Coverage for cuda / core / _kernel_arg_handler.pyx: 85.45%

1# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. (empty)

2# 

3# SPDX-License-Identifier: Apache-2.0 

4  

5from cpython.mem cimport PyMem_Malloc, PyMem_Free 

6from libc.stdint cimport (intptr_t, 

7 int8_t, int16_t, int32_t, int64_t, 

8 uint8_t, uint16_t, uint32_t, uint64_t,) 

9from libcpp cimport bool as cpp_bool 

10from libcpp.complex cimport complex as cpp_complex 

11from libcpp cimport nullptr 

12from libcpp cimport vector 

13  

14import ctypes (empty)

15  

16import numpy (empty)

17  

18from cuda.core._memory import Buffer (empty)

19from cuda.core._tensor_map import TensorMapDescriptor as _TensorMapDescriptor_py (empty)

20from cuda.core._tensor_map cimport TensorMapDescriptor 

21from cuda.core._utils.cuda_utils import driver (empty)

22from cuda.bindings cimport cydriver 

23  

24  

25ctypedef cpp_complex.complex[float] cpp_single_complex 

26ctypedef cpp_complex.complex[double] cpp_double_complex 

27  

28  

29# We need an identifier for fp16 for copying scalars on the host. This is a minimal 

30# implementation borrowed from cuda_fp16.h. 

31cdef extern from *: 

32 """ 

33 #if __cplusplus >= 201103L 

34 #define __CUDA_ALIGN__(n) alignas(n) /* C++11 kindly gives us a keyword for this */ 

35 #else 

36 #if defined(__GNUC__) 

37 #define __CUDA_ALIGN__(n) __attribute__ ((aligned(n))) 

38 #elif defined(_MSC_VER) 

39 #define __CUDA_ALIGN__(n) __declspec(align(n)) 

40 #else 

41 #define __CUDA_ALIGN__(n) 

42 #endif /* defined(__GNUC__) */ 

43 #endif /* __cplusplus >= 201103L */ 

44  

45 typedef struct __CUDA_ALIGN__(2) { 

46 /** 

47 * Storage field contains bits representation of the \p half floating-point number. 

48 */ 

49 unsigned short x; 

50 } __half_raw; 

51 """ 

52 ctypedef struct __half_raw: 

53 unsigned short x 

54  

55  

56ctypedef fused supported_type: 

57 cpp_bool 

58 int8_t 

59 int16_t 

60 int32_t 

61 int64_t 

62 uint8_t 

63 uint16_t 

64 uint32_t 

65 uint64_t 

66 __half_raw 

67 float 

68 double 

69 intptr_t 

70 cpp_single_complex 

71 cpp_double_complex 

72  

73  

74# cache ctypes/numpy type objects to avoid attribute access 

75cdef object ctypes_bool = ctypes.c_bool (empty)

76cdef object ctypes_int8 = ctypes.c_int8 (empty)

77cdef object ctypes_int16 = ctypes.c_int16 (empty)

78cdef object ctypes_int32 = ctypes.c_int32 (empty)

79cdef object ctypes_int64 = ctypes.c_int64 (empty)

80cdef object ctypes_uint8 = ctypes.c_uint8 (empty)

81cdef object ctypes_uint16 = ctypes.c_uint16 (empty)

82cdef object ctypes_uint32 = ctypes.c_uint32 (empty)

83cdef object ctypes_uint64 = ctypes.c_uint64 (empty)

84cdef object ctypes_float = ctypes.c_float (empty)

85cdef object ctypes_double = ctypes.c_double (empty)

86cdef object numpy_bool = numpy.bool_ (empty)

87cdef object numpy_int8 = numpy.int8 (empty)

88cdef object numpy_int16 = numpy.int16 (empty)

89cdef object numpy_int32 = numpy.int32 (empty)

90cdef object numpy_int64 = numpy.int64 (empty)

91cdef object numpy_uint8 = numpy.uint8 (empty)

92cdef object numpy_uint16 = numpy.uint16 (empty)

93cdef object numpy_uint32 = numpy.uint32 (empty)

94cdef object numpy_uint64 = numpy.uint64 (empty)

95cdef object numpy_float16 = numpy.float16 (empty)

96cdef object numpy_float32 = numpy.float32 (empty)

97cdef object numpy_float64 = numpy.float64 (empty)

98cdef object numpy_complex64 = numpy.complex64 (empty)

99cdef object numpy_complex128 = numpy.complex128 (empty)

100  

101  

102cdef object tensor_map_descriptor_type = _TensorMapDescriptor_py (empty)

103  

104  

105# limitation due to cython/cython#534 

106ctypedef void* voidptr 

107  

108  

109# Cython can't infer the overload without at least one input argument with fused type 

110cdef inline int prepare_arg( (empty)

111 vector.vector[void*]& data, 

112 vector.vector[void*]& data_addresses, 

113 arg, # important: keep it a Python object and don't cast 

114 const size_t idx, 

115 const supported_type* __unused=NULL) except -1: 

116 cdef void* ptr = PyMem_Malloc(sizeof(supported_type)) 86 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQa+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

117 # note: this should also work once ctypes has complex support: 

118 # python/cpython#121248 

119 if supported_type is cpp_single_complex: 

120 (<supported_type*>ptr)[0] = cpp_complex.complex[float](arg.real, arg.imag) 1 ctx1S

121 elif supported_type is cpp_double_complex: 

122 (<supported_type*>ptr)[0] = cpp_complex.complex[double](arg.real, arg.imag) 3 ctx1aR5

123 elif supported_type is __half_raw: 

124 (<supported_type*>ptr).x = <int16_t>(arg.view(numpy_int16)) 1 ctx1U

125 else: 

126 (<supported_type*>ptr)[0] = <supported_type>(arg) 87 ctx1O,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQa+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

127 data_addresses[idx] = ptr # take the address to the scalar 86 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQa+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

128 data[idx] = ptr # for later dealloc 86 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQa+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

129 return 0 86 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQa+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

130  

131  

132cdef inline int prepare_tensor_map_arg( (empty)

133 vector.vector[void*]& data, 

134 vector.vector[void*]& data_addresses, 

135 TensorMapDescriptor arg, 

136 const size_t idx) except -1: 

137 # cuLaunchKernel copies argument bytes during launch, so a TensorMap 

138 # descriptor can point directly at its internal CUtensorMap storage. 

139 data_addresses[idx] = arg._get_data_ptr() 

140 return 0 

141  

142  

143cdef inline int prepare_ctypes_arg( (empty)

144 vector.vector[void*]& data, 

145 vector.vector[void*]& data_addresses, 

146 arg, 

147 const size_t idx) except -1: 

148 cdef object arg_type = type(arg) 40 ctx1stbcklmnuvdeopqrwxyzfghijAaBNMDEKJLIGFHC

149 if arg_type is ctypes_bool: 41 ctx1OstbcklmnuvdeopqrwxyzfghijAaBNMDEKJLIGFHC

150 return prepare_arg[cpp_bool](data, data_addresses, arg.value, idx) 5 ctx1klopN

151 elif arg_type is ctypes_int8: 39 ctx1stbcklmnuvdeopqrwxyzfghijAaBMDEKJLIGFHC

152 return prepare_arg[int8_t](data, data_addresses, arg.value, idx) 1 ctx1M

153 elif arg_type is ctypes_int16: 38 ctx1stbcklmnuvdeopqrwxyzfghijAaBDEKJLIGFHC

154 return prepare_arg[int16_t](data, data_addresses, arg.value, idx) 2 ctx1OL

155 elif arg_type is ctypes_int32: 37 ctx1stbcklmnuvdeopqrwxyzfghijAaBDEKJIGFHC

156 return prepare_arg[int32_t](data, data_addresses, arg.value, idx) 1 ctx1K

157 elif arg_type is ctypes_int64: 36 ctx1stbcklmnuvdeopqrwxyzfghijAaBDEJIGFHC

158 return prepare_arg[int64_t](data, data_addresses, arg.value, idx) 1 ctx1J

159 elif arg_type is ctypes_uint8: 36 ctx1OstbcklmnuvdeopqrwxyzfghijAaBDEIGFHC

160 return prepare_arg[uint8_t](data, data_addresses, arg.value, idx) 1 ctx1I

161 elif arg_type is ctypes_uint16: 34 ctx1stbcklmnuvdeopqrwxyzfghijAaBDEGFHC

162 return prepare_arg[uint16_t](data, data_addresses, arg.value, idx) 1 ctx1H

163 elif arg_type is ctypes_uint32: 33 ctx1stbcklmnuvdeopqrwxyzfghijAaBDEGFC

164 return prepare_arg[uint32_t](data, data_addresses, arg.value, idx) 2 ctx1OG

165 elif arg_type is ctypes_uint64: 32 ctx1stbcklmnuvdeopqrwxyzfghijAaBDEFC

166 return prepare_arg[uint64_t](data, data_addresses, arg.value, idx) 1 ctx1F

167 elif arg_type is ctypes_float: 31 ctx1stbcklmnuvdeopqrwxyzfghijAaBDEC

168 return prepare_arg[float](data, data_addresses, arg.value, idx) 1 ctx1E

169 elif arg_type is ctypes_double: 30 ctx1stbcklmnuvdeopqrwxyzfghijAaBDC

170 return prepare_arg[double](data, data_addresses, arg.value, idx) 1 ctx1D

171 else: 

172 # If no exact types are found, fallback to slower `isinstance` check 

173 if isinstance(arg, ctypes_bool): 29 ctx1stbcklmnuvdeopqrwxyzfghijAaBC

174 return prepare_arg[cpp_bool](data, data_addresses, arg.value, idx) 1 ctx1A

175 elif isinstance(arg, ctypes_int8): 29 ctx1stbcklmnuvdeopqrwxyzfghijAaBC

176 return prepare_arg[int8_t](data, data_addresses, arg.value, idx) 

177 elif isinstance(arg, ctypes_int16): 29 ctx1stbcklmnuvdeopqrwxyzfghijAaBC

178 return prepare_arg[int16_t](data, data_addresses, arg.value, idx) 

179 elif isinstance(arg, ctypes_int32): 30 ctx1OstbcklmnuvdeopqrwxyzfghijAaBC

180 return prepare_arg[int32_t](data, data_addresses, arg.value, idx) 2 ctx1AC

181 elif isinstance(arg, ctypes_int64): 28 ctx1stbcklmnuvdeopqrwxyzfghijAaB

182 return prepare_arg[int64_t](data, data_addresses, arg.value, idx) 

183 elif isinstance(arg, ctypes_uint8): 28 ctx1stbcklmnuvdeopqrwxyzfghijAaB

184 return prepare_arg[uint8_t](data, data_addresses, arg.value, idx) 

185 elif isinstance(arg, ctypes_uint16): 29 ctx1OstbcklmnuvdeopqrwxyzfghijAaB

186 return prepare_arg[uint16_t](data, data_addresses, arg.value, idx) 

187 elif isinstance(arg, ctypes_uint32): 28 ctx1stbcklmnuvdeopqrwxyzfghijAaB

188 return prepare_arg[uint32_t](data, data_addresses, arg.value, idx) 

189 elif isinstance(arg, ctypes_uint64): 28 ctx1stbcklmnuvdeopqrwxyzfghijAaB

190 return prepare_arg[uint64_t](data, data_addresses, arg.value, idx) 

191 elif isinstance(arg, ctypes_float): 28 ctx1stbcklmnuvdeopqrwxyzfghijAaB

192 return prepare_arg[float](data, data_addresses, arg.value, idx) 1 ctx1A

193 elif isinstance(arg, ctypes_double): 27 ctx1stbcklmnuvdeopqrwxyzfghijaB

194 return prepare_arg[double](data, data_addresses, arg.value, idx) (empty)

195 else: 

196 return 1 27 ctx1stbcklmnuvdeopqrwxyzfghijaB

197  

198  

199cdef inline int prepare_numpy_arg( (empty)

200 vector.vector[void*]& data, 

201 vector.vector[void*]& data_addresses, 

202 arg, 

203 const size_t idx) except -1: 

204 cdef object arg_type = type(arg) 56 ctx1stbcklmnuvdeopqrwxyzfghijAQaB4NMDEKJLIGFHSRUVT3210ZYXWCP

205 if arg_type is numpy_bool: 56 ctx1stbcklmnuvdeopqrwxyzfghijAQaB4NMDEKJLIGFHSRUVT3210ZYXWCP

206 return prepare_arg[cpp_bool](data, data_addresses, arg, idx) 5 ctx1mnqr4

207 elif arg_type is numpy_int8: 55 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRUVT3210ZYXWCP

208 return prepare_arg[int8_t](data, data_addresses, arg, idx) 1 ctx13

209 elif arg_type is numpy_int16: 54 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRUVT210ZYXWCP

210 return prepare_arg[int16_t](data, data_addresses, arg, idx) 1 ctx12

211 elif arg_type is numpy_int32: 53 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRUVT10ZYXWCP

212 return prepare_arg[int32_t](data, data_addresses, arg, idx) 1 ctx11

213 elif arg_type is numpy_int64: 53 ctx1OstbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRUVT0ZYXWCP

214 return prepare_arg[int64_t](data, data_addresses, arg, idx) 1 ctx10

215 elif arg_type is numpy_uint8: 51 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRUVTZYXWCP

216 return prepare_arg[uint8_t](data, data_addresses, arg, idx) 1 ctx1Z

217 elif arg_type is numpy_uint16: 50 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRUVTYXWCP

218 return prepare_arg[uint16_t](data, data_addresses, arg, idx) 1 ctx1Y

219 elif arg_type is numpy_uint32: 50 ctx1OstbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRUVTXWCP

220 return prepare_arg[uint32_t](data, data_addresses, arg, idx) 1 ctx1X

221 elif arg_type is numpy_uint64: 48 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRUVTWCP

222 return prepare_arg[uint64_t](data, data_addresses, arg, idx) 1 ctx1W

223 elif arg_type is numpy_float16: 48 ctx1OstbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRUVTCP

224 return prepare_arg[__half_raw](data, data_addresses, arg, idx) 1 ctx1U

225 elif arg_type is numpy_float32: 46 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRVTCP

226 return prepare_arg[float](data, data_addresses, arg, idx) 1 ctx1V

227 elif arg_type is numpy_float64: 45 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRTCP

228 return prepare_arg[double](data, data_addresses, arg, idx) 1 ctx1T

229 elif arg_type is numpy_complex64: 44 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHSRCP

230 return prepare_arg[cpp_single_complex](data, data_addresses, arg, idx) 1 ctx1S

231 elif arg_type is numpy_complex128: 43 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHRCP

232 return prepare_arg[cpp_double_complex](data, data_addresses, arg, idx) 2 ctx1OR

233 else: 

234 # If no exact types are found, fallback to slower `isinstance` check 

235 if isinstance(arg, numpy_bool): 42 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

236 return prepare_arg[cpp_bool](data, data_addresses, arg, idx) 

237 elif isinstance(arg, numpy_int8): 42 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

238 return prepare_arg[int8_t](data, data_addresses, arg, idx) 

239 elif isinstance(arg, numpy_int16): 42 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

240 return prepare_arg[int16_t](data, data_addresses, arg, idx) 

241 elif isinstance(arg, numpy_int32): 43 ctx1OstbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

242 return prepare_arg[int32_t](data, data_addresses, arg, idx) 1 ctx1Q

243 elif isinstance(arg, numpy_int64): 42 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

244 return prepare_arg[int64_t](data, data_addresses, arg, idx) 

245 elif isinstance(arg, numpy_uint8): 42 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

246 return prepare_arg[uint8_t](data, data_addresses, arg, idx) (empty)

247 elif isinstance(arg, numpy_uint16): 42 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

248 return prepare_arg[uint16_t](data, data_addresses, arg, idx) 

249 elif isinstance(arg, numpy_uint32): 42 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

250 return prepare_arg[uint32_t](data, data_addresses, arg, idx) 

251 elif isinstance(arg, numpy_uint64): 43 ctx1OstbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

252 return prepare_arg[uint64_t](data, data_addresses, arg, idx) 

253 elif isinstance(arg, numpy_float16): 42 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

254 return prepare_arg[__half_raw](data, data_addresses, arg, idx) 

255 elif isinstance(arg, numpy_float32): 42 ctx1stbcklmnuvdeopqrwxyzfghijAQaBNMDEKJLIGFHCP

256 return prepare_arg[float](data, data_addresses, arg, idx) 2 ctx1QP

257 elif isinstance(arg, numpy_float64): 40 ctx1stbcklmnuvdeopqrwxyzfghijAaBNMDEKJLIGFHC

258 return prepare_arg[double](data, data_addresses, arg, idx) (empty)

259 elif isinstance(arg, numpy_complex64): 40 ctx1stbcklmnuvdeopqrwxyzfghijAaBNMDEKJLIGFHC

260 return prepare_arg[cpp_single_complex](data, data_addresses, arg, idx) 

261 elif isinstance(arg, numpy_complex128): 40 ctx1stbcklmnuvdeopqrwxyzfghijAaBNMDEKJLIGFHC

262 return prepare_arg[cpp_double_complex](data, data_addresses, arg, idx) 

263 else: 

264 return 1 40 ctx1stbcklmnuvdeopqrwxyzfghijAaBNMDEKJLIGFHC

265  

266  

267cdef class ParamHolder: 

268  

269 def __init__(self, kernel_args): 

270 if len(kernel_args) == 0: 246 ctx2, ~ ab- bbcbs t b c k l m n u v d e o p q r w x y z f g h i dbebfbgbhbibjbkblbmbnbobpbqbrbsbtbubvbwbxbybzbAbBbCbDbEbFbGbHb. / : ; = ? 6 7 8 9 ! # $ % ' ( ) @ [ ] IbJbKbLbMbj ^ _ ` { Nb| ObA Q a B Pb+ 4 N M D E K J L I G F H S R 5 U V T * 3 2 1 0 Z Y X W C P } QbRbSbTbUbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?b@b[b]b^b_b`b{b|b}b~bacbcccdcecfcgchcicjckclcmcncocpcqcrcsctcucvcwcxcyczcAcBcCcDcEcFcGcHcIcJcKcLcMcNcOcPcQcRcScTcUcVcWcXcYcZc0c1c2c3c4c5c6c7c8c9c!c#c$c%c'c(c)c

271 self.ptr = 0 164 ctx2~ abbbcbf g h i dbebfbgbhbibjbkblbmbnbobpbqbrbsbtbubvbwbxbybzbAbBbCbDbEbFbGbHbIbJbKbLbMbj NbObPbQbRbSbTbUbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?b@b[b]b^b_b`b{b|b}b~bacbcccdcecfcgchcicjckclcmcncocpcqcrcsctcucvcwcxcyczcAcBcCcDcEcFcGcHcIcJcKcLcMcNcOcPcQcRcScTcUcVcWcXcYcZc0c1c2c3c4c5c6c7c8c9c!c#c$c%c'c(c)c

272 return 164 ctx2~ abbbcbf g h i dbebfbgbhbibjbkblbmbnbobpbqbrbsbtbubvbwbxbybzbAbBbCbDbEbFbGbHbIbJbKbLbMbj NbObPbQbRbSbTbUbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?b@b[b]b^b_b`b{b|b}b~bacbcccdcecfcgchcicjckclcmcncocpcqcrcsctcucvcwcxcyczcAcBcCcDcEcFcGcHcIcJcKcLcMcNcOcPcQcRcScTcUcVcWcXcYcZc0c1c2c3c4c5c6c7c8c9c!c#c$c%c'c(c)c

273  

274 cdef size_t n_args = len(kernel_args) 87 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQaB+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

275 cdef size_t i 

276 cdef int not_prepared 

277 cdef object arg_type 

278 self.data = vector.vector[voidptr](n_args, nullptr) 87 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQaB+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

279 self.data_addresses = vector.vector[voidptr](n_args) 87 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQaB+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

280 for i, arg in enumerate(kernel_args): 87 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQaB+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

281 arg_type = type(arg) 87 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQaB+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

282 if arg_type is Buffer: 87 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQaB+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

283 # we need the address of where the actual buffer address is stored 

284 if type(arg.handle) is int: 12 ctx16789!#$%'()}

285 # see note below on handling int arguments 

286 prepare_arg[intptr_t](self.data, self.data_addresses, arg.handle, i) 12 ctx16789!#$%'()}

287 continue 12 ctx16789!#$%'()}

288 else: 

289 # it's a CUdeviceptr: 

290 self.data_addresses[i] = <void*><intptr_t>(arg.handle.getPtr()) 

291 continue 

292 elif arg_type is bool: 87 ctx1O,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQaB+4NMDEKJLIGFHSR5UVT*3210ZYXWCP

293 prepare_arg[cpp_bool](self.data, self.data_addresses, arg, i) 6 ctx1bcdej+

294 continue 6 ctx1bcdej+

295 elif arg_type is int: 86 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQaB+4NMDEKJLIGFHSR5UVT*3210ZYXWCP

296 # Here's the dilemma: We want to have a fast path to pass in Python 

297 # integers as pointer addresses, but one could also (mistakenly) pass 

298 # it with the intention of passing a scalar integer. It's a mistake 

299 # bacause a Python int is ambiguous (arbitrary width). Our judgement 

300 # call here is to treat it as a pointer address, without any warning! 

301 prepare_arg[intptr_t](self.data, self.data_addresses, arg, i) 81 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]^_`{|+4NMDEKJLIGFHSR5UVT*3210ZYXWCP

302 continue 81 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]^_`{|+4NMDEKJLIGFHSR5UVT*3210ZYXWCP

303 elif arg_type is float: 58 ctx1stbcklmnuvdeopqrwxyzfghijAQaB4NMDEKJLIGFHSR5UVT*3210ZYXWCP

304 prepare_arg[double](self.data, self.data_addresses, arg, i) 1 ctx1*

305 continue 1 ctx1*

306 elif arg_type is complex: 57 ctx1stbcklmnuvdeopqrwxyzfghijAQaB4NMDEKJLIGFHSR5UVT3210ZYXWCP

307 prepare_arg[cpp_double_complex](self.data, self.data_addresses, arg, i) 1 ctx15

308 continue 1 ctx15

309 elif arg_type is tensor_map_descriptor_type: 56 ctx1stbcklmnuvdeopqrwxyzfghijAQaB4NMDEKJLIGFHSRUVT3210ZYXWCP

310 prepare_tensor_map_arg(self.data, self.data_addresses, <TensorMapDescriptor>arg, i) 

311 continue (empty)

312  

313 not_prepared = prepare_numpy_arg(self.data, self.data_addresses, arg, i) 56 ctx1stbcklmnuvdeopqrwxyzfghijAQaB4NMDEKJLIGFHSRUVT3210ZYXWCP

314 if not_prepared: 56 ctx1stbcklmnuvdeopqrwxyzfghijAQaB4NMDEKJLIGFHSRUVT3210ZYXWCP

315 not_prepared = prepare_ctypes_arg(self.data, self.data_addresses, arg, i) 40 ctx1stbcklmnuvdeopqrwxyzfghijAaBNMDEKJLIGFHC

316 if not_prepared: 56 ctx1stbcklmnuvdeopqrwxyzfghijAQaB4NMDEKJLIGFHSRUVT3210ZYXWCP

317 # TODO: revisit this treatment if we decide to cythonize cuda.core 

318 if arg_type is driver.CUgraphConditionalHandle: 27 ctx1stbcklmnuvdeopqrwxyzfghijaB

319 prepare_arg[cydriver.CUgraphConditionalHandle](self.data, self.data_addresses, <intptr_t>int(arg), i) 25 ctx1stbcklmnuvdeopqrwxyzfghij

320 continue 25 ctx1stbcklmnuvdeopqrwxyzfghij

321 # If no exact types are found, fallback to slower `isinstance` check 

322 elif isinstance(arg, Buffer): 2 ctx1aB

323 if isinstance(arg.handle, int): 

324 prepare_arg[intptr_t](self.data, self.data_addresses, arg.handle, i) 

325 continue 

326 else: 

327 self.data_addresses[i] = <void*><intptr_t>(arg.handle.getPtr()) 

328 continue 

329 elif isinstance(arg, bool): 2 ctx1aB

330 prepare_arg[cpp_bool](self.data, self.data_addresses, arg, i) 

331 continue 

332 elif isinstance(arg, int): 2 ctx1aB

333 prepare_arg[intptr_t](self.data, self.data_addresses, arg, i) 1 ctx1a

334 continue 1 ctx1a

335 elif isinstance(arg, float): 2 ctx1aB

336 prepare_arg[double](self.data, self.data_addresses, arg, i) 1 ctx1a

337 continue 1 ctx1a

338 elif isinstance(arg, complex): 3 ctx1OaB

339 prepare_arg[cpp_double_complex](self.data, self.data_addresses, arg, i) 1 ctx1a

340 continue 1 ctx1a

341 elif isinstance(arg, driver.CUgraphConditionalHandle): 1 ctx1B

342 prepare_arg[cydriver.CUgraphConditionalHandle](self.data, self.data_addresses, arg, i) 

343 continue 

344 # TODO: support ctypes/numpy struct 

345 raise TypeError("the argument is of unsupported type: " + str(type(arg))) 1 ctx1B

346  

347 self.kernel_args = kernel_args 86 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQa+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

348 self.ptr = <intptr_t>self.data_addresses.data() 86 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQa+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

349  

350 def __dealloc__(self): 

351 for data in self.data: 247 ctx2O , ~ ab- bbcbs t b c k l m n u v d e o p q r w x y z f g h i dbebfbgbhbibjbkblbmbnbobpbqbrbsbtbubvbwbxbybzbAbBbCbDbEbFbGbHb. / : ; = ? 6 7 8 9 ! # $ % ' ( ) @ [ ] IbJbKbLbMbj ^ _ ` { Nb| ObA Q a B Pb+ 4 N M D E K J L I G F H S R 5 U V T * 3 2 1 0 Z Y X W C P } QbRbSbTbUbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?b@b[b]b^b_b`b{b|b}b~bacbcccdcecfcgchcicjckclcmcncocpcqcrcsctcucvcwcxcyczcAcBcCcDcEcFcGcHcIcJcKcLcMcNcOcPcQcRcScTcUcVcWcXcYcZc0c1c2c3c4c5c6c7c8c9c!c#c$c%c'c(c)c

352 if data: 87 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQaB+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}

353 PyMem_Free(data) 86 ctx1,-stbcklmnuvdeopqrwxyzfghi./:;=?6789!#$%'()@[]j^_`{|AQa+4NMDEKJLIGFHSR5UVT*3210ZYXWCP}