]> git.tdb.fi Git - libs/gl.git/blobdiff - source/backends/vulkan/pipelinestate_backend.cpp
Support compute shaders and compute operations
[libs/gl.git] / source / backends / vulkan / pipelinestate_backend.cpp
index b7e50873aec14f1cb16a841916adb5a6ccfb97cd..7af36eedf82ceadced50b3f576b103ceeff584f8 100644 (file)
@@ -48,37 +48,36 @@ void VulkanPipelineState::update() const
                push_const_compat = hash_update<32>(push_const_compat, self.shprog->get_push_constants_size());
        }
 
-       constexpr unsigned pipeline_mask = PipelineState::SHPROG|PipelineState::VERTEX_SETUP|PipelineState::FACE_CULL|
+       constexpr unsigned graphics_mask = PipelineState::VERTEX_SETUP|PipelineState::FACE_CULL|
                PipelineState::DEPTH_TEST|PipelineState::STENCIL_TEST|PipelineState::BLEND|PipelineState::PRIMITIVE_TYPE;
+       unsigned pipeline_mask = PipelineState::SHPROG;
+       if(!self.shprog->is_compute())
+               pipeline_mask |= graphics_mask;
        if(changes&pipeline_mask)
        {
                handle = device.get_pipeline_cache().get_pipeline(self);
                unapplied |= PipelineState::SHPROG;
        }
 
-       if(changes&(PipelineState::SHPROG|PipelineState::UNIFORMS|PipelineState::TEXTURES))
+       if(changes&(PipelineState::SHPROG|PipelineState::RESOURCES))
        {
                unsigned changed_sets = (changes&PipelineState::SHPROG ? ~0U : 0U);
-               for(const PipelineState::BoundUniformBlock &u: self.uniform_blocks)
-                       if(u.changed || changed_sets==~0U)
+               for(const PipelineState::BoundResource &r: self.resources)
+                       if(r.changed || changed_sets==~0U)
                        {
-                               if(u.block)
-                                       u.used = self.shprog->uses_uniform_block_binding(u.binding);
-                               if(u.binding>=0)
-                                       changed_sets |= 1<<(u.binding>>20);
-                               u.changed = false;
-                       }
-               for(const PipelineState::BoundTexture &t: self.textures)
-                       if(t.changed || changed_sets==~0U)
-                       {
-                               if(t.texture && t.sampler)
-                                       t.used = self.shprog->uses_texture_binding(t.binding);
-                               changed_sets |= 1<<(t.binding>>20);
-                               if(t.texture && t.level>=0)
-                                       t.texture->refresh_mip_views();
-                               if(t.sampler)
-                                       t.sampler->refresh();
-                               t.changed = false;
+                               if(r.type==PipelineState::UNIFORM_BLOCK)
+                                       r.used = self.shprog->uses_uniform_block_binding(r.binding);
+                               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();
+                                       if(r.type==PipelineState::SAMPLED_TEXTURE)
+                                               r.sampler->refresh();
+                               }
+                               if(r.binding>=0)
+                                       changed_sets |= 1<<(r.binding>>20);
+                               r.changed = false;
                        }
 
                if(changed_sets)
@@ -92,7 +91,7 @@ void VulkanPipelineState::update() const
                                        first_changed_desc_set = min(first_changed_desc_set, i);
                                }
 
-                       unapplied |= PipelineState::UNIFORMS;
+                       unapplied |= PipelineState::RESOURCES;
                }
        }
 
@@ -102,50 +101,63 @@ void VulkanPipelineState::update() const
 uint64_t VulkanPipelineState::compute_hash() const
 {
        const PipelineState &self = *static_cast<const PipelineState *>(this);
-       const FrameFormat &format = self.framebuffer->get_format();
 
        uint64_t result = hash<64>(self.shprog);
-       result = hash_update<64>(result, self.vertex_setup->compute_hash());
-       result = hash_round<64>(result, self.primitive_type);
 
-       if(self.front_face!=NON_MANIFOLD && self.face_cull!=NO_CULL)
+       if(!self.shprog->is_compute())
        {
-               result = hash_round<64>(result, self.front_face);
-               result = hash_round<64>(result, self.face_cull);
-       }
+               const FrameFormat &format = self.framebuffer->get_format();
 
-       result = hash_round<64>(result, format.get_samples());
+               result = hash_update<64>(result, self.vertex_setup->compute_hash());
+               result = hash_round<64>(result, self.primitive_type);
 
-       if(self.depth_test.enabled)
-       {
-               result = hash_round<64>(result, self.depth_test.compare);
-               result = hash_update<64>(result, self.depth_test.write);
-       }
+               if(self.front_face!=NON_MANIFOLD && self.face_cull!=NO_CULL)
+               {
+                       result = hash_round<64>(result, self.front_face);
+                       result = hash_round<64>(result, self.face_cull);
+               }
 
-       if(self.stencil_test.enabled)
-       {
-               result = hash_round<64>(result, self.stencil_test.compare);
-               result = hash_round<64>(result, self.stencil_test.stencil_fail_op);
-               result = hash_round<64>(result, self.stencil_test.depth_fail_op);
-               result = hash_round<64>(result, self.stencil_test.depth_pass_op);
-               result = hash_update<64>(result, self.stencil_test.reference);
-       }
+               result = hash_round<64>(result, format.get_samples());
 
-       if(self.blend.enabled)
-       {
-               result = hash_round<64>(result, self.blend.equation);
-               result = hash_round<64>(result, self.blend.src_factor);
-               result = hash_round<64>(result, self.blend.dst_factor);
-               result = hash_round<64>(result, self.blend.write_mask);
-       }
+               if(self.depth_test.enabled)
+               {
+                       result = hash_round<64>(result, self.depth_test.compare);
+                       result = hash_update<64>(result, self.depth_test.write);
+               }
 
-       for(FrameAttachment a: format)
-               result = hash_update<64>(result, a);
+               if(self.stencil_test.enabled)
+               {
+                       result = hash_round<64>(result, self.stencil_test.compare);
+                       result = hash_round<64>(result, self.stencil_test.stencil_fail_op);
+                       result = hash_round<64>(result, self.stencil_test.depth_fail_op);
+                       result = hash_round<64>(result, self.stencil_test.depth_pass_op);
+                       result = hash_update<64>(result, self.stencil_test.reference);
+               }
+
+               if(self.blend.enabled)
+               {
+                       result = hash_round<64>(result, self.blend.equation);
+                       result = hash_round<64>(result, self.blend.src_factor);
+                       result = hash_round<64>(result, self.blend.dst_factor);
+                       result = hash_round<64>(result, self.blend.write_mask);
+               }
+
+               for(FrameAttachment a: format)
+                       result = hash_update<64>(result, a);
+       }
 
        return result;
 }
 
 void VulkanPipelineState::fill_creation_info(vector<char> &buffer) const
+{
+       if(static_cast<const PipelineState *>(this)->shprog->is_compute())
+               fill_compute_creation_info(buffer);
+       else
+               fill_graphics_creation_info(buffer);
+}
+
+void VulkanPipelineState::fill_graphics_creation_info(vector<char> &buffer) const
 {
        const PipelineState &self = *static_cast<const PipelineState *>(this);
 
@@ -270,6 +282,22 @@ void VulkanPipelineState::fill_creation_info(vector<char> &buffer) const
                pipeline_info->pVertexInputState = reinterpret_cast<const VkPipelineVertexInputStateCreateInfo *>(self.vertex_setup->creation_info.data());
 }
 
+void VulkanPipelineState::fill_compute_creation_info(vector<char> &buffer) const
+{
+       const PipelineState &self = *static_cast<const PipelineState *>(this);
+
+       StructureBuilder sb(buffer, 1);
+       VkComputePipelineCreateInfo *const &pipeline_info = sb.add<VkComputePipelineCreateInfo>();
+
+       pipeline_info->sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO;
+
+       if(self.shprog)
+       {
+               pipeline_info->stage = *reinterpret_cast<const VkPipelineShaderStageCreateInfo *>(self.shprog->creation_info.data());
+               pipeline_info->layout = handle_cast<::VkPipelineLayout>(self.shprog->layout_handle);
+       }
+}
+
 uint64_t VulkanPipelineState::compute_descriptor_set_hash(unsigned index) const
 {
        const PipelineState &self = *static_cast<const PipelineState *>(this);
@@ -277,26 +305,29 @@ uint64_t VulkanPipelineState::compute_descriptor_set_hash(unsigned index) const
        uint64_t result = hash<64>(0, 0);
        bool empty = true;
 
-       auto i = lower_bound_member(self.uniform_blocks, static_cast<int>(index)<<20, &PipelineState::BoundUniformBlock::binding);
-       for(; (i!=self.uniform_blocks.end() && static_cast<unsigned>(i->binding)>>20==index); ++i)
-               if(i->used)
+       auto i = lower_bound_member(self.resources, static_cast<int>(index)<<20, &PipelineState::BoundResource::binding);
+       for(; (i!=self.resources.end() && static_cast<unsigned>(i->binding)>>20==index); ++i)
+       {
+               if(!i->used)
+                       continue;
+
+               result = hash_update<64>(result, i->binding);
+               result = hash_update<64>(result, i->type);
+               if(i->type==PipelineState::UNIFORM_BLOCK)
                {
-                       result = hash_update<64>(result, i->binding);
                        result = hash_update<64>(result, reinterpret_cast<uintptr_t>(i->block));
                        result = hash_update<64>(result, reinterpret_cast<uintptr_t>(i->buffer->handle));
-                       empty = false;
                }
-
-       auto j = lower_bound_member(self.textures, index<<20, &PipelineState::BoundTexture::binding);
-       for(; (j!=self.textures.end() && j->binding>>20==index); ++j)
-               if(j->used)
+               else if(i->type==PipelineState::SAMPLED_TEXTURE)
                {
-                       result = hash_update<64>(result, j->binding);
-                       result = hash_update<64>(result, reinterpret_cast<uintptr_t>(j->texture->handle));
-                       result = hash_update<64>(result, reinterpret_cast<uintptr_t>(j->sampler->handle));
-                       result = hash_update<64>(result, j->level);
-                       empty = false;
+                       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;
+       }
 
        if(!empty)
                result = hash_update<64>(result, self.shprog->stage_flags);
@@ -308,9 +339,9 @@ bool VulkanPipelineState::is_descriptor_set_dynamic(unsigned index) const
 {
        const PipelineState &self = *static_cast<const PipelineState *>(this);
 
-       auto i = lower_bound_member(self.uniform_blocks, static_cast<int>(index)<<20, &PipelineState::BoundUniformBlock::binding);
-       for(; (i!=self.uniform_blocks.end() && static_cast<unsigned>(i->binding)>>20==index); ++i)
-               if(i->used && i->buffer && i->buffer->get_usage()==STREAMING)
+       auto i = lower_bound_member(self.resources, static_cast<int>(index)<<20, &PipelineState::BoundResource::binding);
+       for(; (i!=self.resources.end() && static_cast<unsigned>(i->binding)>>20==index); ++i)
+               if(i->used && i->type==PipelineState::UNIFORM_BLOCK && i->buffer->get_usage()==STREAMING)
                        return true;
 
        return false;
@@ -325,17 +356,18 @@ unsigned VulkanPipelineState::fill_descriptor_writes(unsigned index, unsigned fr
 {
        const PipelineState &self = *static_cast<const PipelineState *>(this);
 
-       auto u_begin = lower_bound_member(self.uniform_blocks, static_cast<int>(index)<<20, &PipelineState::BoundUniformBlock::binding);
-       auto t_begin = lower_bound_member(self.textures, index<<20, &PipelineState::BoundTexture::binding);
+       auto begin = lower_bound_member(self.resources, static_cast<int>(index)<<20, &PipelineState::BoundResource::binding);
 
        unsigned n_buffers = 0;
-       for(auto i=u_begin; (i!=self.uniform_blocks.end() && static_cast<unsigned>(i->binding)>>20==index); ++i)
-               if(i->used)
-                       ++n_buffers;
        unsigned n_images = 0;
-       for(auto i=t_begin; (i!=self.textures.end() && i->binding>>20==index); ++i)
+       for(auto i=begin; (i!=self.resources.end() && static_cast<unsigned>(i->binding)>>20==index); ++i)
                if(i->used)
-                       ++n_images;
+               {
+                       if(i->type==PipelineState::UNIFORM_BLOCK)
+                               ++n_buffers;
+                       else if(i->type==PipelineState::SAMPLED_TEXTURE || i->type==PipelineState::STORAGE_TEXTURE)
+                               ++n_images;
+               }
        unsigned n_writes = n_buffers+n_images;
 
        StructureBuilder sb(buffer, 3);
@@ -347,8 +379,12 @@ unsigned VulkanPipelineState::fill_descriptor_writes(unsigned index, unsigned fr
        VkDescriptorBufferInfo *buffer_ptr = buffers;
        VkDescriptorImageInfo *image_ptr = images;
 
-       for(auto i=u_begin; (i!=self.uniform_blocks.end() && static_cast<unsigned>(i->binding)>>20==index); ++i)
-               if(i->used)
+       for(auto i=begin; (i!=self.resources.end() && static_cast<unsigned>(i->binding)>>20==index); ++i)
+       {
+               if(!i->used)
+                       continue;
+
+               if(i->type==PipelineState::UNIFORM_BLOCK)
                {
                        buffer_ptr->buffer = handle_cast<::VkBuffer>(i->buffer->handle);
                        buffer_ptr->offset = i->block->get_offset();
@@ -363,32 +399,49 @@ unsigned VulkanPipelineState::fill_descriptor_writes(unsigned index, unsigned fr
                        write_ptr->pBufferInfo = buffer_ptr;
 
                        ++buffer_ptr;
-                       ++write_ptr;
                }
-
-       for(auto i=t_begin; (i!=self.textures.end() && i->binding>>20==index); ++i)
-               if(i->used)
+               else if(i->type==PipelineState::SAMPLED_TEXTURE || i->type==PipelineState::STORAGE_TEXTURE)
                {
-                       image_ptr->sampler = handle_cast<::VkSampler>(i->sampler->handle);
-                       if(i->level<0)
+                       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->level]);
-                       image_ptr->imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+                               image_ptr->imageView = handle_cast<::VkImageView>(i->texture->mip_view_handles[i->mip_level]);
+
+                       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;
-                       ++write_ptr;
                }
 
+               ++write_ptr;
+       }
+
        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);
@@ -406,7 +459,7 @@ void VulkanPipelineState::apply(const VulkanCommandRecorder &vkCmd, const Vulkan
                        unapplied |= PipelineState::SHPROG;
                        if(self.push_const_compat!=last_ps.push_const_compat)
                        {
-                               unapplied |= PipelineState::UNIFORMS;
+                               unapplied |= PipelineState::RESOURCES;
                                first_changed_desc_set = 0;
                        }
                }
@@ -415,7 +468,7 @@ void VulkanPipelineState::apply(const VulkanCommandRecorder &vkCmd, const Vulkan
                for(unsigned i=0; i<descriptor_set_slots.size(); ++i)
                        if(i>=last->descriptor_set_slots.size() || descriptor_set_slots[i]!=last->descriptor_set_slots[i])
                        {
-                               unapplied |= PipelineState::UNIFORMS;
+                               unapplied |= PipelineState::RESOURCES;
                                first_changed_desc_set = min(first_changed_desc_set, i);
                                break;
                        }
@@ -425,10 +478,11 @@ void VulkanPipelineState::apply(const VulkanCommandRecorder &vkCmd, const Vulkan
                        unapplied |= PipelineState::SCISSOR;
        }
 
+       VkPipelineBindPoint bind_point = (self.shprog->is_compute() ? VK_PIPELINE_BIND_POINT_COMPUTE : VK_PIPELINE_BIND_POINT_GRAPHICS);
        if(unapplied&PipelineState::SHPROG)
-               vkCmd.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, handle);
+               vkCmd.BindPipeline(bind_point, handle);
 
-       if(unapplied&PipelineState::VERTEX_SETUP)
+       if(!self.shprog->is_compute() && (unapplied&PipelineState::VERTEX_SETUP))
                if(const VertexSetup *vs = self.vertex_setup)
                {
                        vkCmd.BindVertexBuffers(0, vs->n_bindings, vs->buffers, vs->offsets);
@@ -436,18 +490,18 @@ void VulkanPipelineState::apply(const VulkanCommandRecorder &vkCmd, const Vulkan
                        vkCmd.BindIndexBuffer(vs->get_index_buffer()->handle, 0, index_type);
                }
 
-       if(!self.uniform_blocks.empty())
+       if(!self.resources.empty())
        {
-               const PipelineState::BoundUniformBlock &first_block = self.uniform_blocks.front();
-               if(first_block.used && first_block.binding==ReflectData::PUSH_CONSTANT)
+               const PipelineState::BoundResource &first_res = self.resources.front();
+               if(first_res.used && first_res.type==PipelineState::UNIFORM_BLOCK && first_res.binding==ReflectData::PUSH_CONSTANT)
                {
-                       const UniformBlock &pc_block = *first_block.block;
+                       const UniformBlock &pc_block = *first_res.block;
                        vkCmd.PushConstants(self.shprog->layout_handle, self.shprog->stage_flags,
                                pc_block.get_offset(), pc_block.get_data_size(), pc_block.get_data_pointer());
                }
        }
 
-       if((unapplied&PipelineState::UNIFORMS) && !descriptor_set_slots.empty())
+       if((unapplied&PipelineState::RESOURCES) && !descriptor_set_slots.empty())
        {
                vector<VkDescriptorSet> descriptor_set_handles;
                descriptor_set_handles.reserve(descriptor_set_slots.size()-first_changed_desc_set);
@@ -455,11 +509,11 @@ void VulkanPipelineState::apply(const VulkanCommandRecorder &vkCmd, const Vulkan
                        descriptor_set_handles.push_back(device.get_descriptor_pool().get_descriptor_set(
                                self.descriptor_set_slots[i], self, i, frame));
 
-               vkCmd.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, self.shprog->layout_handle,
+               vkCmd.BindDescriptorSets(bind_point, self.shprog->layout_handle,
                        first_changed_desc_set, descriptor_set_handles.size(), descriptor_set_handles.data(), 0, 0);
        }
 
-       if(unapplied&(PipelineState::VIEWPORT|PipelineState::SCISSOR))
+       if(!self.shprog->is_compute() && (unapplied&(PipelineState::VIEWPORT|PipelineState::SCISSOR)))
        {
                Rect fb_rect = self.framebuffer->get_rect();