."""
-        for i, arg in enumerate(arguments):
-            if should_sync[i]:
-                self.memcpy_htod(self.allocations[i], arg)
+    @abstractmethod
+    def refresh_memory(self, device_memory, host_arguments, should_sync):
+        """This method must implement refreshing the device memory with a clean copy."""
+        pass
 
 
 class GPUBackend(Backend):
@@ -86,6 +85,12 @@ def copy_texture_memory_args(self, texmem_args):
         """This method must implement the allocation and copy of texture memory to the GPU."""
         pass
 
+    def refresh_memory(self, gpu_memory, host_arguments, should_sync):
+        """Refresh the GPU memory with the untouched host arguments."""
+        for i, arg in enumerate(host_arguments):
+            if should_sync[i]:
+                self.memcpy_htod(gpu_memory[i], arg)
+
 
 class CompilerBackend(Backend):
     """Base class for compiler backends"""
@@ -94,6 +99,12 @@ class CompilerBackend(Backend):
     def __init__(self, iterations, compiler_options, compiler):
         pass
 
+    def refresh_memory(self, gpu_memory, host_arguments, should_sync):
+        """Refresh the GPU memory with the untouched host arguments."""
+        for i, arg in enumerate(host_arguments):
+            if should_sync[i]:
+                self.memcpy_htod(self.allocations[i], arg)
+
     @abstractmethod
     def cleanup_lib(self):
         """Unload the previously loaded shared library"""

kernel_tuner/core.py

@@ -489,7 +489,7 @@ def check_kernel_output(self, func, gpu_args, instance, answer, atol, verify, ve
 
         # re-copy original contents of output arguments to GPU memory, to overwrite any changes
         # by earlier kernel runs
-        self.dev.refresh_memory(instance.arguments, should_sync)
+        self.dev.refresh_memory(gpu_args, instance.arguments, should_sync)
 
         # run the kernel
         check = self.run_kernel(func, gpu_args, instance)