[GPU] Set VGT_DRAW_INITIATOR and use major mode from it

src/xenia/gpu/command_processor.cc

@@ -1153,44 +1153,51 @@ bool CommandProcessor::ExecutePacketType3_DRAW_INDX(RingBuffer* reader,
   // ID = dword0 & 0x3F;
   // use = dword0 & 0x40;
   uint32_t dword0 = reader->ReadAndSwap<uint32_t>();  // viz query info
-  uint32_t dword1 = reader->ReadAndSwap<uint32_t>();
+  reg::VGT_DRAW_INITIATOR vgt_draw_initiator;
-  uint32_t index_count = dword1 >> 16;
+  vgt_draw_initiator.value = reader->ReadAndSwap<uint32_t>();
-  auto prim_type = static_cast<PrimitiveType>(dword1 & 0x3F);
+  WriteRegister(XE_GPU_REG_VGT_DRAW_INITIATOR, vgt_draw_initiator.value);
+
   bool is_indexed = false;
   IndexBufferInfo index_buffer_info;
-  uint32_t src_sel = (dword1 >> 6) & 0x3;
+  switch (vgt_draw_initiator.source_select) {
-  if (src_sel == 0x0) {
+    case xenos::SourceSelect::kDMA: {
-    // DI_SRC_SEL_DMA
       // Indexed draw.
       is_indexed = true;
       index_buffer_info.guest_base = reader->ReadAndSwap<uint32_t>();
       uint32_t index_size = reader->ReadAndSwap<uint32_t>();
       index_buffer_info.endianness = static_cast<Endian>(index_size >> 30);
       index_size &= 0x00FFFFFF;
-    bool index_32bit = (dword1 >> 11) & 0x1;
+      index_buffer_info.format = vgt_draw_initiator.index_size;
-    index_buffer_info.format =
+      index_size *=
-        index_32bit ? IndexFormat::kInt32 : IndexFormat::kInt16;
+          (vgt_draw_initiator.index_size == IndexFormat::kInt32) ? 4 : 2;
-    index_size *= index_32bit ? 4 : 2;
       index_buffer_info.length = index_size;
-    index_buffer_info.count = index_count;
+      index_buffer_info.count = vgt_draw_initiator.num_indices;
-  } else if (src_sel == 0x1) {
+    } break;
-    // DI_SRC_SEL_IMMEDIATE
+    case xenos::SourceSelect::kImmediate: {
+      // TODO(Triang3l): VGT_IMMED_DATA.
       assert_always();
-  } else if (src_sel == 0x2) {
+    } break;
-    // DI_SRC_SEL_AUTO_INDEX
+    case xenos::SourceSelect::kAutoIndex: {
       // Auto draw.
       index_buffer_info.guest_base = 0;
       index_buffer_info.length = 0;
-  } else {
+    } break;
+    default: {
       // Invalid source select.
       assert_always();
+    } break;
   }
 
-  bool success = IssueDraw(prim_type, index_count,
+  bool success =
-                           is_indexed ? &index_buffer_info : nullptr);
+      IssueDraw(vgt_draw_initiator.prim_type, vgt_draw_initiator.num_indices,
+                is_indexed ? &index_buffer_info : nullptr,
+                xenos::IsMajorModeExplicit(vgt_draw_initiator.major_mode,
+                                           vgt_draw_initiator.prim_type));
   if (!success) {
-    XELOGE("PM4_DRAW_INDX(%d, %d, %d): Failed in backend", index_count,
+    XELOGE("PM4_DRAW_INDX(%d, %d, %d): Failed in backend",
-           prim_type, src_sel);
+           vgt_draw_initiator.num_indices,
+           uint32_t(vgt_draw_initiator.prim_type),
+           uint32_t(vgt_draw_initiator.source_select));
   }
 
   return true;
@@ -1200,21 +1207,27 @@ bool CommandProcessor::ExecutePacketType3_DRAW_INDX_2(RingBuffer* reader,
                                                       uint32_t packet,
                                                       uint32_t count) {
   // draw using supplied indices in packet
-  uint32_t dword0 = reader->ReadAndSwap<uint32_t>();
+  reg::VGT_DRAW_INITIATOR vgt_draw_initiator;
-  uint32_t index_count = dword0 >> 16;
+  vgt_draw_initiator.value = reader->ReadAndSwap<uint32_t>();
-  auto prim_type = static_cast<PrimitiveType>(dword0 & 0x3F);
+  WriteRegister(XE_GPU_REG_VGT_DRAW_INITIATOR, vgt_draw_initiator.value);
-  uint32_t src_sel = (dword0 >> 6) & 0x3;
+  assert_true(vgt_draw_initiator.source_select ==
-  assert_true(src_sel == 0x2);  // 'SrcSel=AutoIndex'
+              xenos::SourceSelect::kAutoIndex);
   // Index buffer unused as automatic.
-  // bool index_32bit = (dword0 >> 11) & 0x1;
+  // uint32_t indices_size =
-  // uint32_t indices_size = index_count * (index_32bit ? 4 : 2);
+  //     vgt_draw_initiator.num_indices *
+  //         (vgt_draw_initiator.index_size == IndexFormat::kInt32 ? 4 : 2);
   // uint32_t index_ptr = reader->ptr();
+  // TODO(Triang3l): VGT_IMMED_DATA.
   reader->AdvanceRead((count - 1) * sizeof(uint32_t));
 
-  bool success = IssueDraw(prim_type, index_count, nullptr);
+  bool success = IssueDraw(
+      vgt_draw_initiator.prim_type, vgt_draw_initiator.num_indices, nullptr,
+      xenos::IsMajorModeExplicit(vgt_draw_initiator.major_mode,
+                                 vgt_draw_initiator.prim_type));
   if (!success) {
-    XELOGE("PM4_DRAW_INDX_IMM(%d, %d): Failed in backend", index_count,
+    XELOGE("PM4_DRAW_INDX_IMM(%d, %d): Failed in backend",
-           prim_type);
+           vgt_draw_initiator.num_indices,
+           uint32_t(vgt_draw_initiator.prim_type));
   }
 
   return true;

src/xenia/gpu/command_processor.h

@@ -239,7 +239,8 @@ class CommandProcessor {
                              uint32_t dword_count) = 0;
 
   virtual bool IssueDraw(PrimitiveType prim_type, uint32_t index_count,
-                         IndexBufferInfo* index_buffer_info) = 0;
+                         IndexBufferInfo* index_buffer_info,
+                         bool major_mode_explicit) = 0;
   virtual bool IssueCopy() = 0;
 
   virtual void InitializeTrace() = 0;

src/xenia/gpu/d3d12/d3d12_command_processor.cc

@@ -1245,7 +1245,8 @@ Shader* D3D12CommandProcessor::LoadShader(ShaderType shader_type,
 
 bool D3D12CommandProcessor::IssueDraw(PrimitiveType primitive_type,
                                       uint32_t index_count,
-                                      IndexBufferInfo* index_buffer_info) {
+                                      IndexBufferInfo* index_buffer_info,
+                                      bool major_mode_explicit) {
   auto device = GetD3D12Context()->GetD3D12Provider()->GetDevice();
   auto& regs = *register_file_;
 
@@ -1272,8 +1273,7 @@ bool D3D12CommandProcessor::IssueDraw(PrimitiveType primitive_type,
 
   // Check if using tessellation to get the correct primitive type.
   bool tessellated;
-  if (uint32_t(primitive_type) >=
+  if (major_mode_explicit) {
-      uint32_t(PrimitiveType::kExplicitMajorModeForceStart)) {
     tessellated = regs.Get<reg::VGT_OUTPUT_PATH_CNTL>().path_select ==
                   xenos::VGTOutputPath::kTessellationEnable;
   } else {

src/xenia/gpu/d3d12/d3d12_command_processor.h

@@ -171,7 +171,8 @@ class D3D12CommandProcessor : public CommandProcessor {
                      uint32_t dword_count) override;
 
   bool IssueDraw(PrimitiveType primitive_type, uint32_t index_count,
-                 IndexBufferInfo* index_buffer_info) override;
+                 IndexBufferInfo* index_buffer_info,
+                 bool major_mode_explicit) override;
   bool IssueCopy() override;
 
   void InitializeTrace() override;

src/xenia/gpu/null/null_command_processor.cc

@@ -44,7 +44,8 @@ Shader* NullCommandProcessor::LoadShader(ShaderType shader_type,
 
 bool NullCommandProcessor::IssueDraw(PrimitiveType prim_type,
                                      uint32_t index_count,
-                                     IndexBufferInfo* index_buffer_info) {
+                                     IndexBufferInfo* index_buffer_info,
+                                     bool major_mode_explicit) {
   return true;
 }
 

src/xenia/gpu/null/null_command_processor.h

@@ -41,7 +41,8 @@ class NullCommandProcessor : public CommandProcessor {
                      uint32_t dword_count) override;
 
   bool IssueDraw(PrimitiveType prim_type, uint32_t index_count,
-                 IndexBufferInfo* index_buffer_info) override;
+                 IndexBufferInfo* index_buffer_info,
+                 bool major_mode_explicit) override;
   bool IssueCopy() override;
 
   void InitializeTrace() override;

src/xenia/gpu/register_table.inc

@@ -131,7 +131,10 @@ XE_GPU_REGISTER(0x2182, kDword, SQ_INTERPOLATOR_CNTL)
 XE_GPU_REGISTER(0x2183, kDword, SQ_WRAPPING_0)
 XE_GPU_REGISTER(0x2184, kDword, SQ_WRAPPING_1)
 
+// These three registers are set by the command processor.
 XE_GPU_REGISTER(0x21F9, kDword, VGT_EVENT_INITIATOR)
+XE_GPU_REGISTER(0x21FC, kDword, VGT_DRAW_INITIATOR)
+XE_GPU_REGISTER(0x21FD, kDword, VGT_IMMED_DATA)
 
 XE_GPU_REGISTER(0x2200, kDword, RB_DEPTHCONTROL)
 XE_GPU_REGISTER(0x2201, kDword, RB_BLENDCONTROL0)

src/xenia/gpu/registers.h

@@ -145,6 +145,22 @@ union SQ_CONTEXT_MISC {
 
 *******************************************************************************/
 
+union VGT_DRAW_INITIATOR {
+  // Different than on A2xx and R6xx/R7xx.
+  struct {
+    PrimitiveType prim_type : 6;            // +0
+    xenos::SourceSelect source_select : 2;  // +6
+    xenos::MajorMode major_mode : 2;        // +8
+    uint32_t : 1;                           // +10
+    IndexFormat index_size : 1;             // +11
+    uint32_t not_eop : 1;                   // +12
+    uint32_t : 3;                           // +13
+    uint32_t num_indices : 16;              // +16
+  };
+  uint32_t value;
+  static constexpr Register register_index = XE_GPU_REG_VGT_DRAW_INITIATOR;
+};
+
 union VGT_OUTPUT_PATH_CNTL {
   struct {
     xenos::VGTOutputPath path_select : 2;  // +0

src/xenia/gpu/vk/vulkan_command_processor.cc

@@ -40,7 +40,8 @@ Shader* VulkanCommandProcessor::LoadShader(ShaderType shader_type,
 
 bool VulkanCommandProcessor::IssueDraw(PrimitiveType primitive_type,
                                        uint32_t index_count,
-                                       IndexBufferInfo* index_buffer_info) {
+                                       IndexBufferInfo* index_buffer_info,
+                                       bool major_mode_explicit) {
   return true;
 }
 

src/xenia/gpu/vk/vulkan_command_processor.h

@@ -40,7 +40,8 @@ class VulkanCommandProcessor : public CommandProcessor {
                      uint32_t dword_count) override;
 
   bool IssueDraw(PrimitiveType primitive_type, uint32_t index_count,
-                 IndexBufferInfo* index_buffer_info) override;
+                 IndexBufferInfo* index_buffer_info,
+                 bool major_mode_explicit) override;
   bool IssueCopy() override;
 
   void InitializeTrace() override;

src/xenia/gpu/vulkan/vulkan_command_processor.cc

@@ -598,7 +598,8 @@ Shader* VulkanCommandProcessor::LoadShader(ShaderType shader_type,
 
 bool VulkanCommandProcessor::IssueDraw(PrimitiveType primitive_type,
                                        uint32_t index_count,
-                                       IndexBufferInfo* index_buffer_info) {
+                                       IndexBufferInfo* index_buffer_info,
+                                       bool major_mode_explicit) {
   auto& regs = *register_file_;
 
 #if FINE_GRAINED_DRAW_SCOPES

src/xenia/gpu/vulkan/vulkan_command_processor.h

@@ -81,7 +81,8 @@ class VulkanCommandProcessor : public CommandProcessor {
                      uint32_t dword_count) override;
 
   bool IssueDraw(PrimitiveType primitive_type, uint32_t index_count,
-                 IndexBufferInfo* index_buffer_info) override;
+                 IndexBufferInfo* index_buffer_info,
+                 bool major_mode_explicit) override;
   bool PopulateConstants(VkCommandBuffer command_buffer,
                          VulkanShader* vertex_shader,
                          VulkanShader* pixel_shader);

src/xenia/gpu/xenos.h

@@ -37,12 +37,11 @@ enum class PrimitiveType : uint32_t {
   kQuadStrip = 0x0E,
   kPolygon = 0x0F,
 
-  // Starting with this primitive mode, registers like VGT_OUTPUT_PATH_CNTL have
+  // Starting with this primitive type, explicit major mode is assumed (in the
-  // effect (deduced from R6xx/R7xx registers, and Halo 3 also doesn't reset
+  // R6xx/R7xx registers, k2DCopyRectListV0 is 22, and implicit major mode is
-  // VGT_OUTPUT_PATH_CNTL after the first draw with tessellation).
+  // only used for primitive types 0 through 21) - and tessellation patches use
-  // TODO(Triang3l): Find out if VGT_DRAW_INITIATOR (0x21FC on Adreno 2xx, but
+  // the range that starts from k2DCopyRectListV0.
-  // not seen being used in games) specifies the major mode (or if it's set
+  // TODO(Triang3l): Verify if this is also true for the Xenos.
-  // somewhere else).
   kExplicitMajorModeForceStart = 0x10,
 
   k2DCopyRectListV0 = 0x10,
@@ -460,6 +459,25 @@ typedef enum {
   XE_GPU_INVALIDATE_MASK_ALL = 0x7FFF,
 } XE_GPU_INVALIDATE_MASK;
 
+// VGT_DRAW_INITIATOR::DI_SRC_SEL_*
+enum class SourceSelect : uint32_t {
+  kDMA,
+  kImmediate,
+  kAutoIndex,
+};
+
+// VGT_DRAW_INITIATOR::DI_MAJOR_MODE_*
+enum class MajorMode : uint32_t {
+  kImplicit,
+  kExplicit,
+};
+
+inline bool IsMajorModeExplicit(MajorMode major_mode,
+                                PrimitiveType primitive_type) {
+  return major_mode != MajorMode::kImplicit ||
+         primitive_type >= PrimitiveType::kExplicitMajorModeForceStart;
+}
+
 // instr_arbitrary_filter_t
 enum class ArbitraryFilter : uint32_t {
   k2x4Sym = 0,