Key changes in version cudaprof v3.0 with respect to v2.3:
1) Following new counters are supported:
a) tex_cache_hit, tex_cache_miss : Number of texture cache sector hits, misses
b) "NOP Triggers" : Can be used to get execution counts for certain code paths in the kernel
2) New memory copy option "host mem transfer type" is added for memory transfers.
This specifies whether a memory transfers uses "Pageable" or "Page-locked"
3) Device level summary plot :
One bar for each method is there. Bars are sorted in decreasing GPU time. Bar length
is proportional to cumulative GPU time for a method across all contexts for a device.
4) Session level summary plot :
One bar for each device is there. Bar length is proportional to GPU Utilization.
GPU Utilization is the proportion of time when the GPU was actually executing some method
to total time interval from GPU start to end. The values are presented in percentage.
5) User interface changes:
"Session Settings" Dialog :
a) Added a new device selection option on "Session" tab.
Based on this option the available counters can be selected on "Profiler Counter" tab.
In case of "multi-device" only counters supported by all devices can be selected.
b) All the counters on "Profiler Counter" tab and options on "Other Options" tab are shown
in tree view under different groups.
6) Support for the global memory load & store request counters (gld_request and gst_request) in dropped.
Global memory load and store efficiency which uses these counters is also dropped.
7) Added support for devices of compute capability 2.0.
8) cudaprof MacOS version is updated from v2.0 to v3.0
Key changes in version cudaprof v2.3 with respect to v2.2:
1) The summary table is enhanced to display global memory load efficiency
and global memory store efficiency for each kernel.
These use the following new profiler counters:
gld_request : Number of global memory load requests
gst_request : Number of global memory store requests
This is available only for devices with compute capability 1.2 or higher.
2) The profiler now uses GPU timestamps. This provides more accurate timing information
for GPU events on a timeline.
3) The profiler now supports multiple contexts for each device.
4) The profiler now reports all memory transfers. The method name for memory transfers
is changed from memcopy to memcpy<type> where <type> indicates the direction of transfer
and whether asynchronous memory transfer is used.
5) A new option "Normalize counters" is added in the "Session settings" dialog on
the "Profiler counters" tab. When this option is enabled all counter values are
normalized and per block counts are given.
6) User interface changes:
a) There are changes to the main menu. The "Profile" menu options are moved to
the "Session" menu and "Profiler" menu is dropped. The "Session->View" sub-menu options
are moved a new top level main menu "View".
b) The session list is now a tree with three levels. Sessions at the top level, devices under a session
at the next level and contexts under a device at the lowest level.
Summary session information is displayed when a session is selected in the tree view.
Summary device information is displayed when a device is selected in the tree view.
c) Added shortcut keys Ctrl+MouseScrollWheel for zooming in and out for GPU Time Height plot and GPU Time Width plot
d) A new context menu is added for the profiler table and summary table column header.
It provides options to hide and show columns.
e) The rows and columns in the profiler and summary table are now resized based on the contents.
f) Rows in the profiler output table can now be sorted based on any column by clicking on the column header.
g) A "Apply" button is added to the "Session view settings..." dialog. Also this now a non-modal dialog.
Options for different views can now be changed.
h) The session tree and output windows are now docable and the user can now drag and drop these window
anywhere in the Main Window. New menu options "View->Session Window" and "View->Output Window"
are added to control the visibility of the session and output windows.
i) Close buttons are added for tabs for each view (except for the profiler output table).
Key changes in version cudaprof v2.2 with respect to v1.1:
1) Profiler counters are now also supported on the Windows Vista platform.
2) Support has been added to handle profiler data for multiple CUDA devices.
The session list now uses a tree display and devices are listed as children under each session.
3) Following new profiler counters are supported:
a) gld_32b, gld_64b, gld_128b : Count 32-byte, 64-byte, 128-byte global memory loads
gst_32b, gst_64b, gst_128b : Count 32-byte, 64-byte, 128-byte global memory stores
These are available only for GPUs with compute capability 1.2 or higher.
b) gld_req, gst_req: Count number of global memory load and store requests.
These are available only for GPUs with compute capability 1.2.
4) The summary table is enhanced to display global memory bandwidth values for each kernel. The overall application level global memory bandwidth is also provided using the new menu option "Session->Global Memory Throughput". The global memory bandwidth calculation uses the new counter values gld_32b/64b/128b and gst_32b/64b/128b and is available only for GPUs with compute capability 1.2 or higher.
5) The summary table is enhanced to display instruction bandwidth for each kernel. The instruction bandwidth is a ratio of achieved instruction rate to the peak single instruction issue rate. It uses the "instructions" profiler counter.
6) A new option "Session->Analyze Occupancy" is provided - which reports details of occupancy calculation for each kernel and the factor due to which the maximum occupancy is not achieved.
7) The GPU Time Width Plot is enhanced to display data from multiple streams or from multiple devices. New options "Split On Stream" and "Split on Gpu" are provided in the "Session View Settings" dialog.
8) There is a new option "Timestamp Based Total" provided in the "Session View Settings" dialog for "Summary Plot". This uses the difference in timestamp values for the last and first method to compute the total gpu time instead of using the sum of gpu times for all methods. When this option is used a new bar is displayed to show the "GPU Idle" time.
9) Miscellaneous:
- Handling of large profiler data is improved
- A busy cursor is now displayed for operations which can take long time.
- The host machine configuration can be seen using the new menu option "Help->System Info".
- In session settings dialog all counters, timestamp, kernel and memory transfer options are enabled by default.
- The cputime and profiler counter columns are hidden by default in the summary table.
- CUDA Visual Profiler is now integrated into the CUDA Toolkit installer and the version number has been changed from 1.1 directly to 2.2 to match the CUDA toolkit version.
Key changes in version cudaprof v1.1 with respect to v1.0:
1) Enhancements to profiler output:
a) New columns added for kernel methods:
- Size of grid of blocks (grid size X, grid size Y)
- Size of a thread block (block size X, block size Y, block size Z)
- Register count per thread
- Static shared memory size per block
- Dynamic shared memory size per block
- StreamID of kernel launched
b) New columns added for memcopy methods:
- number of bytes
- direction of transfer (host to device or device to host)
- cputime
2) New view options added:
a) Comparison summary plot : This plot can be used to compare summary profiling data for two sessions.
b) Kernel table : This lists number of calls, grid size, block size, shared memory size per block and register count per thread for each kernel.
c) Memcopy table: This lists number of calls, memory transfer size in bytes and memory transfer direction for each memcopy.
3) cudaprof now detects whether a CUDA capable device is available on the system. If a CUDA device is not found the following message is displayed:
"Unable to load cuda library. CUDA Visual Profiler device features will be disabled."
and certain options like Profile menu options are disabled.
Also based on the device type certain options are enabled or disabled. A new option "Profile->Device Properties" is provided to display cuda device properties.
4) The cputime value displayed is adjusted based on whether kernel execution is asynchronous or not.
5) Summary table has a new method display option. User can choose between "base name", "base name with suffix" or "full name". The "base name" option is useful to combine data for different template based kernel methods having the same name.
6) Width plot
a) display with cpu time enabled is changed. cputime is shown as a separate bar below gputime.
b) A new option is added to use occupancy as a bar height option in width plot.
7) Added height zoom option for height plot.
8) Common improvements to plots
a) Added title for each plot.
b) Added option to display plot configuration options.
9) The error reporting during program execution is changed to help in identifying the specific cause of the error.
10) The cudaprof user document (earlier README file) is now converted to HTML format (cudaprof.html) and can also be viewed using the "Help->Cuda Visual Profiler Help" menu option or using the <F1> function key. This new help option is currently only supported on Windows.
11) The format of the cudaprof .cpj project files is changed from plain text to XML. The information for each session which was earlier in separate .csn files is now part of the .cpj file. The new format is used for any new projects created or when existing projects are updated. Existing projects in the old format can also be opened.
12) CUDA device names for all available CUDA devices are now saved for each session and they are shown in Session Properties.
13) Added menu option "File->Delete" to delete a cuda profiler project.
14) In the Windows version the Microsoft Visual C++ libraries are no longer included in the cudaprof ZIP file. If you do not have Microsoft Visual C++ 2005 SP1 installed you will need to download and install the Microsoft Visual C++ 2005 SP1 Redistributable Package.
