[MLIR] Optimized linear layer in PyTorch frontend

[Moehre2]

• Optimized linear layer in PyTorch frontend by enabling the MatmulNode to expand to GEMM node with transpose
• Restructured daisy workflows for PyTorch frontend
• Now, build time of models in PyTorch daisy workflows is not included in the runtime

[daisytuner]

Daisytuner Report - python_npbench (zinnia)

@@                                   Benchmarks                                   @@
=====================================================================================
  Benchmark              Time        ΔTime       Thr         Energy      ΔEnergy     
=====================================================================================
# adi_numpy              1.32 s      +0.02%      N/A         131.28 J    +0.12%      
# adi_omp                15.95 s     -0.10%      N/A         1505.45 J   -0.11%      
# adi_cuda               4.80 s      +1.09%      N/A         465.19 J    +1.08%      
# adi_seq_tuning         15.85 s     -0.42%      N/A         1495.47 J   -0.62%      
# atax_numpy             2.15 s      -0.50%      N/A         222.81 J    -0.43%      
# atax_omp               2.47 s      +0.06%      N/A         259.24 J    +0.09%      
# atax_cuda              4.11 s      +0.17%      N/A         422.88 J    +0.15%      
# atax_seq_tuning        3.71 s      +0.10%      N/A         375.30 J    +0.07%      
# gemm_numpy             1.23 s      -1.27%      N/A         198.41 J    -1.10%      
# gemm_omp               1.11 s      +0.28%      N/A         162.01 J    +0.28%      
# gemm_cuda              10.63 s     +0.46%      N/A         1010.63 J   +0.51%      
# gemm_seq_tuning        1.11 s      -0.01%      N/A         161.73 J    -0.11%      
# gesummv_numpy          1.75 s      -0.92%      N/A         249.97 J    -1.06%      
# gesummv_omp            5.33 s      +0.35%      N/A         692.01 J    +0.25%      
# gesummv_cuda           8.40 s      +0.82%      N/A         1004.61 J   +0.59%      
# gesummv_seq_tuning     6.55 s      -0.41%      N/A         805.20 J    -0.13%      
# gemver_numpy           1.08 s      -0.38%      N/A         166.97 J    -0.53%      
# gemver_omp             714.68 ms   +0.56%      N/A         81.59 J     +0.97%      
# gemver_cuda            3.86 s      +0.63%      N/A         386.49 J    +0.18%      
# gemver_seq_tuning      4.48 s      +0.84%      N/A         433.74 J    +0.91%      
# k2mm_numpy             1.19 s      -0.78%      N/A         197.27 J    -0.54%      
# k2mm_omp               3.62 s      +0.40%      N/A         469.79 J    +0.47%      
# k2mm_cuda              13.60 s     +0.16%      N/A         1287.13 J   +0.12%      
# k2mm_seq_tuning        3.62 s      -0.11%      N/A         468.00 J    +0.11%      
# k3mm_numpy             1.03 s      -0.24%      N/A         184.19 J    -0.14%      
# k3mm_omp               5.72 s      -0.37%      N/A         796.84 J    -0.13%      
# k3mm_cuda              19.85 s     +0.25%      N/A         1868.92 J   +0.19%      
# k3mm_seq_tuning        5.72 s      -0.16%      N/A         792.39 J    -0.15%      
# mvt_numpy              2.42 s      -0.56%      N/A         247.59 J    -0.71%      
# mvt_omp                2.74 s      -0.09%      N/A         284.61 J    -0.15%      
# mvt_cuda               3.36 s      +0.54%      N/A         342.67 J    +0.35%      
# mvt_seq_tuning         2.75 s      +0.14%      N/A         285.10 J    +0.18%      
# symm_numpy             781.78 ms   -0.59%      N/A         80.62 J     -0.19%      
# symm_omp               8.42 s      +0.41%      N/A         802.01 J    +0.31%      
# symm_seq_tuning        8.36 s      -0.12%      N/A         796.53 J    -0.09%      
# syr2k_numpy            882.10 ms   +0.23%      N/A         89.64 J     -0.23%      
# syr2k_omp              9.81 s      -0.03%      N/A         931.91 J    -0.19%      
# syr2k_cuda             1.64 s      +0.08%      N/A         170.18 J    +0.23%      
# syr2k_seq_tuning       9.80 s      +0.24%      N/A         931.30 J    +0.28%      
# syrk_numpy             775.56 ms   +0.24%      N/A         80.09 J     +0.48%      
# syrk_omp               5.95 s      +0.93%      N/A         572.76 J    +1.03%      
# syrk_cuda              1.52 s      +0.53%      N/A         159.28 J    +0.86%      
# syrk_seq_tuning        5.90 s      +0.33%      N/A         567.60 J    +0.42%      
# trmm_numpy             879.33 ms   +0.22%      N/A         89.52 J     +0.03%      
# trmm_omp               3.10 s      +0.05%      N/A         306.07 J    +0.19%      
# trmm_seq_tuning        3.36 s      +1.10%      N/A         321.51 J    +1.10%      

[daisytuner]

Daisytuner Report - mlir_torch_models (chamomile)

@@                                   Benchmarks                                   @@
=====================================================================================
  Benchmark              Time        ΔTime       Thr         Energy      ΔEnergy     
=====================================================================================
# bn_conv_bn_relu_maxpool_torch18.60 s     N/A         N/A         3616.28 J   N/A         
# bn_conv_bn_relu_maxpool_run_none4.55 s      N/A         N/A         828.56 J    N/A         
# bn_conv_bn_relu_maxpool_run_sequential4.61 s      N/A         N/A         876.99 J    N/A         
# bn_conv_bn_relu_maxpool_run_openmp4.55 s      N/A         N/A         811.26 J    N/A         
# bn_conv_bn_relu_maxpool_run_cuda3.64 s      N/A         N/A         718.39 J    N/A         

[daisytuner]

Daisytuner Report - mlir_torch_layers (chamomile)

@@                                   Benchmarks                                   @@
=====================================================================================
  Benchmark              Time        ΔTime       Thr         Energy      ΔEnergy     
=====================================================================================
# batchnorm_torch        19.27 s     N/A         N/A         3614.31 J   N/A         
# batchnorm_run_none     7.67 s      N/A         N/A         1418.62 J   N/A         
# batchnorm_run_sequential8.48 s      N/A         N/A         1565.67 J   N/A         
# batchnorm_run_openmp   8.51 s      N/A         N/A         1571.35 J   N/A         
# batchnorm_run_cuda     12.71 s     N/A         N/A         2348.82 J   N/A         
# linear_torch           6.13 s      N/A         N/A         1433.52 J   N/A         
# linear_run_none        9.99 s      N/A         N/A         2651.34 J   N/A         
# linear_run_sequential  8.51 s      N/A         N/A         2378.13 J   N/A         
# linear_run_openmp      8.53 s      N/A         N/A         2381.48 J   N/A         
# linear_run_cuda        7.38 s      N/A         N/A         1371.67 J   N/A         
# matmul_torch           6.13 s      N/A         N/A         1434.18 J   N/A         
# matmul_run_none        10.04 s     N/A         N/A         2669.27 J   N/A         
# matmul_run_sequential  8.44 s      N/A         N/A         2374.23 J   N/A         
# matmul_run_openmp      8.48 s      N/A         N/A         2378.82 J   N/A         
# matmul_run_cuda        7.28 s      N/A         N/A         1360.09 J   N/A         
# pooling_torch          25.50 s     N/A         N/A         4811.07 J   N/A         
# pooling_run_none       17.26 s     N/A         N/A         3158.54 J   N/A         
# pooling_run_sequential 17.34 s     N/A         N/A         3174.27 J   N/A         
# pooling_run_openmp     17.30 s     N/A         N/A         3167.63 J   N/A         
# pooling_run_cuda       24.05 s     N/A         N/A         4430.54 J   N/A         
# relu_torch             18.98 s     N/A         N/A         3557.53 J   N/A         
# relu_run_none          4.40 s      N/A         N/A         822.70 J    N/A         
# relu_run_sequential    4.41 s      N/A         N/A         825.30 J    N/A         
# relu_run_openmp        4.41 s      N/A         N/A         825.36 J    N/A         
# relu_run_cuda          6.09 s      N/A         N/A         1134.64 J   N/A         

[ramonwirsch]

"basic" is not the most expressive name. Those checks should have thorough documentation so that there are no misunderstandings for what guarantees they provide

[ramonwirsch]

As static methods with no dependency on MatMulNode, would be more intuitive to find on Tensor sth.
Its applicable to any tensor no?