triton-windows 3.3.1.post19__cp310-cp310-win_amd64.whl → 3.5.0.post21__cp310-cp310-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.

Files changed (225) hide show
  1. triton/_C/libtriton.pyd +0 -0
  2. triton/__init__.py +11 -2
  3. triton/_filecheck.py +97 -0
  4. triton/_internal_testing.py +95 -18
  5. triton/_utils.py +112 -21
  6. triton/backends/__init__.py +20 -23
  7. triton/backends/amd/__init__.py +0 -0
  8. triton/backends/amd/compiler.py +161 -119
  9. triton/backends/amd/driver.c +118 -46
  10. triton/backends/amd/driver.py +274 -96
  11. triton/backends/compiler.py +7 -21
  12. triton/backends/driver.py +13 -0
  13. triton/backends/nvidia/bin/ptxas.exe +0 -0
  14. triton/backends/nvidia/compiler.py +163 -106
  15. triton/backends/nvidia/driver.c +166 -101
  16. triton/backends/nvidia/driver.py +384 -202
  17. triton/compiler/__init__.py +5 -2
  18. triton/compiler/code_generator.py +439 -231
  19. triton/compiler/compiler.py +152 -84
  20. triton/experimental/__init__.py +0 -0
  21. triton/experimental/gluon/__init__.py +5 -0
  22. triton/experimental/gluon/_compiler.py +0 -0
  23. triton/experimental/gluon/_runtime.py +102 -0
  24. triton/experimental/gluon/language/__init__.py +119 -0
  25. triton/experimental/gluon/language/_core.py +490 -0
  26. triton/experimental/gluon/language/_layouts.py +583 -0
  27. triton/experimental/gluon/language/_math.py +20 -0
  28. triton/experimental/gluon/language/_semantic.py +380 -0
  29. triton/experimental/gluon/language/_standard.py +80 -0
  30. triton/experimental/gluon/language/amd/__init__.py +4 -0
  31. triton/experimental/gluon/language/amd/_layouts.py +96 -0
  32. triton/experimental/gluon/language/amd/cdna3/__init__.py +100 -0
  33. triton/experimental/gluon/language/amd/cdna4/__init__.py +48 -0
  34. triton/experimental/gluon/language/amd/cdna4/async_copy.py +151 -0
  35. triton/experimental/gluon/language/extra/__init__.py +3 -0
  36. triton/experimental/gluon/language/nvidia/__init__.py +4 -0
  37. triton/experimental/gluon/language/nvidia/ampere/__init__.py +3 -0
  38. triton/experimental/gluon/language/nvidia/ampere/async_copy.py +74 -0
  39. triton/experimental/gluon/language/nvidia/ampere/mbarrier.py +80 -0
  40. triton/experimental/gluon/language/nvidia/blackwell/__init__.py +387 -0
  41. triton/experimental/gluon/language/nvidia/blackwell/tma.py +52 -0
  42. triton/experimental/gluon/language/nvidia/hopper/__init__.py +132 -0
  43. triton/experimental/gluon/language/nvidia/hopper/mbarrier.py +34 -0
  44. triton/experimental/gluon/language/nvidia/hopper/tma.py +97 -0
  45. triton/experimental/gluon/nvidia/__init__.py +4 -0
  46. triton/experimental/gluon/nvidia/blackwell.py +3 -0
  47. triton/experimental/gluon/nvidia/hopper.py +45 -0
  48. triton/knobs.py +546 -0
  49. triton/language/__init__.py +50 -19
  50. triton/language/core.py +909 -572
  51. triton/language/extra/cuda/__init__.py +10 -7
  52. triton/language/extra/cuda/gdc.py +42 -0
  53. triton/language/extra/cuda/libdevice.py +394 -394
  54. triton/language/extra/cuda/utils.py +21 -21
  55. triton/language/extra/hip/__init__.py +3 -1
  56. triton/language/extra/hip/libdevice.py +120 -104
  57. triton/language/extra/hip/utils.py +35 -0
  58. triton/language/extra/libdevice.py +4 -0
  59. triton/language/math.py +65 -66
  60. triton/language/random.py +12 -2
  61. triton/language/semantic.py +1757 -1768
  62. triton/language/standard.py +127 -62
  63. triton/language/target_info.py +54 -0
  64. triton/runtime/_allocation.py +15 -3
  65. triton/runtime/_async_compile.py +55 -0
  66. triton/runtime/autotuner.py +117 -60
  67. triton/runtime/build.py +83 -17
  68. triton/runtime/cache.py +61 -47
  69. triton/runtime/driver.py +25 -47
  70. triton/runtime/interpreter.py +95 -50
  71. triton/runtime/jit.py +445 -248
  72. triton/runtime/tcc/include/_mingw.h +8 -10
  73. triton/runtime/tcc/include/assert.h +5 -0
  74. triton/runtime/tcc/include/errno.h +1 -1
  75. triton/runtime/tcc/include/float.h +21 -3
  76. triton/runtime/tcc/include/iso646.h +36 -0
  77. triton/runtime/tcc/include/limits.h +5 -0
  78. triton/runtime/tcc/include/malloc.h +2 -2
  79. triton/runtime/tcc/include/math.h +21 -261
  80. triton/runtime/tcc/include/stdalign.h +16 -0
  81. triton/runtime/tcc/include/stdarg.h +5 -70
  82. triton/runtime/tcc/include/stdatomic.h +171 -0
  83. triton/runtime/tcc/include/stddef.h +7 -19
  84. triton/runtime/tcc/include/stdlib.h +15 -4
  85. triton/runtime/tcc/include/stdnoreturn.h +7 -0
  86. triton/runtime/tcc/include/sys/stat.h +2 -2
  87. triton/runtime/tcc/include/sys/types.h +5 -0
  88. triton/runtime/tcc/include/tcc/tcc_libm.h +444 -27
  89. triton/runtime/tcc/include/tccdefs.h +342 -0
  90. triton/runtime/tcc/include/tgmath.h +89 -0
  91. triton/runtime/tcc/include/uchar.h +33 -0
  92. triton/runtime/tcc/include/unistd.h +1 -0
  93. triton/runtime/tcc/include/winapi/qos.h +72 -0
  94. triton/runtime/tcc/include/winapi/shellapi.h +59 -0
  95. triton/runtime/tcc/include/winapi/winbase.h +9 -2
  96. triton/runtime/tcc/include/winapi/wincon.h +8 -0
  97. triton/runtime/tcc/include/winapi/windows.h +1 -1
  98. triton/runtime/tcc/include/winapi/winnls.h +778 -0
  99. triton/runtime/tcc/include/winapi/winnt.h +9 -7
  100. triton/runtime/tcc/include/winapi/winsock2.h +1474 -0
  101. triton/runtime/tcc/include/winapi/ws2ipdef.h +21 -0
  102. triton/runtime/tcc/include/winapi/ws2tcpip.h +391 -0
  103. triton/runtime/tcc/lib/libtcc1.a +0 -0
  104. triton/runtime/tcc/lib/python314.def +1800 -0
  105. triton/runtime/tcc/lib/python314t.def +1809 -0
  106. triton/runtime/tcc/libtcc.dll +0 -0
  107. triton/runtime/tcc/tcc.exe +0 -0
  108. triton/testing.py +16 -12
  109. triton/tools/compile.py +62 -14
  110. triton/tools/disasm.py +3 -4
  111. triton/tools/extra/cuda/compile.c +1 -0
  112. triton/tools/extra/hip/compile.cpp +66 -0
  113. triton/tools/extra/hip/compile.h +13 -0
  114. triton/tools/ragged_tma.py +92 -0
  115. triton/tools/tensor_descriptor.py +34 -0
  116. triton/windows_utils.py +52 -81
  117. {triton_windows-3.3.1.post19.dist-info → triton_windows-3.5.0.post21.dist-info}/METADATA +8 -4
  118. triton_windows-3.5.0.post21.dist-info/RECORD +217 -0
  119. triton_windows-3.5.0.post21.dist-info/entry_points.txt +3 -0
  120. triton_windows-3.5.0.post21.dist-info/licenses/LICENSE +23 -0
  121. triton_windows-3.5.0.post21.dist-info/top_level.txt +1 -0
  122. triton/backends/amd/include/hip/amd_detail/amd_channel_descriptor.h +0 -358
  123. triton/backends/amd/include/hip/amd_detail/amd_device_functions.h +0 -1010
  124. triton/backends/amd/include/hip/amd_detail/amd_hip_atomic.h +0 -1638
  125. triton/backends/amd/include/hip/amd_detail/amd_hip_bf16.h +0 -1814
  126. triton/backends/amd/include/hip/amd_detail/amd_hip_bfloat16.h +0 -293
  127. triton/backends/amd/include/hip/amd_detail/amd_hip_common.h +0 -32
  128. triton/backends/amd/include/hip/amd_detail/amd_hip_complex.h +0 -174
  129. triton/backends/amd/include/hip/amd_detail/amd_hip_cooperative_groups.h +0 -835
  130. triton/backends/amd/include/hip/amd_detail/amd_hip_fp16.h +0 -1809
  131. triton/backends/amd/include/hip/amd_detail/amd_hip_fp8.h +0 -1391
  132. triton/backends/amd/include/hip/amd_detail/amd_hip_gl_interop.h +0 -108
  133. triton/backends/amd/include/hip/amd_detail/amd_hip_math_constants.h +0 -124
  134. triton/backends/amd/include/hip/amd_detail/amd_hip_runtime.h +0 -405
  135. triton/backends/amd/include/hip/amd_detail/amd_hip_runtime_pt_api.h +0 -196
  136. triton/backends/amd/include/hip/amd_detail/amd_hip_unsafe_atomics.h +0 -565
  137. triton/backends/amd/include/hip/amd_detail/amd_hip_vector_types.h +0 -2226
  138. triton/backends/amd/include/hip/amd_detail/amd_math_functions.h +0 -104
  139. triton/backends/amd/include/hip/amd_detail/amd_surface_functions.h +0 -244
  140. triton/backends/amd/include/hip/amd_detail/amd_warp_functions.h +0 -538
  141. triton/backends/amd/include/hip/amd_detail/amd_warp_sync_functions.h +0 -288
  142. triton/backends/amd/include/hip/amd_detail/concepts.hpp +0 -30
  143. triton/backends/amd/include/hip/amd_detail/device_library_decls.h +0 -133
  144. triton/backends/amd/include/hip/amd_detail/functional_grid_launch.hpp +0 -218
  145. triton/backends/amd/include/hip/amd_detail/grid_launch.h +0 -67
  146. triton/backends/amd/include/hip/amd_detail/grid_launch.hpp +0 -50
  147. triton/backends/amd/include/hip/amd_detail/grid_launch_GGL.hpp +0 -26
  148. triton/backends/amd/include/hip/amd_detail/helpers.hpp +0 -137
  149. triton/backends/amd/include/hip/amd_detail/hip_api_trace.hpp +0 -1446
  150. triton/backends/amd/include/hip/amd_detail/hip_assert.h +0 -101
  151. triton/backends/amd/include/hip/amd_detail/hip_cooperative_groups_helper.h +0 -242
  152. triton/backends/amd/include/hip/amd_detail/hip_fp16_gcc.h +0 -254
  153. triton/backends/amd/include/hip/amd_detail/hip_fp16_math_fwd.h +0 -96
  154. triton/backends/amd/include/hip/amd_detail/hip_ldg.h +0 -100
  155. triton/backends/amd/include/hip/amd_detail/hip_prof_str.h +0 -10570
  156. triton/backends/amd/include/hip/amd_detail/hip_runtime_prof.h +0 -78
  157. triton/backends/amd/include/hip/amd_detail/host_defines.h +0 -184
  158. triton/backends/amd/include/hip/amd_detail/hsa_helpers.hpp +0 -102
  159. triton/backends/amd/include/hip/amd_detail/macro_based_grid_launch.hpp +0 -798
  160. triton/backends/amd/include/hip/amd_detail/math_fwd.h +0 -698
  161. triton/backends/amd/include/hip/amd_detail/ockl_image.h +0 -177
  162. triton/backends/amd/include/hip/amd_detail/program_state.hpp +0 -107
  163. triton/backends/amd/include/hip/amd_detail/texture_fetch_functions.h +0 -491
  164. triton/backends/amd/include/hip/amd_detail/texture_indirect_functions.h +0 -478
  165. triton/backends/amd/include/hip/channel_descriptor.h +0 -39
  166. triton/backends/amd/include/hip/device_functions.h +0 -38
  167. triton/backends/amd/include/hip/driver_types.h +0 -468
  168. triton/backends/amd/include/hip/hip_bf16.h +0 -36
  169. triton/backends/amd/include/hip/hip_bfloat16.h +0 -44
  170. triton/backends/amd/include/hip/hip_common.h +0 -100
  171. triton/backends/amd/include/hip/hip_complex.h +0 -38
  172. triton/backends/amd/include/hip/hip_cooperative_groups.h +0 -46
  173. triton/backends/amd/include/hip/hip_deprecated.h +0 -95
  174. triton/backends/amd/include/hip/hip_ext.h +0 -161
  175. triton/backends/amd/include/hip/hip_fp16.h +0 -36
  176. triton/backends/amd/include/hip/hip_fp8.h +0 -33
  177. triton/backends/amd/include/hip/hip_gl_interop.h +0 -32
  178. triton/backends/amd/include/hip/hip_hcc.h +0 -24
  179. triton/backends/amd/include/hip/hip_math_constants.h +0 -36
  180. triton/backends/amd/include/hip/hip_profile.h +0 -27
  181. triton/backends/amd/include/hip/hip_runtime.h +0 -75
  182. triton/backends/amd/include/hip/hip_runtime_api.h +0 -9261
  183. triton/backends/amd/include/hip/hip_texture_types.h +0 -29
  184. triton/backends/amd/include/hip/hip_vector_types.h +0 -41
  185. triton/backends/amd/include/hip/hip_version.h +0 -17
  186. triton/backends/amd/include/hip/hiprtc.h +0 -421
  187. triton/backends/amd/include/hip/library_types.h +0 -78
  188. triton/backends/amd/include/hip/math_functions.h +0 -42
  189. triton/backends/amd/include/hip/surface_types.h +0 -63
  190. triton/backends/amd/include/hip/texture_types.h +0 -194
  191. triton/backends/amd/include/hsa/Brig.h +0 -1131
  192. triton/backends/amd/include/hsa/amd_hsa_common.h +0 -91
  193. triton/backends/amd/include/hsa/amd_hsa_elf.h +0 -462
  194. triton/backends/amd/include/hsa/amd_hsa_kernel_code.h +0 -269
  195. triton/backends/amd/include/hsa/amd_hsa_queue.h +0 -109
  196. triton/backends/amd/include/hsa/amd_hsa_signal.h +0 -80
  197. triton/backends/amd/include/hsa/hsa.h +0 -5738
  198. triton/backends/amd/include/hsa/hsa_amd_tool.h +0 -91
  199. triton/backends/amd/include/hsa/hsa_api_trace.h +0 -579
  200. triton/backends/amd/include/hsa/hsa_api_trace_version.h +0 -68
  201. triton/backends/amd/include/hsa/hsa_ext_amd.h +0 -3146
  202. triton/backends/amd/include/hsa/hsa_ext_finalize.h +0 -531
  203. triton/backends/amd/include/hsa/hsa_ext_image.h +0 -1454
  204. triton/backends/amd/include/hsa/hsa_ven_amd_aqlprofile.h +0 -488
  205. triton/backends/amd/include/hsa/hsa_ven_amd_loader.h +0 -667
  206. triton/backends/amd/include/hsa/hsa_ven_amd_pc_sampling.h +0 -416
  207. triton/backends/amd/include/roctracer/ext/prof_protocol.h +0 -107
  208. triton/backends/amd/include/roctracer/hip_ostream_ops.h +0 -4515
  209. triton/backends/amd/include/roctracer/hsa_ostream_ops.h +0 -1727
  210. triton/backends/amd/include/roctracer/hsa_prof_str.h +0 -3059
  211. triton/backends/amd/include/roctracer/roctracer.h +0 -779
  212. triton/backends/amd/include/roctracer/roctracer_ext.h +0 -81
  213. triton/backends/amd/include/roctracer/roctracer_hcc.h +0 -24
  214. triton/backends/amd/include/roctracer/roctracer_hip.h +0 -37
  215. triton/backends/amd/include/roctracer/roctracer_hsa.h +0 -112
  216. triton/backends/amd/include/roctracer/roctracer_plugin.h +0 -137
  217. triton/backends/amd/include/roctracer/roctracer_roctx.h +0 -67
  218. triton/backends/amd/include/roctracer/roctx.h +0 -229
  219. triton/language/_utils.py +0 -21
  220. triton/language/extra/cuda/_experimental_tma.py +0 -106
  221. triton/runtime/tcc/lib/libtcc1-64.a +0 -0
  222. triton/tools/experimental_descriptor.py +0 -32
  223. triton_windows-3.3.1.post19.dist-info/RECORD +0 -260
  224. triton_windows-3.3.1.post19.dist-info/top_level.txt +0 -14
  225. {triton_windows-3.3.1.post19.dist-info → triton_windows-3.5.0.post21.dist-info}/WHEEL +0 -0
@@ -1,478 +0,0 @@
1
- /*
2
- Copyright (c) 2015 - 2023 Advanced Micro Devices, Inc. All rights reserved.
3
-
4
- Permission is hereby granted, free of charge, to any person obtaining a copy
5
- of this software and associated documentation files (the "Software"), to deal
6
- in the Software without restriction, including without limitation the rights
7
- to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8
- copies of the Software, and to permit persons to whom the Software is
9
- furnished to do so, subject to the following conditions:
10
-
11
- The above copyright notice and this permission notice shall be included in
12
- all copies or substantial portions of the Software.
13
-
14
- THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15
- IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16
- FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
17
- AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18
- LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
19
- OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
20
- THE SOFTWARE.
21
- */
22
-
23
- #pragma once
24
-
25
- #if defined(__cplusplus)
26
-
27
- #if !defined(__HIPCC_RTC__)
28
- #include <hip/hip_vector_types.h>
29
- #include <hip/hip_texture_types.h>
30
- #include <hip/amd_detail/texture_fetch_functions.h>
31
- #include <hip/amd_detail/ockl_image.h>
32
- #include <type_traits>
33
- #endif // !defined(__HIPCC_RTC__)
34
-
35
- #define TEXTURE_OBJECT_PARAMETERS_INIT \
36
- unsigned int ADDRESS_SPACE_CONSTANT* i = (unsigned int ADDRESS_SPACE_CONSTANT*)textureObject; \
37
- unsigned int ADDRESS_SPACE_CONSTANT* s = i + HIP_SAMPLER_OBJECT_OFFSET_DWORD;
38
-
39
- template <
40
- typename T,
41
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
42
- static __device__ __hip_img_chk__ T tex1Dfetch(hipTextureObject_t textureObject, int x)
43
- {
44
- TEXTURE_OBJECT_PARAMETERS_INIT
45
- auto tmp = __ockl_image_load_1Db(i, x);
46
- return __hipMapFrom<T>(tmp);
47
- }
48
-
49
- template <
50
- typename T,
51
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
52
- static __device__ __hip_img_chk__ void tex1Dfetch(T *ptr, hipTextureObject_t textureObject, int x)
53
- {
54
- *ptr = tex1Dfetch<T>(textureObject, x);
55
- }
56
-
57
- template <
58
- typename T,
59
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
60
- static __device__ __hip_img_chk__ T tex1D(hipTextureObject_t textureObject, float x)
61
- {
62
- TEXTURE_OBJECT_PARAMETERS_INIT
63
- auto tmp = __ockl_image_sample_1D(i, s, x);
64
- return __hipMapFrom<T>(tmp);
65
- }
66
-
67
- template <
68
- typename T,
69
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
70
- static __device__ __hip_img_chk__ void tex1D(T *ptr, hipTextureObject_t textureObject, float x)
71
- {
72
- *ptr = tex1D<T>(textureObject, x);
73
- }
74
-
75
- template <
76
- typename T,
77
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
78
- static __device__ __hip_img_chk__ T tex2D(hipTextureObject_t textureObject, float x, float y)
79
- {
80
- TEXTURE_OBJECT_PARAMETERS_INIT
81
- auto tmp = __ockl_image_sample_2D(i, s, float2(x, y).data);
82
- return __hipMapFrom<T>(tmp);
83
- }
84
-
85
- template <
86
- typename T,
87
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
88
- static __device__ __hip_img_chk__ void tex2D(T *ptr, hipTextureObject_t textureObject, float x, float y)
89
- {
90
- *ptr = tex2D<T>(textureObject, x, y);
91
- }
92
-
93
- template <
94
- typename T,
95
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
96
- static __device__ __hip_img_chk__ T tex3D(hipTextureObject_t textureObject, float x, float y, float z)
97
- {
98
- TEXTURE_OBJECT_PARAMETERS_INIT
99
- auto tmp = __ockl_image_sample_3D(i, s, float4(x, y, z, 0.0f).data);
100
- return __hipMapFrom<T>(tmp);
101
- }
102
-
103
- template <
104
- typename T,
105
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
106
- static __device__ __hip_img_chk__ void tex3D(T *ptr, hipTextureObject_t textureObject, float x, float y, float z)
107
- {
108
- *ptr = tex3D<T>(textureObject, x, y, z);
109
- }
110
-
111
- template <
112
- typename T,
113
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
114
- static __device__ __hip_img_chk__ T tex1DLayered(hipTextureObject_t textureObject, float x, int layer)
115
- {
116
- TEXTURE_OBJECT_PARAMETERS_INIT
117
- auto tmp = __ockl_image_sample_1Da(i, s, float2(x, layer).data);
118
- return __hipMapFrom<T>(tmp);
119
- }
120
-
121
- template <
122
- typename T,
123
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
124
- static __device__ __hip_img_chk__ void tex1DLayered(T *ptr, hipTextureObject_t textureObject, float x, int layer)
125
- {
126
- *ptr = tex1DLayered<T>(textureObject, x, layer);
127
- }
128
-
129
- template <
130
- typename T,
131
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
132
- static __device__ __hip_img_chk__ T tex2DLayered(hipTextureObject_t textureObject, float x, float y, int layer)
133
- {
134
- TEXTURE_OBJECT_PARAMETERS_INIT
135
- auto tmp = __ockl_image_sample_2Da(i, s, float4(x, y, layer, 0.0f).data);
136
- return __hipMapFrom<T>(tmp);
137
- }
138
-
139
- template <
140
- typename T,
141
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
142
- static __device__ __hip_img_chk__ void tex2DLayered(T *ptr, hipTextureObject_t textureObject, float x, float y, int layer)
143
- {
144
- *ptr = tex1DLayered<T>(textureObject, x, y, layer);
145
- }
146
-
147
- template <
148
- typename T,
149
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
150
- static __device__ __hip_img_chk__ T texCubemap(hipTextureObject_t textureObject, float x, float y, float z)
151
- {
152
- TEXTURE_OBJECT_PARAMETERS_INIT
153
- auto tmp = __ockl_image_sample_CM(i, s, float4(x, y, z, 0.0f).data);
154
- return __hipMapFrom<T>(tmp);
155
- }
156
-
157
- template <
158
- typename T,
159
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
160
- static __device__ __hip_img_chk__ void texCubemap(T *ptr, hipTextureObject_t textureObject, float x, float y, float z)
161
- {
162
- *ptr = texCubemap<T>(textureObject, x, y, z);
163
- }
164
-
165
- template <
166
- typename T,
167
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
168
- static __device__ __hip_img_chk__ T texCubemapLayered(hipTextureObject_t textureObject, float x, float y, float z, int layer)
169
- {
170
- TEXTURE_OBJECT_PARAMETERS_INIT
171
- auto tmp = __ockl_image_sample_CMa(i, s, float4(x, y, z, layer).data);
172
- return __hipMapFrom<T>(tmp);
173
- }
174
-
175
- template <
176
- typename T,
177
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
178
- static __device__ __hip_img_chk__ void texCubemapLayered(T *ptr, hipTextureObject_t textureObject, float x, float y, float z, int layer)
179
- {
180
- *ptr = texCubemapLayered<T>(textureObject, x, y, z, layer);
181
- }
182
-
183
- template <
184
- typename T,
185
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
186
- static __device__ __hip_img_chk__ T tex2Dgather(hipTextureObject_t textureObject, float x, float y, int comp = 0)
187
- {
188
- TEXTURE_OBJECT_PARAMETERS_INIT
189
- switch (comp) {
190
- case 1: {
191
- auto tmp = __ockl_image_gather4r_2D(i, s, float2(x, y).data);
192
- return __hipMapFrom<T>(tmp);
193
- break;
194
- }
195
- case 2: {
196
- auto tmp = __ockl_image_gather4g_2D(i, s, float2(x, y).data);
197
- return __hipMapFrom<T>(tmp);
198
- break;
199
- }
200
- case 3: {
201
- auto tmp = __ockl_image_gather4b_2D(i, s, float2(x, y).data);
202
- return __hipMapFrom<T>(tmp);
203
- break;
204
- }
205
- default: {
206
- auto tmp = __ockl_image_gather4a_2D(i, s, float2(x, y).data);
207
- return __hipMapFrom<T>(tmp);
208
- break;
209
- }
210
- }
211
- return {};
212
- }
213
-
214
- template <
215
- typename T,
216
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
217
- static __device__ __hip_img_chk__ void tex2Dgather(T *ptr, hipTextureObject_t textureObject, float x, float y, int comp = 0)
218
- {
219
- *ptr = texCubemapLayered<T>(textureObject, x, y, comp);
220
- }
221
-
222
- template <
223
- typename T,
224
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
225
- static __device__ __hip_img_chk__ T tex1DLod(hipTextureObject_t textureObject, float x, float level)
226
- {
227
- TEXTURE_OBJECT_PARAMETERS_INIT
228
- auto tmp = __ockl_image_sample_lod_1D(i, s, x, level);
229
- return __hipMapFrom<T>(tmp);
230
- }
231
-
232
- template <
233
- typename T,
234
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
235
- static __device__ __hip_img_chk__ void tex1DLod(T *ptr, hipTextureObject_t textureObject, float x, float level)
236
- {
237
- *ptr = tex1DLod<T>(textureObject, x, level);
238
- }
239
-
240
- template <
241
- typename T,
242
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
243
- static __device__ __hip_img_chk__ T tex2DLod(hipTextureObject_t textureObject, float x, float y, float level)
244
- {
245
- TEXTURE_OBJECT_PARAMETERS_INIT
246
- auto tmp = __ockl_image_sample_lod_2D(i, s, float2(x, y).data, level);
247
- return __hipMapFrom<T>(tmp);
248
- }
249
-
250
- template <
251
- typename T,
252
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
253
- static __device__ __hip_img_chk__ void tex2DLod(T *ptr, hipTextureObject_t textureObject, float x, float y, float level)
254
- {
255
- *ptr = tex2DLod<T>(textureObject, x, y, level);
256
- }
257
-
258
- template <
259
- typename T,
260
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
261
- static __device__ __hip_img_chk__ T tex3DLod(hipTextureObject_t textureObject, float x, float y, float z, float level)
262
- {
263
- TEXTURE_OBJECT_PARAMETERS_INIT
264
- auto tmp = __ockl_image_sample_lod_3D(i, s, float4(x, y, z, 0.0f).data, level);
265
- return __hipMapFrom<T>(tmp);
266
- }
267
-
268
- template <
269
- typename T,
270
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
271
- static __device__ __hip_img_chk__ void tex3DLod(T *ptr, hipTextureObject_t textureObject, float x, float y, float z, float level)
272
- {
273
- *ptr = tex3DLod<T>(textureObject, x, y, z, level);
274
- }
275
-
276
- template <
277
- typename T,
278
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
279
- static __device__ __hip_img_chk__ T tex1DLayeredLod(hipTextureObject_t textureObject, float x, int layer, float level)
280
- {
281
- TEXTURE_OBJECT_PARAMETERS_INIT
282
- auto tmp = __ockl_image_sample_1Da(i, s, float2(x, layer).data);
283
- return __hipMapFrom<T>(tmp);
284
- }
285
-
286
- template <
287
- typename T,
288
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
289
- static __device__ __hip_img_chk__ void tex1DLayeredLod(T *ptr, hipTextureObject_t textureObject, float x, int layer, float level)
290
- {
291
- *ptr = tex1DLayeredLod<T>(textureObject, x, layer, level);
292
- }
293
-
294
- template <
295
- typename T,
296
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
297
- static __device__ __hip_img_chk__ T tex2DLayeredLod(hipTextureObject_t textureObject, float x, float y, int layer, float level)
298
- {
299
- TEXTURE_OBJECT_PARAMETERS_INIT
300
- auto tmp = __ockl_image_sample_2Da(i, s, float4(x, y, layer, 0.0f).data);
301
- return __hipMapFrom<T>(tmp);
302
- }
303
-
304
- template <
305
- typename T,
306
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
307
- static __device__ __hip_img_chk__ void tex2DLayeredLod(T *ptr, hipTextureObject_t textureObject, float x, float y, int layer, float level)
308
- {
309
- *ptr = tex2DLayeredLod<T>(textureObject, x, y, layer, level);
310
- }
311
-
312
- template <
313
- typename T,
314
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
315
- static __device__ __hip_img_chk__ T texCubemapLod(hipTextureObject_t textureObject, float x, float y, float z, float level)
316
- {
317
- TEXTURE_OBJECT_PARAMETERS_INIT
318
- auto tmp = __ockl_image_sample_lod_CM(i, s, float4(x, y, z, 0.0f).data, level);
319
- return __hipMapFrom<T>(tmp);
320
- }
321
-
322
- template <
323
- typename T,
324
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
325
- static __device__ __hip_img_chk__ void texCubemapLod(T *ptr, hipTextureObject_t textureObject, float x, float y, float z, float level)
326
- {
327
- *ptr = texCubemapLod<T>(textureObject, x, y, z, level);
328
- }
329
-
330
- template <
331
- typename T,
332
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
333
- static __device__ __hip_img_chk__ T texCubemapGrad(hipTextureObject_t textureObject, float x, float y, float z, float4 dPdx, float4 dPdy)
334
- {
335
- TEXTURE_OBJECT_PARAMETERS_INIT
336
- // TODO missing in device libs.
337
- // auto tmp = __ockl_image_sample_grad_CM(i, s, float4(x, y, z, 0.0f).data, float4(dPdx.x, dPdx.y, dPdx.z, 0.0f).data, float4(dPdy.x, dPdy.y, dPdy.z, 0.0f).data);
338
- // return __hipMapFrom<T>(tmp);
339
- return {};
340
- }
341
-
342
- template <
343
- typename T,
344
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
345
- static __device__ __hip_img_chk__ void texCubemapGrad(T *ptr, hipTextureObject_t textureObject, float x, float y, float z, float4 dPdx, float4 dPdy)
346
- {
347
- *ptr = texCubemapGrad<T>(textureObject, x, y, z, dPdx, dPdy);
348
- }
349
-
350
- template <
351
- typename T,
352
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
353
- static __device__ __hip_img_chk__ T texCubemapLayeredLod(hipTextureObject_t textureObject, float x, float y, float z, int layer, float level)
354
- {
355
- TEXTURE_OBJECT_PARAMETERS_INIT
356
- auto tmp = __ockl_image_sample_lod_CMa(i, s, float4(x, y, z, layer).data, level);
357
- return __hipMapFrom<T>(tmp);
358
- }
359
-
360
- template <
361
- typename T,
362
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
363
- static __device__ __hip_img_chk__ void texCubemapLayeredLod(T *ptr, hipTextureObject_t textureObject, float x, float y, float z, int layer, float level)
364
- {
365
- *ptr = texCubemapLayeredLod<T>(textureObject, x, y, z, layer, level);
366
- }
367
-
368
- template <
369
- typename T,
370
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
371
- static __device__ __hip_img_chk__ T tex1DGrad(hipTextureObject_t textureObject, float x, float dPdx, float dPdy)
372
- {
373
- TEXTURE_OBJECT_PARAMETERS_INIT
374
- auto tmp = __ockl_image_sample_grad_1D(i, s, x, dPdx, dPdy);
375
- return __hipMapFrom<T>(tmp);
376
- }
377
-
378
- template <
379
- typename T,
380
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
381
- static __device__ __hip_img_chk__ void tex1DGrad(T *ptr, hipTextureObject_t textureObject, float x, float dPdx, float dPdy)
382
- {
383
- *ptr = tex1DGrad<T>(textureObject, x, dPdx, dPdy);
384
- }
385
-
386
- template <
387
- typename T,
388
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
389
- static __device__ __hip_img_chk__ T tex2DGrad(hipTextureObject_t textureObject, float x, float y, float2 dPdx, float2 dPdy)
390
- {
391
- TEXTURE_OBJECT_PARAMETERS_INIT
392
- auto tmp = __ockl_image_sample_grad_2D(i, s, float2(x, y).data, float2(dPdx.x, dPdx.y).data, float2(dPdy.x, dPdy.y).data);
393
- return __hipMapFrom<T>(tmp);
394
- }
395
-
396
- template <
397
- typename T,
398
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
399
- static __device__ __hip_img_chk__ void tex2DGrad(T *ptr, hipTextureObject_t textureObject, float x, float y, float2 dPdx, float2 dPdy)
400
- {
401
- *ptr = tex2DGrad<T>(textureObject, x, y, dPdx, dPdy);
402
- }
403
-
404
- template <
405
- typename T,
406
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
407
- static __device__ __hip_img_chk__ T tex3DGrad(hipTextureObject_t textureObject, float x, float y, float z, float4 dPdx, float4 dPdy)
408
- {
409
- TEXTURE_OBJECT_PARAMETERS_INIT
410
- auto tmp = __ockl_image_sample_grad_3D(i, s, float4(x, y, z, 0.0f).data, float4(dPdx.x, dPdx.y, dPdx.z, 0.0f).data, float4(dPdy.x, dPdy.y, dPdy.z, 0.0f).data);
411
- return __hipMapFrom<T>(tmp);
412
- }
413
-
414
- template <
415
- typename T,
416
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
417
- static __device__ __hip_img_chk__ void tex3DGrad(T *ptr, hipTextureObject_t textureObject, float x, float y, float z, float4 dPdx, float4 dPdy)
418
- {
419
- *ptr = tex3DGrad<T>(textureObject, x, y, z, dPdx, dPdy);
420
- }
421
-
422
- template <
423
- typename T,
424
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
425
- static __device__ __hip_img_chk__ T tex1DLayeredGrad(hipTextureObject_t textureObject, float x, int layer, float dPdx, float dPdy)
426
- {
427
- TEXTURE_OBJECT_PARAMETERS_INIT
428
- auto tmp = __ockl_image_sample_grad_1Da(i, s, float2(x, layer).data, dPdx, dPdy);
429
- return __hipMapFrom<T>(tmp);
430
- }
431
-
432
- template <
433
- typename T,
434
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
435
- static __device__ __hip_img_chk__ void tex1DLayeredGrad(T *ptr, hipTextureObject_t textureObject, float x, int layer, float dPdx, float dPdy)
436
- {
437
- *ptr = tex1DLayeredGrad<T>(textureObject, x, layer, dPdx, dPdy);
438
- }
439
-
440
- template <
441
- typename T,
442
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
443
- static __device__ __hip_img_chk__ T tex2DLayeredGrad(hipTextureObject_t textureObject, float x, float y, int layer, float2 dPdx, float2 dPdy)
444
- {
445
- TEXTURE_OBJECT_PARAMETERS_INIT
446
- auto tmp = __ockl_image_sample_grad_2Da(i, s, float4(x, y, layer, 0.0f).data, float2(dPdx.x, dPdx.y).data, float2(dPdy.x, dPdy.y).data);
447
- return __hipMapFrom<T>(tmp);
448
- }
449
-
450
- template <
451
- typename T,
452
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
453
- static __device__ __hip_img_chk__ void tex2DLayeredGrad(T *ptr, hipTextureObject_t textureObject, float x, float y, int layer, float2 dPdx, float2 dPdy)
454
- {
455
- *ptr = tex2DLayeredGrad<T>(textureObject, x, y, layer, dPdx, dPdy);
456
- }
457
-
458
- template <
459
- typename T,
460
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
461
- static __device__ __hip_img_chk__ T texCubemapLayeredGrad(hipTextureObject_t textureObject, float x, float y, float z, int layer, float4 dPdx, float4 dPdy)
462
- {
463
- TEXTURE_OBJECT_PARAMETERS_INIT
464
- // TODO missing in device libs.
465
- // auto tmp = __ockl_image_sample_grad_CMa(i, s, float4(x, y, z, layer).data, float4(dPdx.x, dPdx.y, dPdx.z, 0.0f).data, float4(dPdy.x, dPdy.y, dPdy.z, 0.0f).data);
466
- // return __hipMapFrom<T>(tmp);
467
- return {};
468
- }
469
-
470
- template <
471
- typename T,
472
- typename std::enable_if<__hip_is_tex_surf_channel_type<T>::value>::type* = nullptr>
473
- static __device__ __hip_img_chk__ void texCubemapLayeredGrad(T *ptr, hipTextureObject_t textureObject, float x, float y, float z, int layer, float4 dPdx, float4 dPdy)
474
- {
475
- *ptr = texCubemapLayeredGrad<T>(textureObject, x, y, z, layer, dPdx, dPdy);
476
- }
477
-
478
- #endif
@@ -1,39 +0,0 @@
1
- /*
2
- Copyright (c) 2015 - 2021 Advanced Micro Devices, Inc. All rights reserved.
3
-
4
- Permission is hereby granted, free of charge, to any person obtaining a copy
5
- of this software and associated documentation files (the "Software"), to deal
6
- in the Software without restriction, including without limitation the rights
7
- to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8
- copies of the Software, and to permit persons to whom the Software is
9
- furnished to do so, subject to the following conditions:
10
-
11
- The above copyright notice and this permission notice shall be included in
12
- all copies or substantial portions of the Software.
13
-
14
- THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15
- IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16
- FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
17
- AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18
- LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
19
- OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
20
- THE SOFTWARE.
21
- */
22
-
23
- #ifndef HIP_INCLUDE_HIP_CHANNEL_DESCRIPTOR_H
24
- #define HIP_INCLUDE_HIP_CHANNEL_DESCRIPTOR_H
25
-
26
- // Some standard header files, these are included by hc.hpp and so want to make them avail on both
27
- // paths to provide a consistent include env and avoid "missing symbol" errors that only appears
28
- // on NVCC path:
29
-
30
-
31
- #if defined(__HIP_PLATFORM_AMD__) && !defined(__HIP_PLATFORM_NVIDIA__)
32
- #include <hip/amd_detail/amd_channel_descriptor.h>
33
- #elif !defined(__HIP_PLATFORM_AMD__) && defined(__HIP_PLATFORM_NVIDIA__)
34
- #include <hip/nvidia_detail/nvidia_channel_descriptor.h>
35
- #else
36
- #error("Must define exactly one of __HIP_PLATFORM_AMD__ or __HIP_PLATFORM_NVIDIA__");
37
- #endif
38
-
39
- #endif
@@ -1,38 +0,0 @@
1
- /*
2
- Copyright (c) 2015 - 2023 Advanced Micro Devices, Inc. All rights reserved.
3
-
4
- Permission is hereby granted, free of charge, to any person obtaining a copy
5
- of this software and associated documentation files (the "Software"), to deal
6
- in the Software without restriction, including without limitation the rights
7
- to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8
- copies of the Software, and to permit persons to whom the Software is
9
- furnished to do so, subject to the following conditions:
10
-
11
- The above copyright notice and this permission notice shall be included in
12
- all copies or substantial portions of the Software.
13
-
14
- THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15
- IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16
- FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
17
- AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18
- LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
19
- OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
20
- THE SOFTWARE.
21
- */
22
-
23
- #ifndef HIP_INCLUDE_HIP_DEVICE_FUNCTIONS_H
24
- #define HIP_INCLUDE_HIP_DEVICE_FUNCTIONS_H
25
-
26
- #if !defined(__HIPCC_RTC__)
27
- #include <hip/hip_common.h>
28
- #endif
29
-
30
- #if defined(__HIP_PLATFORM_AMD__) && !defined(__HIP_PLATFORM_NVIDIA__)
31
- #include <hip/amd_detail/amd_device_functions.h>
32
- #elif !defined(__HIP_PLATFORM_AMD__) && defined(__HIP_PLATFORM_NVIDIA__)
33
- #include <device_functions.h>
34
- #else
35
- #error("Must define exactly one of __HIP_PLATFORM_AMD__ or __HIP_PLATFORM_NVIDIA__");
36
- #endif
37
-
38
- #endif