shader: Add NVN storage buffer fallbacks

When we can't track the SSBO origin of a global memory instruction,
leave it as a global memory operation and assume these pointers are in
the NVN storage buffer slots, then apply a linear search in the shader's
runtime.

src/shader_recompiler/backend/spirv/emit_context.cpp

@@ -411,6 +411,7 @@ EmitContext::EmitContext(const Profile& profile_, IR::Program& program, u32& bin
     DefineTextures(program.info, binding);
     DefineImages(program.info, binding);
     DefineAttributeMemAccess(program.info);
+    DefineGlobalMemoryFunctions(program.info);
     DefineLabels(program);
 }
 
@@ -762,6 +763,82 @@ void EmitContext::DefineAttributeMemAccess(const Info& info) {
     }
 }
 
+void EmitContext::DefineGlobalMemoryFunctions(const Info& info) {
+    if (!info.uses_global_memory) {
+        return;
+    }
+    using DefPtr = Id StorageDefinitions::*;
+    const Id zero{u32_zero_value};
+    const auto define_body{[&](DefPtr ssbo_member, Id addr, Id element_pointer, u32 shift,
+                               auto&& callback) {
+        AddLabel();
+        const size_t num_buffers{info.storage_buffers_descriptors.size()};
+        for (size_t index = 0; index < num_buffers; ++index) {
+            const auto& ssbo{info.storage_buffers_descriptors[index]};
+            const Id ssbo_addr_cbuf_offset{Const(ssbo.cbuf_offset / 8)};
+            const Id ssbo_size_cbuf_offset{Const(ssbo.cbuf_offset / 4 + 2)};
+            const Id ssbo_addr_pointer{OpAccessChain(
+                uniform_types.U32x2, cbufs[ssbo.cbuf_index].U32x2, zero, ssbo_addr_cbuf_offset)};
+            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 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),
+                                       OpULessThan(U1, addr, ssbo_end))};
+            const Id then_label{OpLabel()};
+            const Id else_label{OpLabel()};
+            OpSelectionMerge(else_label, spv::SelectionControlMask::MaskNone);
+            OpBranchConditional(cond, then_label, else_label);
+            AddLabel(then_label);
+            const Id ssbo_id{ssbos[index].*ssbo_member};
+            const Id ssbo_offset{OpUConvert(U32[1], OpISub(U64, addr, ssbo_addr))};
+            const Id ssbo_index{OpShiftRightLogical(U32[1], ssbo_offset, Const(shift))};
+            const Id ssbo_pointer{OpAccessChain(element_pointer, ssbo_id, zero, ssbo_index)};
+            callback(ssbo_pointer);
+            AddLabel(else_label);
+        }
+    }};
+    const auto define_load{[&](DefPtr ssbo_member, Id element_pointer, Id type, u32 shift) {
+        const Id function_type{TypeFunction(type, U64)};
+        const Id func_id{OpFunction(type, spv::FunctionControlMask::MaskNone, function_type)};
+        const Id addr{OpFunctionParameter(U64)};
+        define_body(ssbo_member, addr, element_pointer, shift,
+                    [&](Id ssbo_pointer) { OpReturnValue(OpLoad(type, ssbo_pointer)); });
+        OpReturnValue(ConstantNull(type));
+        OpFunctionEnd();
+        return func_id;
+    }};
+    const auto define_write{[&](DefPtr ssbo_member, Id element_pointer, Id type, u32 shift) {
+        const Id function_type{TypeFunction(void_id, U64, type)};
+        const Id func_id{OpFunction(void_id, spv::FunctionControlMask::MaskNone, function_type)};
+        const Id addr{OpFunctionParameter(U64)};
+        const Id data{OpFunctionParameter(type)};
+        define_body(ssbo_member, addr, element_pointer, shift, [&](Id ssbo_pointer) {
+            OpStore(ssbo_pointer, data);
+            OpReturn();
+        });
+        OpReturn();
+        OpFunctionEnd();
+        return func_id;
+    }};
+    const auto define{
+        [&](DefPtr ssbo_member, const StorageTypeDefinition& type_def, Id type, size_t size) {
+            const Id element_type{type_def.element};
+            const u32 shift{static_cast<u32>(std::countr_zero(size))};
+            const Id load_func{define_load(ssbo_member, element_type, type, shift)};
+            const Id write_func{define_write(ssbo_member, element_type, type, shift)};
+            return std::make_pair(load_func, write_func);
+        }};
+    std::tie(load_global_func_u32, write_global_func_u32) =
+        define(&StorageDefinitions::U32, storage_types.U32, U32[1], sizeof(u32));
+    std::tie(load_global_func_u32x2, write_global_func_u32x2) =
+        define(&StorageDefinitions::U32x2, storage_types.U32x2, U32[2], sizeof(u32[2]));
+    std::tie(load_global_func_u32x4, write_global_func_u32x4) =
+        define(&StorageDefinitions::U32x4, storage_types.U32x4, U32[4], sizeof(u32[4]));
+}
+
 void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) {
     if (info.constant_buffer_descriptors.empty()) {
         return;

src/shader_recompiler/backend/spirv/emit_context.h

@@ -224,6 +224,13 @@ public:
     Id f32x2_min_cas{};
     Id f32x2_max_cas{};
 
+    Id load_global_func_u32{};
+    Id load_global_func_u32x2{};
+    Id load_global_func_u32x4{};
+    Id write_global_func_u32{};
+    Id write_global_func_u32x2{};
+    Id write_global_func_u32x4{};
+
     Id input_position{};
     std::array<Id, 32> input_generics{};
 
@@ -255,6 +262,7 @@ private:
     void DefineTextures(const Info& info, u32& binding);
     void DefineImages(const Info& info, u32& binding);
     void DefineAttributeMemAccess(const Info& info);
+    void DefineGlobalMemoryFunctions(const Info& info);
     void DefineLabels(IR::Program& program);
 
     void DefineInputs(const Info& info);

src/shader_recompiler/backend/spirv/emit_spirv.h

@@ -84,16 +84,16 @@ void EmitLoadGlobalU8(EmitContext& ctx);
 void EmitLoadGlobalS8(EmitContext& ctx);
 void EmitLoadGlobalU16(EmitContext& ctx);
 void EmitLoadGlobalS16(EmitContext& ctx);
-void EmitLoadGlobal32(EmitContext& ctx);
+Id EmitLoadGlobal32(EmitContext& ctx, Id address);
-void EmitLoadGlobal64(EmitContext& ctx);
+Id EmitLoadGlobal64(EmitContext& ctx, Id address);
-void EmitLoadGlobal128(EmitContext& ctx);
+Id EmitLoadGlobal128(EmitContext& ctx, Id address);
 void EmitWriteGlobalU8(EmitContext& ctx);
 void EmitWriteGlobalS8(EmitContext& ctx);
 void EmitWriteGlobalU16(EmitContext& ctx);
 void EmitWriteGlobalS16(EmitContext& ctx);
-void EmitWriteGlobal32(EmitContext& ctx);
+void EmitWriteGlobal32(EmitContext& ctx, Id address, Id value);
-void EmitWriteGlobal64(EmitContext& ctx);
+void EmitWriteGlobal64(EmitContext& ctx, Id address, Id value);
-void EmitWriteGlobal128(EmitContext& ctx);
+void EmitWriteGlobal128(EmitContext& ctx, Id address, Id value);
 Id EmitLoadStorageU8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
 Id EmitLoadStorageS8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
 Id EmitLoadStorageU16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
@@ -277,9 +277,9 @@ Id EmitFPIsNan16(EmitContext& ctx, Id value);
 Id EmitFPIsNan32(EmitContext& ctx, Id value);
 Id EmitFPIsNan64(EmitContext& ctx, Id value);
 Id EmitIAdd32(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
-void EmitIAdd64(EmitContext& ctx);
+Id EmitIAdd64(EmitContext& ctx, Id a, Id b);
 Id EmitISub32(EmitContext& ctx, Id a, Id b);
-void EmitISub64(EmitContext& ctx);
+Id EmitISub64(EmitContext& ctx, Id a, Id b);
 Id EmitIMul32(EmitContext& ctx, Id a, Id b);
 Id EmitINeg32(EmitContext& ctx, Id value);
 Id EmitINeg64(EmitContext& ctx, Id value);

src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp

@@ -55,16 +55,16 @@ Id EmitIAdd32(EmitContext& ctx, IR::Inst* inst, Id a, Id b) {
     return result;
 }
 
-void EmitIAdd64(EmitContext&) {
+Id EmitIAdd64(EmitContext& ctx, Id a, Id b) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpIAdd(ctx.U64, a, b);
 }
 
 Id EmitISub32(EmitContext& ctx, Id a, Id b) {
     return ctx.OpISub(ctx.U32[1], a, b);
 }
 
-void EmitISub64(EmitContext&) {
+Id EmitISub64(EmitContext& ctx, Id a, Id b) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpISub(ctx.U64, a, b);
 }
 
 Id EmitIMul32(EmitContext& ctx, Id a, Id b) {

src/shader_recompiler/backend/spirv/emit_spirv_memory.cpp

@@ -64,16 +64,16 @@ void EmitLoadGlobalS16(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitLoadGlobal32(EmitContext&) {
+Id EmitLoadGlobal32(EmitContext& ctx, Id address) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpFunctionCall(ctx.U32[1], ctx.load_global_func_u32, address);
 }
 
-void EmitLoadGlobal64(EmitContext&) {
+Id EmitLoadGlobal64(EmitContext& ctx, Id address) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpFunctionCall(ctx.U32[2], ctx.load_global_func_u32x2, address);
 }
 
-void EmitLoadGlobal128(EmitContext&) {
+Id EmitLoadGlobal128(EmitContext& ctx, Id address) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpFunctionCall(ctx.U32[4], ctx.load_global_func_u32x4, address);
 }
 
 void EmitWriteGlobalU8(EmitContext&) {
@@ -92,16 +92,16 @@ void EmitWriteGlobalS16(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitWriteGlobal32(EmitContext&) {
+void EmitWriteGlobal32(EmitContext& ctx, Id address, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    ctx.OpFunctionCall(ctx.void_id, ctx.write_global_func_u32, address, value);
 }
 
-void EmitWriteGlobal64(EmitContext&) {
+void EmitWriteGlobal64(EmitContext& ctx, Id address, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    ctx.OpFunctionCall(ctx.void_id, ctx.write_global_func_u32x2, address, value);
 }
 
-void EmitWriteGlobal128(EmitContext&) {
+void EmitWriteGlobal128(EmitContext& ctx, Id address, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    ctx.OpFunctionCall(ctx.void_id, ctx.write_global_func_u32x4, address, value);
 }
 
 Id EmitLoadStorageU8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {

src/shader_recompiler/frontend/maxwell/program.cpp

@@ -60,6 +60,48 @@ void CollectInterpolationInfo(Environment& env, IR::Program& program) {
         }();
     }
 }
+
+void AddNVNStorageBuffers(IR::Program& program) {
+    if (!program.info.uses_global_memory) {
+        return;
+    }
+    const u32 driver_cbuf{0};
+    const u32 descriptor_size{0x10};
+    const u32 num_buffers{16};
+    const u32 base{[&] {
+        switch (program.stage) {
+        case Stage::VertexA:
+        case Stage::VertexB:
+            return 0x110u;
+        case Stage::TessellationControl:
+            return 0x210u;
+        case Stage::TessellationEval:
+            return 0x310u;
+        case Stage::Geometry:
+            return 0x410u;
+        case Stage::Fragment:
+            return 0x510u;
+        case Stage::Compute:
+            return 0x310u;
+        }
+        throw InvalidArgument("Invalid stage {}", program.stage);
+    }()};
+    auto& descs{program.info.storage_buffers_descriptors};
+    for (u32 index = 0; index < num_buffers; ++index) {
+        const u32 offset{base + index * descriptor_size};
+        const auto it{std::ranges::find(descs, offset, &StorageBufferDescriptor::cbuf_offset)};
+        if (it != descs.end()) {
+            continue;
+        }
+        // Assume these are written for now
+        descs.push_back({
+            .cbuf_index = driver_cbuf,
+            .cbuf_offset = offset,
+            .count = 1,
+            .is_written = true,
+        });
+    }
+}
 } // Anonymous namespace
 
 IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Block>& block_pool,
@@ -105,6 +147,7 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
     Optimization::VerificationPass(program);
     Optimization::CollectShaderInfoPass(env, program);
     CollectInterpolationInfo(env, program);
+    AddNVNStorageBuffers(program);
     return program;
 }
 

src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp

@@ -187,6 +187,8 @@ void VisitUsages(Info& info, IR::Inst& inst) {
     case IR::Opcode::FPUnordGreaterThanEqual16:
     case IR::Opcode::FPIsNan16:
     case IR::Opcode::GlobalAtomicAddF16x2:
+    case IR::Opcode::GlobalAtomicMinF16x2:
+    case IR::Opcode::GlobalAtomicMaxF16x2:
     case IR::Opcode::StorageAtomicAddF16x2:
     case IR::Opcode::StorageAtomicMinF16x2:
     case IR::Opcode::StorageAtomicMaxF16x2:
@@ -373,12 +375,63 @@ void VisitUsages(Info& info, IR::Inst& inst) {
     case IR::Opcode::StorageAtomicAnd64:
     case IR::Opcode::StorageAtomicOr64:
     case IR::Opcode::StorageAtomicXor64:
+    case IR::Opcode::StorageAtomicExchange64:
         info.uses_int64 = true;
         break;
     default:
         break;
     }
     switch (inst.GetOpcode()) {
+    case IR::Opcode::LoadGlobalU8:
+    case IR::Opcode::LoadGlobalS8:
+    case IR::Opcode::LoadGlobalU16:
+    case IR::Opcode::LoadGlobalS16:
+    case IR::Opcode::LoadGlobal32:
+    case IR::Opcode::LoadGlobal64:
+    case IR::Opcode::LoadGlobal128:
+    case IR::Opcode::WriteGlobalU8:
+    case IR::Opcode::WriteGlobalS8:
+    case IR::Opcode::WriteGlobalU16:
+    case IR::Opcode::WriteGlobalS16:
+    case IR::Opcode::WriteGlobal32:
+    case IR::Opcode::WriteGlobal64:
+    case IR::Opcode::WriteGlobal128:
+    case IR::Opcode::GlobalAtomicIAdd32:
+    case IR::Opcode::GlobalAtomicSMin32:
+    case IR::Opcode::GlobalAtomicUMin32:
+    case IR::Opcode::GlobalAtomicSMax32:
+    case IR::Opcode::GlobalAtomicUMax32:
+    case IR::Opcode::GlobalAtomicInc32:
+    case IR::Opcode::GlobalAtomicDec32:
+    case IR::Opcode::GlobalAtomicAnd32:
+    case IR::Opcode::GlobalAtomicOr32:
+    case IR::Opcode::GlobalAtomicXor32:
+    case IR::Opcode::GlobalAtomicExchange32:
+    case IR::Opcode::GlobalAtomicIAdd64:
+    case IR::Opcode::GlobalAtomicSMin64:
+    case IR::Opcode::GlobalAtomicUMin64:
+    case IR::Opcode::GlobalAtomicSMax64:
+    case IR::Opcode::GlobalAtomicUMax64:
+    case IR::Opcode::GlobalAtomicAnd64:
+    case IR::Opcode::GlobalAtomicOr64:
+    case IR::Opcode::GlobalAtomicXor64:
+    case IR::Opcode::GlobalAtomicExchange64:
+    case IR::Opcode::GlobalAtomicAddF32:
+    case IR::Opcode::GlobalAtomicAddF16x2:
+    case IR::Opcode::GlobalAtomicAddF32x2:
+    case IR::Opcode::GlobalAtomicMinF16x2:
+    case IR::Opcode::GlobalAtomicMinF32x2:
+    case IR::Opcode::GlobalAtomicMaxF16x2:
+    case IR::Opcode::GlobalAtomicMaxF32x2:
+        info.uses_int64 = true;
+        info.uses_global_memory = true;
+        info.used_constant_buffer_types |= IR::Type::U32 | IR::Type::U32x2;
+        info.used_storage_buffer_types |= IR::Type::U32 | IR::Type::U32x2 | IR::Type::U32x4;
+        break;
+    default:
+        break;
+    }
+    switch (inst.GetOpcode()) {
     case IR::Opcode::DemoteToHelperInvocation:
         info.uses_demote_to_helper_invocation = true;
         break;

src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp

@@ -11,6 +11,7 @@
 #include <boost/container/flat_set.hpp>
 #include <boost/container/small_vector.hpp>
 
+#include "common/alignment.h"
 #include "shader_recompiler/frontend/ir/basic_block.h"
 #include "shader_recompiler/frontend/ir/breadth_first_search.h"
 #include "shader_recompiler/frontend/ir/ir_emitter.h"
@@ -244,39 +245,6 @@ bool MeetsBias(const StorageBufferAddr& storage_buffer, const Bias& bias) noexce
            storage_buffer.offset < bias.offset_end;
 }
 
-/// Discards a global memory operation, reads return zero and writes are ignored
-void DiscardGlobalMemory(IR::Block& block, IR::Inst& inst) {
-    IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
-    const IR::Value zero{u32{0}};
-    switch (inst.GetOpcode()) {
-    case IR::Opcode::LoadGlobalS8:
-    case IR::Opcode::LoadGlobalU8:
-    case IR::Opcode::LoadGlobalS16:
-    case IR::Opcode::LoadGlobalU16:
-    case IR::Opcode::LoadGlobal32:
-        inst.ReplaceUsesWith(zero);
-        break;
-    case IR::Opcode::LoadGlobal64:
-        inst.ReplaceUsesWith(IR::Value{ir.CompositeConstruct(zero, zero)});
-        break;
-    case IR::Opcode::LoadGlobal128:
-        inst.ReplaceUsesWith(IR::Value{ir.CompositeConstruct(zero, zero, zero, zero)});
-        break;
-    case IR::Opcode::WriteGlobalS8:
-    case IR::Opcode::WriteGlobalU8:
-    case IR::Opcode::WriteGlobalS16:
-    case IR::Opcode::WriteGlobalU16:
-    case IR::Opcode::WriteGlobal32:
-    case IR::Opcode::WriteGlobal64:
-    case IR::Opcode::WriteGlobal128:
-        inst.Invalidate();
-        break;
-    default:
-        throw LogicError("Invalid opcode to discard its global memory operation {}",
-                         inst.GetOpcode());
-    }
-}
-
 struct LowAddrInfo {
     IR::U32 value;
     s32 imm_offset;
@@ -350,6 +318,10 @@ std::optional<StorageBufferAddr> Track(const IR::Value& value, const Bias* bias)
             .index{index.U32()},
             .offset{offset.U32()},
         };
+        if (!Common::IsAligned(storage_buffer.offset, 16)) {
+            // The SSBO pointer has to be aligned
+            return std::nullopt;
+        }
         if (bias && !MeetsBias(storage_buffer, *bias)) {
             // We have to blacklist some addresses in case we wrongly
             // point to them
@@ -372,19 +344,17 @@ void CollectStorageBuffers(IR::Block& block, IR::Inst& inst, StorageInfo& info)
     // Track the low address of the instruction
     const std::optional<LowAddrInfo> low_addr_info{TrackLowAddress(&inst)};
     if (!low_addr_info) {
-        DiscardGlobalMemory(block, inst);
+        // Failed to track the low address, use NVN fallbacks
         return;
     }
     // First try to find storage buffers in the NVN address
     const IR::U32 low_addr{low_addr_info->value};
-    std::optional storage_buffer{Track(low_addr, &nvn_bias)};
+    std::optional<StorageBufferAddr> storage_buffer{Track(low_addr, &nvn_bias)};
     if (!storage_buffer) {
         // If it fails, track without a bias
         storage_buffer = Track(low_addr, nullptr);
         if (!storage_buffer) {
-            // If that also failed, drop the global memory usage
+            // If that also fails, use NVN fallbacks
-            // LOG_ERROR
-            DiscardGlobalMemory(block, inst);
             return;
         }
     }

src/shader_recompiler/shader_info.h

@@ -162,6 +162,7 @@ struct Info {
     bool uses_atomic_f32x2_min{};
     bool uses_atomic_f32x2_max{};
     bool uses_int64_bit_atomics{};
+    bool uses_global_memory{};
 
     IR::Type used_constant_buffer_types{};
     IR::Type used_storage_buffer_types{};