Use F3DEX3 memset to clear Z buffer for RDP time savings

include/functions.h

@@ -1086,6 +1086,7 @@ Gfx* Gfx_TwoTexScrollEnvColor(GraphicsContext* gfxCtx, s32 tile1, u32 x1, u32 y1
                               u32 x2, u32 y2, s32 width2, s32 height2, s32 r, s32 g, s32 b, s32 a);
 Gfx* Gfx_EnvColor(GraphicsContext* gfxCtx, s32 r, s32 g, s32 b, s32 a);
 void Gfx_SetupFrame(GraphicsContext* gfxCtx, s32 clearFB, u8 r, u8 g, u8 b);
+void Gfx_ClearZBuffer(GraphicsContext* gfxCtx);
 void func_80095974(GraphicsContext* gfxCtx);
 void func_80095AA0(PlayState* play, Room* room, Input* input, s32 arg3);
 void Room_DrawBackground2D(Gfx** gfxP, void* tex, void* tlut, u16 width, u16 height, u8 fmt, u8 siz, u16 tlutMode,

include/ultra64/gbi.f3dex3.h

@@ -44,6 +44,7 @@ of warnings if you use -Wpedantic. */
 /*#define G_SPECIAL_3       0xD3  no-op in F3DEX2 */
 /*#define G_SPECIAL_2       0xD4  no-op in F3DEX2 */
 /*#define G_SPECIAL_1       0xD5  triggered MVP recalculation, not supported in F3DEX3 */
+#define G_MEMSET            0xD5
 #define G_DMA_IO            0xD6
 #define G_TEXTURE           0xD7
 #define G_POPMTX            0xD8
@@ -2385,6 +2386,25 @@ _DW({                                               \
 #define gSPDmaWrite(pkt,dmem,dram,size) gSPDma_io((pkt),1,(dmem),(dram),(size))
 #define gsSPDmaWrite(dmem,dram,size)    gsSPDma_io(     1,(dmem),(dram),(size))
 
+/**
+ * Use RSP DMAs to set a region of memory to a repeated 16-bit value. This can
+ * clear the color framebuffer or Z-buffer faster than the RDP can in fill mode.
+ * SPMemset overwrites the DMEM vertex buffer, so vertices loaded before this
+ * command cannot be used after it (though this would not normally be done).
+ * 
+ * dram: Segmented or physical start address. Must be aligned to 16 bytes.
+ * value: 16-bit value to fill the memory with. e.g. 0 for color, 0xFFFC for Z.
+ * size: Size in bytes to fill, must be nonzero and a multiple of 16 bytes.
+ */
+#define gSPMemset(pkt, dram, value, size)               \
+_DW({                                                   \
+    gImmp1(pkt, G_RDPHALF_1, ((value) & 0xFFFF));       \
+    gDma0p(pkt, G_MEMSET, (dram), ((size) & 0xFFFFF0)); \
+})
+
+#define gsSPMemset(pkt, dram, value, size)    \
+    gsImmp1(G_RDPHALF_1, ((value) & 0xFFFF)), \
+    gsDma0p(G_MEMSET, (dram), ((size) & 0xFFFFF0))
 
 /*
  * RSP short command (no DMA required) macros

include/variables.h

@@ -223,13 +223,7 @@ extern volatile s8 gLoadedF3DEX3Version;
 extern volatile s8 gF3DEX3ProfVersion;
 extern volatile s8 gF3DEX3NOCVersion;
 extern s8 gF3DEX3OccMode;
-#endif
-
-#if ENABLE_F3DEX3
-extern u8 gF3DEX3TextBuffer[];
-extern volatile s8 gF3DEX3ProfVersion;
-extern volatile s8 gF3DEX3NOCVersion;
-extern s8 gF3DEX3OccMode;
+extern u8 gUseMemsetForZBuffer;
 #endif
 
 extern SfxBankEntry D_8016BAD0[9];

src/code/z_play.c

@@ -1258,7 +1258,7 @@ void Play_Draw(PlayState* this) {
             clearG = this->lightCtx.fogColor[1];
             clearB = this->lightCtx.fogColor[2];
         }
-        // Clear the fb only if we aren't drawing a skybox, but always clear zb
+        // Clear the fb only if we aren't drawing a skybox
         Gfx_SetupFrame(gfxCtx, clearFB, clearR, clearG, clearB);
     }
 
@@ -1378,6 +1378,15 @@ void Play_Draw(PlayState* this) {
             Environment_DrawSkyboxFilters(this);
         }
 
+        // The Z buffer has to be cleared at some point before anything using it
+        // is drawn (lighting strike is the first which does). But if we are
+        // using F3DEX3's SPMemset to clear it, it should be done as late as
+        // possible, after the RSP has already sent commands to the RDP for the
+        // skybox or framebuffer clear. This is so that the RSP can clear the Z
+        // buffer while the RDP is working on the framebuffer, without making
+        // the RDP wait for new work to be available.
+        Gfx_ClearZBuffer(gfxCtx);
+
         if (!IS_DEBUG || (R_HREG_MODE != HREG_MODE_PLAY) || (R_PLAY_DRAW_ENV_FLAGS & PLAY_ENV_DRAW_LIGHTNING)) {
             Environment_UpdateLightningStrike(this);
             Environment_DrawLightning(this, 0);

src/code/z_prenmi.c

@@ -34,6 +34,7 @@ void PreNMI_Draw(PreNMIState* this) {
 
     gSPSegment(POLY_OPA_DISP++, 0x00, NULL);
     Gfx_SetupFrame(gfxCtx, true, 0, 0, 0);
+    Gfx_ClearZBuffer(gfxCtx);
     Gfx_SetupDL_36Opa(gfxCtx);
     gDPSetFillColor(POLY_OPA_DISP++, (GPACK_RGBA5551(255, 255, 255, 1) << 16) | GPACK_RGBA5551(255, 255, 255, 1));
     gDPFillRectangle(POLY_OPA_DISP++, 0, this->timer + 100, SCREEN_WIDTH - 1, this->timer + 100);

src/code/z_rcp.c

@@ -1,5 +1,9 @@
 #include "global.h"
 
+#if ENABLE_F3DEX3
+u8 gUseMemsetForZBuffer = 1;
+#endif
+
 Gfx sSetupDL[SETUPDL_MAX][6] = {
     {
         /* SETUPDL_0 */
@@ -1472,7 +1476,6 @@ void Gfx_SetupFrame(GraphicsContext* gfxCtx, s32 clearFB, u8 r, u8 g, u8 b) {
 
     // Set up the framebuffer, primitives will be drawn here
     gDPSetColorImage(POLY_OPA_DISP++, G_IM_FMT_RGBA, G_IM_SIZ_16b, gScreenWidth, gfxCtx->curFrameBuffer);
-    gDPSetColorImage(POLY_OPA_DISP++, G_IM_FMT_RGBA, G_IM_SIZ_16b, gScreenWidth, gfxCtx->curFrameBuffer);
     gDPSetColorImage(POLY_XLU_DISP++, G_IM_FMT_RGBA, G_IM_SIZ_16b, gScreenWidth, gfxCtx->curFrameBuffer);
     gDPSetColorImage(OVERLAY_DISP++, G_IM_FMT_RGBA, G_IM_SIZ_16b, gScreenWidth, gfxCtx->curFrameBuffer);
 
@@ -1529,17 +1532,8 @@ void Gfx_SetupFrame(GraphicsContext* gfxCtx, s32 clearFB, u8 r, u8 g, u8 b) {
         }
 #endif
 
-        // Set the whole z buffer to maximum depth
-        // Don't bother with pixels that are being covered by the letterbox
-        gDPSetColorImage(POLY_OPA_DISP++, G_IM_FMT_RGBA, G_IM_SIZ_16b, gScreenWidth, gZBuffer);
-        gDPSetRenderMode(POLY_OPA_DISP++, G_RM_NOOP, G_RM_NOOP2);
-        gDPSetFillColor(POLY_OPA_DISP++, (GPACK_ZDZ(G_MAXFBZ, 0) << 16) | GPACK_ZDZ(G_MAXFBZ, 0));
-        gDPFillRectangle(POLY_OPA_DISP++, 0, letterboxSize, gScreenWidth - 1, gScreenHeight - letterboxSize - 1);
-        gDPPipeSync(POLY_OPA_DISP++);
-
         // Fill the whole screen with the base color, only done when there is no skybox or if it is a solid color.
         // Don't bother with pixels that are being covered by the letterbox
-        gDPSetColorImage(POLY_OPA_DISP++, G_IM_FMT_RGBA, G_IM_SIZ_16b, gScreenWidth, gfxCtx->curFrameBuffer);
         if (clearFB) {
             gDPSetRenderMode(POLY_OPA_DISP++, G_RM_NOOP, G_RM_NOOP2);
             gDPSetFillColor(POLY_OPA_DISP++, (GPACK_RGBA5551(r, g, b, 1) << 16) | GPACK_RGBA5551(r, g, b, 1));
@@ -1560,6 +1554,36 @@ void Gfx_SetupFrame(GraphicsContext* gfxCtx, s32 clearFB, u8 r, u8 g, u8 b) {
     CLOSE_DISPS(gfxCtx, "../z_rcp.c", 2497);
 }
 
+void Gfx_ClearZBuffer(GraphicsContext* gfxCtx) {
+    s32 letterboxSize = Letterbox_GetSize();
+    OPEN_DISPS(gfxCtx, "../z_rcp.c", __LINE__);
+
+    // Set the whole z buffer to maximum depth
+    // Don't bother with pixels that are being covered by the letterbox
+#if ENABLE_F3DEX3
+    if (gUseMemsetForZBuffer) {
+        s32 w2 = gScreenWidth * 2; // 2 bytes per pixel
+        if (letterboxSize < 0 || letterboxSize > 100) {
+            letterboxSize = 0;
+        }
+        gSPMemset(POLY_OPA_DISP++, (u8*)gZBuffer + letterboxSize * w2, GPACK_ZDZ(G_MAXFBZ, 0),
+                  (gScreenHeight - 2 * letterboxSize) * w2);
+    } else {
+#endif
+        gSPDisplayList(POLY_OPA_DISP++, sFillSetupDL);
+        gDPSetColorImage(POLY_OPA_DISP++, G_IM_FMT_RGBA, G_IM_SIZ_16b, gScreenWidth, gZBuffer);
+        gDPSetRenderMode(POLY_OPA_DISP++, G_RM_NOOP, G_RM_NOOP2);
+        gDPSetFillColor(POLY_OPA_DISP++, (GPACK_ZDZ(G_MAXFBZ, 0) << 16) | GPACK_ZDZ(G_MAXFBZ, 0));
+        gDPFillRectangle(POLY_OPA_DISP++, 0, letterboxSize, gScreenWidth - 1, gScreenHeight - letterboxSize - 1);
+        gDPPipeSync(POLY_OPA_DISP++);
+        gDPSetColorImage(POLY_OPA_DISP++, G_IM_FMT_RGBA, G_IM_SIZ_16b, gScreenWidth, gfxCtx->curFrameBuffer);
+#if ENABLE_F3DEX3
+    }
+#endif
+
+    CLOSE_DISPS(gfxCtx, "../z_rcp.c", __LINE__);
+}
+
 void func_80095974(GraphicsContext* gfxCtx) {
     OPEN_DISPS(gfxCtx, "../z_rcp.c", 2503);
 

src/code/z_sample.c

@@ -17,6 +17,7 @@ void Sample_Draw(SampleState* this) {
     gSPSegment(POLY_OPA_DISP++, 0x01, this->staticSegment);
 
     Gfx_SetupFrame(gfxCtx, true, 0, 0, 0);
+    Gfx_ClearZBuffer(gfxCtx);
 
     view->flags = VIEW_VIEWING | VIEW_VIEWPORT | VIEW_PROJECTION_PERSPECTIVE;
     View_Apply(view, VIEW_ALL);

src/overlays/gamestates/ovl_file_choose/z_file_choose.c

@@ -69,6 +69,7 @@ void FileSelect_InitModeDraw(GameState* thisx) {
     FileSelectState* this = (FileSelectState*)thisx;
 
     Gfx_SetupFrame(this->state.gfxCtx, true, 0, 0, 0);
+    Gfx_ClearZBuffer(this->state.gfxCtx);
 }
 
 /**
@@ -1136,6 +1137,7 @@ void FileSelect_ConfigModeDraw(GameState* thisx) {
     Gfx_SetupFrame(this->state.gfxCtx, false, 0, 0, 0);
     Skybox_Draw(&this->skyboxCtx, this->state.gfxCtx, NULL, SKYBOX_NORMAL_SKY, this->envCtx.skyboxBlend, eyeX, eyeY,
                 eyeZ);
+    Gfx_ClearZBuffer(this->state.gfxCtx);
     gDPSetTextureLUT(POLY_OPA_DISP++, G_TT_NONE);
     ZREG(11) += ZREG(10);
     Environment_UpdateSkybox(SKYBOX_NORMAL_SKY, &this->envCtx, &this->skyboxCtx);
@@ -1551,6 +1553,7 @@ void FileSelect_SelectModeDraw(GameState* thisx) {
     Gfx_SetupFrame(this->state.gfxCtx, false, 0, 0, 0);
     Skybox_Draw(&this->skyboxCtx, this->state.gfxCtx, NULL, SKYBOX_NORMAL_SKY, this->envCtx.skyboxBlend, eyeX, eyeY,
                 eyeZ);
+    Gfx_ClearZBuffer(this->state.gfxCtx);
     gDPSetTextureLUT(POLY_OPA_DISP++, G_TT_NONE);
     ZREG(11) += ZREG(10);
     Environment_UpdateSkybox(SKYBOX_NORMAL_SKY, &this->envCtx, &this->skyboxCtx);

src/overlays/gamestates/ovl_opening/z_opening.c

@@ -69,6 +69,7 @@ void TitleSetup_Main(GameState* thisx) {
     TitleSetupState* this = (TitleSetupState*)thisx;
 
     Gfx_SetupFrame(this->state.gfxCtx, true, 0, 0, 0);
+    Gfx_ClearZBuffer(this->state.gfxCtx);
     TitleSetup_SetupTitleScreen(this);
     func_80803C5C(this);
 }

src/overlays/gamestates/ovl_select/z_select.c

@@ -92,6 +92,7 @@ void MapSelect_Draw(MapSelectState* this) {
     Gfx_SetupFrame(gfxCtx, true, 0, 0, 0);
     SET_FULLSCREEN_VIEWPORT(&this->view);
     View_Apply(&this->view, VIEW_ALL);
+    Gfx_SetupDL_28Opa(gfxCtx);
 
     if (!this->state.running) {
         MapSelect_DrawLoadingScreen(this);
@@ -333,12 +334,6 @@ void MapSelect_DrawMenu(MapSelectState* this) {
 
     OPEN_DISPS(gfxCtx, __FILE__, __LINE__);
 
-    gSPSegment(POLY_OPA_DISP++, 0x00, NULL);
-    Gfx_SetupFrame(gfxCtx, true, 0, 0, 0);
-    SET_FULLSCREEN_VIEWPORT(&this->view);
-    View_Apply(&this->view, VIEW_ALL);
-    Gfx_SetupDL_28Opa(gfxCtx);
-
     printer = alloca(sizeof(GfxPrint));
     GfxPrint_Init(printer);
     GfxPrint_Open(printer, POLY_OPA_DISP);
@@ -366,12 +361,6 @@ void MapSelect_DrawLoadingScreen(MapSelectState* this) {
 
     OPEN_DISPS(gfxCtx, __FILE__, __LINE__);
 
-    gSPSegment(POLY_OPA_DISP++, 0x00, NULL);
-    Gfx_SetupFrame(gfxCtx, true, 0, 0, 0);
-    SET_FULLSCREEN_VIEWPORT(&this->view);
-    View_Apply(&this->view, VIEW_ALL);
-    Gfx_SetupDL_28Opa(gfxCtx);
-
     printer = alloca(sizeof(GfxPrint));
     GfxPrint_Init(printer);
     GfxPrint_Open(printer, POLY_OPA_DISP);

src/overlays/gamestates/ovl_title/z_title.c

@@ -155,6 +155,7 @@ void ConsoleLogo_Main(GameState* thisx) {
     gSPSegment(POLY_OPA_DISP++, 0, NULL);
     gSPSegment(POLY_OPA_DISP++, 1, this->staticSegment);
     Gfx_SetupFrame(this->state.gfxCtx, true, 0, 0, 0);
+    Gfx_ClearZBuffer(this->state.gfxCtx);
     ConsoleLogo_Calc(this);
     ConsoleLogo_Draw(this);