From: Mikko Rasa <tdb@tdb.fi>
Date: Fri, 15 Apr 2022 20:10:37 +0000 (+0300)
Subject: Add support for storage images in Renderer and PipelineState
X-Git-Url: http://git.tdb.fi/?a=commitdiff_plain;h=2a70fecfbbe8708be2bdaa75d222dd5a889a8ed3;p=libs%2Fgl.git

Add support for storage images in Renderer and PipelineState
---

diff --git a/extensions/arb_shader_image_load_store.glext b/extensions/arb_shader_image_load_store.glext
new file mode 100644
index 00000000..591da5d5
--- /dev/null
+++ b/extensions/arb_shader_image_load_store.glext
@@ -0,0 +1 @@
+extension ARB_shader_image_load_store
diff --git a/source/backends/opengl/device_backend.cpp b/source/backends/opengl/device_backend.cpp
index db711e51..0fcbe072 100644
--- a/source/backends/opengl/device_backend.cpp
+++ b/source/backends/opengl/device_backend.cpp
@@ -93,6 +93,7 @@ void OpenGLDevice::fill_info()
 	feat.storage_texture_binding_range = lim.max_storage_texture_bindings;
 
 	state.bound_tex_targets.resize(lim.max_texture_bindings);
+	state.bound_storage_textures.resize(lim.max_storage_texture_bindings);
 	state.bound_uniform_blocks.resize(lim.max_uniform_bindings);
 }
 
diff --git a/source/backends/opengl/device_backend.h b/source/backends/opengl/device_backend.h
index 491f1b31..2583d582 100644
--- a/source/backends/opengl/device_backend.h
+++ b/source/backends/opengl/device_backend.h
@@ -16,6 +16,7 @@ struct OpenGLDeviceState
 {
 	const OpenGLPipelineState *last_pipeline = 0;
 	std::vector<int> bound_tex_targets;
+	std::vector<char> bound_storage_textures;
 	std::vector<char> bound_uniform_blocks;
 	unsigned restart_index = 0;
 	unsigned n_clip_distances = 0;
diff --git a/source/backends/opengl/pipelinestate_backend.cpp b/source/backends/opengl/pipelinestate_backend.cpp
index 3b8c6a5a..a2b8ab32 100644
--- a/source/backends/opengl/pipelinestate_backend.cpp
+++ b/source/backends/opengl/pipelinestate_backend.cpp
@@ -1,5 +1,6 @@
 #include <msp/gl/extensions/arb_direct_state_access.h>
 #include <msp/gl/extensions/arb_sampler_objects.h>
+#include <msp/gl/extensions/arb_shader_image_load_store.h>
 #include <msp/gl/extensions/arb_shader_objects.h>
 #include <msp/gl/extensions/arb_uniform_buffer_object.h>
 #include <msp/gl/extensions/arb_vertex_array_object.h>
@@ -128,7 +129,7 @@ void OpenGLPipelineState::apply() const
 							call.func(call.location, call.size, data+call.location*16);
 					}
 				}
-				else if(r.type==PipelineState::TEXTURE)
+				else if(r.type==PipelineState::SAMPLED_TEXTURE)
 				{
 					if(ARB_direct_state_access)
 						glBindTextureUnit(r.binding, r.texture->id);
@@ -145,6 +146,14 @@ void OpenGLPipelineState::apply() const
 					glBindSampler(r.binding, r.sampler->id);
 					r.sampler->refresh();
 				}
+				else if(r.type==PipelineState::STORAGE_TEXTURE)
+				{
+					static Require _req(ARB_shader_image_load_store);
+					GLenum gl_format = get_gl_pixelformat(r.texture->get_format());
+					glBindImageTexture(r.binding, r.texture->id, 0, true, 0, GL_READ_WRITE, gl_format);
+
+					dev_state.bound_storage_textures[r.binding] = 1;
+				}
 			}
 
 			r.changed = false;
@@ -260,6 +269,13 @@ void OpenGLPipelineState::clear()
 				dev_state.bound_tex_targets[i] = 0;
 			}
 
+		for(unsigned i=0; i<dev_state.bound_storage_textures.size(); ++i)
+			if(dev_state.bound_storage_textures[i])
+			{
+				glBindImageTexture(i, 0, 0, true, 0, GL_READ_ONLY, GL_RGBA8);
+				dev_state.bound_storage_textures[i] = 0;
+			}
+
 		for(unsigned i=0; i<dev_state.bound_uniform_blocks.size(); ++i)
 			if(dev_state.bound_uniform_blocks[i])
 			{
diff --git a/source/backends/vulkan/device_backend.cpp b/source/backends/vulkan/device_backend.cpp
index 8e3f2ab6..9a8b930d 100644
--- a/source/backends/vulkan/device_backend.cpp
+++ b/source/backends/vulkan/device_backend.cpp
@@ -56,7 +56,6 @@ void VulkanDevice::fill_info()
 	limits.max_clip_planes = props.limits.maxClipDistances;
 	limits.max_vertex_attributes = props.limits.maxVertexInputAttributes;
 	limits.max_texture_bindings = props.limits.maxDescriptorSetSampledImages;
-	limits.max_storage_texture_bindings = props.limits.maxDescriptorSetStorageImages;
 	limits.max_color_attachments = props.limits.maxColorAttachments;
 	unsigned samples = props.limits.framebufferColorSampleCounts&props.limits.framebufferDepthSampleCounts&props.limits.framebufferStencilSampleCounts;
 	if(samples&VK_SAMPLE_COUNT_64_BIT)
diff --git a/source/backends/vulkan/pipelinestate_backend.cpp b/source/backends/vulkan/pipelinestate_backend.cpp
index f3281804..5f9e7eb9 100644
--- a/source/backends/vulkan/pipelinestate_backend.cpp
+++ b/source/backends/vulkan/pipelinestate_backend.cpp
@@ -64,12 +64,13 @@ void VulkanPipelineState::update() const
 			{
 				if(r.type==PipelineState::UNIFORM_BLOCK)
 					r.used = self.shprog->uses_uniform_block_binding(r.binding);
-				else if(r.type==PipelineState::TEXTURE)
+				else if(r.type==PipelineState::SAMPLED_TEXTURE || r.type==PipelineState::STORAGE_TEXTURE)
 				{
 					r.used = self.shprog->uses_texture_binding(r.binding);
 					if(r.mip_level>=0)
 						r.texture->refresh_mip_views();
-					r.sampler->refresh();
+					if(r.type==PipelineState::SAMPLED_TEXTURE)
+						r.sampler->refresh();
 				}
 				if(r.binding>=0)
 					changed_sets |= 1<<(r.binding>>20);
@@ -285,12 +286,14 @@ uint64_t VulkanPipelineState::compute_descriptor_set_hash(unsigned index) const
 			result = hash_update<64>(result, reinterpret_cast<uintptr_t>(i->block));
 			result = hash_update<64>(result, reinterpret_cast<uintptr_t>(i->buffer->handle));
 		}
-		else if(i->type==PipelineState::TEXTURE)
+		else if(i->type==PipelineState::SAMPLED_TEXTURE)
 		{
 			result = hash_update<64>(result, reinterpret_cast<uintptr_t>(i->texture->handle));
 			result = hash_update<64>(result, reinterpret_cast<uintptr_t>(i->sampler->handle));
 			result = hash_update<64>(result, i->mip_level);
 		}
+		else if(i->type==PipelineState::STORAGE_TEXTURE)
+			result = hash_update<64>(result, reinterpret_cast<uintptr_t>(i->texture->handle));
 		empty = false;
 	}
 
@@ -330,7 +333,7 @@ unsigned VulkanPipelineState::fill_descriptor_writes(unsigned index, unsigned fr
 		{
 			if(i->type==PipelineState::UNIFORM_BLOCK)
 				++n_buffers;
-			else if(i->type==PipelineState::TEXTURE)
+			else if(i->type==PipelineState::SAMPLED_TEXTURE || i->type==PipelineState::STORAGE_TEXTURE)
 				++n_images;
 		}
 	unsigned n_writes = n_buffers+n_images;
@@ -365,19 +368,28 @@ unsigned VulkanPipelineState::fill_descriptor_writes(unsigned index, unsigned fr
 
 			++buffer_ptr;
 		}
-		else if(i->type==PipelineState::TEXTURE)
+		else if(i->type==PipelineState::SAMPLED_TEXTURE || i->type==PipelineState::STORAGE_TEXTURE)
 		{
-			image_ptr->sampler = handle_cast<::VkSampler>(i->sampler->handle);
 			if(i->mip_level<0)
 				image_ptr->imageView = handle_cast<::VkImageView>(i->texture->view_handle);
 			else
 				image_ptr->imageView = handle_cast<::VkImageView>(i->texture->mip_view_handles[i->mip_level]);
-			image_ptr->imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+
+			if(i->type==PipelineState::SAMPLED_TEXTURE)
+			{
+				image_ptr->sampler = handle_cast<::VkSampler>(i->sampler->handle);
+				image_ptr->imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+				write_ptr->descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+			}
+			else if(i->type==PipelineState::STORAGE_TEXTURE)
+			{
+				image_ptr->imageLayout = VK_IMAGE_LAYOUT_GENERAL;
+				write_ptr->descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
+			}
 
 			write_ptr->sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
 			write_ptr->dstBinding = i->binding&0xFFFFF;
 			write_ptr->descriptorCount = 1;
-			write_ptr->descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
 			write_ptr->pImageInfo = image_ptr;
 
 			++image_ptr;
@@ -389,6 +401,15 @@ unsigned VulkanPipelineState::fill_descriptor_writes(unsigned index, unsigned fr
 	return n_writes;
 }
 
+void VulkanPipelineState::synchronize_resources(bool discard_fb_contents) const
+{
+	const PipelineState &self = *static_cast<const PipelineState *>(this);
+
+	for(const PipelineState::BoundResource &r: self.resources)
+		if(r.type==PipelineState::STORAGE_TEXTURE)
+			r.texture->change_layout(-1, VK_IMAGE_LAYOUT_GENERAL, false);
+}
+
 void VulkanPipelineState::apply(const VulkanCommandRecorder &vkCmd, const VulkanPipelineState *last, unsigned frame, bool negative_viewport) const
 {
 	const PipelineState &self = *static_cast<const PipelineState *>(this);
diff --git a/source/backends/vulkan/pipelinestate_backend.h b/source/backends/vulkan/pipelinestate_backend.h
index 892d8b3a..37c96a4e 100644
--- a/source/backends/vulkan/pipelinestate_backend.h
+++ b/source/backends/vulkan/pipelinestate_backend.h
@@ -39,6 +39,7 @@ protected:
 	VkDescriptorSetLayout get_descriptor_set_layout(unsigned) const;
 	unsigned fill_descriptor_writes(unsigned, unsigned, std::vector<char> &) const;
 
+	void synchronize_resources(bool) const;
 	void apply(const VulkanCommandRecorder &, const VulkanPipelineState *, unsigned, bool) const;
 };
 
diff --git a/source/backends/vulkan/program_backend.cpp b/source/backends/vulkan/program_backend.cpp
index 9402226b..a962a78e 100644
--- a/source/backends/vulkan/program_backend.cpp
+++ b/source/backends/vulkan/program_backend.cpp
@@ -138,7 +138,10 @@ void VulkanProgram::finalize_uniforms()
 				bindings.emplace_back();
 				VkDescriptorSetLayoutBinding &binding = bindings.back();
 				binding.binding = u.binding&0xFFFFF;
-				binding.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+				if(is_sampled_image(u.type))
+					binding.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+				else
+					binding.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
 				binding.descriptorCount = 1;
 				binding.stageFlags = stage_flags;
 				binding.pImmutableSamplers = 0;
diff --git a/source/backends/vulkan/texture_backend.cpp b/source/backends/vulkan/texture_backend.cpp
index 5b737498..1367cc8d 100644
--- a/source/backends/vulkan/texture_backend.cpp
+++ b/source/backends/vulkan/texture_backend.cpp
@@ -49,9 +49,13 @@ void VulkanTexture::allocate()
 	const Texture &self = *static_cast<const Texture *>(this);
 	const VulkanFunctions &vk = device.get_functions();
 
+	VkFormat vk_format = static_cast<VkFormat>(get_vulkan_pixelformat(self.storage_fmt));
+	VkFormatProperties props;
+	vk.GetPhysicalDeviceFormatProperties(vk_format, props);
+
 	VkImageCreateInfo image_info = { };
 	image_info.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
-	image_info.format = static_cast<VkFormat>(get_vulkan_pixelformat(self.storage_fmt));
+	image_info.format = vk_format;
 	image_info.extent.width = 1;
 	image_info.extent.height = 1;
 	image_info.extent.depth = 1;
@@ -63,6 +67,9 @@ void VulkanTexture::allocate()
 	image_info.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
 
 	image_info.usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT|VK_IMAGE_USAGE_TRANSFER_DST_BIT|VK_IMAGE_USAGE_SAMPLED_BIT;
+	if(props.optimalTilingFeatures&VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT)
+		image_info.usage |= VK_IMAGE_USAGE_STORAGE_BIT;
+
 	PixelComponents comp = get_components(self.storage_fmt);
 	if(comp==DEPTH_COMPONENT || comp==STENCIL_INDEX)
 		image_info.usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
diff --git a/source/core/datatype.h b/source/core/datatype.h
index d2739c49..29789907 100644
--- a/source/core/datatype.h
+++ b/source/core/datatype.h
@@ -101,6 +101,7 @@ inline bool is_float(DataType t) { return t&0x200; }
 inline bool is_matrix(DataType t) { return t&0xC000; }
 inline bool is_vector(DataType t) { return !is_matrix(t) && (t&0x3000); }
 inline bool is_image(DataType t) { return t&0x70000; }
+inline bool is_sampled_image(DataType t) { return t&0x100000; }
 
 inline DataType get_matrix_column_type(DataType t)
 {
diff --git a/source/core/device.h b/source/core/device.h
index 97a28e73..8c7e92a9 100644
--- a/source/core/device.h
+++ b/source/core/device.h
@@ -17,6 +17,7 @@ struct DeviceLimits
 	unsigned max_clip_planes = 6;
 	unsigned max_vertex_attributes = 16;
 	unsigned max_texture_bindings = 16;
+	unsigned max_storage_texture_bindings = 8;
 	unsigned max_color_attachments = 8;
 	unsigned max_samples = 4;
 	unsigned max_uniform_bindings = 24;
diff --git a/source/core/module.cpp b/source/core/module.cpp
index 5475e434..0870c478 100644
--- a/source/core/module.cpp
+++ b/source/core/module.cpp
@@ -623,11 +623,13 @@ void SpirVModule::Reflection::reflect_matrix_type(CodeIterator op)
 void SpirVModule::Reflection::reflect_image_type(CodeIterator op)
 {
 	TypeInfo &type = types[*(op+1)];
-	DataType sample = types[*(op+2)].type;
+	DataType sample_type = types[*(op+2)].type;
 	unsigned dimensions = *(op+3);
 	bool depth = *(op+4)==1;
 	bool array = *(op+5);
-	type.type = static_cast<DataType>((depth*0x200000) | (array*0x80000) | ((dimensions+1)<<16) | sample);
+	bool sampled = *(op+7)==1;
+	type.type = static_cast<DataType>((depth*0x200000) | (sampled*0x100000) | (array*0x80000) |
+		((dimensions+1)<<16) | sample_type);
 }
 
 void SpirVModule::Reflection::reflect_sampled_image_type(CodeIterator op)
diff --git a/source/core/pipelinestate.cpp b/source/core/pipelinestate.cpp
index 5ad51502..4dc54e24 100644
--- a/source/core/pipelinestate.cpp
+++ b/source/core/pipelinestate.cpp
@@ -70,11 +70,21 @@ void PipelineState::set_texture(unsigned binding, const Texture *tex, int level,
 	if(level>=0 && !can_bind_tex_level(level))
 		throw invalid_operation("PipelineState::set_texture");
 
+	set_texture_resource(binding, tex, level, samp);
+}
+
+void PipelineState::set_storage_texture(unsigned binding, const Texture *tex)
+{
+	set_texture_resource(binding, tex, 0, 0);
+}
+
+void PipelineState::set_texture_resource(unsigned binding, const Texture *tex, int level, const Sampler *samp)
+{
 	auto i = lower_bound_member(resources, static_cast<int>(binding), &BoundResource::binding);
 	if(i==resources.end() || i->binding!=static_cast<int>(binding))
 		i = resources.insert(i, BoundResource(binding));
 
-	ResourceType type = (tex ? TEXTURE : NO_RESOURCE);
+	ResourceType type = (tex ? samp ? SAMPLED_TEXTURE : STORAGE_TEXTURE : NO_RESOURCE);
 	if(i->type!=type || tex!=i->texture || level!=i->mip_level || samp!=i->sampler)
 	{
 		i->type = type;
@@ -82,7 +92,7 @@ void PipelineState::set_texture(unsigned binding, const Texture *tex, int level,
 		i->sampler = samp;
 		i->mip_level = level;
 		i->changed = true;
-		i->used = (tex && samp);
+		i->used = tex;
 		changes |= RESOURCES;
 	}
 }
diff --git a/source/core/pipelinestate.h b/source/core/pipelinestate.h
index e7fdfdc8..965bc15f 100644
--- a/source/core/pipelinestate.h
+++ b/source/core/pipelinestate.h
@@ -36,7 +36,8 @@ private:
 	{
 		NO_RESOURCE,
 		UNIFORM_BLOCK,
-		TEXTURE
+		SAMPLED_TEXTURE,
+		STORAGE_TEXTURE
 	};
 
 	struct BoundResource
@@ -98,6 +99,10 @@ public:
 	void set_uniform_block(int, const UniformBlock *);
 	void set_texture(unsigned, const Texture *, const Sampler *);
 	void set_texture(unsigned, const Texture *, int, const Sampler *);
+	void set_storage_texture(unsigned, const Texture *);
+private:
+	void set_texture_resource(unsigned, const Texture *, int, const Sampler *);
+public:
 	void set_vertex_setup(const VertexSetup *);
 	void set_primitive_type(PrimitiveType);
 	void set_front_face(FaceWinding);
diff --git a/source/render/renderer.cpp b/source/render/renderer.cpp
index f2fb002e..828efd70 100644
--- a/source/render/renderer.cpp
+++ b/source/render/renderer.cpp
@@ -181,6 +181,12 @@ void Renderer::set_texture(Tag tag, const Texture *tex, int level, const Sampler
 	set_resource(texture_stack, state.texture_count, tag, { tex, samp, level });
 }
 
+void Renderer::set_storage_texture(Tag tag, const Texture *tex)
+{
+	State &state = get_state();
+	set_resource(texture_stack, state.texture_count, tag, { tex, 0, 0 });
+}
+
 template<typename T>
 void Renderer::set_resource(vector<BoundResource<T>> &stack, unsigned &count, Tag tag, const T &res)
 {
@@ -413,7 +419,12 @@ void Renderer::apply_state()
 			if(t.binding<0 || shprog_changed)
 				t.binding = state.shprog->get_uniform_binding(t.tag);
 			if(t.binding>=0)
-				ps.set_texture(t.binding, t.resource.texture, t.resource.level, t.resource.sampler);
+			{
+				if(t.resource.sampler)
+					ps.set_texture(t.binding, t.resource.texture, t.resource.level, t.resource.sampler);
+				else
+					ps.set_storage_texture(t.binding, t.resource.texture);
+			}
 		}
 
 	static const DepthTest default_depth_test;
diff --git a/source/render/renderer.h b/source/render/renderer.h
index 9335164b..fb460cf0 100644
--- a/source/render/renderer.h
+++ b/source/render/renderer.h
@@ -197,6 +197,7 @@ public:
 
 	void set_texture(Tag, const Texture *, const Sampler * = 0);
 	void set_texture(Tag, const Texture *, int, const Sampler * = 0);
+	void set_storage_texture(Tag, const Texture *);
 
 private:
 	template<typename T>