Address libmathdx-related compile time regressions introduced with libmathdx 0.2.1 #809
Description
Performance testing revealed some regressions in the compile times with Warp 1.8 resulting from the libmathdx upgrade, for example:
| Change | Before [a447d70c] <v1.7.2^0> | After [e74e05be] <release-1.8~3> | Ratio | Benchmark (Parameter) |
|----------|--------------------------------|------------------------------------|---------|----------------------------------------------------------------|
| + | 2.29±0s | 4.48±0.01s | 1.96 | tile.compile_cholesky.ColdCompileCholeskyLTO.time_cuda_codegen |
| + | 1.29±0s | 1.73±0s | 1.35 | tile.compile_cholesky.WarmCompileCholeskyLTO.time_cuda_codegen |
| + | 7.86±0.02s | 11.8±0.08s | 1.5 | tile.compile_fft.ColdCompileFftLTO.time_cuda_codegen |
| + | 2.02±0s | 2.91±0s | 1.44 | tile.compile_fft.WarmCompileFftLTO.time_cuda_codegen |
| - | 13.8±0.02s | 12.2±0.01s | 0.88 | tile.compile_gemm.ColdCompileGemmLTO.time_cuda_codegen |