dma_pushbuffer: Optimize to avoid loop and copy on Push.

src/core/hle/service/nvdrv/devices/nvhost_gpu.cpp

@@ -128,11 +128,9 @@ u32 nvhost_gpu::AllocateObjectContext(const std::vector<u8>& input, std::vector<
     return 0;
 }
 
-static void PushGPUEntries(const std::vector<Tegra::CommandListHeader>& entries) {
+static void PushGPUEntries(Tegra::CommandList&& entries) {
     auto& dma_pusher{Core::System::GetInstance().GPU().DmaPusher()};
-    for (const auto& entry : entries) {
+    dma_pusher.Push(std::move(entries));
-        dma_pusher.Push(entry);
-    }
     dma_pusher.DispatchCalls();
 }
 
@@ -149,11 +147,11 @@ u32 nvhost_gpu::SubmitGPFIFO(const std::vector<u8>& input, std::vector<u8>& outp
                                    params.num_entries * sizeof(Tegra::CommandListHeader),
                "Incorrect input size");
 
-    std::vector<Tegra::CommandListHeader> entries(params.num_entries);
+    Tegra::CommandList entries(params.num_entries);
     std::memcpy(entries.data(), &input[sizeof(IoctlSubmitGpfifo)],
                 params.num_entries * sizeof(Tegra::CommandListHeader));
 
-    PushGPUEntries(entries);
+    PushGPUEntries(std::move(entries));
 
     params.fence_out.id = 0;
     params.fence_out.value = 0;
@@ -170,11 +168,11 @@ u32 nvhost_gpu::KickoffPB(const std::vector<u8>& input, std::vector<u8>& output)
     LOG_WARNING(Service_NVDRV, "(STUBBED) called, gpfifo={:X}, num_entries={:X}, flags={:X}",
                 params.address, params.num_entries, params.flags);
 
-    std::vector<Tegra::CommandListHeader> entries(params.num_entries);
+    Tegra::CommandList entries(params.num_entries);
     Memory::ReadBlock(params.address, entries.data(),
                       params.num_entries * sizeof(Tegra::CommandListHeader));
 
-    PushGPUEntries(entries);
+    PushGPUEntries(std::move(entries));
 
     params.fence_out.id = 0;
     params.fence_out.value = 0;

src/video_core/dma_pusher.cpp

@@ -23,6 +23,8 @@ void DmaPusher::DispatchCalls() {
     // On entering GPU code, assume all memory may be touched by the ARM core.
     gpu.Maxwell3D().dirty_flags.OnMemoryWrite();
 
+    dma_pushbuffer_subindex = 0;
+
     while (Core::System::GetInstance().IsPoweredOn()) {
         if (!Step()) {
             break;
@@ -89,11 +91,17 @@ bool DmaPusher::Step() {
         }
     } else if (ib_enable && !dma_pushbuffer.empty()) {
         // Current pushbuffer empty, but we have more IB entries to read
-        const CommandListHeader& command_list_header{dma_pushbuffer.front()};
+        const CommandList& command_list{dma_pushbuffer.front()};
+        const CommandListHeader& command_list_header{command_list[dma_pushbuffer_subindex++]};
         dma_get = command_list_header.addr;
         dma_put = dma_get + command_list_header.size * sizeof(u32);
         non_main = command_list_header.is_non_main;
-        dma_pushbuffer.pop();
+
+        if (dma_pushbuffer_subindex >= command_list.size()) {
+            // We've gone through the current list, remove it from the queue
+            dma_pushbuffer.pop();
+            dma_pushbuffer_subindex = 0;
+        }
     } else {
         // Otherwise, pushbuffer empty and IB empty or nonexistent - nothing to do
         return {};

src/video_core/dma_pusher.h

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <vector>
 #include <queue>
 
 #include "common/bit_field.h"
@@ -45,6 +46,8 @@ static_assert(sizeof(CommandHeader) == sizeof(u32), "CommandHeader has incorrect
 
 class GPU;
 
+using CommandList = std::vector<Tegra::CommandListHeader>;
+
 /**
  * The DmaPusher class implements DMA submission to FIFOs, providing an area of memory that the
  * emulated app fills with commands and tells PFIFO to process. The pushbuffers are then assembled
@@ -57,8 +60,8 @@ public:
     explicit DmaPusher(GPU& gpu);
     ~DmaPusher();
 
-    void Push(const CommandListHeader& command_list_header) {
+    void Push(CommandList&& entries) {
-        dma_pushbuffer.push(command_list_header);
+        dma_pushbuffer.push(std::move(entries));
     }
 
     void DispatchCalls();
@@ -72,7 +75,8 @@ private:
 
     GPU& gpu;
 
-    std::queue<CommandListHeader> dma_pushbuffer;
+    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
 
     struct DmaState {
         u32 method;            ///< Current method