Move skymap processing to GPU

sergcpp · May 30, 2024 · e165a51 · e165a51
1 parent 8897b4c
commit e165a51
Show file tree

Hide file tree

Showing 24 changed files with 268 additions and 92 deletions.
diff --git a/internal/AtmosphereRef.cpp b/internal/AtmosphereRef.cpp
@@ -210,7 +210,7 @@ fvec4 SampleTransmittanceLUT(Span<const float> lut, fvec2 uv) {
 }
 
 fvec4 SampleMultiscatterLUT(Span<const float> lut, fvec2 uv) {
-    uv = uv * fvec2(SKY_MULTISCATTER_LUT_RES);
+    uv = fract(uv - 0.5f / SKY_MULTISCATTER_LUT_RES) * fvec2(SKY_MULTISCATTER_LUT_RES);
     auto iuv0 = ivec2(uv);
     iuv0 = clamp(iuv0, ivec2{0, 0}, ivec2{SKY_MULTISCATTER_LUT_RES - 1});
     const ivec2 iuv1 = min(iuv0 + 1, ivec2{SKY_MULTISCATTER_LUT_RES - 1});

diff --git a/internal/Dx/ContextDX.cpp b/internal/Dx/ContextDX.cpp
@@ -432,6 +432,8 @@ void Ray::Dx::EndSingleTimeCommands(void *api, ID3D12Device *device, ID3D12Comma
     command_list->Release();
 }
 
+void Ray::Dx::InsertReadbackMemoryBarrier(void *api, CommandBuffer command_list) {}
+
 int Ray::Dx::Context::WriteTimestamp(ID3D12GraphicsCommandList *cmd_buf, const bool start) {
     const uint32_t query_index = query_counts_[backend_frame]++;
     assert(query_index < MaxTimestampQueries);

diff --git a/internal/Dx/ContextDX.h b/internal/Dx/ContextDX.h
@@ -169,5 +169,7 @@ CommandBuffer BegSingleTimeCommands(void *api, ID3D12Device *device, ID3D12Comma
 void EndSingleTimeCommands(void *api, ID3D12Device *device, ID3D12CommandQueue *cmd_queue, CommandBuffer command_list,
                            ID3D12CommandAllocator *temp_command_allocator);
 
+void InsertReadbackMemoryBarrier(void *api, CommandBuffer command_list);
+
 } // namespace Dx
 } // namespace Ray
diff --git a/internal/RendererDX.cpp b/internal/RendererDX.cpp
@@ -477,8 +477,6 @@ Ray::Dx::Renderer::Renderer(const settings_t &s, ILog *log) {
     indir_args_buf_[1] =
         Buffer{"Indir Args (2/2)", ctx_.get(), eBufType::Indirect, 32 * sizeof(DispatchIndirectCommand)};
 
-    atmosphere_params_buf_ = Buffer{"Atmosphere Params", ctx_.get(), eBufType::Uniform, sizeof(atmosphere_params_t)};
-
     { // zero out counters
         CommandBuffer cmd_buf = BegSingleTimeCommands(ctx_->api(), ctx_->device(), ctx_->temp_command_pool());
 
@@ -666,17 +664,6 @@ void Ray::Dx::Renderer::RenderScene(const SceneBase &scene, RegionContext &regio
     }
 #endif
 
-    if (s.env_.sky_map_spread_angle > 0.0f) {
-        Buffer temp_upload_buf{"Temp atmosphere params upload", ctx_.get(), eBufType::Upload,
-                               sizeof(atmosphere_params_t)};
-        { // update stage buffer
-            uint8_t *mapped_ptr = temp_upload_buf.Map();
-            memcpy(mapped_ptr, &s.env_.atmosphere, sizeof(atmosphere_params_t));
-            temp_upload_buf.Unmap();
-        }
-        CopyBufferToBuffer(temp_upload_buf, 0, atmosphere_params_buf_, 0, sizeof(atmosphere_params_t), cmd_buf);
-    }
-
     //////////////////////////////////////////////////////////////////////////////////
 
     const scene_data_t sc_data = {s.env_,
@@ -705,7 +692,7 @@ void Ray::Dx::Renderer::RenderScene(const SceneBase &scene, RegionContext &regio
                                   cache_grid_params,
                                   s.spatial_cache_entries_.buf(),
                                   s.spatial_cache_voxels_prev_.buf(),
-                                  atmosphere_params_buf_,
+                                  s.atmosphere_params_buf_,
                                   s.sky_transmittance_lut_tex_,
                                   s.sky_multiscatter_lut_tex_,
                                   s.sky_moon_tex_,
@@ -1411,7 +1398,7 @@ void Ray::Dx::Renderer::UpdateSpatialCache(const SceneBase &scene, RegionContext
                                   cache_grid_params,
                                   s.spatial_cache_entries_.buf(),
                                   s.spatial_cache_voxels_curr_.buf(),
-                                  atmosphere_params_buf_,
+                                  s.atmosphere_params_buf_,
                                   {},
                                   {},
                                   {},

diff --git a/internal/RendererGPU.h b/internal/RendererGPU.h
@@ -128,7 +128,6 @@ class Renderer : public RendererBase {
         count_table_buf_, reduce_table_buf_;
     Buffer counters_buf_, indir_args_buf_[2];
 
-    Buffer atmosphere_params_buf_;
     Buffer temp_cache_data_buf_, temp_lock_buf_;
 
     Buffer pixel_readback_buf_, base_color_readback_buf_, depth_normals_readback_buf_;

diff --git a/internal/RendererGPU_kernels.h b/internal/RendererGPU_kernels.h
@@ -826,7 +826,7 @@ inline void Ray::NS::Renderer::kernel_ShadeSky(CommandBuffer cmd_buf, const pass
                                               {&out_img, eResState::UnorderedAccess}};
     TransitionResourceStates(cmd_buf, AllStages, AllStages, res_transitions);
 
-    SmallVector<Binding, 32> bindings = {
+    const Binding bindings[] = {
         {eBindTarget::UBuf, ShadeSky::ATMOSPHERE_PARAMS_BUF_SLOT, sc_data.atmosphere_params},
         {eBindTarget::SBufRO, ShadeSky::RAY_INDICES_BUF_SLOT, ray_indices},
         {eBindTarget::SBufRO, ShadeSky::HITS_BUF_SLOT, hits},

diff --git a/internal/RendererVK.cpp b/internal/RendererVK.cpp
@@ -512,8 +512,6 @@ Ray::Vk::Renderer::Renderer(const settings_t &s, ILog *log) {
     indir_args_buf_[1] =
         Buffer{"Indir Args (2/2)", ctx_.get(), eBufType::Indirect, 32 * sizeof(DispatchIndirectCommand)};
 
-    atmosphere_params_buf_ = Buffer{"Atmosphere Params", ctx_.get(), eBufType::Uniform, sizeof(atmosphere_params_t)};
-
     { // zero out counters, upload random sequence
         CommandBuffer cmd_buf = BegSingleTimeCommands(ctx_->api(), ctx_->device(), ctx_->temp_command_pool());
 
@@ -680,17 +678,6 @@ void Ray::Vk::Renderer::RenderScene(const SceneBase &scene, RegionContext &regio
 
     ctx_->api().vkCmdResetQueryPool(cmd_buf, ctx_->query_pool(ctx_->backend_frame), 0, MaxTimestampQueries);
 
-    if (s.env_.sky_map_spread_angle > 0.0f) {
-        Buffer temp_upload_buf{"Temp atmosphere params upload", ctx_.get(), eBufType::Upload,
-                               sizeof(atmosphere_params_t)};
-        { // update stage buffer
-            uint8_t *mapped_ptr = temp_upload_buf.Map();
-            memcpy(mapped_ptr, &s.env_.atmosphere, sizeof(atmosphere_params_t));
-            temp_upload_buf.Unmap();
-        }
-        CopyBufferToBuffer(temp_upload_buf, 0, atmosphere_params_buf_, 0, sizeof(atmosphere_params_t), cmd_buf);
-    }
-
     //////////////////////////////////////////////////////////////////////////////////
 
     const scene_data_t sc_data = {s.env_,
@@ -719,14 +706,13 @@ void Ray::Vk::Renderer::RenderScene(const SceneBase &scene, RegionContext &regio
                                   cache_grid_params,
                                   s.spatial_cache_entries_.buf(),
                                   s.spatial_cache_voxels_prev_.buf(),
-                                  atmosphere_params_buf_,
+                                  s.atmosphere_params_buf_,
                                   s.sky_transmittance_lut_tex_,
                                   s.sky_multiscatter_lut_tex_,
                                   s.sky_moon_tex_,
                                   s.sky_weather_tex_,
                                   s.sky_cirrus_tex_,
                                   s.sky_noise3d_tex_};
-
     TransitionSceneResources(cmd_buf, sc_data);
 
     VkMemoryBarrier mem_barrier = {VK_STRUCTURE_TYPE_MEMORY_BARRIER};
@@ -1456,7 +1442,7 @@ void Ray::Vk::Renderer::UpdateSpatialCache(const SceneBase &scene, RegionContext
                                   cache_grid_params,
                                   s.spatial_cache_entries_.buf(),
                                   s.spatial_cache_voxels_curr_.buf(),
-                                  atmosphere_params_buf_,
+                                  s.atmosphere_params_buf_,
                                   {},
                                   {},
                                   {},
@@ -1740,12 +1726,7 @@ Ray::color_data_rgba_t Ray::Vk::Renderer::get_pixels_ref(const bool tonemap) con
             CopyImageToBuffer(buffer_to_use, 0, 0, 0, w_, h_, pixel_readback_buf_, cmd_buf, 0);
         }
 
-        VkMemoryBarrier mem_barrier = {VK_STRUCTURE_TYPE_MEMORY_BARRIER};
-        mem_barrier.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
-        mem_barrier.dstAccessMask = VK_ACCESS_HOST_READ_BIT;
-
-        ctx_->api().vkCmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_HOST_BIT, 0, 1,
-                                         &mem_barrier, 0, nullptr, 0, nullptr);
+        InsertReadbackMemoryBarrier(ctx_->api(), cmd_buf);
 
 #if RUN_IN_LOCKSTEP
         EndSingleTimeCommands(ctx_->api(), ctx_->device(), ctx_->graphics_queue(), cmd_buf, ctx_->temp_command_pool());
@@ -1809,12 +1790,7 @@ Ray::color_data_rgba_t Ray::Vk::Renderer::get_aux_pixels_ref(const eAUXBuffer bu
             CopyImageToBuffer(buffer_to_use, 0, 0, 0, w_, h_, readback_buffer_to_use, cmd_buf, 0);
         }
 
-        VkMemoryBarrier mem_barrier = {VK_STRUCTURE_TYPE_MEMORY_BARRIER};
-        mem_barrier.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
-        mem_barrier.dstAccessMask = VK_ACCESS_HOST_READ_BIT;
-
-        ctx_->api().vkCmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_HOST_BIT, 0, 1,
-                                         &mem_barrier, 0, nullptr, 0, nullptr);
+        InsertReadbackMemoryBarrier(ctx_->api(), cmd_buf);
 
 #if RUN_IN_LOCKSTEP
         EndSingleTimeCommands(ctx_->api(), ctx_->device(), ctx_->graphics_queue(), cmd_buf, ctx_->temp_command_pool());

diff --git a/internal/SceneCPU.cpp b/internal/SceneCPU.cpp
@@ -1038,8 +1038,8 @@ void Ray::Cpu::Scene::PrepareEnvMapQTree_nolock() {
             for (int qx = 0; qx < cur_res; ++qx) {
                 for (int jj = -2; jj <= 2; ++jj) {
                     for (int ii = -2; ii <= 2; ++ii) {
-                        const Ref::fvec2 q = {(float(qx) + 0.5f + ii * FilterSize) / cur_res,
-                                              (float(qy) + 0.5f + jj * FilterSize) / cur_res};
+                        const Ref::fvec2 q = {Ref::fract(1.0f + (float(qx) + 0.5f + ii * FilterSize) / cur_res),
+                                              Ref::fract(1.0f + (float(qy) + 0.5f + jj * FilterSize) / cur_res)};
                         Ref::fvec4 dir;
                         CanonicalToDir(value_ptr(q), 0.0f, value_ptr(dir));
 

diff --git a/internal/SceneCommon.cpp b/internal/SceneCommon.cpp
@@ -294,7 +294,7 @@ Ray::SceneCommon::CalcSkyEnvTexture(const atmosphere_params_t &params, const int
     parallel_for(0, res[1], [&](const int y) {
         const float theta = PI * float(y) / float(res[1]);
         for (int x = 0; x < res[0]; ++x) {
-            const uint32_t px_hash = Ref::hash((x << 16)| x);
+            const uint32_t px_hash = Ref::hash((x << 16)| y);
 
             const float phi = 2.0f * PI * (x + 0.5f) / float(res[0]);
             auto ray_dir = Ref::fvec4{sinf(theta) * cosf(phi), cosf(theta), sinf(theta) * sinf(phi), 0.0f};

diff --git a/internal/SceneDX.cpp b/internal/SceneDX.cpp
@@ -21,6 +21,7 @@
 #include "Dx/ContextDX.h"
 #include "TextureParams.h"
 #include "TextureUtils.h"
+#include "inflate/Inflate.h"
 
 namespace Ray {
 uint32_t next_power_of_two(uint32_t v);
@@ -35,7 +36,11 @@ void to_dxr_xform(const float xform[16], float matrix[3][4]) {
     }
 }
 
+namespace Dx {
+#include "shaders/output/bake_sky.comp.cso.inl"
+
 const uint32_t AccStructAlignment = D3D12_RAYTRACING_ACCELERATION_STRUCTURE_BYTE_ALIGNMENT;
+} // namespace Dx
 } // namespace Ray
 
 Ray::Dx::Scene::~Scene() {
@@ -60,6 +65,13 @@ Ray::Dx::Scene::~Scene() {
     bindless_textures_.clear();
 }
 
+bool Ray::Dx::Scene::InitPipelines() {
+    sh_bake_sky_ =
+        Shader{"Bake Sky", ctx_, Inflate(internal_shaders_output_bake_sky_comp_cso), eShaderType::Comp, log_};
+    prog_bake_sky_ = Program{"Bake Sky", ctx_, &sh_bake_sky_, log_};
+    return pi_bake_sky_.Init(ctx_, &prog_bake_sky_, log_);
+}
+
 void Ray::Dx::Scene::GenerateTextureMips_nolock() {}
 
 void Ray::Dx::Scene::PrepareBindlessTextures_nolock() {

diff --git a/internal/SceneDX.h b/internal/SceneDX.h
@@ -2,6 +2,10 @@
 
 #include "Dx/AccStructureDX.h"
 #include "Dx/DescriptorPoolDX.h"
+#include "Dx/DrawCallDX.h"
+#include "Dx/PipelineDX.h"
+#include "Dx/ProgramDX.h"
+#include "Dx/ShaderDX.h"
 #include "Dx/SparseStorageDX.h"
 #include "Dx/TextureAtlasDX.h"
 #include "Dx/TextureDX.h"