maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[libomptarget] Add support for target update non-contiguous 

This patch is the runtime support for https://reviews.llvm.org/D84192.

In order not to modify the tgt_target_data_update information but still be
able to pass the extra information for non-contiguous map item (offset,
count, and stride for each dimension), this patch overload arg when
the maptype is set as OMP_TGT_MAPTYPE_DESCRIPTOR. The origin arg is for
passing the pointer information, however, the overloaded arg is an
array of descriptor_dim:

```
struct descriptor_dim {
  int64_t offset;
  int64_t count;
  int64_t stride
};
```

and the array size is the dimension size. In addition, since we
have count and stride information in descriptor_dim, we can replace/overload the
arg_size parameter by using dimension size.

Reviewed By: grokos

Differential Revision: https://reviews.llvm.org/D82245
This commit is contained in:
cchen 2020-11-06 20:54:49 -06:00
parent e6f3ec6ebb
commit 6847bcec1a
3 changed files with 253 additions and 81 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
openmp/libomptarget/include/omptarget.h
View File

@ -50,6 +50,8 @@ enum tgt_map_type {
OMP_TGT_MAPTYPE_CLOSE = 0x400,
// runtime error if not already allocated
OMP_TGT_MAPTYPE_PRESENT = 0x1000,
// descriptor for non-contiguous target-update
OMP_TGT_MAPTYPE_NON_CONTIG = 0x100000000000,
// member of struct, member given by [16 MSBs] - 1
OMP_TGT_MAPTYPE_MEMBER_OF = 0xffff000000000000
};
@ -121,6 +123,13 @@ struct __tgt_async_info {
void *Queue = nullptr;
};
/// This struct is a record of non-contiguous information
struct __tgt_target_non_contig {
uint64_t offset;
uint64_t count;
uint64_t stride;
};
#ifdef __cplusplus
extern "C" {
#endif

225
openmp/libomptarget/src/omptarget.cpp
View File

@ -635,6 +635,133 @@ int targetDataEnd(DeviceTy &Device, int32_t ArgNum, void **ArgBases,
return OFFLOAD_SUCCESS;
}
static int target_data_contiguous(DeviceTy &Device, void *args_base,
void *HstPtrBegin, int64_t MapSize,
int64_t arg_type) {
bool IsLast, IsHostPtr;
void *TgtPtrBegin = Device.getTgtPtrBegin(HstPtrBegin, MapSize, IsLast, false,
IsHostPtr, /*MustContain=*/true);
if (!TgtPtrBegin) {
DP("hst data:" DPxMOD " not found, becomes a noop\n", DPxPTR(HstPtrBegin));
if (arg_type & OMP_TGT_MAPTYPE_PRESENT) {
MESSAGE("device mapping required by 'present' motion modifier does not "
"exist for host address " DPxMOD " (%" PRId64 " bytes)",
DPxPTR(HstPtrBegin), MapSize);
return OFFLOAD_FAIL;
}
return OFFLOAD_SUCCESS;
}
if (PM->RTLs.RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY &&
TgtPtrBegin == HstPtrBegin) {
DP("hst data:" DPxMOD " unified and shared, becomes a noop\n",
DPxPTR(HstPtrBegin));
return OFFLOAD_SUCCESS;
}
if (arg_type & OMP_TGT_MAPTYPE_FROM) {
DP("Moving %" PRId64 " bytes (tgt:" DPxMOD ") -> (hst:" DPxMOD ")\n",
arg_sizes[i], DPxPTR(TgtPtrBegin), DPxPTR(HstPtrBegin));
int rt = Device.retrieveData(HstPtrBegin, TgtPtrBegin, MapSize, nullptr);
if (rt != OFFLOAD_SUCCESS) {
REPORT("Copying data from device failed.\n");
return OFFLOAD_FAIL;
}
uintptr_t lb = (uintptr_t)HstPtrBegin;
uintptr_t ub = (uintptr_t)HstPtrBegin + MapSize;
Device.ShadowMtx.lock();
for (ShadowPtrListTy::iterator it = Device.ShadowPtrMap.begin();
it != Device.ShadowPtrMap.end(); ++it) {
void **ShadowHstPtrAddr = (void **)it->first;
if ((uintptr_t)ShadowHstPtrAddr < lb)
continue;
if ((uintptr_t)ShadowHstPtrAddr >= ub)
break;
DP("Restoring original host pointer value " DPxMOD
" for host pointer " DPxMOD "\n",
DPxPTR(it->second.HstPtrVal), DPxPTR(ShadowHstPtrAddr));
*ShadowHstPtrAddr = it->second.HstPtrVal;
}
Device.ShadowMtx.unlock();
}
if (arg_type & OMP_TGT_MAPTYPE_TO) {
DP("Moving %" PRId64 " bytes (hst:" DPxMOD ") -> (tgt:" DPxMOD ")\n",
arg_sizes[i], DPxPTR(HstPtrBegin), DPxPTR(TgtPtrBegin));
int rt = Device.submitData(TgtPtrBegin, HstPtrBegin, MapSize, nullptr);
if (rt != OFFLOAD_SUCCESS) {
REPORT("Copying data to device failed.\n");
return OFFLOAD_FAIL;
}
uintptr_t lb = (uintptr_t)HstPtrBegin;
uintptr_t ub = (uintptr_t)HstPtrBegin + MapSize;
Device.ShadowMtx.lock();
for (ShadowPtrListTy::iterator it = Device.ShadowPtrMap.begin();
it != Device.ShadowPtrMap.end(); ++it) {
void **ShadowHstPtrAddr = (void **)it->first;
if ((uintptr_t)ShadowHstPtrAddr < lb)
continue;
if ((uintptr_t)ShadowHstPtrAddr >= ub)
break;
DP("Restoring original target pointer value " DPxMOD " for target "
"pointer " DPxMOD "\n",
DPxPTR(it->second.TgtPtrVal), DPxPTR(it->second.TgtPtrAddr));
rt = Device.submitData(it->second.TgtPtrAddr, &it->second.TgtPtrVal,
sizeof(void *), nullptr);
if (rt != OFFLOAD_SUCCESS) {
REPORT("Copying data to device failed.\n");
Device.ShadowMtx.unlock();
return OFFLOAD_FAIL;
}
}
Device.ShadowMtx.unlock();
}
return OFFLOAD_SUCCESS;
}
static int target_data_non_contiguous(DeviceTy &Device, void *arg_base,
__tgt_target_non_contig *non_contig,
uint64_t size, int64_t arg_type,
int current_dim, int dim_size,
uint64_t offset) {
int rt = OFFLOAD_SUCCESS;
if (current_dim < dim_size) {
for (unsigned int i = 0; i < non_contig[current_dim].count; ++i) {
uint64_t cur_offset =
(non_contig[current_dim].offset + i) * non_contig[current_dim].stride;
// we only need to transfer the first element for the last dimension
// since we've already got a contiguous piece.
if (current_dim != dim_size - 1 || i == 0) {
rt = target_data_non_contiguous(Device, arg_base, non_contig, size,
arg_type, current_dim + 1, dim_size,
offset + cur_offset);
// Stop the whole process if any contiguous piece returns anything
// other than OFFLOAD_SUCCESS.
if (rt != OFFLOAD_SUCCESS)
return rt;
}
}
} else {
char *ptr = (char *)arg_base + offset;
DP("Transfer of non-contiguous : host ptr %lx offset %ld len %ld\n",
(uint64_t)ptr, offset, size);
rt = target_data_contiguous(Device, arg_base, ptr, size, arg_type);
}
return rt;
}
static int get_non_contig_merged_dimension(__tgt_target_non_contig *non_contig,
int32_t dim_size) {
int removed_dim = 0;
for (int i = dim_size - 1; i > 0; --i) {
if (non_contig[i].count * non_contig[i].stride == non_contig[i - 1].stride)
removed_dim++;
}
return removed_dim;
}
/// Internal function to pass data to/from the target.
// async_info_ptr is currently unused, added here so target_data_update has the
// same signature as targetDataBegin and targetDataEnd.
@ -668,88 +795,24 @@ int target_data_update(DeviceTy &Device, int32_t arg_num,
continue;
}
void *HstPtrBegin = args[i];
int64_t MapSize = arg_sizes[i];
bool IsLast, IsHostPtr;
void *TgtPtrBegin = Device.getTgtPtrBegin(
HstPtrBegin, MapSize, IsLast, false, IsHostPtr, /*MustContain=*/true);
if (!TgtPtrBegin) {
DP("hst data:" DPxMOD " not found, becomes a noop\n", DPxPTR(HstPtrBegin));
if (arg_types[i] & OMP_TGT_MAPTYPE_PRESENT) {
MESSAGE("device mapping required by 'present' motion modifier does not "
"exist for host address " DPxMOD " (%" PRId64 " bytes)",
DPxPTR(HstPtrBegin), MapSize);
return OFFLOAD_FAIL;
}
continue;
}
if (PM->RTLs.RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY &&
TgtPtrBegin == HstPtrBegin) {
DP("hst data:" DPxMOD " unified and shared, becomes a noop\n",
DPxPTR(HstPtrBegin));
continue;
}
if (arg_types[i] & OMP_TGT_MAPTYPE_FROM) {
DP("Moving %" PRId64 " bytes (tgt:" DPxMOD ") -> (hst:" DPxMOD ")\n",
arg_sizes[i], DPxPTR(TgtPtrBegin), DPxPTR(HstPtrBegin));
int rt = Device.retrieveData(HstPtrBegin, TgtPtrBegin, MapSize, nullptr);
if (rt != OFFLOAD_SUCCESS) {
REPORT("Copying data from device failed.\n");
return OFFLOAD_FAIL;
}
uintptr_t lb = (uintptr_t) HstPtrBegin;
uintptr_t ub = (uintptr_t) HstPtrBegin + MapSize;
Device.ShadowMtx.lock();
for (ShadowPtrListTy::iterator it = Device.ShadowPtrMap.begin();
it != Device.ShadowPtrMap.end(); ++it) {
void **ShadowHstPtrAddr = (void**) it->first;
if ((uintptr_t) ShadowHstPtrAddr < lb)
continue;
if ((uintptr_t) ShadowHstPtrAddr >= ub)
break;
DP("Restoring original host pointer value " DPxMOD " for host pointer "
DPxMOD "\n", DPxPTR(it->second.HstPtrVal),
DPxPTR(ShadowHstPtrAddr));
*ShadowHstPtrAddr = it->second.HstPtrVal;
}
Device.ShadowMtx.unlock();
}
if (arg_types[i] & OMP_TGT_MAPTYPE_TO) {
DP("Moving %" PRId64 " bytes (hst:" DPxMOD ") -> (tgt:" DPxMOD ")\n",
arg_sizes[i], DPxPTR(HstPtrBegin), DPxPTR(TgtPtrBegin));
int rt = Device.submitData(TgtPtrBegin, HstPtrBegin, MapSize, nullptr);
if (rt != OFFLOAD_SUCCESS) {
REPORT("Copying data to device failed.\n");
return OFFLOAD_FAIL;
}
uintptr_t lb = (uintptr_t) HstPtrBegin;
uintptr_t ub = (uintptr_t) HstPtrBegin + MapSize;
Device.ShadowMtx.lock();
for (ShadowPtrListTy::iterator it = Device.ShadowPtrMap.begin();
it != Device.ShadowPtrMap.end(); ++it) {
void **ShadowHstPtrAddr = (void **)it->first;
if ((uintptr_t)ShadowHstPtrAddr < lb)
continue;
if ((uintptr_t)ShadowHstPtrAddr >= ub)
break;
DP("Restoring original target pointer value " DPxMOD " for target "
"pointer " DPxMOD "\n",
DPxPTR(it->second.TgtPtrVal), DPxPTR(it->second.TgtPtrAddr));
rt = Device.submitData(it->second.TgtPtrAddr, &it->second.TgtPtrVal,
sizeof(void *), nullptr);
if (rt != OFFLOAD_SUCCESS) {
REPORT("Copying data to device failed.\n");
Device.ShadowMtx.unlock();
return OFFLOAD_FAIL;
}
}
Device.ShadowMtx.unlock();
int rt = OFFLOAD_SUCCESS;
if (arg_types[i] & OMP_TGT_MAPTYPE_NON_CONTIG) {
__tgt_target_non_contig *non_contig = (__tgt_target_non_contig *)args[i];
int32_t dim_size = arg_sizes[i];
uint64_t size =
non_contig[dim_size - 1].count * non_contig[dim_size - 1].stride;
int32_t merged_dim =
get_non_contig_merged_dimension(non_contig, dim_size);
rt = target_data_non_contiguous(
Device, args_base[i], non_contig, size, arg_types[i],
/*current_dim=*/0, dim_size - merged_dim, /*offset=*/0);
} else {
rt = target_data_contiguous(Device, args_base[i], args[i], arg_sizes[i],
arg_types[i]);
}
if (rt == OFFLOAD_FAIL)
return OFFLOAD_FAIL;
}
return OFFLOAD_SUCCESS;
}

100
openmp/libomptarget/test/offloading/non_contiguous_update.cpp Normal file
View File

@ -0,0 +1,100 @@
// RUN: %libomptarget-compile-aarch64-unknown-linux-gnu && env LIBOMPTARGET_DEBUG=1 %libomptarget-run-aarch64-unknown-linux-gnu 2>&1 | %fcheck-aarch64-unknown-linux-gnu -allow-empty -check-prefix=DEBUG
// RUN: %libomptarget-compile-powerpc64-ibm-linux-gnu && env LIBOMPTARGET_DEBUG=1 %libomptarget-run-powerpc64-ibm-linux-gnu 2>&1 | %fcheck-powerpc64-ibm-linux-gnu -allow-empty -check-prefix=DEBUG
// RUN: %libomptarget-compile-powerpc64le-ibm-linux-gnu && env LIBOMPTARGET_DEBUG=1 %libomptarget-run-powerpc64le-ibm-linux-gnu 2>&1 | %fcheck-powerpc64le-ibm-linux-gnu -allow-empty -check-prefix=DEBUG
// RUN: %libomptarget-compile-x86_64-pc-linux-gnu && env LIBOMPTARGET_DEBUG=1 %libomptarget-run-x86_64-pc-linux-gnu 2>&1 | %fcheck-x86_64-pc-linux-gnu -allow-empty -check-prefix=DEBUG
// RUN: %libomptarget-compile-nvptx64-nvidia-cuda && env LIBOMPTARGET_DEBUG=1 %libomptarget-run-nvptx64-nvidia-cuda 2>&1 | %fcheck-nvptx64-nvidia-cuda -allow-empty -check-prefix=DEBUG
// REQUIRES: libomptarget-debug
#include <cstdio>
#include <cstdlib>
#include <cassert>
// Data structure definitions copied from OpenMP RTL.
struct __tgt_target_non_contig {
int64_t offset;
int64_t width;
int64_t stride;
};
enum tgt_map_type {
OMP_TGT_MAPTYPE_NON_CONTIG = 0x100000000000
};
// OpenMP RTL interfaces
#ifdef __cplusplus
extern "C" {
#endif
void __tgt_target_data_update(int64_t device_id, int32_t arg_num,
void **args_base, void **args, int64_t *arg_sizes,
int64_t *arg_types);
#ifdef __cplusplus
}
#endif
int main() {
// case 1
// int arr[3][4][5][6];
// #pragma omp target update to(arr[0:2][1:3][1:2][:])
// set up descriptor
__tgt_target_non_contig non_contig[5] = {
{0, 2, 480}, {1, 3, 120}, {1, 2, 24}, {0, 6, 4}, {0, 1, 4}};
int64_t size = 4, type = OMP_TGT_MAPTYPE_NON_CONTIG;
void *base;
void *begin = &non_contig;
int64_t *sizes = &size;
int64_t *types = &type;
// The below diagram is the visualization of the non-contiguous transfer after
// optimization. Note that each element represent the innermost dimension
// (unit size = 24) since the stride * count of last dimension is equal to the
// stride of second last dimension.
//
// OOOOO OOOOO OOOOO
// OXXOO OXXOO OOOOO
// OXXOO OXXOO OOOOO
// OXXOO OXXOO OOOOO
__tgt_target_data_update(/*device_id*/ -1, /*arg_num*/ 1, &base, &begin,
sizes, types);
// DEBUG: offset 144
// DEBUG: offset 264
// DEBUG: offset 384
// DEBUG: offset 624
// DEBUG: offset 744
// DEBUG: offset 864
// case 2
// double darr[3][4][5];
// #pragma omp target update to(darr[0:2:2][2:2][:2:2])
// set up descriptor
__tgt_target_non_contig non_contig_2[4] = {
{0, 2, 320}, {2, 2, 40}, {0, 2, 16}, {0, 1, 8}};
int64_t size_2 = 4, type_2 = OMP_TGT_MAPTYPE_NON_CONTIG;
void *base_2;
void *begin_2 = &non_contig_2;
int64_t *sizes_2 = &size_2;
int64_t *types_2 = &type_2;
// The below diagram is the visualization of the non-contiguous transfer after
// optimization. Note that each element represent the innermost dimension
// (unit size = 24) since the stride * count of last dimension is equal to the
// stride of second last dimension.
//
// OOOOO OOOOO OOOOO
// OOOOO OOOOO OOOOO
// XOXOO OOOOO XOXOO
// XOXOO OOOOO XOXOO
__tgt_target_data_update(/*device_id*/ -1, /*arg_num*/ 1, &base_2, &begin_2,
sizes_2, types_2);
// DEBUG: offset 80
// DEBUG: offset 96
// DEBUG: offset 120
// DEBUG: offset 136
// DEBUG: offset 400
// DEBUG: offset 416
// DEBUG: offset 440
// DEBUG: offset 456
return 0;
}