Hi! Ping. Grüße Thomas On 2022-12-15T19:27:08+0100, I wrote: > Hi Tom! > > First "a bit" of context; skip to "the proposed patch" if you'd like to > see just that. > > > On 2022-02-01T19:31:27+0100, Tom de Vries via Gcc-patches wrote: >> On a GT 1030, with driver version 470.94 and -mptx=3.1 I run into: >> ... >> FAIL: libgomp.oacc-c/../libgomp.oacc-c-c++-common/parallel-dims.c \ >> -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none \ >> -O2 execution test >> ... >> which minimizes to the same test-case as listed in commit "[nvptx] >> Update default ptx isa to 6.3". >> >> The problem is again that the first diverging branch is not handled as such in >> SASS, which causes problems with a subsequent shfl insn, but given that we >> have -mptx=3.1 we can't use the bar.warp.sync insn. >> >> Given that the default is now -mptx=6.3, and consequently -mptx=3.1 is of a >> lesser importance, implement the next best thing: abort when detecting >> non-convergence using this insn: >> ... >> { .reg.b32 act; >> vote.ballot.b32 act,1; >> .reg.pred uni; >> setp.eq.b32 uni,act,0xffffffff; >> @ !uni trap; >> @ !uni exit; >> } >> ... >> >> Interestingly, the effect of this is that rather than aborting, the test-case >> now passes. > > (I suppose this "nudges" the PTX -> SASS compiler into the right > direction?) > > > For avoidance of doubt, my following discussion is not about the specific > (first) use of 'nvptx_uniform_warp_check' introduced here in this > commit r12-6971-gf32f74c2e8cef5fe37af6d4e8d7e8f6b4c8ae9a8 > "[nvptx] Add uniform_warp_check insn": > >> --- a/gcc/config/nvptx/nvptx.cc >> +++ b/gcc/config/nvptx/nvptx.cc >> @@ -4631,15 +4631,29 @@ nvptx_single (unsigned mask, basic_block from, basic_block to) >> if (tail_branch) >> { >> label_insn = emit_label_before (label, before); >> - if (TARGET_PTX_6_0 && mode == GOMP_DIM_VECTOR) >> - warp_sync = emit_insn_after (gen_nvptx_warpsync (), label_insn); >> + if (mode == GOMP_DIM_VECTOR) >> + { >> + if (TARGET_PTX_6_0) >> + warp_sync = emit_insn_after (gen_nvptx_warpsync (), >> + label_insn); >> + else >> + warp_sync = emit_insn_after (gen_nvptx_uniform_warp_check (), >> + label_insn); >> + } >> before = label_insn; >> } >> else >> { >> label_insn = emit_label_after (label, tail); >> - if (TARGET_PTX_6_0 && mode == GOMP_DIM_VECTOR) >> - warp_sync = emit_insn_after (gen_nvptx_warpsync (), label_insn); >> + if (mode == GOMP_DIM_VECTOR) >> + { >> + if (TARGET_PTX_6_0) >> + warp_sync = emit_insn_after (gen_nvptx_warpsync (), >> + label_insn); >> + else >> + warp_sync = emit_insn_after (gen_nvptx_uniform_warp_check (), >> + label_insn); >> + } >> if ((mode == GOMP_DIM_VECTOR || mode == GOMP_DIM_WORKER) >> && CALL_P (tail) && find_reg_note (tail, REG_NORETURN, NULL)) >> emit_insn_after (gen_exit (), label_insn); > > Later, other uses have been added, for example in OpenMP '-muniform-simt' > code generation. > > My following discussion is about the implementation of > 'nvptx_uniform_warp_check', originally introduced as follows: > >> --- a/gcc/config/nvptx/nvptx.md >> +++ b/gcc/config/nvptx/nvptx.md >> @@ -57,6 +57,7 @@ (define_c_enum "unspecv" [ >> UNSPECV_XCHG >> UNSPECV_BARSYNC >> UNSPECV_WARPSYNC >> + UNSPECV_UNIFORM_WARP_CHECK >> UNSPECV_MEMBAR >> UNSPECV_MEMBAR_CTA >> UNSPECV_MEMBAR_GL >> @@ -1985,6 +1986,23 @@ (define_insn "nvptx_warpsync" >> "\\tbar.warp.sync\\t0xffffffff;" >> [(set_attr "predicable" "false")]) >> >> +(define_insn "nvptx_uniform_warp_check" >> + [(unspec_volatile [(const_int 0)] UNSPECV_UNIFORM_WARP_CHECK)] >> + "" >> + { >> + output_asm_insn ("{", NULL); >> + output_asm_insn ("\\t" ".reg.b32" "\\t" "act;", NULL); >> + output_asm_insn ("\\t" "vote.ballot.b32" "\\t" "act,1;", NULL); >> + output_asm_insn ("\\t" ".reg.pred" "\\t" "uni;", NULL); >> + output_asm_insn ("\\t" "setp.eq.b32" "\\t" "uni,act,0xffffffff;", >> + NULL); >> + output_asm_insn ("@ !uni\\t" "trap;", NULL); >> + output_asm_insn ("@ !uni\\t" "exit;", NULL); >> + output_asm_insn ("}", NULL); >> + return ""; >> + } >> + [(set_attr "predicable" "false")]) > > Later adjusted, but the fundamental idea is still the same. > > > Via temporarily disabling 'nvptx_uniform_warp_check': > > (define_insn "nvptx_uniform_warp_check" > [(unspec_volatile [(const_int 0)] UNSPECV_UNIFORM_WARP_CHECK)] > "" > { > +#if 0 > const char *insns[] = { > "{", > "\\t" ".reg.b32" "\\t" "%%r_act;", > "%.\\t" "vote.ballot.b32" "\\t" "%%r_act,1;", > "\\t" ".reg.pred" "\\t" "%%r_do_abort;", > "\\t" "mov.pred" "\\t" "%%r_do_abort,0;", > "%.\\t" "setp.ne.b32" "\\t" "%%r_do_abort,%%r_act," > "0xffffffff;", > "@ %%r_do_abort\\t" "trap;", > "@ %%r_do_abort\\t" "exit;", > "}", > NULL > }; > for (const char **p = &insns[0]; *p != NULL; p++) > output_asm_insn (*p, NULL); > +#endif > return ""; > }) > > ..., I've first tested/confirmed the problem that it was originally > solving. Testing with: > > $ nvidia-smi > [...] > | NVIDIA-SMI 515.65.01 Driver Version: 515.65.01 CUDA Version: 11.7 | > [...] > | 0 Quadro P1000 [...] > > For 'check-gcc' with '--target_board=nvptx-none-run/-mptx=3.1 nvptx.exp', > this (obviously) regresses: > > PASS: gcc.target/nvptx/uniform-simt-2.c (test for excess errors) > PASS: gcc.target/nvptx/uniform-simt-2.c scan-assembler-times @%r[0-9]*\tatom.global.cas 1 > PASS: gcc.target/nvptx/uniform-simt-2.c scan-assembler-times shfl.idx.b32 1 > [-PASS:-]{+FAIL:+} gcc.target/nvptx/uniform-simt-2.c scan-assembler-times vote.ballot.b32 1 > > For 'check-target-libgomp' with > '--target_board=unix/-foffload-options=nvptx-none=-mptx=3.1', there are > not obvious regressions for any OpenMP test cases. > > For example, for the test case 'libgomp.c/pr104783-2.c' of > commit a624388b9546b066250be8baa118b7d50c403c25 > "[nvptx] Add warp sync at simt exit", 'nvptx_uniform_warp_check' is not > applicable per se: this is about an issue with sm_70+ Independent Thread > Scheduling, which is applicable only 'if (TARGET_PTX_6_0)', and in that > case, we emit 'nvptx_warpsync', not 'nvptx_uniform_warp_check'. > > For other OpenMP test cases (which I've not analyzed in detail), we're > maybe simply lucky that 'nvptx_uniform_warp_check' is not relevant > (... at least in this testing configuration). (For avoidance of doubt, I > have no reason to believe that there's any problem with the > PR104783 "[nvptx, openmp] Hang/abort with atomic update in simd construct", > PR104916 "[nvptx] Handle Independent Thread Scheduling for sm_70+ with -muniform-simt", > "[nvptx] Use nvptx_warpsync / nvptx_uniform_warp_check for -muniform-simt", > or other such code changes; mentioning this just for completeness.) > > ..., but as regards OpenACC test cases, this still regresses several: > > [-PASS:-]{+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/parallel-dims.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O2 execution test > > (That's the one cited in the commit log of > commit r12-6971-gf32f74c2e8cef5fe37af6d4e8d7e8f6b4c8ae9a8 > "[nvptx] Add uniform_warp_check insn".) > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-10.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-10.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O2 execution test > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-4.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test[-PASS: libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-4.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 output pattern test-] > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-5.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test[-PASS: libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-5.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 output pattern test-] > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-6.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test[-PASS: libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-6.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 output pattern test-] > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-7.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test[-PASS: libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-128-7.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 output pattern test-] > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-64-1.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/vector-length-64-3.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-c/../libgomp.oacc-c-c++-common/vred2d-128.c -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test > > Same for C++. > > [-PASS:-]{+FAIL:+} libgomp.oacc-c++/ref-1.C -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O2 execution test > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-fortran/gemm-2.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test > > [-PASS:-]{+WARNING: program timed out.+} > {+FAIL:+} libgomp.oacc-fortran/gemm.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/parallel-dims.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O2 execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/parallel-dims.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/parallel-dims.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O3 -g execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/parallel-dims.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -Os execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/routine-7.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O0 execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/routine-7.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O1 execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/routine-7.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O2 execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/routine-7.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/routine-7.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -O3 -g execution test > > [-PASS:-]{+FAIL:+} libgomp.oacc-fortran/routine-7.f90 -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none -Os execution test > > So that's "good": plenty of evidence that 'nvptx_uniform_warp_check' is > necessary and working. > > > Now, "the proposed patch". I'd like to make 'nvptx_uniform_warp_check' > fit for non-full-warp execution. For example, to be able to execute such > code in single-threaded 'cuLaunchKernel' for execution of global > constructors/destructors, where those may, for example, call into nvptx > target libraries compiled with '-mgomp' (thus, '-muniform-simt'). > > OK to push (after proper testing, and with TODO markers adjusted/removed) > the attached > "nvptx: Make 'nvptx_uniform_warp_check' fit for non-full-warp execution"? > > > Grüße > Thomas ----------------- Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955