[v0.11] Adding uarch backend for intel iGPUs

src/cuda/chips.hpp

@@ -1,5 +1,5 @@
-#ifndef __GPUCHIPS__
-#define __GPUCHIPS__
+#ifndef __CUDA_GPUCHIPS__
+#define __CUDA_GPUCHIPS__
 
 typedef uint32_t GPUCHIP;
 

src/cuda/cuda.cpp

@@ -133,7 +133,7 @@ struct gpu_info* get_gpu_info_cuda(int gpu_idx) {
   strcpy(gpu->name, deviceProp.name);
 
   struct pci_dev *devices = get_pci_devices_from_pciutils();
-  gpu->pci = get_pci_from_pciutils(devices);
+  gpu->pci = get_pci_from_pciutils(devices, PCI_VENDOR_ID_NVIDIA);
   gpu->arch = get_uarch_from_cuda(gpu);
   gpu->cach = get_cache_info(deviceProp);
   gpu->mem = get_memory_info(gpu, deviceProp);

src/cuda/pci.cpp

@@ -10,20 +10,6 @@
    else if (pci->device_id == id) return chip;
 #define CHECK_PCI_END else { printBug("TODOO"); return CHIP_UNKNOWN; }
 
-struct pci {
-  uint16_t vendor_id;
-  uint16_t device_id;
-};
-
-struct pci* get_pci_from_pciutils(struct pci_dev *devices) {
-  struct pci* pci = (struct pci*) emalloc(sizeof(struct pci));
-
-  pci->vendor_id = pciutils_get_pci_vendor_id(devices);
-  pci->device_id = pciutils_get_pci_device_id(devices);
-
-  return pci;
-}
-
 /*
  * pci ids were retrieved using https://github.com/pciutils/pciids
  * and parsed using a custom script to take only the relevant

src/cuda/pci.hpp

@@ -6,6 +6,12 @@
 #include "../common/pci.hpp"
 #include "chips.hpp"
 
+/*
+ * doc: https://wiki.osdev.org/PCI#Class_Codes
+ *      https://pci-ids.ucw.cz/read/PC
+ */
+#define PCI_VENDOR_ID_NVIDIA 0x10de
+
 struct pci;
 
 struct pci* get_pci_from_pciutils(struct pci_dev *devices);

src/cuda/uarch.cpp

@@ -3,12 +3,11 @@
 #include <stdint.h>
 #include <cstddef>
 
+#include "../common/uarch.hpp"
 #include "../common/global.hpp"
 #include "../common/gpu.hpp"
 #include "chips.hpp"
 
-typedef uint32_t MICROARCH;
-
 // Any clock multiplier
 #define CM_ANY               -1
 
@@ -43,19 +42,6 @@ static const char *uarch_str[] = {
   /*[ARCH_AMPERE]     = */ "Ampere",
 };
 
-struct uarch {
-  int32_t cc_major;
-  int32_t cc_minor;
-  int32_t compute_capability;
-
-  MICROARCH uarch;
-  GPUCHIP chip;
-
-  int32_t process;
-  char* uarch_str;
-  char* chip_str;
-};
-
 #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);