cpu_pnp_strategy changes

Author: savesanketsw

common/arg.cpp

@@ -564,6 +564,26 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.cpuparams.priority = (enum ggml_sched_priority) prio;
         }
     ));
+    add_opt(common_arg(
+       { "-CPnP", "--cpu-pnp-strategy" }, "N",
+        string_format("set CPU PnP strategy : 0-disabled, 1-efficiency (default: %d)\n", params.cpuparams.cpu_pnp_strategy),
+        [](common_params& params, int strategy) {
+           if (strategy < 0 || strategy > 1) {
+                throw std::invalid_argument("invalid value");
+            }
+            params.cpuparams.cpu_pnp_strategy = (enum ggml_cpu_pnp_strategy)strategy;
+        }
+    ));
+    add_opt(common_arg(
+       { "-CPnPb", "--cpu-pnp-strategy-batch" }, "N",
+        string_format("set CPU PnP strategy batch  : 0-disabled, 1-efficiency (default: %d)\n", params.cpuparams.cpu_pnp_strategy),
+        [](common_params& params, int strategy) {
+           if (strategy < 0 || strategy > 1) {
+                throw std::invalid_argument("invalid value");
+            }
+            params.cpuparams_batch.cpu_pnp_strategy = (enum ggml_cpu_pnp_strategy)strategy;
+        }
+    ));
     add_opt(common_arg(
         {"--poll"}, "<0...100>",
         string_format("use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll),

common/common.cpp

@@ -97,6 +97,77 @@ using json = nlohmann::ordered_json;
 // CPU utils
 //
 
+#if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
+
+// Print CPU Information
+void print_cpu_info(const cpu_info& info) {
+    LOG_INF("CPU Information:\n");
+    LOG_INF("----------------\n");
+    LOG_INF("Is Hybrid Architecture: %s\n", info.is_hybrid ? "Yes" : "No");
+    LOG_INF("Number of Logical Cores: %d\n", info.num_logical_cores);
+    LOG_INF("Number of Physical Cores: %d\n", info.num_physical_cores);
+    LOG_INF("Number of Performance Cores (P-Cores): %d\n", info.num_p_cores);
+    LOG_INF("Number of Efficient Cores (E-Cores): %d\n", info.num_e_cores);
+    LOG_INF("\nE-Core Affinity Mask:\n");
+    LOG_INF("%s\n", info.e_core_affinity_mask.to_string().c_str());
+    LOG_INF("\nP-Core Affinity Mask:\n");
+    LOG_INF("%s\n", info.p_core_affinity_mask.to_string().c_str());
+}
+
+// Populate CPU Information
+int get_cpu_info(cpu_info& c_info) {
+    DWORD buffer_size = 0;
+    
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &buffer_size)) {
+        if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
+            return 0;
+        }
+    }
+
+    std::vector<char> buffer(buffer_size);
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data()), &buffer_size)) {
+        return 0;
+    }
+
+    c_info.num_physical_cores = 0;
+    c_info.num_logical_cores = 0;
+    c_info.num_e_cores = 0;
+    c_info.num_p_cores = 0;
+
+    PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data());
+    while (buffer_size > 0) {
+        if (info->Relationship == RelationProcessorCore) {
+            c_info.num_physical_cores++;
+            for (int i = 0; i < info->Processor.GroupCount; ++i) {
+                GROUP_AFFINITY *groupAffinity = &info->Processor.GroupMask[i];
+                WORD groupNumber = groupAffinity->Group;
+                KAFFINITY mask = groupAffinity->Mask;
+                int baseIndex = groupNumber * 64;
+                c_info.num_logical_cores += __popcnt64(mask);
+                if (info->Processor.EfficiencyClass < 1) {
+                    c_info.e_core_affinity_mask |= (std::bitset<GGML_MAX_N_THREADS>(mask) << baseIndex);
+                    c_info.num_e_cores += __popcnt64(mask);
+                } else {
+                    c_info.p_core_affinity_mask |= (std::bitset<GGML_MAX_N_THREADS>(mask) << baseIndex);
+                    c_info.num_p_cores += __popcnt64(mask);
+                }
+            }
+        }
+
+        buffer_size -= info->Size;
+        info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(reinterpret_cast<char*>(info) + info->Size);
+    }
+
+    if (c_info.num_p_cores > 0 && c_info.num_e_cores > 0)
+        c_info.is_hybrid = true;
+
+    return 1;
+}
+
+#endif
+
+
+
 int32_t cpu_get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -131,29 +202,12 @@ int32_t cpu_get_num_physical_cores() {
     unsigned int n_threads_win = std::thread::hardware_concurrency();
     unsigned int default_threads = n_threads_win > 0 ? (n_threads_win <= 4 ? n_threads_win : n_threads_win / 2) : 4;
 
-    DWORD buffer_size = 0;
-    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &buffer_size)) {
-        if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
-            return default_threads;
-        }
-    }
-
-    std::vector<char> buffer(buffer_size);
-    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data()), &buffer_size)) {
+    cpu_info info;
+    if(!get_cpu_info(info))
         return default_threads;
-    }
+     else 
+        return info.num_physical_cores > 0 ? info.num_physical_cores : default_threads;
 
-    int32_t num_physical_cores = 0;
-    PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data());
-    while (buffer_size > 0) {
-        if (info->Relationship == RelationProcessorCore) {
-            num_physical_cores += info->Processor.GroupCount;
-        }
-        buffer_size -= info->Size;
-        info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(reinterpret_cast<char*>(info) + info->Size);
-    }
-
-    return num_physical_cores > 0 ? num_physical_cores : default_threads;
 #endif
     unsigned int n_threads = std::thread::hardware_concurrency();
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
@@ -291,12 +345,36 @@ bool set_process_priority(enum ggml_sched_priority prio) {
 void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) {
     int32_t n_set = 0;
 
-    if (cpuparams.n_threads < 0) {
+    LOG_INF("n_threads: %d, cpu pnp strategy: %d\n", cpuparams.n_threads, cpuparams.cpu_pnp_strategy);
+
+    if (cpuparams.n_threads < 0 || cpuparams.cpu_pnp_strategy > 0) {
         // Assuming everything about cpuparams is invalid
-        if (role_model != nullptr) {
+        if (role_model != nullptr && cpuparams.cpu_pnp_strategy == 0) {
             cpuparams = *role_model;
         } else {
-            cpuparams.n_threads = cpu_get_num_math();
+            if(cpuparams.n_threads < 0)
+                cpuparams.n_threads = cpu_get_num_math();
+            
+            #if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
+            
+            if(cpuparams.cpu_pnp_strategy == GGML_CPU_PNP_STRATEGY_EFFICIENCY) {
+                cpu_info info;
+                if(get_cpu_info(info)){
+                    print_cpu_info(info);
+                    if(info.is_hybrid){
+                        LOG_INF("hybrid platform detected: applying strategy\n");
+                        if (cpuparams.n_threads > info.num_e_cores) {
+                            LOG_INF("overriding num threads: %d to num efficient cores %d\n", cpuparams.n_threads, info.num_e_cores);
+                            cpuparams.n_threads = info.num_e_cores;
+                        }
+                        for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) {
+                            cpuparams.cpumask[i] = info.e_core_affinity_mask[i];
+                        }
+                        cpuparams.mask_valid = true;
+                    }
+                }
+            }
+            #endif
         }
     }
 

common/common.h

@@ -7,6 +7,7 @@
 #include <string>
 #include <vector>
 #include <sstream>
+#include <bitset>
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -45,11 +46,22 @@ struct common_control_vector_load_info;
 // CPU utils
 //
 
+struct cpu_info {
+    bool is_hybrid  =  false;
+    int num_logical_cores = 0;
+    int num_physical_cores = 0;
+    int num_p_cores = 0;
+    int num_e_cores = 0;
+    std::bitset<GGML_MAX_N_THREADS> e_core_affinity_mask;
+    std::bitset<GGML_MAX_N_THREADS> p_core_affinity_mask;
+};
+
 struct cpu_params {
     int      n_threads                   = -1;
     bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
     bool     mask_valid                  = false;   // Default: any CPU
     enum ggml_sched_priority  priority   = GGML_SCHED_PRIO_NORMAL;  // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
+    enum ggml_cpu_pnp_strategy cpu_pnp_strategy = GGML_CPU_PNP_STRATEGY_DISABLED; // CPU power and performance strategy
     bool     strict_cpu                  = false;   // Use strict CPU placement
     uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
 };

ggml/include/ggml.h

@@ -2169,6 +2169,11 @@ extern "C" {
         GGML_SCHED_PRIO_REALTIME
     };
 
+    enum ggml_cpu_pnp_strategy {
+        GGML_CPU_PNP_STRATEGY_DISABLED,
+        GGML_CPU_PNP_STRATEGY_EFFICIENCY
+    };
+
     // threadpool params
     // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
     struct ggml_threadpool_params {

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1318,10 +1318,11 @@ struct ggml_threadpool {
 struct ggml_compute_state {
 #ifndef GGML_USE_OPENMP
     ggml_thread_t thrd;
-    bool cpumask[GGML_MAX_N_THREADS];
     int  last_graph;
     bool pending;
 #endif
+    bool cpumask[GGML_MAX_N_THREADS];
+    bool mask_valid;
     struct ggml_threadpool * threadpool;
     int ith;
 };
@@ -14044,10 +14045,20 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
     const size_t workers_size = sizeof(struct ggml_compute_state) * tpp->n_threads;
     struct ggml_compute_state * workers = ggml_aligned_malloc(workers_size);
 
+    // Check if cpu mask is valid
+    bool cpumask_valid = false;
+    for (int i = 0;  i < GGML_MAX_N_THREADS; i++) {
+        if (tpp->cpumask[i]) {
+            cpumask_valid = true;
+            break;
+        }
+    }
+
     memset(workers, 0, workers_size);
     for (int j = 0; j < tpp->n_threads; j++) {
         workers[j].threadpool = threadpool;
         workers[j].ith        = j;
+        workers[j].mask_valid = cpumask_valid; // set mask_valid for worker threads use affinity
     }
 
     threadpool->workers = workers;
@@ -14079,6 +14090,12 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
     }
 #endif // GGML_USE_OPENMP
 
+    int32_t cpumask_iter = 0;
+    for (int j = 1; j < tpp->n_threads; j++) {
+        ggml_thread_cpumask_next(tpp->cpumask, workers[j].cpumask, tpp->strict_cpu, &cpumask_iter);
+    }
+    ggml_thread_cpumask_next(tpp->cpumask, workers[0].cpumask, tpp->strict_cpu, &cpumask_iter);
+
     return threadpool;
 }
 
@@ -14125,10 +14142,19 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
                 atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
             }
 
+            // If mask is valid for worker thread apply affinity
+            if(&threadpool->workers[omp_get_thread_num()].mask_valid)
+               ggml_thread_apply_affinity(&threadpool->workers[omp_get_thread_num()].cpumask);
+
             ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]);
         }
     } else {
         atomic_store_explicit(&threadpool->n_threads_cur, 1, memory_order_relaxed);
+        
+        // If mask is valid for main thread apply affinity
+        if(&threadpool->workers[omp_get_thread_num()].mask_valid)
+               ggml_thread_apply_affinity(&threadpool->workers[omp_get_thread_num()].cpumask);
+        
         ggml_graph_compute_thread(&threadpool->workers[0]);
     }
 #else