Remove call to device._poll in _destroy methods (#447)

* Only download release version, we don't use the debug build anyway

* Remove call to device._poll in _destroy methods

* Clean uo

Author: almarklein

download-wgpu-native.py

@@ -102,7 +102,7 @@ def get_arch():
 
 
 def main(version, os_string, arch, upstream):
-    for build in ("release", "debug"):
+    for build in ["release"]:  # ["release", "debug"]
         filename = f"wgpu-{os_string}-{arch}-{build}.zip"
         url = f"https://github.com/{upstream}/releases/download/v{version}/{filename}"
         tmp = tempfile.gettempdir()

tests_mem/test_objects.py

@@ -77,17 +77,13 @@ def test_release_bind_group(n):
 @create_and_release
 def test_release_bind_group_layout(n):
     # Note: when we use the same binding layout descriptor, wgpu-native
-    # re-uses the BindGroupLayout object. On the other hand, it also
-    # does not seem to clean them up. Perhaps it just caches them? There
-    # are only so many possible combinations, and its just 152 bytes
-    # (on Metal) per object.
+    # re-uses the BindGroupLayout object.
 
     global _bind_group_layout_binding
     _bind_group_layout_binding += 1
 
     yield {
         "expected_counts_after_create": {"BindGroupLayout": (n, 1)},
-        "expected_counts_after_release": {"BindGroupLayout": (0, 1)},
     }
 
     binding_layouts = [

tests_mem/testutils.py

@@ -74,11 +74,16 @@ def get_memory_usage():
 def clear_mem():
     time.sleep(0.001)
     gc.collect()
+
     time.sleep(0.001)
     gc.collect()
+
     if is_pypy:
         gc.collect()
 
+    device = wgpu.utils.get_default_device()
+    device._poll()
+
 
 def get_counts():
     """Get a dict that maps object names to a 2-tuple represening
@@ -184,6 +189,9 @@ def core_test_func():
             # Test that class matches function name (should prevent a group of copy-paste errors)
             assert ob_name == cls.__name__[3:]
 
+            # Give wgpu some slack to clean up temporary resources
+            wgpu.utils.get_default_device()._poll()
+
             # Measure peak object counts
             counts2 = get_counts()
             more2 = get_excess_counts(counts1, counts2)

wgpu/backends/wgpu_native/_api.py

@@ -1821,14 +1821,6 @@ def _destroy(self):
             self._internal, internal = None, self._internal
             # H: void f(WGPUBuffer buffer)
             libf.wgpuBufferRelease(internal)
-            self._device._poll()
-            # Note: from the memtests it looks like we need to poll the device
-            # after releasing an object for some objects (buffer, texture,
-            # texture view, sampler, pipeline layout, compute pipeline, and
-            # render pipeline). But not others. Would be nice to at some point
-            # have more clarity on this. In the mean time, we now poll the
-            # device quite a bit, so leaks by not polling the device after
-            # releasing something are highly unlikely.
 
 
 class GPUTexture(classes.GPUTexture, GPUObjectBase):
@@ -1886,7 +1878,6 @@ def _destroy(self):
             self._internal, internal = None, self._internal
             # H: void f(WGPUTexture texture)
             libf.wgpuTextureRelease(internal)
-            self._device._poll()
 
 
 class GPUTextureView(classes.GPUTextureView, GPUObjectBase):
@@ -1895,7 +1886,6 @@ def _destroy(self):
             self._internal, internal = None, self._internal
             # H: void f(WGPUTextureView textureView)
             libf.wgpuTextureViewRelease(internal)
-            self._device._poll()
 
 
 class GPUSampler(classes.GPUSampler, GPUObjectBase):
@@ -1904,7 +1894,6 @@ def _destroy(self):
             self._internal, internal = None, self._internal
             # H: void f(WGPUSampler sampler)
             libf.wgpuSamplerRelease(internal)
-            self._device._poll()
 
 
 class GPUBindGroupLayout(classes.GPUBindGroupLayout, GPUObjectBase):
@@ -1929,7 +1918,6 @@ def _destroy(self):
             self._internal, internal = None, self._internal
             # H: void f(WGPUPipelineLayout pipelineLayout)
             libf.wgpuPipelineLayoutRelease(internal)
-            self._device._poll()
 
 
 class GPUShaderModule(classes.GPUShaderModule, GPUObjectBase):
@@ -1991,7 +1979,6 @@ def _destroy(self):
             self._internal, internal = None, self._internal
             # H: void f(WGPUComputePipeline computePipeline)
             libf.wgpuComputePipelineRelease(internal)
-            self._device._poll()
 
 
 class GPURenderPipeline(classes.GPURenderPipeline, GPUPipelineBase, GPUObjectBase):
@@ -2000,7 +1987,6 @@ def _destroy(self):
             self._internal, internal = None, self._internal
             # H: void f(WGPURenderPipeline renderPipeline)
             libf.wgpuRenderPipelineRelease(internal)
-            self._device._poll()
 
 
 class GPUCommandBuffer(classes.GPUCommandBuffer, GPUObjectBase):
@@ -2504,9 +2490,6 @@ def finish(self, *, label=""):
         )
         # H: WGPUCommandBuffer f(WGPUCommandEncoder commandEncoder, WGPUCommandBufferDescriptor const * descriptor)
         id = libf.wgpuCommandEncoderFinish(self._internal, struct)
-        # WGPU destroys the command encoder when it's finished. So we set
-        # _internal to None to avoid releasing a nonexistent object.
-        self._internal = None
         return GPUCommandBuffer(label, id, self)
 
     def resolve_query_set(
@@ -2650,10 +2633,6 @@ def submit(self, command_buffers):
         c_command_buffers = ffi.new("WGPUCommandBuffer []", command_buffer_ids)
         # H: void f(WGPUQueue queue, size_t commandCount, WGPUCommandBuffer const * commands)
         libf.wgpuQueueSubmit(self._internal, len(command_buffer_ids), c_command_buffers)
-        # WGPU destroys the resource when submitting. We follow this
-        # to avoid releasing a nonexistent object.
-        for cb in command_buffers:
-            cb._internal = None
 
     def write_buffer(self, buffer, buffer_offset, data, data_offset=0, size=None):
         # We support anything that memoryview supports, i.e. anything