Add some comments.

Common/GPU/Vulkan/VulkanMemory.h

@@ -135,6 +135,8 @@ public:
 		return blocks_[curBlockIndex_].writePtr;
 	}
 
+	// NOTE: If you can avoid this by writing the data directly into memory returned from Allocate,
+	// do so. Savings from avoiding memcpy can be significant.
 	VkDeviceSize Push(const void *data, VkDeviceSize numBytes, int alignment, VkBuffer *vkbuf) {
 		uint32_t bindOffset;
 		uint8_t *ptr = Allocate(numBytes, alignment, vkbuf, &bindOffset);

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -825,6 +825,8 @@ void DrawEngineVulkan::DoFlush() {
 
 		VkDescriptorSet ds = GetOrCreateDescriptorSet(imageView, sampler, baseBuf, lightBuf, boneBuf, tess);
 
+		// TODO: Can we avoid binding all three when not needed? Same below for hardware transform.
+		// Think this will require different descriptor set layouts.
 		const uint32_t dynamicUBOOffsets[3] = {
 			baseUBOOffset, lightUBOOffset, boneUBOOffset,
 		};