Enable Geekbench 5 (#304)

vosen · Dec 10, 2024 · 7ac67a8 · 7ac67a8
1 parent 7a6df9d
commit 7ac67a8
Show file tree

Hide file tree

Showing 42 changed files with 24,893 additions and 15,443 deletions.
diff --git a/.devcontainer/Dockerfile b/.devcontainer/Dockerfile
@@ -27,6 +27,7 @@ RUN DEBIAN_FRONTEND=noninteractive apt-get update -y && DEBIAN_FRONTEND=noninter
     libgl-dev libegl-dev libvdpau-dev \
     nvidia-utils-${CUDA_DRIVER} \
     cuda-cudart-dev-${CUDA_PKG_VERSION} \
+    cuda-nvml-dev-${CUDA_PKG_VERSION} \
     cuda-cudart-${CUDA_PKG_VERSION} \
     cuda-profiler-api-${CUDA_PKG_VERSION} \
     cuda-nvcc-${CUDA_PKG_VERSION}

diff --git a/comgr/src/lib.rs b/comgr/src/lib.rs
@@ -1,5 +1,5 @@
 use amd_comgr_sys::*;
-use std::{ffi::CStr, iter, mem, ptr};
+use std::{ffi::CStr, mem, ptr};
 
 struct Data(amd_comgr_data_t);
 
@@ -137,15 +137,29 @@ pub fn compile_bitcode(
     link_with_device_libs_info.set_isa_name(gcn_arch)?;
     link_with_device_libs_info.set_language(amd_comgr_language_t::AMD_COMGR_LANGUAGE_LLVM_IR)?;
     // This makes no sense, but it makes ockl linking work
-    link_with_device_libs_info.set_options([c"-Xclang", c"-mno-link-builtin-bitcode-postopt"].into_iter())?;
+    link_with_device_libs_info
+        .set_options([c"-Xclang", c"-mno-link-builtin-bitcode-postopt"].into_iter())?;
     let with_device_libs = do_action(
         &linked_data_set,
         &link_with_device_libs_info,
         amd_comgr_action_kind_t::AMD_COMGR_ACTION_COMPILE_SOURCE_WITH_DEVICE_LIBS_TO_BC,
     )?;
     let compile_action_info = ActionInfo::new()?;
     compile_action_info.set_isa_name(gcn_arch)?;
-    compile_action_info.set_options(iter::once(c"-O3"))?;
+    let common_options = [c"-O3", c"-mno-wavefrontsize64", c"-mcumode"].into_iter();
+    let opt_options = if cfg!(debug_assertions) {
+        [c"-g", c"", c"", c"", c""]
+    } else {
+        [
+            c"-g0",
+            // default inlining threshold times 10
+            c"-mllvm",
+            c"-inline-threshold=2250",
+            c"-mllvm",
+            c"-inlinehint-threshold=3250",
+        ]
+    };
+    compile_action_info.set_options(common_options.chain(opt_options))?;
     let reloc_data_set = do_action(
         &with_device_libs,
         &compile_action_info,

diff --git a/cuda_base/Cargo.toml b/cuda_base/Cargo.toml
@@ -6,7 +6,7 @@ edition = "2021"
 
 [dependencies]
 quote = "1.0"
-syn = { version = "2.0", features = ["full", "visit-mut"] }
+syn = { version = "2.0", features = ["full", "visit-mut", "extra-traits"] }
 proc-macro2 = "1.0"
 rustc-hash = "1.1.0"