gpuOwl Windows setup for Radeon VII

[kriesel]

Originally Posted by PhilF

Does this build require that MSYS2 be installed? If not, are there any support files that have to be in the same directory as gpuowl-win.exe?

The reason I ask is because I can get gpuowl-win -h to work, but if I supply any exponent using either the -prp or the -pm1 startup option, I get errors and an abrupt Bye.

I am using a Radeon Vii with the latest Radeon software and drivers (dated Dec. 18 2019).

I am a complete beginner with gpuowl so this could easily be my fault.

C:\Users\mrtub\Downloads>gpuowl-win -prp 43000001
2019-12-28 15:41:12 gpuowl v6.11-78-g01d495f-dirty
2019-12-28 15:41:12 Note: no config.txt file found
2019-12-28 15:41:12 config: -prp 43000001
2019-12-28 15:41:12 43000001 FFT 2304K: Width 8x8, Height 256x8, Middle 9; 18.23 bits/word
2019-12-28 15:41:12 OpenCL args "-DEXP=43000001u -DWIDTH=64u -DSMALL_HEIGHT=2048u -DMIDDLE=9u -DWEIGHT_STEP=0xd.aea1d20af415p-3 -DIWEIGHT_STEP=0x9.5af1ae7e10b78p-4 -DWEIGHT_BIGSTEP=0xe.ac0c6e7dd2438p-3 -DIWEIGHT_BIGSTEP=0x8.b95c1e3ea8bd8p-4 -DAMDGPU=1  -I. -cl-fast-relaxed-math -cl-std=CL2.0"
2019-12-28 15:41:13 OpenCL compilation error -11 (args -DEXP=43000001u -DWIDTH=64u -DSMALL_HEIGHT=2048u -DMIDDLE=9u -DWEIGHT_STEP=0xd.aea1d20af415p-3 -DIWEIGHT_STEP=0x9.5af1ae7e10b78p-4 -DWEIGHT_BIGSTEP=0xe.ac0c6e7dd2438p-3 -DIWEIGHT_BIGSTEP=0x8.b95c1e3ea8bd8p-4 -DAMDGPU=1  -I. -cl-fast-relaxed-math -cl-std=CL2.0)
2019-12-28 15:41:13 C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:224:3: error: implicit declaration of function '__asm' is invalid in C99
  X2(u[0], u[2]);
  ^
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:181:2: note: expanded from macro 'X2'
        __asm( "v_add_f64 %0, %1, -%2\n" : "=v" (b.x) : "v" (t.x), "v" (b.x)); \
        ^
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:224:3: error: expected ')'
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:181:35: note: expanded from macro 'X2'
        __asm( "v_add_f64 %0, %1, -%2\n" : "=v" (b.x) : "v" (t.x), "v" (b.x)); \
                                         ^
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:224:3: note: to match this '('
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:181:7: note: expanded from macro 'X2'
        __asm( "v_add_f64 %0, %1, -%2\n" : "=v" (b.x) : "v" (t.x), "v" (b.x)); \
             ^
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:224:3: error: expected ')'
  X2(u[0], u[2]);
  ^
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:182:35: note: expanded from macro 'X2'
        __asm( "v_add_f64 %0, %1, -%2\n" : "=v" (b.y) : "v" (t.y), "v" (b.y)); \
                                         ^
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:224:3: note: to match this '('
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:182:7: note: expanded from macro 'X2'
        __asm( "v_add_f64 %0, %1, -%2\n" : "=v" (b.y) : "v" (t.y), "v" (b.y)); \
             ^
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:225:3: error: expected ')'
  X2_mul_t4(u[1], u[3]);
  ^
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:187:35: note: expanded from macro 'X2_mul_t4'
        __asm( "v_add_f64 %0, %1, -%2\n" : "=v" (t.x) : "v" (b.x), "v" (t.x)); \
                                         ^
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:225:3: note: to match this '('
C:\Users\mrtub\AppData\Local\Temp\\OCL9072T0.cl:187:7: note: expanded from macro 'X2_mul_t4'
        __asm( "v_add_f64 %0, %1, -%2\n" : "=v" (t.x) : "v" (b.x), "v" (t.x)); \
             ^
2019-12-28 15:41:13 Exception gpu_error: BUILD_PROGRAM_FAILURE clBuildProgram at clwrap.cpp:234 build
2019-12-28 15:41:13 Bye

MSYS2 is where to compile it from source for Windows. Executable should be ok standalone, since v6.4 per Preda, not needing a separate gpuowl.cl or other files, except worktodo.txt. https://www.mersenneforum.org/showpo...postcount=1056
-use NO_ASM on the command line or in config.txt on Windows, to avoid all those __asm errors. (Avoiding code intended for ROCm on linux, that's not functional on Windows.) Perhaps other -use directives too, to tune for performance. See https://www.mersenneforum.org/showpo...postcount=1593

[PhilF]

Originally Posted by kriesel

MSYS2 is where to compile it from source for Windows. Executable should be ok standalone, since v6.4 per Preda, not needing a separate gpuowl.cl or other files, except worktodo.txt. https://www.mersenneforum.org/showpo...postcount=1056
-use NO_ASM on the command line or in config.txt on Windows, to avoid all those __asm errors. (Avoiding code intended for ROCm on linux, that's not functional on Windows.) Perhaps other -use directives too, to tune for performance. See https://www.mersenneforum.org/showpo...postcount=1593

I just installed MSYS2, and will try compiling it. But I still don't understand why George's build won't work on my Windows 10 installation, especially since the trivial "-h" option does work with it.

EDIT: I just found that George's build runs for me now (at least for timing purposes with the -time option, haven't tested anything else yet) as long as I include -use ORIG_X2 on the command line. I didn't realize that was required.

[Prime95]

I can no longer build Windows executables. I installed Linux on that box and it wiped out Windows.

My last build was not from the latest source code. Hopefully someone can make available a build from the latest source.

@preda: It would be nice if the make file instructions for gpuowl-win.exe passed in a #define that automatically sets NO_ASM for gpuowl.cl (or sets a #define for WINDOWS=1 which in turn sets NO_ASM in gpuowl.cl. Just a thought.

BTW, I installed Linux on that box because my timings for a 5M FFT were at 830us (and would sometimes go to 11000us when one test ended and the next test began). Whereas after installing Linux and ROCm I was able to get 730us and rock solid behavior. I think the Linux drivers are miles ahead of the Windows drivers.

[preda]

Originally Posted by Prime95

@preda: It would be nice if the make file instructions for gpuowl-win.exe passed in a #define that automatically sets NO_ASM for gpuowl.cl (or sets a #define for WINDOWS=1 which in turn sets NO_ASM in gpuowl.cl. Just a thought.

I just commited a change which automatically retries an OpenCL compilation error with -DNO_ASM=1 ; this should take care of __asm() errors on Windows.