[v1.06][X86] Add Zen5 uarch

src/common/cpu.c

@@ -34,12 +34,6 @@ int64_t get_freq(struct frequency* freq) {
   return freq->max;
 }
 
-#ifdef ARCH_X86
-int64_t get_freq_pp(struct frequency* freq) {
-  return freq->max_pp;
-}
-#endif
-
 #if defined(ARCH_X86) || defined(ARCH_PPC)
 char* get_str_cpu_name(struct cpuInfo* cpu, bool fcpuname) {
   #ifdef ARCH_X86

src/common/cpu.h

@@ -60,11 +60,6 @@ struct frequency {
   int32_t max;
   // Indicates if max frequency was measured
   bool measured;
-#ifdef ARCH_X86
-  // Max frequency when running vectorized code.
-  // Used only for peak performance computation.
-  int32_t max_pp;
-#endif
 };
 
 struct hypervisor {
@@ -193,8 +188,6 @@ struct cpuInfo {
 #ifdef ARCH_X86
   // The index of the first core in the module
   uint32_t first_core_id;
-  // The index of this module
-  uint32_t module_id;
 #endif
 #endif
 };
@@ -207,9 +200,6 @@ uint32_t get_nsockets(struct topology* topo);
 
 VENDOR get_cpu_vendor(struct cpuInfo* cpu);
 int64_t get_freq(struct frequency* freq);
-#ifdef ARCH_X86
-int64_t get_freq_pp(struct frequency* freq);
-#endif
 
 char* get_str_aes(struct cpuInfo* cpu);
 char* get_str_sha(struct cpuInfo* cpu);

src/x86/cpuid.c

@@ -210,14 +210,18 @@ int64_t get_peak_performance(struct cpuInfo* cpu, bool accurate_pp) {
 
   for(int i=0; i < cpu->num_cpus; ptr = ptr->next_cpu, i++) {
     struct topology* topo = ptr->topo;
-    int64_t freq = get_freq(ptr->freq);
+    int64_t max_freq = get_freq(ptr->freq);
 
+    int64_t freq;
   #ifdef __linux__
     if(accurate_pp)
-      freq = get_freq_pp(ptr->freq);
+      freq = measure_frequency(ptr);
+    else
+      freq = max_freq;
   #else
     // Silence compiler warning
     (void)(accurate_pp);
+    freq = max_freq;
   #endif
 
     //First, check we have consistent data
@@ -446,23 +450,6 @@ int32_t get_core_type(void) {
   }
 }
 
-#ifdef __linux__
-// Gets the max frequency for estimating the peak performance
-// and fills in the passed cpuInfo parameter.
-void fill_frequency_info_pp(struct cpuInfo* cpu) {
-  int32_t unused;
-  int32_t *max_freq_pp_vec = malloc(sizeof(int32_t) * cpu->num_cpus);
-  struct cpuInfo* ptr = cpu;
-
-  for (uint32_t i=0; i < cpu->num_cpus; i++) {
-    set_cpu_module(i, cpu->num_cpus, &unused);
-
-    ptr->freq->max_pp = measure_frequency(ptr, max_freq_pp_vec);
-    ptr = ptr->next_cpu;
-  }
-}
-#endif
-
 struct cpuInfo* get_cpu_info(void) {
   struct cpuInfo* cpu = emalloc(sizeof(struct cpuInfo));
   cpu->peak_performance = -1;
@@ -559,7 +546,6 @@ struct cpuInfo* get_cpu_info(void) {
       ptr->core_type = get_core_type();
     }
     ptr->first_core_id = first_core;
-    ptr->module_id = i;
     ptr->feat = get_features_info(ptr);
 
     ptr->arch = get_cpu_uarch(ptr);
@@ -584,13 +570,6 @@ struct cpuInfo* get_cpu_info(void) {
     if(ptr->topo == NULL) return cpu;
   }
 
-#ifdef __linux__
-  // If accurate_pp is requested, we need to get the max frequency
-  // after fetching the topology for all CPU modules, since the topology
-  // is required by fill_frequency_info_pp
-  if (accurate_pp()) fill_frequency_info_pp(cpu);
-#endif
-
   cpu->peak_performance = get_peak_performance(cpu, accurate_pp());
 
   return cpu;
@@ -1026,7 +1005,6 @@ struct frequency* get_frequency_info(struct cpuInfo* cpu) {
     }
   #endif
 
-  freq->max_pp = UNKNOWN_DATA;
   return freq;
 }
 

src/x86/freq/freq.c

@@ -21,12 +21,9 @@
 #define FREQ_VECTOR_SIZE         1<<16
 
 struct freq_thread {
-  // Inputs
-  struct cpuInfo* cpu;
   bool end;
   bool measure;
-  // Output
+  double freq;
-  int32_t *max_pp;
 };
 
 double vector_average_harmonic(double* v, int len) {
@@ -51,7 +48,6 @@ void* measure_freq(void *freq_ptr) {
   char* line = NULL;
   size_t len = 0;
   ssize_t read;
-  struct cpuInfo* cpu = freq->cpu;
 
   int v = 0;
   double* freq_vector = malloc(sizeof(double) * FREQ_VECTOR_SIZE);
@@ -80,43 +76,18 @@ void* measure_freq(void *freq_ptr) {
     sleep_ms(500);
   }
 
-  if (cpu->hybrid_flag) {
+  freq->freq = vector_average_harmonic(freq_vector, v);
-    // We have an heterogeneous architecture. After measuring the
+  printWarn("AVX2 measured freq=%f\n", freq->freq);
-    // frequency for all cores, we now need to compute the average
-    // independently for each CPU module.
-    struct cpuInfo* ptr = cpu;
-    double* freq_vector_ptr = freq_vector;
-
-    for (int i=0; i < cpu->num_cpus; ptr = ptr->next_cpu, i++) {
-      freq->max_pp[i] = vector_average_harmonic(freq_vector_ptr, ptr->topo->total_cores_module);
-      printWarn("AVX2 measured freq=%d (module %d)", freq->max_pp[i], i);
-
-      freq_vector_ptr = freq_vector_ptr + ptr->topo->total_cores_module;
-    }
-  }
-  else {
-    freq->max_pp[0] = vector_average_harmonic(freq_vector, v);
-    printWarn("AVX2 measured freq=%d\n", freq->max_pp[0]);
-  }
 
   return NULL;
 }
 
-int32_t measure_frequency(struct cpuInfo* cpu, int32_t *max_freq_pp_vec) {
+int64_t measure_frequency(struct cpuInfo* cpu) {
-  if (cpu->hybrid_flag && cpu->module_id > 0) {
-    // We have a hybrid architecture and we have already
-    // measured the frequency for this module in a previous
-    // call to this function, so now just return it.
-    return max_freq_pp_vec[cpu->module_id];
-  }
-
   int ret;
   int num_spaces;
   struct freq_thread* freq_struct = malloc(sizeof(struct freq_thread));
   freq_struct->end = false;
   freq_struct->measure = false;
-  freq_struct->cpu = cpu;
-  freq_struct->max_pp = max_freq_pp_vec;
 
   void* (*compute_function)(void*);
 
@@ -188,5 +159,5 @@ int32_t measure_frequency(struct cpuInfo* cpu, int32_t *max_freq_pp_vec) {
   }
 
   printf("\r%*c", num_spaces, ' ');
-  return max_freq_pp_vec[0];
+  return freq_struct->freq;
 }

src/x86/freq/freq.h

@@ -8,6 +8,6 @@
 #define MEASURE_TIME_SECONDS         5
 #define LOOP_ITERS           100000000
 
-int32_t measure_frequency(struct cpuInfo* cpu, int32_t *max_freq_pp_vec);
+int64_t measure_frequency(struct cpuInfo* cpu);
 
 #endif

src/x86/uarch.c

@@ -119,7 +119,9 @@ enum {
   UARCH_ZEN3,
   UARCH_ZEN3_PLUS,
   UARCH_ZEN4,
-  UARCH_ZEN4C
+  UARCH_ZEN4C,
+  UARCH_ZEN5,
+  UARCH_ZEN5C,
 };
 
 struct uarch {
@@ -410,6 +412,12 @@ struct uarch* get_uarch_from_cpuid_amd(uint32_t ef, uint32_t f, uint32_t em, uin
   CHECK_UARCH(arch, 10, 15,  8, NA, NA, "Zen 4",       UARCH_ZEN4,         5) // instlatx64 (AMD MI300C)
   CHECK_UARCH(arch, 10, 15,  9, NA, NA, "Zen 4",       UARCH_ZEN4,         5) // instlatx64 (AMD MI300A)
   CHECK_UARCH(arch, 10, 15, 10, NA, NA, "Zen 4c",      UARCH_ZEN4C,        5) // instlatx64
+  CHECK_UARCH(arch, 11, 15,  0, NA, NA, "Zen 5",       UARCH_ZEN5,         4) // Turin/EPYC (instlatx64)
+  CHECK_UARCH(arch, 11, 15,  1, NA, NA, "Zen 5c",      UARCH_ZEN5C,        3) // Zen5c EPYC (instlatx64, https://en.wikipedia.org/wiki/Zen_5#cite_note-10)
+  CHECK_UARCH(arch, 11, 15,  2, NA, NA, "Zen 5",       UARCH_ZEN5,         4) // Strix Point (instlatx64)
+  CHECK_UARCH(arch, 11, 15,  4, NA, NA, "Zen 5",       UARCH_ZEN5,         4) // Granite Ridge (instlatx64)
+  CHECK_UARCH(arch, 11, 15,  6, NA, NA, "Zen 5",       UARCH_ZEN5,         4) // Krackan Point (instlatx64)
+  CHECK_UARCH(arch, 11, 15,  7, NA, NA, "Zen 5",       UARCH_ZEN5,         4) // Strix Halo (instlatx64)
   UARCH_END
 
   return arch;
@@ -552,6 +560,8 @@ char* infer_cpu_name_from_uarch(struct uarch* arch) {
 }
 
 bool vpus_are_AVX512(struct cpuInfo* cpu) {
+  // Zen5 actually has 2 x AVX512 units
+  // https://www.anandtech.com/show/21469/amd-details-ryzen-ai-300-series-for-mobile-strix-point-with-rdna-35-igpu-xdna-2-npu
   return cpu->arch->uarch != UARCH_ICE_LAKE &&
          cpu->arch->uarch != UARCH_TIGER_LAKE &&
          cpu->arch->uarch != UARCH_ZEN4 &&
@@ -592,6 +602,8 @@ int get_number_of_vpus(struct cpuInfo* cpu) {
       case UARCH_ZEN3_PLUS:
       case UARCH_ZEN4:
       case UARCH_ZEN4C:
+      case UARCH_ZEN5:
+      case UARCH_ZEN5C:
         return 2;
       default:
         return 1;