AI/rocm-systems
2
0
Bifurcation 0
Code Tickets Demandes d'ajout Actions Paquets Projets Publications Wiki Activité

Enabling Standalone GUI on 2.x (#214)

Parcourir la source 
* Initial overhaul of Analyze mode. Basic CLI is enabled.

Signed-off-by: colramos-amd <colramos@amd.com>

* Merge branch '2.x' of github.com:AMDResearch/omniperf into 2.x-dev

Signed-off-by: colramos-amd <colramos@amd.com>

* fix comment typo

Signed-off-by: Karl W Schulz <karl.schulz@amd.com>

* Move error logging to util.py

Signed-off-by: colramos-amd <colramos@amd.com>

* Move perfmon_configs dir into omniperf_soc dir. Rename config dirs for clarity

Signed-off-by: colramos-amd <colramos@amd.com>

* Add a supported_archs property to Omniperf base class

Signed-off-by: colramos-amd <colramos@amd.com>

* Add css assets for GUI styling

Signed-off-by: colramos-amd <colramos@amd.com>

* Re-organize roofline class. Improved useability

Signed-off-by: colramos-amd <colramos@amd.com>

* Enable standalone GUI

Signed-off-by: colramos-amd <colramos@amd.com>

* Remove outdated metric_configs. This was moved to omniperf_soc dir

Signed-off-by: colramos-amd <colramos@amd.com>

* Fix small bug in GUI to enable Mi100 visualization

Signed-off-by: colramos-amd <colramos@amd.com>

---------

Signed-off-by: colramos-amd <colramos@amd.com>
Signed-off-by: Karl W Schulz <karl.schulz@amd.com>
Signed-off-by: Cole Ramos <colramos@amd.com>
Co-authored-by: Karl W Schulz <karl.schulz@amd.com>

[ROCm/rocprofiler-compute commit: 7d93a086c2]
Cette révision appartient à :
Cole Ramos
2023-12-18 16:37:01 -06:00
révisé par GitHub
Parent 4719fa0e3b
révision c73bd2bf91
165 fichiers modifiés avec 37024 ajouts et 14772 suppressions
Télécharger le Fichier Patch Télécharger le Fichier des Différences Développer tous les fichiers Réduire tous les fichiers
projects/rocprofiler-compute/src/utils/gui.py
+379
Voir le fichier
@@ -0,0 +1,379 @@
##############################################################################bl
# MIT License
#
# Copyright (c) 2021 - 2023 Advanced Micro Devices, Inc. All Rights Reserved.
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
##############################################################################el
import sys
import pandas as pd
from dash import html, dash_table
import plotly.express as px
import colorlover
from utils import schema
pd.set_option(
"mode.chained_assignment", None
) # ignore SettingWithCopyWarning pandas warning
IS_DARK = True #TODO: Remove hardcoded in favor of class property
##################
# HELPER FUNCTIONS
##################
def filter_df(column, df, filt):
filt_df = df
if filt != []:
filt_df = df.loc[df[schema.pmc_perf_file_prefix][column].astype(str).isin(filt)]
return filt_df
def multi_bar_chart(table_id, display_df):
if table_id == 1604:
nested_bar = {"NC": {}, "UC": {}, "RW": {}, "CC": {}}
for index, row in display_df.iterrows():
nested_bar[row["Coherency"]][row["Xfer"]] = row["Avg"]
if table_id == 1704:
nested_bar = {"Read": {}, "Write": {}}
for index, row in display_df.iterrows():
nested_bar[row["Transaction"]][row["Type"]] = row["Avg"]
return nested_bar
def discrete_background_color_bins(df, n_bins=5, columns="all"):
bounds = [i * (1.0 / n_bins) for i in range(n_bins + 1)]
if columns == "all":
if "id" in df:
df_numeric_columns = df.select_dtypes("number").drop(["id"], axis=1)
else:
df_numeric_columns = df.select_dtypes("number")
else:
df_numeric_columns = df[columns]
df_max = df_numeric_columns.max().max()
df_min = df_numeric_columns.min().min()
ranges = [((df_max - df_min) * i) + df_min for i in bounds]
styles = []
legend = []
for i in range(1, len(bounds)):
min_bound = ranges[i - 1]
max_bound = ranges[i]
backgroundColor = colorlover.scales[str(n_bins)]["seq"]["Blues"][i - 1]
color = "white" if i > len(bounds) / 2.0 else "inherit"
for column in df_numeric_columns:
styles.append(
{
"if": {
"filter_query": (
"{{{column}}} >= {min_bound}"
+ (
" && {{{column}}} < {max_bound}"
if (i < len(bounds) - 1)
else ""
)
).format(column=column, min_bound=min_bound, max_bound=max_bound),
"column_id": column,
},
"backgroundColor": backgroundColor,
"color": color,
}
)
legend.append(
html.Div(
style={"display": "inline-block", "width": "60px"},
children=[
html.Div(
style={
"backgroundColor": backgroundColor,
"borderLeft": "1px rgb(50, 50, 50) solid",
"height": "10px",
}
),
html.Small(round(min_bound, 2), style={"paddingLeft": "2px"}),
],
)
)
return (styles, html.Div(legend, style={"padding": "5px 0 5px 0"}))
####################
# GRAPHICAL ELEMENTS
####################
def build_bar_chart(display_df, table_config, barchart_elements, norm_filt):
"""
Read data into a bar chart. ID will determine which subtype of barchart.
"""
d_figs = []
# Insr Mix bar chart
if table_config["id"] in barchart_elements["instr_mix"]:
display_df["Avg"] = [
x.astype(int) if x != "" else int(0) for x in display_df["Avg"]
]
df_unit = display_df["Unit"][0]
d_figs.append(
px.bar(
display_df,
x="Avg",
y="Metric",
color="Avg",
labels={"Avg": "# of {}".format(df_unit.lower())},
height=400,
orientation="h",
)
)
# Multi bar chart
elif table_config["id"] in barchart_elements["multi_bar"]:
display_df["Avg"] = [
x.astype(int) if x != "" else int(0) for x in display_df["Avg"]
]
df_unit = display_df["Unit"][0]
nested_bar = multi_bar_chart(table_config["id"], display_df)
# generate chart for each coherency
for group, metric in nested_bar.items():
d_figs.append(
px.bar(
title=group,
x=metric.values(),
y=metric.keys(),
labels={"x": df_unit, "y": ""},
text=metric.values(),
orientation="h",
height=200,
)
.update_xaxes(showgrid=False, rangemode="nonnegative")
.update_yaxes(showgrid=False)
.update_layout(title_x=0.5)
)
# L2 Cache per channel
elif table_config["id"] in barchart_elements["l2_cache_per_chan"]:
nested_bar = {}
channels = []
for colName, colData in display_df.items():
if colName == "Channel":
channels = list(colData.values)
else:
display_df[colName] = [
x.astype(float) if x != "" and x != None else float(0)
for x in display_df[colName]
]
nested_bar[colName] = list(display_df[colName])
for group, metric in nested_bar.items():
d_figs.append(
px.bar(
title=group[0 : group.rfind("(")],
x=channels,
y=metric,
labels={
"x": "Channel",
"y": group[group.rfind("(") + 1 : len(group) - 1].replace(
"per", norm_filt
),
},
).update_yaxes(rangemode="nonnegative")
)
# Speed-of-light bar chart
elif table_config["id"] in barchart_elements["sol"]:
display_df["Value"] = [
x.astype(float) if x != "" else float(0) for x in display_df["Value"]
]
if table_config["id"] == 1701:
# special layout for L2 Cache SOL
d_figs.append(
px.bar(
display_df[display_df["Unit"] == "Pct"],
x="Value",
y="Metric",
color="Value",
range_color=[0, 100],
labels={"Value": "%"},
height=220,
orientation="h",
).update_xaxes(range=[0, 110], ticks="inside")
) # append first % chart
d_figs.append(
px.bar(
display_df[display_df["Unit"] == "Gb/s"],
x="Value",
y="Metric",
color="Value",
range_color=[0, 1638],
labels={"Value": "GB/s"},
height=220,
orientation="h",
).update_xaxes(range=[0, 1638])
) # append second GB/s chart
else:
d_figs.append(
px.bar(
display_df,
x="Value",
y="Metric",
color="Value",
range_color=[0, 100],
labels={"Value": "%"},
height=400,
orientation="h",
).update_xaxes(range=[0, 110])
)
else:
print(
"ERROR: Table id {}. Cannot determine barchart type.".format(
table_config["id"]
)
)
sys.exit(-1)
# update layout for each of the charts
for fig in d_figs:
fig.update_layout(
margin=dict(l=50, r=50, b=50, t=50, pad=4),
paper_bgcolor="rgba(0,0,0,0)",
plot_bgcolor="rgba(0,0,0,0)",
font={"color": "#ffffff"},
)
return d_figs
def build_table_chart(
display_df, table_config, original_df, display_columns, comparable_columns, decimal
):
"""
Read data into a DashTable
"""
d_figs = []
# build comlumns/header with formatting
formatted_columns = []
for col in display_df.columns:
if (
str(col).lower() == "pct"
or str(col).lower() == "pop"
or str(col).lower() == "percentage"
):
formatted_columns.append(
dict(
id=col,
name=col,
type="numeric",
format={"specifier": ".{}f".format(decimal)},
)
)
elif col in comparable_columns:
formatted_columns.append(
dict(
id=col,
name=col,
type="numeric",
format={"specifier": ".{}f".format(decimal)},
)
)
else:
formatted_columns.append(dict(id=col, name=col, type="text"))
# tooltip shows only on the 1st col for now if 'Tips' available
table_tooltip = (
[
{
column: {
"value": str(row["Tips"])
if column == display_columns[0] and row["Tips"]
else "",
"type": "markdown",
}
for column, value in row.items()
}
for row in original_df.to_dict("records")
]
if "Tips" in original_df.columns.values.tolist()
else None
)
# build data table with columns, tooltip, df and other properties
d_t = dash_table.DataTable(
id=str(table_config["id"]),
sort_action="native",
sort_mode="multi",
columns=formatted_columns,
tooltip_data=table_tooltip,
# left-aligning the text of the 1st col
style_cell_conditional=[
{"if": {"column_id": display_columns[0]}, "textAlign": "left"}
],
# style cell
style_cell={"maxWidth": "500px"},
# display style
style_header={
"backgroundColor": "rgb(30, 30, 30)",
"color": "white",
"fontWeight": "bold",
}
if IS_DARK
else {},
style_data={
"backgroundColor": "rgb(50, 50, 50)",
"color": "white",
"whiteSpace": "normal",
"height": "auto",
}
if IS_DARK
else {},
style_data_conditional=[
{"if": {"row_index": "odd"}, "backgroundColor": "rgb(60, 60, 60)"},
{
"if": {"column_id": "PoP", "filter_query": "{PoP} > 50"},
"backgroundColor": "#ffa90a",
"color": "white",
},
{
"if": {"column_id": "PoP", "filter_query": "{PoP} > 80"},
"backgroundColor": "#ff120a",
"color": "white",
},
{
"if": {
"column_id": "Avg",
"filter_query": "{Unit} = Pct && {Avg} > 50",
},
"backgroundColor": "#ffa90a",
"color": "white",
},
{
"if": {
"column_id": "Avg",
"filter_query": "{Unit} = Pct && {Avg} > 80",
},
"backgroundColor": "#ff120a",
"color": "white",
},
]
if IS_DARK
else [],
# the df to display
data=display_df.to_dict("records"),
)
# print("DATA: \n", display_df.to_dict('records'))
d_figs.append(d_t)
return d_figs
# print(d_t.columns)
projects/rocprofiler-compute/src/utils/gui_components/header.py
+341
Voir le fichier
@@ -0,0 +1,341 @@
##############################################################################bl
# MIT License
#
# Copyright (c) 2021 - 2023 Advanced Micro Devices, Inc. All Rights Reserved.
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
##############################################################################el
from dash import html, dcc
import dash_bootstrap_components as dbc
from utils import schema
avail_normalizations = ["per_wave", "per_cycle", "per_second", "per_kernel"]
# List all the unique column values for desired column in df, 'target_col'
def list_unique(orig_list, is_numeric):
list_set = set(orig_list)
unique_list = list(list_set)
if is_numeric:
unique_list.sort()
return unique_list
def create_span(input):
elmt = {}
elmt["label"] = (html.Span(str(input), title=str(input)),)
elmt["value"] = str(input)
return elmt
def get_header(raw_pmc, input_filters, kernel_names):
return html.Header(
id="home",
children=[
html.Nav(
id="nav-wrap",
children=[
html.Ul(
id="nav",
children=[
html.Div(
className="nav-left",
children=[
dbc.DropdownMenu(
[
dbc.DropdownMenuItem("Overview", header=True),
dbc.DropdownMenuItem(
"Roofline",
href="#roofline",
external_link=True,
),
dbc.DropdownMenuItem(
"Top Stat",
href="#top_stat",
external_link=True,
),
dbc.DropdownMenuItem(
"System Info",
href="#system_info",
external_link=True,
),
dbc.DropdownMenuItem(
"System Speed-of-Light",
href="#system_speed-of-light",
external_link=True,
),
dbc.DropdownMenuItem("Compute", header=True),
dbc.DropdownMenuItem(
"Command Processor (CPF/CPC)",
href="#command_processor_cpccpf",
external_link=True,
),
dbc.DropdownMenuItem(
"Shader Processor Input (SPI)",
href="#shader_processor_input_spi",
external_link=True,
),
dbc.DropdownMenuItem(
"Wavefront",
href="#wavefront",
external_link=True,
),
dbc.DropdownMenuItem(
"Compute Units - Instruction Mix",
href="#compute_units_-_instruction_mix",
external_link=True,
),
dbc.DropdownMenuItem(
"Compute Units - Compute Pipeline",
href="#compute_units_-_compute_pipeline",
external_link=True,
),
dbc.DropdownMenuItem("Cache", header=True),
dbc.DropdownMenuItem(
"Local Data Share (LDS)",
href="#local_data_share_lds",
external_link=True,
),
dbc.DropdownMenuItem(
"Instruction Cache",
href="#instruction_cache",
external_link=True,
),
dbc.DropdownMenuItem(
"Scalar L1 Data Cache",
href="#scalar_l1_data_cache",
external_link=True,
),
dbc.DropdownMenuItem(
"Texture Addresser & Texture Data (TA/TD)",
href="#texture_addresser_and_texture_data_tatd",
external_link=True,
),
dbc.DropdownMenuItem(
"Vector L1 Data Cache",
href="#vector_l1_data_cache",
external_link=True,
),
dbc.DropdownMenuItem(
"L2 Cache",
href="#l2_cache",
external_link=True,
),
dbc.DropdownMenuItem(
"L2 Cache (per channel)",
href="#l2_cache_per_channel",
external_link=True,
),
],
label="Menu",
menu_variant="dark",
),
],
),
html.Li(
className="filter",
children=[
html.Div(
children=[
html.A(
className="smoothscroll",
children=["Normalization:"],
),
dcc.Dropdown(
avail_normalizations,
id="norm-filt",
value=input_filters["normalization"],
clearable=False,
style={"width": "150px"},
),
]
)
],
),
html.Li(
className="filter",
children=[
html.Div(
children=[
html.A(
className="smoothscroll",
children=["GCD:"],
),
dcc.Dropdown(
list_unique(
list(
map(
str,
raw_pmc[
schema.pmc_perf_file_prefix
]["gpu-id"],
)
),
True,
), # list avail gcd ids
id="gcd-filt",
multi=True,
value=input_filters[
"gpu"
], # default to any gpu filters passed as args
placeholder="ALL",
clearable=False,
style={"width": "60px"},
),
]
)
],
),
html.Li(
className="filter",
children=[
html.Div(
children=[
html.A(
className="smoothscroll",
children=["Dispatch Filter:"],
),
dcc.Dropdown(
list(
map(
str,
raw_pmc[
schema.pmc_perf_file_prefix
]["Index"],
)
),
id="disp-filt",
multi=True,
value=input_filters[
"dispatch"
], # default to any dispatch filters passed as args
placeholder="ALL",
style={"width": "150px"},
),
]
)
],
),
html.Li(
className="filter",
children=[
html.Div(
children=[
html.A(
className="smoothscroll",
children=["Top N:"],
),
dcc.Dropdown(
[1, 5, 10, 15, 20, 50, 100],
id="top-n-filt",
value=input_filters[
"top_n"
], # default to any dispatch filters passed as args
clearable=False,
style={"width": "50px"},
),
]
)
],
),
html.Li(
className="filter",
children=[
html.Div(
children=[
html.A(
className="smoothscroll",
children=["Kernels:"],
),
dcc.Dropdown(
list(
map(
create_span,
list_unique(
list(
map(
str,
raw_pmc[
schema.pmc_perf_file_prefix
]["KernelName"],
)
),
False,
), # list avail kernel names
)
),
id="kernel-filt",
multi=True,
value=kernel_names,
optionHeight=150,
placeholder="ALL",
style={
"width": "600px", # TODO: Change these widths to % rather than fixed value
},
),
]
)
],
),
html.Div(
className="nav-right",
children=[
html.Li(
children=[
# Report bug button
html.A(
href="",
children=[
html.Button(
className="report",
children=["Report Bug"],
)
],
)
]
)
],
),
],
)
],
),
html.Div(
className="row banner",
children=[
html.H3(
children=["Placeholder. Guided Analysis coming soon..."],
style={"color": "white"},
),
],
),
html.P(
className="scrolldown",
children=[
html.A(
className="smoothscroll",
href="#roofline",
children=[html.I(className="icon-down-circle")],
)
],
),
],
)
projects/rocprofiler-compute/src/utils/gui_components/memchart.py
+1944
Voir le fichier
Fichier diff supprimé car celui-ci est trop grand Voir la Diff
projects/rocprofiler-compute/src/utils/roofline_calc.py
+37 -49
Voir le fichier
@@ -22,9 +22,10 @@
# SOFTWARE.
##############################################################################el
import sys
import os
from dataclasses import dataclass
import logging
import csv
################################################
@@ -94,43 +95,42 @@ def get_color(catagory):
# -------------------------------------------------------------------------------------
# Plot BW at each cache level
# -------------------------------------------------------------------------------------
def calc_ceilings(roof_specs, benchmark_data, targ_mem_level, verbose):
def calc_ceilings(roofline_parameters, dtype, benchmark_data):
"""Given benchmarking data, calculate ceilings (or peak performance) for empirical roofline
"""
# TODO: This is where filtering by memory level will need to occur for standalone
graphPoints = {"hbm": [], "l2": [], "l1": [], "lds": [], "valu": [], "mfma": []}
if targ_mem_level == "ALL":
if roofline_parameters['mem_level'] == "ALL":
cacheHierarchy = ["HBM", "L2", "L1", "LDS"]
else:
cacheHierarchy = targ_mem_level
cacheHierarchy = roofline_parameters['mem_level']
x1 = y1 = x2 = y2 = -1
x1_mfma = y1_mfma = x2_mfma = y2_mfma = -1
target_precision = roof_specs["dtype"][2:]
target_precision = dtype[2:]
if roof_specs["dtype"] != "FP16" and roof_specs["dtype"] != "I8":
if dtype != "FP16" and dtype != "I8":
peakOps = float(
benchmark_data[roof_specs["dtype"] + "Flops"][roof_specs["device"]]
benchmark_data[dtype + "Flops"][roofline_parameters['device_id']]
)
for i in range(0, len(cacheHierarchy)):
# Plot BW line
if verbose >= 3:
print("Current cache level is ", cacheHierarchy[i])
# Plot BW line
logging.debug("[roofline] Current cache level is ", cacheHierarchy[i])
curr_bw = cacheHierarchy[i] + "Bw"
peakBw = float(benchmark_data[curr_bw][roof_specs["device"]])
peakBw = float(benchmark_data[curr_bw][roofline_parameters['device_id']])
if roof_specs["dtype"] == "I8":
peakMFMA = float(benchmark_data["MFMAI8Ops"][roof_specs["device"]])
if dtype == "I8":
peakMFMA = float(benchmark_data["MFMAI8Ops"][roofline_parameters['device_id']])
else:
peakMFMA = float(
benchmark_data["MFMAF{}Flops".format(target_precision)][roof_specs["device"]]
benchmark_data["MFMAF{}Flops".format(target_precision)][roofline_parameters['device_id']]
)
x1 = float(XMIN)
y1 = float(XMIN) * peakBw
# Note: No reg peakOps for FP16 or INT8
if roof_specs["dtype"] != "FP16" and roof_specs["dtype"] != "I8":
if dtype != "FP16" and dtype != "I8":
x2 = peakOps / peakBw
y2 = peakOps
@@ -142,9 +142,9 @@ def calc_ceilings(roof_specs, benchmark_data, targ_mem_level, verbose):
y2_mfma = peakMFMA
# These are the points to use:
if verbose >= 3:
print("x = [{}, {}]".format(x1, x2_mfma))
print("y = [{}, {}]".format(y1, y2_mfma))
logging.debug("[roofline] coordinate points:")
logging.debug("x = [{}, {}]".format(x1, x2_mfma))
logging.debug("y = [{}, {}]".format(y1, y2_mfma))
graphPoints[cacheHierarchy[i].lower()].append([x1, x2_mfma])
graphPoints[cacheHierarchy[i].lower()].append([y1, y2_mfma])
@@ -154,28 +154,26 @@ def calc_ceilings(roof_specs, benchmark_data, targ_mem_level, verbose):
# Plot computing roof
# -------------------------------------------------------------------------------------
# Note: No FMA roof for FP16 or INT8
if roof_specs["dtype"] != "FP16" and roof_specs["dtype"] != "I8":
if dtype != "FP16" and dtype != "I8":
# Plot FMA roof
x0 = XMAX
if x2 < x0:
x0 = x2
if verbose >= 3:
print("FMA ROOF [{}, {}], [{},{}]".format(x0, XMAX, peakOps, peakOps))
logging.debug("FMA ROOF [{}, {}], [{},{}]".format(x0, XMAX, peakOps, peakOps))
graphPoints["valu"].append([x0, XMAX])
graphPoints["valu"].append([peakOps, peakOps])
graphPoints["valu"].append(peakOps)
# Plot MFMA roof
if (
x1_mfma != -1 or roof_specs["dtype"] == "FP16" or roof_specs["dtype"] == "I8"
x1_mfma != -1 or dtype == "FP16" or dtype == "I8"
): # assert that mfma has been assigned
x0_mfma = XMAX
if x2_mfma < x0_mfma:
x0_mfma = x2_mfma
if verbose >= 3:
print("MFMA ROOF [{}, {}], [{},{}]".format(x0_mfma, XMAX, peakMFMA, peakMFMA))
logging.debug("MFMA ROOF [{}, {}], [{},{}]".format(x0_mfma, XMAX, peakMFMA, peakMFMA))
graphPoints["mfma"].append([x0_mfma, XMAX])
graphPoints["mfma"].append([peakMFMA, peakMFMA])
graphPoints["mfma"].append(peakMFMA)
@@ -187,7 +185,7 @@ def calc_ceilings(roof_specs, benchmark_data, targ_mem_level, verbose):
# Overlay application performance
# -------------------------------------------------------------------------------------
# Calculate relevent metrics for ai calculation
def calc_ai(sort_type, ret_df, verbose):
def calc_ai(sort_type, ret_df):
"""Given counter data, caclulate arithmetic intensity for each kernel in the application.
"""
df = ret_df["pmc_perf"]
@@ -263,8 +261,7 @@ def calc_ai(sort_type, ret_df, verbose):
+ (df["SQ_INSTS_VALU_MFMA_MOPS_F64"][idx] * 512)
)
except KeyError:
if verbose >= 3:
print("{}: Skipped total_flops at index {}".format(kernelName[:35], idx))
logging.debug("[roofline] {}: Skipped total_flops at index {}".format(kernelName[:35], idx))
pass
try:
valu_flops += (
@@ -291,8 +288,7 @@ def calc_ai(sort_type, ret_df, verbose):
)
)
except KeyError:
if verbose >= 3:
print("{}: Skipped valu_flops at index {}".format(kernelName[:35], idx))
logging.debug("{}: Skipped valu_flops at index {}".format(kernelName[:35], idx))
pass
try:
@@ -302,8 +298,7 @@ def calc_ai(sort_type, ret_df, verbose):
mfma_flops_f64 += df["SQ_INSTS_VALU_MFMA_MOPS_F64"][idx] * 512
mfma_iops_i8 += df["SQ_INSTS_VALU_MFMA_MOPS_I8"][idx] * 512
except KeyError:
if verbose >= 3:
print("{}: Skipped mfma ops at index {}".format(kernelName[:35], idx))
logging.debug("[roofline] {}: Skipped mfma ops at index {}".format(kernelName[:35], idx))
pass
try:
@@ -313,15 +308,13 @@ def calc_ai(sort_type, ret_df, verbose):
* L2_BANKS
) # L2_BANKS = 32 (since assuming mi200)
except KeyError:
if verbose >= 3:
print("{}: Skipped lds_data at index {}".format(kernelName[:35], idx))
logging.debug("[roofline] {}: Skipped lds_data at index {}".format(kernelName[:35], idx))
pass
try:
L1cache_data += df["TCP_TOTAL_CACHE_ACCESSES_sum"][idx] * 64
except KeyError:
if verbose >= 3:
print("{}: Skipped L1cache_data at index {}".format(kernelName[:35], idx))
logging.debug("[roofline] {}: Skipped L1cache_data at index {}".format(kernelName[:35], idx))
pass
try:
@@ -332,8 +325,7 @@ def calc_ai(sort_type, ret_df, verbose):
+ df["TCP_TCC_READ_REQ_sum"][idx] * 64
)
except KeyError:
if verbose >= 3:
print("{}: Skipped L2cache_data at index {}".format(kernelName[:35], idx))
logging.debug("[roofline] {}: Skipped L2cache_data at index {}".format(kernelName[:35], idx))
pass
try:
hbm_data += (
@@ -343,8 +335,7 @@ def calc_ai(sort_type, ret_df, verbose):
+ ((df["TCC_EA_WRREQ_sum"][idx] - df["TCC_EA_WRREQ_64B_sum"][idx]) * 32)
)
except KeyError:
if verbose >= 3:
print("{}: Skipped hbm_data at index {}".format(kernelName[:35], idx))
logging.debug("[roofline] {}: Skipped hbm_data at index {}".format(kernelName[:35], idx))
pass
totalDuration += df["EndNs"][idx] - df["BeginNs"][idx]
@@ -372,12 +363,11 @@ def calc_ai(sort_type, ret_df, verbose):
avgDuration / calls,
)
)
if verbose >= 2:
print(
"Just added {} to AI_Data at index {}. # of calls: {}".format(
kernelName, idx, calls
)
logging.debug(
"Just added {} to AI_Data at index {}. # of calls: {}".format(
kernelName, idx, calls
)
)
total_flops = (
valu_flops
) = (
@@ -487,8 +477,8 @@ def calc_ai(sort_type, ret_df, verbose):
return intensityPoints
def constuct_roof(roof_specs, targ_mem_level, verbose):
benchmark_results = roof_specs["path"] + "/roofline.csv"
def constuct_roof(roofline_parameters, dtype):
benchmark_results = os.path.join(roofline_parameters["path_to_dir"], "roofline.csv")
# -----------------------------------------------------
# Initialize roofline data dictionary from roofline.csv
# -----------------------------------------------------
@@ -526,8 +516,6 @@ def constuct_roof(roof_specs, targ_mem_level, verbose):
# ------------------
# Generate Roofline
# ------------------
results = calc_ceilings(roof_specs, benchmark_data, targ_mem_level, verbose)
# for key in results:
# print(key, "->", results[key])
results = calc_ceilings(roofline_parameters, dtype, benchmark_data)
return results
projects/rocprofiler-compute/src/utils/tty.py
-1
Voir le fichier
@@ -135,7 +135,6 @@ def show_all(args, runs, archConfigs, output):
)
# show value + percentage
# TODO: better alignment
t_df = (
cur_df[header]
.astype(float)