SWDEV-477218 - Implement hipDeviceGetTexture1DLinearMaxWidth

Change-Id: I8103f710abeb869f5f84be61c57a30b24356def6

CHANGELOG.md

@@ -36,6 +36,10 @@ Full documentation for HIP is available at [rocm.docs.amd.com](https://rocm.docs
 * Soft hang when use AMD_SERIALIZE_KERNEL.
 * Memory leak in hipIpcCloseMemHandle.
 
+### Added
+- Added new HIP APIs
+  - `hipDeviceGetTexture1DLinearMaxWidth` returns the maximum number of elements allocatable in a 1D linear texture for a given element size.
+
 ## HIP 6.2 (For ROCm 6.2)
 
 ### Added

hipamd/include/hip/amd_detail/hip_api_trace.hpp

@@ -61,7 +61,7 @@
 // - Reset any of the *_STEP_VERSION defines to zero if the corresponding *_MAJOR_VERSION increases
 #define HIP_API_TABLE_STEP_VERSION 0
 #define HIP_COMPILER_API_TABLE_STEP_VERSION 0
-#define HIP_RUNTIME_API_TABLE_STEP_VERSION 5
+#define HIP_RUNTIME_API_TABLE_STEP_VERSION 6
 
 // HIP API interface
 typedef hipError_t (*t___hipPopCallConfiguration)(dim3* gridDim, dim3* blockDim, size_t* sharedMem,
@@ -1001,6 +1001,9 @@ typedef hipError_t (*t_hipDrvGraphMemcpyNodeSetParams)(hipGraphNode_t hNode,
 typedef hipError_t (*t_hipExtHostAlloc)(void **ptr, size_t size,
                                          unsigned int flags);
 
+typedef hipError_t (*t_hipDeviceGetTexture1DLinearMaxWidth)(size_t *maxWidthInElements,
+                                                            const hipChannelFormatDesc *fmtDesc,
+                                                            int device);
 // HIP Compiler dispatch table
 struct HipCompilerDispatchTable {
   size_t size;
@@ -1484,4 +1487,5 @@ struct HipDispatchTable {
   t_hipDrvGraphMemcpyNodeGetParams hipDrvGraphMemcpyNodeGetParams_fn;
   t_hipDrvGraphMemcpyNodeSetParams hipDrvGraphMemcpyNodeSetParams_fn;
   t_hipExtHostAlloc hipExtHostAlloc_fn;
+  t_hipDeviceGetTexture1DLinearMaxWidth hipDeviceGetTexture1DLinearMaxWidth_fn;
 };

hipamd/include/hip/amd_detail/hip_prof_str.h

@@ -437,6 +437,7 @@ enum hip_api_id_t {
   HIP_API_ID_hipCreateTextureObject = HIP_API_ID_NONE,
   HIP_API_ID_hipDestroyTextureObject = HIP_API_ID_NONE,
   HIP_API_ID_hipDeviceGetCount = HIP_API_ID_NONE,
+  HIP_API_ID_hipDeviceGetTexture1DLinearMaxWidth = HIP_API_ID_NONE,
   HIP_API_ID_hipGetTextureAlignmentOffset = HIP_API_ID_NONE,
   HIP_API_ID_hipGetTextureObjectResourceDesc = HIP_API_ID_NONE,
   HIP_API_ID_hipGetTextureObjectResourceViewDesc = HIP_API_ID_NONE,
@@ -3977,9 +3978,13 @@ typedef struct hip_api_data_s {
 };
 // hipDrvGraphMemcpyNodeGetParams[('hipGraphNode_t', 'hNode'), ('HIP_MEMCPY3D*', 'nodeParams')]
 #define INIT_hipDrvGraphMemcpyNodeGetParams_CB_ARGS_DATA(cb_data) { \
+  cb_data.args.hipDrvGraphMemcpyNodeGetParams.hNode = (hipGraphNode_t)hNode; \
+  cb_data.args.hipDrvGraphMemcpyNodeGetParams.nodeParams = (HIP_MEMCPY3D*)nodeParams; \
 };
 // hipDrvGraphMemcpyNodeSetParams[('hipGraphNode_t', 'hNode'), ('const HIP_MEMCPY3D*', 'nodeParams')]
 #define INIT_hipDrvGraphMemcpyNodeSetParams_CB_ARGS_DATA(cb_data) { \
+  cb_data.args.hipDrvGraphMemcpyNodeSetParams.hNode = (hipGraphNode_t)hNode; \
+  cb_data.args.hipDrvGraphMemcpyNodeSetParams.nodeParams = (const HIP_MEMCPY3D*)nodeParams; \
 };
 // hipDrvMemcpy2DUnaligned[('const hip_Memcpy2D*', 'pCopy')]
 #define INIT_hipDrvMemcpy2DUnaligned_CB_ARGS_DATA(cb_data) { \
@@ -4455,6 +4460,8 @@ typedef struct hip_api_data_s {
 };
 // hipGraphExecGetFlags[('hipGraphExec_t', 'graphExec'), ('unsigned long long*', 'flags')]
 #define INIT_hipGraphExecGetFlags_CB_ARGS_DATA(cb_data) { \
+  cb_data.args.hipGraphExecGetFlags.graphExec = (hipGraphExec_t)graphExec; \
+  cb_data.args.hipGraphExecGetFlags.flags = (unsigned long long*)flags; \
 };
 // hipGraphExecHostNodeSetParams[('hipGraphExec_t', 'hGraphExec'), ('hipGraphNode_t', 'node'), ('const hipHostNodeParams*', 'pNodeParams')]
 #define INIT_hipGraphExecHostNodeSetParams_CB_ARGS_DATA(cb_data) { \
@@ -4511,6 +4518,9 @@ typedef struct hip_api_data_s {
 };
 // hipGraphExecNodeSetParams[('hipGraphExec_t', 'graphExec'), ('hipGraphNode_t', 'node'), ('hipGraphNodeParams*', 'nodeParams')]
 #define INIT_hipGraphExecNodeSetParams_CB_ARGS_DATA(cb_data) { \
+  cb_data.args.hipGraphExecNodeSetParams.graphExec = (hipGraphExec_t)graphExec; \
+  cb_data.args.hipGraphExecNodeSetParams.node = (hipGraphNode_t)node; \
+  cb_data.args.hipGraphExecNodeSetParams.nodeParams = (hipGraphNodeParams*)nodeParams; \
 };
 // hipGraphExecUpdate[('hipGraphExec_t', 'hGraphExec'), ('hipGraph_t', 'hGraph'), ('hipGraphNode_t*', 'hErrorNode_out'), ('hipGraphExecUpdateResult*', 'updateResult_out')]
 #define INIT_hipGraphExecUpdate_CB_ARGS_DATA(cb_data) { \
@@ -4707,6 +4717,8 @@ typedef struct hip_api_data_s {
 };
 // hipGraphNodeSetParams[('hipGraphNode_t', 'node'), ('hipGraphNodeParams*', 'nodeParams')]
 #define INIT_hipGraphNodeSetParams_CB_ARGS_DATA(cb_data) { \
+  cb_data.args.hipGraphNodeSetParams.node = (hipGraphNode_t)node; \
+  cb_data.args.hipGraphNodeSetParams.nodeParams = (hipGraphNodeParams*)nodeParams; \
 };
 // hipGraphReleaseUserObject[('hipGraph_t', 'graph'), ('hipUserObject_t', 'object'), ('unsigned int', 'count')]
 #define INIT_hipGraphReleaseUserObject_CB_ARGS_DATA(cb_data) { \
@@ -6031,6 +6043,8 @@ typedef struct hip_api_data_s {
 #define INIT_hipDestroyTextureObject_CB_ARGS_DATA(cb_data) {};
 // hipDeviceGetCount()
 #define INIT_hipDeviceGetCount_CB_ARGS_DATA(cb_data) {};
+// hipDeviceGetTexture1DLinearMaxWidth()
+#define INIT_hipDeviceGetTexture1DLinearMaxWidth_CB_ARGS_DATA(cb_data) {};
 // hipGetTextureAlignmentOffset()
 #define INIT_hipGetTextureAlignmentOffset_CB_ARGS_DATA(cb_data) {};
 // hipGetTextureObjectResourceDesc()

hipamd/src/amdhip.def

@@ -29,6 +29,7 @@ hipDeviceGetCacheConfig
 hipDeviceGetStreamPriorityRange
 hipDeviceGetLimit
 hipDeviceGetName
+hipDeviceGetTexture1DLinearMaxWidth
 hipDeviceGetUuid
 hipDeviceGetPCIBusId
 hipDeviceGetSharedMemConfig

hipamd/src/hip_api_trace.cpp

@@ -121,6 +121,9 @@ hipError_t hipDeviceGetP2PAttribute(int* value, hipDeviceP2PAttr attr, int srcDe
 hipError_t hipDeviceGetPCIBusId(char* pciBusId, int len, int device);
 hipError_t hipDeviceGetSharedMemConfig(hipSharedMemConfig* pConfig);
 hipError_t hipDeviceGetStreamPriorityRange(int* leastPriority, int* greatestPriority);
+hipError_t hipDeviceGetTexture1DLinearMaxWidth(size_t* maxWidthInElements,
+                                               const hipChannelFormatDesc* fmtDesc,
+                                               int device);
 hipError_t hipDeviceGetUuid(hipUUID* uuid, hipDevice_t device);
 hipError_t hipDeviceGraphMemTrim(int device);
 hipError_t hipDevicePrimaryCtxGetState(hipDevice_t dev, unsigned int* flags, int* active);
@@ -872,6 +875,7 @@ void UpdateDispatchTable(HipDispatchTable* ptrDispatchTable) {
   ptrDispatchTable->hipDeviceGetPCIBusId_fn = hip::hipDeviceGetPCIBusId;
   ptrDispatchTable->hipDeviceGetSharedMemConfig_fn = hip::hipDeviceGetSharedMemConfig;
   ptrDispatchTable->hipDeviceGetStreamPriorityRange_fn = hip::hipDeviceGetStreamPriorityRange;
+  ptrDispatchTable->hipDeviceGetTexture1DLinearMaxWidth_fn = hip::hipDeviceGetTexture1DLinearMaxWidth;
   ptrDispatchTable->hipDeviceGetUuid_fn = hip::hipDeviceGetUuid;
   ptrDispatchTable->hipDeviceGraphMemTrim_fn = hip::hipDeviceGraphMemTrim;
   ptrDispatchTable->hipDevicePrimaryCtxGetState_fn = hip::hipDevicePrimaryCtxGetState;
@@ -1873,6 +1877,7 @@ HIP_ENFORCE_ABI(HipDispatchTable, hipExternalMemoryGetMappedMipmappedArray_fn, 4
 HIP_ENFORCE_ABI(HipDispatchTable, hipDrvGraphMemcpyNodeGetParams_fn, 459)
 HIP_ENFORCE_ABI(HipDispatchTable, hipDrvGraphMemcpyNodeSetParams_fn, 460)
 HIP_ENFORCE_ABI(HipDispatchTable, hipExtHostAlloc_fn, 461)
+HIP_ENFORCE_ABI(HipDispatchTable, hipDeviceGetTexture1DLinearMaxWidth_fn, 462)
 
 // if HIP_ENFORCE_ABI entries are added for each new function pointer in the table, the number below
 // will be +1 of the number in the last HIP_ENFORCE_ABI line. E.g.:
@@ -1880,9 +1885,9 @@ HIP_ENFORCE_ABI(HipDispatchTable, hipExtHostAlloc_fn, 461)
 //  HIP_ENFORCE_ABI(<table>, <functor>, 8)
 //
 //  HIP_ENFORCE_ABI_VERSIONING(<table>, 9) <- 8 + 1 = 9
-HIP_ENFORCE_ABI_VERSIONING(HipDispatchTable, 462)
+HIP_ENFORCE_ABI_VERSIONING(HipDispatchTable, 463)
 
-static_assert(HIP_RUNTIME_API_TABLE_MAJOR_VERSION == 0 && HIP_RUNTIME_API_TABLE_STEP_VERSION == 5,
+static_assert(HIP_RUNTIME_API_TABLE_MAJOR_VERSION == 0 && HIP_RUNTIME_API_TABLE_STEP_VERSION == 6,
               "If you get this error, add new HIP_ENFORCE_ABI(...) code for the new function "
               "pointers and then update this check so it is true");
 #endif

hipamd/src/hip_device_runtime.cpp

@@ -614,6 +614,24 @@ hipError_t hipDeviceSetSharedMemConfig(hipSharedMemConfig config) {
   HIP_RETURN(hipSuccess);
 }
 
+hipError_t hipDeviceGetTexture1DLinearMaxWidth(size_t* maxWidthInElements,
+ const hipChannelFormatDesc* fmtDesc, int device) {
+  HIP_INIT_API(hipDeviceGetTexture1DLinearMaxWidth, maxWidthInElements, fmtDesc, device);
+  if (maxWidthInElements == nullptr || fmtDesc == nullptr) {
+    HIP_RETURN(hipErrorInvalidValue);
+  }
+  hipDeviceProp_tR0600 prop = {0};
+  HIP_RETURN_ONFAIL(ihipGetDeviceProperties(&prop, device));
+  // Calculate element size according to fmtDesc
+  size_t elementSize = (fmtDesc->x + fmtDesc->y
+  + fmtDesc->z + fmtDesc->w) / 8; // Convert from bits to bytes
+  if (elementSize == 0) {
+    HIP_RETURN(hipErrorInvalidValue);
+  }
+  *maxWidthInElements = prop.maxTexture1DLinear / elementSize;
+  HIP_RETURN(hipSuccess);
+}
+
 hipError_t hipDeviceSynchronize() {
   HIP_INIT_API(hipDeviceSynchronize);
   CHECK_SUPPORTED_DURING_CAPTURE();

hipamd/src/hip_hcc.map.in

@@ -31,6 +31,7 @@ global:
     hipDeviceGetName;
     hipDeviceGetPCIBusId;
     hipDeviceGetSharedMemConfig;
+    hipDeviceGetTexture1DLinearMaxWidth;
     hipDeviceGetP2PAttribute;
     hipDevicePrimaryCtxGetState;
     hipDevicePrimaryCtxRelease;