Fix on real hardware!

Aside from a silly mistake, the main issue why this was crashing on
real hardware before was because the 3DS GPU apparently does not like
rendering using one shader program with a geo component and one without
on the same frame. Making a dummy geoshader for the normal sprite case
appears to fix it (not sure what performance cost this may entail).

Author: TheGag96

source/base.c

@@ -1,5 +1,4 @@
 // citro2d's base.c, altered slightly to work with the geometry shader.
-// Look for to see meaningful changes: !! Changed for geo shader
 
 #include "internal.h"
 #include "render2d_shbin.h"
@@ -34,8 +33,10 @@ bool C2D_Init(size_t maxObjects)
 		switch (shaderId) {
 			case C2D_Normal:
 				vertsPerSprite = 6;
+				break;
 			case C2D_ScanlineOffset:
 				vertsPerSprite = 2;
+				break;
 		}
 
 		ctx->vtxBufSize = vertsPerSprite*maxObjects;
@@ -45,7 +46,7 @@ bool C2D_Init(size_t maxObjects)
 
 		switch (shaderId) {
 			case C2D_Normal:
-				ctx->shader = DVLB_ParseFile((u32*)render2g_shbin, render2d_shbin_size);
+				ctx->shader = DVLB_ParseFile((u32*)render2d_shbin, render2d_shbin_size);
 				break;
 			case C2D_ScanlineOffset:
 				ctx->shader = DVLB_ParseFile((u32*)render2g_shbin, render2g_shbin_size);
@@ -61,8 +62,13 @@ bool C2D_Init(size_t maxObjects)
 		shaderProgramInit(&ctx->program);
 		shaderProgramSetVsh(&ctx->program, &ctx->shader->DVLE[0]);
 
-		if (shaderId == C2D_ScanlineOffset) {
-			shaderProgramSetGsh(&ctx->program, &ctx->shader->DVLE[1], 4*vertsPerSprite);  //!! Changed for geo shader
+		switch (shaderId) {
+			case C2D_Normal:
+				shaderProgramSetGsh(&ctx->program, &ctx->shader->DVLE[1], 4*vertsPerSprite/2);
+				break;
+			case C2D_ScanlineOffset:
+				shaderProgramSetGsh(&ctx->program, &ctx->shader->DVLE[1], 4*vertsPerSprite);
+				break;
 		}
 
 		AttrInfo_Init(&ctx->attrInfo);
@@ -557,14 +563,9 @@ void C2Di_FlushVtxBuf(void)
 
 	GPU_Primitive_t primitive;
 
-	if (__C2Di_CurrentShader == C2D_ScanlineOffset) {
-		primitive = GPU_GEOMETRY_PRIM;
-	}
-	else {
-		primitive = GPU_TRIANGLES;
-	}
+	primitive = GPU_GEOMETRY_PRIM;
 
-	C3D_DrawArrays(primitive, ctx->vtxBufLastPos, len);  //!! Changed for geo shader
+	C3D_DrawArrays(primitive, ctx->vtxBufLastPos, len);
 	ctx->vtxBufLastPos = ctx->vtxBufPos;
 }
 

source/render2d.g.pica

@@ -0,0 +1,60 @@
+; Example PICA200 geometry shader
+.gsh point c0
+
+; Uniforms
+.fvec projection[4]
+
+; Constants
+
+; Outputs - this time the type *is* used
+.out oPosition  position
+.out oTexCoord0 texcoord0
+.out oTexCoord1 texcoord1
+.out oColor     color
+
+; Inputs: we will receive the following inputs:
+; v0-v3:  data of the first vertex
+; v4-v7:  data of the second vertex
+; v8-v11: data of the third vertex
+
+; Inputs
+.in iPosition_1      v0
+.in iTexCoord_1      v1
+.in iProcTexCoord_1  v2
+.in iColor_1         v3
+.in iPosition_2      v4
+.in iTexCoord_2      v5
+.in iProcTexCoord_2  v6
+.in iColor_2         v7
+.in iPosition_3      v8
+.in iTexCoord_3      v9
+.in iProcTexCoord_3  v10
+.in iColor_3         v11
+
+.entry render2d_gmain
+.proc render2d_gmain
+  setemit 0
+  mov oPosition,  iPosition_1
+  mov oTexCoord0, iTexCoord_1
+  mov oTexCoord1, iProcTexCoord_1
+  mov oColor,     iColor_1
+  emit
+
+  ; Emit the second vertex
+  setemit 1
+  mov oPosition,  iPosition_2
+  mov oTexCoord0, iTexCoord_2
+  mov oTexCoord1, iProcTexCoord_2
+  mov oColor,     iColor_2
+  emit
+
+  ; Emit the third vertex and finish the primitive
+  setemit 2, prim
+  mov oPosition,  iPosition_3
+  mov oTexCoord0, iTexCoord_3
+  mov oTexCoord1, iProcTexCoord_3
+  mov oColor,     iColor_3
+  emit
+
+  end
+.end