
// RUN: iree-opt %s --iree-hal-target-backends=cuda \
// RUN:     --iree-abi-transformation-pipeline \
// RUN:     --iree-flow-transformation-pipeline  \
/// This must be used with the custom dispatch region formation
/// because IREE's does not fuse the 6 ops softmax version even with
/// --iree-flow-fuse-multi-use.
// RUN:     --iree-flow-dispatch-use-transform-dialect=%p/softmax_dispatch_spec.mlir \
// RUN:     --iree-stream-transformation-pipeline \
// RUN:     --iree-hal-configuration-pipeline | \
// RUN: iree-opt --pass-pipeline='builtin.module(hal.executable(hal.executable.variant(iree-llvmgpu-lower-executable-target)))' \
// RUN:     --iree-codegen-llvmgpu-use-transform-dialect=%p/softmax_codegen_spec.mlir | \
// RUN: FileCheck %s --check-prefix=CHECK-SHUFFLE

// RUN: iree-compile %s --iree-hal-target-backends=cuda \
// RUN:     --iree-opt-const-expr-hoisting=false --iree-opt-const-eval=false \
/// Constant JIT'ing must be disabled because the transform-dialect debug
/// flags leak to the JIT session, which doesn't know what to do with them.
/// This must be used with the custom dispatch region formation
/// because IREE's does not fuse the 6 ops softmax version even with
/// --iree-flow-fuse-multi-use.
// RUN:     --iree-flow-dispatch-use-transform-dialect=%p/softmax_dispatch_spec.mlir \
// RUN:     --iree-codegen-llvmgpu-use-transform-dialect=%p/softmax_codegen_spec.mlir | \
// RUN: iree-run-module --module=- --function=softmax --device=cuda | \
// RUN: FileCheck %s


!tmp_tensor_t = tensor<16x128xf32>
!in_tensor_t = tensor<16x128x128xf32>
!out_tensor_t = tensor<16x128x128xf32>

// Compilation checks that shuffles are produced.
// CHECK-SHUFFLE: vector.reduction <maxf>
// CHECK-SHUFFLE-COUNT-5: gpu.shuffle  xor
// CHECK-SHUFFLE: vector.reduction <add>
// CHECK-SHUFFLE-COUNT-5: gpu.shuffle  xor

// Execution only checks that @softmax runs.
//      CHECK: EXEC @softmax
//      CHECK: 16x128x128xf32=[
// CHECK-SAME:                [0.0078125 0.0078125 0.0078125 0.0078125

func.func @softmax() -> !out_tensor_t {
  %cst_0 = arith.constant 0.0 : f32
  %cst_1 = arith.constant 1.0 : f32
  %cst_min = arith.constant -3.40282347E+38 : f32
  %input = arith.constant dense<5.000000e+00> : !out_tensor_t
  util.optimization_barrier %input : !in_tensor_t

  %input_max_empty = tensor.empty() : !tmp_tensor_t
  %input_max_filled = linalg.fill ins(%cst_min : f32)
    outs(%input_max_empty : !tmp_tensor_t) -> !tmp_tensor_t
  %input_max = linalg.generic
    {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>,
                      affine_map<(d0, d1, d2) -> (d0, d1)>],
                      iterator_types = ["parallel", "parallel", "reduction"]}
     ins(%input : !in_tensor_t)
    outs(%input_max_filled : !tmp_tensor_t) {
      ^bb0(%arg0: f32, %arg1: f32):
        %max = arith.maxf %arg0, %arg1 : f32
        linalg.yield %max : f32
      } -> !tmp_tensor_t

  // This has been fused manually to avoid the fusion on tensors pass and reduce noise atm.
  %exps_empty = tensor.empty() : !out_tensor_t
  %exps_sum_empty = tensor.empty() : !tmp_tensor_t
  %exps_sum_filled = linalg.fill ins(%cst_0 : f32)
    outs(%exps_sum_empty : !tmp_tensor_t) -> !tmp_tensor_t
  %exps = linalg.generic
    {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>,
                      affine_map<(d0, d1, d2) -> (d0, d1)>,
                      affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
                      iterator_types = ["parallel", "parallel", "parallel"]}
     ins(%input, %input_max : !in_tensor_t, !tmp_tensor_t)
    outs(%exps_empty : !out_tensor_t) {
      ^bb0(%arg0: f32, %arg1: f32, %arg2: f32):
        %sub = arith.subf %arg0, %arg1 : f32
        %exp = math.exp %sub : f32
        linalg.yield %exp: f32
      } -> (!out_tensor_t)

  %exps_sum = linalg.generic
    {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>,
                      affine_map<(d0, d1, d2) -> (d0, d1)>],
                      iterator_types = ["parallel", "parallel", "reduction"]}
     ins(%exps : !out_tensor_t)
    outs(%exps_sum_filled : !tmp_tensor_t) {
      ^bb0(%exp: f32, %acc: f32):
        %add = arith.addf %exp, %acc : f32
        linalg.yield %add : f32
      } -> (!tmp_tensor_t)

  %res_empty = tensor.empty() : !out_tensor_t
  %res = linalg.generic
    {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>,
                      affine_map<(d0, d1, d2) -> (d0, d1)>,
                      affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
                      iterator_types = ["parallel", "parallel", "parallel"]}
     ins(%exps, %exps_sum : !out_tensor_t, !tmp_tensor_t)
    outs(%res_empty : !out_tensor_t) {
      ^bb0(%arg0: f32, %arg1: f32, %arg2: f32):
        // %10 = arith.divf %cst_1, %arg1 : f32
        // %11 = arith.mulf %arg0, %10 : f32
        %div = arith.divf %arg0, %arg1 : f32
        linalg.yield %div : f32
      } -> !out_tensor_t

  return %res: !out_tensor_t
}