More 3D Work

-=-=-=-=-=-

-ESM Shadow Mapping for softer and less glitchy shadows
-HDR Pipeline (convert to Linear on texture import, convert to SRGB at the end)
-Fix to xml parse bug

2 files changed, 204 insertions, 14 deletions

diff --git a/drivers/gles2/shaders/copy.glsl b/drivers/gles2/shaders/copy.glsl
index 2f1b349618..6d78c69e50 100644
--- a/drivers/gles2/shaders/copy.glsl
+++ b/drivers/gles2/shaders/copy.glsl
@@ -54,8 +54,12 @@ varying vec3 cube_interp;
 uniform samplerCube source_cube;
 #else
 varying vec2 uv_interp;
+#ifdef HIGHP_SOURCE
+uniform highp sampler2D source;
+#else
 uniform sampler2D source;
 #endif
+#endif
 varying vec2 uv2_interp;
 
 #ifdef USE_GLOW
@@ -83,12 +87,6 @@ uniform vec3 bcs;
 
 #endif
 
-#ifdef USE_GAMMA
-
-uniform float gamma;
-
-#endif
-
 #ifdef USE_GLOW_COPY
 
 uniform float bloom;
@@ -96,7 +94,7 @@ uniform float bloom_treshold;
 
 #endif
 
-#if defined(BLUR_V_PASS) || defined(BLUR_H_PASS) || defined(USE_HDR_REDUCE)
+#if defined(SHADOW_BLUR_V_PASS) || defined(SHADOW_BLUR_H_PASS) || defined(BLUR_V_PASS) || defined(BLUR_H_PASS) || defined(USE_HDR_REDUCE)
 
 uniform vec2 pixel_size;
 uniform float pixel_scale;
@@ -133,6 +131,17 @@ uniform float custom_alpha;
 void main() {
 
 	//vec4 color = color_interp;
+#ifdef USE_HIGHP_SOURCE
+
+#ifdef USE_CUBEMAP
+	highp vec4 color = textureCube( source_cube,  normalize(cube_interp) );
+
+#else
+	highp vec4 color = texture2D( source,  uv_interp );
+#endif
+
+#else
+
 #ifdef USE_CUBEMAP
 	vec4 color = textureCube( source_cube,  normalize(cube_interp) );
 
@@ -141,6 +150,8 @@ void main() {
 #endif
 
 
+#endif
+
 #ifdef USE_FXAA
 
 #define FXAA_REDUCE_MIN   (1.0/ 128.0)
@@ -226,17 +237,103 @@ void main() {
 
 #endif
 
+#ifdef SHADOW_BLUR_V_PASS
+
+#ifdef USE_RGBA_DEPTH
+
+#define VEC42DEPTH(m_vec4) dot(m_vec4,vec4(1.0 / (256.0 * 256.0 * 256.0),1.0 / (256.0 * 256.0),1.0 / 256.0,1))
+
+	highp float depth = VEC42DEPTH(color)*0.383;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(0.0,pixel_size.y*-3.0)*pixel_scale))*0.006;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(0.0,pixel_size.y*-2.0)*pixel_scale))*0.061;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(0.0,pixel_size.y*-1.0)*pixel_scale))*0.242;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(0.0,pixel_size.y*1.0)*pixel_scale))*0.242;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(0.0,pixel_size.y*2.0)*pixel_scale))*0.061;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(0.0,pixel_size.y*3.0)*pixel_scale))*0.006;
+	highp vec4 comp = fract(depth * vec4(256.0 * 256.0 * 256.0, 256.0 * 256.0, 256.0, 1.0));
+	comp -= comp.xxyz * vec4(0, 1.0 / 256.0, 1.0 / 256.0, 1.0 / 256.0);
+	color=comp;
+
+#else
+
+	highp float depth = color.r*0.383;
+	depth+=texture2D(source,uv_interp+vec2(0.0,pixel_size.y*-3.0)*pixel_scale).r*0.006;
+	depth+=texture2D(source,uv_interp+vec2(0.0,pixel_size.y*-2.0)*pixel_scale).r*0.061;
+	depth+=texture2D(source,uv_interp+vec2(0.0,pixel_size.y*-1.0)*pixel_scale).r*0.242;
+	depth+=texture2D(source,uv_interp+vec2(0.0,pixel_size.y*1.0)*pixel_scale).r*0.242;
+	depth+=texture2D(source,uv_interp+vec2(0.0,pixel_size.y*2.0)*pixel_scale).r*0.061;
+	depth+=texture2D(source,uv_interp+vec2(0.0,pixel_size.y*3.0)*pixel_scale).r*0.006;
+
+#ifdef USE_GLES_OVER_GL
+	gl_FragDepth = depth;
+
+#else
+	gl_FragDepthEXT = depth;
+#endif
+
+	return;
+#endif
+
+#endif
+
+#ifdef SHADOW_BLUR_H_PASS
+
+
+#ifdef USE_RGBA_DEPTH
+
+#define VEC42DEPTH(m_vec4) dot(m_vec4,vec4(1.0 / (256.0 * 256.0 * 256.0),1.0 / (256.0 * 256.0),1.0 / 256.0,1))
+
+	highp float depth = VEC42DEPTH(color)*0.383;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(pixel_size.x*-3.0,0.0)*pixel_scale))*0.006;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(pixel_size.x*-2.0,0.0)*pixel_scale))*0.061;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(pixel_size.x*-1.0,0.0)*pixel_scale))*0.242;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(pixel_size.x*1.0,0.0)*pixel_scale))*0.242;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(pixel_size.x*2.0,0.0)*pixel_scale))*0.061;
+	depth+=VEC42DEPTH(texture2D(source,uv_interp+vec2(pixel_size.x*3.0,0.0)*pixel_scale))*0.006;
+	highp vec4 comp = fract(depth * vec4(256.0 * 256.0 * 256.0, 256.0 * 256.0, 256.0, 1.0));
+	comp -= comp.xxyz * vec4(0, 1.0 / 256.0, 1.0 / 256.0, 1.0 / 256.0);
+	color=comp;
+#else
+
+
+	highp float depth = color.r*0.383;
+	depth+=texture2D(source,uv_interp+vec2(pixel_size.x*-3.0,0.0)*pixel_scale).r*0.006;
+	depth+=texture2D(source,uv_interp+vec2(pixel_size.x*-2.0,0.0)*pixel_scale).r*0.061;
+	depth+=texture2D(source,uv_interp+vec2(pixel_size.x*-1.0,0.0)*pixel_scale).r*0.242;
+	depth+=texture2D(source,uv_interp+vec2(pixel_size.x*1.0,0.0)*pixel_scale).r*0.242;
+	depth+=texture2D(source,uv_interp+vec2(pixel_size.x*2.0,0.0)*pixel_scale).r*0.061;
+	depth+=texture2D(source,uv_interp+vec2(pixel_size.x*3.0,0.0)*pixel_scale).r*0.006;
+
+#ifdef USE_GLES_OVER_GL
+	gl_FragDepth = depth;
+#else
+	gl_FragDepthEXT = depth;
+#endif
+
+	return;
+
+#endif
+
+#endif
+
 #ifdef USE_HDR
 
+
+#ifdef USE_8BIT_HDR
 	highp vec4 _mult = vec4(1.0 / (256.0 * 256.0 * 256.0),1.0 / (256.0 * 256.0),1.0 / 256.0,1);
 	highp float hdr_lum = dot(texture2D( hdr_source, vec2(0.0) ), _mult  );
 	color.rgb*=LUM_RANGE;
 	hdr_lum*=LUM_RANGE; //restore to full range
+#else
+	highp float hdr_lum = texture2D( hdr_source, vec2(0.0) ).r;
+#endif
+
 	highp float tone_scale = tonemap_exposure / hdr_lum; //only linear supported
 	color.rgb*=tone_scale;
 
 #endif
 
+
 #ifdef USE_GLOW_COPY
 
 	highp vec3 glowcol = color.rgb*color.a+step(bloom_treshold,dot(vec3(0.3333,0.3333,0.3333),color.rgb))*bloom*color.rgb;
@@ -281,10 +378,15 @@ void main() {
 
 #endif
 
-#ifdef USE_GAMMA
-
-	color.rgb = pow(color.rgb,gamma);
+#ifdef USE_SRGB
 
+	{ //i have my doubts about how fast this is
+		color.rgb = min(color.rgb,vec3(1.0)); //clamp just in case
+		vec3 S1 = sqrt(color.rgb);
+		vec3 S2 = sqrt(S1);
+		vec3 S3 = sqrt(S2);
+		color.rgb = 0.662002687 * S1 + 0.684122060 * S2 - 0.323583601 * S3 - 0.225411470 * color.rgb;
+	}
 #endif
 
 
@@ -292,13 +394,20 @@ void main() {
 
 	//highp float lum = dot(color.rgb,highp vec3(1.0/3.0,1.0/3.0,1.0/3.0));
 	highp float lum = max(color.r,max(color.g,color.b));
+#ifdef USE_8BIT_HDR
 	highp vec4 comp = fract(lum * vec4(256.0 * 256.0 * 256.0, 256.0 * 256.0, 256.0, 1.0));
 	comp -= comp.xxyz * vec4(0.0, 1.0 / 256.0, 1.0 / 256.0, 1.0 / 256.0);
 	color=comp;
+#else
+	color.rgb=vec3(lum);
+#endif
+
+
 #endif
 
 #ifdef USE_HDR_REDUCE
 
+#ifdef USE_8BIT_HDR
 	highp vec4 _multcv = vec4(1.0 / (256.0 * 256.0 * 256.0),1.0 / (256.0 * 256.0),1.0 / 256.0, 1.0);
 	highp float lum_accum = dot(color,_multcv  );
 	lum_accum += dot(texture2D( source,  uv_interp+vec2(-pixel_size.x,-pixel_size.y) ),_multcv  );
@@ -310,16 +419,42 @@ void main() {
 	lum_accum += dot(texture2D( source,  uv_interp+vec2(0.0,pixel_size.y) ),_multcv  );
 	lum_accum += dot(texture2D( source,  uv_interp+vec2(pixel_size.x,pixel_size.y) ),_multcv  );
 	lum_accum/=9.0;
+#else
+
+	highp float lum_accum = color.r;
+	lum_accum += texture2D( source,  uv_interp+vec2(-pixel_size.x,-pixel_size.y) ).r;
+	lum_accum += texture2D( source,  uv_interp+vec2(0.0,-pixel_size.y) ).r;
+	lum_accum += texture2D( source,  uv_interp+vec2(pixel_size.x,-pixel_size.y) ).r;
+	lum_accum += texture2D( source,  uv_interp+vec2(-pixel_size.x,0.0) ).r;
+	lum_accum += texture2D( source,  uv_interp+vec2(pixel_size.x,0.0) ).r;
+	lum_accum += texture2D( source,  uv_interp+vec2(-pixel_size.x,pixel_size.y) ).r;
+	lum_accum += texture2D( source,  uv_interp+vec2(0.0,pixel_size.y) ).r;
+	lum_accum += texture2D( source,  uv_interp+vec2(pixel_size.x,pixel_size.y) ).r;
+	lum_accum/=9.0;
+
+#endif
 
 #ifdef USE_HDR_STORE
 
+#ifdef USE_8BIT_HDR
 	highp float vd_lum = dot(texture2D( source_vd_lum, vec2(0.0) ), _multcv  );
 	lum_accum = clamp( vd_lum + (lum_accum-vd_lum)*hdr_time_delta*hdr_exp_adj_speed,min_luminance*(1.0/LUM_RANGE),max_luminance*(1.0/LUM_RANGE));
+#else
+	highp float vd_lum=texture2D( source_vd_lum, vec2(0.0) );
+	lum_accum = clamp( vd_lum + (lum_accum-vd_lum)*hdr_time_delta*hdr_exp_adj_speed,min_luminance,max_luminance);
+#endif
+
 #endif
 
+#ifdef USE_8BIT_HDR
 	highp vec4 comp = fract(lum_accum * vec4(256.0 * 256.0 * 256.0, 256.0 * 256.0, 256.0, 1.0));
 	comp -= comp.xxyz * vec4(0.0, 1.0 / 256.0, 1.0 / 256.0, 1.0 / 256.0);
 	color=comp;
+#else
+	color.rgb=vec3(lum_accum);
+#endif
+
+
 #endif
 
 #ifdef USE_RGBE
@@ -338,6 +473,8 @@ void main() {
 #ifdef USE_CUSTOM_ALPHA
 	color.a=custom_alpha;
 #endif
+
+
         gl_FragColor = color;
 }
 
diff --git a/drivers/gles2/shaders/material.glsl b/drivers/gles2/shaders/material.glsl
index 7ed59ae9cd..3aed9820ed 100644
--- a/drivers/gles2/shaders/material.glsl
+++ b/drivers/gles2/shaders/material.glsl
@@ -342,7 +342,7 @@ VERTEX_SHADER_CODE
 
 #ifdef USE_FOG
 
-	fog_interp.a = pow( clamp( (-vertex_interp.z-fog_params.x)/(fog_params.y-fog_params.x), 0.0, 1.0 ), fog_params.z );
+	fog_interp.a = pow( clamp( (length(vertex_interp)-fog_params.x)/(fog_params.y-fog_params.x), 0.0, 1.0 ), fog_params.z );
 	fog_interp.rgb = mix( fog_color_begin, fog_color_end, fog_interp.a );
 #endif
 
@@ -666,9 +666,14 @@ float SAMPLE_SHADOW_TEX( highp vec2 coord, highp float refdepth) {
 
 #ifdef USE_SHADOW_ESM
 
+uniform float esm_multiplier;
 
 float SAMPLE_SHADOW_TEX(vec2 p_uv,float p_depth) {
 
+#if defined (USE_DEPTH_SHADOWS)
+	//these only are used if interpolation exists
+	highp float occluder = SHADOW_DEPTH(shadow_texture, p_uv);
+#else
 	vec2 unnormalized = p_uv/shadow_texel_size;
 	vec2 fractional = fract(unnormalized);
 	unnormalized = floor(unnormalized);
@@ -681,7 +686,8 @@ float SAMPLE_SHADOW_TEX(vec2 p_uv,float p_depth) {
 
 	highp float occluder = (exponent.w + (exponent.x - exponent.w) * fractional.y);
 	occluder = occluder + ((exponent.z + (exponent.y - exponent.z) * fractional.y) - occluder)*fractional.x;
-	return clamp(exp(28.0 * ( occluder - p_depth )),0.0,1.0);
+#endif
+	return clamp(exp(esm_multiplier* ( occluder - p_depth )),0.0,1.0);
 
 }
 
@@ -818,7 +824,7 @@ FRAGMENT_SHADER_CODE
 		vec3 col_up=texture2D(ambient_octree_tex,octant_uv).rgb;
 		octant_uv.y+=ambient_octree_pix_size.y*2.0;
 		vec3 col_down=texture2D(ambient_octree_tex,octant_uv).rgb;
-		ambientmap_color=mix(col_down,col_up,1.0-sub.z);
+		ambientmap_color=mix(col_up,col_down,sub.z);
 
 		ambientmap_color*=diffuse.rgb;
 
@@ -866,6 +872,15 @@ FRAGMENT_SHADER_CODE
 	vec2 pssm_coord;
 	float pssm_z;
 
+#if defined(LIGHT_USE_PSSM) && defined(USE_SHADOW_ESM)
+#define USE_PSSM_BLEND
+	float pssm_blend;
+	vec2 pssm_coord_2;
+	float pssm_z_2;
+	vec3 light_pssm_split_inv = 1.0/light_pssm_split;
+	float w_inv = 1.0/gl_FragCoord.w;
+#endif
+
 #ifdef LIGHT_USE_PSSM4
 
 
@@ -874,10 +889,21 @@ FRAGMENT_SHADER_CODE
 		if (gl_FragCoord.w > light_pssm_split.x) {
 			pssm_coord=shadow_coord.xy;
 			pssm_z=shadow_coord.z;
+#if defined(USE_PSSM_BLEND)
+			pssm_coord_2=shadow_coord2.xy;
+			pssm_z_2=shadow_coord2.z;
+			pssm_blend=smoothstep(0.0,light_pssm_split_inv.x,w_inv);
+#endif
 
 		} else {
 			pssm_coord=shadow_coord2.xy;
 			pssm_z=shadow_coord2.z;
+#if defined(USE_PSSM_BLEND)
+			pssm_coord_2=shadow_coord3.xy;
+			pssm_z_2=shadow_coord3.z;
+			pssm_blend=smoothstep(light_pssm_split_inv.x,light_pssm_split_inv.y,w_inv);
+#endif
+
 		}
 	} else {
 
@@ -885,9 +911,21 @@ FRAGMENT_SHADER_CODE
 		if (gl_FragCoord.w > light_pssm_split.z) {
 			pssm_coord=shadow_coord3.xy;
 			pssm_z=shadow_coord3.z;
+#if defined(USE_PSSM_BLEND)
+			pssm_coord_2=shadow_coord4.xy;
+			pssm_z_2=shadow_coord4.z;
+			pssm_blend=smoothstep(light_pssm_split_inv.y,light_pssm_split_inv.z,w_inv);
+#endif
+
 		} else {
 			pssm_coord=shadow_coord4.xy;
 			pssm_z=shadow_coord4.z;
+#if defined(USE_PSSM_BLEND)
+			pssm_coord_2=shadow_coord4.xy;
+			pssm_z_2=shadow_coord4.z;
+			pssm_blend=0.0;
+#endif
+
 		}
 	}
 
@@ -896,16 +934,31 @@ FRAGMENT_SHADER_CODE
 	if (gl_FragCoord.w > light_pssm_split.x) {
 		pssm_coord=shadow_coord.xy;
 		pssm_z=shadow_coord.z;
+#if defined(USE_PSSM_BLEND)
+		pssm_coord_2=shadow_coord2.xy;
+		pssm_z_2=shadow_coord2.z;
+		pssm_blend=smoothstep(0.0,light_pssm_split_inv.x,w_inv);
+#endif
 
 	} else {
 		pssm_coord=shadow_coord2.xy;
 		pssm_z=shadow_coord2.z;
+#if defined(USE_PSSM_BLEND)
+		pssm_coord_2=shadow_coord2.xy;
+		pssm_z_2=shadow_coord2.z;
+		pssm_blend=0.0;
+#endif
+
 	}
 
 #endif
 
 	//one one sample
 	shadow_attenuation=SAMPLE_SHADOW_TEX(pssm_coord,pssm_z);
+#if defined(USE_PSSM_BLEND)
+	shadow_attenuation=mix(shadow_attenuation,SAMPLE_SHADOW_TEX(pssm_coord_2,pssm_z_2),pssm_blend);
+#endif
+
 
 #endif
 
@@ -1054,7 +1107,7 @@ FRAGMENT_SHADER_CODE
 	diffuse.a=glow;
 #endif
 
-#ifdef USE_HDR
+#ifdef USE_8BIT_HDR
 	diffuse.rgb*=0.25;
 #endif