Improved random sample quality

source/main.cpp

@@ -18,6 +18,7 @@ struct vertex
 	float st[2];
 	float coords[3];
 	float uv[2];
+	float rand[3];
 };
 
 static DVLB_s* vshaderDVLB;
@@ -26,9 +27,10 @@ static s8 spheresUniformLocation;
 static s8 sphereColorsUniformLocation;
 static s8 sphereLightsUniformLocation;
 static s8 randUniformLocation;
+static s8 sampleRandUniformLocation;
 static s8 calcSphereLoopParamsUniformLocation;
 
-static constexpr unsigned int VERTEX_COUNT_W = 150;
+static constexpr unsigned int VERTEX_COUNT_W = 50;
 static constexpr unsigned int VERTEX_COUNT_H = 10 * VERTEX_COUNT_W / 6;
 static constexpr unsigned int VERTEX_COUNT = VERTEX_COUNT_W * VERTEX_COUNT_H;
 static constexpr unsigned int INDICES_COUNT = (VERTEX_COUNT * 2 + VERTEX_COUNT_H * 2);
@@ -135,6 +137,11 @@ static void setupVertices()
 
 			v.uv[0] = s * (240.0f / 256.0f);
 			v.uv[1] = t * (400.0f / 512.0f);
+
+			const C3D_FVec r = randvec();
+			v.rand[0] = r.x;
+			v.rand[1] = r.y;
+			v.rand[2] = r.z;
 		}
 	}
 
@@ -143,12 +150,19 @@ static void setupVertices()
 
 static void setupRandom()
 {
-	constexpr int NUM_RAND = 15; // needs to be the same as rand array length in vshader.v.pica
+	constexpr int NUM_RAND = 10; // needs to be the same as rand array length in vshader.v.pica
 	C3D_FVec* randPtr = C3D_FVUnifWritePtr(GPU_VERTEX_SHADER, randUniformLocation, NUM_RAND);
 	for (int i = 0; i < NUM_RAND; i++)
 	{
 		randPtr[i] = randvec();
 	}
+
+	constexpr int NUM_SAMPLE_RAND = 2;
+	C3D_FVec* sampleRandPtr = C3D_FVUnifWritePtr(GPU_VERTEX_SHADER, sampleRandUniformLocation, NUM_SAMPLE_RAND);
+	for (int i = 0; i < NUM_SAMPLE_RAND; i++)
+	{
+		sampleRandPtr[i] = randvec();
+	}
 }
 
 static void sceneInit()
@@ -178,6 +192,7 @@ static void sceneInit()
 	sphereColorsUniformLocation = shaderInstanceGetUniformLocation(program.vertexShader, "sphereColors");
 	sphereLightsUniformLocation = shaderInstanceGetUniformLocation(program.vertexShader, "sphereLights");
 	randUniformLocation = shaderInstanceGetUniformLocation(program.vertexShader, "rand");
+	sampleRandUniformLocation = shaderInstanceGetUniformLocation(program.vertexShader, "sampleRand");
 	calcSphereLoopParamsUniformLocation = shaderInstanceGetUniformLocation(program.vertexShader, "calcSphereLoopParams");
 
 	C3D_AttrInfo* attrInfo = C3D_GetAttrInfo();
@@ -189,13 +204,14 @@ static void sceneInit()
 	AttrInfo_AddLoader(attrInfo, 4, GPU_FLOAT, 2);
 	AttrInfo_AddLoader(attrInfo, 5, GPU_FLOAT, 3);
 	AttrInfo_AddLoader(attrInfo, 6, GPU_FLOAT, 2);
+	AttrInfo_AddLoader(attrInfo, 7, GPU_FLOAT, 3);
 
 	iboData = linearAlloc(sizeof(vertexIndices));
 	memcpy(iboData, vertexIndices, sizeof(vertexIndices));
 
 	C3D_BufInfo* bufInfo = C3D_GetBufInfo();
 	BufInfo_Init(bufInfo);
-	BufInfo_Add(bufInfo, vboData, sizeof(vertex), 3, 0x654);
+	BufInfo_Add(bufInfo, vboData, sizeof(vertex), 4, 0x7654);
 
 	C3D_TexEnv* env = C3D_GetTexEnv(0);
 	C3D_TexEnvInit(env);
@@ -203,7 +219,7 @@ static void sceneInit()
 	C3D_TexEnvFunc(env, C3D_RGB, GPU_INTERPOLATE);
 	C3D_TexEnvFunc(env, C3D_Alpha, GPU_REPLACE);
 
-	constexpr u16 NUM_SPHERES = 4;
+	constexpr u16 NUM_SPHERES = 5;
 	C3D_IVUnifSet(GPU_VERTEX_SHADER, calcSphereLoopParamsUniformLocation, NUM_SPHERES - 1, 0, 1, 0);
 
 	C3D_FVec* spheres = C3D_FVUnifWritePtr(GPU_VERTEX_SHADER, spheresUniformLocation, NUM_SPHERES);
@@ -211,18 +227,21 @@ static void sceneInit()
 	spheres[1] = FVec4_New(0.6f, 0.0f, 0.0f, 0.5f);
 	spheres[2] = FVec4_New(-0.6f, 0.0f, 0.0f, 0.5f);
 	spheres[3] = FVec4_New(0.0f, 10.0f, 0.0f, 9.0f);
+	spheres[4] = FVec4_New(0.0f, 0.4f, 0.0f, 0.2f);
 
 	C3D_FVec* sphereColors = C3D_FVUnifWritePtr(GPU_VERTEX_SHADER, sphereColorsUniformLocation, NUM_SPHERES);
 	sphereColors[0] = FVec4_New(0.9f, 0.3f, 0.3f,  0);
 	sphereColors[1] = FVec4_New(0.3f, 0.9f, 0.3f,  1);
 	sphereColors[2] = FVec4_New(0.3f, 0.3f, 0.9f,  0);
 	sphereColors[3] = FVec4_New(1.0f, 1.0f, 1.0f, -1);
+	sphereColors[4] = FVec4_New(0.8f, 0.6f, 0.1f,  2);
 
 	C3D_FVec* sphereLights = C3D_FVUnifWritePtr(GPU_VERTEX_SHADER, sphereLightsUniformLocation, NUM_SPHERES);
 	sphereLights[0] = FVec4_New(0.0f, 0.0f, 0.0f, 0);
 	sphereLights[1] = FVec4_New(0.0f, 0.0f, 0.0f, 0.2f);
 	sphereLights[2] = FVec4_New(0.0f, 0.0f, 0.0f, 0);
-	sphereLights[3] = FVec4_New(0.5f, 0.5f, 0.5f, 0);
+	sphereLights[3] = FVec4_New(2.5f, 2.5f, 2.5f, 0);
+	sphereLights[4] = FVec4_New(0.0f, 0.0f, 0.0f, 0);
 }
 
 static void sceneRender()
@@ -276,9 +295,9 @@ int main(int argc, char* argv[])
 
 	sceneInit();
 
-	C3D_FVec lookFrom = FVec3_New(0, 0.0f, 1.0f);
+	C3D_FVec lookFrom = FVec3_New(0.0f, 0.0f, 1.0f);
-	C3D_FVec lookAt = FVec3_New(0, 0, 0.0f);
+	C3D_FVec lookAt = FVec3_New(0.0f, 0.0f, 0.0f);
-	C3D_FVec vup = FVec3_New(0, 1.0f, 0);
+	C3D_FVec vup = FVec3_New(0.0f, 1.0f, 0.0f);
 	bool dirty = true;
 
 	unsigned int frameNum = 0;
@@ -346,10 +365,10 @@ int main(int argc, char* argv[])
 		setupVertices();
 		setupRandom();
 
-		C3D_TexEnvColor(C3D_GetTexEnv(0), getFrameNumColor(frameNum));
+		if (frameNum < 32 && C3D_FrameBegin(C3D_FRAME_SYNCDRAW))
-
-		if (frameNum < 96 && C3D_FrameBegin(C3D_FRAME_SYNCDRAW))
 		{
+			C3D_TexEnvColor(C3D_GetTexEnv(0), getFrameNumColor(frameNum));
+
 			setupCam(lookFrom, lookAt, vup);
 			C3D_TexBind(0, &prevFrame);
 			C3D_RenderTargetClear(target, C3D_CLEAR_ALL, 0, 0);
@@ -366,6 +385,8 @@ int main(int argc, char* argv[])
 
 			frameNum++;
 		}
+
+		C3D_TexEnvColor(C3D_GetTexEnv(0), 0);
 	}
 
 	sceneExit();

source/vshader.v.pica

@@ -1,5 +1,6 @@
 .constf myconst(0.0, 1.0, 0.001, 1000.0)
 .constf myconst2(0.5, 999.0, 1.1, 2.0)
+.constf myconst3(0.5, 0.0, 0.0, 0.0)
 .alias zeros myconst.xxxx
 .alias halfs myconst2.xxxx
 .alias ones myconst.yyyy
@@ -7,10 +8,14 @@
 .alias tooclose myconst.zzzz
 .alias far myconst.wwww
 .alias noHit myconst2.yyyy
+.alias samplesRCP myconst3.xxxx
+
 .alias diffuseID myconst.xxxx
 .alias metallicID myconst.yyyy
+.alias dielectricID myconst2.wwww
 
-.consti bounceLoopParams(14, 0, 1, 0)
+.consti sampleLoopParams(4, 0, 1, 0)
+.consti bounceLoopParams(9, 0, 1, 0)
 
 ; (NUM_SPHERES, 0, 1, 0)
 .ivec calcSphereLoopParams
@@ -31,7 +36,8 @@
 .fvec sphereLights[25]
 
 ; random numbers
-.fvec rand[15]
+.fvec rand[10]
+.fvec sampleRand[5]
 
 .in inOrigin v0
 .in inLowerLeftCorner v1
@@ -40,89 +46,109 @@
 .in inST v4
 .in inPos v5
 .in inUV v6
+.in inRand v7
 
 .out outPos position
 .out outUV texcoord0
 .out outColor color
 
 .proc main
-	; r1 = inOrigin
+	mov r15.xyz, zeros
-	mov r1, inOrigin
 
-	; r2 = inDirection
+	for sampleLoopParams
-	mov r2.xyz, inLowerLeftCorner
+		; r1 = inOrigin
-	mov r3.xy, inST.xy
+		mov r1, inOrigin
-	mad r2.xyz, inHorizontal, r3.x, r2.xyz
-	mad r2.xyz, inVertical, r3.y, r2.xyz
 
-	; set initial color multiplier to (1, 1, 1)
+		; r2 = inDirection
-	mov r4.xyz, ones
+		mov r2.xyz, inLowerLeftCorner
+		mov r3.xy, inST.xy
+		mad r2.xyz, inHorizontal, r3.x, r2.xyz
+		mad r2.xyz, inVertical, r3.y, r2.xyz
 
-	; set initial color to (0, 0, 0)
+		; set initial color multiplier to (1, 1, 1)
-	mov r13.xyz, zeros
+		mov r4.xyz, ones
 
-	; calculate light bounces
+		; set initial color to (0, 0, 0)
-	for bounceLoopParams
+		mov r13.xyz, zeros
-		; setup random numbers for this iteration
-		mov r11, rand[aL]
 
-		; reset max ray distance
+		mov r14.xyz, sampleRand[aL].xyz
-		mov r4.w, far
 
-		; set albedo to a large number for sphere hit check
+		; calculate light bounces
-		mov r10.xyz, far
+		for bounceLoopParams
+			; setup random numbers for this iteration
+			mov r11, rand[aL]
+			add r11.xyz, r11.xyz, r14.xyz
+			add r11.xyz, r11.xyz, inRand.xyz
+			dp3 r14.w, r11.xyz, r11.xyz
+			rsq r14.w, r14.w
+			mul r11.xyz, r11.xyz, r14.w
 
-		; for each sphere
+			; reset max ray distance
-		for calcSphereLoopParams
+			mov r4.w, far
-			; r3 = spheres[i]
-			mov r3, spheres[aL]
 
-			; do calculation
+			; set albedo to a large number for sphere hit check
-			call calcSphere
+			mov r10.xyz, far
+
+			; for each sphere
+			for calcSphereLoopParams
+				; r3 = spheres[i]
+				mov r3, spheres[aL]
+			
+				; do calculation
+				call calcSphere
+			.end
+
+			; check if noHit < albedo
+			; and exit early if true
+			; as albedo has not been set
+			; after the initial set
+			; which only happens when
+			; a ray does not hit any spheres
+			cmp noHit.xyz, lt, lt, r10.xyz
+			; not using breakc. it behaves weird.
+			jmpc cmp.x, labl
+
+			; multiply color by albedo
+			mul r4.xyz, r4.xyz, r10.xyz
+
+			; add emitted light
+			mad r13.xyz, r4.xyz, r12.xyz, r13.xyz
+
+			; set r1 to new ray origin
+			mov r1.xyz, r5.xyz
+
+			; set r2 to new ray direction
+			; try diffuse material
+			cmp diffuseID.w, eq, eq, r10.w
+			callc cmp.x, diffuse
+			jmpc cmp.x, materialFound
+
+			; try metallic material
+			cmp metallicID.w, eq, eq, r10.w
+			callc cmp.x, metallic
+			jmpc cmp.x, materialFound
+
+			; try dielectric material
+			cmp dielectricID.w, eq, eq, r10.w
+			callc cmp.x, dielectric
+			jmpc cmp.x, materialFound
+
+			; material doesn't reflect light
+			; used mostly for light-emitting objects
+			jmpu true, labl
+
+			materialFound:
+			nop
 		.end
 
-		; check if noHit < albedo
+		labl:
-		; and exit early if true
-		; as albedo has not been set
-		; after the initial set
-		; which only happens when
-		; a ray does not hit any spheres
-		cmp noHit.xyz, lt, lt, r10.xyz
-		; not using breakc. it behaves weird.
-		jmpc cmp.x, labl
 
-		; multiply color by albedo
+		; add final color to output
-		mul r4.xyz, r4.xyz, r10.xyz
+		add r15.xyz, r13.xyz, r15.xyz
-
-		; add emitted light
-		mad r13.xyz, r4.xyz, r12.xyz, r13.xyz
-
-		; set r1 to new ray origin
-		mov r1.xyz, r5.xyz
-
-		; set r2 to new ray direction
-		; try diffuse material
-		cmp diffuseID.w, eq, eq, r10.w
-		callc cmp.x, diffuse
-		jmpc cmp.x, materialFound
-
-		; try metallic material
-		cmp metallicID.w, eq, eq, r10.w
-		callc cmp.x, metallic
-		jmpc cmp.x, materialFound
-
-		; material doesn't reflect light
-		; used mostly for light-emitting objects
-		jmpu true, labl
-
-		materialFound:
-		nop
 	.end
 
-	labl:
+	; divide by sample count
-
+	mul outColor.xyz, samplesRCP, r15.xyz
-	; copy final color to output
-	mov outColor.xyz, r13.xyz
 
 	; set alpha to 1
 	mov outColor.w, ones
@@ -133,6 +159,7 @@
 	end
 .end
 
+; -----------------------------
 ; Calculate Sphere Intersection
 ; -----------------------------
 ;   Performs a ray-sphere intersection test.
@@ -159,6 +186,7 @@
 ; -----------
 ;   r5.xyz:   new origin
 ;   r7.xyz:   hit normal
+;   r7.w:     ray is outside sphere (1) or inside sphere (0)
 ;   r8.xyzw:  used for calculations
 ;   r9.xyzw:  used for calculations
 
@@ -253,8 +281,9 @@
 	nop
 .end
 
-; All Materials
+; ---------
-; -------------
+; Materials
+; ---------
 ;   Calculates the new ray direction for a given material.
 ;
 ; Inputs
@@ -274,12 +303,14 @@
 
 ; Diffuse Material
 ; ----------------
+
 .proc diffuse
 	add r2.xyz, r7.xyz, r11.xyz
 .end
 
 ; Metallic Material
 ; -----------------
+
 .proc metallic
 	dp3 r6.xyz, r2, r7
 	mul r6.xyz, twos, r6.xyz
@@ -287,3 +318,11 @@
 	; add a bit of random "fuzziness" to the metal
 	mad r2.xyz, r11.xyz, r12.w, r2.xyz
 .end
+
+; Dielectric Material
+; -------------------
+
+.proc dielectric
+	; currently diffuse material, implement fr later
+	add r2.xyz, r7.xyz, r11.xyz
+.end