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

SWDEV-440866 - [hip-roclr] Adds support to batch memory operations APIs

Browse Source 
Change-Id: I449ffca44bbb04d13348d112e896d603c70fd485
This commit is contained in:
Sourabh Betigeri
2024-03-26 10:22:35 -07:00
committed by Sourabh Betigeri
parent c47f9dda58
commit bd5d8e9baf
18 changed files with 217 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hipamd/include/hip/amd_detail/hip_api_trace.hpp
+6 -1
View File
@@ -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 6
#define HIP_RUNTIME_API_TABLE_STEP_VERSION 7
// HIP API interface
typedef hipError_t (*t___hipPopCallConfiguration)(dim3* gridDim, dim3* blockDim, size_t* sharedMem,
@@ -722,6 +722,8 @@ typedef hipError_t (*t_hipStreamWriteValue32)(hipStream_t stream, void* ptr, uin
unsigned int flags);
typedef hipError_t (*t_hipStreamWriteValue64)(hipStream_t stream, void* ptr, uint64_t value,
unsigned int flags);
typedef hipError_t (*t_hipStreamBatchMemOp)(hipStream_t stream, unsigned int count,
hipStreamBatchMemOpParams* paramArray, unsigned int flags);
typedef hipError_t (*t_hipTexObjectCreate)(hipTextureObject_t* pTexObject,
const HIP_RESOURCE_DESC* pResDesc,
const HIP_TEXTURE_DESC* pTexDesc,
@@ -1519,6 +1521,9 @@ struct HipDispatchTable {
// HIP_RUNTIME_API_TABLE_STEP_VERSION == 6
t_hipDeviceGetTexture1DLinearMaxWidth hipDeviceGetTexture1DLinearMaxWidth_fn;
// HIP_RUNTIME_API_TABLE_STEP_VERSION == 7
t_hipStreamBatchMemOp hipStreamBatchMemOp_fn;
// DO NOT EDIT ABOVE!
// HIP_RUNTIME_API_TABLE_STEP_VERSION == 7
hipamd/include/hip/amd_detail/hip_prof_str.h
+35 -2
View File
@@ -425,7 +425,8 @@ enum hip_api_id_t {
HIP_API_ID_hipMemcpyHtoAAsync = 405,
HIP_API_ID_hipSetValidDevices = 406,
HIP_API_ID_hipExtHostAlloc = 407,
HIP_API_ID_LAST = 407,
HIP_API_ID_hipStreamBatchMemOp = 408,
HIP_API_ID_LAST = 408,
HIP_API_ID_hipChooseDevice = HIP_API_ID_CONCAT(HIP_API_ID_,hipChooseDevice),
HIP_API_ID_hipGetDeviceProperties = HIP_API_ID_CONCAT(HIP_API_ID_,hipGetDeviceProperties),
@@ -545,7 +546,6 @@ static inline const char* hip_api_name(const uint32_t id) {
case HIP_API_ID_hipEventQuery: return "hipEventQuery";
case HIP_API_ID_hipEventRecord: return "hipEventRecord";
case HIP_API_ID_hipEventSynchronize: return "hipEventSynchronize";
case HIP_API_ID_hipExtGetLastError: return "hipExtGetLastError";
case HIP_API_ID_hipExtGetLinkTypeAndHopCount: return "hipExtGetLinkTypeAndHopCount";
case HIP_API_ID_hipExtLaunchKernel: return "hipExtLaunchKernel";
case HIP_API_ID_hipExtLaunchMultiKernelMultiDevice: return "hipExtLaunchMultiKernelMultiDevice";
@@ -859,6 +859,8 @@ static inline const char* hip_api_name(const uint32_t id) {
case HIP_API_ID_hipUserObjectRelease: return "hipUserObjectRelease";
case HIP_API_ID_hipUserObjectRetain: return "hipUserObjectRetain";
case HIP_API_ID_hipWaitExternalSemaphoresAsync: return "hipWaitExternalSemaphoresAsync";
case HIP_API_ID_hipExtGetLastError: return "hipExtGetLastError";
case HIP_API_ID_hipStreamBatchMemOp: return "hipStreamBatchMemOp";
};
return "unknown";
};
@@ -1262,6 +1264,7 @@ static inline uint32_t hipApiIdByName(const char* name) {
if (strcmp("hipUserObjectRelease", name) == 0) return HIP_API_ID_hipUserObjectRelease;
if (strcmp("hipUserObjectRetain", name) == 0) return HIP_API_ID_hipUserObjectRetain;
if (strcmp("hipWaitExternalSemaphoresAsync", name) == 0) return HIP_API_ID_hipWaitExternalSemaphoresAsync;
if (strcmp("hipStreamBatchMemOp", name) == 0) return HIP_API_ID_hipStreamBatchMemOp;
return HIP_API_ID_NONE;
}
@@ -3623,6 +3626,13 @@ typedef struct hip_api_data_s {
unsigned int numExtSems;
hipStream_t stream;
} hipWaitExternalSemaphoresAsync;
struct {
hipStream_t stream;
unsigned int count;
hipStreamBatchMemOpParams* paramArray;
hipStreamBatchMemOpParams paramArray__val;
unsigned int flags;
} hipStreamBatchMemOp;
} args;
uint64_t *phase_data;
} hip_api_data_t;
@@ -5892,6 +5902,15 @@ typedef struct hip_api_data_s {
cb_data.args.hipStreamWriteValue64.value = (uint64_t)value; \
cb_data.args.hipStreamWriteValue64.flags = (unsigned int)flags; \
};
// hipStreamBatchMemOp[('hipStream_t', 'stream'), ('unsigned int', 'count'),
// ('hipStreamBatchMemOpParams*', 'paramArray'), ('unsigned int', 'flags')]
#define INIT_hipStreamBatchMemOp_CB_ARGS_DATA(cb_data) { \
cb_data.args.hipStreamBatchMemOp.stream = (hipStream_t)stream; \
cb_data.args.hipStreamBatchMemOp.count = (unsigned int)count; \
cb_data.args.hipStreamBatchMemOp.paramArray= (hipStreamBatchMemOpParams*)paramArray; \
cb_data.args.hipStreamBatchMemOp.flags = (unsigned int)flags; \
};
// hipTexRefGetAddress[('hipDeviceptr_t*', 'dev_ptr'), ('const textureReference*', 'texRef')]
#define INIT_hipTexRefGetAddress_CB_ARGS_DATA(cb_data) { \
cb_data.args.hipTexRefGetAddress.dev_ptr = (hipDeviceptr_t*)dptr; \
@@ -7527,6 +7546,11 @@ static inline void hipApiArgsInit(hip_api_id_t id, hip_api_data_t* data) {
// hipStreamWriteValue64[('hipStream_t', 'stream'), ('void*', 'ptr'), ('uint64_t', 'value'), ('unsigned int', 'flags')]
case HIP_API_ID_hipStreamWriteValue64:
break;
// hipStreamBatchMemOp[('hipStream_t', 'stream'), ('unsigned int', 'count'),
// ('hipStreamBatchMemOpParams*', 'paramArray'), ('unsigned int', 'flags')]
case HIP_API_ID_hipStreamBatchMemOp:
if (data->args.hipStreamBatchMemOp.paramArray) data->args.hipStreamBatchMemOp.paramArray__val = *(data->args.hipStreamBatchMemOp.paramArray);
break;
// hipTexRefGetAddress[('hipDeviceptr_t*', 'dev_ptr'), ('const textureReference*', 'texRef')]
case HIP_API_ID_hipTexRefGetAddress:
if (data->args.hipTexRefGetAddress.dev_ptr) data->args.hipTexRefGetAddress.dev_ptr__val = *(data->args.hipTexRefGetAddress.dev_ptr);
@@ -10614,6 +10638,15 @@ static inline const char* hipApiString(hip_api_id_t id, const hip_api_data_t* da
oss << ", flags="; roctracer::hip_support::detail::operator<<(oss, data->args.hipStreamWriteValue64.flags);
oss << ")";
break;
case HIP_API_ID_hipStreamBatchMemOp:
oss << "hipStreamBatchMemOp(";
oss << "stream="; roctracer::hip_support::detail::operator<<(oss, data->args.hipStreamBatchMemOp.stream);
oss << ", count="; roctracer::hip_support::detail::operator<<(oss, data->args.hipStreamBatchMemOp.count);
if (data->args.hipStreamBatchMemOp.paramArray == NULL) oss << ", paramArray=NULL";
else { oss << ", paramArray="; roctracer::hip_support::detail::operator<<(oss, data->args.hipStreamBatchMemOp.paramArray__val); }
oss << ", flags="; roctracer::hip_support::detail::operator<<(oss, data->args.hipStreamBatchMemOp.flags);
oss << ")";
break;
case HIP_API_ID_hipTexRefGetAddress:
oss << "hipTexRefGetAddress(";
if (data->args.hipTexRefGetAddress.dev_ptr == NULL) oss << "dev_ptr=NULL";