Eevee: SSR: Add mipmap filtering and bias to reduce noise.

Also fix the roughness factors.

7 files changed, 126 insertions, 25 deletions

diff --git a/source/blender/draw/CMakeLists.txt b/source/blender/draw/CMakeLists.txt
index 3ebdea06093..4098cfe696e 100644
--- a/source/blender/draw/CMakeLists.txt
+++ b/source/blender/draw/CMakeLists.txt
@@ -138,6 +138,7 @@ data_to_c_simple(engines/eevee/shaders/effect_bloom_frag.glsl SRC)
 data_to_c_simple(engines/eevee/shaders/effect_dof_vert.glsl SRC)
 data_to_c_simple(engines/eevee/shaders/effect_dof_geom.glsl SRC)
 data_to_c_simple(engines/eevee/shaders/effect_dof_frag.glsl SRC)
+data_to_c_simple(engines/eevee/shaders/effect_downsample_frag.glsl SRC)
 data_to_c_simple(engines/eevee/shaders/effect_motion_blur_frag.glsl SRC)
 data_to_c_simple(engines/eevee/shaders/effect_ssr_frag.glsl SRC)
 data_to_c_simple(engines/eevee/shaders/lightprobe_planar_downsample_frag.glsl SRC)
diff --git a/source/blender/draw/engines/eevee/eevee_effects.c b/source/blender/draw/engines/eevee/eevee_effects.c
index 6414ef8ff32..e4c4cb5d790 100644
--- a/source/blender/draw/engines/eevee/eevee_effects.c
+++ b/source/blender/draw/engines/eevee/eevee_effects.c
@@ -84,7 +84,10 @@ static struct {
 	struct GPUShader *ssr_raytrace_sh;
 	struct GPUShader *ssr_resolve_sh;
 
+	struct GPUShader *downsample_sh;
+
 	struct GPUTexture *depth_src;
+	struct GPUTexture *color_src;
 
 	float pixelprojmat[4][4];
 } e_data = {NULL}; /* Engine data */
@@ -99,6 +102,7 @@ extern char datatoc_effect_bloom_frag_glsl[];
 extern char datatoc_effect_dof_vert_glsl[];
 extern char datatoc_effect_dof_geom_glsl[];
 extern char datatoc_effect_dof_frag_glsl[];
+extern char datatoc_effect_downsample_frag_glsl[];
 extern char datatoc_lightprobe_lib_glsl[];
 extern char datatoc_raytrace_lib_glsl[];
 extern char datatoc_tonemap_frag_glsl[];
@@ -192,6 +196,8 @@ void EEVEE_effects_init(EEVEE_SceneLayerData *sldata, EEVEE_Data *vedata)
 
 		MEM_freeN(ssr_shader_str);
 
+		e_data.downsample_sh = DRW_shader_create_fullscreen(datatoc_effect_downsample_frag_glsl, NULL);
+
 		e_data.volumetric_upsample_sh = DRW_shader_create_fullscreen(datatoc_volumetric_frag_glsl, "#define STEP_UPSAMPLE\n");
 
 		e_data.minmaxz_downlevel_sh = DRW_shader_create_fullscreen(datatoc_effect_minmaxz_frag_glsl, NULL);
@@ -578,7 +584,7 @@ void EEVEE_effects_init(EEVEE_SceneLayerData *sldata, EEVEE_Data *vedata)
 
 	/* Setup double buffer so we can access last frame as it was before post processes */
 	if ((effects->enabled_effects & EFFECT_DOUBLE_BUFFER) != 0) {
-		DRWFboTexture tex_double_buffer = {&txl->color_double_buffer, DRW_TEX_RGB_11_11_10, DRW_TEX_FILTER};
+		DRWFboTexture tex_double_buffer = {&txl->color_double_buffer, DRW_TEX_RGB_11_11_10, DRW_TEX_FILTER | DRW_TEX_MIPMAP};
 
 		DRW_framebuffer_init(&fbl->double_buffer, &draw_engine_eevee_type,
 		                    (int)viewport_size[0], (int)viewport_size[1],
@@ -711,6 +717,13 @@ void EEVEE_effects_cache_init(EEVEE_SceneLayerData *sldata, EEVEE_Data *vedata)
 	}
 
 	{
+		psl->color_downsample_ps = DRW_pass_create("Downsample", DRW_STATE_WRITE_COLOR);
+		DRWShadingGroup *grp = DRW_shgroup_create(e_data.downsample_sh, psl->color_downsample_ps);
+		DRW_shgroup_uniform_buffer(grp, "source", &e_data.color_src);
+		DRW_shgroup_call_add(grp, quad, NULL);
+	}
+
+	{
 		psl->minmaxz_downlevel = DRW_pass_create("HiZ Down Level", DRW_STATE_WRITE_COLOR);
 		DRWShadingGroup *grp = DRW_shgroup_create(e_data.minmaxz_downlevel_sh, psl->minmaxz_downlevel);
 		DRW_shgroup_uniform_buffer(grp, "depthBuffer", &stl->g_data->minmaxz);
@@ -833,6 +846,12 @@ static void minmax_downsample_cb(void *vedata, int UNUSED(level))
 	DRW_draw_pass(psl->minmaxz_downlevel);
 }
 
+static void simple_downsample_cb(void *vedata, int UNUSED(level))
+{
+	EEVEE_PassList *psl = ((EEVEE_Data *)vedata)->psl;
+	DRW_draw_pass(psl->color_downsample_ps);
+}
+
 void EEVEE_create_minmax_buffer(EEVEE_Data *vedata, GPUTexture *depth_src)
 {
 	EEVEE_PassList *psl = vedata->psl;
@@ -851,6 +870,17 @@ void EEVEE_create_minmax_buffer(EEVEE_Data *vedata, GPUTexture *depth_src)
 	DRW_framebuffer_recursive_downsample(fbl->minmaxz_fb, stl->g_data->minmaxz, 6, &minmax_downsample_cb, vedata);
 }
 
+/**
+ * Simple downsampling algorithm. Reconstruct mip chain up to mip level.
+ **/
+void EEVEE_downsample_buffer(EEVEE_Data *vedata, struct GPUFrameBuffer *fb_src, GPUTexture *texture_src, int level)
+{
+	e_data.color_src = texture_src;
+
+	/* Create lower levels */
+	DRW_framebuffer_recursive_downsample(fb_src, texture_src, level, &simple_downsample_cb, vedata);
+}
+
 void EEVEE_effects_do_volumetrics(EEVEE_SceneLayerData *sldata, EEVEE_Data *vedata)
 {
 	EEVEE_PassList *psl = vedata->psl;
@@ -902,6 +932,8 @@ void EEVEE_effects_do_ssr(EEVEE_SceneLayerData *UNUSED(sldata), EEVEE_Data *veda
 	if ((effects->enabled_effects & EFFECT_SSR) != 0 && stl->g_data->valid_double_buffer) {
 		DefaultTextureList *dtxl = DRW_viewport_texture_list_get();
 
+		EEVEE_downsample_buffer(vedata, fbl->minmaxz_fb, txl->color_double_buffer, 5);
+
 		/* Raytrace at halfres. */
 		e_data.depth_src = dtxl->depth;
 		// e_data.depth_src = stl->g_data->minmaxz;
@@ -1078,6 +1110,7 @@ void EEVEE_draw_effects(EEVEE_Data *vedata)
 
 void EEVEE_effects_free(void)
 {
+	DRW_SHADER_FREE_SAFE(e_data.downsample_sh);
 	DRW_SHADER_FREE_SAFE(e_data.ssr_raytrace_sh);
 	DRW_SHADER_FREE_SAFE(e_data.ssr_resolve_sh);
 
diff --git a/source/blender/draw/engines/eevee/eevee_engine.c b/source/blender/draw/engines/eevee/eevee_engine.c
index 59f2613dd06..4ee3dee655f 100644
--- a/source/blender/draw/engines/eevee/eevee_engine.c
+++ b/source/blender/draw/engines/eevee/eevee_engine.c
@@ -56,7 +56,7 @@ static void EEVEE_engine_init(void *ved)
 	stl->g_data->background_alpha = 1.0f;
 	stl->g_data->valid_double_buffer = (txl->color_double_buffer != NULL);
 
-	DRWFboTexture tex = {&txl->color, DRW_TEX_RGB_11_11_10, DRW_TEX_FILTER};
+	DRWFboTexture tex = {&txl->color, DRW_TEX_RGB_11_11_10, DRW_TEX_FILTER | DRW_TEX_MIPMAP};
 
 	const float *viewport_size = DRW_viewport_size_get();
 	DRW_framebuffer_init(&fbl->main, &draw_engine_eevee_type,
diff --git a/source/blender/draw/engines/eevee/eevee_private.h b/source/blender/draw/engines/eevee/eevee_private.h
index 015bc5b6a9e..f31f4f9fe6c 100644
--- a/source/blender/draw/engines/eevee/eevee_private.h
+++ b/source/blender/draw/engines/eevee/eevee_private.h
@@ -111,6 +111,7 @@ typedef struct EEVEE_PassList {
 	struct DRWPass *volumetric_resolve_transmit_ps;
 	struct DRWPass *ssr_raytrace;
 	struct DRWPass *ssr_resolve;
+	struct DRWPass *color_downsample_ps;
 
 	struct DRWPass *depth_pass;
 	struct DRWPass *depth_pass_cull;
@@ -504,6 +505,7 @@ void EEVEE_lightprobes_free(void);
 void EEVEE_effects_init(EEVEE_SceneLayerData *sldata, EEVEE_Data *vedata);
 void EEVEE_effects_cache_init(EEVEE_SceneLayerData *sldata, EEVEE_Data *vedata);
 void EEVEE_create_minmax_buffer(EEVEE_Data *vedata, struct GPUTexture *depth_src);
+void EEVEE_downsample_buffer(EEVEE_Data *vedata, struct GPUFrameBuffer *fb_src, struct GPUTexture *texture_src, int level);
 void EEVEE_effects_do_volumetrics(EEVEE_SceneLayerData *sldata, EEVEE_Data *vedata);
 void EEVEE_effects_do_ssr(EEVEE_SceneLayerData *sldata, EEVEE_Data *vedata);
 void EEVEE_draw_effects(EEVEE_Data *vedata);
diff --git a/source/blender/draw/engines/eevee/shaders/bsdf_common_lib.glsl b/source/blender/draw/engines/eevee/shaders/bsdf_common_lib.glsl
index 059f0598156..e6cbcde77c7 100644
--- a/source/blender/draw/engines/eevee/shaders/bsdf_common_lib.glsl
+++ b/source/blender/draw/engines/eevee/shaders/bsdf_common_lib.glsl
@@ -374,6 +374,22 @@ void accumulate_light(vec3 light, float fac, inout vec4 accum)
 	accum += vec4(light, 1.0) * min(fac, (1.0 - accum.a));
 }
 
+/* ----------- Cone Apperture Approximation --------- */
+
+/* Return a fitted cone angle given the input roughness */
+float cone_cosine(float r)
+{
+	/* Using phong gloss
+	 * roughness = sqrt(2/(gloss+2)) */
+	float gloss = -2 + 2 / (r * r);
+	/* Drobot 2014 in GPUPro5 */
+	// return cos(2.0 * sqrt(2.0 / (gloss + 2)));
+	/* Uludag 2014 in GPUPro5 */
+	// return pow(0.244, 1 / (gloss + 1));
+	/* Jimenez 2016 in Practical Realtime Strategies for Accurate Indirect Occlusion*/
+	return exp2(-3.32193 * r * r);
+}
+
 /* --------- Closure ---------- */
 #ifdef VOLUMETRICS
 
diff --git a/source/blender/draw/engines/eevee/shaders/effect_downsample_frag.glsl b/source/blender/draw/engines/eevee/shaders/effect_downsample_frag.glsl
new file mode 100644
index 00000000000..4a79fa908b1
--- /dev/null
+++ b/source/blender/draw/engines/eevee/shaders/effect_downsample_frag.glsl
@@ -0,0 +1,15 @@
+/**
+ * Simple downsample shader. Takes the average of the 4 texels of lower mip.
+ **/
+
+uniform sampler2D source;
+
+out vec4 FragColor;
+
+void main()
+{
+	/* Reconstructing Target uvs like this avoid missing pixels if NPO2 */
+	vec2 uvs = gl_FragCoord.xy * 2.0 / vec2(textureSize(source, 0));
+
+	FragColor = texture(source, uvs);
+}
\ No newline at end of file
diff --git a/source/blender/draw/engines/eevee/shaders/effect_ssr_frag.glsl b/source/blender/draw/engines/eevee/shaders/effect_ssr_frag.glsl
index 7337ec792c9..4a1cea1338e 100644
--- a/source/blender/draw/engines/eevee/shaders/effect_ssr_frag.glsl
+++ b/source/blender/draw/engines/eevee/shaders/effect_ssr_frag.glsl
@@ -4,17 +4,25 @@
 uniform sampler2DArray utilTex;
 #endif /* UTIL_TEX */
 
-vec3 generate_ray(ivec2 pix, vec3 V, vec3 N, float roughnessSquared, out float pdf)
+#define BRDF_BIAS 0.7
+
+vec3 generate_ray(ivec2 pix, vec3 V, vec3 N, float a2, out float pdf)
 {
 	float NH;
 	vec3 T, B;
 	make_orthonormal_basis(N, T, B); /* Generate tangent space */
 	vec3 rand = texelFetch(utilTex, ivec3(pix % LUT_SIZE, 2), 0).rba;
-	vec3 H = sample_ggx(rand, roughnessSquared, N, T, B, NH); /* Microfacet normal */
-	pdf = max(32e32, pdf_ggx_reflect(NH, roughnessSquared)); /* Theoretical limit of 10bit float (not in practice?) */
+
+	/* Importance sampling bias */
+	rand.x = mix(rand.x, 0.0, BRDF_BIAS);
+
+	vec3 H = sample_ggx(rand, a2, N, T, B, NH); /* Microfacet normal */
+	pdf = min(1024e32, pdf_ggx_reflect(NH, a2)); /* Theoretical limit of 16bit float */
 	return reflect(-V, H);
 }
 
+#define MAX_MIP 5.0
+
 #ifdef STEP_RAYTRACE
 
 uniform sampler2D depthBuffer;
@@ -44,11 +52,12 @@ void main()
 	/* Retrieve pixel data */
 	vec4 speccol_roughness = texelFetch(specroughBuffer, fullres_texel, 0).rgba;
 	float roughness = speccol_roughness.a;
-	float roughnessSquared = roughness * roughness;
+	float roughnessSquared = max(1e-3, roughness * roughness);
+	float a2 = roughnessSquared * roughnessSquared;
 
 	/* Generate Ray */
 	float pdf;
-	vec3 R = generate_ray(halfres_texel, V, N, roughnessSquared, pdf);
+	vec3 R = generate_ray(halfres_texel, V, N, a2, pdf);
 
 	/* Search for the planar reflection affecting this pixel */
 	/* If no planar is found, fallback to screen space */
@@ -65,7 +74,11 @@ void main()
 	//float hit_dist = raycast(depthBuffer, W, R);
 
 	/* Raycast over screen */
-	float hit_dist = raycast(depthBuffer, viewPosition, R);
+	float hit_dist = -1.0;
+	if (dot(R, N) > 0.0001) {
+		/* Only raytrace if ray is above the surface normal */
+		hit_dist = raycast(depthBuffer, viewPosition, R);
+	}
 
 	vec2 hit_co = project_point(ProjectionMatrix, viewPosition + R * hit_dist).xy * 0.5 + 0.5;
 
@@ -164,7 +177,7 @@ float screen_border_mask(vec2 past_hit_co, vec3 hit)
 	hit_co.zw = past_hit_co;
 
 	const float margin = 0.002;
-	const float atten = 0.05 + margin; /* Screen percentage */
+	const float atten = 0.1 + margin; /* Screen percentage */
 	hit_co = smoothstep(margin, atten, hit_co) * (1 - smoothstep(1.0 - atten, 1.0 - margin, hit_co));
 	vec2 atten_fac = min(hit_co.xy, hit_co.zw);
 
@@ -179,7 +192,9 @@ void main()
 {
 	ivec2 halfres_texel = ivec2(gl_FragCoord.xy / 2.0);
 	ivec2 fullres_texel = ivec2(gl_FragCoord.xy);
-	vec2 uvs = gl_FragCoord.xy / vec2(textureSize(depthBuffer, 0));
+	vec2 texture_size = vec2(textureSize(depthBuffer, 0));
+	vec2 uvs = gl_FragCoord.xy / texture_size;
+	vec3 rand = texelFetch(utilTex, ivec3(fullres_texel % LUT_SIZE, 2), 0).rba;
 
 	float depth = textureLod(depthBuffer, uvs, 0.0).r;
 
@@ -194,48 +209,67 @@ void main()
 	vec3 N = mat3(ViewMatrixInverse) * normal_decode(texelFetch(normalBuffer, fullres_texel, 0).rg, viewCameraVec);
 	vec4 speccol_roughness = texelFetch(specroughBuffer, fullres_texel, 0).rgba;
 	float roughness = speccol_roughness.a;
-	float roughnessSquared = roughness * roughness;
+	float roughnessSquared = max(1e-3, roughness * roughness);
 
 	vec4 spec_accum = vec4(0.0);
 
-	/* Resolve SSR and compute contribution */
+	/* Resolve SSR */
+	float cone_cos = cone_cosine(roughness);
+	float cone_tan = sqrt(1 - cone_cos * cone_cos) / cone_cos;
+	cone_tan *= mix(saturate(dot(N, V) * 2.0), 1.0, sqrt(roughness)); /* Elongation fit */
 
-	/* We generate the same rays that has been generated in the raycast step.
-	 * But we add this ray from our resolve pixel position, increassing accuracy. */
 	vec3 ssr_accum = vec3(0.0);
 	float weight_acc = 0.0;
-	float mask = 0.0;
-	const ivec2 neighbors[7] = ivec2[7](ivec2(0, 0), ivec2(2, 1), ivec2(-2, 1), ivec2(0, -2), ivec2(-2, -1), ivec2(2, -1), ivec2(0, 2));
-	int invert_neighbor = ((((fullres_texel.x & 0x1) + (fullres_texel.y & 0x1)) & 0x1) == 0) ? 1 : -1;
+	float mask_acc = 0.0;
+	float dist_acc = 0.0;
+	float hit_acc = 0.0;
+	const ivec2 neighbors[4] = ivec2[4](ivec2(0, 0), ivec2(1, 1), ivec2(0, 1), ivec2(1, 0));
+	ivec2 invert_neighbor;
+	invert_neighbor.x = ((fullres_texel.x & 0x1) == 0) ? 1 : -1;
+	invert_neighbor.y = ((fullres_texel.y & 0x1) == 0) ? 1 : -1;
 	for (int i = 0; i < NUM_NEIGHBORS; i++) {
 		ivec2 target_texel = halfres_texel + neighbors[i] * invert_neighbor;
 
 		float pdf = texelFetch(pdfBuffer, target_texel, 0).r;
 
 		/* Check if there was a hit */
-		if (pdf != 0.0) {
+		if (pdf > 0.001) {
 			vec2 hit_co = texelFetch(hitBuffer, target_texel, 0).rg;
 
 			/* Reconstruct ray */
 			float hit_depth = textureLod(depthBuffer, hit_co, 0.0).r;
 			vec3 hit_pos = get_world_space_from_depth(hit_co, hit_depth);
 
+			/* Evaluate BSDF */
+			vec3 L = normalize(hit_pos - worldPosition);
+			float bsdf = bsdf_ggx(N, L, V, roughnessSquared);
+
 			/* Find hit position in previous frame */
 			vec2 ref_uvs = get_reprojected_reflection(hit_pos, worldPosition, N);
+			vec2 source_uvs = project_point(PastViewProjectionMatrix, worldPosition).xy * 0.5 + 0.5;
 
-			/* Evaluate BSDF */
-			vec3 L = normalize(hit_pos - worldPosition);
-			float bsdf = bsdf_ggx(N, L, V, max(1e-3, roughness));
+			/* Estimate a cone footprint to sample a corresponding mipmap level */
+			/* compute cone footprint Using UV distance because we are using screen space filtering */
+			float cone_footprint = cone_tan * distance(ref_uvs, source_uvs);
+			float mip = BRDF_BIAS * clamp(log2(cone_footprint * max(texture_size.x, texture_size.y)), 0.0, MAX_MIP);
 
-			float weight = bsdf / max(1e-8, pdf);
-			mask += screen_border_mask(ref_uvs, hit_pos);
-			ssr_accum += textureLod(colorBuffer, ref_uvs, 0.0).rgb * weight;
+			float border_mask = screen_border_mask(ref_uvs, hit_pos);
+			float weight = border_mask * bsdf / max(1e-8, pdf);
+			ssr_accum += textureLod(colorBuffer, ref_uvs, mip).rgb * weight;
 			weight_acc += weight;
+			dist_acc += distance(hit_pos, worldPosition);
+			mask_acc += border_mask;
+			hit_acc += 1.0;
 		}
 	}
 
+	/* Compute SSR contribution */
 	if (weight_acc > 0.0) {
-		accumulate_light(ssr_accum / weight_acc, mask / float(NUM_NEIGHBORS), spec_accum);
+		/* Fade intensity based on roughness and average distance to hit */
+		float fade = saturate(2.0 - roughness * 2.0); /* fade between 0.5 and 1.0 roughness */
+		fade *= mask_acc / hit_acc;
+		fade *= mask_acc / hit_acc;
+		accumulate_light(ssr_accum / weight_acc, fade, spec_accum);
 	}
 
 	/* If SSR contribution is not 1.0, blend with cubemaps */