Remove TODO

We can use hsa_amd_agent_iterate_memory_pools to fetch info about GPU
memory pools in the GPU. HSA_AMD_SEGMENT_GROUP seems to be LDS, and
HSA_AMD_SEGMENT_GLOBAL seems to be global memory.

However, the latter is reported multiple times (I don't know why). The
only solution I found for this is to check for the
HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED flag, which
seems to be reported only once.

For bus width, we simply use HSA_AMD_AGENT_INFO_MEMORY_WIDTH.

src/common/gpu.cpp

@@ -101,6 +101,17 @@ char* get_str_bus_width(struct gpu_info* gpu) {
   return string;
 }
 
+char* get_str_lds_size(struct gpu_info* gpu) {
+  // TODO: Show XX KB (XX MB Total) like in cpufetch
+  uint32_t size = 3+1+3+1;
+  assert(strlen(STRING_UNKNOWN)+1 <= size);
+  char* string = (char *) ecalloc(size, sizeof(char));
+
+  sprintf(string, "%d KB", gpu->mem->lds_size / 1024);
+
+  return string;
+}
+
 char* get_str_memory_clock(struct gpu_info* gpu) {
   return get_freq_as_str_mhz(gpu->mem->freq);
 }

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
@@ -61,6 +65,7 @@ struct memory {
   int32_t bus_width;
   int32_t freq;
   int32_t clk_mul; // clock multiplier
+  int32_t lds_size; // HSA specific for now
 };
 
 struct gpu_info {
@@ -88,6 +93,7 @@ char* get_str_freq(struct gpu_info* gpu);
 char* get_str_memory_size(struct gpu_info* gpu);
 char* get_str_memory_type(struct gpu_info* gpu);
 char* get_str_bus_width(struct gpu_info* gpu);
+char* get_str_lds_size(struct gpu_info* gpu);
 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);

src/common/printer.cpp

@@ -32,64 +32,60 @@
 #define MAX_ATTRIBUTES      100
 #define MAX_TERM_SIZE       1024
 
+typedef struct {
+  int id;
+  const char *name;
+  const char *shortname;
+} AttributeField;
+
+// AttributeField IDs
+//                         Used by
 enum {
-  ATTRIBUTE_NAME,
+  ATTRIBUTE_NAME,          // ALL
-  ATTRIBUTE_CHIP,
+  ATTRIBUTE_CHIP,          // ALL
-  ATTRIBUTE_UARCH,
+  ATTRIBUTE_UARCH,         // ALL
-  ATTRIBUTE_TECHNOLOGY,
+  ATTRIBUTE_TECHNOLOGY,    // ALL
-  ATTRIBUTE_GT,
+  ATTRIBUTE_FREQUENCY,     // ALL
-  ATTRIBUTE_FREQUENCY,
+  ATTRIBUTE_PEAK,          // ALL
-  ATTRIBUTE_STREAMINGMP,
+  ATTRIBUTE_COMPUTE_UNITS, // HSA
-  ATTRIBUTE_CORESPERMP,
+  ATTRIBUTE_MATRIX_CORES,  // HSA
-  ATTRIBUTE_CUDA_CORES,
+  ATTRIBUTE_XCDS,          // HSA
-  ATTRIBUTE_TENSOR_CORES,
+  ATTRIBUTE_LDS_SIZE,      // HSA
-  ATTRIBUTE_EUS,
+  ATTRIBUTE_STREAMINGMP,   // CUDA
-  ATTRIBUTE_L2,
+  ATTRIBUTE_CORESPERMP,    // CUDA
-  ATTRIBUTE_MEMORY,
+  ATTRIBUTE_CUDA_CORES,    // CUDA
-  ATTRIBUTE_MEMORY_FREQ,
+  ATTRIBUTE_TENSOR_CORES,  // CUDA
-  ATTRIBUTE_BUS_WIDTH,
+  ATTRIBUTE_L2,            // CUDA
-  ATTRIBUTE_PEAK,
+  ATTRIBUTE_MEMORY,        // CUDA,HSA
-  ATTRIBUTE_PEAK_TENSOR,
+  ATTRIBUTE_MEMORY_FREQ,   // CUDA
+  ATTRIBUTE_BUS_WIDTH,     // CUDA,HSA
+  ATTRIBUTE_PEAK_TENSOR,   // CUDA
+  ATTRIBUTE_EUS,           // Intel
+  ATTRIBUTE_GT,            // Intel
 };
 
-static const char* ATTRIBUTE_FIELDS [] = {
+static const AttributeField ATTRIBUTE_INFO[] = {
-  "Name:",
+  { ATTRIBUTE_NAME,          "Name:",                   "Name:" },
-  "GPU processor:",
+  { ATTRIBUTE_CHIP,          "GPU processor:",          "Processor:" },
-  "Microarchitecture:",
+  { ATTRIBUTE_UARCH,         "Microarchitecture:",      "uArch:" },
-  "Technology:",
+  { ATTRIBUTE_TECHNOLOGY,    "Technology:",             "Technology:" },
-  "Graphics Tier:",
+  { ATTRIBUTE_FREQUENCY,     "Max Frequency:",          "Max Freq.:" },
-  "Max Frequency:",
+  { ATTRIBUTE_PEAK,          "Peak Performance:",       "Peak Perf.:" },
-  "SMs:",
+  { ATTRIBUTE_COMPUTE_UNITS, "Compute Units (CUs):",    "CUs" },
-  "Cores/SM:",
+  { ATTRIBUTE_MATRIX_CORES,  "Matrix Cores: ",          "Matrix Cores:" },
-  "CUDA Cores:",
+  { ATTRIBUTE_XCDS,          "XCDs:",                   "XCDs" },
-  "Tensor Cores:",
+  { ATTRIBUTE_LDS_SIZE,      "LDS size:",               "LDS:" },
-  "Execution Units:",
+  { ATTRIBUTE_STREAMINGMP,   "SMs:",                    "SMs:" },
-  "L2 Size:",
+  { ATTRIBUTE_CORESPERMP,    "Cores/SM:",               "Cores/SM:" },
-  "Memory:",
+  { ATTRIBUTE_CUDA_CORES,    "CUDA Cores:",             "CUDA Cores:" },
-  "Memory frequency:",
+  { ATTRIBUTE_TENSOR_CORES,  "Tensor Cores:",           "Tensor Cores:" },
-  "Bus width:",
+  { ATTRIBUTE_L2,            "L2 Size:",                "L2 Size:" },
-  "Peak Performance:",
+  { ATTRIBUTE_MEMORY,        "Memory:",                 "Memory:" },
-  "Peak Performance (MMA):",
+  { ATTRIBUTE_MEMORY_FREQ,   "Memory frequency:",       "Memory freq.:" },
-};
+  { ATTRIBUTE_BUS_WIDTH,     "Bus width:",              "Bus width:" },
-
+  { ATTRIBUTE_PEAK_TENSOR,   "Peak Performance (MMA):", "Peak Perf.(MMA):" },
-static const char* ATTRIBUTE_FIELDS_SHORT [] = {
+  { ATTRIBUTE_EUS,           "Execution Units:",        "EUs:" },
-  "Name:",
+  { ATTRIBUTE_GT,            "Graphics Tier:",          "GT:" },
-  "Processor:",
-  "uArch:",
-  "Technology:",
-  "GT:",
-  "Max Freq.:",
-  "SMs:",
-  "Cores/SM:",
-  "CUDA Cores:",
-  "Tensor Cores:",
-  "EUs:",
-  "L2 Size:",
-  "Memory:",
-  "Memory freq.:",
-  "Bus width:",
-  "Peak Perf.:",
-  "Peak Perf.(MMA):",
 };
 
 struct terminal {
@@ -207,8 +203,6 @@ bool ascii_fits_screen(int termw, struct ascii_logo logo, int lf) {
 void replace_bgbyfg_color(struct ascii_logo* logo) {
   // Replace background by foreground color
   for(int i=0; i < 2; i++) {
-    if(logo->color_ascii[i] == NULL) break;
-
     if(strcmp(logo->color_ascii[i], C_BG_BLACK) == 0) strcpy(logo->color_ascii[i], C_FG_BLACK);
     else if(strcmp(logo->color_ascii[i], C_BG_RED) == 0) strcpy(logo->color_ascii[i], C_FG_RED);
     else if(strcmp(logo->color_ascii[i], C_BG_GREEN) == 0) strcpy(logo->color_ascii[i], C_FG_GREEN);
@@ -276,13 +270,14 @@ void choose_ascii_art(struct ascii* art, struct color** cs, struct terminal* ter
   }
 }
 
-uint32_t longest_attribute_length(struct ascii* art, const char** attribute_fields) {
+uint32_t longest_attribute_length(struct ascii* art, bool use_short) {
   uint32_t max = 0;
   uint64_t len = 0;
 
   for(uint32_t i=0; i < art->n_attributes_set; i++) {
     if(art->attributes[i]->value != NULL) {
-      len = strlen(attribute_fields[art->attributes[i]->type]);
+      const char* str = use_short ? ATTRIBUTE_INFO[art->attributes[i]->type].shortname : ATTRIBUTE_INFO[art->attributes[i]->type].name;
+      len = strlen(str);
       if(len > max) max = len;
     }
   }
@@ -306,7 +301,7 @@ uint32_t longest_field_length(struct ascii* art, int la) {
   return max;
 }
 
-void print_ascii_generic(struct ascii* art, uint32_t la, int32_t text_space, const char** attribute_fields) {
+void print_ascii_generic(struct ascii* art, uint32_t la, int32_t text_space, bool use_short) {
   struct ascii_logo* logo = art->art;
   int attr_to_print = 0;
   int attr_type;
@@ -350,11 +345,13 @@ void print_ascii_generic(struct ascii* art, uint32_t la, int32_t text_space, con
       attr_value = art->attributes[attr_to_print]->value;
       attr_to_print++;
 
-      space_right = 1 + (la - strlen(attribute_fields[attr_type]));
+      const char* attr_str = use_short ? ATTRIBUTE_INFO[attr_type].shortname : ATTRIBUTE_INFO[attr_type].name;
+
+      space_right = 1 + (la - strlen(attr_str));
       current_space = max(0, text_space);
 
-      printf("%s%.*s%s", logo->color_text[0], current_space, attribute_fields[attr_type], art->reset);
+      printf("%s%.*s%s", logo->color_text[0], current_space, attr_str, art->reset);
-      current_space = max(0, current_space - (int) strlen(attribute_fields[attr_type]));
+      current_space = max(0, current_space - (int) strlen(attr_str));
       printf("%*s", min(current_space, space_right), "");
       current_space = max(0, current_space - min(current_space, space_right));
       printf("%s%.*s%s", logo->color_text[1], current_space, attr_value, art->reset);
@@ -388,19 +385,19 @@ bool print_gpufetch_intel(struct gpu_info* gpu, STYLE s, struct color** cs, stru
   setAttribute(art, ATTRIBUTE_EUS, eus);
   setAttribute(art, ATTRIBUTE_PEAK, pp);
 
-  const char** attribute_fields = ATTRIBUTE_FIELDS;
+  bool use_short = false;
-  uint32_t longest_attribute = longest_attribute_length(art, attribute_fields);
+  uint32_t longest_attribute = longest_attribute_length(art, use_short);
   uint32_t longest_field = longest_field_length(art, longest_attribute);
   choose_ascii_art(art, cs, term, longest_field);
 
   if(!ascii_fits_screen(term->w, *art->art, longest_field)) {
     // Despite of choosing the smallest logo, the output does not fit
     // Choose the shorter field names and recalculate the longest attr
-    attribute_fields = ATTRIBUTE_FIELDS_SHORT;
+    use_short = true;
-    longest_attribute = longest_attribute_length(art, attribute_fields);
+    longest_attribute = longest_attribute_length(art, use_short);
   }
 
-  print_ascii_generic(art, longest_attribute, term->w - art->art->width, attribute_fields);
+  print_ascii_generic(art, longest_attribute, term->w - art->art->width, use_short);
 
   return true;
 }
@@ -457,19 +454,19 @@ bool print_gpufetch_cuda(struct gpu_info* gpu, STYLE s, struct color** cs, struc
     setAttribute(art, ATTRIBUTE_PEAK_TENSOR, pp_tensor);
   }
 
-  const char** attribute_fields = ATTRIBUTE_FIELDS;
+  bool use_short = false;
-  uint32_t longest_attribute = longest_attribute_length(art, attribute_fields);
+  uint32_t longest_attribute = longest_attribute_length(art, use_short);
   uint32_t longest_field = longest_field_length(art, longest_attribute);
   choose_ascii_art(art, cs, term, longest_field);
 
   if(!ascii_fits_screen(term->w, *art->art, longest_field)) {
     // Despite of choosing the smallest logo, the output does not fit
     // Choose the shorter field names and recalculate the longest attr
-    attribute_fields = ATTRIBUTE_FIELDS_SHORT;
+    use_short = true;
-    longest_attribute = longest_attribute_length(art, attribute_fields);
+    longest_attribute = longest_attribute_length(art, use_short);
   }
 
-  print_ascii_generic(art, longest_attribute, term->w - art->art->width, attribute_fields);
+  print_ascii_generic(art, longest_attribute, term->w - art->art->width, use_short);
 
   free(manufacturing_process);
   free(max_frequency);
@@ -494,8 +491,13 @@ bool print_gpufetch_amd(struct gpu_info* gpu, STYLE s, struct color** cs, struct
   char* gpu_chip = get_str_chip(gpu->arch);
   char* uarch = get_str_uarch_hsa(gpu->arch);
   char* manufacturing_process = get_str_process(gpu->arch);
-  char* sms = get_str_cu(gpu);
+  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);
+  char* lds_size = get_str_lds_size(gpu);
 
   setAttribute(art, ATTRIBUTE_NAME, gpu_name);
   if (gpu_chip != NULL) {
@@ -504,21 +506,28 @@ bool print_gpufetch_amd(struct gpu_info* gpu, STYLE s, struct color** cs, struct
   setAttribute(art, ATTRIBUTE_UARCH, uarch);
   setAttribute(art, ATTRIBUTE_TECHNOLOGY, manufacturing_process);
   setAttribute(art, ATTRIBUTE_FREQUENCY, max_frequency);
-  setAttribute(art, ATTRIBUTE_STREAMINGMP, sms);
+  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);
 
-  const char** attribute_fields = ATTRIBUTE_FIELDS;
+  bool use_short = false;
-  uint32_t longest_attribute = longest_attribute_length(art, attribute_fields);
+  uint32_t longest_attribute = longest_attribute_length(art, use_short);
   uint32_t longest_field = longest_field_length(art, longest_attribute);
   choose_ascii_art(art, cs, term, longest_field);
 
   if(!ascii_fits_screen(term->w, *art->art, longest_field)) {
     // Despite of choosing the smallest logo, the output does not fit
     // Choose the shorter field names and recalculate the longest attr
-    attribute_fields = ATTRIBUTE_FIELDS_SHORT;
+    use_short = true;
-    longest_attribute = longest_attribute_length(art, attribute_fields);
+    longest_attribute = longest_attribute_length(art, use_short);
   }
 
-  print_ascii_generic(art, longest_attribute, term->w - art->art->width, attribute_fields);
+  print_ascii_generic(art, longest_attribute, term->w - art->art->width, use_short);
 
   free(art->attributes);
   free(art);

src/hsa/hsa.cpp

@@ -22,7 +22,16 @@ struct agent_info {
   char vendor_name[64];
   char device_mkt_name[64];
   uint32_t max_clock_freq;
+  // 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) { \
@@ -40,6 +49,51 @@ struct agent_info {
   }                                                                           \
 }
 
+hsa_status_t memory_pool_callback(hsa_amd_memory_pool_t pool, void* data) {
+  struct agent_info* info = reinterpret_cast<struct agent_info *>(data);
+
+  hsa_amd_segment_t segment;
+  hsa_status_t err = hsa_amd_memory_pool_get_info(pool, HSA_AMD_MEMORY_POOL_INFO_SEGMENT, &segment);
+  RET_IF_HSA_ERR(err);
+
+  if (segment == HSA_AMD_SEGMENT_GROUP) {
+    // LDS memory
+    // We want to make sure that this memory pool is not repeated.
+    if (info->lds_size != 0) {
+      printErr("Found HSA_AMD_SEGMENT_GROUP twice!");
+      return HSA_STATUS_ERROR;
+    }
+    uint32_t size = 0;
+
+    err = hsa_amd_memory_pool_get_info(pool, HSA_AMD_MEMORY_POOL_INFO_SIZE, &size);
+    RET_IF_HSA_ERR(err);
+
+    info->lds_size = size;    
+  }
+  else if (segment == HSA_AMD_SEGMENT_GLOBAL) {
+    // Global memory
+    uint32_t global_flags = 0;
+    
+    err = hsa_amd_memory_pool_get_info(pool, HSA_AMD_MEMORY_POOL_INFO_GLOBAL_FLAGS, &global_flags);
+    RET_IF_HSA_ERR(err);
+
+    if (global_flags & HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED) {
+      if (info->global_size != 0) {
+        printErr("Found HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED twice!");
+        return HSA_STATUS_ERROR;
+      }
+
+      uint64_t size = 0;
+
+      err = hsa_amd_memory_pool_get_info(pool, HSA_AMD_MEMORY_POOL_INFO_SIZE, &size);
+      RET_IF_HSA_ERR(err);
+
+      info->global_size = size;
+    }    
+  }
+  return HSA_STATUS_SUCCESS;
+}
+
 hsa_status_t agent_callback(hsa_agent_t agent, void *data) {
   struct agent_info* info = reinterpret_cast<struct agent_info *>(data);
 
@@ -62,6 +116,26 @@ 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_COMPUTE_UNIT_COUNT, &info->compute_unit);
     RET_IF_HSA_ERR(err);
+
+    // According to the documentation, this is deprecated. But what should I be using then?
+    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;
+    // This will fill global_size and lds_size.
+    err = hsa_amd_agent_iterate_memory_pools(agent, memory_pool_callback, data);
+    RET_IF_HSA_ERR(err);
   }
 
   return HSA_STATUS_SUCCESS;
@@ -71,10 +145,26 @@ 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;
 }
 
+struct memory* get_memory_info(struct gpu_info* gpu, struct agent_info info) {
+  struct memory* mem = (struct memory*) emalloc(sizeof(struct memory));
+  
+  mem->bus_width = info.bus_width;
+  mem->lds_size = info.lds_size;
+  mem->size_bytes = info.global_size;
+
+  return mem;
+}
+
 struct gpu_info* get_gpu_info_hsa(int gpu_idx) {
   struct gpu_info* gpu = (struct gpu_info*) emalloc(sizeof(struct gpu_info));
   gpu->pci = NULL;
@@ -118,6 +208,7 @@ struct gpu_info* get_gpu_info_hsa(int gpu_idx) {
   gpu->name = (char *) emalloc(sizeof(char) * (strlen(info.device_mkt_name) + 1));
   strcpy(gpu->name, info.device_mkt_name);
   gpu->arch = get_uarch_from_hsa(gpu, info.gpu_name);
+  gpu->mem = get_memory_info(gpu, info);
 
   if (gpu->arch == NULL) {
     return NULL;
@@ -135,3 +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