Reimplement draw call merging & refactor aurora::gx (#37)

Author: encounter

CMakeLists.txt

@@ -71,13 +71,13 @@ if (AURORA_ENABLE_GX)
     # aurora_gx sources
     lib/gfx/common.cpp
     lib/gfx/texture.cpp
-    lib/gfx/gx.cpp
-    lib/gfx/gx_shader.cpp
     lib/gfx/texture_convert.cpp
-    lib/gfx/shader_info.cpp
-    lib/gfx/model/shader.cpp
-    lib/gfx/fifo.cpp
-    lib/gfx/command_processor.cpp
+    lib/gx/command_processor.cpp
+    lib/gx/fifo.cpp
+    lib/gx/gx.cpp
+    lib/gx/pipeline.cpp
+    lib/gx/shader.cpp
+    lib/gx/shader_info.cpp
     lib/dolphin/gx/GXBump.cpp
     lib/dolphin/gx/GXCull.cpp
     lib/dolphin/gx/GXDispList.cpp

cmake/aurora_gx.cmake

@@ -1,13 +1,13 @@
 add_library(aurora_gx STATIC
         lib/gfx/common.cpp
         lib/gfx/texture.cpp
-        lib/gfx/gx.cpp
-        lib/gfx/gx_shader.cpp
         lib/gfx/texture_convert.cpp
-        lib/gfx/shader_info.cpp
-        lib/gfx/model/shader.cpp
-        lib/gfx/fifo.cpp
-        lib/gfx/command_processor.cpp
+        lib/gx/command_processor.cpp
+        lib/gx/fifo.cpp
+        lib/gx/gx.cpp
+        lib/gx/pipeline.cpp
+        lib/gx/shader.cpp
+        lib/gx/shader_info.cpp
         lib/dolphin/gx/GXBump.cpp
         lib/dolphin/gx/GXCull.cpp
         lib/dolphin/gx/GXDispList.cpp

include/aurora/aurora.h

@@ -102,13 +102,6 @@ void aurora_end_frame();
 AuroraBackend aurora_get_backend();
 const AuroraBackend* aurora_get_available_backends(size_t* count);
 
-#ifndef NDEBUG
-#define AURORA_GFX_DEBUG_GROUPS
-#endif
-
-void push_debug_group(const char* label);
-void pop_debug_group();
-
 #ifdef __cplusplus
 }
 #endif

include/aurora/gfx.h

@@ -0,0 +1,36 @@
+#ifndef AURORA_GFX_H
+#define AURORA_GFX_H
+
+#ifdef __cplusplus
+#include <cstdint>
+
+extern "C" {
+#else
+#include "stdint.h"
+#endif
+
+#ifndef NDEBUG
+#define AURORA_GFX_DEBUG_GROUPS
+#endif
+
+void push_debug_group(const char* label);
+void pop_debug_group();
+
+typedef struct {
+  uint32_t queuedPipelines;
+  uint32_t createdPipelines;
+  uint32_t drawCallCount;
+  uint32_t mergedDrawCallCount;
+  uint32_t lastVertSize;
+  uint32_t lastUniformSize;
+  uint32_t lastIndexSize;
+  uint32_t lastStorageSize;
+} AuroraStats;
+
+const AuroraStats* aurora_get_stats();
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

lib/dolphin/gx/GXDispList.cpp

@@ -1,7 +1,8 @@
 #include "gx.hpp"
 #include "__gx.h"
 
-#include "../../gfx/fifo.hpp"
+#include "../../gx/command_processor.hpp"
+#include "../../gx/fifo.hpp"
 
 #include <cstring>
 
@@ -24,7 +25,7 @@ void GXBeginDisplayList(void* list, u32 size) {
   __gx->inDispList = 1;
 
   // Redirect FIFO writes to the user-provided buffer
-  aurora::gfx::fifo::begin_display_list(static_cast<u8*>(list), size);
+  aurora::gx::fifo::begin_display_list(static_cast<u8*>(list), size);
 }
 
 u32 GXEndDisplayList() {
@@ -34,7 +35,7 @@ u32 GXEndDisplayList() {
   }
 
   // End FIFO redirection and get the byte count (ROUNDUP32)
-  u32 bytesWritten = aurora::gfx::fifo::end_display_list();
+  u32 bytesWritten = aurora::gx::fifo::end_display_list();
 
   // Restore saved shadow register state
   if (__gx->dlSaveContext != 0) {
@@ -60,10 +61,10 @@ void GXCallDisplayList(const void* data, u32 nbytes) {
   // Drain the internal FIFO so that any pending CP register writes
   // (VCD, VAT, etc.) are processed into g_gxState before the display
   // list's draw commands reference them.
-  aurora::gfx::fifo::drain();
+  aurora::gx::fifo::drain();
 
   // Process the display list through the command processor
-  aurora::gfx::fifo::process(static_cast<const u8*>(data), nbytes, true);
+  aurora::gx::fifo::process(static_cast<const u8*>(data), nbytes, true);
 }
 
 void GXCallDisplayListLE(const void* data, u32 nbytes) {
@@ -80,9 +81,9 @@ void GXCallDisplayListLE(const void* data, u32 nbytes) {
   // Drain the internal FIFO so that any pending CP register writes
   // (VCD, VAT, etc.) are processed into g_gxState before the display
   // list's draw commands reference them.
-  aurora::gfx::fifo::drain();
+  aurora::gx::fifo::drain();
 
   // Process the display list through the command processor (little-endian)
-  aurora::gfx::fifo::process(static_cast<const u8*>(data), nbytes, false);
+  aurora::gx::fifo::process(static_cast<const u8*>(data), nbytes, false);
 }
 }

lib/dolphin/gx/GXManage.cpp

@@ -1,7 +1,7 @@
 #include "gx.hpp"
 #include "__gx.h"
 
-#include "../../gfx/fifo.hpp"
+#include "../../gx/fifo.hpp"
 
 #include <cstring>
 
@@ -29,7 +29,7 @@ GXFifoObj* GXInit(void* base, u32 size) {
   __gx->vNum = 0;
 
   // Initialize FIFO subsystem
-  aurora::gfx::fifo::init();
+  aurora::gx::fifo::init();
   GXInitFifoBase(&sFifoObj, base, size);
   GXSetCPUFifo(&sFifoObj);
   GXSetGPFifo(&sFifoObj);

lib/dolphin/gx/GXVert.cpp

@@ -1,6 +1,6 @@
 #include "gx.hpp"
 #include "__gx.h"
-#include "../../gfx/fifo.hpp"
+#include "../../gx/fifo.hpp"
 
 // Track vertex count between GXBegin/GXEnd for mismatch detection
 static u16 sBeginNVerts = 0;
@@ -27,13 +27,13 @@ void GXBegin(GXPrimitive primitive, GXVtxFmt vtxFmt, u16 nVerts) {
 
   // Record state for vertex count validation in GXEnd
   sBeginNVerts = nVerts;
-  sBeginFifoSize = aurora::gfx::fifo::get_buffer_size();
+  sBeginFifoSize = aurora::gx::fifo::get_buffer_size();
   sInBegin = true;
 }
 
 void GXEnd() {
   if (sInBegin) {
-    u32 bytesWritten = aurora::gfx::fifo::get_buffer_size() - sBeginFifoSize;
+    u32 bytesWritten = aurora::gx::fifo::get_buffer_size() - sBeginFifoSize;
     if (sBeginNVerts > 0 && bytesWritten > 0) {
       u32 vtxSize = bytesWritten / sBeginNVerts;
       u32 remainder = bytesWritten % sBeginNVerts;
@@ -48,8 +48,8 @@ void GXEnd() {
     }
     sInBegin = false;
   }
-  if (!aurora::gfx::fifo::in_display_list()) {
-    aurora::gfx::fifo::drain();
+  if (!aurora::gx::fifo::in_display_list()) {
+    aurora::gx::fifo::drain();
   }
 }
 
@@ -201,28 +201,27 @@ void GXTexCoord1x16(u16 index) { GX_WRITE_U16(index); }
 
 void GXTexCoord1x8(u8 index) { GX_WRITE_U8(index); }
 
-void GXCmd1u8(const u8 x) { aurora::gfx::fifo::write_u8(static_cast<uint8_t>(x)); }
-void GXCmd1u16(const u16 x) { aurora::gfx::fifo::write_u16(static_cast<uint16_t>(x)); }
-void GXCmd1u32(const u32 x) { aurora::gfx::fifo::write_u32(static_cast<uint32_t>(x)); }
-
-void GXParam1u8(const u8 x) { aurora::gfx::fifo::write_u8(static_cast<uint8_t>(x)); }
-void GXParam1u16(const u16 x) { aurora::gfx::fifo::write_u16(static_cast<uint16_t>(x)); }
-void GXParam1u32(const u32 x) { aurora::gfx::fifo::write_u32(static_cast<uint32_t>(x)); }
-void GXParam1s8(const s8 x) { aurora::gfx::fifo::write_u8(static_cast<uint8_t>(x)); }
-void GXParam1s16(const s16 x) { aurora::gfx::fifo::write_u16(static_cast<uint16_t>(x)); }
-void GXParam1s32(const s32 x) { aurora::gfx::fifo::write_u32(static_cast<uint32_t>(x)); }
-void GXParam1f32(const f32 x) { aurora::gfx::fifo::write_f32(x); }
+void GXCmd1u8(const u8 x) { aurora::gx::fifo::write_u8(x); }
+void GXCmd1u16(const u16 x) { aurora::gx::fifo::write_u16(x); }
+void GXCmd1u32(const u32 x) { aurora::gx::fifo::write_u32(x); }
+
+void GXParam1u8(const u8 x) { aurora::gx::fifo::write_u8(x); }
+void GXParam1u16(const u16 x) { aurora::gx::fifo::write_u16(x); }
+void GXParam1u32(const u32 x) { aurora::gx::fifo::write_u32(x); }
+void GXParam1s8(const s8 x) { aurora::gx::fifo::write_u8(static_cast<uint8_t>(x)); }
+void GXParam1s16(const s16 x) { aurora::gx::fifo::write_u16(static_cast<uint16_t>(x)); }
+void GXParam1s32(const s32 x) { aurora::gx::fifo::write_u32(static_cast<uint32_t>(x)); }
+void GXParam1f32(const f32 x) { aurora::gx::fifo::write_f32(x); }
 void GXParam3f32(const f32 x, const f32 y, const f32 z) {
-  aurora::gfx::fifo::write_f32(x);
-  aurora::gfx::fifo::write_f32(y);
-  aurora::gfx::fifo::write_f32(z);
+  aurora::gx::fifo::write_f32(x);
+  aurora::gx::fifo::write_f32(y);
+  aurora::gx::fifo::write_f32(z);
 }
 void GXParam4f32(const f32 x, const f32 y, const f32 z, const f32 w) {
-  aurora::gfx::fifo::write_f32(x);
-  aurora::gfx::fifo::write_f32(y);
-  aurora::gfx::fifo::write_f32(z);
-  aurora::gfx::fifo::write_f32(w);
+  aurora::gx::fifo::write_f32(x);
+  aurora::gx::fifo::write_f32(y);
+  aurora::gx::fifo::write_f32(z);
+  aurora::gx::fifo::write_f32(w);
 }
 
-
 } // extern "C"

lib/dolphin/gx/__gx.h

@@ -3,18 +3,18 @@
 #include <dolphin/gx.h>
 
 // Forward declare FIFO write functions
-namespace aurora::gfx::fifo {
+namespace aurora::gx::fifo {
 void write_u8(u8 val);
 void write_u16(u16 val);
 void write_u32(u32 val);
 void write_f32(f32 val);
 } // namespace aurora::gfx::fifo
 
 // FIFO write macros - route through software FIFO buffer
-#define GX_WRITE_U8(ub) aurora::gfx::fifo::write_u8(static_cast<u8>(ub))
-#define GX_WRITE_U16(us) aurora::gfx::fifo::write_u16(static_cast<u16>(us))
-#define GX_WRITE_U32(ui) aurora::gfx::fifo::write_u32(static_cast<u32>(ui))
-#define GX_WRITE_F32(f) aurora::gfx::fifo::write_f32(static_cast<f32>(f))
+#define GX_WRITE_U8(ub) aurora::gx::fifo::write_u8(static_cast<u8>(ub))
+#define GX_WRITE_U16(us) aurora::gx::fifo::write_u16(static_cast<u16>(us))
+#define GX_WRITE_U32(ui) aurora::gx::fifo::write_u32(static_cast<u32>(ui))
+#define GX_WRITE_F32(f) aurora::gx::fifo::write_f32(static_cast<f32>(f))
 
 // XF register write: opcode 0x10, address in XF space (0x1000+addr), then value
 #define GX_WRITE_XF_REG(addr, value)                                                                                   \

lib/dolphin/gx/gx.hpp

@@ -1,11 +1,11 @@
 #pragma once
 
 #include "../../internal.hpp"
-#include "../../gfx/gx.hpp"
+#include "../../gx/gx.hpp"
 
 static aurora::Module Log("aurora::gx");
 
-using aurora::gfx::gx::g_gxState;
+using aurora::gx::g_gxState;
 
 template <typename T>
 static inline void update_gx_state(T& val, T newVal) {