[Offload] Port llvm-offload-device-info to new offload API #155626

RossBrunton · 2025-08-27T14:15:42Z

This is a tool similar to urinfo that simply prints properties of all
devices. The old openMP version has been ported to liboffload.

llvmbot · 2025-08-27T14:16:18Z

@llvm/pr-subscribers-offload

Author: Ross Brunton (RossBrunton)

Changes

This is a tool similar to llvm-offload-device-info, urinfo, etc. that
simply prints properties of all devices. The difference between this
and llvm-offload-device-info is that this uses liboffload rather than
openmp.

Full diff: https://github.com/llvm/llvm-project/pull/155626.diff

3 Files Affected:

(modified) offload/tools/CMakeLists.txt (+1)
(added) offload/tools/liboffload-deviceinfo/CMakeLists.txt (+9)
(added) offload/tools/liboffload-deviceinfo/llvm-liboffload-device-info.cpp (+265)

diff --git a/offload/tools/CMakeLists.txt b/offload/tools/CMakeLists.txt
index 13772a4a675e5..860eed30905f4 100644
--- a/offload/tools/CMakeLists.txt
+++ b/offload/tools/CMakeLists.txt
@@ -14,3 +14,4 @@ endmacro()
 
 add_subdirectory(deviceinfo)
 add_subdirectory(kernelreplay)
+add_subdirectory(liboffload-deviceinfo)
diff --git a/offload/tools/liboffload-deviceinfo/CMakeLists.txt b/offload/tools/liboffload-deviceinfo/CMakeLists.txt
new file mode 100644
index 0000000000000..37ce65721ca0f
--- /dev/null
+++ b/offload/tools/liboffload-deviceinfo/CMakeLists.txt
@@ -0,0 +1,9 @@
+message(STATUS "Building the llvm-liboffload-device-info tool")
+
+add_openmp_tool(llvm-liboffload-device-info llvm-liboffload-device-info.cpp)
+
+llvm_update_compile_flags(llvm-liboffload-device-info)
+
+target_link_libraries(llvm-liboffload-device-info PRIVATE
+  LLVMOffload
+)
diff --git a/offload/tools/liboffload-deviceinfo/llvm-liboffload-device-info.cpp b/offload/tools/liboffload-deviceinfo/llvm-liboffload-device-info.cpp
new file mode 100644
index 0000000000000..93c67af0ecea6
--- /dev/null
+++ b/offload/tools/liboffload-deviceinfo/llvm-liboffload-device-info.cpp
@@ -0,0 +1,265 @@
+//===- llvm-liboffload-device-info.cpp - Print liboffload properties ------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This is a command line utility that, by using the new liboffload API, prints
+// all devices and properties
+//
+//===----------------------------------------------------------------------===//
+
+#include <OffloadAPI.h>
+#include <iostream>
+#include <vector>
+
+#define OFFLOAD_ERR(X)                                                         \
+  if (auto Err = X) {                                                          \
+    return Err;                                                                \
+  }
+
+enum class PrintKind {
+  NORMAL,
+  FP_FLAGS,
+};
+
+template <typename T, PrintKind PK = PrintKind::NORMAL>
+void doWrite(std::ostream &S, T &&Val) {
+  S << Val;
+}
+
+template <>
+void doWrite<ol_platform_backend_t>(std::ostream &S,
+                                    ol_platform_backend_t &&Val) {
+  switch (Val) {
+  case OL_PLATFORM_BACKEND_UNKNOWN:
+    S << "UNKNOWN";
+    break;
+  case OL_PLATFORM_BACKEND_CUDA:
+    S << "CUDA";
+    break;
+  case OL_PLATFORM_BACKEND_AMDGPU:
+    S << "AMDGPU";
+    break;
+  case OL_PLATFORM_BACKEND_HOST:
+    S << "HOST";
+    break;
+  default:
+    S << "<< INVALID >>";
+    break;
+  }
+}
+template <>
+void doWrite<ol_device_type_t>(std::ostream &S, ol_device_type_t &&Val) {
+  switch (Val) {
+  case OL_DEVICE_TYPE_GPU:
+    S << "GPU";
+    break;
+  case OL_DEVICE_TYPE_CPU:
+    S << "CPU";
+    break;
+  case OL_DEVICE_TYPE_HOST:
+    S << "HOST";
+    break;
+  default:
+    S << "<< INVALID >>";
+    break;
+  }
+}
+template <>
+void doWrite<ol_dimensions_t>(std::ostream &S, ol_dimensions_t &&Val) {
+  S << "{x: " << Val.x << ", y: " << Val.y << ", z: " << Val.z << "}";
+}
+template <>
+void doWrite<ol_device_fp_capability_flags_t, PrintKind::FP_FLAGS>(
+    std::ostream &S, ol_device_fp_capability_flags_t &&Val) {
+  S << Val << " {";
+
+  if (Val & OL_DEVICE_FP_CAPABILITY_FLAG_CORRECTLY_ROUNDED_DIVIDE_SQRT) {
+    S << " CORRECTLY_ROUNDED_DIVIDE_SQRT";
+  }
+  if (Val & OL_DEVICE_FP_CAPABILITY_FLAG_ROUND_TO_NEAREST) {
+    S << " ROUND_TO_NEAREST";
+  }
+  if (Val & OL_DEVICE_FP_CAPABILITY_FLAG_ROUND_TO_ZERO) {
+    S << " ROUND_TO_ZERO";
+  }
+  if (Val & OL_DEVICE_FP_CAPABILITY_FLAG_ROUND_TO_INF) {
+    S << " ROUND_TO_INF";
+  }
+  if (Val & OL_DEVICE_FP_CAPABILITY_FLAG_INF_NAN) {
+    S << " INF_NAN";
+  }
+  if (Val & OL_DEVICE_FP_CAPABILITY_FLAG_DENORM) {
+    S << " DENORM";
+  }
+  if (Val & OL_DEVICE_FP_CAPABILITY_FLAG_FMA) {
+    S << " FMA";
+  }
+  if (Val & OL_DEVICE_FP_CAPABILITY_FLAG_SOFT_FLOAT) {
+    S << " SOFT_FLOAT";
+  }
+
+  S << " }";
+}
+
+template <typename T>
+ol_result_t printPlatformValue(std::ostream &S, ol_platform_handle_t Plat,
+                               ol_platform_info_t Info, const char *Desc) {
+  S << Desc << ": ";
+
+  if constexpr (std::is_pointer_v<T>) {
+    std::vector<uint8_t> Val;
+    size_t Size;
+    OFFLOAD_ERR(olGetPlatformInfoSize(Plat, Info, &Size));
+    Val.resize(Size);
+    OFFLOAD_ERR(olGetPlatformInfo(Plat, Info, sizeof(Val), Val.data()));
+    doWrite(S, reinterpret_cast<T>(Val.data()));
+  } else {
+    T Val;
+    OFFLOAD_ERR(olGetPlatformInfo(Plat, Info, sizeof(Val), &Val));
+    doWrite(S, std::move(Val));
+  }
+  S << "\n";
+  return OL_SUCCESS;
+}
+
+template <typename T, PrintKind PK = PrintKind::NORMAL>
+ol_result_t printDeviceValue(std::ostream &S, ol_device_handle_t Dev,
+                             ol_device_info_t Info, const char *Desc,
+                             const char *Units = "") {
+  S << Desc << ": ";
+
+  if constexpr (std::is_pointer_v<T>) {
+    std::vector<uint8_t> Val;
+    size_t Size;
+    OFFLOAD_ERR(olGetDeviceInfoSize(Dev, Info, &Size));
+    Val.resize(Size);
+    OFFLOAD_ERR(olGetDeviceInfo(Dev, Info, sizeof(Val), Val.data()));
+    doWrite<T, PK>(S, reinterpret_cast<T>(Val.data()));
+  } else {
+    T Val;
+    OFFLOAD_ERR(olGetDeviceInfo(Dev, Info, sizeof(Val), &Val));
+    doWrite<T, PK>(S, std::move(Val));
+  }
+  S << Units << "\n";
+  return OL_SUCCESS;
+}
+
+ol_result_t printDevice(std::ostream &S, ol_device_handle_t D) {
+  ol_platform_handle_t Platform;
+  OFFLOAD_ERR(
+      olGetDeviceInfo(D, OL_DEVICE_INFO_PLATFORM, sizeof(Platform), &Platform));
+
+  std::vector<char> Name;
+  size_t NameSize;
+  OFFLOAD_ERR(olGetDeviceInfoSize(D, OL_DEVICE_INFO_NAME, &NameSize))
+  Name.resize(NameSize);
+  OFFLOAD_ERR(olGetDeviceInfo(D, OL_DEVICE_INFO_NAME, NameSize, Name.data()));
+  S << "[" << Name.data() << "]\n";
+
+  OFFLOAD_ERR(printPlatformValue<const char *>(
+      S, Platform, OL_PLATFORM_INFO_NAME, "Platform Name"));
+  OFFLOAD_ERR(printPlatformValue<const char *>(
+      S, Platform, OL_PLATFORM_INFO_VENDOR_NAME, "Platform Vendor Name"));
+  OFFLOAD_ERR(printPlatformValue<const char *>(
+      S, Platform, OL_PLATFORM_INFO_VERSION, "Platform Version"));
+  OFFLOAD_ERR(printPlatformValue<ol_platform_backend_t>(
+      S, Platform, OL_PLATFORM_INFO_BACKEND, "Platform Backend"));
+
+  OFFLOAD_ERR(
+      printDeviceValue<ol_device_type_t>(S, D, OL_DEVICE_INFO_TYPE, "Type"));
+  OFFLOAD_ERR(printDeviceValue<const char *>(
+      S, D, OL_DEVICE_INFO_DRIVER_VERSION, "Driver Version"));
+  OFFLOAD_ERR(printDeviceValue<uint32_t>(
+      S, D, OL_DEVICE_INFO_MAX_WORK_GROUP_SIZE, "Max Work Group Size"));
+  OFFLOAD_ERR(printDeviceValue<ol_dimensions_t>(
+      S, D, OL_DEVICE_INFO_MAX_WORK_GROUP_SIZE_PER_DIMENSION,
+      "Max Work Group Size Per Dimension"));
+  OFFLOAD_ERR(
+      printDeviceValue<uint32_t>(S, D, OL_DEVICE_INFO_VENDOR_ID, "Vendor ID"));
+  OFFLOAD_ERR(printDeviceValue<uint32_t>(S, D, OL_DEVICE_INFO_NUM_COMPUTE_UNITS,
+                                         "Num Compute Units"));
+  OFFLOAD_ERR(printDeviceValue<uint32_t>(
+      S, D, OL_DEVICE_INFO_MAX_CLOCK_FREQUENCY, "Max Clock Frequency", "MHz"));
+  OFFLOAD_ERR(printDeviceValue<uint32_t>(S, D, OL_DEVICE_INFO_MEMORY_CLOCK_RATE,
+                                         "Memory Clock Rate", "MHz"));
+  OFFLOAD_ERR(printDeviceValue<uint32_t>(S, D, OL_DEVICE_INFO_ADDRESS_BITS,
+                                         "Address Bits"));
+  OFFLOAD_ERR(printDeviceValue<uint64_t>(
+      S, D, OL_DEVICE_INFO_MAX_MEM_ALLOC_SIZE, "Max Mem Allocation Size", "B"));
+  OFFLOAD_ERR(printDeviceValue<uint64_t>(S, D, OL_DEVICE_INFO_GLOBAL_MEM_SIZE,
+                                         "Global Mem Size", "B"));
+  OFFLOAD_ERR(
+      (printDeviceValue<ol_device_fp_capability_flags_t, PrintKind::FP_FLAGS>(
+          S, D, OL_DEVICE_INFO_SINGLE_FP_CONFIG,
+          "Single Precision Floating Point Capability")));
+  OFFLOAD_ERR(
+      (printDeviceValue<ol_device_fp_capability_flags_t, PrintKind::FP_FLAGS>(
+          S, D, OL_DEVICE_INFO_DOUBLE_FP_CONFIG,
+          "Double Precision Floating Point Capability")));
+  OFFLOAD_ERR(
+      (printDeviceValue<ol_device_fp_capability_flags_t, PrintKind::FP_FLAGS>(
+          S, D, OL_DEVICE_INFO_HALF_FP_CONFIG,
+          "Half Precision Floating Point Capability")));
+  OFFLOAD_ERR(
+      printDeviceValue<uint32_t>(S, D, OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_CHAR,
+                                 "Native Vector Width For Char"));
+  OFFLOAD_ERR(
+      printDeviceValue<uint32_t>(S, D, OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_SHORT,
+                                 "Native Vector Width For Short"));
+  OFFLOAD_ERR(printDeviceValue<uint32_t>(S, D,
+                                         OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_INT,
+                                         "Native Vector Width For Int"));
+  OFFLOAD_ERR(
+      printDeviceValue<uint32_t>(S, D, OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_LONG,
+                                 "Native Vector Width For Long"));
+  OFFLOAD_ERR(
+      printDeviceValue<uint32_t>(S, D, OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_FLOAT,
+                                 "Native Vector Width For Float"));
+  OFFLOAD_ERR(printDeviceValue<uint32_t>(
+      S, D, OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_DOUBLE,
+      "Native Vector Width For Double"));
+  OFFLOAD_ERR(
+      printDeviceValue<uint32_t>(S, D, OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_HALF,
+                                 "Native Vector Width For Half"));
+
+  return OL_SUCCESS;
+}
+
+ol_result_t printRoot(std::ostream &S) {
+  OFFLOAD_ERR(olInit());
+  S << "Liboffload Version: " << OL_VERSION_MAJOR << "." << OL_VERSION_MINOR
+    << "." << OL_VERSION_PATCH << "\n";
+
+  std::vector<ol_device_handle_t> Devices;
+  OFFLOAD_ERR(olIterateDevices(
+      [](ol_device_handle_t Device, void *UserData) {
+        reinterpret_cast<decltype(Devices) *>(UserData)->push_back(Device);
+        return true;
+      },
+      &Devices));
+
+  S << "Num Devices: " << Devices.size() << "\n";
+
+  for (auto &D : Devices) {
+    S << "\n";
+    OFFLOAD_ERR(printDevice(S, D));
+  }
+
+  OFFLOAD_ERR(olShutDown());
+  return OL_SUCCESS;
+}
+
+int main(int argc, char **argv) {
+  auto Err = printRoot(std::cout);
+
+  if (Err) {
+    std::cerr << "[Liboffload error " << Err->Code << "]: " << Err->Details
+              << "\n";
+    return 1;
+  }
+  return 0;
+}

RossBrunton · 2025-08-27T14:16:41Z

Output on my system:

Liboffload Version: 0.0.1
Num Devices: 2

[gfx1100]
Platform Name: AMDGPU
Platform Vendor Name: Unknown platform vendor
Platform Version: v0.0.1
Platform Backend: AMDGPU
Type: GPU
Driver Version: 1.14
Max Work Group Size: 1024
Max Work Group Size Per Dimension: {x: 1024, y: 1024, z: 1024}
Vendor ID: 4130
Num Compute Units: 80
Max Clock Frequency: 1927MHz
Memory Clock Rate: 1124MHz
Address Bits: 64
Max Mem Allocation Size: 17163091968B
Global Mem Size: 17163091968B
Single Precision Floating Point Capability: 127 { CORRECTLY_ROUNDED_DIVIDE_SQRT ROUND_TO_NEAREST ROUND_TO_ZERO ROUND_TO_INF INF_NAN DENORM FMA }
Double Precision Floating Point Capability: 127 { CORRECTLY_ROUNDED_DIVIDE_SQRT ROUND_TO_NEAREST ROUND_TO_ZERO ROUND_TO_INF INF_NAN DENORM FMA }
Half Precision Floating Point Capability: 0 { }
Native Vector Width For Char: 1
Native Vector Width For Short: 1
Native Vector Width For Int: 1
Native Vector Width For Long: 1
Native Vector Width For Float: 1
Native Vector Width For Double: 1
Native Vector Width For Half: 0

[Virtual Host Device]
Platform Name: Host
Platform Vendor Name: Unknown platform vendor
Platform Version: v0.0.1
Platform Backend: HOST
Type: HOST
Driver Version: 22.0.0git
Max Work Group Size: 1
Max Work Group Size Per Dimension: {x: 1, y: 1, z: 1}
Vendor ID: 0
Num Compute Units: 1
Max Clock Frequency: 64MHz
Memory Clock Rate: 64MHz
Address Bits: 64
Max Mem Allocation Size: 0B
Global Mem Size: 0B
Single Precision Floating Point Capability: 127 { CORRECTLY_ROUNDED_DIVIDE_SQRT ROUND_TO_NEAREST ROUND_TO_ZERO ROUND_TO_INF INF_NAN DENORM FMA }
Double Precision Floating Point Capability: 127 { CORRECTLY_ROUNDED_DIVIDE_SQRT ROUND_TO_NEAREST ROUND_TO_ZERO ROUND_TO_INF INF_NAN DENORM FMA }
Half Precision Floating Point Capability: 0 { }
Native Vector Width For Char: 1
Native Vector Width For Short: 1
Native Vector Width For Int: 1
Native Vector Width For Long: 1
Native Vector Width For Float: 1
Native Vector Width For Double: 1
Native Vector Width For Half: 0

jhuber6

I would rather this one replaces the old one, unless there's some feature it's missing.

RossBrunton · 2025-08-27T14:39:47Z

@jhuber6 I think that makes sense, but llvm-offload-device-info prints out much more information than this does since a lot of properties haven't been mapped to liboffload device infos yet.

shiltian · 2025-08-27T16:55:15Z

I would rather this one replaces the old one, unless there's some feature it's missing.

I don't think it's a good idea to have them co-exist. It'd be super confusing.

jhuber6

Thanks, one fewer thing in offload tied to OpenMP.

offload/tools/deviceinfo/llvm-offload-device-info.cpp

This is a tool similar to llvm-offload-device-info, urinfo, etc. that simply prints properties of all devices. The difference between this and llvm-offload-device-info is that this uses liboffload rather than openmp.

jplehr · 2025-09-04T11:57:32Z

This added a few more fails to the offload staging bot: https://lab.llvm.org/staging/#/builders/64/builds/11129

Can you please take a look at that? I'm trying to get the bot clean again to move it to production.

jplehr · 2025-09-04T12:28:54Z

I opened #156872 which should fix it.

RossBrunton · 2025-09-04T13:20:27Z

@jplehr Ah, sorry about that. Thanks for fixing it.

…lvm#155626)" This reverts commit 4e8b4d6.

…lvm#155626)" breaks offload build, needs prior patches This reverts commit 4e8b4d6.

This is a tool similar to urinfo that simply prints properties of all devices. The old openMP version has been ported to liboffload.

llvmbot added the offload label Aug 27, 2025

jhuber6 reviewed Aug 27, 2025

View reviewed changes

RossBrunton mentioned this pull request Aug 28, 2025

[Offload] For AMDGPU driver, use product name #155632

Merged

RossBrunton marked this pull request as draft August 28, 2025 11:11

RossBrunton force-pushed the getInfo branch from 800988b to 852dbdd Compare August 28, 2025 14:53

RossBrunton changed the title ~~[Offload] Add new llvm-liboffload-device-info tool~~ [Offload] Port llvm-offload-device-info to new offload API Aug 28, 2025

RossBrunton force-pushed the getInfo branch from 6edb0e7 to 37d86d8 Compare August 29, 2025 08:47

RossBrunton marked this pull request as ready for review August 29, 2025 12:11

jhuber6 approved these changes Aug 29, 2025

View reviewed changes

kevinsala reviewed Aug 29, 2025

View reviewed changes

offload/tools/deviceinfo/llvm-offload-device-info.cpp Outdated Show resolved Hide resolved

kevinsala reviewed Aug 30, 2025

View reviewed changes

offload/tools/deviceinfo/llvm-offload-device-info.cpp Outdated Show resolved Hide resolved

RossBrunton added 6 commits September 2, 2025 15:49

[Offload] Add new llvm-liboffload-device-info tool

c5220a4

This is a tool similar to llvm-offload-device-info, urinfo, etc. that simply prints properties of all devices. The difference between this and llvm-offload-device-info is that this uses liboffload rather than openmp.

Add product name

7085c29

Replace binary entirely

050b56a

Fix comment

67663e5

Work size props

3c32aa1

Space before units

3def731

RossBrunton force-pushed the getInfo branch from 37d86d8 to 3def731 Compare September 2, 2025 14:49

RossBrunton merged commit 4e8b4d6 into llvm:main Sep 4, 2025
9 checks passed

searlmc1 pushed a commit to ROCm/llvm-project that referenced this pull request Sep 4, 2025

Revert "[Offload] Port llvm-offload-device-info to new offload API (l…

5b7864d

…lvm#155626)" This reverts commit 4e8b4d6.

searlmc1 pushed a commit to ROCm/llvm-project that referenced this pull request Sep 18, 2025

Revert "[Offload] Port llvm-offload-device-info to new offload API (l…

6143863

…lvm#155626)" breaks offload build, needs prior patches This reverts commit 4e8b4d6.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[Offload] Port llvm-offload-device-info to new offload API #155626

[Offload] Port llvm-offload-device-info to new offload API #155626

Uh oh!

RossBrunton commented Aug 27, 2025 •

edited

Loading

Uh oh!

llvmbot commented Aug 27, 2025

Uh oh!

RossBrunton commented Aug 27, 2025

Uh oh!

jhuber6 left a comment

Uh oh!

RossBrunton commented Aug 27, 2025

Uh oh!

shiltian commented Aug 27, 2025

Uh oh!

jhuber6 left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jplehr commented Sep 4, 2025

Uh oh!

jplehr commented Sep 4, 2025

Uh oh!

RossBrunton commented Sep 4, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

[Offload] Port llvm-offload-device-info to new offload API #155626

[Offload] Port llvm-offload-device-info to new offload API #155626

Uh oh!

Conversation

RossBrunton commented Aug 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Aug 27, 2025

Uh oh!

RossBrunton commented Aug 27, 2025

Uh oh!

jhuber6 left a comment

Choose a reason for hiding this comment

Uh oh!

RossBrunton commented Aug 27, 2025

Uh oh!

shiltian commented Aug 27, 2025

Uh oh!

jhuber6 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jplehr commented Sep 4, 2025

Uh oh!

jplehr commented Sep 4, 2025

Uh oh!

RossBrunton commented Sep 4, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

RossBrunton commented Aug 27, 2025 •

edited

Loading