IREE run module crash due to addition of MFMA op. https://github.com/iree-org/iree/pull/17921

[pashu123]

What happened?

iree-run-module --module=test_rocm.vmfb --device=hip  --input=2x130x130x4xf16=1.0 --input=3x3x4x320xf16=1.0 --input=320xf32=1.0

iree/runtime/src/iree/hal/drivers/hip/native_executable.c:186: INTERNAL; HIP driver error 'hipErrorNoBinaryForGpu' (209): no kernel image is available for execution on the device; mismatched target chip? missing/wrong bitcode directory?; while invoking native function hal.executable.create; while calling import;
[ 1]   native hal.executable.create:0 -
[ 0] bytecode module.__init:306 test.mlir:13:11
      at test.mlir:1:1; creating VM context; creating run context

Steps to reproduce your issue

Cherry-pick the commit from here: https://github.com/iree-org/iree/pull/17921/commits

Repro MLIR:

func.func @test(%5: tensor<2x130x130x4xf16>, %6: tensor<3x3x4x320xf16>, %7: tensor<320xf32>) -> ( tensor<2x128x128x320xf16>, tensor<2x320x128x128xf16> )  {
  %cst = arith.constant 0.000000e+00 : f32
  %c286016 = arith.constant 286016 : index
  %c22976 = arith.constant 22976 : index
  %c21696 = arith.constant 21696 : index
  %c556416 = arith.constant 556416 : index
  %c21527936 = arith.constant 21527936 : index
  %8 = tensor.empty() : tensor<2x320x128x128xf16>
  %9 = tensor.empty() : tensor<2x128x128x320xf16>
  %10 = tensor.empty() : tensor<2x128x128x320xf32>
  %11 = linalg.fill  ins(%cst : f32) outs(%10 : tensor<2x128x128x320xf32>) -> tensor<2x128x128x320xf32>
  %12 = linalg.conv_2d_nhwc_hwcf {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%5, %6 : tensor<2x130x130x4xf16>, tensor<3x3x4x320xf16>) outs(%11 : tensor<2x128x128x320xf32>) -> tensor<2x128x128x320xf32>
  %13:2 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d3, d1, d2)>], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%12, %7 : tensor<2x128x128x320xf32>, tensor<320xf32>) outs(%9, %8 : tensor<2x128x128x320xf16>, tensor<2x320x128x128xf16>)  {
  ^bb0(%in: f32, %in_0: f32, %out: f16, %out_1: f16):
    %14 = arith.addf %in, %in_0 : f32
    %15 = arith.truncf %14 : f32 to f16
    linalg.yield %15, %15 : f16, f16
  } -> (tensor<2x128x128x320xf16>, tensor<2x320x128x128xf16>)
  return %13#0, %13#1 : tensor<2x128x128x320xf16>, tensor<2x320x128x128xf16>
}

Compile command:

iree-compile test.mlir --iree-hal-target-backends=rocm --iree-rocm-target-chip=gfx942 --iree-opt-const-eval=false --iree-global-opt-propagate-transposes=true --iree-global-opt-enable-fuse-horizontal-contractions=true --iree-flow-enable-aggressive-fusion=true --iree-opt-aggressively-propagate-transposes=true --iree-opt-outer-dim-concat=true --iree-vm-target-truncate-unsupported-floats --iree-llvmgpu-enable-prefetch=true --iree-opt-data-tiling=false --iree-codegen-gpu-native-math-precision=true --iree-codegen-llvmgpu-use-vector-distribution --iree-rocm-waves-per-eu=2 --iree-execution-model=async-external "--iree-preprocessing-pass-pipeline=builtin.module(iree-preprocessing-transpose-convolution-pipeline, util.func(iree-preprocessing-pad-to-intrinsics))" --iree-scheduling-dump-statistics-format=json --iree-scheduling-dump-statistics-file=compilation_info.json -o test_rocm.vmfb

Run command:

iree-run-module --module=test_rocm.vmfb --device=hip --input=2x130x130x4xf16=1.0 --input=3x3x4x320xf16=1.0 --input=320xf32=1.0

What component(s) does this issue relate to?

Runtime

Version information

No response

Additional context

No response

[pashu123]

Added the fix here: #17921