[iq136boy]

Motivation

ConvOclDirectFwd1x1 is a legacy OpenCL-based solver for 1x1 forward convolution. As MIOpen continues to migrate away from OpenCL toward HIP, this solver has become a maintenance burden: it relies on three OpenCL kernel files (~1200 lines of OpenCL C), has entries in the find-db for several GPU architectures, and carries the overhead of the OpenCL compilation path. Removing it reduces technical debt and aligns with the broader effort to standardize on HIP-based kernels.

1×1 convolution is mathematically equivalent to a matrix multiply (GEMM), and MIOpen's GEMM-based and CK-based solvers are specifically optimized for exactly this. In practice, those solvers win the find-db tuning for virtually all 1×1 problem shapes. ConvOclDirectFwd1x1 was already rarely — if ever — selected as the best solver, which is precisely why its find-db entries are being cleaned up in this PR.

No HIP replacement solver is introduced. ConvDirectNaiveConvFwd already covers the same problem space (and more — NHWC layout, groups, padding, dilation, INT8/FP8) by dispatching to the same conv_internal::GetConv2DFWDSolution() function that any replacement would use. Since there is no new HIP kernel and no performance difference, a separate solver wrapper for 1×1 problems adds complexity with no benefit.

Additionally, the old OCL solver carried accumulated tech debt that makes it not worth porting:

• It was restricted to a subset of GPU architectures (gfx8, gfx90a, gfx103), requiring explicit device guards
• It was disabled on gfx10/gfx11 via WORKAROUND_SWDEV_271887 due to precision issues
• Padding support was disabled as a workaround for correctness bugs
• It had a history of FP16 precision fixes

Technical Details

• Removed ConvOclDirectFwd1x1 solver (src/solver/conv/conv_ocl_dir2Dfwd1x1.cpp, ~433 lines) and its three OpenCL kernel files (MIOpenConv1x1J1.cl, MIOpenConv1x1J1_stride.cl, MIOpenConv1x1S.cl, ~1200 lines total).
• Skipped the solver ID slot in solver.cpp with ++id; // removed ConvOclDirectFwd1x1 to preserve backward compatibility with existing find-db entries.
• ConvDirectNaiveConvFwd acts as the fallback for any 1×1 problem shapes not claimed by higher-performance solvers.
• Updated find-db binary files for gfx803, gfx900, gfx906, gfx908, and gfx90a to remove entries referencing the old solver.

Test Plan

• Existing unit tests for ConvDirectNaiveConvFwd cover the problem space previously handled by ConvOclDirectFwd1x1.
• CI test

Test Result

CI test pending

Submission Checklist

• Removed legacy OpenCL solver and all associated kernel files.
• Updated find-db, CMake, fin interface, and documentation.
• No new solver introduced; ConvDirectNaiveConvFwd serves as the fallback.

[AnaghaRaoAMD]

LGTM, minor suggestions posted in comments, needs rebase and ensure there is no accuracy/performance regressions on any gfx platform.

[AnaghaRaoAMD]

Rethinking on this some more, we are not introducing any new hip kernel here, there's no real reason for ConvHipDirectFwd1x1 to exist as a separate solver.

1. Same kernel — Both ConvHipDirectFwd1x1 and ConvDirectNaiveConvFwd call the exact same function: conv_internal::GetConv2DFWDSolution()
2. ConvDirectNaiveConvFwd is a superset — It accepts all layouts (NCHW + NHWC), all filter sizes, all strides, all padding, all dilation, groups, more data types (FP8, Int8 in addition to FP16/BFP16/FP32). Every problem that ConvHipDirectFwd1x1 would accept, ConvDirectNaiveConvFwd already accepts.
3. No performance difference — Since they dispatch to the same kernel, there's zero performance advantage to having the 1x1-specific wrapper.

If there's no specific reason to keep a separate solver ID for 1x1 cases, the simpler approach is to just delete ConvOclDirectFwd1x1, skip the ID slot with ++id; in solver.cpp, and let ConvDirectNaiveConvFwd handle the problem space. That would eliminate the new solver file, new test file, and the added complexity entirely.

[iq136boy]

Rethinking on this some more, we are not introducing any new hip kernel here, there's no real reason for ConvHipDirectFwd1x1 to exist as a separate solver.

1. Same kernel — Both ConvHipDirectFwd1x1 and ConvDirectNaiveConvFwd call the exact same function: conv_internal::GetConv2DFWDSolution()
2. ConvDirectNaiveConvFwd is a superset — It accepts all layouts (NCHW + NHWC), all filter sizes, all strides, all padding, all dilation, groups, more data types (FP8, Int8 in addition to FP16/BFP16/FP32). Every problem that ConvHipDirectFwd1x1 would accept, ConvDirectNaiveConvFwd already accepts.
3. No performance difference — Since they dispatch to the same kernel, there's zero performance advantage to having the 1x1-specific wrapper.

If there's no specific reason to keep a separate solver ID for 1x1 cases, the simpler approach is to just delete ConvOclDirectFwd1x1, skip the ID slot with ++id; in solver.cpp, and let ConvDirectNaiveConvFwd handle the problem space. That would eliminate the new solver file, new test file, and the added complexity entirely.

Yes. You are right! The new hip solver is not needed. I will address it. Thanks.