Merge branch 'main' into sub-group-shuffle-broadcast

Author: victor-eds

.github/actions/setup-pytorch/action.yml

@@ -82,7 +82,7 @@ runs:
       uses: ./.github/actions/load
       env:
         # Increase this value to reset cache
-        CACHE_NUMBER: 11
+        CACHE_NUMBER: 12
       with:
         path: pytorch
         key: pytorch-$PYTORCH_CACHE_KEY-$CACHE_NUMBER

.github/pins/pytorch-upstream.txt

@@ -1 +1 @@
-8321eec009c8c79145ebccd51fdfc336e5f8b848
+487873f7cafeb0fd390eaefe40496b804bceabbd

.github/workflows/integration-tests.yml

@@ -10,8 +10,7 @@ name: Integration Tests
 on:
   workflow_dispatch:
   pull_request:
-    # You can name your branch dev-foo to get CI runs.
-    branches: [main, 'dev-**']
+    branches-ignore: ['llvm-**']
   merge_group:
     branches: [main, 'dev-**']
     types: [checks_requested]

.github/workflows/integration-tests.yml.in

@@ -9,8 +9,7 @@ name: Integration Tests
 on:
   workflow_dispatch:
   pull_request:
-    # You can name your branch dev-foo to get CI runs.
-    branches: [main, 'dev-**']
+    branches-ignore: ['llvm-**']
   merge_group:
     branches: [main, 'dev-**']
     types: [checks_requested]

.gitignore

@@ -31,6 +31,9 @@ python/triton/language/extra
 # Proton
 python/triton/profiler
 
+# Instrumentation
+python/triton/instrumentation
+
 # Python caches
 __pycache__/
 *.py[cod]

benchmarks/triton_kernels_benchmark/benchmark_driver.py

@@ -399,19 +399,77 @@ def format_of(ty):
     return src
 
 
+def serialize_kernel_metadata(arg, args_dict):
+    args_dict["num_warps"] = arg.num_warps
+    args_dict["threads_per_warp"] = arg.threads_per_warp
+    args_dict["shared_memory"] = arg.shared
+    args_dict["kernel_name"] = arg.name
+    args_dict["spv_name"] = f"{arg.name}.spv"
+
+
+def serialize_args(args, constants, signature):
+    import numbers
+    dir_path = os.getenv("TRITON_XPU_DUMP_SPIRV_KERNEL_ARGS")
+    if not os.path.exists(dir_path):
+        os.makedirs(dir_path)
+        print(f"Path to directory consisting of SPIR-V Runner data: {dir_path}")
+
+    cnt = 0
+    args_dict = {"gridX": args[cnt], "gridY": args[cnt + 1], "gridZ": args[cnt + 2]}
+    args_dict["argument_list"] = []
+    counts = {"tensors": 0, "scalars": 0, "karg_cnt": 0}
+    cnt = 4
+    for arg in args[cnt:]:
+        if type(arg).__name__ == "KernelMetadata":
+            serialize_kernel_metadata(arg, args_dict)
+
+        if isinstance(arg, torch.Tensor):
+            cpu_tensor = arg.cpu()
+            tensor_path = os.path.join(dir_path, f"tensor_{counts['tensors']}.pt")
+            with open(tensor_path, "wb") as f:
+                torch.save(cpu_tensor, f)
+            new_arg = {
+                "name": f"tensor_{counts['tensors']}", "type": "tensor", "dtype": str(arg.dtype), "ctype":
+                signature[counts["karg_cnt"]]
+            }
+            args_dict["argument_list"].append(new_arg)
+            counts["karg_cnt"] += 1
+            counts["tensors"] += 1
+
+        if isinstance(arg, numbers.Number):
+            if counts["karg_cnt"] not in constants:
+                new_arg = {
+                    "name": f"scalarArg_{counts['scalars']}", "type": "scalar", "value": args[cnt], "ctype":
+                    signature[counts["karg_cnt"]]
+                }
+                args_dict["argument_list"].append(new_arg)
+            counts["karg_cnt"] += 1
+            counts["scalars"] += 1
+        cnt += 1
+    # Dump argument info as a JSON file
+    json_path = os.path.join(dir_path, "args_data.json")
+    with open(json_path, "w", encoding="utf-8") as json_file:
+        import json
+        json.dump(args_dict, json_file, indent=4)
+
+
 class XPULauncher:
 
     def __init__(self, src, metadata):  # pylint: disable=unused-argument
         ids = {"ids_of_const_exprs": src.fn.constexprs if hasattr(src, "fn") else tuple()}
         constants = src.constants if hasattr(src, "constants") else {}
         cst_key = lambda i: src.fn.arg_names.index(i) if isinstance(i, str) else i
-        constants = {cst_key(key): value for key, value in constants.items()}
-        signature = {cst_key(key): value for key, value in src.signature.items()}
-        src = make_launcher(constants, signature, ids)
+        self.constants = {cst_key(key): value for key, value in constants.items()}
+        self.signature = {cst_key(key): value for key, value in src.signature.items()}
+        src = make_launcher(self.constants, self.signature, ids)
         mod = compile_module_from_src(src, "__triton_launcher")
         self.launch = mod.launch
 
     def __call__(self, *args, **kwargs):
+        # Serialize KernelArguments for SPIR-V Runner
+        serialize_kernel_args = os.getenv("TRITON_XPU_DUMP_SPIRV_KERNEL_ARGS", None)
+        if serialize_kernel_args:
+            serialize_args(args, self.constants, self.signature)
         self.launch(*args, **kwargs)
 
 

bin/RegisterTritonDialects.h

@@ -88,8 +88,6 @@ inline void registerTritonDialects(mlir::DialectRegistry &registry) {
   mlir::registerTritonAMDGPUStreamPipeline();
   mlir::registerTritonAMDGPUStreamPipelineV2();
   mlir::registerTritonAMDGPUCanonicalizePointers();
-  mlir::triton::registerTritonAMDGPUInsertInstructionSchedHints();
-  mlir::triton::registerTritonAMDGPULowerInstructionSchedHints();
 
   // TODO: register Triton & TritonGPU passes
   registry.insert<mlir::triton::TritonDialect, mlir::cf::ControlFlowDialect,

include/triton/Conversion/TritonGPUToLLVM/ElementwiseOpToLLVMBase.h

@@ -88,10 +88,11 @@ class ElementwiseOpConversionBase : public ConvertOpToLLVMPattern<SourceOp> {
       // encoding not available
       return resultVals;
     Attribute baseEncoding = encoding;
-    if (isa<AMDMfmaEncodingAttr>(baseEncoding))
-      // TODO: this logic seems incorrect for mfma layout. Skip for now.
-      // We saw mismatches for some flash-attention tests on AMD backend.
-      // Note that this logic works for sliced layout whose parent is
+    if (isa<AMDMfmaEncodingAttr>(baseEncoding) ||
+        isa<AMDWmmaEncodingAttr>(baseEncoding))
+      // TODO: this logic seems incorrect for mfma and wmma layout. Skip for
+      // now. We saw mismatches for some flash-attention and dot tests on AMD
+      // backend. Note that this logic works for sliced layout whose parent is
       // mfma layout. Therefore, this is not combined with the following check.
       return resultVals;
     while (auto sliced = dyn_cast<SliceEncodingAttr>(baseEncoding))

include/triton/Conversion/TritonGPUToLLVM/PatternTritonGPUOpToLLVM.h

@@ -27,33 +27,15 @@ constexpr int patternBenefitPrioritizeOverLLVMConversions = 10;
 constexpr int patternBenefitClampOptimizedPattern = 20;
 constexpr int patternBenefitConvertLayoutOptimizedPattern = 20;
 
-struct BackendCallbacks {
-  /**
-   * A backend-specific callback for appending auxiliary data during
-   * `LocalStoreOp` conversion.
-   *
-   * @param[in] op The reference to the re-written `LocalStoreOp`.
-   * @param[in] count The number of issued LLVM instructions.
-   * @param[in] type The input type of issued LLVM instructions.
-   */
-  std::function<void(triton::gpu::LocalStoreOp op, size_t llvmOpCount,
-                     Type llvmOpType)>
-      localStoreOpConversion = nullptr;
-};
-
 void populateElementwiseOpToLLVMPatterns(
     LLVMTypeConverter &typeConverter, RewritePatternSet &patterns,
     ModuleAxisInfoAnalysis &axisInfoAnalysis, const TargetInfoBase &targetInfo,
     PatternBenefit benefit);
 
-// The given callback is invoked at the end of a successful rewrite. The
-// callback receives 1) the current source op, 2) the number of issued LLVM
-// instructions and 3) their input types. Each MLIR backend can provide a
-// callback and, thus, handle backend-specific behaviors.
-void populateMemoryOpToLLVMPattern(
-    LLVMTypeConverter &typeConverter, const TargetInfoBase &targetInfo,
-    RewritePatternSet &patterns, PatternBenefit benefit,
-    std::optional<BackendCallbacks> backendCallbacks = std::nullopt);
+void populateMemoryOpToLLVMPattern(LLVMTypeConverter &typeConverter,
+                                   const TargetInfoBase &targetInfo,
+                                   RewritePatternSet &patterns,
+                                   PatternBenefit benefit);
 
 void populateAssertOpToLLVMPattern(LLVMTypeConverter &typeConverter,
                                    RewritePatternSet &patterns,

include/triton/Conversion/TritonGPUToLLVM/Utility.h

@@ -1366,11 +1366,11 @@ SmallVector<Value> loadSharedToDistributed(RankedTensorType dstTy,
                                            Location loc, RewriterBase &rewriter,
                                            const TargetInfoBase &target);
 
-void storeDistributedToShared(
-    MemDescType dstTy, RankedTensorType srcTy, Type elemLlvmTy,
-    ArrayRef<Value> srcVals, Value smemBase, ArrayRef<Value> dstStrides,
-    Location loc, RewriterBase &rewriter, const TargetInfoBase &target,
-    std::pair<size_t, Type> *const llvmOpCount = nullptr);
+void storeDistributedToShared(MemDescType dstTy, RankedTensorType srcTy,
+                              Type elemLlvmTy, ArrayRef<Value> srcVals,
+                              Value smemBase, ArrayRef<Value> dstStrides,
+                              Location loc, RewriterBase &rewriter,
+                              const TargetInfoBase &target);
 
 inline Value getStructFromSharedMemoryObject(Location loc,
                                              const SharedMemoryObject &smemObj,