[v0.11] Add support for printing EUs (currently only in Gen9/Gen9.5)

src/common/gpu.cpp

@@ -148,3 +148,10 @@ char* get_str_peak_performance_tensor(struct gpu_info* gpu) {
   return get_str_peak_performance_generic(gpu->peak_performance_t);
 }
 
+char* get_str_generic(int32_t data) {
+  // Largest int is 10, +1 for possible negative, +1 for EOL
+  uint32_t max_size = 12;
+  char* dummy = (char *) ecalloc(max_size, sizeof(char));
+  snprintf(dummy, max_size, "%d", data);
+  return dummy;
+}

src/common/gpu.hpp

@@ -44,6 +44,12 @@ struct topology {
   int32_t tensor_cores;
 };
 
+struct topology_i {
+  int32_t slices;
+  int32_t subslices;
+  int32_t eu_subslice;
+};
+
 struct memory {
   int64_t size_bytes;
   MEMTYPE type;
@@ -59,6 +65,7 @@ struct gpu_info {
   int64_t freq;
   struct pci* pci;
   struct topology* topo;
+  struct topology_i* topo_i;
   struct memory* mem;
   struct cache* cach;
   int64_t peak_performance;
@@ -76,5 +83,6 @@ char* get_str_memory_clock(struct gpu_info* gpu);
 char* get_str_l2(struct gpu_info* gpu);
 char* get_str_peak_performance(struct gpu_info* gpu);
 char* get_str_peak_performance_tensor(struct gpu_info* gpu);
+char* get_str_generic(int32_t data);
 
 #endif

src/common/printer.cpp

@@ -10,6 +10,7 @@
 #include "../common/gpu.hpp"
 
 #include "../intel/uarch.hpp"
+#include "../intel/intel.hpp"
 #include "../cuda/cuda.hpp"
 #include "../cuda/uarch.hpp"
 
@@ -41,6 +42,7 @@ enum {
   ATTRIBUTE_CORESPERMP,
   ATTRIBUTE_CUDA_CORES,
   ATTRIBUTE_TENSOR_CORES,
+  ATTRIBUTE_EUS,
   ATTRIBUTE_L2,
   ATTRIBUTE_MEMORY,
   ATTRIBUTE_MEMORY_FREQ,
@@ -60,6 +62,7 @@ static const char* ATTRIBUTE_FIELDS [] = {
   "Cores/SM:",
   "CUDA Cores:",
   "Tensor Cores:",
+  "Execution Units:",
   "L2 Size:",
   "Memory:",
   "Memory frequency:",
@@ -79,6 +82,7 @@ static const char* ATTRIBUTE_FIELDS_SHORT [] = {
   "Cores/SM:",
   "CUDA Cores:",
   "Tensor Cores:",
+  "EUs:",
   "L2 Size:",
   "Memory:",
   "Memory freq.:",
@@ -366,11 +370,13 @@ bool print_gpufetch_intel(struct gpu_info* gpu, STYLE s, struct color** cs, stru
   char* uarch = get_str_uarch_intel(gpu->arch);
   char* gt = get_str_gt(gpu->arch);
   char* manufacturing_process = get_str_process(gpu->arch);
+  char* eus = get_str_eu(gpu);
 
   setAttribute(art, ATTRIBUTE_NAME, gpu_name);
   setAttribute(art, ATTRIBUTE_UARCH, uarch);
   setAttribute(art, ATTRIBUTE_TECHNOLOGY, manufacturing_process);
   setAttribute(art, ATTRIBUTE_GT, gt);
+  setAttribute(art, ATTRIBUTE_EUS, eus);
 
   const char** attribute_fields = ATTRIBUTE_FIELDS;
   uint32_t longest_attribute = longest_attribute_length(art, attribute_fields);