From 2ce5bb9bd6cddb83810eab98f54b0e4120e55c20 Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Sat, 25 Feb 2023 16:24:21 -0500 Subject: [PATCH] buffer_cache: Add logic for non-NVN storage buffer tracking --- .../global_memory_to_storage_buffer_pass.cpp | 8 ++++++- src/video_core/buffer_cache/buffer_cache.h | 24 +++++++++++++++---- 2 files changed, 27 insertions(+), 5 deletions(-) diff --git a/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp b/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp index 336338e628..d1e59f22eb 100644 --- a/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp +++ b/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp @@ -35,6 +35,7 @@ struct Bias { u32 index; u32 offset_begin; u32 offset_end; + u32 alignment; }; using boost::container::flat_set; @@ -349,7 +350,8 @@ std::optional Track(const IR::Value& value, const Bias* bias) .index = index.U32(), .offset = offset.U32(), }; - if (!Common::IsAligned(storage_buffer.offset, 16)) { + const u32 alignment{bias ? bias->alignment : 8U}; + if (!Common::IsAligned(storage_buffer.offset, alignment)) { // The SSBO pointer has to be aligned return std::nullopt; } @@ -371,6 +373,7 @@ void CollectStorageBuffers(IR::Block& block, IR::Inst& inst, StorageInfo& info) .index = 0, .offset_begin = 0x110, .offset_end = 0x610, + .alignment = 16, }; // Track the low address of the instruction const std::optional low_addr_info{TrackLowAddress(&inst)}; @@ -386,8 +389,11 @@ void CollectStorageBuffers(IR::Block& block, IR::Inst& inst, StorageInfo& info) storage_buffer = Track(low_addr, nullptr); if (!storage_buffer) { // If that also fails, use NVN fallbacks + LOG_WARNING(Shader, "Storage buffer failed to track, using global memory fallbacks"); return; } + LOG_WARNING(Shader, "Storage buffer tracked without bias, index {} offset {}", + storage_buffer->index, storage_buffer->offset); } // Collect storage buffer and the instruction if (IsGlobalMemoryWrite(inst)) { diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h index 06fd40851a..20faa65daa 100644 --- a/src/video_core/buffer_cache/buffer_cache.h +++ b/src/video_core/buffer_cache/buffer_cache.h @@ -366,7 +366,8 @@ private: void NotifyBufferDeletion(); - [[nodiscard]] Binding StorageBufferBinding(GPUVAddr ssbo_addr, bool is_written = false) const; + [[nodiscard]] Binding StorageBufferBinding(GPUVAddr ssbo_addr, u32 cbuf_index, + bool is_written = false) const; [[nodiscard]] TextureBufferBinding GetTextureBufferBinding(GPUVAddr gpu_addr, u32 size, PixelFormat format); @@ -749,7 +750,7 @@ void BufferCache

::BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index, const auto& cbufs = maxwell3d->state.shader_stages[stage]; const GPUVAddr ssbo_addr = cbufs.const_buffers[cbuf_index].address + cbuf_offset; - storage_buffers[stage][ssbo_index] = StorageBufferBinding(ssbo_addr, is_written); + storage_buffers[stage][ssbo_index] = StorageBufferBinding(ssbo_addr, cbuf_index, is_written); } template @@ -789,7 +790,7 @@ void BufferCache

::BindComputeStorageBuffer(size_t ssbo_index, u32 cbuf_index, const auto& cbufs = launch_desc.const_buffer_config; const GPUVAddr ssbo_addr = cbufs[cbuf_index].Address() + cbuf_offset; - compute_storage_buffers[ssbo_index] = StorageBufferBinding(ssbo_addr, is_written); + compute_storage_buffers[ssbo_index] = StorageBufferBinding(ssbo_addr, cbuf_index, is_written); } template @@ -1935,11 +1936,26 @@ void BufferCache

::NotifyBufferDeletion() { template typename BufferCache

::Binding BufferCache

::StorageBufferBinding(GPUVAddr ssbo_addr, + u32 cbuf_index, bool is_written) const { const GPUVAddr gpu_addr = gpu_memory->Read(ssbo_addr); - const u32 size = gpu_memory->Read(ssbo_addr + 8); + const auto size = [&]() { + const bool is_nvn_cbuf = cbuf_index == 0; + // The NVN driver buffer (index 0) is known to pack the SSBO address followed by its size. + if (is_nvn_cbuf) { + return gpu_memory->Read(ssbo_addr + 8); + } + // Other titles (notably Doom Eternal) may use STG/LDG on buffer addresses in custom defined + // cbufs, which do not store the sizes adjacent to the addresses, so use the fully + // mapped buffer size for now. + const u32 memory_layout_size = static_cast(gpu_memory->GetMemoryLayoutSize(gpu_addr)); + LOG_INFO(HW_GPU, "Binding storage buffer for cbuf index {}, MemoryLayoutSize 0x{:X}", + cbuf_index, memory_layout_size); + return memory_layout_size; + }(); const std::optional cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr); if (!cpu_addr || size == 0) { + LOG_WARNING(HW_GPU, "Failed to find storage buffer for cbuf index {}", cbuf_index); return NULL_BINDING; } const VAddr cpu_end = Common::AlignUp(*cpu_addr + size, Core::Memory::YUZU_PAGESIZE);