AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Make doorbell-size ASIC specific

Browse Source 
This is in preparation for gfx900, which uses 64-bit doorbells. We
maintain the same number of doorbells per process by making the
doorbell page size bigger.

KFD will need to implement the same rule.

Change-Id: I3c4110869b191b83943b5a390a48edfc94d941d8


[ROCm/ROCR-Runtime commit: 48207af92a]
This commit is contained in:
Felix Kuehling
2016-10-28 18:48:44 -04:00
parent 77b6cca404
commit 3ca06e0e94
1 changed files with 43 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocr-runtime/src/queues.c
+43 -21
View File
@@ -35,8 +35,11 @@
#include <sys/mman.h>
#include <fcntl.h>
/* 1024 doorbells, 4 bytes each doorbell */
#define DOORBELLS_PAGE_SIZE 1024 * 4
/* 1024 doorbells, 4 or 8 bytes each doorbell depending on ASIC generation */
#define DOORBELL_SIZE_GFX7 4
#define DOORBELL_SIZE_GFX8 4
#define DOORBELL_SIZE_GFX9 8
#define DOORBELLS_PAGE_SIZE(ds) (1024 * (ds))
enum asic_family_type {
CHIP_KAVERI = 0,
@@ -60,46 +63,55 @@ struct device_info
{
enum asic_family_type asic_family;
uint32_t eop_buffer_size;
uint32_t doorbell_size;
};
struct device_info kaveri_device_info = {
.asic_family = CHIP_KAVERI,
.eop_buffer_size = 0,
.doorbell_size = DOORBELL_SIZE_GFX7,
};
struct device_info hawaii_device_info = {
.asic_family = CHIP_HAWAII,
.eop_buffer_size = 0,
.doorbell_size = DOORBELL_SIZE_GFX7,
};
struct device_info carrizo_device_info = {
.asic_family = CHIP_CARRIZO,
.eop_buffer_size = 4096,
.doorbell_size = DOORBELL_SIZE_GFX8,
};
struct device_info tonga_device_info = {
.asic_family = CHIP_TONGA,
.eop_buffer_size = TONGA_PAGE_SIZE,
.doorbell_size = DOORBELL_SIZE_GFX8,
};
struct device_info fiji_device_info = {
.asic_family = CHIP_FIJI,
.eop_buffer_size = TONGA_PAGE_SIZE,
.doorbell_size = DOORBELL_SIZE_GFX8,
};
struct device_info polaris10_device_info = {
.asic_family = CHIP_POLARIS10,
.eop_buffer_size = TONGA_PAGE_SIZE,
.doorbell_size = DOORBELL_SIZE_GFX8,
};
struct device_info polaris11_device_info = {
.asic_family = CHIP_POLARIS11,
.eop_buffer_size = TONGA_PAGE_SIZE,
.doorbell_size = DOORBELL_SIZE_GFX8,
};
struct device_info vega10_device_info = {
.asic_family = CHIP_VEGA10,
.eop_buffer_size = 4096,
.doorbell_size = DOORBELL_SIZE_GFX9,
};
struct device_id
@@ -199,7 +211,8 @@ struct queue
struct process_doorbells
{
bool need_mmap;
bool use_gpuvm;
uint32_t size;
void* doorbells;
pthread_mutex_t doorbells_mutex;
};
@@ -219,7 +232,8 @@ HSAKMT_STATUS init_process_doorbells(unsigned int NumNodes)
return HSAKMT_STATUS_NO_MEMORY;
for (i = 0; i < NumNodes; i++) {
doorbells[i].need_mmap = true;
doorbells[i].use_gpuvm = false;
doorbells[i].size = 0;
doorbells[i].doorbells = NULL;
pthread_mutex_init(&doorbells[i].doorbells_mutex, NULL);
}
@@ -242,7 +256,8 @@ static struct device_info *get_device_info_by_dev_id(uint16_t dev_id)
return NULL;
}
static bool use_gpuvm_doorbell(uint16_t dev_id)
static void get_doorbell_map_info(uint16_t dev_id,
struct process_doorbells *doorbell)
{
struct device_info *dev_info;
@@ -252,8 +267,9 @@ static bool use_gpuvm_doorbell(uint16_t dev_id)
* GPUVM doorbell on Tonga requires a workaround for VM TLB ACTIVE bit
* lookup bug. Remove ASIC check when this is implemented in amdgpu.
*/
return (topology_is_dgpu(dev_id) &&
dev_info->asic_family != CHIP_TONGA);
doorbell->use_gpuvm = (topology_is_dgpu(dev_id) &&
dev_info->asic_family != CHIP_TONGA);
doorbell->size = DOORBELLS_PAGE_SIZE(dev_info->doorbell_size);
}
void destroy_process_doorbells(void)
@@ -264,14 +280,14 @@ void destroy_process_doorbells(void)
return;
for (i = 0; i < num_doorbells; i++) {
if (doorbells[i].need_mmap)
if (!doorbells[i].size)
continue;
if (use_gpuvm_doorbell(get_device_id_by_node(i))) {
if (doorbells[i].use_gpuvm) {
fmm_unmap_from_gpu(doorbells[i].doorbells);
fmm_release(doorbells[i].doorbells);
} else
munmap(doorbells[i].doorbells, DOORBELLS_PAGE_SIZE);
munmap(doorbells[i].doorbells, doorbells[i].size);
}
free(doorbells);
@@ -290,11 +306,11 @@ void clear_process_doorbells(void)
return;
for (i = 0; i < num_doorbells; i++) {
if (doorbells[i].need_mmap)
if (!doorbells[i].size)
continue;
if (!use_gpuvm_doorbell(get_device_id_by_node(i)))
munmap(doorbells[i].doorbells, DOORBELLS_PAGE_SIZE);
if (!doorbells[i].use_gpuvm)
munmap(doorbells[i].doorbells, doorbells[i].size);
}
free(doorbells);
@@ -307,13 +323,12 @@ static HSAKMT_STATUS map_doorbell_apu(HSAuint32 NodeId, HSAuint32 gpu_id,
{
void *ptr;
ptr = mmap(0, DOORBELLS_PAGE_SIZE, PROT_READ|PROT_WRITE,
ptr = mmap(0, doorbells[NodeId].size, PROT_READ|PROT_WRITE,
MAP_SHARED, kfd_fd, doorbell_offset);
if (ptr == MAP_FAILED)
return HSAKMT_STATUS_ERROR;
doorbells[NodeId].need_mmap = false;
doorbells[NodeId].doorbells = ptr;
return HSAKMT_STATUS_SUCCESS;
@@ -324,19 +339,18 @@ static HSAKMT_STATUS map_doorbell_dgpu(HSAuint32 NodeId, HSAuint32 gpu_id,
{
void *ptr;
ptr = fmm_allocate_doorbell(gpu_id, DOORBELLS_PAGE_SIZE,
ptr = fmm_allocate_doorbell(gpu_id, doorbells[NodeId].size,
doorbell_offset);
if (ptr == NULL)
return HSAKMT_STATUS_ERROR;
/* map for GPU access */
if (fmm_map_to_gpu(ptr, DOORBELLS_PAGE_SIZE, NULL)) {
if (fmm_map_to_gpu(ptr, doorbells[NodeId].size, NULL)) {
fmm_release(ptr);
return HSAKMT_STATUS_ERROR;
}
doorbells[NodeId].need_mmap = false;
doorbells[NodeId].doorbells = ptr;
return HSAKMT_STATUS_SUCCESS;
@@ -348,16 +362,22 @@ static HSAKMT_STATUS map_doorbell(HSAuint32 NodeId, HSAuint32 gpu_id,
HSAKMT_STATUS status = HSAKMT_STATUS_SUCCESS;
pthread_mutex_lock(&doorbells[NodeId].doorbells_mutex);
if (!doorbells[NodeId].need_mmap) {
if (doorbells[NodeId].size) {
pthread_mutex_unlock(&doorbells[NodeId].doorbells_mutex);
return HSAKMT_STATUS_SUCCESS;
}
if (use_gpuvm_doorbell(get_device_id_by_node(NodeId)))
get_doorbell_map_info(get_device_id_by_node(NodeId),
&doorbells[NodeId]);
if (doorbells[NodeId].use_gpuvm)
status = map_doorbell_dgpu(NodeId, gpu_id, doorbell_offset);
else
status = map_doorbell_apu(NodeId, gpu_id, doorbell_offset);
if (status != HSAKMT_STATUS_SUCCESS)
doorbells[NodeId].size = 0;
pthread_mutex_unlock(&doorbells[NodeId].doorbells_mutex);
return status;
@@ -619,7 +639,9 @@ hsaKmtCreateQueue(
}
QueueResource->QueueId = PORT_VPTR_TO_UINT64(q);
QueueResource->Queue_DoorBell = VOID_PTR_ADD32(doorbells[NodeId].doorbells, q->queue_id);
QueueResource->Queue_DoorBell = VOID_PTR_ADD(doorbells[NodeId].doorbells,
(q->queue_id *
dev_info->doorbell_size));
return HSAKMT_STATUS_SUCCESS;
}