Details
Diff Detail
- Build Status
Buildable 8412 Build 8412: arc lint + arc unit
Event Timeline
Nice. Is there any reason this is still marked as WIP? Sould I already have a look.
(Marking this as requesting changes to make sure this does not ye show up in my review queue)
@grosser: I wanted to rebase and get a clean history, but PPCGCodeGeneration can be reviewed. The test cases I am dubious about are:
- test/GPGPU/mostly-sequential.ll since we pull in an entire loop into the kernel
- test/GPGPU/non-read-only-scalars.ll since the data flow is wrong.
- test/GPGPU/phi-nodes-in-kernel.ll since we stopped sending a phi node (this makes sense, but I want a double check)
lib/CodeGen/PPCGCodeGeneration.cpp | ||
---|---|---|
58 | DO NOT COMMIT! :) | |
1169 | Sure. For now, I'm relying on the fact that it worked before, so extra annotations shouldn't actually hurt. However, we do seem to mis-compile in 2 of the test cases. I'll have to see what code PPCG emits. | |
test/GPGPU/host-control-flow.ll | ||
27 | Yes, possibly. However, we generate this already without PPCG helping us, do we not? |
lib/CodeGen/PPCGCodeGeneration.cpp | ||
---|---|---|
1169 | GPUNodeBuilder has already been handling these cases. I think the schedule tree just become more detailed. Do these nodes reside in a Block ? in the order { "init_device", "to_device"+, "kernel", "from_device"+, "clear_device" } ? | |
lib/External/ppcg/cuda.c | ||
530 | You were right ! We've already been doing stuff under "init_device" and "clear_device". You could do something like, if (!strcmp(Str, "init_device") ) { initializeAfterRTH(); preloadInvariantLoads(); isl_ast_node_free(UserStmt); isl_ast_expr_free(Expr); return; } if ( !strcmp(Str, "clear_device") ) { finalize(); isl_ast_node_free(UserStmt); isl_ast_expr_free(Expr); return; } |
[Polly] [PPCG] [1/3] Bump up PPCG to vanilla 0.07
This commit *WILL BREAK*. We are first upgrading to PPCG. We make Polly
specific changes in a separate commit.
This makes upgrading PPCG the next time easier since the changes made to PPCG
for Polly will be clearly documented in a separate commit.
Hm, perhaps updating this diff is too confusing. I am abandoning this huge change and splitting this into 3 chunks.
"polly-debug-print-bounds" ;-)