1 files changed, 34 insertions, 3 deletions

diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index 2ee980bab..c7ec1eac9 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -4,6 +4,7 @@
 
 #include "common/assert.h"
 #include "common/logging/log.h"
+#include "common/microprofile.h"
 #include "common/settings.h"
 #include "core/core.h"
 #include "video_core/engines/maxwell_3d.h"
@@ -12,6 +13,9 @@
 #include "video_core/renderer_base.h"
 #include "video_core/textures/decoders.h"
 
+MICROPROFILE_DECLARE(GPU_DMAEngine);
+MICROPROFILE_DEFINE(GPU_DMAEngine, "GPU", "DMA Engine", MP_RGB(224, 224, 128));
+
 namespace Tegra::Engines {
 
 using namespace Texture;
@@ -21,6 +25,10 @@ MaxwellDMA::MaxwellDMA(Core::System& system_, MemoryManager& memory_manager_)
 
 MaxwellDMA::~MaxwellDMA() = default;
 
+void MaxwellDMA::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_) {
+    rasterizer = rasterizer_;
+}
+
 void MaxwellDMA::CallMethod(u32 method, u32 method_argument, bool is_last_call) {
     ASSERT_MSG(method < NUM_REGS, "Invalid MaxwellDMA register");
 
@@ -39,12 +47,12 @@ void MaxwellDMA::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
 }
 
 void MaxwellDMA::Launch() {
+    MICROPROFILE_SCOPE(GPU_DMAEngine);
     LOG_TRACE(Render_OpenGL, "DMA copy 0x{:x} -> 0x{:x}", static_cast<GPUVAddr>(regs.offset_in),
               static_cast<GPUVAddr>(regs.offset_out));
 
     // TODO(Subv): Perform more research and implement all features of this engine.
     const LaunchDMA& launch = regs.launch_dma;
-    ASSERT(launch.remap_enable == 0);
     ASSERT(launch.semaphore_type == LaunchDMA::SemaphoreType::NONE);
     ASSERT(launch.interrupt_type == LaunchDMA::InterruptType::NONE);
     ASSERT(launch.data_transfer_type == LaunchDMA::DataTransferType::NON_PIPELINED);
@@ -77,11 +85,31 @@ void MaxwellDMA::CopyPitchToPitch() {
     // When `multi_line_enable` bit is disabled the copy is performed as if we were copying a 1D
     // buffer of length `line_length_in`.
     // Otherwise we copy a 2D image of dimensions (line_length_in, line_count).
+    auto& accelerate = rasterizer->AccessAccelerateDMA();
     if (!regs.launch_dma.multi_line_enable) {
-        memory_manager.CopyBlock(regs.offset_out, regs.offset_in, regs.line_length_in);
+        const bool is_buffer_clear = regs.launch_dma.remap_enable != 0 &&
+                                     regs.remap_const.dst_x == RemapConst::Swizzle::CONST_A;
+        // TODO: allow multisized components.
+        if (is_buffer_clear) {
+            ASSERT(regs.remap_const.component_size_minus_one == 3);
+            accelerate.BufferClear(regs.offset_out, regs.line_length_in, regs.remap_consta_value);
+            std::vector<u32> tmp_buffer(regs.line_length_in, regs.remap_consta_value);
+            memory_manager.WriteBlockUnsafe(regs.offset_out,
+                                            reinterpret_cast<u8*>(tmp_buffer.data()),
+                                            regs.line_length_in * sizeof(u32));
+            return;
+        }
+        UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
+        if (!accelerate.BufferCopy(regs.offset_in, regs.offset_out, regs.line_length_in)) {
+            std::vector<u8> tmp_buffer(regs.line_length_in);
+            memory_manager.ReadBlockUnsafe(regs.offset_in, tmp_buffer.data(), regs.line_length_in);
+            memory_manager.WriteBlock(regs.offset_out, tmp_buffer.data(), regs.line_length_in);
+        }
         return;
     }
 
+    UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
+
     // Perform a line-by-line copy.
     // We're going to take a subrect of size (line_length_in, line_count) from the source rectangle.
     // There is no need to manually flush/invalidate the regions because CopyBlock does that for us.
@@ -99,12 +127,14 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
 
     // Optimized path for micro copies.
     const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
-    if (dst_size < GOB_SIZE && regs.pitch_out <= GOB_SIZE_X) {
+    if (dst_size < GOB_SIZE && regs.pitch_out <= GOB_SIZE_X &&
+        regs.src_params.height > GOB_SIZE_Y) {
         FastCopyBlockLinearToPitch();
         return;
     }
 
     // Deswizzle the input and copy it over.
+    UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
     const u32 bytes_per_pixel = regs.pitch_out / regs.line_length_in;
     const Parameters& src_params = regs.src_params;
     const u32 width = src_params.width;
@@ -134,6 +164,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
 
 void MaxwellDMA::CopyPitchToBlockLinear() {
     UNIMPLEMENTED_IF_MSG(regs.dst_params.block_size.width != 0, "Block width is not one");
+    UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
 
     const auto& dst_params = regs.dst_params;
     const u32 bytes_per_pixel = regs.pitch_in / regs.line_length_in;