Merge pull request #130 from HackerN64/F3DEX3Memset

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);