roofline estimator: simplify

[vkuzo]

Summary:

1. remove estimating torch.compile limitations (interesting but hasn't
   been useful)
2. make clearer distinction between roofline and benchmarked values
3. fix formatting of sympy-generated numbers in pandas by casting to float
4. general var names cleanup

Test Plan:

(pytorch) [[email protected] ~/local/ao (20250225_mx_roofline)]$ python benchmarks/float8/float8_roofline.py ~/local/tmp/20250225_test.csv
do_benchmarks: True
shape_gen_name: square
bf16_gemm_time_sympy 4.44444444444444e-15*K*M*N
fp8_gemm_time_sympy 2.22222222222222e-15*K*M*N

100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 8/8 [00:17<00:00,  2.14s/it]
   fwd_M  fwd_K  fwd_N  r_bf16_gemm_s  r_fp8_gemm_s  r_fp8_ovhd_s  r_fp8_gemm_and_ovhd_s  r_fp8_gemm_and_ovhd_spdp  b_bf16_gemm_s  b_fp8_gemm_s  b_bf16_e2e_s  b_fp8_e2e_s  b_fp8_e2e_spdp
0    256    256    256       7.46e-08      3.73e-08      6.12e-06               6.16e-06                      0.01       9.41e-06      9.72e-06      2.35e-05     4.73e-05            0.50
1    512    512    512       5.97e-07      2.98e-07      6.50e-06               6.80e-06                      0.09       1.10e-05      1.03e-05      2.63e-05     5.00e-05            0.53
2   1024   1024   1024       4.77e-06      2.39e-06      7.99e-06               1.04e-05                      0.46       1.71e-05      1.29e-05      3.41e-05     6.02e-05            0.57
3   2048   2048   2048       3.82e-05      1.91e-05      1.40e-05               3.31e-05                      1.15       4.61e-05      2.83e-05      8.72e-05     1.29e-04            0.67
4   4096   4096   4096       3.05e-04      1.53e-04      3.79e-05               1.91e-04                      1.60       3.31e-04      1.67e-04      4.08e-04     4.22e-04            0.97
5   8192   8192   8192       2.44e-03      1.22e-03      1.34e-04               1.36e-03                      1.80       2.69e-03      8.37e-04      2.71e-03     2.01e-03            1.35
6  16384  16384  16384       1.95e-02      9.77e-03      5.17e-04               1.03e-02                      1.90       2.19e-02      1.01e-02      2.75e-02     1.51e-02            1.82
7  32768  32768  32768       1.56e-01      7.82e-02      2.05e-03               8.02e-02                      1.95       2.67e-01      6.22e-02      2.51e-01     1.27e-01            1.99

Reviewers:

Subscribers:

Tasks:

Tags:

[vkuzo]

Stack from ghstack (oldest at bottom):

• roofline estimator: add float8 rowwise and mxfp8 recipe support #1789
• -> roofline estimator: simplify #1783
• roofline estimation: delete axiswise scaling, for now #1782
• roofline estimation: delete scaling type #1781
• [1/x] mx roofline: make the script work on NVIDIA B200 #1778

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1783

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit e9ab762 with merge base 1ab1b77 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.