From ab43a11ef22f200a325c31bed1dda3054cbd3fb3 Mon Sep 17 00:00:00 2001
From: Dr-Noob <peibolms@gmail.com>
Date: Tue, 10 Sep 2024 22:43:23 +0100
Subject: [PATCH] [v1.06][X86] Fix accurate-pp in hybrid architectures (fixes
 #169)

Overview of changes:
- Adds field max_pp in frequency struct to hold the max freq for peak-performance estimation.
- Instead of getting the max frequency in get_peak_performance, we get it in get_cpu_info (more natural).
- Adds fill_frequency_info_pp which fills the max_pp of the passed cpu by calling measure_frequency.

The approach is to call measure_frequency with a vector where the max frequencies are stored. Then,
the first time measure_frequency is called, the frequency is measured while running all the cores,
and the max frequency is computed per module (e.g., in the case of 2 modules, we would compute
the freq for the first and for the second module), and saved into this vector. Subsequent calls to
measure_frequency will just read the corresponding value for the vector. In other words, the frequency
is only measured once for the whole CPU.
---
 src/common/cpu.c    |  6 ++++++
 src/common/cpu.h    | 10 ++++++++++
 src/x86/cpuid.c     | 34 ++++++++++++++++++++++++++++------
 src/x86/freq/freq.c | 39 ++++++++++++++++++++++++++++++++++-----
 src/x86/freq/freq.h |  2 +-
 5 files changed, 79 insertions(+), 12 deletions(-)

diff --git a/src/common/cpu.c b/src/common/cpu.c
index 946c9905..e3f6bcb5 100644
--- a/src/common/cpu.c
+++ b/src/common/cpu.c
@@ -34,6 +34,12 @@ 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
diff --git a/src/common/cpu.h b/src/common/cpu.h
index aabcee7e..3e63e8bc 100644
--- a/src/common/cpu.h
+++ b/src/common/cpu.h
@@ -60,6 +60,11 @@ 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 {
@@ -188,6 +193,8 @@ 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
 };
@@ -200,6 +207,9 @@ 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);
diff --git a/src/x86/cpuid.c b/src/x86/cpuid.c
index dd67ee09..dfa9aa19 100644
--- a/src/x86/cpuid.c
+++ b/src/x86/cpuid.c
@@ -210,18 +210,14 @@ 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 max_freq = get_freq(ptr->freq);
+    int64_t freq = get_freq(ptr->freq);
 
-    int64_t freq;
   #ifdef __linux__
     if(accurate_pp)
-      freq = measure_frequency(ptr);
-    else
-      freq = max_freq;
+      freq = get_freq_pp(ptr->freq);
   #else
     // Silence compiler warning
     (void)(accurate_pp);
-    freq = max_freq;
   #endif
 
     //First, check we have consistent data
@@ -450,6 +446,23 @@ int32_t get_core_type(void) {
   }
 }
 
+#ifdef __linux__
+// Gets the max frequency for estimating the peak performance,
+// filling in the passed cpuInfo parameter with this information.
+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;
@@ -546,6 +559,7 @@ 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);
@@ -570,6 +584,13 @@ 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;
@@ -1005,6 +1026,7 @@ struct frequency* get_frequency_info(struct cpuInfo* cpu) {
     }
   #endif
 
+  freq->max_pp = UNKNOWN_DATA;
   return freq;
 }
 
diff --git a/src/x86/freq/freq.c b/src/x86/freq/freq.c
index 615b66ab..308b8aff 100644
--- a/src/x86/freq/freq.c
+++ b/src/x86/freq/freq.c
@@ -21,9 +21,12 @@
 #define FREQ_VECTOR_SIZE         1<<16
 
 struct freq_thread {
+  // Inputs
+  struct cpuInfo* cpu;
   bool end;
   bool measure;
-  double freq;
+  // Output
+  int32_t *max_pp;
 };
 
 double vector_average_harmonic(double* v, int len) {
@@ -48,6 +51,7 @@ 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);
@@ -76,18 +80,43 @@ void* measure_freq(void *freq_ptr) {
     sleep_ms(500);
   }
 
-  freq->freq = vector_average_harmonic(freq_vector, v);
-  printWarn("AVX2 measured freq=%f\n", freq->freq);
+  if (cpu->hybrid_flag) {
+    // We have an heterogeneous architecture. After measuring the
+    // 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;
 }
 
-int64_t measure_frequency(struct cpuInfo* cpu) {
+int32_t measure_frequency(struct cpuInfo* cpu, int32_t *max_freq_pp_vec) {
+  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*);
 
@@ -159,5 +188,5 @@ int64_t measure_frequency(struct cpuInfo* cpu) {
   }
 
   printf("\r%*c", num_spaces, ' ');
-  return freq_struct->freq;
+  return max_freq_pp_vec[0];
 }
diff --git a/src/x86/freq/freq.h b/src/x86/freq/freq.h
index de62916c..cc6f98cd 100644
--- a/src/x86/freq/freq.h
+++ b/src/x86/freq/freq.h
@@ -8,6 +8,6 @@
 #define MEASURE_TIME_SECONDS         5
 #define LOOP_ITERS           100000000
 
-int64_t measure_frequency(struct cpuInfo* cpu);
+int32_t measure_frequency(struct cpuInfo* cpu, int32_t *max_freq_pp_vec);
 
 #endif