Skip to content

Commit dd4e015

Browse files
RossBruntonRossBrunton
committed
Replace loader handles with field at start of handle data
This replaces the handle logic in the loader from wrapped pointers to a ddi table at the start of the handle struct itself.
1 parent d277bc4 commit dd4e015

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

81 files changed

+1312
-4369
lines changed

‎unified-runtime/scripts/generate_code.py

Lines changed: 0 additions & 19 deletions
Original file line numberDiff line numberDiff line change
@@ -238,25 +238,6 @@ def _mako_loader_cpp(path, namespace, tags, version, specs, meta):
238238
"make_loader_cpp path %s namespace %s version %s\n" % (path, namespace, version)
239239
)
240240
loc = 0
241-
template = "ldrddi.hpp.mako"
242-
fin = os.path.join(templates_dir, template)
243-
244-
name = "%s_ldrddi" % (namespace)
245-
filename = "%s.hpp" % (name)
246-
fout = os.path.join(path, filename)
247-
248-
print("Generating %s..." % fout)
249-
loc += util.makoWrite(
250-
fin,
251-
fout,
252-
name=name,
253-
ver=version,
254-
namespace=namespace,
255-
tags=tags,
256-
specs=specs,
257-
meta=meta,
258-
)
259-
260241
template = "ldrddi.cpp.mako"
261242
fin = os.path.join(templates_dir, template)
262243

‎unified-runtime/scripts/templates/helper.py

Lines changed: 0 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -1487,20 +1487,6 @@ def get_initial_null_set(obj):
14871487
return ""
14881488

14891489

1490-
"""
1491-
Public:
1492-
returns true if the function always wraps output pointers in loader handles
1493-
"""
1494-
1495-
1496-
def always_wrap_outputs(obj):
1497-
cname = obj_traits.class_name(obj)
1498-
return (cname, obj["name"]) in [
1499-
("$xProgram", "Link"),
1500-
("$xProgram", "LinkExp"),
1501-
]
1502-
1503-
15041490
"""
15051491
Private:
15061492
returns the list of parameters, filtering based on desc tags

‎unified-runtime/scripts/templates/ldrddi.cpp.mako

Lines changed: 22 additions & 231 deletions
Large diffs are not rendered by default.

‎unified-runtime/scripts/templates/ldrddi.hpp.mako

Lines changed: 0 additions & 57 deletions
This file was deleted.

‎unified-runtime/scripts/templates/ur_interface_loader.cpp.mako

Lines changed: 29 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -20,6 +20,7 @@ from templates import helper as th
2020
//===----------------------------------------------------------------------===//
2121
#include <${n}_api.h>
2222
#include <${n}_ddi.h>
23+
#include <mutex>
2324

2425
#include "ur_interface_loader.hpp"
2526

@@ -68,22 +69,45 @@ ${X}_APIEXPORT ${x}_result_t ${X}_APICALL ${tbl['export']['name']}(
6869
} // extern "C"
6970
#endif
7071

71-
#ifdef UR_STATIC_ADAPTER_${Adapter}
72-
namespace ur::${adapter} {
73-
ur_result_t urAdapterGetDdiTables(ur_dditable_t *ddi) {
74-
if (ddi == nullptr) {
72+
namespace {
73+
ur_result_t populateDdiTable(ur_dditable_t *ddi) {
74+
if (ddi == nullptr) {
7575
return UR_RESULT_ERROR_INVALID_NULL_POINTER;
7676
}
7777

7878
ur_result_t result;
7979

80+
#ifdef UR_STATIC_ADAPTER_${Adapter}
81+
#define NAMESPACE_ ::ur::${adapter}
82+
#else
83+
#define NAMESPACE_
84+
#endif
85+
8086
%for tbl in th.get_pfntables(specs, meta, n, tags):
81-
result = ${n}::${adapter}::${tbl['export']['name']}( ${X}_API_VERSION_CURRENT, &ddi->${tbl['name']} );
87+
result = NAMESPACE_::${tbl['export']['name']}( ${X}_API_VERSION_CURRENT, &ddi->${tbl['name']} );
8288
if (result != UR_RESULT_SUCCESS)
8389
return result;
8490
%endfor
8591

92+
#undef NAMESPACE_
93+
8694
return result;
8795
}
8896
}
97+
98+
99+
namespace ur::${adapter} {
100+
const ${x}_dditable_t *ddi_getter::value() {
101+
static std::once_flag flag;
102+
static ${x}_dditable_t table;
103+
104+
std::call_once(flag, []() { populateDdiTable(&table); });
105+
return &table;
106+
}
107+
108+
#ifdef UR_STATIC_ADAPTER_${Adapter}
109+
ur_result_t urAdapterGetDdiTables(${x}_dditable_t *ddi) {
110+
return populateDdiTable(ddi);
111+
}
89112
#endif
113+
}

‎unified-runtime/scripts/templates/ur_interface_loader.hpp.mako

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -18,6 +18,8 @@ from templates import helper as th
1818
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
1919
//
2020
//===----------------------------------------------------------------------===//
21+
#pragma once
22+
2123
#include <${n}_api.h>
2224
#include <${n}_ddi.h>
2325

@@ -36,4 +38,8 @@ ${x}_result_t ${th.make_func_name(n, tags, obj)}(
3638
#ifdef UR_STATIC_ADAPTER_LEVEL_ZERO
3739
ur_result_t urAdapterGetDdiTables(ur_dditable_t *ddi);
3840
#endif
41+
42+
struct ddi_getter {
43+
const static ${x}_dditable_t *value();
44+
};
3945
}

‎unified-runtime/source/adapters/cuda/adapter.cpp

Lines changed: 3 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -14,7 +14,7 @@
1414
#include "logger/ur_logger.hpp"
1515
#include "tracing.hpp"
1616

17-
struct ur_adapter_handle_t_ {
17+
struct ur_adapter_handle_t_ : ur_handle_t_ {
1818
std::atomic<uint32_t> RefCount = 0;
1919
std::mutex Mutex;
2020
struct cuda_tracing_context_t_ *TracingCtx = nullptr;
@@ -41,7 +41,8 @@ class ur_legacy_sink : public logger::Sink {
4141
// through UR entry points.
4242
// https://github.com/oneapi-src/unified-runtime/issues/1330
4343
ur_adapter_handle_t_::ur_adapter_handle_t_()
44-
: logger(logger::get_logger("cuda",
44+
: ur_handle_t_(),
45+
logger(logger::get_logger("cuda",
4546
/*default_log_level*/ logger::Level::ERR)) {
4647

4748
if (std::getenv("UR_LOG_CUDA") != nullptr)

‎unified-runtime/source/adapters/cuda/common.hpp

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -93,3 +93,8 @@ inline umf_result_t setCUMemoryProviderParams(
9393
}
9494

9595
} // namespace umf
96+
97+
struct cuda_ddi_getter {
98+
const static ur_dditable_t *value();
99+
};
100+
using ur_handle_t_ = ur_handle_base_t_<cuda_ddi_getter>;

‎unified-runtime/source/adapters/cuda/context.hpp

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -108,7 +108,7 @@ CreateHostMemoryProviderPool(ur_device_handle_t_ *DeviceHandle,
108108
return UR_RESULT_SUCCESS;
109109
}
110110

111-
struct ur_context_handle_t_ {
111+
struct ur_context_handle_t_ : ur_handle_t_ {
112112

113113
struct deleter_data {
114114
ur_context_extended_deleter_t Function;
@@ -126,7 +126,7 @@ struct ur_context_handle_t_ {
126126
umf_memory_pool_handle_t MemoryPoolHost = nullptr;
127127

128128
ur_context_handle_t_(const ur_device_handle_t *Devs, uint32_t NumDevices)
129-
: Devices{Devs, Devs + NumDevices}, RefCount{1} {
129+
: ur_handle_t_(), Devices{Devs, Devs + NumDevices}, RefCount{1} {
130130
for (auto &Dev : Devices) {
131131
urDeviceRetain(Dev);
132132
}

‎unified-runtime/source/adapters/cuda/device.hpp

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -16,7 +16,7 @@
1616

1717
#include "common.hpp"
1818

19-
struct ur_device_handle_t_ {
19+
struct ur_device_handle_t_ : ur_handle_t_ {
2020
private:
2121
using native_type = CUdevice;
2222

@@ -40,8 +40,8 @@ struct ur_device_handle_t_ {
4040
public:
4141
ur_device_handle_t_(native_type cuDevice, CUcontext cuContext, CUevent evBase,
4242
ur_platform_handle_t platform, uint32_t DevIndex)
43-
: CuDevice(cuDevice), CuContext(cuContext), EvBase(evBase), RefCount{1},
44-
Platform(platform), DeviceIndex{DevIndex} {
43+
: ur_handle_t_(), CuDevice(cuDevice), CuContext(cuContext),
44+
EvBase(evBase), RefCount{1}, Platform(platform), DeviceIndex{DevIndex} {
4545

4646
UR_CHECK_ERROR(cuDeviceGetAttribute(
4747
&MaxRegsPerBlock, CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK,

‎unified-runtime/source/adapters/cuda/event.cpp

Lines changed: 7 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -24,7 +24,7 @@ ur_event_handle_t_::ur_event_handle_t_(ur_command_t Type,
2424
native_type EvEnd, native_type EvQueued,
2525
native_type EvStart, CUstream Stream,
2626
uint32_t StreamToken)
27-
: CommandType{Type}, RefCount{1}, HasOwnership{true},
27+
: ur_handle_t_(), CommandType{Type}, RefCount{1}, HasOwnership{true},
2828
HasBeenWaitedOn{false}, IsRecorded{false}, IsStarted{false},
2929
StreamToken{StreamToken}, EventID{0}, EvEnd{EvEnd}, EvStart{EvStart},
3030
EvQueued{EvQueued}, Queue{Queue}, Stream{Stream}, Context{Context} {
@@ -34,11 +34,12 @@ ur_event_handle_t_::ur_event_handle_t_(ur_command_t Type,
3434

3535
ur_event_handle_t_::ur_event_handle_t_(ur_context_handle_t Context,
3636
CUevent EventNative)
37-
: CommandType{UR_COMMAND_EVENTS_WAIT}, RefCount{1}, HasOwnership{false},
38-
HasBeenWaitedOn{false}, IsRecorded{false}, IsStarted{false},
39-
IsInterop{true}, StreamToken{std::numeric_limits<uint32_t>::max()},
40-
EventID{0}, EvEnd{EventNative}, EvStart{nullptr}, EvQueued{nullptr},
41-
Queue{nullptr}, Stream{nullptr}, Context{Context} {
37+
: ur_handle_t_(), CommandType{UR_COMMAND_EVENTS_WAIT}, RefCount{1},
38+
HasOwnership{false}, HasBeenWaitedOn{false}, IsRecorded{false},
39+
IsStarted{false}, IsInterop{true},
40+
StreamToken{std::numeric_limits<uint32_t>::max()}, EventID{0},
41+
EvEnd{EventNative}, EvStart{nullptr}, EvQueued{nullptr}, Queue{nullptr},
42+
Stream{nullptr}, Context{Context} {
4243
urContextRetain(Context);
4344
}
4445

‎unified-runtime/source/adapters/cuda/event.hpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -17,7 +17,7 @@
1717

1818
/// UR Event mapping to CUevent
1919
///
20-
struct ur_event_handle_t_ {
20+
struct ur_event_handle_t_ : ur_handle_t_ {
2121
public:
2222
using native_type = CUevent;
2323

‎unified-runtime/source/adapters/cuda/kernel.hpp

Lines changed: 4 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -34,7 +34,7 @@
3434
/// A compiler pass converts the UR API local memory model into the
3535
/// CUDA shared model. This object simply calculates the total of
3636
/// shared memory, and the initial offsets of each parameter.
37-
struct ur_kernel_handle_t_ {
37+
struct ur_kernel_handle_t_ : ur_handle_t_ {
3838
using native_type = CUfunction;
3939

4040
native_type Function;
@@ -250,8 +250,9 @@ struct ur_kernel_handle_t_ {
250250
ur_kernel_handle_t_(CUfunction Func, CUfunction FuncWithOffsetParam,
251251
const char *Name, ur_program_handle_t Program,
252252
ur_context_handle_t Context)
253-
: Function{Func}, FunctionWithOffsetParam{FuncWithOffsetParam},
254-
Name{Name}, Context{Context}, Program{Program}, RefCount{1} {
253+
: ur_handle_t_(), Function{Func},
254+
FunctionWithOffsetParam{FuncWithOffsetParam}, Name{Name},
255+
Context{Context}, Program{Program}, RefCount{1} {
255256
urProgramRetain(Program);
256257
urContextRetain(Context);
257258
/// Note: this code assumes that there is only one device per context

‎unified-runtime/source/adapters/cuda/memory.hpp

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -310,7 +310,7 @@ struct SurfaceMem {
310310
/// is on a different device, marked by
311311
/// LastQueueWritingToMemObj->getDevice()
312312
///
313-
struct ur_mem_handle_t_ {
313+
struct ur_mem_handle_t_ : ur_handle_t_ {
314314
// Context where the memory object is accessible
315315
ur_context_handle_t Context;
316316

@@ -345,17 +345,17 @@ struct ur_mem_handle_t_ {
345345
/// Constructs the UR mem handler for a non-typed allocation ("buffer")
346346
ur_mem_handle_t_(ur_context_handle_t Ctxt, ur_mem_flags_t MemFlags,
347347
BufferMem::AllocMode Mode, void *HostPtr, size_t Size)
348-
: Context{Ctxt}, RefCount{1}, MemFlags{MemFlags},
348+
: ur_handle_t_(), Context{Ctxt}, RefCount{1}, MemFlags{MemFlags},
349349
HaveMigratedToDeviceSinceLastWrite(Context->Devices.size(), false),
350350
Mem{std::in_place_type<BufferMem>, Ctxt, this, Mode, HostPtr, Size} {
351351
urContextRetain(Context);
352352
};
353353

354354
// Subbuffer constructor
355355
ur_mem_handle_t_(ur_mem_handle_t Parent, size_t SubBufferOffset)
356-
: Context{Parent->Context}, RefCount{1}, MemFlags{Parent->MemFlags},
357-
HaveMigratedToDeviceSinceLastWrite(Parent->Context->Devices.size(),
358-
false),
356+
: ur_handle_t_(), Context{Parent->Context}, RefCount{1},
357+
MemFlags{Parent->MemFlags}, HaveMigratedToDeviceSinceLastWrite(
358+
Parent->Context->Devices.size(), false),
359359
Mem{BufferMem{std::get<BufferMem>(Parent->Mem)}} {
360360
auto &SubBuffer = std::get<BufferMem>(Mem);
361361
SubBuffer.Parent = Parent;
@@ -376,7 +376,7 @@ struct ur_mem_handle_t_ {
376376
ur_mem_handle_t_(ur_context_handle_t Ctxt, ur_mem_flags_t MemFlags,
377377
ur_image_format_t ImageFormat, ur_image_desc_t ImageDesc,
378378
void *HostPtr)
379-
: Context{Ctxt}, RefCount{1}, MemFlags{MemFlags},
379+
: ur_handle_t_(), Context{Ctxt}, RefCount{1}, MemFlags{MemFlags},
380380
HaveMigratedToDeviceSinceLastWrite(Context->Devices.size(), false),
381381
Mem{std::in_place_type<SurfaceMem>,
382382
Ctxt,

‎unified-runtime/source/adapters/cuda/physical_mem.hpp

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -20,7 +20,7 @@
2020
/// UR queue mapping on physical memory allocations used in virtual memory
2121
/// management.
2222
///
23-
struct ur_physical_mem_handle_t_ {
23+
struct ur_physical_mem_handle_t_ : ur_handle_t_ {
2424
using native_type = CUmemGenericAllocationHandle;
2525

2626
std::atomic_uint32_t RefCount;
@@ -33,8 +33,8 @@ struct ur_physical_mem_handle_t_ {
3333
ur_physical_mem_handle_t_(native_type PhysMem, ur_context_handle_t_ *Ctx,
3434
ur_device_handle_t Device, size_t Size,
3535
ur_physical_mem_properties_t Properties)
36-
: RefCount(1), PhysicalMem(PhysMem), Context(Ctx), Device(Device),
37-
Size(Size), Properties(Properties) {
36+
: ur_handle_t_(), RefCount(1), PhysicalMem(PhysMem), Context(Ctx),
37+
Device(Device), Size(Size), Properties(Properties) {
3838
urContextRetain(Context);
3939
urDeviceRetain(Device);
4040
}

‎unified-runtime/source/adapters/cuda/platform.hpp

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -9,9 +9,10 @@
99
//===----------------------------------------------------------------------===//
1010
#pragma once
1111

12+
#include "common.hpp"
1213
#include <ur/ur.hpp>
1314
#include <vector>
1415

15-
struct ur_platform_handle_t_ {
16+
struct ur_platform_handle_t_ : ur_handle_t_ {
1617
std::vector<std::unique_ptr<ur_device_handle_t_>> Devices;
1718
};

‎unified-runtime/source/adapters/cuda/program.hpp

Lines changed: 5 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -17,7 +17,7 @@
1717

1818
#include "context.hpp"
1919

20-
struct ur_program_handle_t_ {
20+
struct ur_program_handle_t_ : ur_handle_t_ {
2121
using native_type = CUmodule;
2222
native_type Module;
2323
const char *Binary;
@@ -47,9 +47,10 @@ struct ur_program_handle_t_ {
4747
ur_program_build_status_t BuildStatus = UR_PROGRAM_BUILD_STATUS_NONE;
4848

4949
ur_program_handle_t_(ur_context_handle_t Context, ur_device_handle_t Device)
50-
: Module{nullptr}, Binary{}, BinarySizeInBytes{0}, RefCount{1},
51-
Context{Context}, Device{Device}, KernelReqdWorkGroupSizeMD{},
52-
KernelMaxWorkGroupSizeMD{}, KernelMaxLinearWorkGroupSizeMD{} {
50+
: ur_handle_t_(), Module{nullptr}, Binary{}, BinarySizeInBytes{0},
51+
RefCount{1}, Context{Context}, Device{Device},
52+
KernelReqdWorkGroupSizeMD{}, KernelMaxWorkGroupSizeMD{},
53+
KernelMaxLinearWorkGroupSizeMD{} {
5354
urContextRetain(Context);
5455
urDeviceRetain(Device);
5556
}

0 commit comments

Comments
 (0)