[tfrt:jitrt] Add rt.status support to JitRt API intrinsics

PiperOrigin-RevId: 445052118

Author: ezhulenev

backends/jitrt/include/tfrt/jitrt/opdefs/rt_base.td

@@ -44,11 +44,23 @@ class RT_Type<string name, string typeMnemonic> : TypeDef<RuntimeDialect,
   let mnemonic = typeMnemonic;
 }
 
+// -------------------------------------------------------------------------- //
+// Types for integrating JitRt kernels with the runtime.
+// -------------------------------------------------------------------------- //
+
 // This is an opaque handle to tfrt::jitrt::KernelContextType.
 def KernelContextType : RT_Type<"KernelContext", "kernel_context"> {
   let summary = "Kernel Context type";
   let description = [{
-    Opaque handle used for interacting with the TFRT run-time.
+    Opaque handle used for interacting with the JitRt runtime.
+  }];
+}
+
+// This is an opaque handle to tfrt::jitrt::StatusType.
+def StatusType : RT_Type<"Status", "status"> {
+  let summary = "Status type";
+  let description = [{
+    A status type returned from the JitRt runtime API intrinsics.
   }];
 }
 

backends/jitrt/include/tfrt/jitrt/opdefs/rt_ops.h

@@ -22,6 +22,7 @@
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/Dialect.h"
 #include "mlir/IR/OpDefinition.h"
+#include "mlir/IR/OpImplementation.h"
 #include "mlir/IR/Types.h"
 #include "tfrt/jitrt/opdefs/rt_dialect.h.inc"
 

backends/jitrt/include/tfrt/jitrt/opdefs/rt_ops.td

@@ -117,6 +117,20 @@ def SetErrorOp : RT_Op<"set_error"> {
   let assemblyFormat = "$ctx `,` $error attr-dict";
 }
 
+//===----------------------------------------------------------------------===//
+// IsOkOp
+//===----------------------------------------------------------------------===//
+
+def IsOkOp : RT_Op<"is_ok"> {
+  let summary = "returns true if status is ok";
+  let description = "Checks if the runtime status is ok.";
+
+  let arguments = (ins StatusType:$status);
+  let results = (outs I1:$ok);
+
+  let assemblyFormat = "$status attr-dict";
+}
+
 //===----------------------------------------------------------------------===//
 // CustomCallOp
 //===----------------------------------------------------------------------===//
@@ -130,13 +144,18 @@ def CustomCallOp : RT_Op<"custom_call"> {
     on top of the JitRt, for example this can be used as an extension mechanism
     to register vendor specific kernels (e.g. call oneDNN convolution).
 
+    Returns `!rt.status` value which can be checked to see if the custom call
+    was successful.
+
     Example:
 
       ```mlir
       func @compute(%ctx: !rt.kernel_context, %arg0: memref<?xf32>,
                                               %arg1: memref<?xf32>) {
-        %0 = rt.custom_call "one_dnn.some_operation"(%arg0, %arg1)
-          : (memref<?xf32>, memref<?xf32>) -> !one_dnn.status
+        %status = rt.custom_call "one_dnn.some_operation"(%arg0, %arg1)
+          : (memref<?xf32>, memref<?xf32>) -> ()
+        %0 = rt.is_ok %status
+        cf.assert %0, "failed to call one_dnn custom call"
         return
       }
       ```
@@ -152,10 +171,13 @@ def CustomCallOp : RT_Op<"custom_call"> {
     Variadic<AnyType>:$operands
   );
 
-  let results = (outs Variadic<AnyType>);
+  let results = (outs
+    StatusType:$status,
+    Variadic<AnyType>:$results
+  );
 
   let assemblyFormat = [{
-    $callee `(` $operands `)` attr-dict `:` functional-type($operands, results)
+    $callee `(` $operands `)` attr-dict `:` functional-type($operands, $results)
   }];
 }
 

backends/jitrt/include/tfrt/jitrt/runtime.h

@@ -56,8 +56,9 @@ extern "C" void *runtimeGetResultStorage(KernelContext *, int64_t);
 // Sets kernel context to an error state.
 extern "C" void runtimeSetError(KernelContext *, const char *);
 
-// Calls the custom call function registered with the runtime.
-extern "C" void runtimeCustomCall(const char *, void **args);
+// Calls the custom call function registered with the runtime. Returns true
+// if the custom call was successful.
+extern "C" bool runtimeCustomCall(const char *, void **args);
 
 }  // namespace runtime
 }  // namespace jitrt

backends/jitrt/lib/conversion/rt_to_llvm.cc

@@ -113,7 +113,8 @@ struct RuntimeAPI {
   static FunctionType CustomCallFunctionType(MLIRContext *ctx) {
     auto callee = OpaquePointerType(ctx);
     auto args = CustomCallArgumentsType(ctx);
-    return FunctionType::get(ctx, {callee, args}, {});
+    auto i1 = IntegerType::get(ctx, 1);
+    return FunctionType::get(ctx, {callee, args}, {i1});
   }
 };
 
@@ -139,11 +140,16 @@ class RuntimeTypeConverter : public TypeConverter {
   RuntimeTypeConverter() {
     addConversion([](Type type) { return type; });
     addConversion(ConvertKernelContextType);
+    addConversion(ConvertStatusType);
   }
 
   static llvm::Optional<Type> ConvertKernelContextType(KernelContextType type) {
     return LLVM::LLVMPointerType::get(IntegerType::get(type.getContext(), 8));
   }
+
+  static llvm::Optional<Type> ConvertStatusType(StatusType type) {
+    return IntegerType::get(type.getContext(), 1);
+  }
 };
 
 // -------------------------------------------------------------------------- //
@@ -253,6 +259,23 @@ class SetErrorOpLowering : public OpConversionPattern<SetErrorOp> {
   }
 };
 
+//===----------------------------------------------------------------------===//
+// Convert rt.is_ok to the corresponding runtime API call.
+//===----------------------------------------------------------------------===//
+
+class IsOkOpLowering : public OpConversionPattern<IsOkOp> {
+ public:
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult matchAndRewrite(
+      IsOkOp op, OpAdaptor adaptor,
+      ConversionPatternRewriter &rewriter) const override {
+    // Just pass through the converted operand.
+    rewriter.replaceOp(op, adaptor.status());
+    return success();
+  }
+};
+
 //===----------------------------------------------------------------------===//
 // Convert rt.custom_call to the corresponding runtime API call.
 //===----------------------------------------------------------------------===//
@@ -444,7 +467,8 @@ class CustomCallOpLowering : public OpConversionPattern<CustomCallOp> {
     if (failed(args)) return op.emitOpError() << "failed to encode arguments";
 
     // Call runtime API to call the custom call target.
-    rewriter.replaceOpWithNewOp<CallOp>(op, kCustomCall, TypeRange(),
+    auto i1 = rewriter.getI1Type();
+    rewriter.replaceOpWithNewOp<CallOp>(op, kCustomCall, TypeRange(i1),
                                         ValueRange({callee, *args}));
 
     return success();
@@ -471,10 +495,11 @@ void ConvertRuntimeToLLVMPass::runOnOperation() {
   // We use conversion to LLVM type to lower all runtime operands to LLVM types.
   LLVMTypeConverter llvm_converter(ctx);
   llvm_converter.addConversion(RuntimeTypeConverter::ConvertKernelContextType);
+  llvm_converter.addConversion(RuntimeTypeConverter::ConvertStatusType);
 
   // Lower from the runtime operations to the runtime API function calls.
-  patterns.add<SetOutputOpLowering, SetErrorOpLowering, CustomCallOpLowering>(
-      llvm_converter, ctx);
+  patterns.add<SetOutputOpLowering, SetErrorOpLowering, IsOkOpLowering,
+               CustomCallOpLowering>(llvm_converter, ctx);
 
   // Convert function signatures and call sites.
   mlir::populateFunctionOpInterfaceTypeConversionPattern<FuncOp>(patterns,

backends/jitrt/lib/jitrt.cc

@@ -1257,7 +1257,7 @@ extern "C" void runtimeSetError(KernelContext* ctx, const char* error) {
   ctx->call_frame->error = {error};
 }
 
-extern "C" void runtimeCustomCall(const char* callee, void** args) {
+extern "C" bool runtimeCustomCall(const char* callee, void** args) {
   assert(callee && "callee must be not null");
 
   // Default custom calls registry for the JitRt kernels.
@@ -1267,14 +1267,13 @@ extern "C" void runtimeCustomCall(const char* callee, void** args) {
     return registry;
   }();
 
-  // TODO(ezhulenev): Return failure if custom call is not registered.
   auto* custom_call = registry->Find(callee);
-  assert(custom_call && "unknown custom call");
+  if (custom_call == nullptr) return false;
 
-  // TODO(ezhulenev): Handle failures in custom calls.
   auto result = custom_call->call(args);
-  assert(mlir::succeeded(result) && "failed custom call");
-  (void)result;
+  if (mlir::failed(result)) return false;
+
+  return true;
 }
 
 llvm::orc::SymbolMap RuntimeApiSymbolMap(llvm::orc::MangleAndInterner mangle) {

backends/jitrt/mlir_tests/jitrt/compile.assert.mlir

@@ -31,7 +31,7 @@ module @kernels attributes { tfrt.compiled } {
 }
 
 // CHECK: --- Running 'runtime_error'
-func.func @runtime_error() -> !tfrt.chain {
+func.func @runtime_error() -> !t.tensor {
   %ch0 = tfrt.new.chain
 
   // Allocate and initialize input tensor.
@@ -40,9 +40,10 @@ func.func @runtime_error() -> !tfrt.chain {
 
   %executable = jitrt.compile { kernel = @kernels::@main }
 
-  // expected-error @+1 {{Dimension 0 must have size 0}}
   %output = jitrt.execute %executable[%input_ready](%input)
               : (!t.tensor) -> (!t.tensor)
 
-  tfrt.return %ch0 : !tfrt.chain
+  // CHECK: returned <<error: compiled kernel run time error:
+  // CHECK-SAME: Dimension 0 must have size 0>>
+  tfrt.return %output : !t.tensor
 }

backends/jitrt/mlir_tests/jitrt/compile.custom_call.mlir

@@ -20,10 +20,14 @@ module @kernels attributes { tfrt.compiled } {
     %c1 = arith.constant 1 : index
     %0 = memref.dim %input, %c0 : memref<?x?xf32>
     %1 = memref.dim %input, %c1 : memref<?x?xf32>
-    %output = memref.alloc(%0, %1) : memref<?x?xf32>
 
-    rt.custom_call "testlib.times_two"(%input, %output)
+    // Reverse dimension order to test invalid custom call arguments below.
+    %output = memref.alloc(%1, %0) : memref<?x?xf32>
+
+    %status = rt.custom_call "testlib.times_two"(%input, %output)
       : (memref<?x?xf32>, memref<?x?xf32>) -> ()
+    %ok = rt.is_ok %status
+    cf.assert %ok, "failed to call custom call 'testlib.times_two'"
 
     func.return %output : memref<?x?xf32>
   }
@@ -55,4 +59,23 @@ func.func @compiled_custom_call() -> !tfrt.chain {
   %printed = tfrt.print.i1 %cmp, %cmp_ch
 
   tfrt.return %printed : !tfrt.chain
-}
+}
+
+// CHECK: --- Running 'compiled_custom_call_error'
+func.func @compiled_custom_call_error() -> !t.tensor {
+  %ch0 = tfrt.new.chain
+
+  // Allocate and initialize input tensor.
+  %input = tfrt_dht.create_uninitialized_tensor.f32.2 [16 : i64, 4 : i64]
+  %ch1 = tfrt_dht.fill_tensor_with_constant.f32 %input, %ch0 1.0 : f32
+
+  // Compile a kernel with a custom call.
+  %executable = jitrt.compile { kernel = @kernels::@main }
+
+  // Execute compiled kernel with tensor operands.
+  %output = jitrt.execute %executable[%ch1](%input) : (!t.tensor) -> !t.tensor
+
+  // CHECK: returned <<error: compiled kernel run time error:
+  // CHECK-SAME: failed to call custom call 'testlib.times_two'>>
+  tfrt.return %output : !t.tensor
+}

backends/jitrt/mlir_tests/rt/ops.mlir

@@ -43,6 +43,8 @@ func.func @set_error(%arg0: !rt.kernel_context) {
 // CHECK:  %[[MEMREF:.*]]: memref<?xf32>
 func.func @custom_call(%arg0: !rt.kernel_context, %arg1: memref<?xf32>) -> f32 {
   // CHECK: rt.custom_call "f32_reduce"(%[[MEMREF]]) : (memref<?xf32>) -> f32
-  %0 = rt.custom_call "f32_reduce"(%arg1) : (memref<?xf32>) -> f32
+  %status, %0 = rt.custom_call "f32_reduce"(%arg1) : (memref<?xf32>) -> f32
+  %ok = rt.is_ok %status
+  cf.assert %ok, "failed to call custom call"
   func.return %0 : f32
 }

backends/jitrt/mlir_tests/rt/rt_to_llvm.mlir

@@ -79,8 +79,11 @@ func.func @custom_call(%arg0: !rt.kernel_context, %arg1: memref<?xf32>) {
   // CHECK: %[[C3:.*]] = arith.constant 3 : i32
   // CHECK: %[[ARGS:.*]] = llvm.alloca %[[C3]] x !llvm.ptr<i8>
 
-  // CHECK: call @runtimeCustomCall(%[[CALLEE]], %[[ARGS]])
-  rt.custom_call "f32_reduce"(%arg1) : (memref<?xf32>) -> ()
+  // CHECK: %[[STATUS:.*]] = call @runtimeCustomCall(%[[CALLEE]], %[[ARGS]])
+  // CHECK: cf.assert %[[STATUS]], "oops"
+  %status = rt.custom_call "f32_reduce"(%arg1) : (memref<?xf32>) -> ()
+  %ok = rt.is_ok %status
+  cf.assert %ok, "oops"
 
   func.return
 }