[XPU][OptRed] Define triton_intel_gpu.simd_reduce and use in optimi…

…zed transposed reduction

Define SIMD transpose-reduce operation performing a SIMD reduction while transposing
the implicit SIMD matrix. See description definition for further context.

Using this operation in the transpose reduction pass allows us to perform the optimization
while not using SLM.

Signed-off-by: victor-eds <victor.perez@codeplay.com>
Co-authored-by: chengjunlu <chengjun.lu@intel.com>
Signed-off-by: Victor Perez <victor.perez@codeplay.com>

test/Conversion/intel/simd-reduce.mlir

@@ -0,0 +1,24 @@
+// RUN: triton-opt %s -split-input-file --intel-allocate-shared-memory --convert-triton-intel-gpu-to-llvm | FileCheck %s
+
+// Basic 16x16 SIMD reduction.
+
+#blocked = #ttg.blocked<{sizePerThread = [1, 16], threadsPerWarp = [16, 1], warpsPerCTA = [1, 1], order = [0, 1]}>
+#blocked1 = #ttg.blocked<{sizePerThread = [1], threadsPerWarp = [16], warpsPerCTA = [1], order = [0]}>
+
+module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, "ttg.threads-per-warp" = 16 : i32} {
+// CHECK-LABEL:   llvm.func spir_kernelcc @test_single(
+// CHECK-SAME:                                         %[[VAL_0:.*]]: !llvm.struct
+// CHECK:           %[[VAL_17:.*]] = llvm.mlir.poison : vector<16xf32>
+// COM: Check we insert all tensor elements in a vector:
+// CHECK-COUNT-16:  llvm.insertelement
+// CHECK:           %[[VAL_50:.*]] = llvm.inline_asm has_side_effects asm_dialect = att operand_attrs = [] "{\0A.decl temp_result v_type=G type=f num_elts=128 align=wordx32\0Aadd (M1_NM, 16) temp_result(0, 0)<1> $1(0, 0)<16;8,1> $1(0, 8)<16;8,1>\0Aadd (M1_NM, 16) temp_result(1, 0)<1> $1(2, 0)<16;8,1> $1(2, 8)<16;8,1>\0Aadd (M1_NM, 16) temp_result(2, 0)<1> $1(4, 0)<16;8,1> $1(4, 8)<16;8,1>\0Aadd (M1_NM, 16) temp_result(3, 0)<1> $1(6, 0)<16;8,1> $1(6, 8)<16;8,1>\0Aadd (M1_NM, 16) temp_result(4, 0)<1> $1(8, 0)<16;8,1> $1(8, 8)<16;8,1>\0Aadd (M1_NM, 16) temp_result(5, 0)<1> $1(10, 0)<16;8,1> $1(10, 8)<16;8,1>\0Aadd (M1_NM, 16) temp_result(6, 0)<1> $1(12, 0)<16;8,1> $1(12, 8)<16;8,1>\0Aadd (M1_NM, 16) temp_result(7, 0)<1> $1(14, 0)<16;8,1> $1(14, 8)<16;8,1>\0Aadd (M1_NM, 16) temp_result(0, 0)<1> temp_result(0, 0)<8;4,1> temp_result(0, 4)<8;4,1>\0Aadd (M1_NM, 16) temp_result(1, 0)<1> temp_result(2, 0)<8;4,1> temp_result(2, 4)<8;4,1>\0Aadd (M1_NM, 16) temp_result(2, 0)<1> temp_result(4, 0)<8;4,1> temp_result(4, 4)<8;4,1>\0Aadd (M1_NM, 16) temp_result(3, 0)<1> temp_result(6, 0)<8;4,1> temp_result(6, 4)<8;4,1>\0Aadd (M1_NM, 16) temp_result(0, 0)<1> temp_result(0, 0)<4;2,1> temp_result(0, 2)<4;2,1>\0Aadd (M1_NM, 16) temp_result(1, 0)<1> temp_result(2, 0)<4;2,1> temp_result(2, 2)<4;2,1>\0Aadd (M1_NM, 16) $0(0, 0)<1> temp_result(0, 0)<2;1,0> temp_result(0, 1)<2;1,0>\0A}", "=rw,rw" %{{.*}} : (vector<16xf32>) -> f32
+// COM: Check we obtain a single result, i.e., the SIMD reduction minimizes register usage.
+// CHECK:           %[[VAL_51:.*]] = llvm.mlir.undef : !llvm.struct<(f32)>
+// CHECK:           %[[VAL_52:.*]] = llvm.insertvalue %[[VAL_50]], %[[VAL_51]][0] : !llvm.struct<(f32)>
+// CHECK:           llvm.return %[[VAL_52]] : !llvm.struct<(f32)>
+// CHECK:         }
+  tt.func @test_single(%arg0: tensor<16x16xf32, #blocked>) -> tensor<16xf32, #blocked1> {
+    %0 = triton_intel_gpu.simd_reduce add %arg0 axis = 0 : tensor<16x16xf32, #blocked> -> tensor<16xf32, #blocked1>
+    tt.return %0 : tensor<16xf32, #blocked1>
+  }
+}

test/TritonIntelGPU/tritonintelgpu.mlir

@@ -58,3 +58,17 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 8 : i32, "ttg.thr
     tt.return %res : tensor<16x16xf16>
   }
 }
+
+// -----
+
+#blocked = #ttg.blocked<{sizePerThread = [1, 16], threadsPerWarp = [16, 1], warpsPerCTA = [1, 1], order = [0, 1]}>
+#blocked1 = #ttg.blocked<{sizePerThread = [1], threadsPerWarp = [16], warpsPerCTA = [1], order = [0]}>
+
+module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, "ttg.threads-per-warp" = 16 : i32} {
+  tt.func @triton_intel_gpu.simd_reduce(%arg0: tensor<16x16xf32, #blocked>) -> tensor<16xf32, #blocked1> {
+    // CHECK-LABEL: @triton_intel_gpu.simd_reduce
+    // CHECK:         triton_intel_gpu.simd_reduce add %{{.*}} axis = 0 : tensor<16x16xf32, #blocked> -> tensor<16xf32, #blocked1>
+    %0 = triton_intel_gpu.simd_reduce add %arg0 axis = 0 : tensor<16x16xf32, #blocked> -> tensor<16xf32, #blocked1>
+    tt.return %0 : tensor<16xf32, #blocked1>
+  }
+}

third_party/intel/include/Dialect/TritonIntelGPU/IR/TritonIntelGPUOps.td

@@ -12,6 +12,7 @@
 include "triton/Dialect/Triton/IR/TritonTypes.td"
 include "triton/Dialect/Triton/IR/TritonAttrDefs.td"
 include "triton/Dialect/TritonGPU/IR/TritonGPUTypes.td"
+include "intel/include/Dialect/TritonGEN/IR/TritonGENAttrDefs.td"
 include "intel/include/Dialect/TritonIntelGPU/IR/TritonIntelGPUAttrDefs.td"
 include "intel/include/Dialect/TritonIntelGPU/IR/TritonIntelGPUDialect.td"
 include "mlir/Interfaces/SideEffectInterfaces.td"
@@ -202,4 +203,66 @@ def TTIG_SubGroupTransposeOp
   let hasVerifier = 1;
 }
 
+def TTIG_SIMDReduceOp : TTIG_Op<"simd_reduce", [Pure, SameOperandsAndResultElementType]> {
+  let summary = "SIMD reduction.";
+  let description = [{
+    The `triton_intel_gpu.simd_reduce` operation performs a SIMD reduction.
+    Contrary to `tt.reduce`, when performing a warp reduction, the result is
+    non-uniform.
+
+    The reduction axis must be in such a way that only a warp reduction is
+    performed, i.e., `sizePerThread[axis]`, `warpsPerCTA[axis]` and
+    `CTAsPerCGA[axis]` must be 1; and `shape[axis]` and `threadsPerWarp[axis]`
+    must be equal to the sub-group size.
+
+    The output type must be compatible with the performed reduction. However,
+    ensuring this is up to the user. As a rule of thumb, the total number of
+    elements in the output tensor must be sub-group size smaller than in the
+    original one. Users should bear in mind a tensor like:
+
+    ```
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    ```
+
+    would be reduced to:
+
+    ```
+    t0 t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15
+    ```
+
+    Example:
+    ```mlir
+    #blocked = #ttg.blocked<{sizePerThread = [1, 16], threadsPerWarp = [16, 1], warpsPerCTA = [1, 1], order = [0, 1]}>
+    #blocked1 = #ttg.blocked<{sizePerThread = [1], threadsPerWarp = [16], warpsPerCTA = [1], order = [0]}>
+    triton_intel_gpu.simd_reduce add %0 axis = 0 : tensor<16x16xf32, #blocked> -> tensor<16xf32, #blocked1>
+    // # 3D reduction:
+    #blocked = #ttg.blocked<{sizePerThread = [1, 16, 1], threadsPerWarp = [16, 1, 1], warpsPerCTA = [1, 1, 2], order = [0, 1, 2]}>
+    #blocked1 = #ttg.blocked<{sizePerThread = [1, 1], threadsPerWarp = [16, 1], warpsPerCTA = [1, 2], order = [0, 1]}>
+    triton_intel_gpu.simd_reduce add %0 axis = 0 : tensor<16x16x2xf32, #blocked> -> tensor<16x2xf32, #blocked1>
+    ```
+  }];
+  let arguments = (ins TT_Tensor:$src,
+		       TritonGEN_ReduceKindAttr: $op,
+		       I32Attr:$axis);
+  let results = (outs TT_Tensor:$res);
+  let assemblyFormat = [{
+    $op $src `axis` `=` $axis attr-dict `:` type($src) `->` type($res)
+  }];
+}
+
 #endif