Skip to content

feat(ascend): support DDR memory and fix device enumeration for 310P3 #491

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
yulink1981 wants to merge 1 commit into
base: master
Choose a base branch
from
+65 −17
Open

feat(ascend): support DDR memory and fix device enumeration for 310P3 #491

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions include/nvtop/extract_gpuinfo_common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,6 +61,7 @@ enum gpuinfo_static_info_valid {
gpuinfo_l2cache_size_valid,
gpuinfo_n_exec_engines_valid,
gpuinfo_engine_count_valid,
gpuinfo_memory_type_valid,
gpuinfo_static_info_count,
};

Expand All @@ -78,6 +79,7 @@ struct gpuinfo_static_info {
unsigned engine_count;
bool integrated_graphics;
bool encode_decode_shared;
char memory_type[8];
unsigned char valid[(gpuinfo_static_info_count + CHAR_BIT - 1) / CHAR_BIT];
};

Expand Down
68 changes: 53 additions & 15 deletions src/extract_gpuinfo_ascend.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -127,7 +127,7 @@ static bool gpuinfo_ascend_get_device_handles(struct list_head *devices, unsigne
for (int i = 0; i < num_cards; ++i) {
for (int j = 0; j < card_device_list[i]; ++j) {
gpu_infos[*count].base.vendor = &gpu_vendor_ascend;
_encode_card_device_id_to_pdev(gpu_infos[*count].base.pdev, i, j);
_encode_card_device_id_to_pdev(gpu_infos[*count].base.pdev, card_list[i], j);
list_add_tail(&gpu_infos[*count].base.list, devices);
*count += 1;
}
Expand Down Expand Up @@ -163,7 +163,23 @@ static void gpuinfo_ascend_populate_static_info(struct gpu_info *_gpu_info) {
SET_VALID(gpuinfo_device_name_valid, static_info->valid);
}
free(chip_info);
// todo: it seems that other static infos are not supported by Ascend DCMI for now, will add if possible in future

/* Detect memory type: HBM (910B) vs DDR (310P3/300i DUO) */
struct dsmi_hbm_info_stru hbm_info;
int detect_ret = dcmi_get_hbm_info(card_id, device_id, &hbm_info);
if (detect_ret == DCMI_SUCCESS) {
strncpy(static_info->memory_type, "HBM", sizeof(static_info->memory_type) - 1);
static_info->memory_type[sizeof(static_info->memory_type) - 1] = '\0';
SET_VALID(gpuinfo_memory_type_valid, static_info->valid);
} else {
struct dcmi_memory_info_stru ddr_info;
detect_ret = dcmi_get_memory_info(card_id, device_id, &ddr_info);
if (detect_ret == DCMI_SUCCESS) {
strncpy(static_info->memory_type, "DDR", sizeof(static_info->memory_type) - 1);
static_info->memory_type[sizeof(static_info->memory_type) - 1] = '\0';
SET_VALID(gpuinfo_memory_type_valid, static_info->valid);
}
}
}

static void gpuinfo_ascend_refresh_dynamic_info(struct gpu_info *_gpu_info) {
Expand All @@ -188,11 +204,40 @@ static void gpuinfo_ascend_refresh_dynamic_info(struct gpu_info *_gpu_info) {
SET_VALID(gpuinfo_gpu_clock_speed_max_valid, dynamic_info->valid);
}

unsigned hbm_freq;
last_dcmi_return_status = dcmi_get_device_frequency(card_id, device_id, DCMI_FREQ_HBM, &hbm_freq);
/* Try HBM memory (Ascend 910B/910A) first, then DDR (Ascend 310P3/300i DUO) */
struct dsmi_hbm_info_stru hbm_info;
last_dcmi_return_status = dcmi_get_hbm_info(card_id, device_id, &hbm_info);
if (last_dcmi_return_status == DCMI_SUCCESS) {
dynamic_info->mem_clock_speed = hbm_freq;
SET_VALID(gpuinfo_mem_clock_speed_valid, dynamic_info->valid);
SET_GPUINFO_DYNAMIC(dynamic_info, total_memory, hbm_info.memory_size * KB_TO_GB);
SET_GPUINFO_DYNAMIC(dynamic_info, used_memory, hbm_info.memory_usage * KB_TO_GB);
SET_GPUINFO_DYNAMIC(dynamic_info, free_memory, (hbm_info.memory_size - hbm_info.memory_usage) * KB_TO_GB);
SET_GPUINFO_DYNAMIC(dynamic_info, mem_util_rate, hbm_info.memory_usage * 100 / hbm_info.memory_size);

/* HBM memory clock */
unsigned hbm_freq;
last_dcmi_return_status = dcmi_get_device_frequency(card_id, device_id, DCMI_FREQ_HBM, &hbm_freq);
if (last_dcmi_return_status == DCMI_SUCCESS) {
dynamic_info->mem_clock_speed = hbm_freq;
SET_VALID(gpuinfo_mem_clock_speed_valid, dynamic_info->valid);
}
} else {
/* Fallback to DDR memory (300i DUO / 310P3) */
struct dcmi_memory_info_stru ddr_info;
last_dcmi_return_status = dcmi_get_memory_info(card_id, device_id, &ddr_info);
if (last_dcmi_return_status == DCMI_SUCCESS) {
/* memory_size is in MB, convert to bytes for nvtop */
unsigned long long total_bytes = ddr_info.memory_size * 1024 * 1024;
unsigned long long used_bytes = total_bytes * ddr_info.utilize / 100;
SET_GPUINFO_DYNAMIC(dynamic_info, total_memory, total_bytes);
SET_GPUINFO_DYNAMIC(dynamic_info, used_memory, used_bytes);
SET_GPUINFO_DYNAMIC(dynamic_info, free_memory, total_bytes - used_bytes);
SET_GPUINFO_DYNAMIC(dynamic_info, mem_util_rate, ddr_info.utilize);

if (ddr_info.freq > 0) {
dynamic_info->mem_clock_speed = ddr_info.freq;
SET_VALID(gpuinfo_mem_clock_speed_valid, dynamic_info->valid);
}
}
}

unsigned aicore_util_rate;
Expand All @@ -202,15 +247,6 @@ static void gpuinfo_ascend_refresh_dynamic_info(struct gpu_info *_gpu_info) {
SET_VALID(gpuinfo_gpu_util_rate_valid, dynamic_info->valid);
}

struct dsmi_hbm_info_stru hbm_info;
last_dcmi_return_status = dcmi_get_hbm_info(card_id, device_id, &hbm_info);
if (last_dcmi_return_status == DCMI_SUCCESS) {
SET_GPUINFO_DYNAMIC(dynamic_info, total_memory, hbm_info.memory_size * KB_TO_GB);
SET_GPUINFO_DYNAMIC(dynamic_info, used_memory, hbm_info.memory_usage * KB_TO_GB);
SET_GPUINFO_DYNAMIC(dynamic_info, free_memory, (hbm_info.memory_size - hbm_info.memory_usage) * KB_TO_GB);
SET_GPUINFO_DYNAMIC(dynamic_info, mem_util_rate, hbm_info.memory_usage * 100 / hbm_info.memory_size);
}

int device_temperature;
last_dcmi_return_status = dcmi_get_device_temperature(card_id, device_id, &device_temperature);
if (last_dcmi_return_status == DCMI_SUCCESS) {
Expand Down Expand Up @@ -241,6 +277,8 @@ static void gpuinfo_ascend_get_running_processes(struct gpu_info *_gpu_info) {
perror("Could not allocate memory: ");
exit(EXIT_FAILURE);
}
// Zero-initialize to prevent garbage values for unset fields (e.g. gpu_usage)
memset(_gpu_info->processes, 0, _gpu_info->processes_array_size * sizeof(*_gpu_info->processes));
for (int i = 0; i < proc_num; i++) {
_gpu_info->processes[i].type = gpu_process_compute;
_gpu_info->processes[i].pid = proc_info[i].proc_id;
Expand Down
12 changes: 10 additions & 2 deletions src/interface.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -806,10 +806,12 @@ static void draw_devices(struct list_head *devices, struct nvtop_interface *inte

// MEM CLOCK
werase(dev->mem_clock_info);
const char *mem_label = GPUINFO_STATIC_FIELD_VALID(&device->static_info, memory_type)
? device->static_info.memory_type : "MEM";
if (GPUINFO_DYNAMIC_FIELD_VALID(&device->dynamic_info, mem_clock_speed))
mvwprintw(dev->mem_clock_info, 0, 0, "MEM %uMHz", device->dynamic_info.mem_clock_speed);
mvwprintw(dev->mem_clock_info, 0, 0, "%s %uMHz", mem_label, device->dynamic_info.mem_clock_speed);
else
mvwprintw(dev->mem_clock_info, 0, 0, "MEM N/A MHz");
mvwprintw(dev->mem_clock_info, 0, 0, "%s N/A MHz", mem_label);
mvwchgat(dev->mem_clock_info, 0, 0, 3, 0, cyan_color, NULL);
wnoutrefresh(dev->mem_clock_info);

Expand Down Expand Up @@ -2172,6 +2174,12 @@ void print_snapshot(struct list_head *devices, bool use_fahrenheit_option, bool
else
printf("%s\"%s\": null,\n", indent_level_four, device_name_field);

// Memory Type
if (GPUINFO_STATIC_FIELD_VALID(&device->static_info, memory_type))
printf("%s\"memory_type\": \"%s\",\n", indent_level_four, device->static_info.memory_type);
else
printf("%s\"memory_type\": null,\n", indent_level_four);

// GPU Clock Speed
if (GPUINFO_DYNAMIC_FIELD_VALID(&device->dynamic_info, gpu_clock_speed))
printf("%s\"%s\": \"%uMHz\",\n", indent_level_four, gpu_clock_field, device->dynamic_info.gpu_clock_speed);
Expand Down