Skip to content

queue improvements #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
stlemme wants to merge 6 commits into
base: main
Choose a base branch
from
Open

queue improvements #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
6766c43
update dependencies
michael-kenzel Nov 24, 2023
31f764d
fix warning
michael-kenzel Nov 24, 2023
ffbca77
improved memory barriers
michael-kenzel Nov 24, 2023
87905f5
improved atomic loads and stores
michael-kenzel Nov 24, 2023
08ddcc9
improved broker queue implementation
michael-kenzel Nov 24, 2023
57bbedb
udpate .gitignore
michael-kenzel Nov 24, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions .gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,3 +3,4 @@ __pycache__/**
dependencies/**
build/**
html/**
plot_data/**
15 changes: 8 additions & 7 deletions build.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,8 +102,8 @@ def __init__(self, dir, buildsystem):
self.rv_source_dir = self.source_dir/"rv"

def pull(self):
pull_git_dependency(self.source_dir, "https://github.com/llvm/llvm-project.git", branch="llvmorg-14.0.1")
pull_git_dependency(self.rv_source_dir, "https://github.com/cdl-saarland/rv.git", "--recurse-submodules", branch="release/14.x")
pull_git_dependency(self.source_dir, "https://github.com/llvm/llvm-project.git", branch="llvmorg-16.0.6")
pull_git_dependency(self.rv_source_dir, "https://github.com/cdl-saarland/rv.git", "--recurse-submodules", branch="release/16.x")

def configure(self, configs, anydsl):
for patch in [
Expand All @@ -129,14 +129,15 @@ def configure(self, configs, anydsl):
LLVM_TOOL_RV_BUILD=True,
CLANG_INCLUDE_DOCS=False,
CLANG_INCLUDE_TESTS=False,
RV_ENABLE_PLUGIN=False,
LLVM_RVPLUG_LINK_INTO_TOOLS=False,
LLVM_BUILD_LLVM_DYLIB=False if windows else True,
LLVM_LINK_LLVM_DYLIB=False if windows else True
)

def build(self, config):
self.buildsystem.build(self.build_dir, config, "clang", "llvm-as", "lld", "LLVMExecutionEngine", "LLVMRuntimeDyld")
self.buildsystem.build(self.build_dir, config, "RV", "LLVMRV", "LLVMMCJIT")
self.buildsystem.build(self.build_dir, config, "RV", "LLVMMCJIT")

@component(depends_on=(LLVM,))
class Thorin(Build):
Expand All @@ -146,7 +147,7 @@ def __init__(self, dir, buildsystem):
self.half_source_dir = self.source_dir/"half"

def pull(self):
pull_git_dependency(self.source_dir, "https://github.com/AnyDSL/thorin.git", branch="ipdps23")
pull_git_dependency(self.source_dir, "https://github.com/AnyDSL/thorin.git", branch="master")
pull_svn_dependency(self.half_source_dir, "https://svn.code.sf.net/p/half/code/tags/release-2.2.0")

def configure(self, configs, llvm):
Expand All @@ -167,7 +168,7 @@ def __init__(self, dir, buildsystem):
self.source_dir = dir/"dependencies"/"anydsl"/"artic"

def pull(self):
pull_git_dependency(self.source_dir, "https://github.com/AnyDSL/artic.git", branch="ipdps23")
pull_git_dependency(self.source_dir, "https://github.com/AnyDSL/artic.git", branch="master")

def configure(self, configs, thorin):
self.buildsystem.configure(self.build_dir, configs, self.source_dir,
Expand All @@ -184,7 +185,7 @@ def __init__(self, dir, buildsystem):
self.source_dir = dir/"dependencies"/"anydsl"/"runtime"

def pull(self):
pull_git_dependency(self.source_dir, "https://github.com/AnyDSL/runtime.git", branch="ipdps23")
pull_git_dependency(self.source_dir, "https://github.com/AnyDSL/runtime.git", branch="master")

def configure(self, configs, anydsl, llvm, thorin, artic):
if windows:
Expand Down Expand Up @@ -217,7 +218,7 @@ def __init__(self, dir, buildsystem):
def configure(self, configs, boost, runtime):
self.buildsystem.configure(self.build_dir, configs, self.source_dir,
Boost_NO_SYSTEM_PATHS=True,
BOOST_ROOT=boost.build_dir,
Boost_ROOT=boost.build_dir,
AnyDSL_runtime_DIR=runtime.build_dir/"share"/"anydsl"/"cmake",
BUILD_TESTING=True,
CMAKE_EXPORT_COMPILE_COMMANDS=True,
Expand Down
139 changes: 119 additions & 20 deletions src/mapping_cuda.art
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,20 +38,37 @@ fn @cuda_pred(b: bool) -> i32 {
if b { 1 } else { 0 }
}

fn @cuda_memory_barrier(order: memory_order) {
if cuda_device_arch() >= 700 {
(@|| {
match order {
memory_order::acquire => asm("fence.acq_rel.gpu;" :::: "volatile"),
memory_order::release => asm("fence.acq_rel.gpu;" :::: "volatile"),
memory_order::acq_rel => asm("fence.acq_rel.gpu;" :::: "volatile"),
memory_order::seq_cst => asm("fence.sc.gpu;" :::: "volatile"),
_ => ()
}
})()
}
else {
cuda_threadfence()
}
}

fn @cuda_legacy_atomic_memory_barrier_pre(order:memory_order) {
match order {
memory_order::release => cuda_threadfence(),
memory_order::acq_rel => cuda_threadfence(),
memory_order::seq_cst => cuda_threadfence(),
memory_order::release => cuda_memory_barrier(memory_order::release),
memory_order::acq_rel => cuda_memory_barrier(memory_order::acq_rel),
memory_order::seq_cst => cuda_memory_barrier(memory_order::seq_cst),
_ => ()
}
}

fn @cuda_legacy_atomic_memory_barrier_post(order:memory_order) {
match order {
memory_order::acquire => cuda_threadfence(),
memory_order::acq_rel => cuda_threadfence(),
memory_order::seq_cst => cuda_threadfence(),
memory_order::acquire => cuda_memory_barrier(memory_order::acquire),
memory_order::acq_rel => cuda_memory_barrier(memory_order::acq_rel),
memory_order::seq_cst => cuda_memory_barrier(memory_order::seq_cst),
_ => ()
}
}
Expand Down Expand Up @@ -99,16 +116,98 @@ fn @cuda_legacy_atomic_memory_order_wrap_cas[T](f: fn(&mut addrspace(1) T, T, T)
}
}

fn @cuda_atomic_wait_and_transition[T](f: fn(&mut addrspace(1) T, T, T) -> T, cmp: fn(T, T) -> bool) {
@|location:&mut addrspace(1) T, expected:T, desired:T, order:memory_order, _debug_msg:&[u8]| -> () {

fn @atomic_load_global_i32(location: &addrspace(1) i32, order: memory_order) -> i32 = atomic_load_global_u32(location as &addrspace(1) u32, order) as i32;
fn @atomic_load_global_u32(location: &addrspace(1) u32, order: memory_order) -> u32 {
let mut value: u32;

if cuda_device_arch() >= 700 {
(@|| {
match order {
memory_order::relaxed => asm("ld.relaxed.gpu.b32 %0, [%1];" : "=r"(value) : "l"(location) : "memory"),
memory_order::acquire => asm("ld.acquire.gpu.b32 %0, [%1];" : "=r"(value) : "l"(location) : "memory"),
_ => cuda_trap()
}
})();
}
else {
cuda_legacy_atomic_memory_barrier_pre(order);
while !cmp(f(location, expected, desired), expected) {
cuda_threadfence();
}
asm("ld.volatile.global.b32 %0, [%1];" : "=r"(value) : "l"(location) : "memory");
cuda_legacy_atomic_memory_barrier_post(order);
}

value
}

fn @atomic_load_global_i64(location: &addrspace(1) i64, order: memory_order) = atomic_load_global_u64(location as &addrspace(1) u64, order) as i64;
fn @atomic_load_global_u64(location: &addrspace(1) u64, order: memory_order) -> u64 {
let mut value: u64;

if cuda_device_arch() >= 700 {
(@|| {
match order {
memory_order::relaxed => asm("ld.relaxed.gpu.b64 %0, [%1];" : "=l"(value) : "l"(location) : "memory"),
memory_order::acquire => asm("ld.acquire.gpu.b64 %0, [%1];" : "=l"(value) : "l"(location) : "memory"),
_ => cuda_trap()
}
})();
}
else {
cuda_legacy_atomic_memory_barrier_pre(order);
asm("ld.volatile.global.b64 %0, [%1];" : "=l"(value) : "l"(location) : "memory");
cuda_legacy_atomic_memory_barrier_post(order);
}

value
}

fn @atomic_store_global_i32(location: &addrspace(1) i32, value: i32, order: memory_order) = atomic_store_global_u32(location as &addrspace(1) u32, value as u32, order);
fn @atomic_store_global_u32(location: &addrspace(1) u32, value: u32, order: memory_order) -> () {
if cuda_device_arch() >= 700 {
(@|| {
match order {
memory_order::relaxed => asm("st.relaxed.gpu.b32 [%0], %1;" : : "l"(location), "r"(value) : "memory"),
memory_order::release => asm("st.release.gpu.b32 [%0], %1;" : : "l"(location), "r"(value) : "memory"),
_ => cuda_trap()
}
})()
}
else {
cuda_legacy_atomic_memory_barrier_pre(order);
asm("st.volatile.global.b32 [%0], %1;" : : "l"(location), "r"(value) : "memory");
cuda_legacy_atomic_memory_barrier_post(order);
}
}

fn @atomic_store_global_i64(location: &addrspace(1) i64, value: i64, order: memory_order) = atomic_store_global_u64(location as &addrspace(1) u64, value as u64, order);
fn @atomic_store_global_u64(location: &addrspace(1) u64, value: u64, order: memory_order) -> () {
if cuda_device_arch() >= 700 {
(@|| {
match order {
memory_order::relaxed => asm("st.relaxed.gpu.b64 [%0], %1;" : : "l"(location), "l"(value) : "memory"),
memory_order::release => asm("st.release.gpu.b64 [%0], %1;" : : "l"(location), "l"(value) : "memory"),
_ => cuda_trap()
}
})()
}
else {
cuda_legacy_atomic_memory_barrier_pre(order);
asm("st.volatile.global.b64 [%0], %1;" : : "l"(location), "l"(value) : "memory");
cuda_legacy_atomic_memory_barrier_post(order);
}
}


// fn @cuda_atomic_wait_and_transition[T](f: fn(&mut addrspace(1) T, T, T) -> T, cmp: fn(T, T) -> bool) {
// @|location:&mut addrspace(1) T, expected:T, desired:T, order:memory_order, _debug_msg:&[u8]| -> () {
// cuda_legacy_atomic_memory_barrier_pre(order);
// while !cmp(f(location, expected, desired), expected) {
// cuda_threadfence();
// }
// cuda_legacy_atomic_memory_barrier_post(order);
// }
// }

fn @cuda_thread(idx: fn(i32) -> u32, gid: fn() -> u32, body: fn(thread_context) -> ()) -> () {
let sleep = @|t: u32| {
if cuda_device_arch() >= 700 {
Expand All @@ -125,20 +224,20 @@ fn @cuda_thread(idx: fn(i32) -> u32, gid: fn() -> u32, body: fn(thread_context)
gid = gid,
uid = @|| gid() as i32,

atomic_load_global_i32 = cuda_legacy_atomic_memory_order_wrap_load[i32](@|location| { let mut res:i32; asm("atom.global.add.s32 %0, [%1], 0;" : "=r"(res) : "l"(location) : "memory"); res }),
atomic_load_global_u32 = cuda_legacy_atomic_memory_order_wrap_load[u32](@|location| { let mut res:u32; asm("atom.global.add.u32 %0, [%1], 0;" : "=r"(res) : "l"(location) : "memory"); res }),
atomic_load_global_i64 = cuda_legacy_atomic_memory_order_wrap_load[i64](@|location| { let mut res:i64; asm("atom.global.add.u64 %0, [%1], 0;" : "=l"(res) : "l"(location) : "memory"); res }),
atomic_load_global_u64 = cuda_legacy_atomic_memory_order_wrap_load[u64](@|location| { let mut res:u64; asm("atom.global.add.u64 %0, [%1], 0;" : "=l"(res) : "l"(location) : "memory"); res }),
atomic_load_global_i32 = atomic_load_global_i32,
atomic_load_global_u32 = atomic_load_global_u32,
atomic_load_global_i64 = atomic_load_global_i64,
atomic_load_global_u64 = atomic_load_global_u64,

atomic_load_global_i32_coalesced = cuda_legacy_atomic_memory_order_wrap_load[i32](@|location| { let mut res:i32; asm("ld.volatile.global.b32 %0, [%1];" : "=r"(res) : "l"(location) : "memory"); res }),
atomic_load_global_u32_coalesced = cuda_legacy_atomic_memory_order_wrap_load[u32](@|location| { let mut res:u32; asm("ld.volatile.global.b32 %0, [%1];" : "=r"(res) : "l"(location) : "memory"); res }),
atomic_load_global_i64_coalesced = cuda_legacy_atomic_memory_order_wrap_load[i64](@|location| { let mut res:i64; asm("ld.volatile.global.b64 %0, [%1];" : "=l"(res) : "l"(location) : "memory"); res }),
atomic_load_global_u64_coalesced = cuda_legacy_atomic_memory_order_wrap_load[u64](@|location| { let mut res:u64; asm("ld.volatile.global.b64 %0, [%1];" : "=l"(res) : "l"(location) : "memory"); res }),

atomic_store_global_i32 = cuda_legacy_atomic_memory_order_wrap_store[i32](@|location, value| { let mut res:i32; asm("atom.global.exch.b32 %0, [%1], %2;" : "=r"(res) : "l"(location), "r"(value) : "memory"); }),
atomic_store_global_u32 = cuda_legacy_atomic_memory_order_wrap_store[u32](@|location, value| { let mut res:u32; asm("atom.global.exch.b32 %0, [%1], %2;" : "=r"(res) : "l"(location), "r"(value) : "memory"); }),
atomic_store_global_i64 = cuda_legacy_atomic_memory_order_wrap_store[i64](@|location, value| { let mut res:i64; asm("atom.global.exch.b64 %0, [%1], %2;" : "=l"(res) : "l"(location), "l"(value) : "memory"); }),
atomic_store_global_u64 = cuda_legacy_atomic_memory_order_wrap_store[u64](@|location, value| { let mut res:u64; asm("atom.global.exch.b64 %0, [%1], %2;" : "=l"(res) : "l"(location), "l"(value) : "memory"); }),
atomic_store_global_i32 = atomic_store_global_i32,
atomic_store_global_u32 = atomic_store_global_u32,
atomic_store_global_i64 = atomic_store_global_i64,
atomic_store_global_u64 = atomic_store_global_u64,

atomic_store_global_i32_coalesced = cuda_legacy_atomic_memory_order_wrap_store[i32](@|location, value| asm("st.volatile.global.b32 [%0], %1;" : : "l"(location), "r"(value) : "memory")),
atomic_store_global_u32_coalesced = cuda_legacy_atomic_memory_order_wrap_store[u32](@|location, value| asm("st.volatile.global.b32 [%0], %1;" : : "l"(location), "r"(value) : "memory")),
Expand Down Expand Up @@ -189,7 +288,7 @@ fn @cuda_thread(idx: fn(i32) -> u32, gid: fn() -> u32, body: fn(thread_context)

atomic_inc_global_u32 = cuda_atomic_inc_global_u32,

memory_barrier = @|_| cuda_threadfence(),
memory_barrier = cuda_memory_barrier,

timestamp = @|| cuda_globaltimer() as i64,
timestamp32 = @|| cuda_globaltimer_lo() as i32,
Expand Down
Loading