blender
/
blender
136
408
Fork 591
Code Issues 6.3k Pull Requests 775 Projects 19 Wiki Activity

CMake: improve selection/filtering of archs of Intel GPU binaries #112669

Merged
Xavier Hallade merged 3 commits from xavierh/blender:cmake_ocloc_updates into main 2023-10-11 14:59:46 +02:00
Conversation 3 Commits 3 Files Changed 2 +27 -4
2 changed files with 27 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
CMakeLists.txt
View File

@ -684,8 +684,10 @@ This option is only for debugging purposes."
)
# https://www.intel.com/content/www/us/en/develop/documentation/oneapi-dpcpp-cpp-compiler-dev-guide-and-reference/top/compilation/ahead-of-time-compilation.html
# acm-g10 is the target for the first Intel Arc Alchemist GPUs.
set(CYCLES_ONEAPI_SPIR64_GEN_DEVICES "acm-g10" CACHE STRING "\
# The target architectures levels can be retrieved from `ocloc` output when running
xavierh marked this conversation as resolved Outdated
Ray molenkamp commented 2023-09-21 15:41:29 +02:00
Outdated
Review
Copy Link

Where do these magic numbers come from? I was expecting to find them on the page linked in the comment right above them, but that doesn't seem to be the case

Where do these magic numbers come from? I was expecting to find them on the page linked in the comment right above them, but that doesn't seem to be the case
Xavier Hallade commented 2023-09-21 16:26:45 +02:00
Outdated
Review
Copy Link

the page linked is the most user friendly page to explain how to pick targets, using ocloc compile --help.
It doesn't say anything about the magic numbers (architecture versions), I'll ask that it gets improved.

I've retrieved the architecture version number using ocloc ids acm-g10 for Alchemist, and the Meteor Lake one is a bit more convoluted to get yet.. you get it from ocloc when compiling with ocloc for a given meteor lake device id, ie ocloc compile -device 0x7d40 -file test.c, it's also visible from 5569dac6d0/cmake/setup_platform_flags.cmake (L288)

That's maybe a lot of info, which options do you prefer:
a. no change here, I get the webpage improved
b. all the above explanation in the commit message
c. all the above explanation in a comment
d. something else?

the page linked is the most user friendly page to explain how to pick targets, using `ocloc compile --help`. It doesn't say anything about the magic numbers (architecture versions), I'll ask that it gets improved. I've retrieved the architecture version number using `ocloc ids acm-g10` for Alchemist, and the Meteor Lake one is a bit more convoluted to get yet.. you get it from ocloc when compiling with ocloc for a given meteor lake device id, ie `ocloc compile -device 0x7d40 -file test.c`, it's also visible from https://github.com/intel/compute-runtime/blob/5569dac6d0c80ec49b8dac118135ec9ec008cd4b/cmake/setup_platform_flags.cmake#L288 That's maybe a lot of info, which options do you prefer: a. no change here, I get the webpage improved b. all the above explanation in the commit message c. all the above explanation in a comment d. something else?
Xavier Hallade commented 2023-09-26 11:16:04 +02:00
Outdated
Review
Copy Link

I went with c. in latest update.

I went with c. in latest update.
# `ocloc compile -device {device_id} test.c` for given GPUs PCI device IDs.
# 12.55.8 is for Arc Alchemist GPUs. 12.70.0 for Meteor Lake iGPUs.
set(CYCLES_ONEAPI_INTEL_BINARIES_ARCH 12.55.8 12.70.0 CACHE STRING "\
oneAPI Intel GPU architectures to build binaries for"
)
set(CYCLES_ONEAPI_SYCL_TARGETS spir64 spir64_gen CACHE STRING "\
@ -693,7 +695,7 @@ oneAPI targets to build AOT binaries for"
)
mark_as_advanced(WITH_CYCLES_ONEAPI_HOST_TASK_EXECUTION)
mark_as_advanced(CYCLES_ONEAPI_SPIR64_GEN_DEVICES)
mark_as_advanced(CYCLES_ONEAPI_INTEL_BINARIES_ARCH)
mark_as_advanced(CYCLES_ONEAPI_SYCL_TARGETS)
endif()

23
intern/cycles/kernel/CMakeLists.txt
View File

@ -884,7 +884,28 @@ if(WITH_CYCLES_DEVICE_ONEAPI)
endif()
# Enable zebin, a graphics binary format with improved compatibility.
string(PREPEND CYCLES_ONEAPI_SYCL_OPTIONS_spir64_gen "--format zebin ")
string(PREPEND CYCLES_ONEAPI_SYCL_OPTIONS_spir64_gen "-device ${CYCLES_ONEAPI_SPIR64_GEN_DEVICES} ")
# Add the list of Intel devices to build binaries for.
foreach(device ${CYCLES_ONEAPI_INTEL_BINARIES_ARCH})
# Run ocloc ids to test if the device is supported.
if(WIN32)
execute_process(COMMAND ${OCLOC_INSTALL_DIR}/ocloc.exe ids ${device}
RESULT_VARIABLE oclocids_ret
OUTPUT_QUIET
ERROR_QUIET)
else()
execute_process(COMMAND ${OCLOC_INSTALL_DIR}/bin/ocloc ids ${device}
RESULT_VARIABLE oclocids_ret
OUTPUT_QUIET
ERROR_QUIET)
endif()
if(NOT oclocids_ret EQUAL 0)
list(REMOVE_ITEM CYCLES_ONEAPI_INTEL_BINARIES_ARCH ${device})
message(STATUS "binaries for ${device} not supported by Intel Graphics Compiler/ocloc, skipped.")
endif()
endforeach()
list(JOIN CYCLES_ONEAPI_INTEL_BINARIES_ARCH "," gen_devices_string)
string(PREPEND CYCLES_ONEAPI_SYCL_OPTIONS_spir64_gen "-device ${gen_devices_string} ")
# Host execution won't use GPU binaries, no need to compile them.
if(WITH_CYCLES_ONEAPI_BINARIES AND NOT WITH_CYCLES_ONEAPI_HOST_TASK_EXECUTION)