msaa

7 files changed, 172 insertions, 30 deletions

diff --git a/c2.cpp b/c2.cpp
index ac9e7d4..94e2e38 100644
--- a/c2.cpp
+++ b/c2.cpp
@@ -24,6 +24,8 @@ extern "C" {
 #define scene_arena_size (1024 * 4)
 #define per_frame_memory_size (1024 * 1024)
 
+#define MSAA_SAMPLES 4
+
 static float verts[] = {
 	 0.5f,  0.5f, 1.0f, 0.0f, 0.0f, 1.0f, 1.0f,
 	-0.5f,  0.5f, 0.0f, 1.0f, 0.0f, 0.0f, 1.0f,
@@ -653,7 +655,7 @@ struct C2 : public App {
 	Tonemap tonemap;
 	Buffer_Id vbo, cbuf;
 	Sampler_Id clamped_linear;
-	Texture_Id hdr_target;
+	Texture_Id hdr_target, hdr_resolved, ms_depth;
 	Texture_Id ui_texture;
 	Buffer_Id ui_buffer;
 	Line_Renderer lr;
@@ -876,7 +878,7 @@ struct C2 : public App {
 
 		pb.begin_rp();
 		pb.rp_target(hdr_target, Clear_Mode::restore);
-		pb.rp_depth_target(dev->get_depth_target(), Clear_Mode::restore);
+		pb.rp_depth_target(ms_depth, Clear_Mode::restore);
 		Render_Pass& sky_pass = pb.build_rp();
 
 		pb.begin_rp();
@@ -903,6 +905,7 @@ struct C2 : public App {
 			dev,
 			&frame_arena,
 			hdr_target,
+			ms_depth,
 			&lighting
 		);
 		ctx.debug_pop();
@@ -919,6 +922,8 @@ struct C2 : public App {
 		);
 		ctx.debug_pop();
 
+		ctx.resolve(hdr_resolved, hdr_target);
+
 		tonemap.update(dev, 0.2f);
 
 		ctx.debug_push("TONEMAP");
@@ -927,7 +932,7 @@ struct C2 : public App {
 			&frame_arena,
 			quad,
 			tonemap_pass,
-			hdr_target,
+			hdr_resolved,
 			clamped_linear
 		);
 		ctx.debug_pop();
@@ -1014,6 +1019,8 @@ struct C2 : public App {
 		assets.destroy();
 		dev->destroy_texture(default_texture);
 		dev->destroy_texture(hdr_target);
+		dev->destroy_texture(hdr_resolved);
+		dev->destroy_texture(ms_depth);
 		dev->destroy_texture(ui_texture);
 		dev->destroy_buffer(ui_buffer);
 		dev->destroy_sampler(clamped_linear);
@@ -1025,7 +1032,9 @@ struct C2 : public App {
 	void on_resize() override {
 		ui->layout(w, h);
 		dev->on_resize();
+		dev->destroy_texture(ms_depth);
 		dev->destroy_texture(hdr_target);
+		dev->destroy_texture(hdr_resolved);
 		dev->destroy_texture(ui_texture);
 		dev->destroy_buffer(ui_buffer);
 		make_hdr_target();
@@ -1033,16 +1042,43 @@ struct C2 : public App {
 	}
 
 	void make_hdr_target() {
+		int sw = dev->swap_w();
+		int sh = dev->swap_h();
 		hdr_target = dev->create_texture(
-			"HDR target",
+			"MS HDR target",
 			texture_format_rgba16f,
-			Texture_Flags::sampleable | Texture_Flags::colour_target,
-			dev->swap_w(),
-			dev->swap_h(),
+			Texture_Flags::copy_src | Texture_Flags::colour_target,
+			sw,
+			sh,
 			1,
 			1,
 			1,
-			Buffer_Id(0)
+			Buffer_Id(0),
+			MSAA_SAMPLES
+		);
+		hdr_resolved = dev->create_texture(
+			"Resolved HDR",
+			texture_format_rgba16f,
+			Texture_Flags::sampleable | Texture_Flags::copy_dst,
+			sw,
+			sh,
+			1,
+			1,
+			1,
+			Buffer_Id(0),
+			1
+		);
+		ms_depth = dev->create_texture(
+			"MS depth",
+			texture_format_d24s8,
+			Texture_Flags::sampleable | Texture_Flags::depth_stencil_target,
+			sw,
+			sh,
+			1,
+			1,
+			1,
+			0,
+			MSAA_SAMPLES
 		);
 	}
 
diff --git a/model.cpp b/model.cpp
index f24c0eb..cb9c23e 100644
--- a/model.cpp
+++ b/model.cpp
@@ -384,7 +384,7 @@ void Model_Instance::render(
 				res.vp
 			);
 		} else {
-			pb.depth(true, false, Depth_Mode::equal);
+			pb.depth(true, true, Depth_Mode::less);
 			pb.shader(mesh.shader);
 			pb.sbuffer(mesh.light_binding, lighting->lights.gpuonly);
 			pb.sbuffer(mesh.casters_binding, lighting->casters.gpuonly);
diff --git a/pipeline.cpp b/pipeline.cpp
index 0636a1c..374dad8 100644
--- a/pipeline.cpp
+++ b/pipeline.cpp
@@ -23,6 +23,7 @@ Pipeline_Builder& Pipeline_Builder::rp_target(Texture_Id id, Colour clear) {
 	Render_Pass::Target t{
 		.id = id,
 		.fmt = texture.fmt,
+		.samples = texture.samples,
 		.mode = Clear_Mode::clear,
 		.clear = { .colour = clear }
 	};
@@ -37,6 +38,7 @@ Pipeline_Builder& Pipeline_Builder::rp_target(Texture_Id id, Clear_Mode clear) {
 	Render_Pass::Target t{
 		.id = id,
 		.fmt = texture.fmt,
+		.samples = texture.samples,
 		.mode = clear,
 		.clear = { .depth = 0.0f }
 	};
@@ -51,6 +53,7 @@ Pipeline_Builder& Pipeline_Builder::rp_depth_target(Texture_Id id, float clear)
 	Render_Pass::Target t{
 		.id = id,
 		.fmt = texture.fmt,
+		.samples = texture.samples,
 		.mode = Clear_Mode::clear,
 		.clear = { .depth = clear }
 	};
@@ -63,6 +66,7 @@ Pipeline_Builder& Pipeline_Builder::rp_depth_target(Texture_Id id, Clear_Mode mo
 	Render_Pass::Target t{
 		.id = id,
 		.fmt = texture.fmt,
+		.samples = texture.samples,
 		.mode = mode,
 		.clear = { .depth = 0.0f }
 	};
@@ -73,7 +77,7 @@ Pipeline_Builder& Pipeline_Builder::rp_depth_target(Texture_Id id, Clear_Mode mo
 
 void Pipeline_Builder::validate_rp() {
 	int i, c = pass->colour_count;
-	int w, h;
+	int w, h, s;
 	assert(c || pass->depth.id);
 	if (c) {
 		Texture& tex = dev->get_texture(pass->colours[0].id);
@@ -81,7 +85,8 @@ void Pipeline_Builder::validate_rp() {
 		assert(pass->colours[0].fmt == tex.fmt);
 		w = tex.w;
 		h = tex.h;
-		assert(w && h);
+		s = tex.samples;
+		assert(w && h && s);
 	}
 	for (i = 1; i < c; i++) {
 		Texture& tex = dev->get_texture(pass->colours[i].id);
@@ -89,6 +94,7 @@ void Pipeline_Builder::validate_rp() {
 		assert(pass->colours[i].fmt == tex.fmt);
 		assert(tex.w == w);
 		assert(tex.h == h);
+		assert(tex.samples == s);
 	}
 	if (pass->depth.id) {
 		Texture& d = dev->get_texture(pass->depth.id);
@@ -101,6 +107,7 @@ void Pipeline_Builder::validate_rp() {
 		if (c) {
 			assert(d.w == w);
 			assert(d.h == h);
+			assert(d.samples == s);
 		}
 	}
 }
diff --git a/renderer.cpp b/renderer.cpp
index d8c4895..c7040ec 100644
--- a/renderer.cpp
+++ b/renderer.cpp
@@ -108,7 +108,9 @@ void Renderer::update_globals(
 
 void Renderer::render(
 	Device* dev,
-	Arena* a, Texture_Id hdr_target,
+	Arena* a,
+	Texture_Id hdr_target,
+	Texture_Id depth_target,
 	const Lighting* l
 ) {
 	int i, j;
@@ -124,7 +126,7 @@ void Renderer::render(
 	Render_Pass& forward_pass = pb
 		.begin_rp()
 		.rp_target(hdr_target, Clear_Mode::restore)
-		.rp_depth_target(dev->get_depth_target(), Clear_Mode::restore)
+		.rp_depth_target(depth_target, 1.0f)
 		.build_rp();
 	Render_Pass* shadow_passes[Lighting::max_shadows];
 	for (i = 0; i < Lighting::max_shadows; i++) {
diff --git a/renderer.hpp b/renderer.hpp
index c49c804..6bfe551 100644
--- a/renderer.hpp
+++ b/renderer.hpp
@@ -50,6 +50,7 @@ struct Renderer {
 		Device* dev,
 		Arena* a,
 		Texture_Id hdr_target,
+		Texture_Id depth_target,
 		const Lighting* l
 	);
 
diff --git a/video.cpp b/video.cpp
index 08ee606..9e2699e 100644
--- a/video.cpp
+++ b/video.cpp
@@ -561,7 +561,8 @@ struct Texture_Vk : public Texture, public Late_Terminated {
 		int array_size,
 		int start_mip,
 		int start_array,
-		bool alias
+		bool alias,
+		int samples
 	);
 	void destroy(Device_Vk*) override;
 	void set_name(Device_Vk* dev, const char* name);
@@ -732,7 +733,8 @@ struct Pipeline_Vk {
 		Arena& scope,
 		Device_Vk* dev,
 		VkGraphicsPipelineCreateInfo& info,
-		const Pipeline& desc
+		const Pipeline& desc,
+		const Render_Pass& rpo
 	);
 	void init_depthstencil(
 		Arena& scope,
@@ -934,6 +936,7 @@ struct Device_Vk : public Device {
 	VkPhysicalDevice phys_dev;
 	VkSurfaceKHR surf;
 	uint32_t backbuffer_index;
+	VkSampleCountFlagBits max_samples;
 	Texture_Id backbuffer_id;
 	Swap_Cap swap_cap;
 	VkPhysicalDeviceMemoryProperties mem_props;
@@ -1021,6 +1024,9 @@ struct Device_Vk : public Device {
 	void create_terminators();
 	void create_depth(int w, int h);
 
+	VkSampleCountFlagBits get_max_samples();
+	VkSampleCountFlagBits get_samples(int);
+
 	int find_memory_type(
 		uint32_t filter,
 		VkMemoryPropertyFlags flags
@@ -1114,6 +1120,37 @@ void Device_Vk::init_validation() {
 
 #endif
 
+VkSampleCountFlagBits Device_Vk::get_max_samples() {
+	VkPhysicalDeviceProperties p;
+	VkSampleCountFlagBits
+		i = VK_SAMPLE_COUNT_64_BIT,
+		e = VK_SAMPLE_COUNT_1_BIT;
+	VkSampleCountFlags c;
+	vkGetPhysicalDeviceProperties(phys_dev, &p);
+	c =
+		p.limits.framebufferColorSampleCounts &
+		p.limits.framebufferDepthSampleCounts;
+	for (; i >= e; i = (VkSampleCountFlagBits)(i >> 1))
+		if (c & i) return i;
+	return VK_SAMPLE_COUNT_1_BIT;
+}
+
+VkSampleCountFlagBits Device_Vk::get_samples(
+	int c
+) {
+	VkSampleCountFlagBits b = VK_SAMPLE_COUNT_1_BIT;
+	switch (c) {
+		case 1:  b = VK_SAMPLE_COUNT_1_BIT;  break;
+		case 2:  b = VK_SAMPLE_COUNT_2_BIT;  break;
+		case 4:  b = VK_SAMPLE_COUNT_4_BIT;  break;
+		case 8:  b = VK_SAMPLE_COUNT_8_BIT;  break;
+		case 16: b = VK_SAMPLE_COUNT_16_BIT; break;
+		case 32: b = VK_SAMPLE_COUNT_32_BIT; break;
+		case 64: b = VK_SAMPLE_COUNT_64_BIT; break;
+		default: break;
+	}
+	return std::min(max_samples, b);
+}
 
 bool Device_Vk::has_validation() {
 	unsigned count, i;
@@ -1352,6 +1389,7 @@ void Device_Vk::init_internal() {
 #endif
 	surf = app_create_vk_surface(app, inst);
 	create_dev(&swap_cap);
+	max_samples = get_max_samples();
 	vrama.init(this);
 	gladLoaderLoadVulkan(inst, phys_dev, dev);
 	vkGetDeviceQueue(dev, (uint32_t)queue_index, 0, &queue);
@@ -1642,7 +1680,7 @@ void Renderpass_Vk::init(
 		auto& colour = rp.colours[i];
 		auto& ad = ads[index];
 		ad.format = get_vk_format(colour.fmt);
-		ad.samples = VK_SAMPLE_COUNT_1_BIT;
+		ad.samples = dev->get_samples(colour.samples);
 		ad.loadOp = load_op_from_mode(colour.mode);
 		ad.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
 		ad.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
@@ -1659,9 +1697,10 @@ void Renderpass_Vk::init(
 	if (has_depth) {
 		int i = count++;
 		auto& ad = ads[i];
-		ad.format = get_vk_format(dev->get_texture(rp.depth.id).fmt);
-		ad.samples = VK_SAMPLE_COUNT_1_BIT;
-		ad.loadOp = load_op_from_mode(rp.depth.mode);
+		auto& depth = rp.depth;
+		ad.format = get_vk_format(dev->get_texture(depth.id).fmt);
+		ad.samples = dev->get_samples(depth.samples);
+		ad.loadOp = load_op_from_mode(depth.mode);
 		ad.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
 		ad.stencilLoadOp = ad.loadOp;
 		ad.stencilStoreOp = VK_ATTACHMENT_STORE_OP_STORE;
@@ -1887,7 +1926,8 @@ Texture_Id Swapchain::create_image(
 		1,
 		0,
 		0,
-		true
+		true,
+		1
 	);
 	return id;
 }
@@ -2286,6 +2326,37 @@ void Context::copy(
 	);
 }
 
+void Context::resolve(Texture_Id dst, Texture_Id src) {
+	Context_Vk* ctx = (Context_Vk*)this;
+	Device_Vk* dev = ctx->dev;
+	Texture_Vk& d = *(Texture_Vk*)&dev->get_texture(dst);
+	Texture_Vk& s = *(Texture_Vk*)&dev->get_texture(src);
+	VkImageResolve r{};
+	r.srcSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+	r.srcSubresource.layerCount = 1;
+	r.dstSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+	r.dstSubresource.layerCount = 1;
+	r.extent.width = d.w;
+	r.extent.height = d.h;
+	r.extent.depth = 1;
+	assert(d.w == s.w);
+	assert(d.h == s.h);
+	assert(d.d == 1 && s.d == 1);
+	assert(d.samples == 1 && s.samples > 1);
+	ctx->check_end_rp();
+	transition(src, Resource_State::copy_src);
+	transition(dst, Resource_State::copy_dst);
+	vkCmdResolveImage(
+		ctx->cb,
+		s.image,
+		state_to_image_layout(s.state),
+		d.image,
+		state_to_image_layout(d.state),
+		1,
+		&r
+	);
+}
+
 void Context::transition(Texture_Id id, Resource_State state) {
 	Context_Vk* ctx = (Context_Vk*)this;
 	Device_Vk* dev = ctx->dev;
@@ -2434,6 +2505,14 @@ void Context::transition(Texture_Id id, Resource_State state) {
 		b.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
 		src_stage = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
 		dst_stage = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+	} else if (
+		src_layout == VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL &&
+		dst_layout == VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL
+	) {
+		b.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
+		b.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
+		src_stage = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+		dst_stage = VK_PIPELINE_STAGE_TRANSFER_BIT;
 	} else {
 		print_err("Bad resource transition.\n");
 		assert(0);
@@ -2857,7 +2936,8 @@ void Pipeline_Vk::init_msaa(
 	Arena& scope,
 	Device_Vk* dev,
 	VkGraphicsPipelineCreateInfo& info,
-	const Pipeline& desc
+	const Pipeline& desc,
+	const Render_Pass& rpo
 ) {
 	VkPipelineMultisampleStateCreateInfo& mi = 
 		*(VkPipelineMultisampleStateCreateInfo*)arena_alloc(
@@ -2869,7 +2949,7 @@ void Pipeline_Vk::init_msaa(
 	zero(&mi, sizeof mi);
 	mi.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO;
 	mi.sampleShadingEnable = VK_FALSE;
-	mi.rasterizationSamples = VK_SAMPLE_COUNT_1_BIT;
+	mi.rasterizationSamples = dev->get_samples(rpo.get_samples());
 	info.pMultisampleState = &mi;
 }
 
@@ -3064,7 +3144,7 @@ void Pipeline_Vk::init(Device_Vk* dev, const Pso_Key& key) {
 	init_input_assembly(scope, dev, info, desc);
 	init_viewport(scope, info, desc);
 	init_rasterisation(scope, dev, info, desc);
-	init_msaa(scope, dev, info, desc);
+	init_msaa(scope, dev, info, desc, key.rpo.rpo);
 	init_depthstencil(scope, dev, info, desc);
 	init_blending(scope, dev, info, key.rpo.rpo, desc);
 	info.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO;
@@ -3594,7 +3674,8 @@ Texture_Id Device::create_texture(
 	int d,
 	int mip_count,
 	int array_size,
-	Buffer_Id init
+	Buffer_Id init,
+	int samples
 ) {
 	VkImageCreateInfo ii{};
 	VkResult r;
@@ -3626,7 +3707,7 @@ Texture_Id Device::create_texture(
 	ii.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
 	ii.usage = get_texture_usage(flags);
 	ii.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
-	ii.samples = VK_SAMPLE_COUNT_1_BIT;
+	ii.samples = dev->get_samples(samples);
 	ii.flags = image_flags;
 	r = vkCreateImage(dev->dev, &ii, &dev->ac, &image);
 	if (r != VK_SUCCESS) {
@@ -3684,7 +3765,8 @@ Texture_Id Device::create_texture(
 		array_size,
 		0,
 		0,
-		false
+		false,
+		samples
 	);
 	if (init) {
 		Context& ctx = dev->acquire();
@@ -3747,7 +3829,8 @@ Texture_Id Device::alias_texture(
 		array_size,
 		start_mip,
 		start_array,
-		true
+		true,
+		texture.samples
 	);
 	nt.set_name(dev, name);
 	return ntid;
@@ -4049,7 +4132,8 @@ void Texture_Vk::init(
 	int array_size,
 	int start_mip,
 	int start_array,
-	bool alias
+	bool alias,
+	int samples
 ) {
 	t->id = id;
 	t->parent = parent;
@@ -4069,6 +4153,7 @@ void Texture_Vk::init(
 	t->start_mip = start_mip;
 	t->start_array = start_array;
 	t->alias = alias;
+	t->samples = samples;
 }
 
 void Texture_Vk::destroy(Device_Vk* dev) {
diff --git a/video.hpp b/video.hpp
index 4afb6ba..530b126 100644
--- a/video.hpp
+++ b/video.hpp
@@ -133,6 +133,7 @@ struct Pipeline {
 	Cull_Mode cull_mode;
 	Vertex_Format_Id vertex_format;
 	Shader_Id shader;
+	int samples;
 	Descriptor descriptors[pipeline_max_descriptors];
 	int descriptor_count;
 
@@ -161,7 +162,8 @@ struct Pipeline {
 			blend_dst_alpha == other.blend_dst_alpha &&
 			blend_mode == other.blend_mode &&
 			blend_mode_alpha == other.blend_mode_alpha &&
-			cull_mode == other.cull_mode;
+			cull_mode == other.cull_mode &&
+			samples == other.samples;
 	}
 
 	bool desc_layout_eq(const Pipeline& other) const {
@@ -228,6 +230,7 @@ struct Render_Pass {
 	struct Target {
 		Texture_Id id;
 		Texture_Format fmt;
+		int samples;
 		Clear_Mode mode;
 		union {
 			Colour colour;
@@ -266,6 +269,11 @@ struct Render_Pass {
 		}
 		return true;
 	}
+
+	int get_samples() const {
+		if (colour_count) return colours[0].samples;
+		return depth.samples;
+	}
 };
 
 struct Vertex_Buffer_Binding {
@@ -359,6 +367,7 @@ struct Texture : public Asset {
 	int w, h, d;
 	int mip_count, array_size;
 	int start_mip, start_array;
+	int samples;
 	bool alias;
 };
 
@@ -486,7 +495,8 @@ struct Device {
 		int d,
 		int mip_count,
 		int array_size,
-		Buffer_Id init
+		Buffer_Id init,
+		int samples = 1
 	);
 	Texture_Id alias_texture(
 		Texture_Id o,
@@ -562,6 +572,7 @@ struct Context {
 		int w,
 		int h
 	);
+	void resolve(Texture_Id dst, Texture_Id src);
 	void transition(Texture_Id id, Resource_State state);
 	void debug_push(const char* name);
 	void debug_pop();