SWDEV-558359 - [Topaz][Gigapixel][Performance] The clc 512x512 models are running with overall lower performance for MIGraphX EP than for DirectML EP

[KenLagos]

Description

Bug Fix: large integers incorrectly converted to -1 in IsUnsupportedOpMode() function

prevent casting of int64 to int32 in IsUnsupportedOpMode() function for "Slice" operator

Motivation and Context

Currently clc models in topaz underperform DML by 52x and the root cause is due to the "Slice" node being flagged as unsupported.

This change fixes an issue where start and end attributes for the slice node were being converted to -1 instead of their real int64 value. The -1 end or start value was causing migraphx/onnxruntime to incorrectly flag this node as unsupported.

Local testing on my machine suggests inference time goes from 90 seconds to 3 seconds which is massive boost in performance and closes the gap between DML and Migraphx performance considerably.

[Zhaeong]

This will need to be cherry-picked into wml-main after merging

[TedThemistokleous]

What in the slice check is being unsupported? Why are you modifying code that's doing index checks?

[TedThemistokleous]

The fix is in toVector call to use a larger size, its using int which may default to 32 bits on some systems. If you're seeing rollover that's likely the culprit

Either make this a larger type to avoid casting things down to int32 (since int-> unsigned) or use size_t/int64_t

std::vector<int> toVector(const ONNX_NAMESPACE::int64s& nums) {
  std::vector<int> result;
  int num = nums.size();
  for (int i = 0; i < num; ++i) {
    result.push_back(static_cast<int>(nums[i]));
  }

  return result;
}

Also don't remove the .at() bounds checks here. These shouldn't be removed for direct access [ ]

[TedThemistokleous]

@Zhaeong good to cherry-pick this off the tip of rocm7.1_internal_testing. I'll upstream this to OnnxRT

[TedThemistokleous]

Upstreamed here - microsoft#26403

[Zhaeong]

Merged to wml-main:
#189