<div dir="ltr"><div>Can you remove AC_TM_SLP_VECTORIZE and run the pass unconditionally on >= gfx9?</div><div> </div><div>Thanks, </div><div>Marek </div></div> <div class="gmail_quote"><div dir="ltr">On Fri, Dec 7, 2018 at 12:23 PM Rhys Perry <<a href="mailto:pendingchaos02@gmail.com">pendingchaos02@gmail.com</a>> wrote: </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Signed-off-by: Rhys Perry <<a href="mailto:pendingchaos02@gmail.com" target="_blank">pendingchaos02@gmail.com</a>> --- src/amd/common/ac_llvm_util.c | 9 ++++++--- src/amd/common/ac_llvm_util.h | 1 + src/amd/vulkan/radv_shader.c | 3 +++ 3 files changed, 10 insertions(+), 3 deletions(-) diff --git a/src/amd/common/ac_llvm_util.c b/src/amd/common/ac_llvm_util.c index dc9b684e9d..3219126188 100644 --- a/src/amd/common/ac_llvm_util.c +++ b/src/amd/common/ac_llvm_util.c @@ -33,6 +33,7 @@ #if HAVE_LLVM >= 0x0700 #include <llvm-c/Transforms/Utils.h> #endif +#include <llvm-c/Transforms/Vectorize.h> #include "c11/threads.h" #include "gallivm/lp_bld_misc.h" #include "util/u_math.h" @@ -177,7 +178,7 @@ static LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family, } static LLVMPassManagerRef ac_create_passmgr(LLVMTargetLibraryInfoRef target_library_info, - bool check_ir) + enum ac_target_machine_options tm_options) { LLVMPassManagerRef passmgr = LLVMCreatePassManager(); if (!passmgr) @@ -187,7 +188,7 @@ static LLVMPassManagerRef ac_create_passmgr(LLVMTargetLibraryInfoRef target_libr LLVMAddTargetLibraryInfo(target_library_info, passmgr); - if (check_ir) + if (tm_options & AC_TM_CHECK_IR) LLVMAddVerifierPass(passmgr); LLVMAddAlwaysInlinerPass(passmgr); /* Normally, the pass manager runs all passes on one function before @@ -203,6 +204,8 @@ static LLVMPassManagerRef ac_create_passmgr(LLVMTargetLibraryInfoRef target_libr LLVMAddLICMPass(passmgr); LLVMAddAggressiveDCEPass(passmgr); LLVMAddCFGSimplificationPass(passmgr); + if (tm_options & AC_TM_SLP_VECTORIZE) + LLVMAddSLPVectorizePass(passmgr); /* This is recommended by the instruction combining pass. */ LLVMAddEarlyCSEMemSSAPass(passmgr); LLVMAddInstructionCombiningPass(passmgr); @@ -332,7 +335,7 @@ ac_init_llvm_compiler(struct ac_llvm_compiler *compiler, } compiler->passmgr = ac_create_passmgr(compiler->target_library_info, - tm_options & AC_TM_CHECK_IR); + tm_options); if (!compiler->passmgr) goto fail; diff --git a/src/amd/common/ac_llvm_util.h b/src/amd/common/ac_llvm_util.h index eaf5f21876..35b12c6c6c 100644 --- a/src/amd/common/ac_llvm_util.h +++ b/src/amd/common/ac_llvm_util.h @@ -65,6 +65,7 @@ enum ac_target_machine_options { AC_TM_CHECK_IR = (1 << 5), AC_TM_ENABLE_GLOBAL_ISEL = (1 << 6), AC_TM_CREATE_LOW_OPT = (1 << 7), + AC_TM_SLP_VECTORIZE = (1 << 8), }; enum ac_float_mode { diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c index 9ba20ac72e..a2ddf17680 100644 --- a/src/amd/vulkan/radv_shader.c +++ b/src/amd/vulkan/radv_shader.c @@ -598,6 +598,9 @@ shader_variant_create(struct radv_device *device, tm_options |= AC_TM_SISCHED; if (options->check_ir) tm_options |= AC_TM_CHECK_IR; + /* vectorization is disabled on pre-GFX9 because it's not very useful there */ + if (device->physical_device->rad_info.chip_class >= GFX9) + tm_options |= AC_TM_SLP_VECTORIZE; thread_compiler = !(device->instance->debug_flags & RADV_DEBUG_NOTHREADLLVM); radv_init_llvm_once(); -- 2.19.2 _______________________________________________ mesa-dev mailing list <a href="mailto:mesa-dev@lists.freedesktop.org" target="_blank">mesa-dev@lists.freedesktop.org</a> <a href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev" rel="noreferrer" target="_blank">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a> </blockquote></div>