UPSTREAM PR #17502: HIP: Patch failed testcase in WMMA-MMQ kernels for RDNA 4

[loci-dev]

Mirrored from ggml-org/llama.cpp#17502

1. Patched failed test case MUL_MAT(type_a=q4_0,type_b=f32,m=576,n=512,k=576,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1) for enabling WMMA on RDNA4 (verified all test cases passing when running ./build/bin/test-backend-ops test -o MUL_MAT

2. Quick clean up on mma.cuh to add ggml_cuda_memcpy_1 back in for half2 and bfloat162

for ggml-org/llama.cpp#17156

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Review Summary: PR #326

Analysis Scope: Comparing version 6b1d7254-fefa-44ba-8e76-256501ca6ef9 against baseline aab9b31c-ad35-48ba-b9fe-4c0fd3dc2df2

Condition Assessment: Condition 1 applies - no measurable performance changes detected.

PR #326 introduces AMD RDNA4 WMMA support through two targeted code changes in CUDA kernel files. Analysis of 15 performance-critical functions across 16 binaries shows zero measurable impact on Response Time, Throughput Time, and power consumption. All functions report is_modified: false, indicating the changes are conditionally compiled and inactive in the analyzed build configuration.

Code Changes:

• ggml/src/ggml-cuda/mma.cuh: Added type-based branching for FP16/BF16 tile loading using ggml_cuda_memcpy_1
• ggml/src/ggml-cuda/mmq.cuh: Added amd_wmma_available(cc) to shared memory calculation

Performance Metrics:

• Power consumption: 12 of 16 binaries unchanged; 4 binaries show sub-nanojoule deltas
• Core inference functions unchanged: llama_decode (44752296 ns), llama_encode (11253996 ns), llama_tokenize (899199 ns)
• Tokens per second: No impact - inference functions show identical response times

Conclusion: Changes are architecture-specific correctness fixes with no runtime impact on current build. RDNA4-targeted builds would benefit from enabled WMMA functionality.