GnmDriver: Clear context support (#567)

* gnmdriver: added support for gpu context reset

* shader_recompiler: minor validation fixes

* shader_recompiler: added `V_CMPX_GT_I32`

* shader_recompiler: fix for crash on inline sampler access

* compilation warnings and dead code elimination

* amdgpu: fix for registers addressing

* libraries: videoout: reduce logging pressure

* shader_recompiler: fix for devergence scope detection

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -179,6 +179,7 @@ void DefineEntryPoint(const IR::Program& program, EmitContext& ctx, Id main) {
     spv::ExecutionModel execution_model{};
     ctx.AddCapability(spv::Capability::Image1D);
     ctx.AddCapability(spv::Capability::Sampled1D);
+    ctx.AddCapability(spv::Capability::ImageQuery);
     if (info.uses_fp16) {
         ctx.AddCapability(spv::Capability::Float16);
         ctx.AddCapability(spv::Capability::Int16);

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -405,6 +405,10 @@ spv::ImageFormat GetFormat(const AmdGpu::Image& image) {
         image.GetNumberFmt() == AmdGpu::NumberFormat::Float) {
         return spv::ImageFormat::Rg16f;
     }
+    if (image.GetDataFmt() == AmdGpu::DataFormat::Format16_16 &&
+        image.GetNumberFmt() == AmdGpu::NumberFormat::Snorm) {
+        return spv::ImageFormat::Rg16Snorm;
+    }
     if (image.GetDataFmt() == AmdGpu::DataFormat::Format8_8 &&
         image.GetNumberFmt() == AmdGpu::NumberFormat::Unorm) {
         return spv::ImageFormat::Rg8;

src/shader_recompiler/frontend/control_flow_graph.cpp

@@ -21,8 +21,13 @@ struct Compare {
     }
 };
 
-static IR::Condition MakeCondition(Opcode opcode) {
-    switch (opcode) {
+static IR::Condition MakeCondition(const GcnInst& inst) {
+    if (inst.IsCmpx()) {
+        ASSERT(inst.opcode == Opcode::V_CMPX_NE_U32);
+        return IR::Condition::Execnz;
+    }
+
+    switch (inst.opcode) {
     case Opcode::S_CBRANCH_SCC0:
         return IR::Condition::Scc0;
     case Opcode::S_CBRANCH_SCC1:
@@ -37,7 +42,6 @@ static IR::Condition MakeCondition(Opcode opcode) {
         return IR::Condition::Execnz;
     case Opcode::S_AND_SAVEEXEC_B64:
     case Opcode::S_ANDN2_B64:
-    case Opcode::V_CMPX_NE_U32:
         return IR::Condition::Execnz;
     default:
         return IR::Condition::True;
@@ -94,7 +98,8 @@ void CFG::EmitDivergenceLabels() {
                // While this instruction does not save EXEC it is often used paired
                // with SAVEEXEC to mask the threads that didn't pass the condition
                // of initial branch.
-               inst.opcode == Opcode::S_ANDN2_B64 || inst.opcode == Opcode::V_CMPX_NE_U32;
+               (inst.opcode == Opcode::S_ANDN2_B64 && inst.dst[0].field == OperandField::ExecLo) ||
+               inst.opcode == Opcode::V_CMPX_NE_U32;
     };
     const auto is_close_scope = [](const GcnInst& inst) {
         // Closing an EXEC scope can be either a branch instruction
@@ -104,7 +109,8 @@ void CFG::EmitDivergenceLabels() {
                // Sometimes compiler might insert instructions between the SAVEEXEC and the branch.
                // Those instructions need to be wrapped in the condition as well so allow branch
                // as end scope instruction.
-               inst.opcode == Opcode::S_CBRANCH_EXECZ || inst.opcode == Opcode::S_ANDN2_B64;
+               inst.opcode == Opcode::S_CBRANCH_EXECZ ||
+               (inst.opcode == Opcode::S_ANDN2_B64 && inst.dst[0].field == OperandField::ExecLo);
     };
 
     // Since we will be adding new labels, avoid iterating those as well.
@@ -171,7 +177,7 @@ void CFG::EmitBlocks() {
         block->begin_index = GetIndex(start);
         block->end_index = end_index;
         block->end_inst = end_inst;
-        block->cond = MakeCondition(end_inst.opcode);
+        block->cond = MakeCondition(end_inst);
         blocks.insert(*block);
     }
 }

src/shader_recompiler/frontend/instruction.cpp

@@ -47,4 +47,18 @@ bool GcnInst::IsConditionalBranch() const {
     return false;
 }
 
+bool GcnInst::IsCmpx() const {
+    if ((opcode >= Opcode::V_CMPX_F_F32 && opcode <= Opcode::V_CMPX_T_F32) ||
+        (opcode >= Opcode::V_CMPX_F_F64 && opcode <= Opcode::V_CMPX_T_F64) ||
+        (opcode >= Opcode::V_CMPSX_F_F32 && opcode <= Opcode::V_CMPSX_T_F32) ||
+        (opcode >= Opcode::V_CMPSX_F_F64 && opcode <= Opcode::V_CMPSX_T_F64) ||
+        (opcode >= Opcode::V_CMPX_F_I32 && opcode <= Opcode::V_CMPX_CLASS_F32) ||
+        (opcode >= Opcode::V_CMPX_F_I64 && opcode <= Opcode::V_CMPX_CLASS_F64) ||
+        (opcode >= Opcode::V_CMPX_F_U32 && opcode <= Opcode::V_CMPX_T_U32) ||
+        (opcode >= Opcode::V_CMPX_F_U64 && opcode <= Opcode::V_CMPX_T_U64)) {
+        return true;
+    }
+    return false;
+}
+
 } // namespace Shader::Gcn

src/shader_recompiler/frontend/instruction.h

@@ -203,6 +203,7 @@ struct GcnInst {
     bool IsUnconditionalBranch() const;
     bool IsConditionalBranch() const;
     bool IsFork() const;
+    bool IsCmpx() const;
 };
 
 } // namespace Shader::Gcn

src/shader_recompiler/frontend/translate/vector_alu.cpp

@@ -280,6 +280,8 @@ void Translator::EmitVectorAlu(const GcnInst& inst) {
         return V_CMP_U32(ConditionOp::GT, true, false, inst);
     case Opcode::V_CMP_LT_I32:
         return V_CMP_U32(ConditionOp::LT, true, false, inst);
+    case Opcode::V_CMPX_GT_I32:
+        return V_CMP_U32(ConditionOp::GT, true, true, inst);
     case Opcode::V_CMPX_LT_I32:
         return V_CMP_U32(ConditionOp::LT, true, true, inst);
     case Opcode::V_CMPX_F_U32:

src/shader_recompiler/ir/passes/resource_tracking_pass.cpp

@@ -246,10 +246,7 @@ public:
                 return true;
             }
             // Samplers with different bindings might still be the same.
-            const auto old_sharp =
-                info.ReadUd<AmdGpu::Sampler>(existing.sgpr_base, existing.dword_offset);
-            const auto new_sharp = info.ReadUd<AmdGpu::Sampler>(desc.sgpr_base, desc.dword_offset);
-            return old_sharp == new_sharp;
+            return existing.GetSsharp(info) == desc.GetSsharp(info);
         })};
         return index;
     }