AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix test cases (#837)

Browse Source 
* Fix formatting

[ROCm/rocprofiler-compute commit: 80ea339217]
This commit is contained in:
vedithal-amd
2025-07-28 16:19:45 -04:00
committed by GitHub
parent 17e5892614
commit 2966867099
7 changed files with 90 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler-compute/src/utils/roofline_calc.py
+56 -14
View File
@@ -259,11 +259,25 @@ def calc_ai(mspec, sort_type, ret_df):
df = df.sort_values(by=["Kernel_Name"])
df = df.reset_index(drop=True)
total_flops = valu_flops = mfma_flops_f6f4 = mfma_flops_f8 = mfma_flops_bf16 = (
total_flops = (
valu_flops
) = (
mfma_flops_f6f4
) = (
mfma_flops_f8
) = (
mfma_flops_bf16
) = (
mfma_flops_f16
) = mfma_iops_i8 = mfma_flops_f32 = mfma_flops_f64 = lds_data = L1cache_data = (
L2cache_data
) = hbm_data = calls = totalDuration = avgDuration = 0.0
) = (
mfma_iops_i8
) = (
mfma_flops_f32
) = (
mfma_flops_f64
) = (
lds_data
) = L1cache_data = L2cache_data = hbm_data = calls = totalDuration = avgDuration = 0.0
kernelName = ""
@@ -484,13 +498,27 @@ def calc_ai(mspec, sort_type, ret_df):
kernelName, idx, calls
)
)
total_flops = valu_flops = mfma_flops_f6f4 = mfma_flops_f8 = (
total_flops = (
valu_flops
) = (
mfma_flops_f6f4
) = (
mfma_flops_f8
) = (
mfma_flops_bf16
) = mfma_flops_f16 = mfma_iops_i8 = mfma_flops_f32 = mfma_flops_f64 = (
) = (
mfma_flops_f16
) = (
mfma_iops_i8
) = (
mfma_flops_f32
) = (
mfma_flops_f64
) = (
lds_data
) = L1cache_data = L2cache_data = hbm_data = calls = totalDuration = (
avgDuration
) = 0.0
) = (
L1cache_data
) = L2cache_data = hbm_data = calls = totalDuration = avgDuration = 0.0
if sort_type == "dispatches":
myList.append(
@@ -514,13 +542,27 @@ def calc_ai(mspec, sort_type, ret_df):
avgDuration,
)
)
total_flops = valu_flops = mfma_flops_f6f4 = mfma_flops_f8 = (
total_flops = (
valu_flops
) = (
mfma_flops_f6f4
) = (
mfma_flops_f8
) = (
mfma_flops_bf16
) = mfma_flops_f16 = mfma_iops_i8 = mfma_flops_f32 = mfma_flops_f64 = (
) = (
mfma_flops_f16
) = (
mfma_iops_i8
) = (
mfma_flops_f32
) = (
mfma_flops_f64
) = (
lds_data
) = L1cache_data = L2cache_data = hbm_data = calls = totalDuration = (
avgDuration
) = 0.0
) = (
L1cache_data
) = L2cache_data = hbm_data = calls = totalDuration = avgDuration = 0.0
myList.sort(key=lambda x: x.totalDuration, reverse=True)