Merge branch 'slow-startup' into 'master'

Optimize PNG detection and generate blendtables on parallel thread

See merge request KartKrew/Kart!498

src/d_main.c

@@ -1417,6 +1417,9 @@ void D_SRB2Main(void)
 	// setup loading screen
 	SCR_Startup();
 
+	// Do this in background; lots of number crunching
+	R_InitTranslucencyTables();
+
 	CON_SetLoadingProgress(LOADED_ISTARTUPGRAPHICS);
 
 	CONS_Printf("HU_Init()...\n");

src/r_draw.c

@@ -26,6 +26,7 @@
 #include "z_zone.h"
 #include "console.h" // Until buffering gets finished
 #include "k_color.h" // SRB2kart
+#include "i_threads.h"
 
 #ifdef HWRENDER
 #include "hardware/hw_main.h"
@@ -192,6 +193,29 @@ CV_PossibleValue_t Followercolor_cons_t[MAXSKINCOLORS+3];	// +3 to account for "
 
 #define TRANSTAB_AMTMUL10 (255.0f / 10.0f)
 
+static void R_AllocateBlendTables(void)
+{
+	INT32 i;
+
+	for (i = 0; i < NUMBLENDMAPS; i++)
+	{
+		if (i == blendtab_modulate)
+			continue;
+		blendtables[i] = Z_MallocAlign((NUMTRANSTABLES + 1) * 0x10000, PU_STATIC, NULL, 16);
+	}
+
+	// Modulation blending only requires a single table
+	blendtables[blendtab_modulate] = Z_MallocAlign(0x10000, PU_STATIC, NULL, 16);
+}
+
+#ifdef HAVE_THREADS
+static void R_GenerateBlendTables_Thread(void *userdata)
+{
+	(void)userdata;
+	R_GenerateBlendTables();
+}
+#endif
+
 /** \brief Initializes the translucency tables used by the Software renderer.
 */
 void R_InitTranslucencyTables(void)
@@ -212,20 +236,20 @@ void R_InitTranslucencyTables(void)
 	W_ReadLump(W_GetNumForName("TRANS80"), transtables+0x70000);
 	W_ReadLump(W_GetNumForName("TRANS90"), transtables+0x80000);
 
+	R_AllocateBlendTables();
+
+#ifdef HAVE_THREADS
+	I_spawn_thread("blend-tables",
+			R_GenerateBlendTables_Thread, NULL);
+#else
 	R_GenerateBlendTables();
+#endif
 }
 
 void R_GenerateBlendTables(void)
 {
 	INT32 i;
 
-	for (i = 0; i < NUMBLENDMAPS; i++)
-	{
-		if (i == blendtab_modulate)
-			continue;
-		blendtables[i] = Z_MallocAlign((NUMTRANSTABLES + 1) * 0x10000, PU_STATIC, NULL, 16);
-	}
-
 	for (i = 0; i <= 9; i++)
 	{
 		const size_t offs = (0x10000 * i);
@@ -236,8 +260,6 @@ void R_GenerateBlendTables(void)
 		R_GenerateTranslucencyTable(blendtables[blendtab_reversesubtract] + offs, AST_REVERSESUBTRACT, alpha);
 	}
 
-	// Modulation blending only requires a single table
-	blendtables[blendtab_modulate] = Z_MallocAlign(0x10000, PU_STATIC, NULL, 16);
 	R_GenerateTranslucencyTable(blendtables[blendtab_modulate], AST_MODULATE, 0);
 }
 

src/r_main.c

@@ -1137,7 +1137,7 @@ void R_Init(void)
 	R_InitLightTables();
 
 	//I_OutputMsg("\nR_InitTranslucencyTables\n");
-	R_InitTranslucencyTables();
+	//R_InitTranslucencyTables();
 
 	R_InitDrawNodes();
 

src/r_picformats.h

@@ -105,6 +105,7 @@ typedef struct
 } spriteinfo_t;
 
 // Portable Network Graphics
+#define PNG_HEADER_SIZE (8)
 boolean Picture_IsLumpPNG(const UINT8 *d, size_t s);
 #define Picture_ThrowPNGError(lumpname, wadfilename) I_Error("W_Wad: Lump \"%s\" in file \"%s\" is a .png - please convert to either Doom or Flat (raw) image format.", lumpname, wadfilename); // Fears Of LJ Sonic
 

src/r_textures.c

@@ -725,6 +725,7 @@ Rloadflats (INT32 i, INT32 w)
 	UINT16 texstart, texend;
 	texture_t *texture;
 	texpatch_t *patch;
+	UINT8 header[PNG_HEADER_SIZE];
 
 	// Yes
 	if (wadfiles[w]->type == RET_PK3)
@@ -743,7 +744,6 @@ Rloadflats (INT32 i, INT32 w)
 		// Work through each lump between the markers in the WAD.
 		for (j = 0; j < (texend - texstart); j++)
 		{
-			UINT8 *flatlump;
 			UINT16 wadnum = (UINT16)w;
 			lumpnum_t lumpnum = texstart + j;
 			size_t lumplength;
@@ -755,7 +755,7 @@ Rloadflats (INT32 i, INT32 w)
 					continue; // If it is then SKIP IT
 			}
 
-			flatlump = W_CacheLumpNumPwad(wadnum, lumpnum, PU_CACHE);
+			W_ReadLumpHeaderPwad(wadnum, lumpnum, header, sizeof header, 0);
 			lumplength = W_LumpLengthPwad(wadnum, lumpnum);
 
 			switch (lumplength)
@@ -790,12 +790,14 @@ Rloadflats (INT32 i, INT32 w)
 			M_Memcpy(texture->name, W_CheckNameForNumPwad(wadnum, lumpnum), sizeof(texture->name));
 
 #ifndef NO_PNG_LUMPS
-			if (Picture_IsLumpPNG((UINT8 *)flatlump, lumplength))
+			if (Picture_IsLumpPNG(header, lumplength))
 			{
+				UINT8 *flatlump = W_CacheLumpNumPwad(wadnum, lumpnum, PU_CACHE);
 				INT32 width, height;
 				Picture_PNGDimensions((UINT8 *)flatlump, &width, &height, NULL, NULL, lumplength);
 				texture->width = (INT16)width;
 				texture->height = (INT16)height;
+				Z_Free(flatlump);
 			}
 			else
 #endif
@@ -814,8 +816,6 @@ Rloadflats (INT32 i, INT32 w)
 			patch->lump = texstart + j;
 			patch->flip = 0;
 
-			Z_Free(flatlump);
-
 			texturewidth[i] = texture->width;
 			textureheight[i] = texture->height << FRACBITS;
 			i++;
@@ -835,8 +835,8 @@ Rloadtextures (INT32 i, INT32 w)
 	UINT16 j;
 	UINT16 texstart, texend, texturesLumpPos;
 	texture_t *texture;
-	softwarepatch_t *patchlump;
 	texpatch_t *patch;
+	softwarepatch_t patchlump;
 
 	// Get the lump numbers for the markers in the WAD, if they exist.
 	if (wadfiles[w]->type == RET_PK3)
@@ -876,7 +876,7 @@ Rloadtextures (INT32 i, INT32 w)
 					continue; // If it is then SKIP IT
 			}
 
-			patchlump = W_CacheLumpNumPwad(wadnum, lumpnum, PU_CACHE);
+			W_ReadLumpHeaderPwad(wadnum, lumpnum, &patchlump, PNG_HEADER_SIZE, 0);
 #ifndef NO_PNG_LUMPS
 			lumplength = W_LumpLengthPwad(wadnum, lumpnum);
 #endif
@@ -888,18 +888,20 @@ Rloadtextures (INT32 i, INT32 w)
 			M_Memcpy(texture->name, W_CheckNameForNumPwad(wadnum, lumpnum), sizeof(texture->name));
 
 #ifndef NO_PNG_LUMPS
-			if (Picture_IsLumpPNG((UINT8 *)patchlump, lumplength))
+			if (Picture_IsLumpPNG((UINT8 *)&patchlump, lumplength))
 			{
+				UINT8 *png = W_CacheLumpNumPwad(wadnum, lumpnum, PU_CACHE);
 				INT32 width, height;
-				Picture_PNGDimensions((UINT8 *)patchlump, &width, &height, NULL, NULL, lumplength);
+				Picture_PNGDimensions(png, &width, &height, NULL, NULL, lumplength);
 				texture->width = (INT16)width;
 				texture->height = (INT16)height;
+				Z_Free(png);
 			}
 			else
 #endif
 			{
-				texture->width = SHORT(patchlump->width);
+				texture->width = SHORT(patchlump.width);
-				texture->height = SHORT(patchlump->height);
+				texture->height = SHORT(patchlump.height);
 			}
 
 			texture->type = TEXTURETYPE_SINGLEPATCH;
@@ -915,8 +917,6 @@ Rloadtextures (INT32 i, INT32 w)
 			patch->lump = texstart + j;
 			patch->flip = 0;
 
-			Z_Free(patchlump);
-
 			texturewidth[i] = texture->width;
 			textureheight[i] = texture->height << FRACBITS;
 			i++;

src/r_things.c

@@ -286,16 +286,18 @@ boolean R_AddSingleSpriteDef(const char *sprname, spritedef_t *spritedef, UINT16
 
 #ifndef NO_PNG_LUMPS
 			{
-				softwarepatch_t *png = W_CacheLumpNumPwad(wadnum, l, PU_STATIC);
+				UINT8 header[PNG_HEADER_SIZE];
 				size_t len = W_LumpLengthPwad(wadnum, l);
 
-				if (Picture_IsLumpPNG((UINT8 *)png, len))
+				W_ReadLumpHeaderPwad(wadnum, l, header, sizeof header, 0);
+
+				if (Picture_IsLumpPNG(header, len))
 				{
+					UINT8 *png = W_CacheLumpNumPwad(wadnum, l, PU_STATIC);
 					Picture_PNGDimensions((UINT8 *)png, &width, &height, &topoffset, &leftoffset, len);
 					isPNG = true;
+					Z_Free(png);
 				}
-
-				Z_Free(png);
 			}
 
 			if (!isPNG)

src/w_wad.c

@@ -1480,10 +1480,10 @@ size_t W_ReadLumpHeaderPwad(UINT16 wad, UINT16 lump, void *dest, size_t size, si
 			int zErr; // Helper var.
 			z_stream strm;
 			unsigned long rawSize = l->disksize;
-			unsigned long decSize = l->size;
+			unsigned long decSize = size;
 
 			rawData = Z_Malloc(rawSize, PU_STATIC, NULL);
-			decData = Z_Malloc(decSize, PU_STATIC, NULL);
+			decData = dest;
 
 			if (fread(rawData, 1, rawSize, handle) < rawSize)
 				I_Error("wad %d, lump %d: cannot read compressed data", wad, lump);
@@ -1501,12 +1501,8 @@ size_t W_ReadLumpHeaderPwad(UINT16 wad, UINT16 lump, void *dest, size_t size, si
 			zErr = inflateInit2(&strm, -15);
 			if (zErr == Z_OK)
 			{
-				zErr = inflate(&strm, Z_FINISH);
+				zErr = inflate(&strm, Z_SYNC_FLUSH);
-				if (zErr == Z_STREAM_END)
+				if (zErr != Z_OK && zErr != Z_STREAM_END)
-				{
-					M_Memcpy(dest, decData, size);
-				}
-				else
 				{
 					size = 0;
 					zerr(zErr);
@@ -1520,7 +1516,6 @@ size_t W_ReadLumpHeaderPwad(UINT16 wad, UINT16 lump, void *dest, size_t size, si
 			}
 
 			Z_Free(rawData);
-			Z_Free(decData);
 
 #ifdef NO_PNG_LUMPS
 			if (Picture_IsLumpPNG((UINT8 *)dest, size))