radv: add a workaround for DXVK hangs by setting amdgpu-skip-threshold
authorSamuel Pitoiset <samuel.pitoiset@gmail.com>
Fri, 8 Jun 2018 09:38:01 +0000 (11:38 +0200)
committerSamuel Pitoiset <samuel.pitoiset@gmail.com>
Sat, 9 Jun 2018 12:16:49 +0000 (14:16 +0200)
Workaround for bug in llvm that causes the GPU to hang in presence
of nested loops because there is an exec mask issue. The proper
solution is to fix LLVM but this might require a bunch of work.

This fixes a bunch of GPU hangs that happen with DXVK.

Vega10:
Totals from affected shaders:
SGPRS: 110456 -> 110456 (0.00 %)
VGPRS: 122800 -> 122800 (0.00 %)
Spilled SGPRs: 7478 -> 7478 (0.00 %)
Spilled VGPRs: 36 -> 36 (0.00 %)
Code Size: 9901104 -> 9922928 (0.22 %) bytes
Max Waves: 7143 -> 7143 (0.00 %)

Code size slightly increases because it inserts more branch
instructions but that's expected. I don't see any real performance
changes.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105613
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
src/amd/vulkan/radv_shader.c

index 68f2a55e810ac47d47be297249c72317368ab2c7..76790a19047a86abdad5668690675f0679f1f572 100644 (file)
@@ -36,6 +36,7 @@
 
 #include <llvm-c/Core.h>
 #include <llvm-c/TargetMachine.h>
+#include <llvm-c/Support.h>
 
 #include "sid.h"
 #include "gfx9d.h"
@@ -467,6 +468,82 @@ radv_fill_shader_variant(struct radv_device *device,
        memcpy(ptr, binary->code, binary->code_size);
 }
 
+static void radv_init_llvm_target()
+{
+       LLVMInitializeAMDGPUTargetInfo();
+       LLVMInitializeAMDGPUTarget();
+       LLVMInitializeAMDGPUTargetMC();
+       LLVMInitializeAMDGPUAsmPrinter();
+
+       /* For inline assembly. */
+       LLVMInitializeAMDGPUAsmParser();
+
+       /* Workaround for bug in llvm 4.0 that causes image intrinsics
+        * to disappear.
+        * https://reviews.llvm.org/D26348
+        *
+        * Workaround for bug in llvm that causes the GPU to hang in presence
+        * of nested loops because there is an exec mask issue. The proper
+        * solution is to fix LLVM but this might require a bunch of work.
+        * https://bugs.llvm.org/show_bug.cgi?id=37744
+        *
+        * "mesa" is the prefix for error messages.
+        */
+       const char *argv[3] = { "mesa", "-simplifycfg-sink-common=false",
+                               "-amdgpu-skip-threshold=1" };
+       LLVMParseCommandLineOptions(3, argv, NULL);
+}
+
+static once_flag radv_init_llvm_target_once_flag = ONCE_FLAG_INIT;
+
+static LLVMTargetRef radv_get_llvm_target(const char *triple)
+{
+       LLVMTargetRef target = NULL;
+       char *err_message = NULL;
+
+       call_once(&radv_init_llvm_target_once_flag, radv_init_llvm_target);
+
+       if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
+               fprintf(stderr, "Cannot find target for triple %s ", triple);
+               if (err_message) {
+                       fprintf(stderr, "%s\n", err_message);
+               }
+               LLVMDisposeMessage(err_message);
+               return NULL;
+       }
+       return target;
+}
+
+static LLVMTargetMachineRef radv_create_target_machine(enum radeon_family family,
+                                                      enum ac_target_machine_options tm_options,
+                                                      const char **out_triple)
+{
+       assert(family >= CHIP_TAHITI);
+       char features[256];
+       const char *triple = (tm_options & AC_TM_SUPPORTS_SPILL) ? "amdgcn-mesa-mesa3d" : "amdgcn--";
+       LLVMTargetRef target = radv_get_llvm_target(triple);
+
+       snprintf(features, sizeof(features),
+                "+DumpCode,+vgpr-spilling,-fp32-denormals,+fp64-denormals%s%s%s%s",
+                tm_options & AC_TM_SISCHED ? ",+si-scheduler" : "",
+                tm_options & AC_TM_FORCE_ENABLE_XNACK ? ",+xnack" : "",
+                tm_options & AC_TM_FORCE_DISABLE_XNACK ? ",-xnack" : "",
+                tm_options & AC_TM_PROMOTE_ALLOCA_TO_SCRATCH ? ",-promote-alloca" : "");
+
+       LLVMTargetMachineRef tm = LLVMCreateTargetMachine(
+                                    target,
+                                    triple,
+                                    ac_get_llvm_processor_name(family),
+                                    features,
+                                    LLVMCodeGenLevelDefault,
+                                    LLVMRelocDefault,
+                                    LLVMCodeModelDefault);
+
+       if (out_triple)
+               *out_triple = triple;
+       return tm;
+}
+
 static struct radv_shader_variant *
 shader_variant_create(struct radv_device *device,
                      struct radv_shader_module *module,
@@ -501,7 +578,7 @@ shader_variant_create(struct radv_device *device,
                tm_options |= AC_TM_SUPPORTS_SPILL;
        if (device->instance->perftest_flags & RADV_PERFTEST_SISCHED)
                tm_options |= AC_TM_SISCHED;
-       tm = ac_create_target_machine(chip_family, tm_options, NULL);
+       tm = radv_create_target_machine(chip_family, tm_options, NULL);
 
        if (gs_copy_shader) {
                assert(shader_count == 1);