triton-windows 3.3.1.post19__cp311-cp311-win_amd64.whl → 3.5.0.post21__cp311-cp311-win_amd64.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Potentially problematic release.
This version of triton-windows might be problematic. Click here for more details.
- triton/_C/libtriton.pyd +0 -0
- triton/__init__.py +11 -2
- triton/_filecheck.py +97 -0
- triton/_internal_testing.py +95 -18
- triton/_utils.py +112 -21
- triton/backends/__init__.py +20 -23
- triton/backends/amd/__init__.py +0 -0
- triton/backends/amd/compiler.py +161 -119
- triton/backends/amd/driver.c +118 -46
- triton/backends/amd/driver.py +274 -96
- triton/backends/compiler.py +7 -21
- triton/backends/driver.py +13 -0
- triton/backends/nvidia/bin/ptxas.exe +0 -0
- triton/backends/nvidia/compiler.py +163 -106
- triton/backends/nvidia/driver.c +166 -101
- triton/backends/nvidia/driver.py +384 -202
- triton/compiler/__init__.py +5 -2
- triton/compiler/code_generator.py +439 -231
- triton/compiler/compiler.py +152 -84
- triton/experimental/__init__.py +0 -0
- triton/experimental/gluon/__init__.py +5 -0
- triton/experimental/gluon/_compiler.py +0 -0
- triton/experimental/gluon/_runtime.py +102 -0
- triton/experimental/gluon/language/__init__.py +119 -0
- triton/experimental/gluon/language/_core.py +490 -0
- triton/experimental/gluon/language/_layouts.py +583 -0
- triton/experimental/gluon/language/_math.py +20 -0
- triton/experimental/gluon/language/_semantic.py +380 -0
- triton/experimental/gluon/language/_standard.py +80 -0
- triton/experimental/gluon/language/amd/__init__.py +4 -0
- triton/experimental/gluon/language/amd/_layouts.py +96 -0
- triton/experimental/gluon/language/amd/cdna3/__init__.py +100 -0
- triton/experimental/gluon/language/amd/cdna4/__init__.py +48 -0
- triton/experimental/gluon/language/amd/cdna4/async_copy.py +151 -0
- triton/experimental/gluon/language/extra/__init__.py +3 -0
- triton/experimental/gluon/language/nvidia/__init__.py +4 -0
- triton/experimental/gluon/language/nvidia/ampere/__init__.py +3 -0
- triton/experimental/gluon/language/nvidia/ampere/async_copy.py +74 -0
- triton/experimental/gluon/language/nvidia/ampere/mbarrier.py +80 -0
- triton/experimental/gluon/language/nvidia/blackwell/__init__.py +387 -0
- triton/experimental/gluon/language/nvidia/blackwell/tma.py +52 -0
- triton/experimental/gluon/language/nvidia/hopper/__init__.py +132 -0
- triton/experimental/gluon/language/nvidia/hopper/mbarrier.py +34 -0
- triton/experimental/gluon/language/nvidia/hopper/tma.py +97 -0
- triton/experimental/gluon/nvidia/__init__.py +4 -0
- triton/experimental/gluon/nvidia/blackwell.py +3 -0
- triton/experimental/gluon/nvidia/hopper.py +45 -0
- triton/knobs.py +546 -0
- triton/language/__init__.py +50 -19
- triton/language/core.py +909 -572
- triton/language/extra/cuda/__init__.py +10 -7
- triton/language/extra/cuda/gdc.py +42 -0
- triton/language/extra/cuda/libdevice.py +394 -394
- triton/language/extra/cuda/utils.py +21 -21
- triton/language/extra/hip/__init__.py +3 -1
- triton/language/extra/hip/libdevice.py +120 -104
- triton/language/extra/hip/utils.py +35 -0
- triton/language/extra/libdevice.py +4 -0
- triton/language/math.py +65 -66
- triton/language/random.py +12 -2
- triton/language/semantic.py +1757 -1768
- triton/language/standard.py +127 -62
- triton/language/target_info.py +54 -0
- triton/runtime/_allocation.py +15 -3
- triton/runtime/_async_compile.py +55 -0
- triton/runtime/autotuner.py +117 -60
- triton/runtime/build.py +83 -17
- triton/runtime/cache.py +61 -47
- triton/runtime/driver.py +25 -47
- triton/runtime/interpreter.py +95 -50
- triton/runtime/jit.py +445 -248
- triton/runtime/tcc/include/_mingw.h +8 -10
- triton/runtime/tcc/include/assert.h +5 -0
- triton/runtime/tcc/include/errno.h +1 -1
- triton/runtime/tcc/include/float.h +21 -3
- triton/runtime/tcc/include/iso646.h +36 -0
- triton/runtime/tcc/include/limits.h +5 -0
- triton/runtime/tcc/include/malloc.h +2 -2
- triton/runtime/tcc/include/math.h +21 -261
- triton/runtime/tcc/include/stdalign.h +16 -0
- triton/runtime/tcc/include/stdarg.h +5 -70
- triton/runtime/tcc/include/stdatomic.h +171 -0
- triton/runtime/tcc/include/stddef.h +7 -19
- triton/runtime/tcc/include/stdlib.h +15 -4
- triton/runtime/tcc/include/stdnoreturn.h +7 -0
- triton/runtime/tcc/include/sys/stat.h +2 -2
- triton/runtime/tcc/include/sys/types.h +5 -0
- triton/runtime/tcc/include/tcc/tcc_libm.h +444 -27
- triton/runtime/tcc/include/tccdefs.h +342 -0
- triton/runtime/tcc/include/tgmath.h +89 -0
- triton/runtime/tcc/include/uchar.h +33 -0
- triton/runtime/tcc/include/unistd.h +1 -0
- triton/runtime/tcc/include/winapi/qos.h +72 -0
- triton/runtime/tcc/include/winapi/shellapi.h +59 -0
- triton/runtime/tcc/include/winapi/winbase.h +9 -2
- triton/runtime/tcc/include/winapi/wincon.h +8 -0
- triton/runtime/tcc/include/winapi/windows.h +1 -1
- triton/runtime/tcc/include/winapi/winnls.h +778 -0
- triton/runtime/tcc/include/winapi/winnt.h +9 -7
- triton/runtime/tcc/include/winapi/winsock2.h +1474 -0
- triton/runtime/tcc/include/winapi/ws2ipdef.h +21 -0
- triton/runtime/tcc/include/winapi/ws2tcpip.h +391 -0
- triton/runtime/tcc/lib/libtcc1.a +0 -0
- triton/runtime/tcc/lib/python314.def +1800 -0
- triton/runtime/tcc/lib/python314t.def +1809 -0
- triton/runtime/tcc/libtcc.dll +0 -0
- triton/runtime/tcc/tcc.exe +0 -0
- triton/testing.py +16 -12
- triton/tools/compile.py +62 -14
- triton/tools/disasm.py +3 -4
- triton/tools/extra/cuda/compile.c +1 -0
- triton/tools/extra/hip/compile.cpp +66 -0
- triton/tools/extra/hip/compile.h +13 -0
- triton/tools/ragged_tma.py +92 -0
- triton/tools/tensor_descriptor.py +34 -0
- triton/windows_utils.py +52 -81
- {triton_windows-3.3.1.post19.dist-info → triton_windows-3.5.0.post21.dist-info}/METADATA +8 -4
- triton_windows-3.5.0.post21.dist-info/RECORD +217 -0
- triton_windows-3.5.0.post21.dist-info/entry_points.txt +3 -0
- triton_windows-3.5.0.post21.dist-info/licenses/LICENSE +23 -0
- triton_windows-3.5.0.post21.dist-info/top_level.txt +1 -0
- triton/backends/amd/include/hip/amd_detail/amd_channel_descriptor.h +0 -358
- triton/backends/amd/include/hip/amd_detail/amd_device_functions.h +0 -1010
- triton/backends/amd/include/hip/amd_detail/amd_hip_atomic.h +0 -1638
- triton/backends/amd/include/hip/amd_detail/amd_hip_bf16.h +0 -1814
- triton/backends/amd/include/hip/amd_detail/amd_hip_bfloat16.h +0 -293
- triton/backends/amd/include/hip/amd_detail/amd_hip_common.h +0 -32
- triton/backends/amd/include/hip/amd_detail/amd_hip_complex.h +0 -174
- triton/backends/amd/include/hip/amd_detail/amd_hip_cooperative_groups.h +0 -835
- triton/backends/amd/include/hip/amd_detail/amd_hip_fp16.h +0 -1809
- triton/backends/amd/include/hip/amd_detail/amd_hip_fp8.h +0 -1391
- triton/backends/amd/include/hip/amd_detail/amd_hip_gl_interop.h +0 -108
- triton/backends/amd/include/hip/amd_detail/amd_hip_math_constants.h +0 -124
- triton/backends/amd/include/hip/amd_detail/amd_hip_runtime.h +0 -405
- triton/backends/amd/include/hip/amd_detail/amd_hip_runtime_pt_api.h +0 -196
- triton/backends/amd/include/hip/amd_detail/amd_hip_unsafe_atomics.h +0 -565
- triton/backends/amd/include/hip/amd_detail/amd_hip_vector_types.h +0 -2226
- triton/backends/amd/include/hip/amd_detail/amd_math_functions.h +0 -104
- triton/backends/amd/include/hip/amd_detail/amd_surface_functions.h +0 -244
- triton/backends/amd/include/hip/amd_detail/amd_warp_functions.h +0 -538
- triton/backends/amd/include/hip/amd_detail/amd_warp_sync_functions.h +0 -288
- triton/backends/amd/include/hip/amd_detail/concepts.hpp +0 -30
- triton/backends/amd/include/hip/amd_detail/device_library_decls.h +0 -133
- triton/backends/amd/include/hip/amd_detail/functional_grid_launch.hpp +0 -218
- triton/backends/amd/include/hip/amd_detail/grid_launch.h +0 -67
- triton/backends/amd/include/hip/amd_detail/grid_launch.hpp +0 -50
- triton/backends/amd/include/hip/amd_detail/grid_launch_GGL.hpp +0 -26
- triton/backends/amd/include/hip/amd_detail/helpers.hpp +0 -137
- triton/backends/amd/include/hip/amd_detail/hip_api_trace.hpp +0 -1446
- triton/backends/amd/include/hip/amd_detail/hip_assert.h +0 -101
- triton/backends/amd/include/hip/amd_detail/hip_cooperative_groups_helper.h +0 -242
- triton/backends/amd/include/hip/amd_detail/hip_fp16_gcc.h +0 -254
- triton/backends/amd/include/hip/amd_detail/hip_fp16_math_fwd.h +0 -96
- triton/backends/amd/include/hip/amd_detail/hip_ldg.h +0 -100
- triton/backends/amd/include/hip/amd_detail/hip_prof_str.h +0 -10570
- triton/backends/amd/include/hip/amd_detail/hip_runtime_prof.h +0 -78
- triton/backends/amd/include/hip/amd_detail/host_defines.h +0 -184
- triton/backends/amd/include/hip/amd_detail/hsa_helpers.hpp +0 -102
- triton/backends/amd/include/hip/amd_detail/macro_based_grid_launch.hpp +0 -798
- triton/backends/amd/include/hip/amd_detail/math_fwd.h +0 -698
- triton/backends/amd/include/hip/amd_detail/ockl_image.h +0 -177
- triton/backends/amd/include/hip/amd_detail/program_state.hpp +0 -107
- triton/backends/amd/include/hip/amd_detail/texture_fetch_functions.h +0 -491
- triton/backends/amd/include/hip/amd_detail/texture_indirect_functions.h +0 -478
- triton/backends/amd/include/hip/channel_descriptor.h +0 -39
- triton/backends/amd/include/hip/device_functions.h +0 -38
- triton/backends/amd/include/hip/driver_types.h +0 -468
- triton/backends/amd/include/hip/hip_bf16.h +0 -36
- triton/backends/amd/include/hip/hip_bfloat16.h +0 -44
- triton/backends/amd/include/hip/hip_common.h +0 -100
- triton/backends/amd/include/hip/hip_complex.h +0 -38
- triton/backends/amd/include/hip/hip_cooperative_groups.h +0 -46
- triton/backends/amd/include/hip/hip_deprecated.h +0 -95
- triton/backends/amd/include/hip/hip_ext.h +0 -161
- triton/backends/amd/include/hip/hip_fp16.h +0 -36
- triton/backends/amd/include/hip/hip_fp8.h +0 -33
- triton/backends/amd/include/hip/hip_gl_interop.h +0 -32
- triton/backends/amd/include/hip/hip_hcc.h +0 -24
- triton/backends/amd/include/hip/hip_math_constants.h +0 -36
- triton/backends/amd/include/hip/hip_profile.h +0 -27
- triton/backends/amd/include/hip/hip_runtime.h +0 -75
- triton/backends/amd/include/hip/hip_runtime_api.h +0 -9261
- triton/backends/amd/include/hip/hip_texture_types.h +0 -29
- triton/backends/amd/include/hip/hip_vector_types.h +0 -41
- triton/backends/amd/include/hip/hip_version.h +0 -17
- triton/backends/amd/include/hip/hiprtc.h +0 -421
- triton/backends/amd/include/hip/library_types.h +0 -78
- triton/backends/amd/include/hip/math_functions.h +0 -42
- triton/backends/amd/include/hip/surface_types.h +0 -63
- triton/backends/amd/include/hip/texture_types.h +0 -194
- triton/backends/amd/include/hsa/Brig.h +0 -1131
- triton/backends/amd/include/hsa/amd_hsa_common.h +0 -91
- triton/backends/amd/include/hsa/amd_hsa_elf.h +0 -462
- triton/backends/amd/include/hsa/amd_hsa_kernel_code.h +0 -269
- triton/backends/amd/include/hsa/amd_hsa_queue.h +0 -109
- triton/backends/amd/include/hsa/amd_hsa_signal.h +0 -80
- triton/backends/amd/include/hsa/hsa.h +0 -5738
- triton/backends/amd/include/hsa/hsa_amd_tool.h +0 -91
- triton/backends/amd/include/hsa/hsa_api_trace.h +0 -579
- triton/backends/amd/include/hsa/hsa_api_trace_version.h +0 -68
- triton/backends/amd/include/hsa/hsa_ext_amd.h +0 -3146
- triton/backends/amd/include/hsa/hsa_ext_finalize.h +0 -531
- triton/backends/amd/include/hsa/hsa_ext_image.h +0 -1454
- triton/backends/amd/include/hsa/hsa_ven_amd_aqlprofile.h +0 -488
- triton/backends/amd/include/hsa/hsa_ven_amd_loader.h +0 -667
- triton/backends/amd/include/hsa/hsa_ven_amd_pc_sampling.h +0 -416
- triton/backends/amd/include/roctracer/ext/prof_protocol.h +0 -107
- triton/backends/amd/include/roctracer/hip_ostream_ops.h +0 -4515
- triton/backends/amd/include/roctracer/hsa_ostream_ops.h +0 -1727
- triton/backends/amd/include/roctracer/hsa_prof_str.h +0 -3059
- triton/backends/amd/include/roctracer/roctracer.h +0 -779
- triton/backends/amd/include/roctracer/roctracer_ext.h +0 -81
- triton/backends/amd/include/roctracer/roctracer_hcc.h +0 -24
- triton/backends/amd/include/roctracer/roctracer_hip.h +0 -37
- triton/backends/amd/include/roctracer/roctracer_hsa.h +0 -112
- triton/backends/amd/include/roctracer/roctracer_plugin.h +0 -137
- triton/backends/amd/include/roctracer/roctracer_roctx.h +0 -67
- triton/backends/amd/include/roctracer/roctx.h +0 -229
- triton/language/_utils.py +0 -21
- triton/language/extra/cuda/_experimental_tma.py +0 -106
- triton/runtime/tcc/lib/libtcc1-64.a +0 -0
- triton/tools/experimental_descriptor.py +0 -32
- triton_windows-3.3.1.post19.dist-info/RECORD +0 -260
- triton_windows-3.3.1.post19.dist-info/top_level.txt +0 -14
- {triton_windows-3.3.1.post19.dist-info → triton_windows-3.5.0.post21.dist-info}/WHEEL +0 -0
|
@@ -2,363 +2,363 @@ from triton.language import core
|
|
|
2
2
|
|
|
3
3
|
|
|
4
4
|
@core.extern
|
|
5
|
-
def clz(arg0,
|
|
5
|
+
def clz(arg0, _semantic=None):
|
|
6
6
|
return core.extern_elementwise(
|
|
7
7
|
"", "", [arg0], {
|
|
8
8
|
(core.dtype("int32"), ): ("__nv_clz", core.dtype("int32")),
|
|
9
9
|
(core.dtype("int64"), ): ("__nv_clzll", core.dtype("int32")),
|
|
10
|
-
}, is_pure=True,
|
|
10
|
+
}, is_pure=True, _semantic=_semantic)
|
|
11
11
|
|
|
12
12
|
|
|
13
13
|
@core.extern
|
|
14
|
-
def popc(arg0,
|
|
14
|
+
def popc(arg0, _semantic=None):
|
|
15
15
|
return core.extern_elementwise(
|
|
16
16
|
"", "", [arg0], {
|
|
17
17
|
(core.dtype("int32"), ): ("__nv_popc", core.dtype("int32")),
|
|
18
18
|
(core.dtype("int64"), ): ("__nv_popcll", core.dtype("int32")),
|
|
19
|
-
}, is_pure=True,
|
|
19
|
+
}, is_pure=True, _semantic=_semantic)
|
|
20
20
|
|
|
21
21
|
|
|
22
22
|
@core.extern
|
|
23
|
-
def byte_perm(arg0, arg1, arg2,
|
|
23
|
+
def byte_perm(arg0, arg1, arg2, _semantic=None):
|
|
24
24
|
return core.extern_elementwise("", "", [arg0, arg1, arg2], {
|
|
25
25
|
(core.dtype("int32"), core.dtype("int32"), core.dtype("int32")): ("__nv_byte_perm", core.dtype("int32")),
|
|
26
|
-
}, is_pure=True,
|
|
26
|
+
}, is_pure=True, _semantic=_semantic)
|
|
27
27
|
|
|
28
28
|
|
|
29
29
|
@core.extern
|
|
30
|
-
def mulhi(arg0, arg1,
|
|
30
|
+
def mulhi(arg0, arg1, _semantic=None):
|
|
31
31
|
return core.extern_elementwise(
|
|
32
32
|
"", "", [arg0, arg1], {
|
|
33
33
|
(core.dtype("int32"), core.dtype("int32")): ("__nv_mulhi", core.dtype("int32")),
|
|
34
34
|
(core.dtype("uint32"), core.dtype("uint32")): ("__nv_umulhi", core.dtype("uint32")),
|
|
35
35
|
(core.dtype("int64"), core.dtype("int64")): ("__nv_mul64hi", core.dtype("int64")),
|
|
36
36
|
(core.dtype("uint64"), core.dtype("uint64")): ("__nv_umul64hi", core.dtype("uint64")),
|
|
37
|
-
}, is_pure=True,
|
|
37
|
+
}, is_pure=True, _semantic=_semantic)
|
|
38
38
|
|
|
39
39
|
|
|
40
40
|
@core.extern
|
|
41
|
-
def mul24(arg0, arg1,
|
|
41
|
+
def mul24(arg0, arg1, _semantic=None):
|
|
42
42
|
return core.extern_elementwise(
|
|
43
43
|
"", "", [arg0, arg1], {
|
|
44
44
|
(core.dtype("int32"), core.dtype("int32")): ("__nv_mul24", core.dtype("int32")),
|
|
45
45
|
(core.dtype("uint32"), core.dtype("uint32")): ("__nv_umul24", core.dtype("uint32")),
|
|
46
|
-
}, is_pure=True,
|
|
46
|
+
}, is_pure=True, _semantic=_semantic)
|
|
47
47
|
|
|
48
48
|
|
|
49
49
|
@core.extern
|
|
50
|
-
def brev(arg0,
|
|
50
|
+
def brev(arg0, _semantic=None):
|
|
51
51
|
return core.extern_elementwise(
|
|
52
52
|
"", "", [arg0], {
|
|
53
53
|
(core.dtype("int32"), ): ("__nv_brev", core.dtype("int32")),
|
|
54
54
|
(core.dtype("int64"), ): ("__nv_brevll", core.dtype("int64")),
|
|
55
|
-
}, is_pure=True,
|
|
55
|
+
}, is_pure=True, _semantic=_semantic)
|
|
56
56
|
|
|
57
57
|
|
|
58
58
|
@core.extern
|
|
59
|
-
def sad(arg0, arg1, arg2,
|
|
59
|
+
def sad(arg0, arg1, arg2, _semantic=None):
|
|
60
60
|
return core.extern_elementwise(
|
|
61
61
|
"", "", [arg0, arg1, arg2], {
|
|
62
62
|
(core.dtype("int32"), core.dtype("int32"), core.dtype("uint32")): ("__nv_sad", core.dtype("int32")),
|
|
63
63
|
(core.dtype("uint32"), core.dtype("uint32"), core.dtype("uint32")): ("__nv_usad", core.dtype("uint32")),
|
|
64
|
-
}, is_pure=True,
|
|
64
|
+
}, is_pure=True, _semantic=_semantic)
|
|
65
65
|
|
|
66
66
|
|
|
67
67
|
@core.extern
|
|
68
|
-
def abs(arg0,
|
|
68
|
+
def abs(arg0, _semantic=None):
|
|
69
69
|
return core.extern_elementwise(
|
|
70
70
|
"", "", [arg0], {
|
|
71
71
|
(core.dtype("int32"), ): ("__nv_abs", core.dtype("int32")),
|
|
72
72
|
(core.dtype("int64"), ): ("__nv_llabs", core.dtype("int64")),
|
|
73
73
|
(core.dtype("fp32"), ): ("__nv_fabsf", core.dtype("fp32")),
|
|
74
74
|
(core.dtype("fp64"), ): ("__nv_fabs", core.dtype("fp64")),
|
|
75
|
-
}, is_pure=True,
|
|
75
|
+
}, is_pure=True, _semantic=_semantic)
|
|
76
76
|
|
|
77
77
|
|
|
78
78
|
@core.extern
|
|
79
|
-
def floor(arg0,
|
|
79
|
+
def floor(arg0, _semantic=None):
|
|
80
80
|
return core.extern_elementwise(
|
|
81
81
|
"", "", [arg0], {
|
|
82
82
|
(core.dtype("fp32"), ): ("__nv_floorf", core.dtype("fp32")),
|
|
83
83
|
(core.dtype("fp64"), ): ("__nv_floor", core.dtype("fp64")),
|
|
84
|
-
}, is_pure=True,
|
|
84
|
+
}, is_pure=True, _semantic=_semantic)
|
|
85
85
|
|
|
86
86
|
|
|
87
87
|
@core.extern
|
|
88
|
-
def rcp64h(arg0,
|
|
88
|
+
def rcp64h(arg0, _semantic=None):
|
|
89
89
|
return core.extern_elementwise("", "", [arg0], {
|
|
90
90
|
(core.dtype("fp64"), ): ("__nv_rcp64h", core.dtype("fp64")),
|
|
91
|
-
}, is_pure=True,
|
|
91
|
+
}, is_pure=True, _semantic=_semantic)
|
|
92
92
|
|
|
93
93
|
|
|
94
94
|
@core.extern
|
|
95
|
-
def rsqrt(arg0,
|
|
95
|
+
def rsqrt(arg0, _semantic=None):
|
|
96
96
|
return core.extern_elementwise(
|
|
97
97
|
"", "", [arg0], {
|
|
98
98
|
(core.dtype("fp32"), ): ("__nv_rsqrtf", core.dtype("fp32")),
|
|
99
99
|
(core.dtype("fp64"), ): ("__nv_rsqrt", core.dtype("fp64")),
|
|
100
|
-
}, is_pure=True,
|
|
100
|
+
}, is_pure=True, _semantic=_semantic)
|
|
101
101
|
|
|
102
102
|
|
|
103
103
|
@core.extern
|
|
104
|
-
def ceil(arg0,
|
|
104
|
+
def ceil(arg0, _semantic=None):
|
|
105
105
|
return core.extern_elementwise(
|
|
106
106
|
"", "", [arg0], {
|
|
107
107
|
(core.dtype("fp64"), ): ("__nv_ceil", core.dtype("fp64")),
|
|
108
108
|
(core.dtype("fp32"), ): ("__nv_ceilf", core.dtype("fp32")),
|
|
109
|
-
}, is_pure=True,
|
|
109
|
+
}, is_pure=True, _semantic=_semantic)
|
|
110
110
|
|
|
111
111
|
|
|
112
112
|
@core.extern
|
|
113
|
-
def trunc(arg0,
|
|
113
|
+
def trunc(arg0, _semantic=None):
|
|
114
114
|
return core.extern_elementwise(
|
|
115
115
|
"", "", [arg0], {
|
|
116
116
|
(core.dtype("fp64"), ): ("__nv_trunc", core.dtype("fp64")),
|
|
117
117
|
(core.dtype("fp32"), ): ("__nv_truncf", core.dtype("fp32")),
|
|
118
|
-
}, is_pure=True,
|
|
118
|
+
}, is_pure=True, _semantic=_semantic)
|
|
119
119
|
|
|
120
120
|
|
|
121
121
|
@core.extern
|
|
122
|
-
def exp2(arg0,
|
|
122
|
+
def exp2(arg0, _semantic=None):
|
|
123
123
|
return core.extern_elementwise(
|
|
124
124
|
"", "", [arg0], {
|
|
125
125
|
(core.dtype("fp32"), ): ("__nv_exp2f", core.dtype("fp32")),
|
|
126
126
|
(core.dtype("fp64"), ): ("__nv_exp2", core.dtype("fp64")),
|
|
127
|
-
}, is_pure=True,
|
|
127
|
+
}, is_pure=True, _semantic=_semantic)
|
|
128
128
|
|
|
129
129
|
|
|
130
130
|
@core.extern
|
|
131
|
-
def saturatef(arg0,
|
|
131
|
+
def saturatef(arg0, _semantic=None):
|
|
132
132
|
return core.extern_elementwise("", "", [arg0], {
|
|
133
133
|
(core.dtype("fp32"), ): ("__nv_saturatef", core.dtype("fp32")),
|
|
134
|
-
}, is_pure=True,
|
|
134
|
+
}, is_pure=True, _semantic=_semantic)
|
|
135
135
|
|
|
136
136
|
|
|
137
137
|
@core.extern
|
|
138
|
-
def fma_rn(arg0, arg1, arg2,
|
|
138
|
+
def fma_rn(arg0, arg1, arg2, _semantic=None):
|
|
139
139
|
return core.extern_elementwise(
|
|
140
140
|
"", "", [arg0, arg1, arg2], {
|
|
141
141
|
(core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32")): ("__nv_fmaf_rn", core.dtype("fp32")),
|
|
142
142
|
(core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64")): ("__nv_fma_rn", core.dtype("fp64")),
|
|
143
|
-
}, is_pure=True,
|
|
143
|
+
}, is_pure=True, _semantic=_semantic)
|
|
144
144
|
|
|
145
145
|
|
|
146
146
|
@core.extern
|
|
147
|
-
def fma_rz(arg0, arg1, arg2,
|
|
147
|
+
def fma_rz(arg0, arg1, arg2, _semantic=None):
|
|
148
148
|
return core.extern_elementwise(
|
|
149
149
|
"", "", [arg0, arg1, arg2], {
|
|
150
150
|
(core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32")): ("__nv_fmaf_rz", core.dtype("fp32")),
|
|
151
151
|
(core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64")): ("__nv_fma_rz", core.dtype("fp64")),
|
|
152
|
-
}, is_pure=True,
|
|
152
|
+
}, is_pure=True, _semantic=_semantic)
|
|
153
153
|
|
|
154
154
|
|
|
155
155
|
@core.extern
|
|
156
|
-
def fma_rd(arg0, arg1, arg2,
|
|
156
|
+
def fma_rd(arg0, arg1, arg2, _semantic=None):
|
|
157
157
|
return core.extern_elementwise(
|
|
158
158
|
"", "", [arg0, arg1, arg2], {
|
|
159
159
|
(core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32")): ("__nv_fmaf_rd", core.dtype("fp32")),
|
|
160
160
|
(core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64")): ("__nv_fma_rd", core.dtype("fp64")),
|
|
161
|
-
}, is_pure=True,
|
|
161
|
+
}, is_pure=True, _semantic=_semantic)
|
|
162
162
|
|
|
163
163
|
|
|
164
164
|
@core.extern
|
|
165
|
-
def fma_ru(arg0, arg1, arg2,
|
|
165
|
+
def fma_ru(arg0, arg1, arg2, _semantic=None):
|
|
166
166
|
return core.extern_elementwise(
|
|
167
167
|
"", "", [arg0, arg1, arg2], {
|
|
168
168
|
(core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32")): ("__nv_fmaf_ru", core.dtype("fp32")),
|
|
169
169
|
(core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64")): ("__nv_fma_ru", core.dtype("fp64")),
|
|
170
|
-
}, is_pure=True,
|
|
170
|
+
}, is_pure=True, _semantic=_semantic)
|
|
171
171
|
|
|
172
172
|
|
|
173
173
|
@core.extern
|
|
174
|
-
def fast_dividef(arg0, arg1,
|
|
174
|
+
def fast_dividef(arg0, arg1, _semantic=None):
|
|
175
175
|
return core.extern_elementwise("", "", [arg0, arg1], {
|
|
176
176
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fast_fdividef", core.dtype("fp32")),
|
|
177
|
-
}, is_pure=True,
|
|
177
|
+
}, is_pure=True, _semantic=_semantic)
|
|
178
178
|
|
|
179
179
|
|
|
180
180
|
@core.extern
|
|
181
|
-
def div_rn(arg0, arg1,
|
|
181
|
+
def div_rn(arg0, arg1, _semantic=None):
|
|
182
182
|
return core.extern_elementwise(
|
|
183
183
|
"", "", [arg0, arg1], {
|
|
184
184
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fdiv_rn", core.dtype("fp32")),
|
|
185
185
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_ddiv_rn", core.dtype("fp64")),
|
|
186
|
-
}, is_pure=True,
|
|
186
|
+
}, is_pure=True, _semantic=_semantic)
|
|
187
187
|
|
|
188
188
|
|
|
189
189
|
@core.extern
|
|
190
|
-
def div_rz(arg0, arg1,
|
|
190
|
+
def div_rz(arg0, arg1, _semantic=None):
|
|
191
191
|
return core.extern_elementwise(
|
|
192
192
|
"", "", [arg0, arg1], {
|
|
193
193
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fdiv_rz", core.dtype("fp32")),
|
|
194
194
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_ddiv_rz", core.dtype("fp64")),
|
|
195
|
-
}, is_pure=True,
|
|
195
|
+
}, is_pure=True, _semantic=_semantic)
|
|
196
196
|
|
|
197
197
|
|
|
198
198
|
@core.extern
|
|
199
|
-
def div_rd(arg0, arg1,
|
|
199
|
+
def div_rd(arg0, arg1, _semantic=None):
|
|
200
200
|
return core.extern_elementwise(
|
|
201
201
|
"", "", [arg0, arg1], {
|
|
202
202
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fdiv_rd", core.dtype("fp32")),
|
|
203
203
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_ddiv_rd", core.dtype("fp64")),
|
|
204
|
-
}, is_pure=True,
|
|
204
|
+
}, is_pure=True, _semantic=_semantic)
|
|
205
205
|
|
|
206
206
|
|
|
207
207
|
@core.extern
|
|
208
|
-
def div_ru(arg0, arg1,
|
|
208
|
+
def div_ru(arg0, arg1, _semantic=None):
|
|
209
209
|
return core.extern_elementwise(
|
|
210
210
|
"", "", [arg0, arg1], {
|
|
211
211
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fdiv_ru", core.dtype("fp32")),
|
|
212
212
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_ddiv_ru", core.dtype("fp64")),
|
|
213
|
-
}, is_pure=True,
|
|
213
|
+
}, is_pure=True, _semantic=_semantic)
|
|
214
214
|
|
|
215
215
|
|
|
216
216
|
@core.extern
|
|
217
|
-
def rcp_rn(arg0,
|
|
217
|
+
def rcp_rn(arg0, _semantic=None):
|
|
218
218
|
return core.extern_elementwise(
|
|
219
219
|
"", "", [arg0], {
|
|
220
220
|
(core.dtype("fp32"), ): ("__nv_frcp_rn", core.dtype("fp32")),
|
|
221
221
|
(core.dtype("fp64"), ): ("__nv_drcp_rn", core.dtype("fp64")),
|
|
222
|
-
}, is_pure=True,
|
|
222
|
+
}, is_pure=True, _semantic=_semantic)
|
|
223
223
|
|
|
224
224
|
|
|
225
225
|
@core.extern
|
|
226
|
-
def rcp_rz(arg0,
|
|
226
|
+
def rcp_rz(arg0, _semantic=None):
|
|
227
227
|
return core.extern_elementwise(
|
|
228
228
|
"", "", [arg0], {
|
|
229
229
|
(core.dtype("fp32"), ): ("__nv_frcp_rz", core.dtype("fp32")),
|
|
230
230
|
(core.dtype("fp64"), ): ("__nv_drcp_rz", core.dtype("fp64")),
|
|
231
|
-
}, is_pure=True,
|
|
231
|
+
}, is_pure=True, _semantic=_semantic)
|
|
232
232
|
|
|
233
233
|
|
|
234
234
|
@core.extern
|
|
235
|
-
def rcp_rd(arg0,
|
|
235
|
+
def rcp_rd(arg0, _semantic=None):
|
|
236
236
|
return core.extern_elementwise(
|
|
237
237
|
"", "", [arg0], {
|
|
238
238
|
(core.dtype("fp32"), ): ("__nv_frcp_rd", core.dtype("fp32")),
|
|
239
239
|
(core.dtype("fp64"), ): ("__nv_drcp_rd", core.dtype("fp64")),
|
|
240
|
-
}, is_pure=True,
|
|
240
|
+
}, is_pure=True, _semantic=_semantic)
|
|
241
241
|
|
|
242
242
|
|
|
243
243
|
@core.extern
|
|
244
|
-
def rcp_ru(arg0,
|
|
244
|
+
def rcp_ru(arg0, _semantic=None):
|
|
245
245
|
return core.extern_elementwise(
|
|
246
246
|
"", "", [arg0], {
|
|
247
247
|
(core.dtype("fp32"), ): ("__nv_frcp_ru", core.dtype("fp32")),
|
|
248
248
|
(core.dtype("fp64"), ): ("__nv_drcp_ru", core.dtype("fp64")),
|
|
249
|
-
}, is_pure=True,
|
|
249
|
+
}, is_pure=True, _semantic=_semantic)
|
|
250
250
|
|
|
251
251
|
|
|
252
252
|
@core.extern
|
|
253
|
-
def sqrt_rn(arg0,
|
|
253
|
+
def sqrt_rn(arg0, _semantic=None):
|
|
254
254
|
return core.extern_elementwise(
|
|
255
255
|
"", "", [arg0], {
|
|
256
256
|
(core.dtype("fp32"), ): ("__nv_fsqrt_rn", core.dtype("fp32")),
|
|
257
257
|
(core.dtype("fp64"), ): ("__nv_dsqrt_rn", core.dtype("fp64")),
|
|
258
|
-
}, is_pure=True,
|
|
258
|
+
}, is_pure=True, _semantic=_semantic)
|
|
259
259
|
|
|
260
260
|
|
|
261
261
|
@core.extern
|
|
262
|
-
def sqrt_rz(arg0,
|
|
262
|
+
def sqrt_rz(arg0, _semantic=None):
|
|
263
263
|
return core.extern_elementwise(
|
|
264
264
|
"", "", [arg0], {
|
|
265
265
|
(core.dtype("fp32"), ): ("__nv_fsqrt_rz", core.dtype("fp32")),
|
|
266
266
|
(core.dtype("fp64"), ): ("__nv_dsqrt_rz", core.dtype("fp64")),
|
|
267
|
-
}, is_pure=True,
|
|
267
|
+
}, is_pure=True, _semantic=_semantic)
|
|
268
268
|
|
|
269
269
|
|
|
270
270
|
@core.extern
|
|
271
|
-
def sqrt_rd(arg0,
|
|
271
|
+
def sqrt_rd(arg0, _semantic=None):
|
|
272
272
|
return core.extern_elementwise(
|
|
273
273
|
"", "", [arg0], {
|
|
274
274
|
(core.dtype("fp32"), ): ("__nv_fsqrt_rd", core.dtype("fp32")),
|
|
275
275
|
(core.dtype("fp64"), ): ("__nv_dsqrt_rd", core.dtype("fp64")),
|
|
276
|
-
}, is_pure=True,
|
|
276
|
+
}, is_pure=True, _semantic=_semantic)
|
|
277
277
|
|
|
278
278
|
|
|
279
279
|
@core.extern
|
|
280
|
-
def sqrt_ru(arg0,
|
|
280
|
+
def sqrt_ru(arg0, _semantic=None):
|
|
281
281
|
return core.extern_elementwise(
|
|
282
282
|
"", "", [arg0], {
|
|
283
283
|
(core.dtype("fp32"), ): ("__nv_fsqrt_ru", core.dtype("fp32")),
|
|
284
284
|
(core.dtype("fp64"), ): ("__nv_dsqrt_ru", core.dtype("fp64")),
|
|
285
|
-
}, is_pure=True,
|
|
285
|
+
}, is_pure=True, _semantic=_semantic)
|
|
286
286
|
|
|
287
287
|
|
|
288
288
|
@core.extern
|
|
289
|
-
def sqrt(arg0,
|
|
289
|
+
def sqrt(arg0, _semantic=None):
|
|
290
290
|
return core.extern_elementwise(
|
|
291
291
|
"", "", [arg0], {
|
|
292
292
|
(core.dtype("fp32"), ): ("__nv_sqrtf", core.dtype("fp32")),
|
|
293
293
|
(core.dtype("fp64"), ): ("__nv_sqrt", core.dtype("fp64")),
|
|
294
|
-
}, is_pure=True,
|
|
294
|
+
}, is_pure=True, _semantic=_semantic)
|
|
295
295
|
|
|
296
296
|
|
|
297
297
|
@core.extern
|
|
298
|
-
def add_rn(arg0, arg1,
|
|
298
|
+
def add_rn(arg0, arg1, _semantic=None):
|
|
299
299
|
return core.extern_elementwise(
|
|
300
300
|
"", "", [arg0, arg1], {
|
|
301
301
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dadd_rn", core.dtype("fp64")),
|
|
302
302
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fadd_rn", core.dtype("fp32")),
|
|
303
|
-
}, is_pure=True,
|
|
303
|
+
}, is_pure=True, _semantic=_semantic)
|
|
304
304
|
|
|
305
305
|
|
|
306
306
|
@core.extern
|
|
307
|
-
def add_rz(arg0, arg1,
|
|
307
|
+
def add_rz(arg0, arg1, _semantic=None):
|
|
308
308
|
return core.extern_elementwise(
|
|
309
309
|
"", "", [arg0, arg1], {
|
|
310
310
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dadd_rz", core.dtype("fp64")),
|
|
311
311
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fadd_rz", core.dtype("fp32")),
|
|
312
|
-
}, is_pure=True,
|
|
312
|
+
}, is_pure=True, _semantic=_semantic)
|
|
313
313
|
|
|
314
314
|
|
|
315
315
|
@core.extern
|
|
316
|
-
def add_rd(arg0, arg1,
|
|
316
|
+
def add_rd(arg0, arg1, _semantic=None):
|
|
317
317
|
return core.extern_elementwise(
|
|
318
318
|
"", "", [arg0, arg1], {
|
|
319
319
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dadd_rd", core.dtype("fp64")),
|
|
320
320
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fadd_rd", core.dtype("fp32")),
|
|
321
|
-
}, is_pure=True,
|
|
321
|
+
}, is_pure=True, _semantic=_semantic)
|
|
322
322
|
|
|
323
323
|
|
|
324
324
|
@core.extern
|
|
325
|
-
def add_ru(arg0, arg1,
|
|
325
|
+
def add_ru(arg0, arg1, _semantic=None):
|
|
326
326
|
return core.extern_elementwise(
|
|
327
327
|
"", "", [arg0, arg1], {
|
|
328
328
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dadd_ru", core.dtype("fp64")),
|
|
329
329
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fadd_ru", core.dtype("fp32")),
|
|
330
|
-
}, is_pure=True,
|
|
330
|
+
}, is_pure=True, _semantic=_semantic)
|
|
331
331
|
|
|
332
332
|
|
|
333
333
|
@core.extern
|
|
334
|
-
def mul_rn(arg0, arg1,
|
|
334
|
+
def mul_rn(arg0, arg1, _semantic=None):
|
|
335
335
|
return core.extern_elementwise(
|
|
336
336
|
"", "", [arg0, arg1], {
|
|
337
337
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dmul_rn", core.dtype("fp64")),
|
|
338
338
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fmul_rn", core.dtype("fp32")),
|
|
339
|
-
}, is_pure=True,
|
|
339
|
+
}, is_pure=True, _semantic=_semantic)
|
|
340
340
|
|
|
341
341
|
|
|
342
342
|
@core.extern
|
|
343
|
-
def mul_rz(arg0, arg1,
|
|
343
|
+
def mul_rz(arg0, arg1, _semantic=None):
|
|
344
344
|
return core.extern_elementwise(
|
|
345
345
|
"", "", [arg0, arg1], {
|
|
346
346
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dmul_rz", core.dtype("fp64")),
|
|
347
347
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fmul_rz", core.dtype("fp32")),
|
|
348
|
-
}, is_pure=True,
|
|
348
|
+
}, is_pure=True, _semantic=_semantic)
|
|
349
349
|
|
|
350
350
|
|
|
351
351
|
@core.extern
|
|
352
|
-
def mul_rd(arg0, arg1,
|
|
352
|
+
def mul_rd(arg0, arg1, _semantic=None):
|
|
353
353
|
return core.extern_elementwise(
|
|
354
354
|
"", "", [arg0, arg1], {
|
|
355
355
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dmul_rd", core.dtype("fp64")),
|
|
356
356
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fmul_rd", core.dtype("fp32")),
|
|
357
|
-
}, is_pure=True,
|
|
357
|
+
}, is_pure=True, _semantic=_semantic)
|
|
358
358
|
|
|
359
359
|
|
|
360
360
|
@core.extern
|
|
361
|
-
def mul_ru(arg0, arg1,
|
|
361
|
+
def mul_ru(arg0, arg1, _semantic=None):
|
|
362
362
|
return core.extern_elementwise(
|
|
363
363
|
"", "", [
|
|
364
364
|
arg0,
|
|
@@ -372,1258 +372,1258 @@ def mul_ru(arg0, arg1, _builder=None):
|
|
|
372
372
|
core.dtype("fp32"),
|
|
373
373
|
core.dtype("fp32"),
|
|
374
374
|
): ("__nv_fmul_ru", core.dtype("fp32")),
|
|
375
|
-
}, is_pure=True,
|
|
375
|
+
}, is_pure=True, _semantic=_semantic)
|
|
376
376
|
|
|
377
377
|
|
|
378
378
|
@core.extern
|
|
379
|
-
def double2float_rn(arg0,
|
|
379
|
+
def double2float_rn(arg0, _semantic=None):
|
|
380
380
|
return core.extern_elementwise("", "", [arg0], {
|
|
381
381
|
(core.dtype("fp64"), ): ("__nv_double2float_rn", core.dtype("fp32")),
|
|
382
|
-
}, is_pure=True,
|
|
382
|
+
}, is_pure=True, _semantic=_semantic)
|
|
383
383
|
|
|
384
384
|
|
|
385
385
|
@core.extern
|
|
386
|
-
def double2float_rz(arg0,
|
|
386
|
+
def double2float_rz(arg0, _semantic=None):
|
|
387
387
|
return core.extern_elementwise("", "", [arg0], {
|
|
388
388
|
(core.dtype("fp64"), ): ("__nv_double2float_rz", core.dtype("fp32")),
|
|
389
|
-
}, is_pure=True,
|
|
389
|
+
}, is_pure=True, _semantic=_semantic)
|
|
390
390
|
|
|
391
391
|
|
|
392
392
|
@core.extern
|
|
393
|
-
def double2float_rd(arg0,
|
|
393
|
+
def double2float_rd(arg0, _semantic=None):
|
|
394
394
|
return core.extern_elementwise("", "", [arg0], {
|
|
395
395
|
(core.dtype("fp64"), ): ("__nv_double2float_rd", core.dtype("fp32")),
|
|
396
|
-
}, is_pure=True,
|
|
396
|
+
}, is_pure=True, _semantic=_semantic)
|
|
397
397
|
|
|
398
398
|
|
|
399
399
|
@core.extern
|
|
400
|
-
def double2float_ru(arg0,
|
|
400
|
+
def double2float_ru(arg0, _semantic=None):
|
|
401
401
|
return core.extern_elementwise("", "", [arg0], {
|
|
402
402
|
(core.dtype("fp64"), ): ("__nv_double2float_ru", core.dtype("fp32")),
|
|
403
|
-
}, is_pure=True,
|
|
403
|
+
}, is_pure=True, _semantic=_semantic)
|
|
404
404
|
|
|
405
405
|
|
|
406
406
|
@core.extern
|
|
407
|
-
def double2int_rn(arg0,
|
|
407
|
+
def double2int_rn(arg0, _semantic=None):
|
|
408
408
|
return core.extern_elementwise("", "", [arg0], {
|
|
409
409
|
(core.dtype("fp64"), ): ("__nv_double2int_rn", core.dtype("int32")),
|
|
410
|
-
}, is_pure=True,
|
|
410
|
+
}, is_pure=True, _semantic=_semantic)
|
|
411
411
|
|
|
412
412
|
|
|
413
413
|
@core.extern
|
|
414
|
-
def double2int_rz(arg0,
|
|
414
|
+
def double2int_rz(arg0, _semantic=None):
|
|
415
415
|
return core.extern_elementwise("", "", [arg0], {
|
|
416
416
|
(core.dtype("fp64"), ): ("__nv_double2int_rz", core.dtype("int32")),
|
|
417
|
-
}, is_pure=True,
|
|
417
|
+
}, is_pure=True, _semantic=_semantic)
|
|
418
418
|
|
|
419
419
|
|
|
420
420
|
@core.extern
|
|
421
|
-
def double2int_rd(arg0,
|
|
421
|
+
def double2int_rd(arg0, _semantic=None):
|
|
422
422
|
return core.extern_elementwise("", "", [arg0], {
|
|
423
423
|
(core.dtype("fp64"), ): ("__nv_double2int_rd", core.dtype("int32")),
|
|
424
|
-
}, is_pure=True,
|
|
424
|
+
}, is_pure=True, _semantic=_semantic)
|
|
425
425
|
|
|
426
426
|
|
|
427
427
|
@core.extern
|
|
428
|
-
def double2int_ru(arg0,
|
|
428
|
+
def double2int_ru(arg0, _semantic=None):
|
|
429
429
|
return core.extern_elementwise("", "", [arg0], {
|
|
430
430
|
(core.dtype("fp64"), ): ("__nv_double2int_ru", core.dtype("int32")),
|
|
431
|
-
}, is_pure=True,
|
|
431
|
+
}, is_pure=True, _semantic=_semantic)
|
|
432
432
|
|
|
433
433
|
|
|
434
434
|
@core.extern
|
|
435
|
-
def double2uint_rn(arg0,
|
|
435
|
+
def double2uint_rn(arg0, _semantic=None):
|
|
436
436
|
return core.extern_elementwise("", "", [arg0], {
|
|
437
437
|
(core.dtype("fp64"), ): ("__nv_double2uint_rn", core.dtype("int32")),
|
|
438
|
-
}, is_pure=True,
|
|
438
|
+
}, is_pure=True, _semantic=_semantic)
|
|
439
439
|
|
|
440
440
|
|
|
441
441
|
@core.extern
|
|
442
|
-
def double2uint_rz(arg0,
|
|
442
|
+
def double2uint_rz(arg0, _semantic=None):
|
|
443
443
|
return core.extern_elementwise("", "", [arg0], {
|
|
444
444
|
(core.dtype("fp64"), ): ("__nv_double2uint_rz", core.dtype("int32")),
|
|
445
|
-
}, is_pure=True,
|
|
445
|
+
}, is_pure=True, _semantic=_semantic)
|
|
446
446
|
|
|
447
447
|
|
|
448
448
|
@core.extern
|
|
449
|
-
def double2uint_rd(arg0,
|
|
449
|
+
def double2uint_rd(arg0, _semantic=None):
|
|
450
450
|
return core.extern_elementwise("", "", [arg0], {
|
|
451
451
|
(core.dtype("fp64"), ): ("__nv_double2uint_rd", core.dtype("int32")),
|
|
452
|
-
}, is_pure=True,
|
|
452
|
+
}, is_pure=True, _semantic=_semantic)
|
|
453
453
|
|
|
454
454
|
|
|
455
455
|
@core.extern
|
|
456
|
-
def double2uint_ru(arg0,
|
|
456
|
+
def double2uint_ru(arg0, _semantic=None):
|
|
457
457
|
return core.extern_elementwise("", "", [arg0], {
|
|
458
458
|
(core.dtype("fp64"), ): ("__nv_double2uint_ru", core.dtype("int32")),
|
|
459
|
-
}, is_pure=True,
|
|
459
|
+
}, is_pure=True, _semantic=_semantic)
|
|
460
460
|
|
|
461
461
|
|
|
462
462
|
@core.extern
|
|
463
|
-
def int2double_rn(arg0,
|
|
463
|
+
def int2double_rn(arg0, _semantic=None):
|
|
464
464
|
return core.extern_elementwise("", "", [arg0], {
|
|
465
465
|
(core.dtype("int32"), ): ("__nv_int2double_rn", core.dtype("fp64")),
|
|
466
|
-
}, is_pure=True,
|
|
466
|
+
}, is_pure=True, _semantic=_semantic)
|
|
467
467
|
|
|
468
468
|
|
|
469
469
|
@core.extern
|
|
470
|
-
def uint2double_rn(arg0,
|
|
470
|
+
def uint2double_rn(arg0, _semantic=None):
|
|
471
471
|
return core.extern_elementwise("", "", [arg0], {
|
|
472
472
|
(core.dtype("uint32"), ): ("__nv_uint2double_rn", core.dtype("fp64")),
|
|
473
|
-
}, is_pure=True,
|
|
473
|
+
}, is_pure=True, _semantic=_semantic)
|
|
474
474
|
|
|
475
475
|
|
|
476
476
|
@core.extern
|
|
477
|
-
def float2int_rn(arg0,
|
|
477
|
+
def float2int_rn(arg0, _semantic=None):
|
|
478
478
|
return core.extern_elementwise("", "", [arg0], {
|
|
479
479
|
(core.dtype("fp32"), ): ("__nv_float2int_rn", core.dtype("int32")),
|
|
480
|
-
}, is_pure=True,
|
|
480
|
+
}, is_pure=True, _semantic=_semantic)
|
|
481
481
|
|
|
482
482
|
|
|
483
483
|
@core.extern
|
|
484
|
-
def float2int_rz(arg0,
|
|
484
|
+
def float2int_rz(arg0, _semantic=None):
|
|
485
485
|
return core.extern_elementwise("", "", [arg0], {
|
|
486
486
|
(core.dtype("fp32"), ): ("__nv_float2int_rz", core.dtype("int32")),
|
|
487
|
-
}, is_pure=True,
|
|
487
|
+
}, is_pure=True, _semantic=_semantic)
|
|
488
488
|
|
|
489
489
|
|
|
490
490
|
@core.extern
|
|
491
|
-
def float2int_rd(arg0,
|
|
491
|
+
def float2int_rd(arg0, _semantic=None):
|
|
492
492
|
return core.extern_elementwise("", "", [arg0], {
|
|
493
493
|
(core.dtype("fp32"), ): ("__nv_float2int_rd", core.dtype("int32")),
|
|
494
|
-
}, is_pure=True,
|
|
494
|
+
}, is_pure=True, _semantic=_semantic)
|
|
495
495
|
|
|
496
496
|
|
|
497
497
|
@core.extern
|
|
498
|
-
def float2int_ru(arg0,
|
|
498
|
+
def float2int_ru(arg0, _semantic=None):
|
|
499
499
|
return core.extern_elementwise("", "", [arg0], {
|
|
500
500
|
(core.dtype("fp32"), ): ("__nv_float2int_ru", core.dtype("int32")),
|
|
501
|
-
}, is_pure=True,
|
|
501
|
+
}, is_pure=True, _semantic=_semantic)
|
|
502
502
|
|
|
503
503
|
|
|
504
504
|
@core.extern
|
|
505
|
-
def float2uint_rn(arg0,
|
|
505
|
+
def float2uint_rn(arg0, _semantic=None):
|
|
506
506
|
return core.extern_elementwise("", "", [arg0], {
|
|
507
507
|
(core.dtype("fp32"), ): ("__nv_float2uint_rn", core.dtype("int32")),
|
|
508
|
-
}, is_pure=True,
|
|
508
|
+
}, is_pure=True, _semantic=_semantic)
|
|
509
509
|
|
|
510
510
|
|
|
511
511
|
@core.extern
|
|
512
|
-
def float2uint_rz(arg0,
|
|
512
|
+
def float2uint_rz(arg0, _semantic=None):
|
|
513
513
|
return core.extern_elementwise("", "", [arg0], {
|
|
514
514
|
(core.dtype("fp32"), ): ("__nv_float2uint_rz", core.dtype("int32")),
|
|
515
|
-
}, is_pure=True,
|
|
515
|
+
}, is_pure=True, _semantic=_semantic)
|
|
516
516
|
|
|
517
517
|
|
|
518
518
|
@core.extern
|
|
519
|
-
def float2uint_rd(arg0,
|
|
519
|
+
def float2uint_rd(arg0, _semantic=None):
|
|
520
520
|
return core.extern_elementwise("", "", [arg0], {
|
|
521
521
|
(core.dtype("fp32"), ): ("__nv_float2uint_rd", core.dtype("int32")),
|
|
522
|
-
}, is_pure=True,
|
|
522
|
+
}, is_pure=True, _semantic=_semantic)
|
|
523
523
|
|
|
524
524
|
|
|
525
525
|
@core.extern
|
|
526
|
-
def float2uint_ru(arg0,
|
|
526
|
+
def float2uint_ru(arg0, _semantic=None):
|
|
527
527
|
return core.extern_elementwise("", "", [arg0], {
|
|
528
528
|
(core.dtype("fp32"), ): ("__nv_float2uint_ru", core.dtype("int32")),
|
|
529
|
-
}, is_pure=True,
|
|
529
|
+
}, is_pure=True, _semantic=_semantic)
|
|
530
530
|
|
|
531
531
|
|
|
532
532
|
@core.extern
|
|
533
|
-
def int2float_rn(arg0,
|
|
533
|
+
def int2float_rn(arg0, _semantic=None):
|
|
534
534
|
return core.extern_elementwise("", "", [arg0], {
|
|
535
535
|
(core.dtype("int32"), ): ("__nv_int2float_rn", core.dtype("fp32")),
|
|
536
|
-
}, is_pure=True,
|
|
536
|
+
}, is_pure=True, _semantic=_semantic)
|
|
537
537
|
|
|
538
538
|
|
|
539
539
|
@core.extern
|
|
540
|
-
def int2float_rz(arg0,
|
|
540
|
+
def int2float_rz(arg0, _semantic=None):
|
|
541
541
|
return core.extern_elementwise("", "", [arg0], {
|
|
542
542
|
(core.dtype("int32"), ): ("__nv_int2float_rz", core.dtype("fp32")),
|
|
543
|
-
}, is_pure=True,
|
|
543
|
+
}, is_pure=True, _semantic=_semantic)
|
|
544
544
|
|
|
545
545
|
|
|
546
546
|
@core.extern
|
|
547
|
-
def int2float_rd(arg0,
|
|
547
|
+
def int2float_rd(arg0, _semantic=None):
|
|
548
548
|
return core.extern_elementwise("", "", [arg0], {
|
|
549
549
|
(core.dtype("int32"), ): ("__nv_int2float_rd", core.dtype("fp32")),
|
|
550
|
-
}, is_pure=True,
|
|
550
|
+
}, is_pure=True, _semantic=_semantic)
|
|
551
551
|
|
|
552
552
|
|
|
553
553
|
@core.extern
|
|
554
|
-
def int2float_ru(arg0,
|
|
554
|
+
def int2float_ru(arg0, _semantic=None):
|
|
555
555
|
return core.extern_elementwise("", "", [arg0], {
|
|
556
556
|
(core.dtype("int32"), ): ("__nv_int2float_ru", core.dtype("fp32")),
|
|
557
|
-
}, is_pure=True,
|
|
557
|
+
}, is_pure=True, _semantic=_semantic)
|
|
558
558
|
|
|
559
559
|
|
|
560
560
|
@core.extern
|
|
561
|
-
def uint2float_rn(arg0,
|
|
561
|
+
def uint2float_rn(arg0, _semantic=None):
|
|
562
562
|
return core.extern_elementwise("", "", [arg0], {
|
|
563
563
|
(core.dtype("uint32"), ): ("__nv_uint2float_rn", core.dtype("fp32")),
|
|
564
|
-
}, is_pure=True,
|
|
564
|
+
}, is_pure=True, _semantic=_semantic)
|
|
565
565
|
|
|
566
566
|
|
|
567
567
|
@core.extern
|
|
568
|
-
def uint2float_rz(arg0,
|
|
568
|
+
def uint2float_rz(arg0, _semantic=None):
|
|
569
569
|
return core.extern_elementwise("", "", [arg0], {
|
|
570
570
|
(core.dtype("uint32"), ): ("__nv_uint2float_rz", core.dtype("fp32")),
|
|
571
|
-
}, is_pure=True,
|
|
571
|
+
}, is_pure=True, _semantic=_semantic)
|
|
572
572
|
|
|
573
573
|
|
|
574
574
|
@core.extern
|
|
575
|
-
def uint2float_rd(arg0,
|
|
575
|
+
def uint2float_rd(arg0, _semantic=None):
|
|
576
576
|
return core.extern_elementwise("", "", [arg0], {
|
|
577
577
|
(core.dtype("uint32"), ): ("__nv_uint2float_rd", core.dtype("fp32")),
|
|
578
|
-
}, is_pure=True,
|
|
578
|
+
}, is_pure=True, _semantic=_semantic)
|
|
579
579
|
|
|
580
580
|
|
|
581
581
|
@core.extern
|
|
582
|
-
def uint2float_ru(arg0,
|
|
582
|
+
def uint2float_ru(arg0, _semantic=None):
|
|
583
583
|
return core.extern_elementwise("", "", [arg0], {
|
|
584
584
|
(core.dtype("uint32"), ): ("__nv_uint2float_ru", core.dtype("fp32")),
|
|
585
|
-
}, is_pure=True,
|
|
585
|
+
}, is_pure=True, _semantic=_semantic)
|
|
586
586
|
|
|
587
587
|
|
|
588
588
|
@core.extern
|
|
589
|
-
def hiloint2double(arg0, arg1,
|
|
589
|
+
def hiloint2double(arg0, arg1, _semantic=None):
|
|
590
590
|
return core.extern_elementwise("", "", [arg0, arg1], {
|
|
591
591
|
(core.dtype("int32"), core.dtype("int32")): ("__nv_hiloint2double", core.dtype("fp64")),
|
|
592
|
-
}, is_pure=True,
|
|
592
|
+
}, is_pure=True, _semantic=_semantic)
|
|
593
593
|
|
|
594
594
|
|
|
595
595
|
@core.extern
|
|
596
|
-
def double2loint(arg0,
|
|
596
|
+
def double2loint(arg0, _semantic=None):
|
|
597
597
|
return core.extern_elementwise("", "", [arg0], {
|
|
598
598
|
(core.dtype("fp64"), ): ("__nv_double2loint", core.dtype("int32")),
|
|
599
|
-
}, is_pure=True,
|
|
599
|
+
}, is_pure=True, _semantic=_semantic)
|
|
600
600
|
|
|
601
601
|
|
|
602
602
|
@core.extern
|
|
603
|
-
def double2hiint(arg0,
|
|
603
|
+
def double2hiint(arg0, _semantic=None):
|
|
604
604
|
return core.extern_elementwise("", "", [arg0], {
|
|
605
605
|
(core.dtype("fp64"), ): ("__nv_double2hiint", core.dtype("int32")),
|
|
606
|
-
}, is_pure=True,
|
|
606
|
+
}, is_pure=True, _semantic=_semantic)
|
|
607
607
|
|
|
608
608
|
|
|
609
609
|
@core.extern
|
|
610
|
-
def float2ll_rn(arg0,
|
|
610
|
+
def float2ll_rn(arg0, _semantic=None):
|
|
611
611
|
return core.extern_elementwise("", "", [arg0], {
|
|
612
612
|
(core.dtype("fp32"), ): ("__nv_float2ll_rn", core.dtype("int64")),
|
|
613
|
-
}, is_pure=True,
|
|
613
|
+
}, is_pure=True, _semantic=_semantic)
|
|
614
614
|
|
|
615
615
|
|
|
616
616
|
@core.extern
|
|
617
|
-
def float2ll_rz(arg0,
|
|
617
|
+
def float2ll_rz(arg0, _semantic=None):
|
|
618
618
|
return core.extern_elementwise("", "", [arg0], {
|
|
619
619
|
(core.dtype("fp32"), ): ("__nv_float2ll_rz", core.dtype("int64")),
|
|
620
|
-
}, is_pure=True,
|
|
620
|
+
}, is_pure=True, _semantic=_semantic)
|
|
621
621
|
|
|
622
622
|
|
|
623
623
|
@core.extern
|
|
624
|
-
def float2ll_rd(arg0,
|
|
624
|
+
def float2ll_rd(arg0, _semantic=None):
|
|
625
625
|
return core.extern_elementwise("", "", [arg0], {
|
|
626
626
|
(core.dtype("fp32"), ): ("__nv_float2ll_rd", core.dtype("int64")),
|
|
627
|
-
}, is_pure=True,
|
|
627
|
+
}, is_pure=True, _semantic=_semantic)
|
|
628
628
|
|
|
629
629
|
|
|
630
630
|
@core.extern
|
|
631
|
-
def float2ll_ru(arg0,
|
|
631
|
+
def float2ll_ru(arg0, _semantic=None):
|
|
632
632
|
return core.extern_elementwise("", "", [arg0], {
|
|
633
633
|
(core.dtype("fp32"), ): ("__nv_float2ll_ru", core.dtype("int64")),
|
|
634
|
-
}, is_pure=True,
|
|
634
|
+
}, is_pure=True, _semantic=_semantic)
|
|
635
635
|
|
|
636
636
|
|
|
637
637
|
@core.extern
|
|
638
|
-
def float2ull_rn(arg0,
|
|
638
|
+
def float2ull_rn(arg0, _semantic=None):
|
|
639
639
|
return core.extern_elementwise("", "", [arg0], {
|
|
640
640
|
(core.dtype("fp32"), ): ("__nv_float2ull_rn", core.dtype("int64")),
|
|
641
|
-
}, is_pure=True,
|
|
641
|
+
}, is_pure=True, _semantic=_semantic)
|
|
642
642
|
|
|
643
643
|
|
|
644
644
|
@core.extern
|
|
645
|
-
def float2ull_rz(arg0,
|
|
645
|
+
def float2ull_rz(arg0, _semantic=None):
|
|
646
646
|
return core.extern_elementwise("", "", [arg0], {
|
|
647
647
|
(core.dtype("fp32"), ): ("__nv_float2ull_rz", core.dtype("int64")),
|
|
648
|
-
}, is_pure=True,
|
|
648
|
+
}, is_pure=True, _semantic=_semantic)
|
|
649
649
|
|
|
650
650
|
|
|
651
651
|
@core.extern
|
|
652
|
-
def float2ull_rd(arg0,
|
|
652
|
+
def float2ull_rd(arg0, _semantic=None):
|
|
653
653
|
return core.extern_elementwise("", "", [arg0], {
|
|
654
654
|
(core.dtype("fp32"), ): ("__nv_float2ull_rd", core.dtype("int64")),
|
|
655
|
-
}, is_pure=True,
|
|
655
|
+
}, is_pure=True, _semantic=_semantic)
|
|
656
656
|
|
|
657
657
|
|
|
658
658
|
@core.extern
|
|
659
|
-
def float2ull_ru(arg0,
|
|
659
|
+
def float2ull_ru(arg0, _semantic=None):
|
|
660
660
|
return core.extern_elementwise("", "", [arg0], {
|
|
661
661
|
(core.dtype("fp32"), ): ("__nv_float2ull_ru", core.dtype("int64")),
|
|
662
|
-
}, is_pure=True,
|
|
662
|
+
}, is_pure=True, _semantic=_semantic)
|
|
663
663
|
|
|
664
664
|
|
|
665
665
|
@core.extern
|
|
666
|
-
def double2ll_rn(arg0,
|
|
666
|
+
def double2ll_rn(arg0, _semantic=None):
|
|
667
667
|
return core.extern_elementwise("", "", [arg0], {
|
|
668
668
|
(core.dtype("fp64"), ): ("__nv_double2ll_rn", core.dtype("int64")),
|
|
669
|
-
}, is_pure=True,
|
|
669
|
+
}, is_pure=True, _semantic=_semantic)
|
|
670
670
|
|
|
671
671
|
|
|
672
672
|
@core.extern
|
|
673
|
-
def double2ll_rz(arg0,
|
|
673
|
+
def double2ll_rz(arg0, _semantic=None):
|
|
674
674
|
return core.extern_elementwise("", "", [arg0], {
|
|
675
675
|
(core.dtype("fp64"), ): ("__nv_double2ll_rz", core.dtype("int64")),
|
|
676
|
-
}, is_pure=True,
|
|
676
|
+
}, is_pure=True, _semantic=_semantic)
|
|
677
677
|
|
|
678
678
|
|
|
679
679
|
@core.extern
|
|
680
|
-
def double2ll_rd(arg0,
|
|
680
|
+
def double2ll_rd(arg0, _semantic=None):
|
|
681
681
|
return core.extern_elementwise("", "", [arg0], {
|
|
682
682
|
(core.dtype("fp64"), ): ("__nv_double2ll_rd", core.dtype("int64")),
|
|
683
|
-
}, is_pure=True,
|
|
683
|
+
}, is_pure=True, _semantic=_semantic)
|
|
684
684
|
|
|
685
685
|
|
|
686
686
|
@core.extern
|
|
687
|
-
def double2ll_ru(arg0,
|
|
687
|
+
def double2ll_ru(arg0, _semantic=None):
|
|
688
688
|
return core.extern_elementwise("", "", [arg0], {
|
|
689
689
|
(core.dtype("fp64"), ): ("__nv_double2ll_ru", core.dtype("int64")),
|
|
690
|
-
}, is_pure=True,
|
|
690
|
+
}, is_pure=True, _semantic=_semantic)
|
|
691
691
|
|
|
692
692
|
|
|
693
693
|
@core.extern
|
|
694
|
-
def double2ull_rn(arg0,
|
|
694
|
+
def double2ull_rn(arg0, _semantic=None):
|
|
695
695
|
return core.extern_elementwise("", "", [arg0], {
|
|
696
696
|
(core.dtype("fp64"), ): ("__nv_double2ull_rn", core.dtype("int64")),
|
|
697
|
-
}, is_pure=True,
|
|
697
|
+
}, is_pure=True, _semantic=_semantic)
|
|
698
698
|
|
|
699
699
|
|
|
700
700
|
@core.extern
|
|
701
|
-
def double2ull_rz(arg0,
|
|
701
|
+
def double2ull_rz(arg0, _semantic=None):
|
|
702
702
|
return core.extern_elementwise("", "", [arg0], {
|
|
703
703
|
(core.dtype("fp64"), ): ("__nv_double2ull_rz", core.dtype("int64")),
|
|
704
|
-
}, is_pure=True,
|
|
704
|
+
}, is_pure=True, _semantic=_semantic)
|
|
705
705
|
|
|
706
706
|
|
|
707
707
|
@core.extern
|
|
708
|
-
def double2ull_rd(arg0,
|
|
708
|
+
def double2ull_rd(arg0, _semantic=None):
|
|
709
709
|
return core.extern_elementwise("", "", [arg0], {
|
|
710
710
|
(core.dtype("fp64"), ): ("__nv_double2ull_rd", core.dtype("int64")),
|
|
711
|
-
}, is_pure=True,
|
|
711
|
+
}, is_pure=True, _semantic=_semantic)
|
|
712
712
|
|
|
713
713
|
|
|
714
714
|
@core.extern
|
|
715
|
-
def double2ull_ru(arg0,
|
|
715
|
+
def double2ull_ru(arg0, _semantic=None):
|
|
716
716
|
return core.extern_elementwise("", "", [arg0], {
|
|
717
717
|
(core.dtype("fp64"), ): ("__nv_double2ull_ru", core.dtype("int64")),
|
|
718
|
-
}, is_pure=True,
|
|
718
|
+
}, is_pure=True, _semantic=_semantic)
|
|
719
719
|
|
|
720
720
|
|
|
721
721
|
@core.extern
|
|
722
|
-
def ll2float_rn(arg0,
|
|
722
|
+
def ll2float_rn(arg0, _semantic=None):
|
|
723
723
|
return core.extern_elementwise("", "", [arg0], {
|
|
724
724
|
(core.dtype("int64"), ): ("__nv_ll2float_rn", core.dtype("fp32")),
|
|
725
|
-
}, is_pure=True,
|
|
725
|
+
}, is_pure=True, _semantic=_semantic)
|
|
726
726
|
|
|
727
727
|
|
|
728
728
|
@core.extern
|
|
729
|
-
def ll2float_rz(arg0,
|
|
729
|
+
def ll2float_rz(arg0, _semantic=None):
|
|
730
730
|
return core.extern_elementwise("", "", [arg0], {
|
|
731
731
|
(core.dtype("int64"), ): ("__nv_ll2float_rz", core.dtype("fp32")),
|
|
732
|
-
}, is_pure=True,
|
|
732
|
+
}, is_pure=True, _semantic=_semantic)
|
|
733
733
|
|
|
734
734
|
|
|
735
735
|
@core.extern
|
|
736
|
-
def ll2float_rd(arg0,
|
|
736
|
+
def ll2float_rd(arg0, _semantic=None):
|
|
737
737
|
return core.extern_elementwise("", "", [arg0], {
|
|
738
738
|
(core.dtype("int64"), ): ("__nv_ll2float_rd", core.dtype("fp32")),
|
|
739
|
-
}, is_pure=True,
|
|
739
|
+
}, is_pure=True, _semantic=_semantic)
|
|
740
740
|
|
|
741
741
|
|
|
742
742
|
@core.extern
|
|
743
|
-
def ll2float_ru(arg0,
|
|
743
|
+
def ll2float_ru(arg0, _semantic=None):
|
|
744
744
|
return core.extern_elementwise("", "", [arg0], {
|
|
745
745
|
(core.dtype("int64"), ): ("__nv_ll2float_ru", core.dtype("fp32")),
|
|
746
|
-
}, is_pure=True,
|
|
746
|
+
}, is_pure=True, _semantic=_semantic)
|
|
747
747
|
|
|
748
748
|
|
|
749
749
|
@core.extern
|
|
750
|
-
def ull2float_rn(arg0,
|
|
750
|
+
def ull2float_rn(arg0, _semantic=None):
|
|
751
751
|
return core.extern_elementwise("", "", [arg0], {
|
|
752
752
|
(core.dtype("uint64"), ): ("__nv_ull2float_rn", core.dtype("fp32")),
|
|
753
|
-
}, is_pure=True,
|
|
753
|
+
}, is_pure=True, _semantic=_semantic)
|
|
754
754
|
|
|
755
755
|
|
|
756
756
|
@core.extern
|
|
757
|
-
def ull2float_rz(arg0,
|
|
757
|
+
def ull2float_rz(arg0, _semantic=None):
|
|
758
758
|
return core.extern_elementwise("", "", [arg0], {
|
|
759
759
|
(core.dtype("uint64"), ): ("__nv_ull2float_rz", core.dtype("fp32")),
|
|
760
|
-
}, is_pure=True,
|
|
760
|
+
}, is_pure=True, _semantic=_semantic)
|
|
761
761
|
|
|
762
762
|
|
|
763
763
|
@core.extern
|
|
764
|
-
def ull2float_rd(arg0,
|
|
764
|
+
def ull2float_rd(arg0, _semantic=None):
|
|
765
765
|
return core.extern_elementwise("", "", [arg0], {
|
|
766
766
|
(core.dtype("uint64"), ): ("__nv_ull2float_rd", core.dtype("fp32")),
|
|
767
|
-
}, is_pure=True,
|
|
767
|
+
}, is_pure=True, _semantic=_semantic)
|
|
768
768
|
|
|
769
769
|
|
|
770
770
|
@core.extern
|
|
771
|
-
def ull2float_ru(arg0,
|
|
771
|
+
def ull2float_ru(arg0, _semantic=None):
|
|
772
772
|
return core.extern_elementwise("", "", [arg0], {
|
|
773
773
|
(core.dtype("uint64"), ): ("__nv_ull2float_ru", core.dtype("fp32")),
|
|
774
|
-
}, is_pure=True,
|
|
774
|
+
}, is_pure=True, _semantic=_semantic)
|
|
775
775
|
|
|
776
776
|
|
|
777
777
|
@core.extern
|
|
778
|
-
def ll2double_rn(arg0,
|
|
778
|
+
def ll2double_rn(arg0, _semantic=None):
|
|
779
779
|
return core.extern_elementwise("", "", [arg0], {
|
|
780
780
|
(core.dtype("int64"), ): ("__nv_ll2double_rn", core.dtype("fp64")),
|
|
781
|
-
}, is_pure=True,
|
|
781
|
+
}, is_pure=True, _semantic=_semantic)
|
|
782
782
|
|
|
783
783
|
|
|
784
784
|
@core.extern
|
|
785
|
-
def ll2double_rz(arg0,
|
|
785
|
+
def ll2double_rz(arg0, _semantic=None):
|
|
786
786
|
return core.extern_elementwise("", "", [arg0], {
|
|
787
787
|
(core.dtype("int64"), ): ("__nv_ll2double_rz", core.dtype("fp64")),
|
|
788
|
-
}, is_pure=True,
|
|
788
|
+
}, is_pure=True, _semantic=_semantic)
|
|
789
789
|
|
|
790
790
|
|
|
791
791
|
@core.extern
|
|
792
|
-
def ll2double_rd(arg0,
|
|
792
|
+
def ll2double_rd(arg0, _semantic=None):
|
|
793
793
|
return core.extern_elementwise("", "", [arg0], {
|
|
794
794
|
(core.dtype("int64"), ): ("__nv_ll2double_rd", core.dtype("fp64")),
|
|
795
|
-
}, is_pure=True,
|
|
795
|
+
}, is_pure=True, _semantic=_semantic)
|
|
796
796
|
|
|
797
797
|
|
|
798
798
|
@core.extern
|
|
799
|
-
def ll2double_ru(arg0,
|
|
799
|
+
def ll2double_ru(arg0, _semantic=None):
|
|
800
800
|
return core.extern_elementwise("", "", [arg0], {
|
|
801
801
|
(core.dtype("int64"), ): ("__nv_ll2double_ru", core.dtype("fp64")),
|
|
802
|
-
}, is_pure=True,
|
|
802
|
+
}, is_pure=True, _semantic=_semantic)
|
|
803
803
|
|
|
804
804
|
|
|
805
805
|
@core.extern
|
|
806
|
-
def ull2double_rn(arg0,
|
|
806
|
+
def ull2double_rn(arg0, _semantic=None):
|
|
807
807
|
return core.extern_elementwise("", "", [arg0], {
|
|
808
808
|
(core.dtype("uint64"), ): ("__nv_ull2double_rn", core.dtype("fp64")),
|
|
809
|
-
}, is_pure=True,
|
|
809
|
+
}, is_pure=True, _semantic=_semantic)
|
|
810
810
|
|
|
811
811
|
|
|
812
812
|
@core.extern
|
|
813
|
-
def ull2double_rz(arg0,
|
|
813
|
+
def ull2double_rz(arg0, _semantic=None):
|
|
814
814
|
return core.extern_elementwise("", "", [arg0], {
|
|
815
815
|
(core.dtype("uint64"), ): ("__nv_ull2double_rz", core.dtype("fp64")),
|
|
816
|
-
}, is_pure=True,
|
|
816
|
+
}, is_pure=True, _semantic=_semantic)
|
|
817
817
|
|
|
818
818
|
|
|
819
819
|
@core.extern
|
|
820
|
-
def ull2double_rd(arg0,
|
|
820
|
+
def ull2double_rd(arg0, _semantic=None):
|
|
821
821
|
return core.extern_elementwise("", "", [arg0], {
|
|
822
822
|
(core.dtype("uint64"), ): ("__nv_ull2double_rd", core.dtype("fp64")),
|
|
823
|
-
}, is_pure=True,
|
|
823
|
+
}, is_pure=True, _semantic=_semantic)
|
|
824
824
|
|
|
825
825
|
|
|
826
826
|
@core.extern
|
|
827
|
-
def ull2double_ru(arg0,
|
|
827
|
+
def ull2double_ru(arg0, _semantic=None):
|
|
828
828
|
return core.extern_elementwise("", "", [arg0], {
|
|
829
829
|
(core.dtype("uint64"), ): ("__nv_ull2double_ru", core.dtype("fp64")),
|
|
830
|
-
}, is_pure=True,
|
|
830
|
+
}, is_pure=True, _semantic=_semantic)
|
|
831
831
|
|
|
832
832
|
|
|
833
833
|
@core.extern
|
|
834
|
-
def int_as_float(arg0,
|
|
834
|
+
def int_as_float(arg0, _semantic=None):
|
|
835
835
|
return core.extern_elementwise("", "", [arg0], {
|
|
836
836
|
(core.dtype("int32"), ): ("__nv_int_as_float", core.dtype("fp32")),
|
|
837
|
-
}, is_pure=True,
|
|
837
|
+
}, is_pure=True, _semantic=_semantic)
|
|
838
838
|
|
|
839
839
|
|
|
840
840
|
@core.extern
|
|
841
|
-
def float_as_int(arg0,
|
|
841
|
+
def float_as_int(arg0, _semantic=None):
|
|
842
842
|
return core.extern_elementwise("", "", [arg0], {
|
|
843
843
|
(core.dtype("fp32"), ): ("__nv_float_as_int", core.dtype("int32")),
|
|
844
|
-
}, is_pure=True,
|
|
844
|
+
}, is_pure=True, _semantic=_semantic)
|
|
845
845
|
|
|
846
846
|
|
|
847
847
|
@core.extern
|
|
848
|
-
def uint_as_float(arg0,
|
|
848
|
+
def uint_as_float(arg0, _semantic=None):
|
|
849
849
|
return core.extern_elementwise("", "", [arg0], {
|
|
850
850
|
(core.dtype("uint32"), ): ("__nv_uint_as_float", core.dtype("fp32")),
|
|
851
|
-
}, is_pure=True,
|
|
851
|
+
}, is_pure=True, _semantic=_semantic)
|
|
852
852
|
|
|
853
853
|
|
|
854
854
|
@core.extern
|
|
855
|
-
def float_as_uint(arg0,
|
|
855
|
+
def float_as_uint(arg0, _semantic=None):
|
|
856
856
|
return core.extern_elementwise("", "", [arg0], {
|
|
857
857
|
(core.dtype("fp32"), ): ("__nv_float_as_uint", core.dtype("int32")),
|
|
858
|
-
}, is_pure=True,
|
|
858
|
+
}, is_pure=True, _semantic=_semantic)
|
|
859
859
|
|
|
860
860
|
|
|
861
861
|
@core.extern
|
|
862
|
-
def longlong_as_double(arg0,
|
|
862
|
+
def longlong_as_double(arg0, _semantic=None):
|
|
863
863
|
return core.extern_elementwise("", "", [arg0], {
|
|
864
864
|
(core.dtype("int64"), ): ("__nv_longlong_as_double", core.dtype("fp64")),
|
|
865
|
-
}, is_pure=True,
|
|
865
|
+
}, is_pure=True, _semantic=_semantic)
|
|
866
866
|
|
|
867
867
|
|
|
868
868
|
@core.extern
|
|
869
|
-
def double_as_longlong(arg0,
|
|
869
|
+
def double_as_longlong(arg0, _semantic=None):
|
|
870
870
|
return core.extern_elementwise("", "", [arg0], {
|
|
871
871
|
(core.dtype("fp64"), ): ("__nv_double_as_longlong", core.dtype("int64")),
|
|
872
|
-
}, is_pure=True,
|
|
872
|
+
}, is_pure=True, _semantic=_semantic)
|
|
873
873
|
|
|
874
874
|
|
|
875
875
|
@core.extern
|
|
876
|
-
def fast_sinf(arg0,
|
|
876
|
+
def fast_sinf(arg0, _semantic=None):
|
|
877
877
|
return core.extern_elementwise("", "", [arg0], {
|
|
878
878
|
(core.dtype("fp32"), ): ("__nv_fast_sinf", core.dtype("fp32")),
|
|
879
|
-
}, is_pure=True,
|
|
879
|
+
}, is_pure=True, _semantic=_semantic)
|
|
880
880
|
|
|
881
881
|
|
|
882
882
|
@core.extern
|
|
883
|
-
def fast_cosf(arg0,
|
|
883
|
+
def fast_cosf(arg0, _semantic=None):
|
|
884
884
|
return core.extern_elementwise("", "", [arg0], {
|
|
885
885
|
(core.dtype("fp32"), ): ("__nv_fast_cosf", core.dtype("fp32")),
|
|
886
|
-
}, is_pure=True,
|
|
886
|
+
}, is_pure=True, _semantic=_semantic)
|
|
887
887
|
|
|
888
888
|
|
|
889
889
|
@core.extern
|
|
890
|
-
def fast_log2f(arg0,
|
|
890
|
+
def fast_log2f(arg0, _semantic=None):
|
|
891
891
|
return core.extern_elementwise("", "", [arg0], {
|
|
892
892
|
(core.dtype("fp32"), ): ("__nv_fast_log2f", core.dtype("fp32")),
|
|
893
|
-
}, is_pure=True,
|
|
893
|
+
}, is_pure=True, _semantic=_semantic)
|
|
894
894
|
|
|
895
895
|
|
|
896
896
|
@core.extern
|
|
897
|
-
def fast_logf(arg0,
|
|
897
|
+
def fast_logf(arg0, _semantic=None):
|
|
898
898
|
return core.extern_elementwise("", "", [arg0], {
|
|
899
899
|
(core.dtype("fp32"), ): ("__nv_fast_logf", core.dtype("fp32")),
|
|
900
|
-
}, is_pure=True,
|
|
900
|
+
}, is_pure=True, _semantic=_semantic)
|
|
901
901
|
|
|
902
902
|
|
|
903
903
|
@core.extern
|
|
904
|
-
def fast_expf(arg0,
|
|
904
|
+
def fast_expf(arg0, _semantic=None):
|
|
905
905
|
return core.extern_elementwise("", "", [arg0], {
|
|
906
906
|
(core.dtype("fp32"), ): ("__nv_fast_expf", core.dtype("fp32")),
|
|
907
|
-
}, is_pure=True,
|
|
907
|
+
}, is_pure=True, _semantic=_semantic)
|
|
908
908
|
|
|
909
909
|
|
|
910
910
|
@core.extern
|
|
911
|
-
def fast_tanf(arg0,
|
|
911
|
+
def fast_tanf(arg0, _semantic=None):
|
|
912
912
|
return core.extern_elementwise("", "", [arg0], {
|
|
913
913
|
(core.dtype("fp32"), ): ("__nv_fast_tanf", core.dtype("fp32")),
|
|
914
|
-
}, is_pure=True,
|
|
914
|
+
}, is_pure=True, _semantic=_semantic)
|
|
915
915
|
|
|
916
916
|
|
|
917
917
|
@core.extern
|
|
918
|
-
def fast_exp10f(arg0,
|
|
918
|
+
def fast_exp10f(arg0, _semantic=None):
|
|
919
919
|
return core.extern_elementwise("", "", [arg0], {
|
|
920
920
|
(core.dtype("fp32"), ): ("__nv_fast_exp10f", core.dtype("fp32")),
|
|
921
|
-
}, is_pure=True,
|
|
921
|
+
}, is_pure=True, _semantic=_semantic)
|
|
922
922
|
|
|
923
923
|
|
|
924
924
|
@core.extern
|
|
925
|
-
def fast_log10f(arg0,
|
|
925
|
+
def fast_log10f(arg0, _semantic=None):
|
|
926
926
|
return core.extern_elementwise("", "", [arg0], {
|
|
927
927
|
(core.dtype("fp32"), ): ("__nv_fast_log10f", core.dtype("fp32")),
|
|
928
|
-
}, is_pure=True,
|
|
928
|
+
}, is_pure=True, _semantic=_semantic)
|
|
929
929
|
|
|
930
930
|
|
|
931
931
|
@core.extern
|
|
932
|
-
def fast_powf(arg0, arg1,
|
|
932
|
+
def fast_powf(arg0, arg1, _semantic=None):
|
|
933
933
|
return core.extern_elementwise("", "", [arg0, arg1], {
|
|
934
934
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fast_powf", core.dtype("fp32")),
|
|
935
|
-
}, is_pure=True,
|
|
935
|
+
}, is_pure=True, _semantic=_semantic)
|
|
936
936
|
|
|
937
937
|
|
|
938
938
|
@core.extern
|
|
939
|
-
def hadd(arg0, arg1,
|
|
939
|
+
def hadd(arg0, arg1, _semantic=None):
|
|
940
940
|
return core.extern_elementwise(
|
|
941
941
|
"", "", [arg0, arg1], {
|
|
942
942
|
(core.dtype("int32"), core.dtype("int32")): ("__nv_hadd", core.dtype("int32")),
|
|
943
943
|
(core.dtype("uint32"), core.dtype("uint32")): ("__nv_uhadd", core.dtype("uint32")),
|
|
944
|
-
}, is_pure=True,
|
|
944
|
+
}, is_pure=True, _semantic=_semantic)
|
|
945
945
|
|
|
946
946
|
|
|
947
947
|
@core.extern
|
|
948
|
-
def rhadd(arg0, arg1,
|
|
948
|
+
def rhadd(arg0, arg1, _semantic=None):
|
|
949
949
|
return core.extern_elementwise(
|
|
950
950
|
"", "", [arg0, arg1], {
|
|
951
951
|
(core.dtype("int32"), core.dtype("int32")): ("__nv_rhadd", core.dtype("int32")),
|
|
952
952
|
(core.dtype("uint32"), core.dtype("uint32")): ("__nv_urhadd", core.dtype("uint32")),
|
|
953
|
-
}, is_pure=True,
|
|
953
|
+
}, is_pure=True, _semantic=_semantic)
|
|
954
954
|
|
|
955
955
|
|
|
956
956
|
@core.extern
|
|
957
|
-
def sub_rn(arg0, arg1,
|
|
957
|
+
def sub_rn(arg0, arg1, _semantic=None):
|
|
958
958
|
return core.extern_elementwise(
|
|
959
959
|
"", "", [arg0, arg1], {
|
|
960
960
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fsub_rn", core.dtype("fp32")),
|
|
961
961
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dsub_rn", core.dtype("fp64")),
|
|
962
|
-
}, is_pure=True,
|
|
962
|
+
}, is_pure=True, _semantic=_semantic)
|
|
963
963
|
|
|
964
964
|
|
|
965
965
|
@core.extern
|
|
966
|
-
def sub_rz(arg0, arg1,
|
|
966
|
+
def sub_rz(arg0, arg1, _semantic=None):
|
|
967
967
|
return core.extern_elementwise(
|
|
968
968
|
"", "", [arg0, arg1], {
|
|
969
969
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fsub_rz", core.dtype("fp32")),
|
|
970
970
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dsub_rz", core.dtype("fp64")),
|
|
971
|
-
}, is_pure=True,
|
|
971
|
+
}, is_pure=True, _semantic=_semantic)
|
|
972
972
|
|
|
973
973
|
|
|
974
974
|
@core.extern
|
|
975
|
-
def sub_rd(arg0, arg1,
|
|
975
|
+
def sub_rd(arg0, arg1, _semantic=None):
|
|
976
976
|
return core.extern_elementwise(
|
|
977
977
|
"", "", [arg0, arg1], {
|
|
978
978
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fsub_rd", core.dtype("fp32")),
|
|
979
979
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dsub_rd", core.dtype("fp64")),
|
|
980
|
-
}, is_pure=True,
|
|
980
|
+
}, is_pure=True, _semantic=_semantic)
|
|
981
981
|
|
|
982
982
|
|
|
983
983
|
@core.extern
|
|
984
|
-
def sub_ru(arg0, arg1,
|
|
984
|
+
def sub_ru(arg0, arg1, _semantic=None):
|
|
985
985
|
return core.extern_elementwise(
|
|
986
986
|
"", "", [arg0, arg1], {
|
|
987
987
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fsub_ru", core.dtype("fp32")),
|
|
988
988
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_dsub_ru", core.dtype("fp64")),
|
|
989
|
-
}, is_pure=True,
|
|
989
|
+
}, is_pure=True, _semantic=_semantic)
|
|
990
990
|
|
|
991
991
|
|
|
992
992
|
@core.extern
|
|
993
|
-
def rsqrt_rn(arg0,
|
|
993
|
+
def rsqrt_rn(arg0, _semantic=None):
|
|
994
994
|
return core.extern_elementwise("", "", [
|
|
995
995
|
arg0,
|
|
996
996
|
], {
|
|
997
997
|
(core.dtype("fp32"), ): ("__nv_frsqrt_rn", core.dtype("fp32")),
|
|
998
|
-
}, is_pure=True,
|
|
998
|
+
}, is_pure=True, _semantic=_semantic)
|
|
999
999
|
|
|
1000
1000
|
|
|
1001
1001
|
@core.extern
|
|
1002
|
-
def ffs(arg0,
|
|
1002
|
+
def ffs(arg0, _semantic=None):
|
|
1003
1003
|
return core.extern_elementwise(
|
|
1004
1004
|
"", "", [
|
|
1005
1005
|
arg0,
|
|
1006
1006
|
], {
|
|
1007
1007
|
(core.dtype("int32"), ): ("__nv_ffs", core.dtype("int32")),
|
|
1008
1008
|
(core.dtype("int64"), ): ("__nv_ffsll", core.dtype("int32")),
|
|
1009
|
-
}, is_pure=True,
|
|
1009
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1010
1010
|
|
|
1011
1011
|
|
|
1012
1012
|
@core.extern
|
|
1013
|
-
def rint(arg0,
|
|
1013
|
+
def rint(arg0, _semantic=None):
|
|
1014
1014
|
return core.extern_elementwise(
|
|
1015
1015
|
"", "", [
|
|
1016
1016
|
arg0,
|
|
1017
1017
|
], {
|
|
1018
1018
|
(core.dtype("fp32"), ): ("__nv_rintf", core.dtype("fp32")),
|
|
1019
1019
|
(core.dtype("fp64"), ): ("__nv_rint", core.dtype("fp64")),
|
|
1020
|
-
}, is_pure=True,
|
|
1020
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1021
1021
|
|
|
1022
1022
|
|
|
1023
1023
|
@core.extern
|
|
1024
|
-
def llrint(arg0,
|
|
1024
|
+
def llrint(arg0, _semantic=None):
|
|
1025
1025
|
return core.extern_elementwise(
|
|
1026
1026
|
"", "", [
|
|
1027
1027
|
arg0,
|
|
1028
1028
|
], {
|
|
1029
1029
|
(core.dtype("fp32"), ): ("__nv_llrintf", core.dtype("int64")),
|
|
1030
1030
|
(core.dtype("fp64"), ): ("__nv_llrint", core.dtype("int64")),
|
|
1031
|
-
}, is_pure=True,
|
|
1031
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1032
1032
|
|
|
1033
1033
|
|
|
1034
1034
|
@core.extern
|
|
1035
|
-
def nearbyint(arg0,
|
|
1035
|
+
def nearbyint(arg0, _semantic=None):
|
|
1036
1036
|
return core.extern_elementwise(
|
|
1037
1037
|
"", "", [
|
|
1038
1038
|
arg0,
|
|
1039
1039
|
], {
|
|
1040
1040
|
(core.dtype("fp32"), ): ("__nv_nearbyintf", core.dtype("fp32")),
|
|
1041
1041
|
(core.dtype("fp64"), ): ("__nv_nearbyint", core.dtype("fp64")),
|
|
1042
|
-
}, is_pure=True,
|
|
1042
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1043
1043
|
|
|
1044
1044
|
|
|
1045
1045
|
@core.extern
|
|
1046
|
-
def isnan(arg0,
|
|
1046
|
+
def isnan(arg0, _semantic=None):
|
|
1047
1047
|
return core.extern_elementwise(
|
|
1048
1048
|
"", "", [
|
|
1049
1049
|
arg0,
|
|
1050
1050
|
], {
|
|
1051
1051
|
(core.dtype("fp32"), ): ("__nv_isnanf", core.dtype("int32")),
|
|
1052
1052
|
(core.dtype("fp64"), ): ("__nv_isnand", core.dtype("int32")),
|
|
1053
|
-
}, is_pure=True,
|
|
1053
|
+
}, is_pure=True, _semantic=_semantic).to(core.int1, _semantic=_semantic)
|
|
1054
1054
|
|
|
1055
1055
|
|
|
1056
1056
|
@core.extern
|
|
1057
|
-
def signbit(arg0,
|
|
1057
|
+
def signbit(arg0, _semantic=None):
|
|
1058
1058
|
return core.extern_elementwise(
|
|
1059
1059
|
"", "", [
|
|
1060
1060
|
arg0,
|
|
1061
1061
|
], {
|
|
1062
1062
|
(core.dtype("fp32"), ): ("__nv_signbitf", core.dtype("int32")),
|
|
1063
1063
|
(core.dtype("fp64"), ): ("__nv_signbitd", core.dtype("int32")),
|
|
1064
|
-
}, is_pure=True,
|
|
1064
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1065
1065
|
|
|
1066
1066
|
|
|
1067
1067
|
@core.extern
|
|
1068
|
-
def copysign(arg0, arg1,
|
|
1068
|
+
def copysign(arg0, arg1, _semantic=None):
|
|
1069
1069
|
return core.extern_elementwise(
|
|
1070
1070
|
"", "", [arg0, arg1], {
|
|
1071
1071
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_copysignf", core.dtype("fp32")),
|
|
1072
1072
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_copysign", core.dtype("fp64")),
|
|
1073
|
-
}, is_pure=True,
|
|
1073
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1074
1074
|
|
|
1075
1075
|
|
|
1076
1076
|
@core.extern
|
|
1077
|
-
def finitef(arg0,
|
|
1077
|
+
def finitef(arg0, _semantic=None):
|
|
1078
1078
|
return core.extern_elementwise("", "", [arg0], {
|
|
1079
1079
|
(core.dtype("fp32"), ): ("__nv_finitef", core.dtype("int32")),
|
|
1080
|
-
}, is_pure=True,
|
|
1080
|
+
}, is_pure=True, _semantic=_semantic).to(core.int1, _semantic=_semantic)
|
|
1081
1081
|
|
|
1082
1082
|
|
|
1083
1083
|
@core.extern
|
|
1084
|
-
def isinf(arg0,
|
|
1084
|
+
def isinf(arg0, _semantic=None):
|
|
1085
1085
|
return core.extern_elementwise(
|
|
1086
1086
|
"", "", [arg0], {
|
|
1087
1087
|
(core.dtype("fp32"), ): ("__nv_isinff", core.dtype("int32")),
|
|
1088
1088
|
(core.dtype("fp64"), ): ("__nv_isinfd", core.dtype("int32")),
|
|
1089
|
-
}, is_pure=True,
|
|
1089
|
+
}, is_pure=True, _semantic=_semantic).to(core.int1, _semantic=_semantic)
|
|
1090
1090
|
|
|
1091
1091
|
|
|
1092
1092
|
@core.extern
|
|
1093
|
-
def nextafter(arg0, arg1,
|
|
1093
|
+
def nextafter(arg0, arg1, _semantic=None):
|
|
1094
1094
|
return core.extern_elementwise(
|
|
1095
1095
|
"", "", [arg0, arg1], {
|
|
1096
1096
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_nextafterf", core.dtype("fp32")),
|
|
1097
1097
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_nextafter", core.dtype("fp64")),
|
|
1098
|
-
}, is_pure=True,
|
|
1098
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1099
1099
|
|
|
1100
1100
|
|
|
1101
1101
|
@core.extern
|
|
1102
|
-
def sin(arg0,
|
|
1102
|
+
def sin(arg0, _semantic=None):
|
|
1103
1103
|
return core.extern_elementwise(
|
|
1104
1104
|
"", "", [arg0], {
|
|
1105
1105
|
(core.dtype("fp32"), ): ("__nv_sinf", core.dtype("fp32")),
|
|
1106
1106
|
(core.dtype("fp64"), ): ("__nv_sin", core.dtype("fp64")),
|
|
1107
|
-
}, is_pure=True,
|
|
1107
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1108
1108
|
|
|
1109
1109
|
|
|
1110
1110
|
@core.extern
|
|
1111
|
-
def cos(arg0,
|
|
1111
|
+
def cos(arg0, _semantic=None):
|
|
1112
1112
|
return core.extern_elementwise(
|
|
1113
1113
|
"", "", [arg0], {
|
|
1114
1114
|
(core.dtype("fp32"), ): ("__nv_cosf", core.dtype("fp32")),
|
|
1115
1115
|
(core.dtype("fp64"), ): ("__nv_cos", core.dtype("fp64")),
|
|
1116
|
-
}, is_pure=True,
|
|
1116
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1117
1117
|
|
|
1118
1118
|
|
|
1119
1119
|
@core.extern
|
|
1120
|
-
def sinpi(arg0,
|
|
1120
|
+
def sinpi(arg0, _semantic=None):
|
|
1121
1121
|
return core.extern_elementwise(
|
|
1122
1122
|
"", "", [arg0], {
|
|
1123
1123
|
(core.dtype("fp32"), ): ("__nv_sinpif", core.dtype("fp32")),
|
|
1124
1124
|
(core.dtype("fp64"), ): ("__nv_sinpi", core.dtype("fp64")),
|
|
1125
|
-
}, is_pure=True,
|
|
1125
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1126
1126
|
|
|
1127
1127
|
|
|
1128
1128
|
@core.extern
|
|
1129
|
-
def cospi(arg0,
|
|
1129
|
+
def cospi(arg0, _semantic=None):
|
|
1130
1130
|
return core.extern_elementwise(
|
|
1131
1131
|
"", "", [arg0], {
|
|
1132
1132
|
(core.dtype("fp32"), ): ("__nv_cospif", core.dtype("fp32")),
|
|
1133
1133
|
(core.dtype("fp64"), ): ("__nv_cospi", core.dtype("fp64")),
|
|
1134
|
-
}, is_pure=True,
|
|
1134
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1135
1135
|
|
|
1136
1136
|
|
|
1137
1137
|
@core.extern
|
|
1138
|
-
def tan(arg0,
|
|
1138
|
+
def tan(arg0, _semantic=None):
|
|
1139
1139
|
return core.extern_elementwise(
|
|
1140
1140
|
"", "", [arg0], {
|
|
1141
1141
|
(core.dtype("fp32"), ): ("__nv_tanf", core.dtype("fp32")),
|
|
1142
1142
|
(core.dtype("fp64"), ): ("__nv_tan", core.dtype("fp64")),
|
|
1143
|
-
}, is_pure=True,
|
|
1143
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1144
1144
|
|
|
1145
1145
|
|
|
1146
1146
|
@core.extern
|
|
1147
|
-
def log2(arg0,
|
|
1147
|
+
def log2(arg0, _semantic=None):
|
|
1148
1148
|
return core.extern_elementwise(
|
|
1149
1149
|
"", "", [arg0], {
|
|
1150
1150
|
(core.dtype("fp32"), ): ("__nv_log2f", core.dtype("fp32")),
|
|
1151
1151
|
(core.dtype("fp64"), ): ("__nv_log2", core.dtype("fp64")),
|
|
1152
|
-
}, is_pure=True,
|
|
1152
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1153
1153
|
|
|
1154
1154
|
|
|
1155
1155
|
@core.extern
|
|
1156
|
-
def exp(arg0,
|
|
1156
|
+
def exp(arg0, _semantic=None):
|
|
1157
1157
|
return core.extern_elementwise(
|
|
1158
1158
|
"", "", [arg0], {
|
|
1159
1159
|
(core.dtype("fp32"), ): ("__nv_expf", core.dtype("fp32")),
|
|
1160
1160
|
(core.dtype("fp64"), ): ("__nv_exp", core.dtype("fp64")),
|
|
1161
|
-
}, is_pure=True,
|
|
1161
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1162
1162
|
|
|
1163
1163
|
|
|
1164
1164
|
@core.extern
|
|
1165
|
-
def exp10(arg0,
|
|
1165
|
+
def exp10(arg0, _semantic=None):
|
|
1166
1166
|
return core.extern_elementwise(
|
|
1167
1167
|
"", "", [arg0], {
|
|
1168
1168
|
(core.dtype("fp32"), ): ("__nv_exp10f", core.dtype("fp32")),
|
|
1169
1169
|
(core.dtype("fp64"), ): ("__nv_exp10", core.dtype("fp64")),
|
|
1170
|
-
}, is_pure=True,
|
|
1170
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1171
1171
|
|
|
1172
1172
|
|
|
1173
1173
|
@core.extern
|
|
1174
|
-
def cosh(arg0,
|
|
1174
|
+
def cosh(arg0, _semantic=None):
|
|
1175
1175
|
return core.extern_elementwise(
|
|
1176
1176
|
"", "", [arg0], {
|
|
1177
1177
|
(core.dtype("fp32"), ): ("__nv_coshf", core.dtype("fp32")),
|
|
1178
1178
|
(core.dtype("fp64"), ): ("__nv_cosh", core.dtype("fp64")),
|
|
1179
|
-
}, is_pure=True,
|
|
1179
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1180
1180
|
|
|
1181
1181
|
|
|
1182
1182
|
@core.extern
|
|
1183
|
-
def sinh(arg0,
|
|
1183
|
+
def sinh(arg0, _semantic=None):
|
|
1184
1184
|
return core.extern_elementwise(
|
|
1185
1185
|
"", "", [arg0], {
|
|
1186
1186
|
(core.dtype("fp32"), ): ("__nv_sinhf", core.dtype("fp32")),
|
|
1187
1187
|
(core.dtype("fp64"), ): ("__nv_sinh", core.dtype("fp64")),
|
|
1188
|
-
}, is_pure=True,
|
|
1188
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1189
1189
|
|
|
1190
1190
|
|
|
1191
1191
|
@core.extern
|
|
1192
|
-
def tanh(arg0,
|
|
1192
|
+
def tanh(arg0, _semantic=None):
|
|
1193
1193
|
return core.extern_elementwise(
|
|
1194
1194
|
"", "", [arg0], {
|
|
1195
1195
|
(core.dtype("fp32"), ): ("__nv_tanhf", core.dtype("fp32")),
|
|
1196
1196
|
(core.dtype("fp64"), ): ("__nv_tanh", core.dtype("fp64")),
|
|
1197
|
-
}, is_pure=True,
|
|
1197
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1198
1198
|
|
|
1199
1199
|
|
|
1200
1200
|
@core.extern
|
|
1201
|
-
def atan2(arg0, arg1,
|
|
1201
|
+
def atan2(arg0, arg1, _semantic=None):
|
|
1202
1202
|
return core.extern_elementwise(
|
|
1203
1203
|
"", "", [arg0, arg1], {
|
|
1204
1204
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_atan2f", core.dtype("fp32")),
|
|
1205
1205
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_atan2", core.dtype("fp64")),
|
|
1206
|
-
}, is_pure=True,
|
|
1206
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1207
1207
|
|
|
1208
1208
|
|
|
1209
1209
|
@core.extern
|
|
1210
|
-
def atan(arg0,
|
|
1210
|
+
def atan(arg0, _semantic=None):
|
|
1211
1211
|
return core.extern_elementwise(
|
|
1212
1212
|
"", "", [arg0], {
|
|
1213
1213
|
(core.dtype("fp32"), ): ("__nv_atanf", core.dtype("fp32")),
|
|
1214
1214
|
(core.dtype("fp64"), ): ("__nv_atan", core.dtype("fp64")),
|
|
1215
|
-
}, is_pure=True,
|
|
1215
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1216
1216
|
|
|
1217
1217
|
|
|
1218
1218
|
@core.extern
|
|
1219
|
-
def asin(arg0,
|
|
1219
|
+
def asin(arg0, _semantic=None):
|
|
1220
1220
|
return core.extern_elementwise(
|
|
1221
1221
|
"", "", [arg0], {
|
|
1222
1222
|
(core.dtype("fp32"), ): ("__nv_asinf", core.dtype("fp32")),
|
|
1223
1223
|
(core.dtype("fp64"), ): ("__nv_asin", core.dtype("fp64")),
|
|
1224
|
-
}, is_pure=True,
|
|
1224
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1225
1225
|
|
|
1226
1226
|
|
|
1227
1227
|
@core.extern
|
|
1228
|
-
def acos(arg0,
|
|
1228
|
+
def acos(arg0, _semantic=None):
|
|
1229
1229
|
return core.extern_elementwise(
|
|
1230
1230
|
"", "", [arg0], {
|
|
1231
1231
|
(core.dtype("fp32"), ): ("__nv_acosf", core.dtype("fp32")),
|
|
1232
1232
|
(core.dtype("fp64"), ): ("__nv_acos", core.dtype("fp64")),
|
|
1233
|
-
}, is_pure=True,
|
|
1233
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1234
1234
|
|
|
1235
1235
|
|
|
1236
1236
|
@core.extern
|
|
1237
|
-
def log(arg0,
|
|
1237
|
+
def log(arg0, _semantic=None):
|
|
1238
1238
|
return core.extern_elementwise(
|
|
1239
1239
|
"", "", [arg0], {
|
|
1240
1240
|
(core.dtype("fp32"), ): ("__nv_logf", core.dtype("fp32")),
|
|
1241
1241
|
(core.dtype("fp64"), ): ("__nv_log", core.dtype("fp64")),
|
|
1242
|
-
}, is_pure=True,
|
|
1242
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1243
1243
|
|
|
1244
1244
|
|
|
1245
1245
|
@core.extern
|
|
1246
|
-
def log10(arg0,
|
|
1246
|
+
def log10(arg0, _semantic=None):
|
|
1247
1247
|
return core.extern_elementwise(
|
|
1248
1248
|
"", "", [arg0], {
|
|
1249
1249
|
(core.dtype("fp32"), ): ("__nv_log10f", core.dtype("fp32")),
|
|
1250
1250
|
(core.dtype("fp64"), ): ("__nv_log10", core.dtype("fp64")),
|
|
1251
|
-
}, is_pure=True,
|
|
1251
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1252
1252
|
|
|
1253
1253
|
|
|
1254
1254
|
@core.extern
|
|
1255
|
-
def log1p(arg0,
|
|
1255
|
+
def log1p(arg0, _semantic=None):
|
|
1256
1256
|
return core.extern_elementwise(
|
|
1257
1257
|
"", "", [arg0], {
|
|
1258
1258
|
(core.dtype("fp32"), ): ("__nv_log1pf", core.dtype("fp32")),
|
|
1259
1259
|
(core.dtype("fp64"), ): ("__nv_log1p", core.dtype("fp64")),
|
|
1260
|
-
}, is_pure=True,
|
|
1260
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1261
1261
|
|
|
1262
1262
|
|
|
1263
1263
|
@core.extern
|
|
1264
|
-
def acosh(arg0,
|
|
1264
|
+
def acosh(arg0, _semantic=None):
|
|
1265
1265
|
return core.extern_elementwise(
|
|
1266
1266
|
"", "", [arg0], {
|
|
1267
1267
|
(core.dtype("fp32"), ): ("__nv_acoshf", core.dtype("fp32")),
|
|
1268
1268
|
(core.dtype("fp64"), ): ("__nv_acosh", core.dtype("fp64")),
|
|
1269
|
-
}, is_pure=True,
|
|
1269
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1270
1270
|
|
|
1271
1271
|
|
|
1272
1272
|
@core.extern
|
|
1273
|
-
def asinh(arg0,
|
|
1273
|
+
def asinh(arg0, _semantic=None):
|
|
1274
1274
|
return core.extern_elementwise(
|
|
1275
1275
|
"", "", [arg0], {
|
|
1276
1276
|
(core.dtype("fp32"), ): ("__nv_asinhf", core.dtype("fp32")),
|
|
1277
1277
|
(core.dtype("fp64"), ): ("__nv_asinh", core.dtype("fp64")),
|
|
1278
|
-
}, is_pure=True,
|
|
1278
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1279
1279
|
|
|
1280
1280
|
|
|
1281
1281
|
@core.extern
|
|
1282
|
-
def atanh(arg0,
|
|
1282
|
+
def atanh(arg0, _semantic=None):
|
|
1283
1283
|
return core.extern_elementwise(
|
|
1284
1284
|
"", "", [arg0], {
|
|
1285
1285
|
(core.dtype("fp32"), ): ("__nv_atanhf", core.dtype("fp32")),
|
|
1286
1286
|
(core.dtype("fp64"), ): ("__nv_atanh", core.dtype("fp64")),
|
|
1287
|
-
}, is_pure=True,
|
|
1287
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1288
1288
|
|
|
1289
1289
|
|
|
1290
1290
|
@core.extern
|
|
1291
|
-
def expm1(arg0,
|
|
1291
|
+
def expm1(arg0, _semantic=None):
|
|
1292
1292
|
return core.extern_elementwise(
|
|
1293
1293
|
"", "", [arg0], {
|
|
1294
1294
|
(core.dtype("fp32"), ): ("__nv_expm1f", core.dtype("fp32")),
|
|
1295
1295
|
(core.dtype("fp64"), ): ("__nv_expm1", core.dtype("fp64")),
|
|
1296
|
-
}, is_pure=True,
|
|
1296
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1297
1297
|
|
|
1298
1298
|
|
|
1299
1299
|
@core.extern
|
|
1300
|
-
def hypot(arg0, arg1,
|
|
1300
|
+
def hypot(arg0, arg1, _semantic=None):
|
|
1301
1301
|
return core.extern_elementwise(
|
|
1302
1302
|
"", "", [arg0, arg1], {
|
|
1303
1303
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_hypotf", core.dtype("fp32")),
|
|
1304
1304
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_hypot", core.dtype("fp64")),
|
|
1305
|
-
}, is_pure=True,
|
|
1305
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1306
1306
|
|
|
1307
1307
|
|
|
1308
1308
|
@core.extern
|
|
1309
|
-
def rhypot(arg0, arg1,
|
|
1309
|
+
def rhypot(arg0, arg1, _semantic=None):
|
|
1310
1310
|
return core.extern_elementwise(
|
|
1311
1311
|
"", "", [arg0, arg1], {
|
|
1312
1312
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_rhypotf", core.dtype("fp32")),
|
|
1313
1313
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_rhypot", core.dtype("fp64")),
|
|
1314
|
-
}, is_pure=True,
|
|
1314
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1315
1315
|
|
|
1316
1316
|
|
|
1317
1317
|
@core.extern
|
|
1318
|
-
def norm3d(arg0, arg1, arg2,
|
|
1318
|
+
def norm3d(arg0, arg1, arg2, _semantic=None):
|
|
1319
1319
|
return core.extern_elementwise(
|
|
1320
1320
|
"", "", [arg0, arg1, arg2], {
|
|
1321
1321
|
(core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32")): ("__nv_norm3df", core.dtype("fp32")),
|
|
1322
1322
|
(core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64")): ("__nv_norm3d", core.dtype("fp64")),
|
|
1323
|
-
}, is_pure=True,
|
|
1323
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1324
1324
|
|
|
1325
1325
|
|
|
1326
1326
|
@core.extern
|
|
1327
|
-
def rnorm3d(arg0, arg1, arg2,
|
|
1327
|
+
def rnorm3d(arg0, arg1, arg2, _semantic=None):
|
|
1328
1328
|
return core.extern_elementwise(
|
|
1329
1329
|
"", "", [arg0, arg1, arg2], {
|
|
1330
1330
|
(core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32")): ("__nv_rnorm3df", core.dtype("fp32")),
|
|
1331
1331
|
(core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64")): ("__nv_rnorm3d", core.dtype("fp64")),
|
|
1332
|
-
}, is_pure=True,
|
|
1332
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1333
1333
|
|
|
1334
1334
|
|
|
1335
1335
|
@core.extern
|
|
1336
|
-
def norm4d(arg0, arg1, arg2, arg3,
|
|
1336
|
+
def norm4d(arg0, arg1, arg2, arg3, _semantic=None):
|
|
1337
1337
|
return core.extern_elementwise(
|
|
1338
1338
|
"", "", [arg0, arg1, arg2, arg3], {
|
|
1339
1339
|
(core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32")):
|
|
1340
1340
|
("__nv_norm4df", core.dtype("fp32")),
|
|
1341
1341
|
(core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64")):
|
|
1342
1342
|
("__nv_norm4d", core.dtype("fp64")),
|
|
1343
|
-
}, is_pure=True,
|
|
1343
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1344
1344
|
|
|
1345
1345
|
|
|
1346
1346
|
@core.extern
|
|
1347
|
-
def rnorm4d(arg0, arg1, arg2, arg3,
|
|
1347
|
+
def rnorm4d(arg0, arg1, arg2, arg3, _semantic=None):
|
|
1348
1348
|
return core.extern_elementwise(
|
|
1349
1349
|
"", "", [arg0, arg1, arg2, arg3], {
|
|
1350
1350
|
(core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32")):
|
|
1351
1351
|
("__nv_rnorm4df", core.dtype("fp32")),
|
|
1352
1352
|
(core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64")):
|
|
1353
1353
|
("__nv_rnorm4d", core.dtype("fp64")),
|
|
1354
|
-
}, is_pure=True,
|
|
1354
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1355
1355
|
|
|
1356
1356
|
|
|
1357
1357
|
@core.extern
|
|
1358
|
-
def cbrt(arg0,
|
|
1358
|
+
def cbrt(arg0, _semantic=None):
|
|
1359
1359
|
return core.extern_elementwise(
|
|
1360
1360
|
"", "", [arg0], {
|
|
1361
1361
|
(core.dtype("fp32"), ): ("__nv_cbrtf", core.dtype("fp32")),
|
|
1362
1362
|
(core.dtype("fp64"), ): ("__nv_cbrt", core.dtype("fp64")),
|
|
1363
|
-
}, is_pure=True,
|
|
1363
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1364
1364
|
|
|
1365
1365
|
|
|
1366
1366
|
@core.extern
|
|
1367
|
-
def rcbrt(arg0,
|
|
1367
|
+
def rcbrt(arg0, _semantic=None):
|
|
1368
1368
|
return core.extern_elementwise(
|
|
1369
1369
|
"", "", [arg0], {
|
|
1370
1370
|
(core.dtype("fp32"), ): ("__nv_rcbrtf", core.dtype("fp32")),
|
|
1371
1371
|
(core.dtype("fp64"), ): ("__nv_rcbrt", core.dtype("fp64")),
|
|
1372
|
-
}, is_pure=True,
|
|
1372
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1373
1373
|
|
|
1374
1374
|
|
|
1375
1375
|
@core.extern
|
|
1376
|
-
def j0(arg0,
|
|
1376
|
+
def j0(arg0, _semantic=None):
|
|
1377
1377
|
return core.extern_elementwise("", "", [arg0], {
|
|
1378
1378
|
(core.dtype("fp32"), ): ("__nv_j0f", core.dtype("fp32")),
|
|
1379
1379
|
(core.dtype("fp64"), ): ("__nv_j0", core.dtype("fp64")),
|
|
1380
|
-
}, is_pure=True,
|
|
1380
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1381
1381
|
|
|
1382
1382
|
|
|
1383
1383
|
@core.extern
|
|
1384
|
-
def j1(arg0,
|
|
1384
|
+
def j1(arg0, _semantic=None):
|
|
1385
1385
|
return core.extern_elementwise("", "", [arg0], {
|
|
1386
1386
|
(core.dtype("fp32"), ): ("__nv_j1f", core.dtype("fp32")),
|
|
1387
1387
|
(core.dtype("fp64"), ): ("__nv_j1", core.dtype("fp64")),
|
|
1388
|
-
}, is_pure=True,
|
|
1388
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1389
1389
|
|
|
1390
1390
|
|
|
1391
1391
|
@core.extern
|
|
1392
|
-
def y0(arg0,
|
|
1392
|
+
def y0(arg0, _semantic=None):
|
|
1393
1393
|
return core.extern_elementwise("", "", [arg0], {
|
|
1394
1394
|
(core.dtype("fp32"), ): ("__nv_y0f", core.dtype("fp32")),
|
|
1395
1395
|
(core.dtype("fp64"), ): ("__nv_y0", core.dtype("fp64")),
|
|
1396
|
-
}, is_pure=True,
|
|
1396
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1397
1397
|
|
|
1398
1398
|
|
|
1399
1399
|
@core.extern
|
|
1400
|
-
def y1(arg0,
|
|
1400
|
+
def y1(arg0, _semantic=None):
|
|
1401
1401
|
return core.extern_elementwise("", "", [arg0], {
|
|
1402
1402
|
(core.dtype("fp32"), ): ("__nv_y1f", core.dtype("fp32")),
|
|
1403
1403
|
(core.dtype("fp64"), ): ("__nv_y1", core.dtype("fp64")),
|
|
1404
|
-
}, is_pure=True,
|
|
1404
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1405
1405
|
|
|
1406
1406
|
|
|
1407
1407
|
@core.extern
|
|
1408
|
-
def yn(arg0, arg1,
|
|
1408
|
+
def yn(arg0, arg1, _semantic=None):
|
|
1409
1409
|
return core.extern_elementwise(
|
|
1410
1410
|
"", "", [arg0, arg1], {
|
|
1411
1411
|
(core.dtype("int32"), core.dtype("fp32")): ("__nv_ynf", core.dtype("fp32")),
|
|
1412
1412
|
(core.dtype("int32"), core.dtype("fp64")): ("__nv_yn", core.dtype("fp64")),
|
|
1413
|
-
}, is_pure=True,
|
|
1413
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1414
1414
|
|
|
1415
1415
|
|
|
1416
1416
|
@core.extern
|
|
1417
|
-
def jn(arg0, arg1,
|
|
1417
|
+
def jn(arg0, arg1, _semantic=None):
|
|
1418
1418
|
return core.extern_elementwise(
|
|
1419
1419
|
"", "", [arg0, arg1], {
|
|
1420
1420
|
(core.dtype("int32"), core.dtype("fp32")): ("__nv_jnf", core.dtype("fp32")),
|
|
1421
1421
|
(core.dtype("int32"), core.dtype("fp64")): ("__nv_jn", core.dtype("fp64")),
|
|
1422
|
-
}, is_pure=True,
|
|
1422
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1423
1423
|
|
|
1424
1424
|
|
|
1425
1425
|
@core.extern
|
|
1426
|
-
def cyl_bessel_i0(arg0,
|
|
1426
|
+
def cyl_bessel_i0(arg0, _semantic=None):
|
|
1427
1427
|
return core.extern_elementwise(
|
|
1428
1428
|
"", "", [arg0], {
|
|
1429
1429
|
(core.dtype("fp32"), ): ("__nv_cyl_bessel_i0f", core.dtype("fp32")),
|
|
1430
1430
|
(core.dtype("fp64"), ): ("__nv_cyl_bessel_i0", core.dtype("fp64")),
|
|
1431
|
-
}, is_pure=True,
|
|
1431
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1432
1432
|
|
|
1433
1433
|
|
|
1434
1434
|
@core.extern
|
|
1435
|
-
def cyl_bessel_i1(arg0,
|
|
1435
|
+
def cyl_bessel_i1(arg0, _semantic=None):
|
|
1436
1436
|
return core.extern_elementwise(
|
|
1437
1437
|
"", "", [arg0], {
|
|
1438
1438
|
(core.dtype("fp32"), ): ("__nv_cyl_bessel_i1f", core.dtype("fp32")),
|
|
1439
1439
|
(core.dtype("fp64"), ): ("__nv_cyl_bessel_i1", core.dtype("fp64")),
|
|
1440
|
-
}, is_pure=True,
|
|
1440
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1441
1441
|
|
|
1442
1442
|
|
|
1443
1443
|
@core.extern
|
|
1444
|
-
def erf(arg0,
|
|
1444
|
+
def erf(arg0, _semantic=None):
|
|
1445
1445
|
return core.extern_elementwise(
|
|
1446
1446
|
"", "", [arg0], {
|
|
1447
1447
|
(core.dtype("fp32"), ): ("__nv_erff", core.dtype("fp32")),
|
|
1448
1448
|
(core.dtype("fp64"), ): ("__nv_erf", core.dtype("fp64")),
|
|
1449
|
-
}, is_pure=True,
|
|
1449
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1450
1450
|
|
|
1451
1451
|
|
|
1452
1452
|
@core.extern
|
|
1453
|
-
def erfinv(arg0,
|
|
1453
|
+
def erfinv(arg0, _semantic=None):
|
|
1454
1454
|
return core.extern_elementwise(
|
|
1455
1455
|
"", "", [arg0], {
|
|
1456
1456
|
(core.dtype("fp32"), ): ("__nv_erfinvf", core.dtype("fp32")),
|
|
1457
1457
|
(core.dtype("fp64"), ): ("__nv_erfinv", core.dtype("fp64")),
|
|
1458
|
-
}, is_pure=True,
|
|
1458
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1459
1459
|
|
|
1460
1460
|
|
|
1461
1461
|
@core.extern
|
|
1462
|
-
def erfc(arg0,
|
|
1462
|
+
def erfc(arg0, _semantic=None):
|
|
1463
1463
|
return core.extern_elementwise(
|
|
1464
1464
|
"", "", [arg0], {
|
|
1465
1465
|
(core.dtype("fp32"), ): ("__nv_erfcf", core.dtype("fp32")),
|
|
1466
1466
|
(core.dtype("fp64"), ): ("__nv_erfc", core.dtype("fp64")),
|
|
1467
|
-
}, is_pure=True,
|
|
1467
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1468
1468
|
|
|
1469
1469
|
|
|
1470
1470
|
@core.extern
|
|
1471
|
-
def erfcx(arg0,
|
|
1471
|
+
def erfcx(arg0, _semantic=None):
|
|
1472
1472
|
return core.extern_elementwise(
|
|
1473
1473
|
"", "", [arg0], {
|
|
1474
1474
|
(core.dtype("fp32"), ): ("__nv_erfcxf", core.dtype("fp32")),
|
|
1475
1475
|
(core.dtype("fp64"), ): ("__nv_erfcx", core.dtype("fp64")),
|
|
1476
|
-
}, is_pure=True,
|
|
1476
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1477
1477
|
|
|
1478
1478
|
|
|
1479
1479
|
@core.extern
|
|
1480
|
-
def erfcinv(arg0,
|
|
1480
|
+
def erfcinv(arg0, _semantic=None):
|
|
1481
1481
|
return core.extern_elementwise(
|
|
1482
1482
|
"", "", [arg0], {
|
|
1483
1483
|
(core.dtype("fp32"), ): ("__nv_erfcinvf", core.dtype("fp32")),
|
|
1484
1484
|
(core.dtype("fp64"), ): ("__nv_erfcinv", core.dtype("fp64")),
|
|
1485
|
-
}, is_pure=True,
|
|
1485
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1486
1486
|
|
|
1487
1487
|
|
|
1488
1488
|
@core.extern
|
|
1489
|
-
def normcdfinv(arg0,
|
|
1489
|
+
def normcdfinv(arg0, _semantic=None):
|
|
1490
1490
|
return core.extern_elementwise(
|
|
1491
1491
|
"", "", [arg0], {
|
|
1492
1492
|
(core.dtype("fp32"), ): ("__nv_normcdfinvf", core.dtype("fp32")),
|
|
1493
1493
|
(core.dtype("fp64"), ): ("__nv_normcdfinv", core.dtype("fp64")),
|
|
1494
|
-
}, is_pure=True,
|
|
1494
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1495
1495
|
|
|
1496
1496
|
|
|
1497
1497
|
@core.extern
|
|
1498
|
-
def normcdf(arg0,
|
|
1498
|
+
def normcdf(arg0, _semantic=None):
|
|
1499
1499
|
return core.extern_elementwise(
|
|
1500
1500
|
"", "", [arg0], {
|
|
1501
1501
|
(core.dtype("fp32"), ): ("__nv_normcdff", core.dtype("fp32")),
|
|
1502
1502
|
(core.dtype("fp64"), ): ("__nv_normcdf", core.dtype("fp64")),
|
|
1503
|
-
}, is_pure=True,
|
|
1503
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1504
1504
|
|
|
1505
1505
|
|
|
1506
1506
|
@core.extern
|
|
1507
|
-
def lgamma(arg0,
|
|
1507
|
+
def lgamma(arg0, _semantic=None):
|
|
1508
1508
|
return core.extern_elementwise(
|
|
1509
1509
|
"", "", [arg0], {
|
|
1510
1510
|
(core.dtype("fp32"), ): ("__nv_lgammaf", core.dtype("fp32")),
|
|
1511
1511
|
(core.dtype("fp64"), ): ("__nv_lgamma", core.dtype("fp64")),
|
|
1512
|
-
}, is_pure=True,
|
|
1512
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1513
1513
|
|
|
1514
1514
|
|
|
1515
1515
|
@core.extern
|
|
1516
|
-
def ldexp(arg0, arg1,
|
|
1516
|
+
def ldexp(arg0, arg1, _semantic=None):
|
|
1517
1517
|
return core.extern_elementwise(
|
|
1518
1518
|
"", "", [arg0, arg1], {
|
|
1519
1519
|
(core.dtype("fp32"), core.dtype("int32")): ("__nv_ldexpf", core.dtype("fp32")),
|
|
1520
1520
|
(core.dtype("fp64"), core.dtype("int32")): ("__nv_ldexp", core.dtype("fp64")),
|
|
1521
|
-
}, is_pure=True,
|
|
1521
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1522
1522
|
|
|
1523
1523
|
|
|
1524
1524
|
@core.extern
|
|
1525
|
-
def scalbn(arg0, arg1,
|
|
1525
|
+
def scalbn(arg0, arg1, _semantic=None):
|
|
1526
1526
|
return core.extern_elementwise(
|
|
1527
1527
|
"", "", [arg0, arg1], {
|
|
1528
1528
|
(core.dtype("fp32"), core.dtype("int32")): ("__nv_scalbnf", core.dtype("fp32")),
|
|
1529
1529
|
(core.dtype("fp64"), core.dtype("int32")): ("__nv_scalbn", core.dtype("fp64")),
|
|
1530
|
-
}, is_pure=True,
|
|
1530
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1531
1531
|
|
|
1532
1532
|
|
|
1533
1533
|
@core.extern
|
|
1534
|
-
def fmod(arg0, arg1,
|
|
1534
|
+
def fmod(arg0, arg1, _semantic=None):
|
|
1535
1535
|
return core.extern_elementwise(
|
|
1536
1536
|
"", "", [arg0, arg1], {
|
|
1537
1537
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fmodf", core.dtype("fp32")),
|
|
1538
1538
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_fmod", core.dtype("fp64")),
|
|
1539
|
-
}, is_pure=True,
|
|
1539
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1540
1540
|
|
|
1541
1541
|
|
|
1542
1542
|
@core.extern
|
|
1543
|
-
def remainder(arg0, arg1,
|
|
1543
|
+
def remainder(arg0, arg1, _semantic=None):
|
|
1544
1544
|
return core.extern_elementwise(
|
|
1545
1545
|
"", "", [arg0, arg1], {
|
|
1546
1546
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_remainderf", core.dtype("fp32")),
|
|
1547
1547
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_remainder", core.dtype("fp64")),
|
|
1548
|
-
}, is_pure=True,
|
|
1548
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1549
1549
|
|
|
1550
1550
|
|
|
1551
1551
|
@core.extern
|
|
1552
|
-
def fma(arg0, arg1, arg2,
|
|
1552
|
+
def fma(arg0, arg1, arg2, _semantic=None):
|
|
1553
1553
|
return core.extern_elementwise(
|
|
1554
1554
|
"", "", [arg0, arg1, arg2], {
|
|
1555
1555
|
(core.dtype("fp32"), core.dtype("fp32"), core.dtype("fp32")): ("__nv_fmaf", core.dtype("fp32")),
|
|
1556
1556
|
(core.dtype("fp64"), core.dtype("fp64"), core.dtype("fp64")): ("__nv_fma", core.dtype("fp64")),
|
|
1557
|
-
}, is_pure=True,
|
|
1557
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1558
1558
|
|
|
1559
1559
|
|
|
1560
1560
|
@core.extern
|
|
1561
|
-
def pow(arg0, arg1,
|
|
1561
|
+
def pow(arg0, arg1, _semantic=None):
|
|
1562
1562
|
return core.extern_elementwise(
|
|
1563
1563
|
"", "", [arg0, arg1], {
|
|
1564
1564
|
(core.dtype("fp32"), core.dtype("int32")): ("__nv_powif", core.dtype("fp32")),
|
|
1565
1565
|
(core.dtype("fp64"), core.dtype("int32")): ("__nv_powi", core.dtype("fp64")),
|
|
1566
1566
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_powf", core.dtype("fp32")),
|
|
1567
1567
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_pow", core.dtype("fp64")),
|
|
1568
|
-
}, is_pure=True,
|
|
1568
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1569
1569
|
|
|
1570
1570
|
|
|
1571
1571
|
@core.extern
|
|
1572
|
-
def tgamma(arg0,
|
|
1572
|
+
def tgamma(arg0, _semantic=None):
|
|
1573
1573
|
return core.extern_elementwise(
|
|
1574
1574
|
"", "", [arg0], {
|
|
1575
1575
|
(core.dtype("fp32"), ): ("__nv_tgammaf", core.dtype("fp32")),
|
|
1576
1576
|
(core.dtype("fp64"), ): ("__nv_tgamma", core.dtype("fp64")),
|
|
1577
|
-
}, is_pure=True,
|
|
1577
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1578
1578
|
|
|
1579
1579
|
|
|
1580
1580
|
@core.extern
|
|
1581
|
-
def round(arg0,
|
|
1581
|
+
def round(arg0, _semantic=None):
|
|
1582
1582
|
return core.extern_elementwise(
|
|
1583
1583
|
"", "", [arg0], {
|
|
1584
1584
|
(core.dtype("fp32"), ): ("__nv_roundf", core.dtype("fp32")),
|
|
1585
1585
|
(core.dtype("fp64"), ): ("__nv_round", core.dtype("fp64")),
|
|
1586
|
-
}, is_pure=True,
|
|
1586
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1587
1587
|
|
|
1588
1588
|
|
|
1589
1589
|
@core.extern
|
|
1590
|
-
def llround(arg0,
|
|
1590
|
+
def llround(arg0, _semantic=None):
|
|
1591
1591
|
return core.extern_elementwise(
|
|
1592
1592
|
"", "", [arg0], {
|
|
1593
1593
|
(core.dtype("fp32"), ): ("__nv_llroundf", core.dtype("int64")),
|
|
1594
1594
|
(core.dtype("fp64"), ): ("__nv_llround", core.dtype("int64")),
|
|
1595
|
-
}, is_pure=True,
|
|
1595
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1596
1596
|
|
|
1597
1597
|
|
|
1598
1598
|
@core.extern
|
|
1599
|
-
def fdim(arg0, arg1,
|
|
1599
|
+
def fdim(arg0, arg1, _semantic=None):
|
|
1600
1600
|
return core.extern_elementwise(
|
|
1601
1601
|
"", "", [arg0, arg1], {
|
|
1602
1602
|
(core.dtype("fp32"), core.dtype("fp32")): ("__nv_fdimf", core.dtype("fp32")),
|
|
1603
1603
|
(core.dtype("fp64"), core.dtype("fp64")): ("__nv_fdim", core.dtype("fp64")),
|
|
1604
|
-
}, is_pure=True,
|
|
1604
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1605
1605
|
|
|
1606
1606
|
|
|
1607
1607
|
@core.extern
|
|
1608
|
-
def ilogb(arg0,
|
|
1608
|
+
def ilogb(arg0, _semantic=None):
|
|
1609
1609
|
return core.extern_elementwise(
|
|
1610
1610
|
"", "", [arg0], {
|
|
1611
1611
|
(core.dtype("fp32"), ): ("__nv_ilogbf", core.dtype("int32")),
|
|
1612
1612
|
(core.dtype("fp64"), ): ("__nv_ilogb", core.dtype("int32")),
|
|
1613
|
-
}, is_pure=True,
|
|
1613
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1614
1614
|
|
|
1615
1615
|
|
|
1616
1616
|
@core.extern
|
|
1617
|
-
def logb(arg0,
|
|
1617
|
+
def logb(arg0, _semantic=None):
|
|
1618
1618
|
return core.extern_elementwise(
|
|
1619
1619
|
"", "", [arg0], {
|
|
1620
1620
|
(core.dtype("fp32"), ): ("__nv_logbf", core.dtype("fp32")),
|
|
1621
1621
|
(core.dtype("fp64"), ): ("__nv_logb", core.dtype("fp64")),
|
|
1622
|
-
}, is_pure=True,
|
|
1622
|
+
}, is_pure=True, _semantic=_semantic)
|
|
1623
1623
|
|
|
1624
1624
|
|
|
1625
1625
|
@core.extern
|
|
1626
|
-
def isfinited(arg0,
|
|
1626
|
+
def isfinited(arg0, _semantic=None):
|
|
1627
1627
|
return core.extern_elementwise("", "", [arg0], {
|
|
1628
1628
|
(core.dtype("fp64"), ): ("__nv_isfinited", core.dtype("int32")),
|
|
1629
|
-
}, is_pure=True,
|
|
1629
|
+
}, is_pure=True, _semantic=_semantic).to(core.int1, _semantic=_semantic)
|