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

Differentiate gfx700 and improve the logic by introducing is_gfx700()

Browse Source 
Because gfx700 has local memory but other APUs don't, we should reflect
that in the code. Meanwhile, fix a bug that on gfx902 svm aperture is not
added when calling hsaKmtGetNodeMemoryProperties().

Change-Id: Id840f2db0b14fda9ee713b219a9474c15f8a9771
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
This commit is contained in:
Yong Zhao
2018-08-07 16:03:24 -04:00
committed by Yong Zhao
parent 8fbf4a26ec
commit 110e754f64
3 changed files with 11 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/libhsakmt.h
+1
View File
@@ -107,6 +107,7 @@ HSAKMT_STATUS validate_nodeid(uint32_t nodeid, uint32_t *gpu_id);
HSAKMT_STATUS gpuid_to_nodeid(uint32_t gpu_id, uint32_t* node_id); HSAKMT_STATUS gpuid_to_nodeid(uint32_t gpu_id, uint32_t* node_id);
bool prefer_ats(HSAuint32 node_id); bool prefer_ats(HSAuint32 node_id);
uint16_t get_device_id_by_node_id(HSAuint32 node_id); uint16_t get_device_id_by_node_id(HSAuint32 node_id);
bool is_kaveri(HSAuint32 node_id);
uint16_t get_device_id_by_gpu_id(HSAuint32 gpu_id); uint16_t get_device_id_by_gpu_id(HSAuint32 gpu_id);
int get_drm_render_fd_by_gpu_id(HSAuint32 gpu_id); int get_drm_render_fd_by_gpu_id(HSAuint32 gpu_id);
HSAKMT_STATUS validate_nodeid_array(uint32_t **gpu_id_array, HSAKMT_STATUS validate_nodeid_array(uint32_t **gpu_id_array,
src/memory.c
+1 -1
View File
@@ -49,7 +49,7 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtSetMemoryPolicy(HSAuint32 Node,
pr_debug("[%s] node %d; default %d; alternate %d\n", pr_debug("[%s] node %d; default %d; alternate %d\n",
__func__, Node, DefaultPolicy, AlternatePolicy); __func__, Node, DefaultPolicy, AlternatePolicy);
if (is_dgpu) if (!is_kaveri(Node))
/* This is a legacy API useful on Kaveri only. On dGPU /* This is a legacy API useful on Kaveri only. On dGPU
* the alternate aperture is setup and used * the alternate aperture is setup and used
* automatically for coherent allocations. Don't let * automatically for coherent allocations. Don't let
src/topology.c
+9 -6
View File
@@ -1780,7 +1780,6 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtGetNodeMemoryProperties(HSAuint32 NodeId,
HSAKMT_STATUS err = HSAKMT_STATUS_SUCCESS; HSAKMT_STATUS err = HSAKMT_STATUS_SUCCESS;
uint32_t i, gpu_id; uint32_t i, gpu_id;
HSAuint64 aperture_limit; HSAuint64 aperture_limit;
bool nodeIsDGPU;
if (!MemoryProperties) if (!MemoryProperties)
return HSAKMT_STATUS_INVALID_PARAMETER; return HSAKMT_STATUS_INVALID_PARAMETER;
@@ -1816,8 +1815,6 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtGetNodeMemoryProperties(HSAuint32 NodeId,
if (gpu_id == 0) if (gpu_id == 0)
goto out; goto out;
nodeIsDGPU = topology_is_dgpu(get_device_id_by_gpu_id(gpu_id));
/*Add LDS*/ /*Add LDS*/
if (i < NumBanks && if (i < NumBanks &&
fmm_get_aperture_base_and_limit(FMM_LDS, gpu_id, fmm_get_aperture_base_and_limit(FMM_LDS, gpu_id,
@@ -1831,7 +1828,7 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtGetNodeMemoryProperties(HSAuint32 NodeId,
* For dGPU the topology node contains Local Memory and it is added by * For dGPU the topology node contains Local Memory and it is added by
* the for loop above * the for loop above
*/ */
if (!nodeIsDGPU && i < NumBanks && g_props[NodeId].node.LocalMemSize > 0 && if (is_kaveri(NodeId) && i < NumBanks && g_props[NodeId].node.LocalMemSize > 0 &&
fmm_get_aperture_base_and_limit(FMM_GPUVM, gpu_id, fmm_get_aperture_base_and_limit(FMM_GPUVM, gpu_id,
&MemoryProperties[i].VirtualBaseAddress, &aperture_limit) == HSAKMT_STATUS_SUCCESS) { &MemoryProperties[i].VirtualBaseAddress, &aperture_limit) == HSAKMT_STATUS_SUCCESS) {
MemoryProperties[i].HeapType = HSA_HEAPTYPE_FRAME_BUFFER_PRIVATE; MemoryProperties[i].HeapType = HSA_HEAPTYPE_FRAME_BUFFER_PRIVATE;
@@ -1848,8 +1845,8 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtGetNodeMemoryProperties(HSAuint32 NodeId,
i++; i++;
} }
/* On dGPUs add SVM aperture */ /* Add SVM aperture */
if (nodeIsDGPU && i < NumBanks && if (topology_is_svm_needed(get_device_id_by_gpu_id(gpu_id)) && i < NumBanks &&
fmm_get_aperture_base_and_limit( fmm_get_aperture_base_and_limit(
FMM_SVM, gpu_id, &MemoryProperties[i].VirtualBaseAddress, FMM_SVM, gpu_id, &MemoryProperties[i].VirtualBaseAddress,
&aperture_limit) == HSAKMT_STATUS_SUCCESS) { &aperture_limit) == HSAKMT_STATUS_SUCCESS) {
@@ -1954,6 +1951,12 @@ bool prefer_ats(HSAuint32 node_id)
&& g_props[node_id].node.NumFComputeCores; && g_props[node_id].node.NumFComputeCores;
} }
bool is_kaveri(HSAuint32 node_id)
{
return g_props[node_id].node.EngineId.ui32.Major == 7
&& g_props[node_id].node.EngineId.ui32.Minor == 0;
}
uint16_t get_device_id_by_gpu_id(HSAuint32 gpu_id) uint16_t get_device_id_by_gpu_id(HSAuint32 gpu_id)
{ {
unsigned int i; unsigned int i;