stapip patch

engine/inc/renderer/3d/pipeline/static/core/bag/packaging/stapip_bag_packages_bbox.hpp

@@ -42,6 +42,8 @@ class StaPipBagPackagesBBox {
   /** @brief Get amount of vertices */
   const u32& getVertexCount() const;
 
+  const std::vector<CoreBBox>& getParts() const { return *bboxParts; }
+
   /**
    * @brief For example if we want to build a bbox from 96-112 vertices, we
    * should input index = 3, partsSize = 1

engine/inc/renderer/3d/pipeline/static/core/stapip_qbuffer_renderer.hpp

@@ -97,7 +97,6 @@ class StaPipQBufferRenderer {
   packet2_t** packets;
   StaPipVU1Program** dBufferPrograms;
   StaPipQBuffer** buffers;
-  packet2_t* currentPacket;
   packet2_t* staticDataPacket;
   packet2_t* objectDataPacket;
 

engine/inc/renderer/core/3d/bbox/core_bbox.hpp

@@ -31,6 +31,7 @@ class CoreBBox {
                     const u32& startIndex, const u32& stopIndex);
 
   const Vec4& operator[](const u8& i) { return _vertices[i]; }
+  const u8 getVertexCount() { return 8; }
   const Vec4& getVertex(const u8& i) { return _vertices[i]; }
   Vec4* getVertices() { return _vertices; }
 

engine/src/renderer/3d/pipeline/shared/vcl_sml.i

@@ -1381,8 +1381,6 @@ QSL3@:
 ;//---------------------------------------------------------
 ;// PerformClipCheck - checks if the vertex is outside the viewing frustum. 
 ;// If it is, then the appropriate clipping flags are set.
-;// Not using sqi instruction, because VCL cannot optimize it.
-;// Primtag contains information about how many polys we will send
 ;// 2 - Bitwise AND the clipping flags with 0x3FFFF, this makes sure that 
 ;//     we get the clipping judgement for the last three verts 
 ;//     (i.e. that make up the triangle we are about to draw)

engine/src/renderer/3d/pipeline/static/core/bag/packaging/stapip_bag_packager.cpp

@@ -106,10 +106,9 @@ CoreBBoxFrustum StaPipBagPackager::checkFrustum(const StaPipBagPackage& pkg) {
     auto& bbox = renderBBox->getChildBBox1By3(pkg.indexOf1By3BBox);
     return bbox.clipIsInFrustum(frustumPlanes->getAll(), *pkg.bag->info->model);
   } else {  // Is package
-    auto bbox = renderBBox->createChildBBox(
+    const auto& indexOfPart = pkg.indexOf1By3BBox;
-        pkg.indexOf1By3BBox,
+    auto partSize = ceil(pkg.size / static_cast<float>(maxVertCount / 3));
-        ceil(pkg.size / static_cast<float>(maxVertCount / 3)));
+    auto bbox = renderBBox->createChildBBox(indexOfPart, partSize);
-
     return bbox.clipIsInFrustum(frustumPlanes->getAll(), *pkg.bag->info->model);
   }
 }

engine/src/renderer/3d/pipeline/static/core/bag/packaging/stapip_bag_packages_bbox.cpp

@@ -72,7 +72,7 @@ const u32& StaPipBagPackagesBBox::getVertexCount() const { return vertexCount; }
 
 RenderBBox StaPipBagPackagesBBox::createChildBBox(const u32& index,
                                                   const u16& partsSize) const {
-  return RenderBBox(*bboxParts, index, partsSize);
+  return RenderBBox(*bboxParts, index, index + partsSize);
 }
 
 void StaPipBagPackagesBBox::print() const {

engine/src/renderer/3d/pipeline/static/core/programs/as_is/stapip_as_is_c_vu1.vclpp

@@ -90,6 +90,11 @@ processing:
         ScaleVertexToGSFormat{ scale, vertex3 }
         FixColor{ color3 }
 
+        iaddiu      adcBit,     VI00, 0x0
+        isw.w       adcBit,     XYZ2_STORE_OFFSET(destAddress)
+        isw.w       adcBit,     XYZ2_STORE_OFFSET+2(destAddress)
+        isw.w       adcBit,     XYZ2_STORE_OFFSET+4(destAddress)
+
         ;--- Store vertex1
         sq      color1,         RGBA_STORE_OFFSET(destAddress)
         sq.xyz  vertex1,        XYZ2_STORE_OFFSET(destAddress)

engine/src/renderer/3d/pipeline/static/core/programs/as_is/stapip_as_is_d_vu1.vclpp

@@ -76,6 +76,11 @@ vertexLoop:
         CalculateTyraDirectionalLights{ outputColor3, normal3, lightDirections, lightColors, lightMatrix, ambientColor }
         FixColor{ outputColor3 }
 
+        iaddiu      adcBit,     VI00, 0x0
+        isw.w       adcBit,     XYZ2_STORE_OFFSET(destAddress)
+        isw.w       adcBit,     XYZ2_STORE_OFFSET+2(destAddress)
+        isw.w       adcBit,     XYZ2_STORE_OFFSET+4(destAddress)
+
         ;--- Store vertex1
         sq      outputColor1,   RGBA_STORE_OFFSET(destAddress)
         sq.xyz  vertex1,        XYZ2_STORE_OFFSET(destAddress)

engine/src/renderer/3d/pipeline/static/core/programs/as_is/stapip_as_is_tc_vu1.vclpp

@@ -101,6 +101,11 @@ processing:
         PerformTexturePerspectiveCorrection{ outputStq3, stq3 }
         FixColor{ color3 }
 
+        iaddiu      adcBit,     VI00, 0x0
+        isw.w       adcBit,     XYZ2_STORE_OFFSET(destAddress)
+        isw.w       adcBit,     XYZ2_STORE_OFFSET+3(destAddress)
+        isw.w       adcBit,     XYZ2_STORE_OFFSET+6(destAddress)
+
         ;--- Store vertex1
         sq      outputStq1,     STQ_STORE_OFFSET(destAddress)
         sq      color1,         RGBA_STORE_OFFSET(destAddress)

engine/src/renderer/3d/pipeline/static/core/programs/as_is/stapip_as_is_td_vu1.vclpp

@@ -87,6 +87,11 @@ vertexLoop:
         CalculateTyraDirectionalLights{ outputColor3, normal3, lightDirections, lightColors, lightMatrix, ambientColor }
         FixColor{ outputColor3 }
 
+        iaddiu      adcBit,     VI00, 0x0
+        isw.w       adcBit,     XYZ2_STORE_OFFSET(destAddress)
+        isw.w       adcBit,     XYZ2_STORE_OFFSET+3(destAddress)
+        isw.w       adcBit,     XYZ2_STORE_OFFSET+6(destAddress)
+
         ;--- Store vertex1
         sq      outputStq1,     STQ_STORE_OFFSET(destAddress)
         sq      outputColor1,   RGBA_STORE_OFFSET(destAddress)

engine/src/renderer/3d/pipeline/static/core/stapip_qbuffer_renderer.cpp

@@ -11,7 +11,7 @@
 #include "renderer/3d/pipeline/static/core/stapip_qbuffer_renderer.hpp"
 #include "renderer/3d/pipeline/static/core/programs/stapip_vu1_shared_defines.h"
 
-#define TYRA_QBUFF_RENDERER_VERBOSE_LOG 1
+// #define TYRA_QBUFF_RENDERER_VERBOSE_LOG 1
 
 #ifdef TYRA_QBUFF_RENDERER_VERBOSE_LOG
 #define Verbose(...) Debug::writeLines("VRB: ", ##__VA_ARGS__, "\n")
@@ -272,11 +272,11 @@ void StaPipQBufferRenderer::cull(StaPipQBuffer* buffer) {
 
   dBufferPrograms[getQBufferIndex(buffer)] = getCullProgramByBag(buffer->bag);
 
+  Verbose("Add cull[", getQBufferIndex(buffer), "]: ", buffer->size);
+
   auto is1stDBuffer = is1stDBufferFlushTime();
   auto is2ndDBuffer = is2ndDBufferFlushTime();
 
-  Verbose("Add cull[", getQBufferIndex(buffer), "]: ", buffer->size);
-
   if (is1stDBuffer || is2ndDBuffer) {
     auto from = is1stDBuffer ? 0 : buffersCount / 2;
     auto to = from + buffersCount / 2;
@@ -321,7 +321,7 @@ void StaPipQBufferRenderer::clearLastProgramName() {
 
 void StaPipQBufferRenderer::addBuffersDataToPacket(const u32& from,
                                                    const u32& to) {
-  currentPacket = packets[context];
+  auto* currentPacket = packets[context];
   packet2_reset(currentPacket, false);
 
   for (u32 i = from; i < to; i++) {
@@ -346,6 +346,7 @@ void StaPipQBufferRenderer::addBuffersDataToPacket(const u32& from,
 }
 
 void StaPipQBufferRenderer::sendPacket() {
+  auto* currentPacket = packets[context];
   dma_channel_wait(DMA_CHANNEL_VIF1, 0);
   dma_channel_send_packet2(currentPacket, DMA_CHANNEL_VIF1, true);
 

engine/src/renderer/core/3d/bbox/render_bbox.cpp

@@ -46,6 +46,7 @@ CoreBBoxFrustum RenderBBox::clipIsInFrustum(const Plane* frustumPlanes,
 
   // Oh no, it probably needs clipping
 
+  // This is crappy guard band, but it works xd
   float guardBand[6];  // This probably needs more calibration
 
   guardBand[0] = -15.0F;  // Top