Eevee: Render: Add Subsurface Pass support.

6 files changed, 174 insertions, 12 deletions

diff --git a/source/blender/draw/engines/eevee/eevee_materials.c b/source/blender/draw/engines/eevee/eevee_materials.c
index 84627e03137..8c0ba921419 100644
--- a/source/blender/draw/engines/eevee/eevee_materials.c
+++ b/source/blender/draw/engines/eevee/eevee_materials.c
@@ -1056,9 +1056,16 @@ static void material_opaque(
 						DRW_shgroup_uniform_texture(*shgrp, "sssTexProfile", sss_tex_profile);
 					}
 
-					DRW_shgroup_stencil_mask(*shgrp, e_data.sss_count + 1);
-					EEVEE_subsurface_add_pass(sldata, vedata, e_data.sss_count + 1, sss_profile);
-					e_data.sss_count++;
+					/* Limit of 8 bit stencil buffer. ID 255 is refraction. */
+					if (e_data.sss_count < 254) {
+						DRW_shgroup_stencil_mask(*shgrp, e_data.sss_count + 1);
+						EEVEE_subsurface_add_pass(sldata, vedata, e_data.sss_count + 1, sss_profile);
+						e_data.sss_count++;
+					}
+					else {
+						/* TODO : display message. */
+						printf("Error: Too many different Subsurface shader in the scene.\n");
+					}
 				}
 			}
 		}
diff --git a/source/blender/draw/engines/eevee/eevee_private.h b/source/blender/draw/engines/eevee/eevee_private.h
index 17d41cabe66..1c35d5e7262 100644
--- a/source/blender/draw/engines/eevee/eevee_private.h
+++ b/source/blender/draw/engines/eevee/eevee_private.h
@@ -178,6 +178,7 @@ typedef struct EEVEE_PassList {
 	struct DRWPass *ssr_resolve;
 	struct DRWPass *sss_blur_ps;
 	struct DRWPass *sss_resolve_ps;
+	struct DRWPass *sss_accum_ps;
 	struct DRWPass *color_downsample_ps;
 	struct DRWPass *color_downsample_cube_ps;
 	struct DRWPass *taa_resolve;
@@ -220,6 +221,7 @@ typedef struct EEVEE_FramebufferList {
 	struct GPUFrameBuffer *bloom_accum_fb[MAX_BLOOM_STEP - 1];
 	struct GPUFrameBuffer *sss_blur_fb;
 	struct GPUFrameBuffer *sss_clear_fb;
+	struct GPUFrameBuffer *sss_accum_fb;
 	struct GPUFrameBuffer *dof_down_fb;
 	struct GPUFrameBuffer *dof_scatter_far_fb;
 	struct GPUFrameBuffer *dof_scatter_near_fb;
@@ -249,6 +251,8 @@ typedef struct EEVEE_TextureList {
 	struct GPUTexture *bloom_blit; /* R16_G16_B16 */
 	struct GPUTexture *bloom_downsample[MAX_BLOOM_STEP]; /* R16_G16_B16 */
 	struct GPUTexture *bloom_upsample[MAX_BLOOM_STEP - 1]; /* R16_G16_B16 */
+	struct GPUTexture *sss_dir_accum;
+	struct GPUTexture *sss_col_accum;
 	struct GPUTexture *ssr_normal_input;
 	struct GPUTexture *ssr_specrough_input;
 	struct GPUTexture *ssr_hit_output;
@@ -820,10 +824,12 @@ void EEVEE_screen_raytrace_free(void);
 /* eevee_subsurface.c */
 int EEVEE_subsurface_init(EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata);
 void EEVEE_subsurface_cache_init(EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata);
+void EEVEE_subsurface_output_init(EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata);
 void EEVEE_subsurface_add_pass(
         EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata, unsigned int sss_id, struct GPUUniformBuffer *sss_profile);
 void EEVEE_subsurface_data_render(EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata);
 void EEVEE_subsurface_compute(EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata);
+void EEVEE_subsurface_output_accumulate(EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata);
 void EEVEE_subsurface_free(void);
 
 /* eevee_motion_blur.c */
diff --git a/source/blender/draw/engines/eevee/eevee_render.c b/source/blender/draw/engines/eevee/eevee_render.c
index f7ebded2408..e3d6d182053 100644
--- a/source/blender/draw/engines/eevee/eevee_render.c
+++ b/source/blender/draw/engines/eevee/eevee_render.c
@@ -160,6 +160,51 @@ static void eevee_render_result_combined(
 	DRW_framebuffer_read_data(rr->xof, rr->yof, rr->rectx, rr->recty, 4, 0, rp->rect);
 }
 
+static void eevee_render_result_subsurface(
+        RenderResult *rr, const char *viewname,
+        EEVEE_Data *vedata, EEVEE_ViewLayerData *UNUSED(sldata))
+{
+	const DRWContextState *draw_ctx = DRW_context_state_get();
+	ViewLayer *view_layer = draw_ctx->view_layer;
+
+	if ((view_layer->passflag & SCE_PASS_SUBSURFACE_COLOR) != 0) {
+		RenderLayer *rl = rr->layers.first;
+		RenderPass *rp = RE_pass_find_by_name(rl, RE_PASSNAME_SUBSURFACE_COLOR, viewname);
+
+		IDProperty *props = BKE_view_layer_engine_evaluated_get(view_layer, COLLECTION_MODE_NONE, RE_engine_id_BLENDER_EEVEE);
+		float render_samples = (float)BKE_collection_engine_property_value_get_int(props, "taa_render_samples");
+
+		DRW_framebuffer_bind(vedata->fbl->sss_accum_fb);
+		DRW_framebuffer_read_data(rr->xof, rr->yof, rr->rectx, rr->recty, 3, 1, rp->rect);
+
+		/* This is the accumulated color. Divide by the number of samples. */
+		for (int i = 0; i < rr->rectx * rr->recty * 3; i++) {
+			rp->rect[i] /= render_samples;
+		}
+	}
+
+	if ((view_layer->passflag & SCE_PASS_SUBSURFACE_DIRECT) != 0) {
+		RenderLayer *rl = rr->layers.first;
+		RenderPass *rp = RE_pass_find_by_name(rl, RE_PASSNAME_SUBSURFACE_DIRECT, viewname);
+
+		IDProperty *props = BKE_view_layer_engine_evaluated_get(view_layer, COLLECTION_MODE_NONE, RE_engine_id_BLENDER_EEVEE);
+		float render_samples = (float)BKE_collection_engine_property_value_get_int(props, "taa_render_samples");
+
+		DRW_framebuffer_bind(vedata->fbl->sss_accum_fb);
+		DRW_framebuffer_read_data(rr->xof, rr->yof, rr->rectx, rr->recty, 3, 0, rp->rect);
+
+		/* This is the accumulated color. Divide by the number of samples. */
+		for (int i = 0; i < rr->rectx * rr->recty * 3; i++) {
+			rp->rect[i] /= render_samples;
+		}
+	}
+
+	if ((view_layer->passflag & SCE_PASS_SUBSURFACE_INDIRECT) != 0) {
+		/* Do nothing as all the lighting is in the direct pass.
+		 * TODO : Separate Direct from indirect lighting. */
+	}
+}
+
 static void eevee_render_result_normal(
         RenderResult *rr, const char *viewname,
         EEVEE_Data *vedata, EEVEE_ViewLayerData *UNUSED(sldata))
@@ -253,6 +298,13 @@ void EEVEE_render_draw(EEVEE_Data *vedata, struct RenderEngine *engine, struct D
 	EEVEE_lights_cache_finish(sldata);
 	EEVEE_lightprobes_cache_finish(sldata, vedata);
 
+	if ((view_layer->passflag & (SCE_PASS_SUBSURFACE_COLOR |
+	                             SCE_PASS_SUBSURFACE_DIRECT |
+	                             SCE_PASS_SUBSURFACE_INDIRECT)) != 0)
+	{
+		EEVEE_subsurface_output_init(sldata, vedata);
+	}
+
 	/* Init render result. */
 	const char *viewname = NULL;
 	const float *render_size = DRW_viewport_size_get();
@@ -304,13 +356,12 @@ void EEVEE_render_draw(EEVEE_Data *vedata, struct RenderEngine *engine, struct D
 		/* Effects pre-transparency */
 		EEVEE_subsurface_compute(sldata, vedata);
 		EEVEE_reflection_compute(sldata, vedata);
-		EEVEE_occlusion_draw_debug(sldata, vedata);
-		DRW_draw_pass(psl->probe_display);
 		EEVEE_refraction_compute(sldata, vedata);
 		/* Opaque refraction */
 		DRW_draw_pass(psl->refract_depth_pass);
 		DRW_draw_pass(psl->refract_depth_pass_cull);
 		DRW_draw_pass(psl->refract_pass);
+		EEVEE_subsurface_output_accumulate(sldata, vedata);
 		/* Result NORMAL */
 		eevee_render_result_normal(rr, viewname, vedata, sldata);
 		/* Volumetrics Resolve Opaque */
@@ -324,8 +375,8 @@ void EEVEE_render_draw(EEVEE_Data *vedata, struct RenderEngine *engine, struct D
 		EEVEE_draw_effects(sldata, vedata);
 	}
 
-	/* Result Combined */
 	eevee_render_result_combined(rr, viewname, vedata, sldata);
+	eevee_render_result_subsurface(rr, viewname, vedata, sldata);
 
 	RE_engine_end_result(engine, rr, false, false, false);
 }
diff --git a/source/blender/draw/engines/eevee/eevee_subsurface.c b/source/blender/draw/engines/eevee/eevee_subsurface.c
index e93e9574acd..4ca1f0f42a1 100644
--- a/source/blender/draw/engines/eevee/eevee_subsurface.c
+++ b/source/blender/draw/engines/eevee/eevee_subsurface.c
@@ -33,7 +33,7 @@
 #include "GPU_texture.h"
 
 static struct {
-	struct GPUShader *sss_sh[3];
+	struct GPUShader *sss_sh[4];
 } e_data = {NULL}; /* Engine data */
 
 extern char datatoc_common_uniforms_lib_glsl[];
@@ -49,6 +49,9 @@ static void eevee_create_shader_subsurface(void)
 	e_data.sss_sh[1] = DRW_shader_create_fullscreen(frag_str, "#define SECOND_PASS\n");
 	e_data.sss_sh[2] = DRW_shader_create_fullscreen(frag_str, "#define SECOND_PASS\n"
 	                                                          "#define USE_SEP_ALBEDO\n");
+	e_data.sss_sh[3] = DRW_shader_create_fullscreen(frag_str, "#define SECOND_PASS\n"
+	                                                          "#define USE_SEP_ALBEDO\n"
+	                                                          "#define RESULT_ACCUM\n");
 
 	MEM_freeN(frag_str);
 }
@@ -71,6 +74,11 @@ int EEVEE_subsurface_init(EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata)
 		effects->sss_separate_albedo = BKE_collection_engine_property_value_get_bool(props, "sss_separate_albedo");
 		common_data->sss_jitter_threshold = BKE_collection_engine_property_value_get_float(props, "sss_jitter_threshold");
 
+		/* Force separate albedo for final render */
+		if (DRW_state_is_image_render()) {
+			effects->sss_separate_albedo = true;
+		}
+
 		/* Shaders */
 		if (!e_data.sss_sh[0]) {
 			eevee_create_shader_subsurface();
@@ -109,6 +117,47 @@ int EEVEE_subsurface_init(EEVEE_ViewLayerData *sldata, EEVEE_Data *vedata)
 	return 0;
 }
 
+static void set_shgrp_stencil(void *UNUSED(userData), DRWShadingGroup *shgrp)
+{
+	DRW_shgroup_stencil_mask(shgrp, 255);
+}
+
+void EEVEE_subsurface_output_init(EEVEE_ViewLayerData *UNUSED(sldata), EEVEE_Data *vedata)
+{
+	EEVEE_FramebufferList *fbl = vedata->fbl;
+	EEVEE_TextureList *txl = vedata->txl;
+	const float *viewport_size = DRW_viewport_size_get();
+
+	const DRWContextState *draw_ctx = DRW_context_state_get();
+	ViewLayer *view_layer = draw_ctx->view_layer;
+	IDProperty *props = BKE_view_layer_engine_evaluated_get(view_layer, COLLECTION_MODE_NONE, RE_engine_id_BLENDER_EEVEE);
+
+	if (BKE_collection_engine_property_value_get_bool(props, "sss_enable")) {
+		float clear[4] = {0.0f, 0.0f, 0.0f, 0.0f};
+
+		DRWFboTexture tex_data[2] = {{&txl->sss_dir_accum, DRW_TEX_RGBA_16, 0},
+		                             {&txl->sss_col_accum, DRW_TEX_RGBA_16, 0}};
+		DRW_framebuffer_init(&fbl->sss_accum_fb, &draw_engine_eevee_type, (int)viewport_size[0], (int)viewport_size[1],
+		                     tex_data, 2);
+
+		/* Clear texture. */
+		DRW_framebuffer_bind(fbl->sss_accum_fb);
+		DRW_framebuffer_clear(true, false, false, clear, 0.0f);
+
+		/* Make the opaque refraction pass mask the sss. */
+		DRWState state = DRW_STATE_WRITE_COLOR | DRW_STATE_DEPTH_EQUAL | DRW_STATE_CLIP_PLANES |
+		                 DRW_STATE_WIRE | DRW_STATE_WRITE_STENCIL;
+		DRW_pass_state_set(vedata->psl->refract_pass, state);
+		DRW_pass_foreach_shgroup(vedata->psl->refract_pass, &set_shgrp_stencil, NULL);
+	}
+	else {
+		/* Cleanup to release memory */
+		DRW_TEXTURE_FREE_SAFE(txl->sss_dir_accum);
+		DRW_TEXTURE_FREE_SAFE(txl->sss_col_accum);
+		DRW_FRAMEBUFFER_FREE_SAFE(fbl->sss_accum_fb);
+	}
+}
+
 void EEVEE_subsurface_cache_init(EEVEE_ViewLayerData *UNUSED(sldata), EEVEE_Data *vedata)
 {
 	EEVEE_PassList *psl = vedata->psl;
@@ -121,7 +170,9 @@ void EEVEE_subsurface_cache_init(EEVEE_ViewLayerData *UNUSED(sldata), EEVEE_Data
 		 */
 		psl->sss_blur_ps = DRW_pass_create("Blur Horiz", DRW_STATE_WRITE_COLOR | DRW_STATE_STENCIL_EQUAL);
 
-		psl->sss_resolve_ps = DRW_pass_create("Blur Vert", DRW_STATE_WRITE_COLOR | DRW_STATE_ADDITIVE | DRW_STATE_STENCIL_EQUAL);
+		DRWState state = DRW_STATE_WRITE_COLOR | DRW_STATE_ADDITIVE | DRW_STATE_STENCIL_EQUAL;
+		psl->sss_resolve_ps = DRW_pass_create("Blur Vert", state);
+		psl->sss_accum_ps = DRW_pass_create("Resolve Accum", state);
 	}
 }
 
@@ -157,6 +208,18 @@ void EEVEE_subsurface_add_pass(
 	if (effects->sss_separate_albedo) {
 		DRW_shgroup_uniform_buffer(grp, "sssAlbedo", &txl->sss_albedo);
 	}
+
+	if (DRW_state_is_image_render()) {
+		grp = DRW_shgroup_create(e_data.sss_sh[3], psl->sss_accum_ps);
+		DRW_shgroup_uniform_texture(grp, "utilTex", EEVEE_materials_get_util_tex());
+		DRW_shgroup_uniform_buffer(grp, "depthBuffer", &dtxl->depth);
+		DRW_shgroup_uniform_buffer(grp, "sssData", &txl->sss_blur);
+		DRW_shgroup_uniform_buffer(grp, "sssAlbedo", &txl->sss_albedo);
+		DRW_shgroup_uniform_block(grp, "sssProfile", sss_profile);
+		DRW_shgroup_uniform_block(grp, "common_block", sldata->common_ubo);
+		DRW_shgroup_stencil_mask(grp, sss_id);
+		DRW_shgroup_call_add(grp, quad, NULL);
+	}
 }
 
 void EEVEE_subsurface_data_render(EEVEE_ViewLayerData *UNUSED(sldata), EEVEE_Data *vedata)
@@ -271,9 +334,35 @@ void EEVEE_subsurface_compute(EEVEE_ViewLayerData *UNUSED(sldata), EEVEE_Data *v
 	}
 }
 
+void EEVEE_subsurface_output_accumulate(EEVEE_ViewLayerData *UNUSED(sldata), EEVEE_Data *vedata)
+{
+	EEVEE_PassList *psl = vedata->psl;
+	EEVEE_FramebufferList *fbl = vedata->fbl;
+	EEVEE_TextureList *txl = vedata->txl;
+	EEVEE_StorageList *stl = vedata->stl;
+	EEVEE_EffectsInfo *effects = stl->effects;
+
+	if (((effects->enabled_effects & EFFECT_SSS) != 0) && (fbl->sss_accum_fb != NULL)) {
+		/* Copy stencil channel, could be avoided (see EEVEE_subsurface_init) */
+		DRW_framebuffer_blit(fbl->main, fbl->sss_blur_fb, false, true);
+
+		/* Only do vertical pass + Resolve */
+		DRW_framebuffer_texture_detach(txl->sss_stencil);
+		DRW_framebuffer_texture_attach(fbl->sss_accum_fb, txl->sss_stencil, 0, 0);
+		DRW_framebuffer_bind(fbl->sss_accum_fb);
+		DRW_draw_pass(psl->sss_accum_ps);
+
+		/* Restore */
+		DRW_framebuffer_texture_detach(txl->sss_stencil);
+		DRW_framebuffer_texture_attach(fbl->sss_blur_fb, txl->sss_stencil, 0, 0);
+		DRW_framebuffer_bind(fbl->main);
+	}
+}
+
 void EEVEE_subsurface_free(void)
 {
 	DRW_SHADER_FREE_SAFE(e_data.sss_sh[0]);
 	DRW_SHADER_FREE_SAFE(e_data.sss_sh[1]);
 	DRW_SHADER_FREE_SAFE(e_data.sss_sh[2]);
+	DRW_SHADER_FREE_SAFE(e_data.sss_sh[3]);
 }
diff --git a/source/blender/draw/engines/eevee/shaders/effect_subsurface_frag.glsl b/source/blender/draw/engines/eevee/shaders/effect_subsurface_frag.glsl
index 184eac54c26..d9e12bcc4bd 100644
--- a/source/blender/draw/engines/eevee/shaders/effect_subsurface_frag.glsl
+++ b/source/blender/draw/engines/eevee/shaders/effect_subsurface_frag.glsl
@@ -18,7 +18,10 @@ uniform sampler2DArray utilTex;
 #define texelfetch_noise_tex(coord) texelFetch(utilTex, ivec3(ivec2(coord) % LUT_SIZE, 2.0), 0)
 #endif /* UTIL_TEX */
 
-out vec4 FragColor;
+layout(location = 0) out vec4 FragColor;
+#ifdef RESULT_ACCUM
+layout(location = 1) out vec4 sssColor;
+#endif
 
 uniform mat4 ProjectionMatrix;
 
@@ -84,10 +87,15 @@ void main(void)
 #ifdef FIRST_PASS
 	FragColor = vec4(accum, sss_data.a);
 #else /* SECOND_PASS */
-	#ifdef USE_SEP_ALBEDO
+# ifdef USE_SEP_ALBEDO
+#  ifdef RESULT_ACCUM
+	FragColor = vec4(accum, 1.0);
+	sssColor = texture(sssAlbedo, uvs);
+#  else
 	FragColor = vec4(accum * texture(sssAlbedo, uvs).rgb, 1.0);
-	#else
+#  endif
+# else
 	FragColor = vec4(accum, 1.0);
-	#endif
+# endif
 #endif
 }
diff --git a/source/blender/draw/intern/draw_manager.c b/source/blender/draw/intern/draw_manager.c
index b9873ff3926..0e8861b48bb 100644
--- a/source/blender/draw/intern/draw_manager.c
+++ b/source/blender/draw/intern/draw_manager.c
@@ -1105,6 +1105,7 @@ void DRW_shgroup_state_disable(DRWShadingGroup *shgroup, DRWState state)
 
 void DRW_shgroup_stencil_mask(DRWShadingGroup *shgroup, unsigned int mask)
 {
+	BLI_assert(mask <= 255);
 	shgroup->stencil_mask = mask;
 }