Quick patches to make it work after rebasing (triton-lang#3)

Devjiu · Aug 13, 2024 · dfe2ad7 · dfe2ad7
1 parent 100c775
commit dfe2ad7
Show file tree

Hide file tree

Showing 6 changed files with 13 additions and 11 deletions.
diff --git a/lib/Conversion/TritonCPUToLLVM/PrintOpToLLVM.cpp b/lib/Conversion/TritonCPUToLLVM/PrintOpToLLVM.cpp
@@ -39,7 +39,7 @@ struct PrintOpConversion : public ConvertOpToLLVMPattern<triton::PrintOp> {
 
     for (size_t i = 0; i < op.getNumOperands(); i++) {
       auto elems = unpackLLElements(loc, adaptor.getOperands()[i], rewriter);
-      if (op.getOperand(i).getType().dyn_cast<RankedTensorType>()) {
+      if (dyn_cast<RankedTensorType>(op.getOperand(i).getType())) {
         llvm_unreachable("Not implemented for tensor types");
       }
 
@@ -61,7 +61,7 @@ struct PrintOpConversion : public ConvertOpToLLVMPattern<triton::PrintOp> {
   std::string getFormatSubstr(Value value, bool hex = false,
                               std::optional<int> width = std::nullopt) const {
     Type type = value.getType();
-    if (type.isa<LLVM::LLVMPointerType>()) {
+    if (isa<LLVM::LLVMPointerType>(type)) {
       return "%p";
     }
     // Hex is "0x%0nx" or "0x%0nllx", where n is the number of hex digits in the

diff --git a/lib/Conversion/TritonCPUToLLVM/TypeConverter.cpp b/lib/Conversion/TritonCPUToLLVM/TypeConverter.cpp
@@ -24,7 +24,7 @@ Type TritonCPUToLLVMTypeConverter::convertTritonPointerType(
     triton::PointerType type) {
   auto ctx = type.getContext();
   auto pointeeType = type.getPointeeType();
-  if (pointeeType.isa<RankedTensorType>()) {
+  if (isa<RankedTensorType>(pointeeType)) {
     llvm_unreachable("Not implemented");
   }
   return LLVM::LLVMPointerType::get(ctx, type.getAddressSpace());

diff --git a/lib/Conversion/TritonToTritonCPU/TritonCPUConversion.cpp b/lib/Conversion/TritonToTritonCPU/TritonCPUConversion.cpp
@@ -26,7 +26,7 @@ TritonCPUTypeConverter::TritonCPUTypeConverter(MLIRContext *context)
   addConversion([this](triton::PointerType ptrType) -> triton::PointerType {
     // Check whether tensor pointer `tt.ptr<tensor<>>`
     auto pointeeTensorType =
-        ptrType.getPointeeType().dyn_cast<RankedTensorType>();
+        dyn_cast<RankedTensorType>(ptrType.getPointeeType());
     if (pointeeTensorType == nullptr)
       return ptrType;
 
@@ -99,9 +99,9 @@ TritonCPUConversionTarget::TritonCPUConversionTarget(
   // We have requirements for the data layouts
   addDynamicallyLegalOp<triton::DotOp>([](triton::DotOp dotOp) -> bool {
     Attribute aEncoding =
-        dotOp.getA().getType().cast<RankedTensorType>().getEncoding();
+        cast<RankedTensorType>(dotOp.getA().getType()).getEncoding();
     Attribute bEncoding =
-        dotOp.getB().getType().cast<RankedTensorType>().getEncoding();
+        cast<RankedTensorType>(dotOp.getB().getType()).getEncoding();
     // TODO:
     return false;
   });

diff --git a/python/triton/runtime/jit.py b/python/triton/runtime/jit.py
@@ -633,7 +633,7 @@ def run(self, *args, grid, warmup, **kwargs):
 
             # The CPU launcher will provide the grid ids directly to the kernel.
             # Note that this design is interim and subject to change.
-            if target[0] == 'cpu':
+            if target.backend == 'cpu':
                 signature["__grid0"] = 'i32'
                 signature["__grid1"] = 'i32'
                 signature["__grid2"] = 'i32'

diff --git a/third_party/cpu/backend/compiler.py b/third_party/cpu/backend/compiler.py
@@ -7,7 +7,7 @@
 from typing import Any
 
 from triton._C.libtriton import cpu, ir, llvm, passes
-from triton.backends.compiler import BaseBackend
+from triton.backends.compiler import BaseBackend, GPUTarget
 
 
 @dataclass(frozen=True)
@@ -35,8 +35,8 @@ def hash(self):
 class CPUBackend(BaseBackend):
 
     @staticmethod
-    def supports_target(target: tuple):
-        return target[0] == "cpu"
+    def supports_target(target: GPUTarget):
+        return target.backend == "cpu"
 
     def __init__(self, target: tuple) -> None:
         super().__init__(target)

diff --git a/third_party/cpu/backend/driver.py b/third_party/cpu/backend/driver.py
@@ -1,3 +1,4 @@
+from triton.backends.compiler import GPUTarget
 from triton.backends.driver import CPUDriverBase
 
 # ------------------------
@@ -60,7 +61,8 @@ def __init__(self):
 
     def get_current_target(self):
         # Capability and warp size are zeros for CPU.
-        return ("cpu", 0, 0)
+        # TODO: GPUTarget naming isn't obviously good.
+        return GPUTarget("cpu", 0, 0)
 
     @staticmethod
     def is_active():