Merge pull request #15 from AMDResearch/8-progress-indicator

Improve GUI Useage

[ROCm/rocprofiler-compute commit: ea67c3fe50]
This commit is contained in:
Cole Ramos
2022-11-09 12:58:08 -06:00
zatwierdzone przez GitHub
7 zmienionych plików z 63 dodań i 39 usunięć
@@ -175,9 +175,7 @@ def omniperf_cli(args):
args.path[0][0]
) # create mega df
is_gui = False
# parser.load_table_data(
# runs[args.path[0][0]], args.path[0][0], is_gui, args.g
# ) # create the loaded table
parser.load_kernel_top(runs[args.path[0][0]], args.path[0][0])
input_filters = {
"kernel": runs[args.path[0][0]].filter_kernel_ids,
@@ -29,6 +29,7 @@ from dash.dash_table import FormatTemplate
from dash.dash_table.Format import Format, Scheme, Symbol
from dash import html, dash_table
from dash.dependencies import Input, Output, State
import dash_bootstrap_components as dbc
from dash import dcc
import plotly.express as px
@@ -307,9 +308,7 @@ def build_layout(
Build gui layout
"""
comparable_columns = parser.build_comparable_columns(time_unit)
base_run, base_data = next(iter(runs.items()))
app.layout = html.Div(style={"backgroundColor": "rgb(50, 50, 50)" if IS_DARK else ""})
filt_kernel_names = []
@@ -319,8 +318,17 @@ def build_layout(
app.layout.children = html.Div(
children=[
get_header(runs[path_to_dir].raw_pmc, input_filters, filt_kernel_names),
html.Div(id="container", children=[]),
dbc.Spinner(
children=[
get_header(
runs[path_to_dir].raw_pmc, input_filters, filt_kernel_names
),
html.Div(id="container", children=[]),
],
fullscreen=True,
color="primary",
spinner_style={"width": "6rem", "height": "6rem"},
)
]
)
@@ -333,12 +341,13 @@ def build_layout(
)
def generate_from_filter(disp_filt, kernel_filter, gcd_filter, div_children):
runs[path_to_dir].dfs = copy.deepcopy(archConfigs.dfs) # reset the equations
panel_configs = copy.deepcopy(archConfigs.panel_configs)
# Generate original raw df
runs[path_to_dir].raw_pmc = file_io.create_df_pmc(path_to_dir)
if verbose:
if verbose >= 1:
print("disp-filter is ", disp_filt)
print("kernel-filter is ", kernel_filter)
print("gpu-filter is ", gcd_filter)
print("gpu-filter is ", gcd_filter, "\n")
runs[path_to_dir].filter_kernel_ids = kernel_filter
runs[path_to_dir].filter_gpu_ids = gcd_filter
runs[path_to_dir].filter_dispatch_ids = disp_filt
@@ -351,11 +360,27 @@ def build_layout(
time_unit,
num_results,
)
# Evaluate metrics and table data from the raw df
is_gui = True
# Only display basic metrics if no filters are applied
if not (disp_filt or kernel_filter or gcd_filter):
temp = {}
keep = [1, 201, 101, 1901]
for key in runs[path_to_dir].dfs:
if keep.count(key) != 0:
temp[key] = runs[path_to_dir].dfs[key]
runs[path_to_dir].dfs = temp
temp = {}
keep = [0, 100, 200, 1900]
for key in panel_configs:
if keep.count(key) != 0:
temp[key] = panel_configs[key]
panel_configs = temp
parser.load_table_data(
runs[path_to_dir], path_to_dir, True, debug
) # Note: All the filtering happens in this function
div_children = []
div_children.append(
get_memchart(archConfigs.panel_configs[1900]["data source"], base_data)
@@ -369,7 +394,7 @@ def build_layout(
)
)
# Iterate over each section as defined in panel configs
for panel_id, panel in archConfigs.panel_configs.items():
for panel_id, panel in panel_configs.items():
title = str(panel_id // 100) + ". " + panel["title"]
section_title = (
panel["title"]
@@ -183,7 +183,7 @@ def get_roofline(path_to_dir, ret_df, verbose):
# Generate roofline plots
print("Path: ", path_to_dir)
ai_data = roofline_calc.plot_application("kernels", ret_df, verbose)
if verbose:
if verbose >= 1:
# print AI data for each mem level
for i in ai_data:
print(i, "->", ai_data[i])
@@ -650,7 +650,6 @@ def apply_filters(workload, is_gui, debug):
# NB: support ignoring the 1st n dispatched execution by '> n'
# The better way may be parsing python slice string
for d in workload.filter_dispatch_ids:
print("len of ret_df is ", len(ret_df))
if int(d) > len(ret_df) - 2: # subtract 2 bc of the two header rows
print("{} is an invalid dispatch id.".format(d))
sys.exit(1)
@@ -674,12 +673,7 @@ def apply_filters(workload, is_gui, debug):
return ret_df
def load_table_data(workload, dir, is_gui, debug):
"""
Load data for all "raw_csv_table".
Calculate mertric value for all "metric_table".
"""
def load_kernel_top(workload, dir):
# NB:
# - Do pmc_kernel_top.csv loading before eval_metric because we need the kernel names.
# - There might be a better way/timing to load raw_csv_table.
@@ -698,9 +692,16 @@ def load_table_data(workload, dir, is_gui, debug):
# All transposed columns should be marked with a general header,
# so tty could detect them and show them correctly in comparison.
tmp[id].columns = ["Info"]
workload.dfs.update(tmp)
def load_table_data(workload, dir, is_gui, debug):
"""
Load data for all "raw_csv_table".
Calculate mertric value for all "metric_table".
"""
load_kernel_top(workload, dir)
eval_metric(
workload.dfs,
workload.dfs_type,
@@ -239,7 +239,7 @@ def plot_application(sortType, ret_df, verbose):
avgDuration / calls,
)
)
if verbose:
if verbose >= 2:
print(
"Just added {} to AI_Data at index {}. # of calls: {}".format(
kernelName, index, calls
@@ -299,7 +299,7 @@ def plot_application(sortType, ret_df, verbose):
+ (row["SQ_INSTS_VALU_MFMA_MOPS_F64"] * 512)
)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped total_flops at index {}".format(index))
pass
try:
@@ -327,7 +327,7 @@ def plot_application(sortType, ret_df, verbose):
)
)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped valu_flops at index {}".format(index))
pass
@@ -338,7 +338,7 @@ def plot_application(sortType, ret_df, verbose):
mfma_flops_f64 += row["SQ_INSTS_VALU_MFMA_MOPS_F64"] * 512
mfma_iops_i8 += row["SQ_INSTS_VALU_MFMA_MOPS_I8"] * 512
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped mfma ops at index {}".format(index))
pass
@@ -347,14 +347,14 @@ def plot_application(sortType, ret_df, verbose):
(row["SQ_LDS_IDX_ACTIVE"] - row["SQ_LDS_BANK_CONFLICT"]) * 4 * L2_BANKS
) # L2_BANKS = 32 (since assuming mi200)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped lds_data at index {}".format(index))
pass
try:
L1cache_data += row["TCP_TOTAL_CACHE_ACCESSES_sum"] * 64
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped L1cache_data at index {}".format(index))
pass
@@ -366,7 +366,7 @@ def plot_application(sortType, ret_df, verbose):
+ row["TCP_TCC_READ_REQ_sum"] * 64
)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped L2cache_data at index {}".format(index))
pass
try:
@@ -377,7 +377,7 @@ def plot_application(sortType, ret_df, verbose):
+ ((row["TCC_EA_WRREQ_sum"] - row["TCC_EA_WRREQ_64B_sum"]) * 32)
)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped hbm_data at index {}".format(index))
pass
@@ -80,7 +80,7 @@ def parse(my_parser):
"-v", "--version", action="version", version="%(PROG)s (" + VER + ")"
)
general_group.add_argument(
"-V", "--verbose", help="Increase output verbosity", action="store_true"
"-V", "--verbose", help="Increase output verbosity", action="count", default=0
)
profile_group.add_argument(
@@ -231,7 +231,7 @@ def parse(my_parser):
"-v", "--version", action="version", version="%(PROG)s (" + VER + ")"
)
general_group.add_argument(
"-V", "--verbose", help="Increase output verbosity", action="store_true"
"-V", "--verbose", help="Increase output verbosity", action="count", default=0
)
interaction_group.add_argument(
@@ -331,7 +331,7 @@ def parse(my_parser):
"-v", "--version", action="version", version="%(PROG)s (" + VER + ")"
)
general_group.add_argument(
"-V", "--verbose", help="Increase output verbosity", action="store_true"
"-V", "--verbose", help="Increase output verbosity", action="count", default=0
)
analyze_group.add_argument(
@@ -294,7 +294,7 @@ def plot_application(inputs, verbose):
avgDuration / calls,
)
)
if verbose:
if verbose >= 2:
print(
"Just added {} to AI_Data at index {}. # of calls: {}".format(
kernelName, index, calls
@@ -354,7 +354,7 @@ def plot_application(inputs, verbose):
+ (row["SQ_INSTS_VALU_MFMA_MOPS_F64"] * 512)
)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped total_flops at index {}".format(index))
pass
try:
@@ -382,7 +382,7 @@ def plot_application(inputs, verbose):
)
)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped valu_flops at index {}".format(index))
pass
@@ -393,7 +393,7 @@ def plot_application(inputs, verbose):
mfma_flops_f64 += row["SQ_INSTS_VALU_MFMA_MOPS_F64"] * 512
mfma_iops_i8 += row["SQ_INSTS_VALU_MFMA_MOPS_I8"] * 512
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped mfma ops at index {}".format(index))
pass
@@ -402,14 +402,14 @@ def plot_application(inputs, verbose):
(row["SQ_LDS_IDX_ACTIVE"] - row["SQ_LDS_BANK_CONFLICT"]) * 4 * L2_BANKS
) # L2_BANKS = 32 (since assuming mi200)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped lds_data at index {}".format(index))
pass
try:
L1cache_data += row["TCP_TOTAL_CACHE_ACCESSES_sum"] * 64
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped L1cache_data at index {}".format(index))
pass
@@ -421,7 +421,7 @@ def plot_application(inputs, verbose):
+ row["TCP_TCC_READ_REQ_sum"] * 64
)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped L2cache_data at index {}".format(index))
pass
try:
@@ -432,7 +432,7 @@ def plot_application(inputs, verbose):
+ ((row["TCC_EA_WRREQ_sum"] - row["TCC_EA_WRREQ_64B_sum"]) * 32)
)
except KeyError:
if verbose:
if verbose >= 2:
print("Skipped hbm_data at index {}".format(index))
pass