summaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_crat.c')
-rw-r--r--drivers/gpu/drm/amd/amdkfd/kfd_crat.c76
1 files changed, 57 insertions, 19 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
index 7f2ae0d15d4a..4a7180b46b71 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
@@ -28,6 +28,7 @@
#include "kfd_topology.h"
#include "amdgpu.h"
#include "amdgpu_amdkfd.h"
+#include "amdgpu_xgmi.h"
/* GPU Processor ID base for dGPUs for which VCRAT needs to be created.
* GPU processor ID are expressed with Bit[31]=1.
@@ -1422,6 +1423,7 @@ err:
static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
+ bool cache_line_size_missing,
struct kfd_gpu_cache_info *pcache_info)
{
struct amdgpu_device *adev = kdev->adev;
@@ -1434,7 +1436,10 @@ static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
CRAT_CACHE_FLAGS_DATA_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
- pcache_info[0].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2;
+ pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2;
+ pcache_info[i].cache_line_size = adev->gfx.config.gc_tcp_cache_line_size;
+ if (cache_line_size_missing && !pcache_info[i].cache_line_size)
+ pcache_info[i].cache_line_size = 128;
i++;
}
/* Scalar L1 Instruction Cache per SQC */
@@ -1446,6 +1451,9 @@ static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
CRAT_CACHE_FLAGS_INST_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
+ pcache_info[i].cache_line_size = adev->gfx.config.gc_instruction_cache_line_size;
+ if (cache_line_size_missing && !pcache_info[i].cache_line_size)
+ pcache_info[i].cache_line_size = 128;
i++;
}
/* Scalar L1 Data Cache per SQC */
@@ -1456,6 +1464,9 @@ static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
CRAT_CACHE_FLAGS_DATA_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
+ pcache_info[i].cache_line_size = adev->gfx.config.gc_scalar_data_cache_line_size;
+ if (cache_line_size_missing && !pcache_info[i].cache_line_size)
+ pcache_info[i].cache_line_size = 64;
i++;
}
/* GL1 Data Cache per SA */
@@ -1468,6 +1479,8 @@ static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
CRAT_CACHE_FLAGS_DATA_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
+ if (cache_line_size_missing)
+ pcache_info[i].cache_line_size = 128;
i++;
}
/* L2 Data Cache per GPU (Total Tex Cache) */
@@ -1478,6 +1491,9 @@ static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
CRAT_CACHE_FLAGS_DATA_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
+ pcache_info[i].cache_line_size = adev->gfx.config.gc_tcc_cache_line_size;
+ if (cache_line_size_missing && !pcache_info[i].cache_line_size)
+ pcache_info[i].cache_line_size = 128;
i++;
}
/* L3 Data Cache per GPU */
@@ -1488,6 +1504,7 @@ static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
CRAT_CACHE_FLAGS_DATA_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
+ pcache_info[i].cache_line_size = 64;
i++;
}
return i;
@@ -1503,6 +1520,8 @@ static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
if (adev->gfx.config.gc_tcp_size_per_cu) {
pcache_info[i].cache_size = adev->gfx.config.gc_tcp_size_per_cu;
pcache_info[i].cache_level = 1;
+ /* Cacheline size not available in IP discovery for gc943,gc944 */
+ pcache_info[i].cache_line_size = 128;
pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
CRAT_CACHE_FLAGS_DATA_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
@@ -1514,6 +1533,7 @@ static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
pcache_info[i].cache_size =
adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
pcache_info[i].cache_level = 1;
+ pcache_info[i].cache_line_size = 64;
pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
CRAT_CACHE_FLAGS_INST_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
@@ -1524,6 +1544,7 @@ static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
pcache_info[i].cache_level = 1;
+ pcache_info[i].cache_line_size = 64;
pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
CRAT_CACHE_FLAGS_DATA_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
@@ -1534,6 +1555,7 @@ static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
if (adev->gfx.config.gc_tcc_size) {
pcache_info[i].cache_size = adev->gfx.config.gc_tcc_size;
pcache_info[i].cache_level = 2;
+ pcache_info[i].cache_line_size = 128;
pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
CRAT_CACHE_FLAGS_DATA_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
@@ -1544,6 +1566,7 @@ static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
if (adev->gmc.mall_size) {
pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
pcache_info[i].cache_level = 3;
+ pcache_info[i].cache_line_size = 64;
pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
CRAT_CACHE_FLAGS_DATA_CACHE |
CRAT_CACHE_FLAGS_SIMD_CACHE);
@@ -1556,6 +1579,7 @@ static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info)
{
int num_of_cache_types = 0;
+ bool cache_line_size_missing = false;
switch (kdev->adev->asic_type) {
case CHIP_KAVERI:
@@ -1614,6 +1638,8 @@ int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pc
num_of_cache_types = ARRAY_SIZE(aldebaran_cache_info);
break;
case IP_VERSION(9, 4, 3):
+ case IP_VERSION(9, 4, 4):
+ case IP_VERSION(9, 5, 0):
num_of_cache_types =
kfd_fill_gpu_cache_info_from_gfx_config_v2(kdev->kfd,
*pcache_info);
@@ -1677,8 +1703,19 @@ int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pc
case IP_VERSION(11, 0, 4):
case IP_VERSION(11, 5, 0):
case IP_VERSION(11, 5, 1):
+ case IP_VERSION(11, 5, 2):
+ case IP_VERSION(11, 5, 3):
+ /* Cacheline size not available in IP discovery for gc11.
+ * kfd_fill_gpu_cache_info_from_gfx_config to hard code it
+ */
+ cache_line_size_missing = true;
+ fallthrough;
+ case IP_VERSION(12, 0, 0):
+ case IP_VERSION(12, 0, 1):
num_of_cache_types =
- kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd, *pcache_info);
+ kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd,
+ cache_line_size_missing,
+ *pcache_info);
break;
default:
*pcache_info = dummy_cache_info;
@@ -2096,9 +2133,6 @@ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3);
int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT :
KFD_CRAT_INTRA_SOCKET_WEIGHT;
- uint32_t bandwidth = ext_cpu ? amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(
- kdev->adev, NULL, true) : mem_bw;
-
/*
* with host gpu xgmi link, host can access gpu memory whether
* or not pcie bar type is large, so always create bidirectional
@@ -2107,8 +2141,16 @@ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
sub_type_hdr->weight_xgmi = weight;
- sub_type_hdr->minimum_bandwidth_mbs = bandwidth;
- sub_type_hdr->maximum_bandwidth_mbs = bandwidth;
+ if (ext_cpu) {
+ amdgpu_xgmi_get_bandwidth(kdev->adev, NULL,
+ AMDGPU_XGMI_BW_MODE_PER_LINK,
+ AMDGPU_XGMI_BW_UNIT_MBYTES,
+ &sub_type_hdr->minimum_bandwidth_mbs,
+ &sub_type_hdr->maximum_bandwidth_mbs);
+ } else {
+ sub_type_hdr->minimum_bandwidth_mbs = mem_bw;
+ sub_type_hdr->maximum_bandwidth_mbs = mem_bw;
+ }
} else {
sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
sub_type_hdr->minimum_bandwidth_mbs =
@@ -2161,12 +2203,12 @@ static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
if (use_ta_info) {
sub_type_hdr->weight_xgmi = KFD_CRAT_XGMI_WEIGHT *
- amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev);
- sub_type_hdr->maximum_bandwidth_mbs =
- amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev,
- peer_kdev->adev, false);
- sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ?
- amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0;
+ amdgpu_xgmi_get_hops_count(kdev->adev, peer_kdev->adev);
+ amdgpu_xgmi_get_bandwidth(kdev->adev, peer_kdev->adev,
+ AMDGPU_XGMI_BW_MODE_PER_PEER,
+ AMDGPU_XGMI_BW_UNIT_MBYTES,
+ &sub_type_hdr->minimum_bandwidth_mbs,
+ &sub_type_hdr->maximum_bandwidth_mbs);
} else {
bool is_single_hop = kdev->kfd == peer_kdev->kfd;
int weight = is_single_hop ? KFD_CRAT_INTRA_SOCKET_WEIGHT :
@@ -2210,9 +2252,6 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image,
* Modify length and total_entries as subunits are added.
*/
avail_size -= sizeof(struct crat_header);
- if (avail_size < 0)
- return -ENOMEM;
-
memset(crat_table, 0, sizeof(struct crat_header));
memcpy(&crat_table->signature, CRAT_SIGNATURE,
@@ -2226,9 +2265,6 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image,
* First fill in the sub type header and then sub type data
*/
avail_size -= sizeof(struct crat_subtype_computeunit);
- if (avail_size < 0)
- return -ENOMEM;
-
sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
@@ -2325,6 +2361,8 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image,
continue;
if (peer_dev->gpu->kfd->hive_id != kdev->kfd->hive_id)
continue;
+ if (!amdgpu_xgmi_get_is_sharing_enabled(kdev->adev, peer_dev->gpu->adev))
+ continue;
sub_type_hdr = (typeof(sub_type_hdr))(
(char *)sub_type_hdr +
sizeof(struct crat_subtype_iolink));
generated by cgit (git 2.34.1) at 2025-07-30 10:11:38 +0000