Fix memory leaks

Author: junjihashimoto

gpu.hpp

@@ -1002,6 +1002,7 @@ inline void wait(Context &ctx, std::future<void> &future) {
 inline void toCPU(Context &ctx, Tensor &tensor, void *data, size_t bufferSize,
                   CopyData &op) {
   wgpuQueueSubmit(ctx.queue, 1, &op.commandBuffer);
+  wgpuCommandBufferRelease(op.commandBuffer);
   CallbackData callbackData = {op.readbackBuffer, bufferSize, data, &op.promise,
                                &op.future};
   wgpuQueueOnSubmittedWorkDone(
@@ -1062,9 +1063,14 @@ inline void toCPU(Context &ctx, Tensor &tensor, void *data, size_t bufferSize) {
     wgpuCommandEncoderCopyBufferToBuffer(commandEncoder, tensor.data.buffer, 0,
                                          op.readbackBuffer, 0, bufferSize);
     op.commandBuffer = wgpuCommandEncoderFinish(commandEncoder, nullptr);
+    wgpuComputePassEncoderRelease(computePassEncoder);
+    wgpuCommandEncoderRelease(commandEncoder);
     check(op.commandBuffer, "Create command buffer", __FILE__, __LINE__);
   }
   toCPU(ctx, tensor, data, bufferSize, op);
+  if (op.readbackBuffer) {
+    wgpuBufferRelease(op.readbackBuffer);
+  }
 }
 
 /**
@@ -1103,9 +1109,12 @@ inline void toCPU(Context &ctx, WGPUBuffer buffer, void *data,
     wgpuCommandEncoderCopyBufferToBuffer(commandEncoder, buffer, 0,
                                          op.readbackBuffer, 0, bufferSize);
     op.commandBuffer = wgpuCommandEncoderFinish(commandEncoder, nullptr);
+    wgpuComputePassEncoderRelease(computePassEncoder);
+    wgpuCommandEncoderRelease(commandEncoder);
     check(op.commandBuffer, "Create command buffer", __FILE__, __LINE__);
   }
   wgpuQueueSubmit(ctx.queue, 1, &op.commandBuffer);
+  wgpuCommandBufferRelease(op.commandBuffer);
   CallbackData callbackData = {op.readbackBuffer, bufferSize, data, &op.promise,
                                &op.future};
   wgpuQueueOnSubmittedWorkDone(
@@ -1131,6 +1140,9 @@ inline void toCPU(Context &ctx, WGPUBuffer buffer, void *data,
       },
       &callbackData);
   wait(ctx, op.future);
+  if (op.readbackBuffer) {
+    wgpuBufferRelease(op.readbackBuffer);
+  }
 }
 
 
@@ -1217,6 +1229,9 @@ inline void resetCommandBuffer(WGPUDevice &device, Kernel &op) {
         op->totalWorkgroups[2]);
     wgpuComputePassEncoderEnd(computePassEncoder);
     op->commandBuffer = wgpuCommandEncoderFinish(commandEncoder, nullptr);
+    wgpuComputePassEncoderRelease(computePassEncoder);
+    wgpuCommandEncoderRelease(commandEncoder);
+
     op->used = false;
   }
 }
@@ -1507,6 +1522,7 @@ inline void dispatchKernel(Context &ctx, Kernel &kernel,
     resetCommandBuffer(ctx.device, kernel);
   }
   wgpuQueueSubmit(ctx.queue, 1, &kernel->commandBuffer);
+  wgpuCommandBufferRelease(kernel->commandBuffer);
   kernel->used = true;
   wgpuQueueOnSubmittedWorkDone(
       ctx.queue,