Fixed dolphin rendering.

cmake/targets/dolphin.cmake

@@ -16,6 +16,9 @@ target_compile_definitions(${DUSK_LIBRARY_TARGET_NAME} PUBLIC
   DOL=1
   ISO=2
   DUSK_DOLPHIN_BUILD_TYPE=${DUSK_DOLPHIN_BUILD_TYPE}
+  # GameCube/Wii PowerPC is always big-endian; declare it at compile time
+  # like every other target instead of relying on endian.h's runtime probe.
+  DUSK_PLATFORM_ENDIAN_BIG
 )
 
 # Custom compiler flags

src/dusk/asset/loader/chunk/assetchunkloader.c

@@ -115,6 +115,11 @@ errorret_t assetChunkLoaderSync(assetloading_t *loading) {
   memoryCopy(out->tiles, data + offset, tileSize);
   offset += tileSize;
 
+  for(size_t t = 0; t < CHUNK_TILE_COUNT; t++) {
+    uint32_t *shape = (uint32_t *)&out->tiles[t].shape;
+    *shape = endianLittleToHost32(*shape);
+  }
+
   out->meshCount = data[offset];
   offset += sizeof(uint8_t);
   assertTrue(
@@ -136,6 +141,9 @@ errorret_t assetChunkLoaderSync(assetloading_t *loading) {
 
     memoryCopy(out->meshOffsets[m], data + offset, sizeof(vec3));
     offset += sizeof(vec3);
+    out->meshOffsets[m][0] = endianLittleToHostFloat(out->meshOffsets[m][0]);
+    out->meshOffsets[m][1] = endianLittleToHostFloat(out->meshOffsets[m][1]);
+    out->meshOffsets[m][2] = endianLittleToHostFloat(out->meshOffsets[m][2]);
   }
 
   memoryFree(data);

src/dusk/asset/loader/dmf/assetmeshloader.c

@@ -48,8 +48,20 @@ errorret_t assetMeshLoaderAsync(assetloading_t *loading) {
   uint32_t vertCount = endianLittleToHost32(*(uint32_t *)(raw + 8));
   meshvertex_t *vertices = NULL;
   if(vertCount > 0) {
-    vertices = memoryAllocate(vertCount * sizeof(meshvertex_t));
+    // 32-byte (cache-line) aligned: GX_SetArray + DCFlushRange on Dolphin
+    // require this for the DMA'd vertex data to actually reach the GPU
+    // coherently. Static compiled-in vertex arrays happen to get this from
+    // the linker; a plain memoryAllocate here would not.
+    vertices = memoryAlign(32, vertCount * sizeof(meshvertex_t));
     memoryCopy(vertices, raw + 12, vertCount * sizeof(meshvertex_t));
+
+    for(uint32_t v = 0; v < vertCount; v++) {
+      vertices[v].uv[0] = endianLittleToHostFloat(vertices[v].uv[0]);
+      vertices[v].uv[1] = endianLittleToHostFloat(vertices[v].uv[1]);
+      vertices[v].pos[0] = endianLittleToHostFloat(vertices[v].pos[0]);
+      vertices[v].pos[1] = endianLittleToHostFloat(vertices[v].pos[1]);
+      vertices[v].pos[2] = endianLittleToHostFloat(vertices[v].pos[2]);
+    }
   }
   memoryFree(raw);
 
@@ -103,18 +115,22 @@ errorret_t assetMeshLoaderSync(assetloading_t *loading) {
     out->vertices = NULL;
     errorChain(ret);
   }
+  out->meshInitialized = true;
 
   ret = meshFlush(&out->mesh, 0, (int32_t)vertCount);
   if(errorIsNotOk(ret)) {
     loading->entry->state = ASSET_ENTRY_STATE_ERROR;
     meshDispose(&out->mesh);
+    out->meshInitialized = false;
     memoryFree(out->vertices);
     out->vertices = NULL;
     errorChain(ret);
   }
 
-  #ifndef DUSK_OPENGL_LEGACY
-    // VBO owns the data now; CPU copy is no longer needed.
+  #if defined(DUSK_OPENGL) && !defined(DUSK_OPENGL_LEGACY)
+    // VBO owns the data now; CPU copy is no longer needed. The platform
+    // mesh object itself still needs meshDispose later - tracked via
+    // out->meshInitialized, independent of the CPU buffer's lifetime.
     memoryFree(out->vertices);
     out->vertices = NULL;
   #endif
@@ -129,8 +145,11 @@ errorret_t assetMeshDispose(assetentry_t *entry) {
   assertIsMainThread("Must be called from the main thread.");
 
   assetmeshoutput_t *out = &entry->data.mesh;
-  if(out->vertices != NULL) {
+  if(out->meshInitialized) {
     errorChain(meshDispose(&out->mesh));
+    out->meshInitialized = false;
+  }
+  if(out->vertices != NULL) {
     memoryFree(out->vertices);
     out->vertices = NULL;
   }

src/dusk/asset/loader/dmf/assetmeshloader.h

@@ -30,6 +30,7 @@ typedef struct {
 
 typedef struct {
   mesh_t mesh;
+  bool_t meshInitialized;
   meshvertex_t *vertices;
 } assetmeshoutput_t;
 

src/duskdolphin/display/displaydolphin.c

@@ -78,6 +78,12 @@ errorret_t displayInitDolphin(void) {
   GX_SetDispCopyGamma(GX_GM_1_0);
   GX_SetColorUpdate(GX_TRUE);
 
+  // Without this, the EFB's Z-buffer format/compression is left at whatever
+  // GX_Init() defaulted to, so depth testing and the GX_MAX_Z24 clear value
+  // used every frame in frameBufferClearDolphin aren't guaranteed to line up
+  // with the actual EFB Z format.
+  GX_SetPixelFmt(GX_PF_RGB8_Z24, GX_ZC_LINEAR);
+
   // Describe mesh vertex format.
   GX_ClearVtxDesc();
   GX_SetVtxDesc(GX_VA_POS,  GX_INDEX16);
@@ -101,7 +107,7 @@ errorret_t displaySetStateDolphin(displaystate_t state) {
   }
 
   if(state.flags & DISPLAY_STATE_FLAG_DEPTH_TEST) {
-    GX_SetZMode(GX_TRUE, GX_LEQUAL, GX_TRUE);
+    GX_SetZMode(GX_TRUE, GX_LEQUAL, GX_FALSE);
   } else {
     GX_SetZMode(GX_FALSE, GX_ALWAYS, GX_FALSE);
   }