video_core/dma_pusher: The full list of headers at once.

Fetching every u32 from memory leads to a big overhead. So let's fetch all of them as a block if possible.
This reduces the Memory::* calls by the dma_pusher by a factor of 10.

src/video_core/dma_pusher.cpp

@@ -38,13 +38,13 @@ bool DmaPusher::Step() {
         const auto address = gpu.MemoryManager().GpuToCpuAddress(dma_get);
         ASSERT_MSG(address, "Invalid GPU address");
 
-        const CommandHeader command_header{Memory::Read32(*address)};
+        GPUVAddr size = dma_put - dma_get;
+        ASSERT_MSG(size % sizeof(CommandHeader) == 0, "Invalid aligned GPU addresses");
+        command_headers.resize(size / sizeof(CommandHeader));
 
-        dma_get += sizeof(u32);
+        Memory::ReadBlock(*address, command_headers.data(), size);
 
-        if (!non_main) {
-            dma_mget = dma_get;
-        }
+        for (const CommandHeader& command_header : command_headers) {
 
             // now, see if we're in the middle of a command
             if (dma_state.length_pending) {
@@ -91,6 +91,14 @@ bool DmaPusher::Step() {
                     break;
                 }
             }
+        }
+
+        dma_get = dma_put;
+
+        if (!non_main) {
+            // TODO (degasus): This is dead code, as dma_mget is never read.
+            dma_mget = dma_get;
+        }
     } else if (ib_enable && !dma_pushbuffer.empty()) {
         // Current pushbuffer empty, but we have more IB entries to read
         const CommandList& command_list{dma_pushbuffer.front()};

src/video_core/dma_pusher.h

@@ -75,6 +75,8 @@ private:
 
     GPU& gpu;
 
+    std::vector<CommandHeader> command_headers; ///< Buffer for list of commands fetched at once
+
     std::queue<CommandList> dma_pushbuffer; ///< Queue of command lists to be processed
     std::size_t dma_pushbuffer_subindex{};  ///< Index within a command list within the pushbuffer