[v0.30] Add support for XCDs and matrix cores

For XCDs, we dont show them if the GPU is made of a single
XCD, as it adds little value

For matrix cores, we assume it can be computed as
compute_units * simds_per_cu, it seems to work for the GPUs
I checked from CDNA3 and RDNA3. Not sure what would happen for
older GPUs that do not have matrix cores though.

src/common/gpu.hpp

@@ -46,6 +46,10 @@ struct topology_c {
 // HSA topology
 struct topology_h {
   int32_t compute_units;
+  int32_t num_shader_engines;
+  int32_t simds_per_cu;
+  int32_t num_xcc;
+  int32_t matrix_cores;
 };
 
 // Intel topology

src/common/printer.cpp

@@ -48,6 +48,8 @@ enum {
   ATTRIBUTE_FREQUENCY,     // ALL
   ATTRIBUTE_PEAK,          // ALL
   ATTRIBUTE_COMPUTE_UNITS, // HSA
+  ATTRIBUTE_MATRIX_CORES,  // HSA
+  ATTRIBUTE_XCDS,          // HSA
   ATTRIBUTE_LDS_SIZE,      // HSA
   ATTRIBUTE_STREAMINGMP,   // CUDA
   ATTRIBUTE_CORESPERMP,    // CUDA
@@ -70,6 +72,8 @@ static const AttributeField ATTRIBUTE_INFO[] = {
   { ATTRIBUTE_FREQUENCY,     "Max Frequency:",          "Max Freq.:" },
   { ATTRIBUTE_PEAK,          "Peak Performance:",       "Peak Perf.:" },
   { ATTRIBUTE_COMPUTE_UNITS, "Compute Units (CUs):",    "CUs" },
+  { ATTRIBUTE_MATRIX_CORES,  "Matrix Cores:",           "Matrix Cores:" },
+  { ATTRIBUTE_XCDS,          "XCDs:",                   "XCDs" },
   { ATTRIBUTE_LDS_SIZE,      "LDS size:",               "LDS:" },
   { ATTRIBUTE_STREAMINGMP,   "SMs:",                    "SMs:" },
   { ATTRIBUTE_CORESPERMP,    "Cores/SM:",               "Cores/SM:" },
@@ -488,6 +492,8 @@ bool print_gpufetch_amd(struct gpu_info* gpu, STYLE s, struct color** cs, struct
   char* uarch = get_str_uarch_hsa(gpu->arch);
   char* manufacturing_process = get_str_process(gpu->arch);
   char* cus = get_str_cu(gpu);
+  char* matrix_cores = get_str_matrix_cores(gpu);
+  char* xcds = get_str_xcds(gpu);
   char* max_frequency = get_str_freq(gpu);
   char* bus_width = get_str_bus_width(gpu);
   char* mem_size = get_str_memory_size(gpu);
@@ -501,6 +507,10 @@ bool print_gpufetch_amd(struct gpu_info* gpu, STYLE s, struct color** cs, struct
   setAttribute(art, ATTRIBUTE_TECHNOLOGY, manufacturing_process);
   setAttribute(art, ATTRIBUTE_FREQUENCY, max_frequency);
   setAttribute(art, ATTRIBUTE_COMPUTE_UNITS, cus);
+  setAttribute(art, ATTRIBUTE_MATRIX_CORES, matrix_cores);
+  if (xcds != NULL) {
+    setAttribute(art, ATTRIBUTE_XCDS, xcds);
+  }
   setAttribute(art, ATTRIBUTE_LDS_SIZE, lds_size);
   setAttribute(art, ATTRIBUTE_MEMORY, mem_size);
   setAttribute(art, ATTRIBUTE_BUS_WIDTH, bus_width);

src/hsa/hsa.cpp

@@ -22,10 +22,16 @@ struct agent_info {
   char vendor_name[64];
   char device_mkt_name[64];
   uint32_t max_clock_freq;
-  uint32_t compute_unit;
+  // Memory
   uint32_t bus_width;
   uint32_t lds_size;
   uint64_t global_size;
+  // Topology
+  uint32_t compute_unit;
+  uint32_t num_shader_engines;
+  uint32_t simds_per_cu;
+  uint32_t num_xcc;            // Acccelerator Complex Dies (XCDs)
+  uint32_t matrix_cores;       // Cores with WMMA/MFMA capabilities
 };
 
 #define RET_IF_HSA_ERR(err) { \
@@ -115,6 +121,15 @@ hsa_status_t agent_callback(hsa_agent_t agent, void *data) {
     err = hsa_agent_get_info(agent, (hsa_agent_info_t) HSA_AMD_AGENT_INFO_MEMORY_WIDTH, &info->bus_width);
     RET_IF_HSA_ERR(err);
 
+    err = hsa_agent_get_info(agent, (hsa_agent_info_t) HSA_AMD_AGENT_INFO_NUM_SHADER_ENGINES, &info->num_shader_engines);
+    RET_IF_HSA_ERR(err);
+
+    err = hsa_agent_get_info(agent, (hsa_agent_info_t) HSA_AMD_AGENT_INFO_NUM_SIMDS_PER_CU, &info->simds_per_cu);
+    RET_IF_HSA_ERR(err);
+
+    err = hsa_agent_get_info(agent, (hsa_agent_info_t) HSA_AMD_AGENT_INFO_NUM_XCC, &info->num_xcc);
+    RET_IF_HSA_ERR(err);
+
     // We will check against zero to see if it was set beforehand.
     info->global_size = 0;
     info->lds_size = 0;
@@ -130,6 +145,12 @@ struct topology_h* get_topology_info(struct agent_info info) {
   struct topology_h* topo = (struct topology_h*) emalloc(sizeof(struct topology_h));
 
   topo->compute_units = info.compute_unit;
+  topo->num_shader_engines = info.num_shader_engines; // not printed at the moment
+  topo->simds_per_cu = info.simds_per_cu;             // not printed at the moment
+  topo->num_xcc = info.num_xcc;
+  // Old GPUs (GCN I guess) might not have matrix cores.
+  // Not sure what would happen here?
+  topo->matrix_cores = topo->compute_units * topo->simds_per_cu;
 
   return topo;
 }
@@ -205,3 +226,17 @@ struct gpu_info* get_gpu_info_hsa(int gpu_idx) {
 char* get_str_cu(struct gpu_info* gpu) {
   return get_str_generic(gpu->topo_h->compute_units);
 }
+
+char* get_str_xcds(struct gpu_info* gpu) {
+  // If there is a single XCD, then we dont want to
+  // print it.
+  if (gpu->topo_h->num_xcc == 1) {
+    return NULL;
+  }
+  return get_str_generic(gpu->topo_h->num_xcc);
+}
+
+char* get_str_matrix_cores(struct gpu_info* gpu) {
+  // TODO: Show XX (WMMA/MFMA)
+  return get_str_generic(gpu->topo_h->matrix_cores);
+}

src/hsa/hsa.hpp

@@ -5,5 +5,7 @@
 
 struct gpu_info* get_gpu_info_hsa(int gpu_idx);
 char* get_str_cu(struct gpu_info* gpu);
+char* get_str_xcds(struct gpu_info* gpu);
+char* get_str_matrix_cores(struct gpu_info* gpu);
 
 #endif