Merge OpenAI Triton commit 4dfdc32 (#4445)

This PR change the Triton base from
6af4919 to
4dfdc32 (Jun 5).
Pass rate: 97.23%

Author: whitneywhtsang

.github/workflows/documentation.yml

@@ -4,7 +4,8 @@ on:
   schedule:
     - cron: "0 0 * * *"
 
-permissions: read-all
+permissions:
+  contents: write
 
 jobs:
   Build-Documentation:
@@ -15,7 +16,7 @@ jobs:
       - name: Checkout branch
         uses: actions/checkout@v4
         with:
-          token: ${{ secrets.CI_PAT }}
+          token: ${{ secrets.GITHUB_TOKEN }}
           fetch-depth: 0
 
       - name: Clear docs

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

@@ -60,26 +60,6 @@ jobs:
             ~/.triton/nvidia
             ~/.triton/json
           key: ${{ runner.os }}-${{ runner.arch }}-llvm-${{ steps.cache-key.outputs.llvm }}-nvidia-${{ steps.cache-key.outputs.nvidia }}-json-${{ steps.cache-key.outputs.json }}
-      - # Cache ~/.cache/ccache to speed up compilation.
-        #
-        # On branch `main` we always start from an empty cache, i.e. we skip the
-        # "restore" step.  This is to prevent the caches from accumulating stale
-        # files over time.
-        name: Restore cache of ccache and Triton compilation artifacts
-        id: restore-build-cache
-        if: github.ref != 'refs/heads/main'
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            ~/.ccache
-          # Restore the most recent cache entry.
-          restore-keys: |
-            triton-artifacts-${{ runner.os }}-${{ runner.arch }}-${{ env.RUNNER_TYPE }}-llvm-${{ steps.cache-key.outputs.llvm }}-
-            triton-artifacts-${{ runner.os }}-${{ runner.arch }}-${{ env.RUNNER_TYPE }}-
-          # We expect this cache key never to hit and for us to fall back
-          # unconditionally to the restore-key, so it doesn't actually matter
-          # what we put here (so long as it doesn't hit an existing key).
-          key: triton-artifacts-${{ runner.os }}-${{ runner.arch }}-${{ env.RUNNER_TYPE }}-llvm-${{ steps.cache-key.outputs.llvm }}-${{ steps.cache-key.outputs.datetime }}
       - name: Inspect cache directories
         run: |
           mkdir -p ~/.triton
@@ -152,18 +132,6 @@ jobs:
 
           mkdir -p ~/.ccache
           du -h -d 1 ~/.ccache
-      - # If we're on branch `main`, save the ccache Triton compilation artifacts
-        # to the cache so they can be used by other (non-main) CI runs.
-        #
-        # (It wouldn't be a problem to save the cache on every run, because github
-        # evicts cache entries LRU, but maybe this saves a bit of time in CI.)
-        name: Save ccache and Triton compilation artifacts to cache
-        if: github.ref == 'refs/heads/main'
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            ~/.ccache
-          key: triton-artifacts-${{ runner.os }}-${{ runner.arch }}-${{ env.RUNNER_TYPE }}-llvm-${{ steps.cache-key.outputs.llvm }}-${{ steps.cache-key.outputs.datetime }}
       - name: Clean up caches
         # Always cleanup the worker, even if builds or tests failed
         if: always()

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

@@ -57,26 +57,6 @@ jobs:
             ~/.triton/nvidia
             ~/.triton/json
           key: ${{ runner.os }}-${{ runner.arch }}-llvm-${{ steps.cache-key.outputs.llvm }}-nvidia-${{ steps.cache-key.outputs.nvidia }}-json-${{ steps.cache-key.outputs.json }}
-      - # Cache ~/.cache/ccache to speed up compilation.
-        #
-        # On branch `main` we always start from an empty cache, i.e. we skip the
-        # "restore" step.  This is to prevent the caches from accumulating stale
-        # files over time.
-        name: Restore cache of ccache and Triton compilation artifacts
-        id: restore-build-cache
-        if: github.ref != 'refs/heads/main'
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            ~/.ccache
-          # Restore the most recent cache entry.
-          restore-keys: |
-            triton-artifacts-${{ runner.os }}-${{ runner.arch }}-${{ env.RUNNER_TYPE }}-llvm-${{ steps.cache-key.outputs.llvm }}-
-            triton-artifacts-${{ runner.os }}-${{ runner.arch }}-${{ env.RUNNER_TYPE }}-
-          # We expect this cache key never to hit and for us to fall back
-          # unconditionally to the restore-key, so it doesn't actually matter
-          # what we put here (so long as it doesn't hit an existing key).
-          key: triton-artifacts-${{ runner.os }}-${{ runner.arch }}-${{ env.RUNNER_TYPE }}-llvm-${{ steps.cache-key.outputs.llvm }}-${{ steps.cache-key.outputs.datetime }}
       - name: Inspect cache directories
         run: |
           mkdir -p ~/.triton
@@ -130,15 +110,3 @@ jobs:
 
           mkdir -p ~/.ccache
           du -h -d 1 ~/.ccache
-      - # If we're on branch `main`, save the ccache Triton compilation artifacts
-        # to the cache so they can be used by other (non-main) CI runs.
-        #
-        # (It wouldn't be a problem to save the cache on every run, because github
-        # evicts cache entries LRU, but maybe this saves a bit of time in CI.)
-        name: Save ccache and Triton compilation artifacts to cache
-        if: github.ref == 'refs/heads/main'
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            ~/.ccache
-          key: triton-artifacts-${{ runner.os }}-${{ runner.arch }}-${{ env.RUNNER_TYPE }}-llvm-${{ steps.cache-key.outputs.llvm }}-${{ steps.cache-key.outputs.datetime }}

Makefile

@@ -106,9 +106,9 @@ dev-install-llvm:
 
 .PHONY: golden-samples
 golden-samples: triton-opt
-	$(TRITON_OPT) test/TritonGPU/samples/simulated-grouped-gemm.mlir.in -tritongpu-pipeline -canonicalize | \
+	$(TRITON_OPT) test/TritonGPU/samples/simulated-grouped-gemm.mlir.in -tritongpu-assign-latencies -tritongpu-schedule-loops -tritongpu-pipeline -canonicalize | \
 		$(PYTHON) utils/generate-test-checks.py --source test/TritonGPU/samples/simulated-grouped-gemm.mlir.in --source_delim_regex="\bmodule" \
 		-o test/TritonGPU/samples/simulated-grouped-gemm.mlir
-	$(TRITON_OPT) test/TritonGPU/samples/descriptor-matmul-pipeline.mlir.in -tritongpu-pipeline -canonicalize | \
+	$(TRITON_OPT) test/TritonGPU/samples/descriptor-matmul-pipeline.mlir.in -tritongpu-assign-latencies -tritongpu-schedule-loops -tritongpu-pipeline -canonicalize | \
 		$(PYTHON) utils/generate-test-checks.py --source test/TritonGPU/samples/descriptor-matmul-pipeline.mlir.in --source_delim_regex="\bmodule" \
 		-o test/TritonGPU/samples/descriptor-matmul-pipeline.mlir

include/triton/Dialect/Triton/IR/TritonOps.td

@@ -865,19 +865,24 @@ def TT_ElementwiseInlineAsmOp : TT_Op<"elementwise_inline_asm", [
 //
 // Histogram Op
 //
-def TT_HistogramOp : TT_Op<"histogram", [Pure]> {
+def TT_HistogramOp : TT_Op<"histogram", [Pure,
+    TypesMatchWith<"mask type matches src type",
+                 "src", "mask", "getI1SameShape($_self)",
+                 "($_op.getOperands().size() <= 1) || std::equal_to<>()">]> {
   let summary = "return a histogram of the inputs.";
   let description = [{
     Return the histogram of the input tensor. The number of bins is equal to
     the dimension of the output tensor. Each bins has a width of 1 and bins
     start at 0.
   }];
 
-  let arguments = (ins TT_IntTensor:$src);
+  let arguments = (ins TT_IntTensor:$src,
+    Optional<TT_BoolLike>:$mask);
+
   let results = (outs TT_IntTensor:$result);
 
   let assemblyFormat = [{
-    $src attr-dict `:` type($src) `->` type($result)
+    $src (`,` $mask^)? attr-dict `:` type($src) `->` type($result)
   }];
 }
 
@@ -1028,22 +1033,6 @@ def TT_MakeTensorDescOp : TT_Op<"make_tensor_descriptor", [
   }];
 }
 
-def ReinterpretTensorDescOp : TT_Op<"reinterpret_tensor_descriptor", [Pure]> {
-  let summary = "Reinterpret a pointer as a tensor descriptor";
-
-  let description = [{
-     This Op exists to help the transition from untyped raw TMA objects to typed Tensor descriptor objects.
-     Ideally, we can remove this once the APIs are fully fleshed out.
-  }];
-
-  let arguments = (ins TT_Ptr:$rawDesc);
-  let results = (outs TT_TensorDescType:$result);
-
-  let assemblyFormat = [{
-    $rawDesc attr-dict `:` qualified(type($rawDesc))  `to` qualified(type($result))
-  }];
-}
-
 // The following ops, including `call`, `func`, and `return` are copied and modified from
 // https://github.com/llvm/llvm-project/blob/main/mlir/include/mlir/Dialect/Func/IR/FuncOps.td
 // We could revert it back once MLIR has a better inliner interface.
@@ -1385,54 +1374,5 @@ def TT_DescriptorScatterOp : TT_Op<"descriptor_scatter", [TT_DescriptorStoreLike
   let hasVerifier = 1;
 }
 
-def TT_ExperimentalTensormapCreateOp: TT_Op<
-  "experimental_tensormap_create",
-  [
-    MemoryEffects<[MemRead<GlobalMemory>, MemWrite<GlobalMemory>]>,
-    AttrSizedOperandSegments,
-  ]
-> {
-  let summary = "Create a new TMA descriptor on device";
-  let arguments = (
-      ins
-      TT_PtrType:$desc_ptr,
-      TT_PtrType:$global_address,
-      Variadic<I32>:$box_dim,
-      Variadic<I32>:$global_dim,
-      Variadic<I64>:$global_stride,
-      Variadic<I32>:$element_stride,
-      ConfinedAttr<I32Attr, [IntNonNegative, IntMaxValue<15>]>:$elem_type,
-      ConfinedAttr<I32Attr, [IntNonNegative, IntMaxValue<2>]>:$interleave_layout,
-      ConfinedAttr<I32Attr, [IntNonNegative, IntMaxValue<3>]>:$swizzle_mode,
-      ConfinedAttr<I32Attr, [IntNonNegative, IntMaxValue<1>]>:$fill_mode
-  );
-  let extraClassDeclaration = [{
-      int32_t getRank() {
-          return getBoxDim().size();
-      }
-  }];
-  let assemblyFormat = [{
-    $desc_ptr `,` $global_address `,`
-    `[` $box_dim `]` `,`
-    `[` $global_dim `]` `,`
-    `[` $global_stride `]` `,`
-    `[` $element_stride `]`
-    attr-dict `:` functional-type(operands, results)
-  }];
-
-  let hasVerifier = 1;
-}
-
-def TT_ExperimentalTensormapFenceproxyAcquireOp: TT_Op<
-  "experimental_tensormap_fenceproxy_acquire",
-  [MemoryEffects<[MemWrite<GlobalMemory>]>]
-> {
-  let summary = "Acquire fence on a tensormap object";
-  let arguments = (ins TT_PtrType:$desc_ptr);
-  let assemblyFormat = [{
-    $desc_ptr attr-dict `:` qualified(type($desc_ptr))
-  }];
-}
-
 
 #endif // Triton_OPS

include/triton/Dialect/Triton/IR/TritonTypes.td

@@ -92,7 +92,7 @@ def TT_TensorPtr : TT_PtrOf<[TT_Tensor]>;
 // Any Type in Triton IR
 def TT_Type : AnyTypeOf<[TT_FloatLike, TT_IntLike, TT_PtrLike, TT_TensorPtr]>;
 
-// Result type of ExperimentalMakeTensorDescriptor
+// Result type of MakeTensorDescriptor
 def TT_TensorDescType : TritonTypeDef<"TensorDesc", "tensordesc", []> {
   let summary = "Tensor descriptor type (`::mlir::triton::TensorDescType`) in Triton IR type system";
 

include/triton/Dialect/TritonGPU/IR/LinearLayoutConversions.h

@@ -287,6 +287,11 @@ LinearLayout chooseScaledMfmaScaleLayout(
     const std::vector<std::vector<int32_t>> &dotOperandWarpBasis,
     ArrayRef<int64_t> dotOperandShape, unsigned mfmaMDim);
 
+// Create LinearLayout for nvidia mma tile.
+LinearLayout nvidiaMmaTile(MLIRContext *ctx, ArrayRef<unsigned> tileShape,
+                           unsigned kWidth, ArrayRef<unsigned> order,
+                           ArrayRef<unsigned> repOrder);
+
 // Create a LinearLayout similar to mfmaLayout, but changing each thread to hold
 // 8 elements. This layout is useful for emitting the widest 128-bit global
 // store instructions. Since it closely resembles mfmaLayout, conversion between

include/triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td

@@ -312,54 +312,46 @@ When vec=2, elements are swizzled in pairs of 2.  In other words, the element at
         if(!mmaEnc)
           return get(context, 1, 1, 1, order, CTALayout);
 
-        int opIdx = dotOpEnc.getOpIdx();
-        auto shapePerCTA = getShapePerCTA(CTALayout.getCTASplitNum(), shape);
-
-        // number of rows per phase
-
-        // index of the inner dimension in `order`
-        unsigned inner = (opIdx == 0) ? 0 : 1;
-
         // ---- begin Ampere & Hopper ----
         if (mmaEnc.isAmpere() || mmaEnc.isHopper()) {
-          int perPhase = 128 / (std::max<int>(1, shapePerCTA[order[0]] * 4 / dotOpEnc.getKWidth()));
-          perPhase = std::max<int>(perPhase, 1);
-          std::vector<size_t> matShape = {8, 8, 4 * dotOpEnc.getKWidth()};
-          int vecWidth = 32 / typeWidthInBit;
-          if (vecWidth != dotOpEnc.getKWidth() && order[0] == inner) {
-              perPhase = std::max<int>(perPhase, 2 * vecWidth);
-          }
-          int rank = order.size();
-          // --- handle A operand ---
-          if (opIdx == 0) { // compute swizzling for A operand
-              int m = (needTrans) ? matShape[2] : matShape[0];
-              int k = (needTrans) ? matShape[0] : matShape[2];
-              int vec = (order[0] == rank-1) ? k : m;
-              int mmaStride = (order[0] == rank-1) ? m : k;
-              int maxPhase = std::max(mmaStride / perPhase, 1);
-              return get(context, vec, perPhase, maxPhase, order, CTALayout);
-          }
-
-          // --- handle B operand ---
-          if (opIdx == 1) {
-              // we compute vec and maxPhase m, n and k size of the mma
-              // instruction. when matmul operands is transposed, we should
-              // consider that to get m, n and k.
-              int n = needTrans ? matShape[2] : matShape[1];
-              int k = needTrans ? matShape[1] : matShape[2];
-              int vec = (order[0] == rank-1) ? n : k;
-              int mmaStride = (order[0] == rank-1) ? k : n;
-              int maxPhase = std::max(mmaStride / perPhase, 1);
-              return get(context, vec, perPhase, maxPhase, order, CTALayout);
-          }
-
-          llvm_unreachable("invalid operand index");
+          return get(context, dotOpEnc.getOpIdx(), dotOpEnc.getKWidth(), shape, order, CTALayout, typeWidthInBit, needTrans);
         }
 
         // ---- not implemented ----
         llvm_unreachable("unsupported swizzling for provided MMA version");
     }]>,
 
+    // NVIDIA constructor!
+    // TODO(lezcano): We should totally get rid of all these constructors...
+    AttrBuilder<(ins "int":$opIdx,
+                     "unsigned":$kWidth,
+                     "ArrayRef<int64_t>":$shape,
+                     "ArrayRef<unsigned>":$order,
+                     "CTALayoutAttr":$CTALayout,
+                     "unsigned":$bitwidth,
+                     "bool":$needTrans), [{
+        int K =  getShapePerCTA(CTALayout.getCTASplitNum(), shape)[order[0]];
+        // Elems necessary to cover all the banks divided by the inner dimension
+        // This packs a few rows together for small K
+        int perPhase = std::max<int>(1024 / (bitwidth * K), 1);
+
+        int mmaStride = 8;
+        int vec = 4 * kWidth;
+        // needsTrans is equiv. to flipping the opIdx
+        if (needTrans)
+          std::swap(vec, mmaStride);
+        assert(opIdx == 0 || opIdx == 1);
+        int rank = order.size();
+        int kDim = opIdx == 0 ? rank-1 : rank-2;
+        if (order[0] != kDim)
+          std::swap(vec, mmaStride);
+        // Count how many vec elements are needed to cover all the banks
+        int maxPhase = std::max(std::min<int>(mmaStride, 1024 / (vec * bitwidth)), 1);
+        // Account for the row packing from perPhase: mmaStride / perPhase
+        maxPhase = std::max(maxPhase / perPhase, 1);
+        return get(context, vec, perPhase, maxPhase, order, CTALayout);
+    }]>,
+
     AttrBuilder<(ins "DotOperandEncodingAttr":$dotOpEnc,
                      "ArrayRef<int64_t>":$shape,
                      "ArrayRef<unsigned>":$order,