Fixes

Similarly to NVIDIA and Intel GPUs, we now detect microarchitecture,
also with manufacturing process and specific chip name. We infer all
of this from the gfx name (in the code we use the term llvm_target),
altough it's not clear yet that this method is completely reliable (see
comments for more details). In the future we might want to replace that
with a better way. Once we have the gfx name, we *should* be able to
infer the specific chip, and from the chip we can easily infer the
microarchitecture.

This commit also includes some refactorings and code improvements on
the HSA backend.

CMakeLists.txt

@@ -10,9 +10,10 @@ set(CUDA_DIR "${SRC_DIR}/cuda")
 set(HSA_DIR "${SRC_DIR}/hsa")
 set(INTEL_DIR "${SRC_DIR}/intel")
 
-# Enable Intel backend by default
-if(NOT DEFINED ENABLE_INTEL_BACKEND)
-    set(ENABLE_INTEL_BACKEND true)
+# Make sure that at least one backend is enabled.
+# It does not make sense that the user has not specified any backend.
+if(NOT ENABLE_INTEL_BACKEND AND NOT ENABLE_CUDA_BACKEND AND NOT ENABLE_HSA_BACKEND)
+  message(FATAL_ERROR "No backend was enabled! Please enable at least one backend with -DENABLE_XXX_BACKEND")
 endif()
 
 if(ENABLE_CUDA_BACKEND)
@@ -27,8 +28,7 @@ if(ENABLE_CUDA_BACKEND)
 endif()
 
 if(ENABLE_HSA_BACKEND)
-  # TODO: Needs rocm-cmake, what if its not insalled?
-  find_package(ROCmCMakeBuildTools)
+  find_package(ROCmCMakeBuildTools QUIET)
   if (ROCmCMakeBuildTools_FOUND)
     find_package(hsa-runtime64 1.0 REQUIRED)    
     link_directories(hsa_backend hsa-runtime64::hsa-runtime64)
@@ -48,15 +48,51 @@ if(ENABLE_HSA_BACKEND)
       message(STATUS "${BoldYellow}HSA not found, disabling HSA backend${ColorReset}")
       set(ENABLE_HSA_BACKEND false)
     endif()
+  else()
+    # rocm-cmake is not installed, try to manually find neccesary files.
+    message(STATUS "${BoldYellow}Could NOT find HSA automatically, running manual search...${ColorReset}")
+    if (NOT DEFINED ROCM_PATH)
+      set(ROCM_PATH "/opt/rocm" CACHE PATH "Path to ROCm")
+    endif()
+
+    find_path(HSA_INCLUDE_DIR hsa/hsa.h HINTS ${ROCM_PATH}/include)
+    find_library(HSA_LIBRARY hsa-runtime64 HINTS ${ROCM_PATH}/lib ${ROCM_PATH}/lib64)
+
+    if (HSA_INCLUDE_DIR AND HSA_LIBRARY)
+      message(STATUS "${BoldYellow}HSA was found manually${ColorReset}")
     else()
       set(ENABLE_HSA_BACKEND false)
-    message(STATUS "${BoldYellow}ROCm not found${ColorReset}")
+      message(STATUS "${BoldYellow}HSA was not found manually${ColorReset}")
+    endif()
   endif()
 endif()
 
-list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_LIST_DIR}/cmake")
-find_package(PCIUTILS)
-if(NOT ${PCIUTILS_FOUND})
+set(GPUFECH_COMMON
+    ${COMMON_DIR}/main.cpp
+    ${COMMON_DIR}/args.cpp
+    ${COMMON_DIR}/gpu.cpp
+    ${COMMON_DIR}/global.cpp
+    ${COMMON_DIR}/printer.cpp
+    ${COMMON_DIR}/master.cpp
+    ${COMMON_DIR}/uarch.cpp
+)
+
+set(GPUFETCH_LINK_TARGETS z)
+
+if(NOT(ENABLE_HSA_BACKEND AND NOT ENABLE_CUDA_BACKEND AND NOT ENABLE_INTEL_BACKEND))
+  # Look for pciutils only if not building HSA only.
+  #
+  # This has the (intented) secondary effect that if only HSA backend is enabled
+  # by the user, but ROCm cannot be found, pciutils will still be compiled in
+  # order to show the list of GPUs available on the system, so that the user will
+  # get at least some feedback even if HSA is not found.
+  list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_LIST_DIR}/cmake")
+  list(APPEND GPUFECH_COMMON ${COMMON_DIR}/pci.cpp ${COMMON_DIR}/sort.cpp)
+  list(APPEND GPUFETCH_LINK_TARGETS pci)
+  set(CMAKE_ENABLE_PCIUTILS ON)
+
+  find_package(PCIUTILS)
+  if(NOT ${PCIUTILS_FOUND})
     message(STATUS "${BoldYellow}pciutils not found, downloading and building a local copy...${ColorReset}")
 
     # Download and build pciutils
@@ -71,18 +107,23 @@ if(NOT ${PCIUTILS_FOUND})
 
     include_directories(${PCIUTILS_INSTALL_LOCATION}/include)
     link_directories(${PCIUTILS_INSTALL_LOCATION}/lib)
-else()
+  else()
     include_directories(${PCIUTILS_INCLUDE_DIR})
     link_libraries(${PCIUTILS_LIBRARIES})
     # Needed for linking libpci in FreeBSD
     link_directories(/usr/local/lib/)
+  endif()
 endif()
 
-add_executable(gpufetch ${COMMON_DIR}/main.cpp ${COMMON_DIR}/args.cpp ${COMMON_DIR}/gpu.cpp ${COMMON_DIR}/pci.cpp ${COMMON_DIR}/sort.cpp ${COMMON_DIR}/global.cpp ${COMMON_DIR}/printer.cpp ${COMMON_DIR}/master.cpp ${COMMON_DIR}/uarch.cpp)
+add_executable(gpufetch ${GPUFECH_COMMON})
 set(SANITY_FLAGS -Wfloat-equal -Wshadow -Wpointer-arith -Wall -Wextra -pedantic -fstack-protector-all -pedantic)
 target_compile_features(gpufetch PRIVATE cxx_std_11)
 target_compile_options(gpufetch PRIVATE ${SANITY_FLAGS})
 
+if (CMAKE_ENABLE_PCIUTILS)
+  target_compile_definitions(gpufetch PUBLIC BACKEND_USE_PCI)
+endif()
+
 if(ENABLE_INTEL_BACKEND)
   target_compile_definitions(gpufetch PUBLIC BACKEND_INTEL)
 
@@ -127,20 +168,24 @@ endif()
 if(ENABLE_HSA_BACKEND)
   target_compile_definitions(gpufetch PUBLIC BACKEND_HSA)
 
-  add_library(hsa_backend STATIC ${HSA_DIR}/hsa.cpp)
+  add_library(hsa_backend STATIC ${HSA_DIR}/hsa.cpp ${HSA_DIR}/uarch.cpp)
 
   if(NOT ${PCIUTILS_FOUND})
     add_dependencies(hsa_backend pciutils)
   endif()
 
   target_include_directories(hsa_backend PRIVATE "${HSA_INCLUDE_DIR}")
-  message(STATUS "Found HSA: ${HSA_INCLUDE_DIR}")
 
+  if (HSA_LIBRARY)
+    target_link_libraries(hsa_backend PRIVATE ${HSA_LIBRARY})
+  else()
     target_link_libraries(hsa_backend PRIVATE hsa-runtime64::hsa-runtime64)
+  endif()
+
   target_link_libraries(gpufetch hsa_backend)
 endif()
 
-target_link_libraries(gpufetch pci z)
+target_link_libraries(gpufetch ${GPUFETCH_LINK_TARGETS})
 install(TARGETS gpufetch DESTINATION bin)
 
 if(NOT WIN32)

src/common/gpu.hpp

@@ -3,8 +3,6 @@
 
 #include <cstdint>
 
-#include "../cuda/pci.hpp"
-
 #define UNKNOWN_FREQ -1
 
 enum {

src/common/main.cpp

@@ -8,7 +8,11 @@
 #include "../cuda/cuda.hpp"
 #include "../cuda/uarch.hpp"
 
-static const char* VERSION = "0.25";
+#ifdef BACKEND_USE_PCI
+#include "pci.hpp"
+#endif
+
+static const char* VERSION = "0.30";
 
 void print_help(char *argv[]) {
   const char **t = args_str;
@@ -79,8 +83,12 @@ int main(int argc, char* argv[]) {
   }
 
   if(get_num_gpus_available(list) == 0) {
+#ifdef BACKEND_USE_PCI    
     printErr("No GPU was detected! Available GPUs are:");
     print_gpus_list_pci();
+#else
+    printErr("No GPU was detected!");
+#endif    
     printf("Please, make sure that the appropiate backend is enabled:\n");
     print_enabled_backends();
     printf("Visit https://github.com/Dr-Noob/gpufetch#2-backends for more information\n");

src/common/master.cpp

@@ -1,7 +1,10 @@
 #include <cstdlib>
 #include <cstdio>
 
-#include "pci.hpp"
+#ifdef BACKEND_USE_PCI
+  #include "pci.hpp"
+#endif  
+
 #include "global.hpp"
 #include "colors.hpp"
 #include "master.hpp"
@@ -19,7 +22,9 @@ struct gpu_list {
 
 struct gpu_list* get_gpu_list() {
   int idx = 0;
+#ifdef BACKEND_USE_PCI
   struct pci_dev *devices = get_pci_devices_from_pciutils();
+#endif
   struct gpu_list* list = (struct gpu_list*) malloc(sizeof(struct gpu_list));
   list->num_gpus = 0;
   list->gpus = (struct gpu_info**) malloc(sizeof(struct info*) * MAX_GPUS);
@@ -40,7 +45,7 @@ struct gpu_list* get_gpu_list() {
   bool valid = true;
 
   while(valid) {
-    list->gpus[idx] = get_gpu_info_hsa(devices, idx);
+    list->gpus[idx] = get_gpu_info_hsa(idx);
     if(list->gpus[idx] != NULL) idx++;
     else valid = false;
   }

src/common/printer.cpp

@@ -11,6 +11,7 @@
 #include "../intel/uarch.hpp"
 #include "../intel/intel.hpp"
 #include "../hsa/hsa.hpp"
+#include "../hsa/uarch.hpp"
 #include "../cuda/cuda.hpp"
 #include "../cuda/uarch.hpp"
 
@@ -490,10 +491,18 @@ bool print_gpufetch_amd(struct gpu_info* gpu, STYLE s, struct color** cs, struct
     return false;
 
   char* gpu_name = get_str_gpu_name(gpu);
+  char* gpu_chip = get_str_chip(gpu->arch);
+  char* uarch = get_str_uarch_hsa(gpu->arch);
+  char* manufacturing_process = get_str_process(gpu->arch);
   char* sms = get_str_cu(gpu);
   char* max_frequency = get_str_freq(gpu);
 
   setAttribute(art, ATTRIBUTE_NAME, gpu_name);
+  if (gpu_chip != NULL) {
+    setAttribute(art, ATTRIBUTE_CHIP, gpu_chip);
+  }
+  setAttribute(art, ATTRIBUTE_UARCH, uarch);
+  setAttribute(art, ATTRIBUTE_TECHNOLOGY, manufacturing_process);
   setAttribute(art, ATTRIBUTE_FREQUENCY, max_frequency);
   setAttribute(art, ATTRIBUTE_STREAMINGMP, sms);
 

src/common/uarch.hpp

@@ -16,6 +16,9 @@ struct uarch {
   int32_t cc_minor;
   int32_t compute_capability;
 
+  // HSA specific
+  int32_t llvm_target;
+
   // Intel specific
   int32_t gt;
   int32_t eu;

src/cuda/cuda.cpp

@@ -5,8 +5,8 @@
 
 #include "cuda.hpp"
 #include "uarch.hpp"
+#include "pci.hpp"
 #include "gpufetch_helper_cuda.hpp"
-#include "../common/pci.hpp"
 #include "../common/global.hpp"
 #include "../common/uarch.hpp"
 
@@ -33,10 +33,8 @@ int get_tensor_cores(struct uarch* arch, int sm, int major) {
   if(major == 7) {
     // TU116 does not have tensor cores!
     // https://www.anandtech.com/show/13973/nvidia-gtx-1660-ti-review-feat-evga-xc-gaming/2
-    if(arch->chip == CHIP_TU116   || arch->chip == CHIP_TU116BM ||
-       arch->chip == CHIP_TU116GL || arch->chip == CHIP_TU116M) {
+    if (is_chip_TU116(arch))
       return 0;
-    }
     return sm * 8;
   }
   else if(major == 8) return sm * 4;

src/cuda/uarch.cpp

@@ -8,6 +8,7 @@
 #include "../common/uarch.hpp"
 #include "../common/global.hpp"
 #include "../common/gpu.hpp"
+#include "pci.hpp"
 #include "chips.hpp"
 
 // Any clock multiplier
@@ -361,3 +362,8 @@ void free_uarch_struct(struct uarch* arch) {
   free(arch->chip_str);
   free(arch);
 }
+
+bool is_chip_TU116(struct uarch* arch) {
+  return arch->chip == CHIP_TU116   || arch->chip == CHIP_TU116BM ||
+         arch->chip == CHIP_TU116GL || arch->chip == CHIP_TU116M;
+}

src/cuda/uarch.hpp

@@ -13,5 +13,6 @@ char* get_str_cc(struct uarch* arch);
 char* get_str_chip(struct uarch* arch);
 char* get_str_process(struct uarch* arch);
 void free_uarch_struct(struct uarch* arch);
+bool is_chip_TU116(struct uarch* arch);
 
 #endif

src/hsa/chips.hpp

@@ -0,0 +1,37 @@
+#ifndef __HSA_GPUCHIPS__
+#define __HSA_GPUCHIPS__
+
+typedef uint32_t GPUCHIP;
+
+enum {
+  CHIP_UNKNOWN_HSA,
+  // VEGA (TODO)
+  // ...
+  // RDNA
+  CHIP_NAVI_10,
+  CHIP_NAVI_12,
+  CHIP_NAVI_14,
+  // RDNA2
+  // There are way more (eg Oberon)
+  // Maybe we'll add them in the future.
+  CHIP_NAVI_21,
+  CHIP_NAVI_22,
+  CHIP_NAVI_23,
+  CHIP_NAVI_24,
+  // RDNA3
+  // There are way more as well.
+  // Supporting Navi only for now.
+  CHIP_NAVI_31,
+  CHIP_NAVI_32,
+  CHIP_NAVI_33,
+  // RDNA4
+  CHIP_NAVI_44,
+  CHIP_NAVI_48,
+  // CDNA
+  CHIP_ARCTURUS,      // MI100 series
+  CHIP_ALDEBARAN,     // MI200 series
+  CHIP_AQUA_VANJARAM, // MI300 series
+  CHIP_CDNA_NEXT      // MI350 series
+};
+
+#endif

src/hsa/hsa.cpp

@@ -12,7 +12,7 @@
 #include <hsa/hsa_ext_amd.h>
 
 #include "hsa.hpp"
-#include "../common/pci.hpp"
+#include "uarch.hpp"
 #include "../common/global.hpp"
 #include "../common/uarch.hpp"
 
@@ -34,8 +34,7 @@ struct agent_info {
       snprintf(&(err_val[0]), sizeof(err_val), "%#x", (uint32_t)err);         \
       err_str = &(err_val[0]);                                                \
     }                                                                         \
-    printErr("HSA failure at: %s:%d\n",                              \
-                      __FILE__, __LINE__);                           \
+    printErr("HSA failure at: %s:%d\n", __FILE__, __LINE__);                  \
     printErr("Call returned %s\n", err_str);                                  \
     return (err);                                                             \
   }                                                                           \
@@ -52,7 +51,6 @@ hsa_status_t agent_callback(hsa_agent_t agent, void *data) {
     err = hsa_agent_get_info(agent, HSA_AGENT_INFO_NAME, info->gpu_name);
     RET_IF_HSA_ERR(err);
 
-    // TODO: What if vendor_name is not AMD?
     err = hsa_agent_get_info(agent, HSA_AGENT_INFO_VENDOR_NAME, info->vendor_name);
     RET_IF_HSA_ERR(err);
 
@@ -77,7 +75,7 @@ struct topology_h* get_topology_info(struct agent_info info) {
   return topo;
 }
 
-struct gpu_info* get_gpu_info_hsa(struct pci_dev *devices, int gpu_idx) {
+struct gpu_info* get_gpu_info_hsa(int gpu_idx) {
   struct gpu_info* gpu = (struct gpu_info*) emalloc(sizeof(struct gpu_info));
   gpu->pci = NULL;
   gpu->idx = gpu_idx;
@@ -92,11 +90,8 @@ struct gpu_info* get_gpu_info_hsa(struct pci_dev *devices, int gpu_idx) {
     return NULL;
   }
 
-  hsa_status_t status;
-
-  // Initialize the HSA runtime
-  status = hsa_init();
-  if (status != HSA_STATUS_SUCCESS) {
+  hsa_status_t err = hsa_init();
+  if (err != HSA_STATUS_SUCCESS) {
     printErr("Failed to initialize HSA runtime");
     return NULL;
   }
@@ -105,23 +100,35 @@ struct gpu_info* get_gpu_info_hsa(struct pci_dev *devices, int gpu_idx) {
   info.deviceId = gpu_idx;
 
   // Iterate over all agents in the system
-  status = hsa_iterate_agents(agent_callback, &info);
-  if (status != HSA_STATUS_SUCCESS) {
+  err = hsa_iterate_agents(agent_callback, &info);
+  if (err != HSA_STATUS_SUCCESS) {
     printErr("Failed to iterate HSA agents");
     hsa_shut_down();
     return NULL;
   }
 
-  gpu->freq = info.max_clock_freq;
+  if (strcmp(info.vendor_name, "AMD") != 0) {
+    printErr("HSA vendor name is: '%s'. Only AMD is supported!", info.vendor_name);
+    return NULL;
+  }
   gpu->vendor = GPU_VENDOR_AMD;
+
+  gpu->freq = info.max_clock_freq;
+  gpu->topo_h = get_topology_info(info);
   gpu->name = (char *) emalloc(sizeof(char) * (strlen(info.device_mkt_name) + 1));
   strcpy(gpu->name, info.device_mkt_name);
-  gpu->topo_h = get_topology_info(info);
+  gpu->arch = get_uarch_from_hsa(gpu, info.gpu_name);
 
-  // TODO: Use gpu_name for uarch detection
+  if (gpu->arch == NULL) {
+    return NULL;
+  }
 
   // Shut down the HSA runtime
-  hsa_shut_down();
+  err = hsa_shut_down();
+  if (err != HSA_STATUS_SUCCESS) {
+    printErr("Failed to shutdown HSA runtime");
+    return NULL;
+  }
   return gpu;
 }
 

src/hsa/hsa.hpp

@@ -3,7 +3,7 @@
 
 #include "../common/gpu.hpp"
 
-struct gpu_info* get_gpu_info_hsa(struct pci_dev *devices, int gpu_idx);
+struct gpu_info* get_gpu_info_hsa(int gpu_idx);
 char* get_str_cu(struct gpu_info* gpu);
 
 #endif

src/hsa/uarch.cpp

@@ -0,0 +1,321 @@
+#include <cstdlib>
+#include <cstdint>
+#include <cstring>
+
+#include "../common/uarch.hpp"
+#include "../common/global.hpp"
+#include "../common/gpu.hpp"
+#include "chips.hpp"
+
+// MICROARCH values
+enum {
+  UARCH_UNKNOWN,
+  // GCN (Graphics Core Next)
+  // Empty for now
+  // ...
+  // RDNA (Radeon DNA)
+  UARCH_RDNA,
+  UARCH_RDNA2,
+  UARCH_RDNA3,
+  UARCH_RDNA4,
+  // CDNA (Compute DNA)
+  UARCH_CDNA,
+  UARCH_CDNA2,
+  UARCH_CDNA3,
+  UARCH_CDNA4
+};
+
+static const char *uarch_str[] = {
+  /*[ARCH_UNKNOWN]    = */ STRING_UNKNOWN,
+  /*[UARCH_RDNA]      = */ "RDNA",
+  /*[UARCH_RDNA2]     = */ "RDNA2",
+  /*[UARCH_RDNA3]     = */ "RDNA3",
+  /*[UARCH_RDNA4]     = */ "RDNA4",
+  /*[UARCH_CDNA]      = */ "CDNA",
+  /*[UARCH_CDNA2]     = */ "CDNA2",
+  /*[UARCH_CDNA3]     = */ "CDNA3",
+  /*[UARCH_CDNA4]     = */ "CDNA4",
+};
+
+// Sources: 
+// - https://rocm.docs.amd.com/en/latest/reference/gpu-arch-specs.html
+// - https://www.techpowerup.com
+//
+// This is sometimes refered to as LLVM target, but also shader ISA.
+//
+// LLVM target *usually* maps to a specific architecture. However there
+// are case where this is not true:
+// MI8 is GCN3.0 with LLVM target gfx803
+// MI6 is GCN4.0 with LLVM target gfx803
+// or
+// Strix Point can be gfx1150 or gfx1151
+//
+// NOTE: GCN chips are stored for completeness, but they are
+// not actively supported.
+enum {
+  TARGET_UNKNOWN_HSA,
+  /// GCN (Graphics Core Next)
+  /// ------------------------
+  // GCN 1.0 
+  TARGET_GFX600,
+  TARGET_GFX601,
+  TARGET_GFX602,
+  // GCN 2.0
+  TARGET_GFX700,
+  TARGET_GFX701,
+  TARGET_GFX702,
+  TARGET_GFX703,
+  TARGET_GFX704,
+  TARGET_GFX705,
+  // GCN 3.0 / 4.0
+  TARGET_GFX801,
+  TARGET_GFX802,
+  TARGET_GFX803,
+  TARGET_GFX805,
+  TARGET_GFX810,
+  // GCN 5.0
+  TARGET_GFX900,
+  TARGET_GFX902,
+  TARGET_GFX904,
+  // GCN 5.1
+  TARGET_GFX906,
+  // ???
+  TARGET_GFX909,
+  TARGET_GFX90C,
+  /// RDNA (Radeon DNA)
+  /// -----------------
+  // RDNA1
+  TARGET_GFX1010,
+  TARGET_GFX1011,
+  TARGET_GFX1012,
+  // RDNA2
+  TARGET_GFX1013, // Oberon
+  TARGET_GFX1030,
+  TARGET_GFX1031,
+  TARGET_GFX1032,
+  TARGET_GFX1033,
+  TARGET_GFX1034,
+  TARGET_GFX1035, // ??
+  TARGET_GFX1036, // ??
+  // RDNA3
+  TARGET_GFX1100,
+  TARGET_GFX1101,
+  TARGET_GFX1102,
+  TARGET_GFX1103, // ???
+  // RDNA3.5
+  TARGET_GFX1150, // Strix Point
+  TARGET_GFX1151, // Strix Halo / Strix Point
+  TARGET_GFX1152, // Krackan Point
+  TARGET_GFX1153, // ???
+  // RDNA4
+  TARGET_GFX1200,
+  TARGET_GFX1201,
+  TARGET_GFX1250, // ???
+  TARGET_GFX1251, // ???
+  /// CDNA (Compute DNA)
+  /// ------------------
+  // CDNA
+  TARGET_GFX908,
+  // CDNA2
+  TARGET_GFX90A,
+  // CDNA3
+  TARGET_GFX942,
+  // CDNA4
+  TARGET_GFX950  
+};
+
+#define CHECK_UARCH_START if (false) {}
+#define CHECK_UARCH(arch, chip_, str, uarch, process) \
+   else if (arch->chip == chip_) fill_uarch(arch, str, uarch, process);
+#define CHECK_UARCH_END else { if(arch->chip != CHIP_UNKNOWN_HSA) printBug("map_chip_to_uarch_hsa: Unknown chip id: %d", arch->chip); fill_uarch(arch, STRING_UNKNOWN, UARCH_UNKNOWN, UNK); }
+
+void fill_uarch(struct uarch* arch, char const *str, MICROARCH u, uint32_t process) {
+  arch->chip_str = (char *) emalloc(sizeof(char) * (strlen(str)+1));
+  strcpy(arch->chip_str, str);
+  arch->uarch = u;
+  arch->process = process;
+}
+
+// On chiplet based chips (such as Navi31, Navi32, etc),
+// we have 2 different processes: The MCD process and the
+// rest of the chip process. They might be different and here
+// we just take one - let's take MCD process for now.
+//
+// TODO: Should we differentiate?
+void map_chip_to_uarch_hsa(struct uarch* arch) {
+  CHECK_UARCH_START
+
+  // RDNA
+  CHECK_UARCH(arch, CHIP_NAVI_10,  "Navi 10", UARCH_RDNA,  7)
+  CHECK_UARCH(arch, CHIP_NAVI_12,  "Navi 12", UARCH_RDNA,  7)
+  CHECK_UARCH(arch, CHIP_NAVI_14,  "Navi 14", UARCH_RDNA,  7)
+  CHECK_UARCH(arch, CHIP_NAVI_21,  "Navi 21", UARCH_RDNA2, 7)
+  CHECK_UARCH(arch, CHIP_NAVI_22,  "Navi 22", UARCH_RDNA2, 7)
+  CHECK_UARCH(arch, CHIP_NAVI_23,  "Navi 23", UARCH_RDNA2, 7)
+  CHECK_UARCH(arch, CHIP_NAVI_24,  "Navi 24", UARCH_RDNA2, 6)
+  CHECK_UARCH(arch, CHIP_NAVI_31,  "Navi 31", UARCH_RDNA3, 6)
+  CHECK_UARCH(arch, CHIP_NAVI_32,  "Navi 32", UARCH_RDNA3, 6)
+  CHECK_UARCH(arch, CHIP_NAVI_33,  "Navi 33", UARCH_RDNA3, 6)
+  CHECK_UARCH(arch, CHIP_NAVI_44,  "Navi 44", UARCH_RDNA4, 4)
+  CHECK_UARCH(arch, CHIP_NAVI_48,  "Navi 48", UARCH_RDNA4, 4)
+  // CDNA
+  // NOTE: We will not show chip name for CDNA, thus use empty str
+  CHECK_UARCH(arch, CHIP_ARCTURUS,        "", UARCH_CDNA,  7)
+  CHECK_UARCH(arch, CHIP_ALDEBARAN,       "", UARCH_CDNA2, 6)
+  CHECK_UARCH(arch, CHIP_AQUA_VANJARAM,   "", UARCH_CDNA3, 6)
+  CHECK_UARCH(arch, CHIP_CDNA_NEXT,       "", UARCH_CDNA4, 6) // big difference between MCD and rest of the chip process
+  
+  CHECK_UARCH_END
+}
+
+#define CHECK_TGT_START if (false) {}
+#define CHECK_TGT(target, llvm_target, chip) \
+  else if (target == llvm_target) return chip;
+#define CHECK_TGT_END else { printBug("LLVM target '%d' has no matching chip", target); return CHIP_UNKNOWN_HSA; }
+
+// We have at least 2 choices to infer the chip:
+//
+// - LLVM target (e.g., gfx1101 is Navi 32)
+// - PCI ID (e.g., 0x7470 is Navi 32)
+//
+// For now we will use the first approach, which seems to have
+// some issues like mentioned in the enum.
+// However PCI detection is also not perfect, since it is
+// quite hard to find PCI ids from old hardware.
+GPUCHIP get_chip_from_target_hsa(int32_t target) {
+  CHECK_TGT_START
+  /// RDNA
+  /// -------------------------------------------
+  CHECK_TGT(target, TARGET_GFX1010, CHIP_NAVI_10)
+  CHECK_TGT(target, TARGET_GFX1011, CHIP_NAVI_12)
+  CHECK_TGT(target, TARGET_GFX1012, CHIP_NAVI_14)
+  // CHECK_TGT(target, TARGET_GFX1013, TODO)
+  /// RDNA2
+  /// -------------------------------------------
+  CHECK_TGT(target, TARGET_GFX1030, CHIP_NAVI_21)
+  CHECK_TGT(target, TARGET_GFX1031, CHIP_NAVI_22)
+  CHECK_TGT(target, TARGET_GFX1032, CHIP_NAVI_23)
+  CHECK_TGT(target, TARGET_GFX1033, CHIP_NAVI_21)
+  CHECK_TGT(target, TARGET_GFX1034, CHIP_NAVI_24)
+  // CHECK_TGT(target, TARGET_GFX1035, TODO)
+  // CHECK_TGT(target, TARGET_GFX1036, TODO)
+  /// RDNA3
+  /// -------------------------------------------
+  CHECK_TGT(target, TARGET_GFX1100, CHIP_NAVI_31)
+  CHECK_TGT(target, TARGET_GFX1101, CHIP_NAVI_32)
+  CHECK_TGT(target, TARGET_GFX1102, CHIP_NAVI_33)
+  // CHECK_TGT(target, TARGET_GFX1103, TODO)
+  /// RDNA3.5
+  /// -------------------------------------------
+  // CHECK_TGT(target, TARGET_GFX1150, TODO)
+  // CHECK_TGT(target, TARGET_GFX1151, TODO)
+  // CHECK_TGT(target, TARGET_GFX1152, TODO)
+  // CHECK_TGT(target, TARGET_GFX1153, TODO)
+  /// RDNA4
+  /// -------------------------------------------
+  CHECK_TGT(target, TARGET_GFX1200, CHIP_NAVI_44)
+  CHECK_TGT(target, TARGET_GFX1201, CHIP_NAVI_48)
+  // CHECK_TGT(target, TARGET_GFX1250, TODO)
+  // CHECK_TGT(target, TARGET_GFX1251, TODO)
+  /// CDNA
+  /// -------------------------------------------
+  CHECK_TGT(target, TARGET_GFX908, CHIP_ARCTURUS)
+  /// CDNA2
+  /// -------------------------------------------
+  CHECK_TGT(target, TARGET_GFX90A, CHIP_ALDEBARAN)
+  /// CDNA3
+  /// -------------------------------------------
+  CHECK_TGT(target, TARGET_GFX942, CHIP_AQUA_VANJARAM)
+  /// CDNA4
+  /// -------------------------------------------
+  CHECK_TGT(target, TARGET_GFX950, CHIP_CDNA_NEXT)
+  CHECK_TGT_END
+}
+
+#define CHECK_TGT_STR_START if (false) {}
+#define CHECK_TGT_STR(target, llvm_target, chip) \
+  else if (strcmp(target, llvm_target) == 0) return chip;
+#define CHECK_TGT_STR_END else { return TARGET_UNKNOWN_HSA; }
+
+// Maps the LLVM target string to the enum value
+int32_t get_llvm_target_from_str(char* target) {
+  // TODO: Autogenerate this
+  // TODO: Add all, not only the ones we support in get_chip_from_target_hsa
+  CHECK_TGT_STR_START
+  CHECK_TGT_STR(target, "gfx1010", TARGET_GFX1010)
+  CHECK_TGT_STR(target, "gfx1011", TARGET_GFX1011)
+  CHECK_TGT_STR(target, "gfx1012", TARGET_GFX1012)
+  CHECK_TGT_STR(target, "gfx1013", TARGET_GFX1013)
+  CHECK_TGT_STR(target, "gfx1030", TARGET_GFX1030)
+  CHECK_TGT_STR(target, "gfx1031", TARGET_GFX1031)
+  CHECK_TGT_STR(target, "gfx1032", TARGET_GFX1032)
+  CHECK_TGT_STR(target, "gfx1033", TARGET_GFX1033)
+  CHECK_TGT_STR(target, "gfx1034", TARGET_GFX1034)
+  CHECK_TGT_STR(target, "gfx1035", TARGET_GFX1035)
+  CHECK_TGT_STR(target, "gfx1036", TARGET_GFX1036)
+  CHECK_TGT_STR(target, "gfx1100", TARGET_GFX1100)
+  CHECK_TGT_STR(target, "gfx1101", TARGET_GFX1101)
+  CHECK_TGT_STR(target, "gfx1102", TARGET_GFX1102)
+  CHECK_TGT_STR(target, "gfx1103", TARGET_GFX1103)
+  CHECK_TGT_STR(target, "gfx1200", TARGET_GFX1200)
+  CHECK_TGT_STR(target, "gfx1201", TARGET_GFX1201)
+  CHECK_TGT_STR(target, "gfx1250", TARGET_GFX1250)
+  CHECK_TGT_STR(target, "gfx1251", TARGET_GFX1251)
+  CHECK_TGT_STR(target, "gfx908",  TARGET_GFX908)
+  CHECK_TGT_STR(target, "gfx90a",  TARGET_GFX90A)
+  CHECK_TGT_STR(target, "gfx942",  TARGET_GFX942)
+  CHECK_TGT_STR(target, "gfx950",  TARGET_GFX950)
+  CHECK_TGT_STR_END
+}
+
+struct uarch* get_uarch_from_hsa(struct gpu_info* gpu, char* gpu_name) {
+  struct uarch* arch = (struct uarch*) emalloc(sizeof(struct uarch));
+
+  arch->llvm_target = get_llvm_target_from_str(gpu_name);
+  if (arch->llvm_target == TARGET_UNKNOWN_HSA) {
+    printErr("Unknown LLVM target: '%s'", gpu_name);
+    return NULL;
+  }
+
+  arch->chip_str = NULL;
+  arch->chip = get_chip_from_target_hsa(arch->llvm_target);
+  map_chip_to_uarch_hsa(arch);
+
+  return arch;
+}
+
+bool is_uarch_valid(struct uarch* arch) {
+  if (arch == NULL) {
+    printBug("Invalid uarch: arch is NULL");
+    return false;
+  }
+  if (arch->uarch >= UARCH_UNKNOWN && arch->uarch <= UARCH_CDNA4) {
+    return true;
+  }
+  else {
+    printBug("Invalid uarch: %d", arch->uarch);
+    return false;
+  }
+}
+
+bool is_cdna(struct uarch* arch) {
+  return arch->uarch == UARCH_CDNA ||
+         arch->uarch == UARCH_CDNA2 ||
+         arch->uarch == UARCH_CDNA3 ||
+         arch->uarch == UARCH_CDNA4;
+}
+
+char* get_str_chip(struct uarch* arch) {
+  // We dont want to show CDNA chip names as they add
+  // no value, since each architecture maps one to one
+  // to a chip.
+  if (is_cdna(arch)) return NULL;
+  return arch->chip_str;
+}
+
+const char* get_str_uarch_hsa(struct uarch* arch) {
+  if (!is_uarch_valid(arch)) {
+    return NULL;
+  }
+  return uarch_str[arch->uarch];
+}

src/hsa/uarch.hpp

@@ -0,0 +1,13 @@
+#ifndef __HSA_UARCH__
+#define __HSA_UARCH__
+
+#include "../common/gpu.hpp"
+
+struct uarch;
+
+struct uarch* get_uarch_from_hsa(struct gpu_info* gpu, char* gpu_name);
+char* get_str_uarch_hsa(struct uarch* arch);
+char* get_str_process(struct uarch* arch); // TODO: Shouldnt we define this in the cpp?
+char* get_str_chip(struct uarch* arch);
+
+#endif