sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count",
dev->node_props.cpu_cores_count);
sysfs_show_32bit_prop(buffer, offs, "simd_count",
- dev->gpu ? (dev->node_props.simd_count *
- NUM_XCC(dev->gpu->xcc_mask)) : 0);
+ dev->gpu ? dev->node_props.simd_count : 0);
sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
dev->node_props.mem_banks_count);
sysfs_show_32bit_prop(buffer, offs, "caches_count",
/* Helper function. See kfd_fill_gpu_cache_info for parameter description */
static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext,
struct kfd_gpu_cache_info *pcache_info,
- struct kfd_cu_info *cu_info,
int cu_bitmask,
int cache_type, unsigned int cu_processor_id,
int cu_block)
/* Helper function. See kfd_fill_gpu_cache_info for parameter description */
static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext,
struct kfd_gpu_cache_info *pcache_info,
- struct kfd_cu_info *cu_info,
- int cache_type, unsigned int cu_processor_id)
+ struct amdgpu_cu_info *cu_info,
+ struct amdgpu_gfx_config *gfx_info,
+ int cache_type, unsigned int cu_processor_id,
+ struct kfd_node *knode)
{
unsigned int cu_sibling_map_mask;
int first_active_cu;
- int i, j, k;
+ int i, j, k, xcc, start, end;
struct kfd_cache_properties *pcache = NULL;
- cu_sibling_map_mask = cu_info->cu_bitmap[0][0];
+ start = ffs(knode->xcc_mask) - 1;
+ end = start + NUM_XCC(knode->xcc_mask);
+ cu_sibling_map_mask = cu_info->bitmap[start][0][0];
cu_sibling_map_mask &=
((1 << pcache_info[cache_type].num_cu_shared) - 1);
first_active_cu = ffs(cu_sibling_map_mask);
cu_sibling_map_mask = cu_sibling_map_mask >> (first_active_cu - 1);
k = 0;
- for (i = 0; i < cu_info->num_shader_engines; i++) {
- for (j = 0; j < cu_info->num_shader_arrays_per_engine; j++) {
- pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF);
- pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
- pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
- pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
- k += 4;
-
- cu_sibling_map_mask = cu_info->cu_bitmap[i % 4][j + i / 4];
- cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
+ for (xcc = start; xcc < end; xcc++) {
+ for (i = 0; i < gfx_info->max_shader_engines; i++) {
+ for (j = 0; j < gfx_info->max_sh_per_se; j++) {
+ pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF);
+ pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
+ pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
+ pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
+ k += 4;
+
+ cu_sibling_map_mask = cu_info->bitmap[xcc][i % 4][j + i / 4];
+ cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
+ }
}
}
pcache->sibling_map_size = k;
static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_node *kdev)
{
struct kfd_gpu_cache_info *pcache_info = NULL;
- int i, j, k;
+ int i, j, k, xcc, start, end;
int ct = 0;
unsigned int cu_processor_id;
int ret;
unsigned int num_cu_shared;
- struct kfd_cu_info cu_info;
- struct kfd_cu_info *pcu_info;
+ struct amdgpu_cu_info *cu_info = &kdev->adev->gfx.cu_info;
+ struct amdgpu_gfx_config *gfx_info = &kdev->adev->gfx.config;
int gpu_processor_id;
struct kfd_cache_properties *props_ext;
int num_of_entries = 0;
int num_of_cache_types = 0;
struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES];
- amdgpu_amdkfd_get_cu_info(kdev->adev, &cu_info);
- pcu_info = &cu_info;
gpu_processor_id = dev->node_props.simd_id_base;
* then it will consider only one CU from
* the shared unit
*/
+ start = ffs(kdev->xcc_mask) - 1;
+ end = start + NUM_XCC(kdev->xcc_mask);
+
for (ct = 0; ct < num_of_cache_types; ct++) {
cu_processor_id = gpu_processor_id;
if (pcache_info[ct].cache_level == 1) {
- for (i = 0; i < pcu_info->num_shader_engines; i++) {
- for (j = 0; j < pcu_info->num_shader_arrays_per_engine; j++) {
- for (k = 0; k < pcu_info->num_cu_per_sh; k += pcache_info[ct].num_cu_shared) {
+ for (xcc = start; xcc < end; xcc++) {
+ for (i = 0; i < gfx_info->max_shader_engines; i++) {
+ for (j = 0; j < gfx_info->max_sh_per_se; j++) {
+ for (k = 0; k < gfx_info->max_cu_per_sh; k += pcache_info[ct].num_cu_shared) {
- ret = fill_in_l1_pcache(&props_ext, pcache_info, pcu_info,
- pcu_info->cu_bitmap[i % 4][j + i / 4], ct,
+ ret = fill_in_l1_pcache(&props_ext, pcache_info,
+ cu_info->bitmap[xcc][i % 4][j + i / 4], ct,
cu_processor_id, k);
- if (ret < 0)
- break;
+ if (ret < 0)
+ break;
- if (!ret) {
- num_of_entries++;
- list_add_tail(&props_ext->list, &dev->cache_props);
- }
+ if (!ret) {
+ num_of_entries++;
+ list_add_tail(&props_ext->list, &dev->cache_props);
+ }
- /* Move to next CU block */
- num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <=
- pcu_info->num_cu_per_sh) ?
- pcache_info[ct].num_cu_shared :
- (pcu_info->num_cu_per_sh - k);
- cu_processor_id += num_cu_shared;
+ /* Move to next CU block */
+ num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <=
+ gfx_info->max_cu_per_sh) ?
+ pcache_info[ct].num_cu_shared :
+ (gfx_info->max_cu_per_sh - k);
+ cu_processor_id += num_cu_shared;
+ }
}
}
}
} else {
ret = fill_in_l2_l3_pcache(&props_ext, pcache_info,
- pcu_info, ct, cu_processor_id);
+ cu_info, gfx_info, ct, cu_processor_id, kdev);
if (ret < 0)
break;
{
uint32_t gpu_id;
struct kfd_topology_device *dev;
- struct kfd_cu_info cu_info;
int res = 0;
int i;
const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type];
+ struct amdgpu_gfx_config *gfx_info = &gpu->adev->gfx.config;
+ struct amdgpu_cu_info *cu_info = &gpu->adev->gfx.cu_info;
gpu_id = kfd_generate_gpu_id(gpu);
if (gpu->xcp && !gpu->xcp->ddev) {
/* Fill-in additional information that is not available in CRAT but
* needed for the topology
*/
-
- amdgpu_amdkfd_get_cu_info(dev->gpu->adev, &cu_info);
-
for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) {
dev->node_props.name[i] = __tolower(asic_name[i]);
if (asic_name[i] == '\0')
dev->node_props.name[i] = '\0';
dev->node_props.simd_arrays_per_engine =
- cu_info.num_shader_arrays_per_engine;
+ gfx_info->max_sh_per_se;
dev->node_props.gfx_target_version =
gpu->kfd->device_info.gfx_target_version;
*/
if (dev->gpu->adev->asic_type == CHIP_CARRIZO) {
dev->node_props.simd_count =
- cu_info.simd_per_cu * cu_info.cu_active_number;
+ cu_info->simd_per_cu * cu_info->number;
dev->node_props.max_waves_per_simd = 10;
}