[Mesa-dev] [PATCH 11/11] ac/radv: using tls to store llvm related info and speed up compiles (v3)
Dave Airlie
airlied at gmail.com
Wed Jun 27 03:58:45 UTC 2018
From: Dave Airlie <airlied at redhat.com>
I'd like to encourage people to test this to see if it helps (like
does it make app startup better or less hitching in dxvk).
The basic idea is to store a bunch of LLVM related data structs
in thread local storage so we can avoid reiniting them every time
we compile a shader. Since we know llvm objects aren't thread safe
it has to be stored using TLS to avoid any collisions.
This should remove all the fixed overheads setup costs of creating
the pass manager each time.
This takes a demo app time to compile the radv meta shaders on nocache
and exit from 1.7s to 1s.
TODO: this doesn't work for radeonsi yet, but I'm not sure how TLS
works if you have radeonsi and radv loaded at the same time, if
they'll magically try and use the same tls stuff, in which case
this might explode all over the place.
v2: fix llvm6 build, inline emit function, handle multiple targets
in one thread
v3: rebase and port onto new structure
---
src/amd/common/ac_llvm_helper.cpp | 120 ++++++++++++++++++++++++++++--
src/amd/common/ac_llvm_util.c | 10 +--
src/amd/common/ac_llvm_util.h | 9 +++
src/amd/vulkan/radv_debug.h | 1 +
src/amd/vulkan/radv_device.c | 1 +
src/amd/vulkan/radv_shader.c | 2 +
6 files changed, 132 insertions(+), 11 deletions(-)
diff --git a/src/amd/common/ac_llvm_helper.cpp b/src/amd/common/ac_llvm_helper.cpp
index 27403dbe085..f1f1399b3fb 100644
--- a/src/amd/common/ac_llvm_helper.cpp
+++ b/src/amd/common/ac_llvm_helper.cpp
@@ -31,12 +31,21 @@
#include "ac_llvm_util.h"
#include <llvm-c/Core.h>
-#include <llvm/Target/TargetOptions.h>
-#include <llvm/ExecutionEngine/ExecutionEngine.h>
-#include <llvm/IR/Attributes.h>
-#include <llvm/IR/CallSite.h>
+#include <llvm/Target/TargetMachine.h>
#include <llvm/IR/IRBuilder.h>
#include <llvm/Analysis/TargetLibraryInfo.h>
+#include <llvm/IR/LegacyPassManager.h>
+
+#include <llvm-c/Transforms/IPO.h>
+#include <llvm-c/Transforms/Scalar.h>
+#if HAVE_LLVM >= 0x0700
+#include <llvm-c/Transforms/Utils.h>
+#endif
+
+#if HAVE_LLVM < 0x0700
+#include "llvm/Support/raw_ostream.h"
+#endif
+#include <list>
void ac_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes)
{
@@ -101,11 +110,110 @@ ac_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info)
delete reinterpret_cast<llvm::TargetLibraryInfoImpl *>(library_info);
}
+class ac_llvm_per_thread_info {
+public:
+ ac_llvm_per_thread_info(enum radeon_family arg_family,
+ enum ac_target_machine_options arg_tm_options)
+ : family(arg_family), tm_options(arg_tm_options),
+ OStream(CodeString) {}
+ ~ac_llvm_per_thread_info() {
+ ac_llvm_compiler_dispose_internal(&llvm_info);
+ }
+
+ struct ac_llvm_compiler_info llvm_info;
+ enum radeon_family family;
+ enum ac_target_machine_options tm_options;
+ llvm::SmallString<0> CodeString;
+ llvm::raw_svector_ostream OStream;
+ llvm::legacy::PassManager pass;
+};
+
+/* we have to store a linked list per thread due to the possiblity of multiple gpus being required */
+static thread_local std::list<ac_llvm_per_thread_info> ac_llvm_per_thread_list;
+
bool ac_compile_to_memory_buffer(struct ac_llvm_compiler_info *info,
LLVMModuleRef M,
char **ErrorMessage,
LLVMMemoryBufferRef *OutMemBuf)
{
- return LLVMTargetMachineEmitToMemoryBuffer(info->tm, M, LLVMObjectFile,
- ErrorMessage, OutMemBuf);
+ ac_llvm_per_thread_info *thread_info = nullptr;
+ if (info->thread_stored) {
+ for (auto &I : ac_llvm_per_thread_list) {
+ if (I.llvm_info.tm == info->tm) {
+ thread_info = &I;
+ break;
+ }
+ }
+
+ if (!thread_info) {
+ assert(0);
+ return false;
+ }
+ } else {
+ return LLVMTargetMachineEmitToMemoryBuffer(info->tm, M, LLVMObjectFile,
+ ErrorMessage, OutMemBuf);
+ }
+
+ llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(thread_info->llvm_info.tm);
+ llvm::Module *Mod = llvm::unwrap(M);
+ llvm::StringRef Data;
+
+ Mod->setDataLayout(TM->createDataLayout());
+
+ thread_info->pass.run(*Mod);
+
+ Data = thread_info->OStream.str();
+ *OutMemBuf = LLVMCreateMemoryBufferWithMemoryRangeCopy(Data.data(), Data.size(), "");
+ thread_info->CodeString = "";
+ return false;
+}
+
+bool ac_llvm_compiler_init(struct ac_llvm_compiler_info *info,
+ bool add_target_library_info,
+ enum radeon_family family,
+ enum ac_target_machine_options tm_options)
+{
+ if (tm_options & AC_TM_THREAD_LLVM) {
+ for (auto &I : ac_llvm_per_thread_list) {
+ if (I.family == family &&
+ I.tm_options == tm_options) {
+ *info = I.llvm_info;
+ return true;
+ }
+ }
+
+ ac_llvm_per_thread_list.emplace_back(family, tm_options);
+ ac_llvm_per_thread_info &tinfo = ac_llvm_per_thread_list.back();
+ if (!ac_llvm_compiler_init_internal(&tinfo.llvm_info,
+ true,
+ family,
+ tm_options))
+ return false;
+
+ tinfo.llvm_info.thread_stored = true;
+ *info = tinfo.llvm_info;
+
+ llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tinfo.llvm_info.tm);
+ if (TM->addPassesToEmitFile(tinfo.pass, tinfo.OStream,
+#if HAVE_LLVM >= 0x0700
+ nullptr,
+#endif
+ llvm::TargetMachine::CGFT_ObjectFile)) {
+ assert(0);
+ return false;
+ }
+ } else {
+ if (!ac_llvm_compiler_init_internal(info,
+ add_target_library_info,
+ family,
+ tm_options))
+ return false;
+ }
+ return true;
+}
+
+void ac_llvm_compiler_dispose(struct ac_llvm_compiler_info *info)
+{
+ if (!info->thread_stored)
+ ac_llvm_compiler_dispose_internal(info);
}
diff --git a/src/amd/common/ac_llvm_util.c b/src/amd/common/ac_llvm_util.c
index 7f8c5746b37..d8ec4ee0092 100644
--- a/src/amd/common/ac_llvm_util.c
+++ b/src/amd/common/ac_llvm_util.c
@@ -188,10 +188,10 @@ static LLVMPassManagerRef ac_init_passmgr(LLVMTargetLibraryInfoRef target_librar
return passmgr;
}
-bool ac_llvm_compiler_init(struct ac_llvm_compiler_info *info,
- bool add_target_library_info,
- enum radeon_family family,
- enum ac_target_machine_options tm_options)
+bool ac_llvm_compiler_init_internal(struct ac_llvm_compiler_info *info,
+ bool add_target_library_info,
+ enum radeon_family family,
+ enum ac_target_machine_options tm_options)
{
memset(info, 0, sizeof(*info));
info->tm = ac_create_target_machine(family, tm_options, &info->triple);
@@ -223,7 +223,7 @@ fail:
return false;
}
-void ac_llvm_compiler_dispose(struct ac_llvm_compiler_info *info)
+void ac_llvm_compiler_dispose_internal(struct ac_llvm_compiler_info *info)
{
if (info->data_layout)
LLVMDisposeMessage((char*)info->data_layout);
diff --git a/src/amd/common/ac_llvm_util.h b/src/amd/common/ac_llvm_util.h
index 1ba972dab87..c579e6a9ec3 100644
--- a/src/amd/common/ac_llvm_util.h
+++ b/src/amd/common/ac_llvm_util.h
@@ -60,6 +60,7 @@ enum ac_target_machine_options {
AC_TM_FORCE_DISABLE_XNACK = (1 << 3),
AC_TM_PROMOTE_ALLOCA_TO_SCRATCH = (1 << 4),
AC_TM_CHECK_IR = (1 << 5),
+ AC_TM_THREAD_LLVM = (1 << 6),
};
enum ac_float_mode {
@@ -74,6 +75,7 @@ struct ac_llvm_compiler_info {
LLVMTargetLibraryInfoRef target_library_info;
const char *triple;
const char *data_layout;
+ bool thread_stored;
};
const char *ac_get_llvm_processor_name(enum radeon_family family);
@@ -114,6 +116,12 @@ ac_get_store_intr_attribs(bool writeonly_memory)
unsigned
ac_count_scratch_private_memory(LLVMValueRef function);
+bool ac_llvm_compiler_init_internal(struct ac_llvm_compiler_info *info,
+ bool add_target_library_info, /* crash workaround */
+ enum radeon_family family,
+ enum ac_target_machine_options tm_options);
+void ac_llvm_compiler_dispose_internal(struct ac_llvm_compiler_info *info);
+
bool ac_llvm_compiler_init(struct ac_llvm_compiler_info *info,
bool add_target_library_info, /* crash workaround */
enum radeon_family family,
@@ -127,6 +135,7 @@ bool ac_compile_to_memory_buffer(struct ac_llvm_compiler_info *info,
LLVMModuleRef M,
char **ErrorMessage,
LLVMMemoryBufferRef *OutMemBuf);
+
#ifdef __cplusplus
}
#endif
diff --git a/src/amd/vulkan/radv_debug.h b/src/amd/vulkan/radv_debug.h
index f1b0dc26a63..9fe4c3b7404 100644
--- a/src/amd/vulkan/radv_debug.h
+++ b/src/amd/vulkan/radv_debug.h
@@ -49,6 +49,7 @@ enum {
RADV_DEBUG_ERRORS = 0x80000,
RADV_DEBUG_STARTUP = 0x100000,
RADV_DEBUG_CHECKIR = 0x200000,
+ RADV_DEBUG_NOTHREADLLVM = 0x400000,
};
enum {
diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
index ad3465f594e..73c48cef1f0 100644
--- a/src/amd/vulkan/radv_device.c
+++ b/src/amd/vulkan/radv_device.c
@@ -436,6 +436,7 @@ static const struct debug_control radv_debug_options[] = {
{"errors", RADV_DEBUG_ERRORS},
{"startup", RADV_DEBUG_STARTUP},
{"checkir", RADV_DEBUG_CHECKIR},
+ {"nothreadllvm", RADV_DEBUG_NOTHREADLLVM},
{NULL, 0}
};
diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index 3bbb7867640..f7033aff771 100644
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -547,6 +547,8 @@ shader_variant_create(struct radv_device *device,
tm_options |= AC_TM_SISCHED;
if (options->check_ir)
tm_options |= AC_TM_CHECK_IR;
+ if (!(device->instance->debug_flags & RADV_DEBUG_NOTHREADLLVM))
+ tm_options |= AC_TM_THREAD_LLVM;
radv_init_llvm_once();
--
2.17.1
More information about the mesa-dev
mailing list