Merge pull request #11902 from ameerj/ssbo-align

shader_recompiler: Align SSBO offsets to meet host requirements

src/shader_recompiler/backend/glasm/emit_glasm_memory.cpp

@@ -5,6 +5,7 @@
 #include "shader_recompiler/backend/glasm/glasm_emit_context.h"
 #include "shader_recompiler/frontend/ir/program.h"
 #include "shader_recompiler/frontend/ir/value.h"
+#include "shader_recompiler/profile.h"
 #include "shader_recompiler/runtime_info.h"
 
 namespace Shader::Backend::GLASM {
@@ -35,7 +36,9 @@ void GlobalStorageOp(EmitContext& ctx, Register address, bool pointer_based, std
             continue;
         }
         const auto& ssbo{ctx.info.storage_buffers_descriptors[index]};
-        ctx.Add("LDC.U64 DC.x,c{}[{}];"    // ssbo_addr
+        const u64 ssbo_align_mask{~(ctx.profile.min_ssbo_alignment - 1U)};
+        ctx.Add("LDC.U64 DC.x,c{}[{}];"    // unaligned_ssbo_addr
+                "AND.U64 DC.x,DC.x,{};"    // ssbo_addr = unaligned_ssbo_addr & ssbo_align_mask
                 "LDC.U32 RC.x,c{}[{}];"    // ssbo_size_u32
                 "CVT.U64.U32 DC.y,RC.x;"   // ssbo_size = ssbo_size_u32
                 "ADD.U64 DC.y,DC.y,DC.x;"  // ssbo_end = ssbo_addr + ssbo_size
@@ -44,8 +47,8 @@ void GlobalStorageOp(EmitContext& ctx, Register address, bool pointer_based, std
                 "AND.U.CC RC.x,RC.x,RC.y;" // cond = a && b
                 "IF NE.x;"                 // if cond
                 "SUB.U64 DC.x,{}.x,DC.x;", // offset = input_addr - ssbo_addr
-                ssbo.cbuf_index, ssbo.cbuf_offset, ssbo.cbuf_index, ssbo.cbuf_offset + 8, address,
+                ssbo.cbuf_index, ssbo.cbuf_offset, ssbo_align_mask, ssbo.cbuf_index,
-                address, address);
+                ssbo.cbuf_offset + 8, address, address, address);
         if (pointer_based) {
             ctx.Add("PK64.U DC.y,c[{}];"      // host_ssbo = cbuf
                     "ADD.U64 DC.x,DC.x,DC.y;" // host_addr = host_ssbo + offset

src/shader_recompiler/backend/glsl/glsl_emit_context.cpp

@@ -601,7 +601,10 @@ std::string EmitContext::DefineGlobalMemoryFunctions() {
             addr_xy[i] = fmt::format("ftou({}[{}].{})", cbuf, addr_loc / 16, Swizzle(addr_loc));
             size_xy[i] = fmt::format("ftou({}[{}].{})", cbuf, size_loc / 16, Swizzle(size_loc));
         }
-        const auto addr_pack{fmt::format("packUint2x32(uvec2({},{}))", addr_xy[0], addr_xy[1])};
+        const u32 ssbo_align_mask{~(static_cast<u32>(profile.min_ssbo_alignment) - 1U)};
+        const auto aligned_low_addr{fmt::format("{}&{}", addr_xy[0], ssbo_align_mask)};
+        const auto aligned_addr{fmt::format("uvec2({},{})", aligned_low_addr, addr_xy[1])};
+        const auto addr_pack{fmt::format("packUint2x32({})", aligned_addr)};
         const auto addr_statment{fmt::format("uint64_t {}={};", ssbo_addr, addr_pack)};
         func += addr_statment;
 

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -891,7 +891,9 @@ void EmitContext::DefineGlobalMemoryFunctions(const Info& info) {
             const Id ssbo_size_pointer{OpAccessChain(uniform_types.U32, cbufs[ssbo.cbuf_index].U32,
                                                      zero, ssbo_size_cbuf_offset)};
 
-            const Id ssbo_addr{OpBitcast(U64, OpLoad(U32[2], ssbo_addr_pointer))};
+            const u64 ssbo_align_mask{~(profile.min_ssbo_alignment - 1U)};
+            const Id unaligned_addr{OpBitcast(U64, OpLoad(U32[2], ssbo_addr_pointer))};
+            const Id ssbo_addr{OpBitwiseAnd(U64, unaligned_addr, Constant(U64, ssbo_align_mask))};
             const Id ssbo_size{OpUConvert(U64, OpLoad(U32[1], ssbo_size_pointer))};
             const Id ssbo_end{OpIAdd(U64, ssbo_addr, ssbo_size)};
             const Id cond{OpLogicalAnd(U1, OpUGreaterThanEqual(U1, addr, ssbo_addr),

src/shader_recompiler/frontend/maxwell/translate_program.cpp

@@ -298,7 +298,7 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
 
     Optimization::PositionPass(env, program);
 
-    Optimization::GlobalMemoryToStorageBufferPass(program);
+    Optimization::GlobalMemoryToStorageBufferPass(program, host_info);
     Optimization::TexturePass(env, program, host_info);
 
     if (Settings::values.resolution_info.active) {

src/shader_recompiler/host_translate_info.h

@@ -16,6 +16,7 @@ struct HostTranslateInfo {
     bool needs_demote_reorder{}; ///< True when the device needs DemoteToHelperInvocation reordered
     bool support_snorm_render_buffer{};  ///< True when the device supports SNORM render buffers
     bool support_viewport_index_layer{}; ///< True when the device supports gl_Layer in VS
+    u32 min_ssbo_alignment{};            ///< Minimum alignment supported by the device for SSBOs
     bool support_geometry_shader_passthrough{}; ///< True when the device supports geometry
                                                 ///< passthrough shaders
     bool support_conditional_barrier{}; ///< True when the device supports barriers in conditional

src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp

@@ -11,6 +11,7 @@
 #include "shader_recompiler/frontend/ir/breadth_first_search.h"
 #include "shader_recompiler/frontend/ir/ir_emitter.h"
 #include "shader_recompiler/frontend/ir/value.h"
+#include "shader_recompiler/host_translate_info.h"
 #include "shader_recompiler/ir_opt/passes.h"
 
 namespace Shader::Optimization {
@@ -408,7 +409,7 @@ void CollectStorageBuffers(IR::Block& block, IR::Inst& inst, StorageInfo& info)
 }
 
 /// Returns the offset in indices (not bytes) for an equivalent storage instruction
-IR::U32 StorageOffset(IR::Block& block, IR::Inst& inst, StorageBufferAddr buffer) {
+IR::U32 StorageOffset(IR::Block& block, IR::Inst& inst, StorageBufferAddr buffer, u32 alignment) {
     IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
     IR::U32 offset;
     if (const std::optional<LowAddrInfo> low_addr{TrackLowAddress(&inst)}) {
@@ -421,7 +422,10 @@ IR::U32 StorageOffset(IR::Block& block, IR::Inst& inst, StorageBufferAddr buffer
     }
     // Subtract the least significant 32 bits from the guest offset. The result is the storage
     // buffer offset in bytes.
-    const IR::U32 low_cbuf{ir.GetCbuf(ir.Imm32(buffer.index), ir.Imm32(buffer.offset))};
+    IR::U32 low_cbuf{ir.GetCbuf(ir.Imm32(buffer.index), ir.Imm32(buffer.offset))};
+
+    // Align the offset base to match the host alignment requirements
+    low_cbuf = ir.BitwiseAnd(low_cbuf, ir.Imm32(~(alignment - 1U)));
     return ir.ISub(offset, low_cbuf);
 }
 
@@ -516,7 +520,7 @@ void Replace(IR::Block& block, IR::Inst& inst, const IR::U32& storage_index,
 }
 } // Anonymous namespace
 
-void GlobalMemoryToStorageBufferPass(IR::Program& program) {
+void GlobalMemoryToStorageBufferPass(IR::Program& program, const HostTranslateInfo& host_info) {
     StorageInfo info;
     for (IR::Block* const block : program.post_order_blocks) {
         for (IR::Inst& inst : block->Instructions()) {
@@ -540,7 +544,8 @@ void GlobalMemoryToStorageBufferPass(IR::Program& program) {
         const IR::U32 index{IR::Value{static_cast<u32>(info.set.index_of(it))}};
         IR::Block* const block{storage_inst.block};
         IR::Inst* const inst{storage_inst.inst};
-        const IR::U32 offset{StorageOffset(*block, *inst, storage_buffer)};
+        const IR::U32 offset{
+            StorageOffset(*block, *inst, storage_buffer, host_info.min_ssbo_alignment)};
         Replace(*block, *inst, index, offset);
     }
 }

src/shader_recompiler/ir_opt/passes.h

@@ -16,7 +16,7 @@ void CollectShaderInfoPass(Environment& env, IR::Program& program);
 void ConditionalBarrierPass(IR::Program& program);
 void ConstantPropagationPass(Environment& env, IR::Program& program);
 void DeadCodeEliminationPass(IR::Program& program);
-void GlobalMemoryToStorageBufferPass(IR::Program& program);
+void GlobalMemoryToStorageBufferPass(IR::Program& program, const HostTranslateInfo& host_info);
 void IdentityRemovalPass(IR::Program& program);
 void LowerFp64ToFp32(IR::Program& program);
 void LowerFp16ToFp32(IR::Program& program);

src/shader_recompiler/profile.h

@@ -85,6 +85,8 @@ struct Profile {
 
     /// Maxwell and earlier nVidia architectures have broken robust support
     bool has_broken_robust{};
+
+    u64 min_ssbo_alignment{};
 };
 
 } // namespace Shader

src/video_core/buffer_cache/buffer_cache.h

@@ -1753,15 +1753,25 @@ Binding BufferCache<P>::StorageBufferBinding(GPUVAddr ssbo_addr, u32 cbuf_index,
         const u32 memory_layout_size = static_cast<u32>(gpu_memory->GetMemoryLayoutSize(gpu_addr));
         return std::min(memory_layout_size, static_cast<u32>(8_MiB));
     }();
-    const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
+    // Alignment only applies to the offset of the buffer
-    if (!cpu_addr || size == 0) {
+    const u32 alignment = runtime.GetStorageBufferAlignment();
+    const GPUVAddr aligned_gpu_addr = Common::AlignDown(gpu_addr, alignment);
+    const u32 aligned_size = static_cast<u32>(gpu_addr - aligned_gpu_addr) + size;
+
+    const std::optional<VAddr> aligned_cpu_addr = gpu_memory->GpuToCpuAddress(aligned_gpu_addr);
+    if (!aligned_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, YUZU_PAGESIZE);
+    const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
+    ASSERT_MSG(cpu_addr, "Unaligned storage buffer address not found for cbuf index {}",
+               cbuf_index);
+    // The end address used for size calculation does not need to be aligned
+    const VAddr cpu_end = Common::AlignUp(*cpu_addr + size, Core::Memory::YUZU_PAGESIZE);
+
     const Binding binding{
-        .cpu_addr = *cpu_addr,
+        .cpu_addr = *aligned_cpu_addr,
-        .size = is_written ? size : static_cast<u32>(cpu_end - *cpu_addr),
+        .size = is_written ? aligned_size : static_cast<u32>(cpu_end - *aligned_cpu_addr),
         .buffer_id = BufferId{},
     };
     return binding;

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -191,6 +191,10 @@ public:
         return device.CanReportMemoryUsage();
     }
 
+    u32 GetStorageBufferAlignment() const {
+        return static_cast<u32>(device.GetShaderStorageBufferAlignment());
+    }
+
 private:
     static constexpr std::array PABO_LUT{
         GL_VERTEX_PROGRAM_PARAMETER_BUFFER_NV,          GL_TESS_CONTROL_PROGRAM_PARAMETER_BUFFER_NV,

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -232,6 +232,7 @@ ShaderCache::ShaderCache(RasterizerOpenGL& rasterizer_, Core::Frontend::EmuWindo
           .has_gl_bool_ref_bug = device.HasBoolRefBug(),
           .ignore_nan_fp_comparisons = true,
           .gl_max_compute_smem_size = device.GetMaxComputeSharedMemorySize(),
+          .min_ssbo_alignment = device.GetShaderStorageBufferAlignment(),
       },
       host_info{
           .support_float64 = true,
@@ -240,6 +241,7 @@ ShaderCache::ShaderCache(RasterizerOpenGL& rasterizer_, Core::Frontend::EmuWindo
           .needs_demote_reorder = device.IsAmd(),
           .support_snorm_render_buffer = false,
           .support_viewport_index_layer = device.HasVertexViewportLayer(),
+          .min_ssbo_alignment = static_cast<u32>(device.GetShaderStorageBufferAlignment()),
           .support_geometry_shader_passthrough = device.HasGeometryShaderPassthrough(),
           .support_conditional_barrier = device.SupportsConditionalBarriers(),
       } {

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -355,6 +355,10 @@ bool BufferCacheRuntime::CanReportMemoryUsage() const {
     return device.CanReportMemoryUsage();
 }
 
+u32 BufferCacheRuntime::GetStorageBufferAlignment() const {
+    return static_cast<u32>(device.GetStorageBufferAlignment());
+}
+
 void BufferCacheRuntime::TickFrame(VideoCommon::SlotVector<Buffer>& slot_buffers) noexcept {
     for (auto it = slot_buffers.begin(); it != slot_buffers.end(); it++) {
         it->ResetUsageTracking();

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -91,6 +91,8 @@ public:
 
     bool CanReportMemoryUsage() const;
 
+    u32 GetStorageBufferAlignment() const;
+
     [[nodiscard]] StagingBufferRef UploadStagingBuffer(size_t size);
 
     [[nodiscard]] StagingBufferRef DownloadStagingBuffer(size_t size, bool deferred = false);

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -373,6 +373,7 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, const Device& device
             driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY,
         .has_broken_robust =
             device.IsNvidia() && device.GetNvidiaArch() <= NvidiaArchitecture::Arch_Pascal,
+        .min_ssbo_alignment = device.GetStorageBufferAlignment(),
     };
 
     host_info = Shader::HostTranslateInfo{
@@ -383,6 +384,7 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, const Device& device
             driver_id == VK_DRIVER_ID_AMD_PROPRIETARY || driver_id == VK_DRIVER_ID_AMD_OPEN_SOURCE,
         .support_snorm_render_buffer = true,
         .support_viewport_index_layer = device.IsExtShaderViewportIndexLayerSupported(),
+        .min_ssbo_alignment = static_cast<u32>(device.GetStorageBufferAlignment()),
         .support_geometry_shader_passthrough = device.IsNvGeometryShaderPassthroughSupported(),
         .support_conditional_barrier = device.SupportsConditionalBarriers(),
     };