msaa

1 files changed, 100 insertions, 15 deletions

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) {