SWDEV-463401: amdsmi_get_gpu_asic_info() adds num_of_compute_units

number of compute units `amdgpu_gpu_info.num_of_compute_units` is exposed through amdsmi_get_gpu_asic_info().

Code changes related to the following:
  * API
  * CLI
  * Unit tests
  * Examples

Change-Id: Ibeb612d079ed87437a0e56124b8504098fc2dcfd
Signed-off-by: Oliveira, Daniel <daniel.oliveira@amd.com>

example/amd_smi_drm_example.cc

@@ -301,7 +301,8 @@ int main() {
             printf("\tDeviceID: 0x%lx\n", asic_info.device_id);
             printf("\tVendorID: 0x%x\n", asic_info.vendor_id);
             printf("\tRevisionID: 0x%x\n", asic_info.rev_id);
-            printf("\tAsic serial: 0x%s\n\n", asic_info.asic_serial);
+            printf("\tAsic serial: 0x%s\n", asic_info.asic_serial);
+            printf("\tNum of Computes: %d\n\n", asic_info.num_of_compute_units);
 
             // Get VRAM info
             amdsmi_vram_info_t vram_info = {};

example/amd_smi_nodrm_example.cc

@@ -151,7 +151,8 @@ int main() {
             printf("\tVendorID: 0x%x\n", asic_info.vendor_id);
             printf("\tRevisionID: 0x%x\n", asic_info.rev_id);
             printf("\tAsic serial: 0x%s\n", asic_info.asic_serial);
-            printf("\tOAM id: 0x%x\n\n", asic_info.oam_id);
+            printf("\tOAM id: 0x%x\n", asic_info.oam_id);
+            printf("\tNum of Computes: %d\n\n", asic_info.num_of_compute_units);
 
             // Get VBIOS info
             amdsmi_vbios_info_t vbios_info = {};

include/amd_smi/amdsmi.h

@@ -588,7 +588,8 @@ typedef struct {
   uint32_t rev_id;
   char asic_serial[AMDSMI_NORMAL_STRING_LENGTH];
   uint32_t oam_id;   //< 0xFFFF if not supported
-  uint32_t reserved[18];
+  uint32_t num_of_compute_units;   //< 0xFFFFFFFF if not supported
+  uint32_t reserved[17];
 } amdsmi_asic_info_t;
 
 typedef enum {

py-interface/amdsmi_interface.py

@@ -1643,7 +1643,8 @@ def amdsmi_get_gpu_asic_info(
         "device_id": asic_info_struct.device_id,
         "rev_id": _padHexValue(hex(asic_info_struct.rev_id), 2),
         "asic_serial": asic_info_struct.asic_serial.decode("utf-8"),
-        "oam_id": asic_info_struct.oam_id
+        "oam_id": asic_info_struct.oam_id,
+        "num_compute_units": asic_info_struct.num_of_compute_units
     }
 
     string_values = ["market_name", "vendor_name"]
@@ -1670,6 +1671,10 @@ def amdsmi_get_gpu_asic_info(
     if asic_info["oam_id"] == 0xFFFF: # uint 16 max
         asic_info["oam_id"] = "N/A"
 
+    # Check for max value as a sign for not applicable
+    if asic_info["num_compute_units"] == 0xFFFFFFFF: # uint 32 max
+        asic_info["num_compute_units"] = "N/A"
+
     # Remove commas from vendor name for clean output
     asic_info["vendor_name"] = asic_info["vendor_name"].replace(',', '')
 

py-interface/amdsmi_wrapper.py

@@ -901,7 +901,8 @@ struct_amdsmi_asic_info_t._fields_ = [
     ('rev_id', ctypes.c_uint32),
     ('asic_serial', ctypes.c_char * 32),
     ('oam_id', ctypes.c_uint32),
-    ('reserved', ctypes.c_uint32 * 18),
+    ('num_of_compute_units', ctypes.c_uint32),
+    ('reserved', ctypes.c_uint32 * 17),
 ]
 
 amdsmi_asic_info_t = struct_amdsmi_asic_info_t

rocm_smi/include/rocm_smi/rocm_smi_kfd.h

@@ -90,6 +90,10 @@ class KFDNode {
     // Get gfx target version from kfd
     int get_gfx_target_version(uint64_t* gfx_target_version);
 
+    // Get simd_per_cu from kfd
+    int32_t get_simd_per_cu(uint64_t* simd_per_cu) const;
+    int32_t get_simd_count(uint64_t* simd_count) const;
+
  private:
     uint32_t node_indx_;
     uint32_t amdgpu_dev_index_;

rocm_smi/include/rocm_smi/rocm_smi_utils.h

@@ -126,8 +126,8 @@ std::string print_rsmi_od_volt_freq_data_t(rsmi_od_volt_freq_data_t *odv);
 std::string print_rsmi_od_volt_freq_regions(uint32_t num_regions,
                                             rsmi_freq_volt_region_t *regions);
 bool is_sudo_user();
-rsmi_status_t rsmi_get_gfx_target_version(uint32_t dv_ind,
-  std::string *gfx_version);
+rsmi_status_t rsmi_get_gfx_target_version(uint32_t dv_ind, std::string *gfx_version);
+rsmi_status_t rsmi_dev_number_of_computes_get(uint32_t dv_ind, uint32_t* num_computes);
 
 std::string leftTrim(const std::string &s);
 std::string rightTrim(const std::string &s);

rocm_smi/src/rocm_smi_kfd.cc

@@ -1077,5 +1077,32 @@ int KFDNode::get_gfx_target_version(uint64_t *gfx_target_version) {
   return ret;
 }
 
+int32_t KFDNode::get_simd_per_cu(uint64_t* simd_per_cu) const
+{
+    const std::string properties_path("/sys/class/kfd/kfd/topology/nodes/" +
+                                      std::to_string(this->node_indx_) +
+                                      "/properties");
+
+    auto tmp_simd_per_cu = uint64_t(0);
+    auto ret = read_node_properties(this->node_indx_, "simd_per_cu",
+                                    &tmp_simd_per_cu);
+    *simd_per_cu = tmp_simd_per_cu;
+    return ret;
+}
+
+int32_t KFDNode::get_simd_count(uint64_t* simd_count) const
+{
+    const std::string properties_path("/sys/class/kfd/kfd/topology/nodes/" +
+                                      std::to_string(this->node_indx_) +
+                                      "/properties");
+
+    auto tmp_simd_count = uint64_t(0);
+    auto ret = read_node_properties(this->node_indx_, "simd_count",
+                                    &tmp_simd_count);
+    *simd_count = tmp_simd_count;
+    return ret;
+}
+
+
 }  // namespace smi
 }  // namespace amd

rocm_smi/src/rocm_smi_utils.cc

@@ -65,6 +65,7 @@
 #include <vector>
 
 #include "rocm_smi/rocm_smi.h"
+#include "rocm_smi/rocm_smi_kfd.h"
 #include "rocm_smi/rocm_smi_utils.h"
 #include "rocm_smi/rocm_smi_exception.h"
 #include "rocm_smi/rocm_smi_main.h"
@@ -1198,6 +1199,25 @@ rsmi_status_t rsmi_get_gfx_target_version(uint32_t dv_ind,
   }
 }
 
+rsmi_status_t rsmi_dev_number_of_computes_get(uint32_t dv_ind, uint32_t* num_computes)
+{
+  GET_DEV_AND_KFDNODE_FROM_INDX
+
+  auto tmp_simd_per_cu = uint64_t(0);
+  auto tmp_simd_count = uint64_t(0);
+  auto ret_simd_per_cu = kfd_node->get_simd_per_cu(&tmp_simd_per_cu);
+  auto ret_simd_count  = kfd_node->get_simd_count(&tmp_simd_count);
+
+  if (((ret_simd_per_cu != 0) || (ret_simd_count != 0)) ||
+      ((tmp_simd_per_cu == 0) || (tmp_simd_count == 0))) {
+    return rsmi_status_t::RSMI_STATUS_NOT_SUPPORTED;
+  }
+
+  *num_computes = (tmp_simd_count / tmp_simd_per_cu);
+  return rsmi_status_t::RSMI_STATUS_SUCCESS;
+}
+
+
 std::queue<std::string> getAllDeviceGfxVers() {
   uint32_t num_monitor_devs = 0;
   rsmi_status_t ret;

src/amd_smi/amd_smi.cc

@@ -757,6 +757,15 @@ amdsmi_get_gpu_asic_info(amdsmi_processor_handle processor_handle, amdsmi_asic_i
                     &(tmp_oam_id));
     info->oam_id = tmp_oam_id;
 
+    // default to 0xffffffff as not supported
+    info->num_of_compute_units = std::numeric_limits<uint32_t>::max();
+    auto tmp_num_of_compute_units = uint32_t(0);
+    status = rsmi_wrapper(amd::smi::rsmi_dev_number_of_computes_get, processor_handle,
+                          &tmp_num_of_compute_units);
+    if (status == amdsmi_status_t::AMDSMI_STATUS_SUCCESS) {
+        info->num_of_compute_units = tmp_num_of_compute_units;
+    }
+
     return AMDSMI_STATUS_SUCCESS;
 }
 

src/amd_smi/amd_smi_utils.cc

@@ -615,3 +615,4 @@ amdsmi_status_t smi_amdgpu_is_gpu_power_management_enabled(amd::smi::AMDSmiGPUDe
     *enabled = false;
     return AMDSMI_STATUS_SUCCESS;
 }
+