contrib/llvm-project/clang/lib/Headers/opencl-c-base.h

   1 //===----- opencl-c-base.h - OpenCL C language base definitions -----------===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8
   9 #ifndef _OPENCL_BASE_H_
  10 #define _OPENCL_BASE_H_
  11
  12 // Define extension macros
  13
  14 #if (defined(__OPENCL_CPP_VERSION__) || __OPENCL_C_VERSION__ >= 200)
  15 // For SPIR and SPIR-V all extensions are supported.
  16 #if defined(__SPIR__) || defined(__SPIRV__)
  17 #define cl_khr_subgroup_extended_types 1
  18 #define cl_khr_subgroup_non_uniform_vote 1
  19 #define cl_khr_subgroup_ballot 1
  20 #define cl_khr_subgroup_non_uniform_arithmetic 1
  21 #define cl_khr_subgroup_shuffle 1
  22 #define cl_khr_subgroup_shuffle_relative 1
  23 #define cl_khr_subgroup_clustered_reduce 1
  24 #define cl_khr_extended_bit_ops 1
  25 #define cl_khr_integer_dot_product 1
  26 #define __opencl_c_integer_dot_product_input_4x8bit 1
  27 #define __opencl_c_integer_dot_product_input_4x8bit_packed 1
  28 #define cl_ext_float_atomics 1
  29 #ifdef cl_khr_fp16
  30 #define __opencl_c_ext_fp16_global_atomic_load_store 1
  31 #define __opencl_c_ext_fp16_local_atomic_load_store 1
  32 #define __opencl_c_ext_fp16_global_atomic_add 1
  33 #define __opencl_c_ext_fp16_local_atomic_add 1
  34 #define __opencl_c_ext_fp16_global_atomic_min_max 1
  35 #define __opencl_c_ext_fp16_local_atomic_min_max 1
  36 #endif
  37 #ifdef cl_khr_fp64
  38 #define __opencl_c_ext_fp64_global_atomic_add 1
  39 #define __opencl_c_ext_fp64_local_atomic_add 1
  40 #define __opencl_c_ext_fp64_global_atomic_min_max 1
  41 #define __opencl_c_ext_fp64_local_atomic_min_max 1
  42 #endif
  43 #define __opencl_c_ext_fp32_global_atomic_add 1
  44 #define __opencl_c_ext_fp32_local_atomic_add 1
  45 #define __opencl_c_ext_fp32_global_atomic_min_max 1
  46 #define __opencl_c_ext_fp32_local_atomic_min_max 1
  47
  48 #endif // defined(__SPIR__) || defined(__SPIRV__)
  49 #endif // (defined(__OPENCL_CPP_VERSION__) || __OPENCL_C_VERSION__ >= 200)
  50
  51 // Define feature macros for OpenCL C 2.0
  52 #if (__OPENCL_CPP_VERSION__ == 100 || __OPENCL_C_VERSION__ == 200)
  53 #define __opencl_c_pipes 1
  54 #define __opencl_c_generic_address_space 1
  55 #define __opencl_c_work_group_collective_functions 1
  56 #define __opencl_c_atomic_order_acq_rel 1
  57 #define __opencl_c_atomic_order_seq_cst 1
  58 #define __opencl_c_atomic_scope_device 1
  59 #define __opencl_c_atomic_scope_all_devices 1
  60 #define __opencl_c_device_enqueue 1
  61 #define __opencl_c_read_write_images 1
  62 #define __opencl_c_program_scope_global_variables 1
  63 #define __opencl_c_images 1
  64 #endif
  65
  66 // Define header-only feature macros for OpenCL C 3.0.
  67 #if (__OPENCL_CPP_VERSION__ == 202100 || __OPENCL_C_VERSION__ == 300)
  68 // For the SPIR and SPIR-V target all features are supported.
  69 #if defined(__SPIR__) || defined(__SPIRV__)
  70 #define __opencl_c_atomic_order_seq_cst 1
  71 #define __opencl_c_atomic_scope_device 1
  72 #define __opencl_c_atomic_scope_all_devices 1
  73 #define __opencl_c_read_write_images 1
  74 #endif // defined(__SPIR__)
  75 #endif // (__OPENCL_CPP_VERSION__ == 202100 || __OPENCL_C_VERSION__ == 300)
  76
  77 #if !defined(__opencl_c_generic_address_space)
  78 // Internal feature macro to provide named (global, local, private) address
  79 // space overloads for builtin functions that take a pointer argument.
  80 #define __opencl_c_named_address_space_builtins 1
  81 #endif // !defined(__opencl_c_generic_address_space)
  82
  83 // built-in scalar data types:
  84
  85 /**
  86  * An unsigned 8-bit integer.
  87  */
  88 typedef unsigned char uchar;
  89
  90 /**
  91  * An unsigned 16-bit integer.
  92  */
  93 typedef unsigned short ushort;
  94
  95 /**
  96  * An unsigned 32-bit integer.
  97  */
  98 typedef unsigned int uint;
  99
 100 /**
 101  * An unsigned 64-bit integer.
 102  */
 103 typedef unsigned long ulong;
 104
 105 /**
 106  * The unsigned integer type of the result of the sizeof operator. This
 107  * is a 32-bit unsigned integer if CL_DEVICE_ADDRESS_BITS
 108  * defined in table 4.3 is 32-bits and is a 64-bit unsigned integer if
 109  * CL_DEVICE_ADDRESS_BITS is 64-bits.
 110  */
 111 typedef __SIZE_TYPE__ size_t;
 112
 113 /**
 114  * A signed integer type that is the result of subtracting two pointers.
 115  * This is a 32-bit signed integer if CL_DEVICE_ADDRESS_BITS
 116  * defined in table 4.3 is 32-bits and is a 64-bit signed integer if
 117  * CL_DEVICE_ADDRESS_BITS is 64-bits.
 118  */
 119 typedef __PTRDIFF_TYPE__ ptrdiff_t;
 120
 121 /**
 122  * A signed integer type with the property that any valid pointer to
 123  * void can be converted to this type, then converted back to pointer
 124  * to void, and the result will compare equal to the original pointer.
 125  */
 126 typedef __INTPTR_TYPE__ intptr_t;
 127
 128 /**
 129  * An unsigned integer type with the property that any valid pointer to
 130  * void can be converted to this type, then converted back to pointer
 131  * to void, and the result will compare equal to the original pointer.
 132  */
 133 typedef __UINTPTR_TYPE__ uintptr_t;
 134
 135 // built-in vector data types:
 136 typedef char char2 __attribute__((ext_vector_type(2)));
 137 typedef char char3 __attribute__((ext_vector_type(3)));
 138 typedef char char4 __attribute__((ext_vector_type(4)));
 139 typedef char char8 __attribute__((ext_vector_type(8)));
 140 typedef char char16 __attribute__((ext_vector_type(16)));
 141 typedef uchar uchar2 __attribute__((ext_vector_type(2)));
 142 typedef uchar uchar3 __attribute__((ext_vector_type(3)));
 143 typedef uchar uchar4 __attribute__((ext_vector_type(4)));
 144 typedef uchar uchar8 __attribute__((ext_vector_type(8)));
 145 typedef uchar uchar16 __attribute__((ext_vector_type(16)));
 146 typedef short short2 __attribute__((ext_vector_type(2)));
 147 typedef short short3 __attribute__((ext_vector_type(3)));
 148 typedef short short4 __attribute__((ext_vector_type(4)));
 149 typedef short short8 __attribute__((ext_vector_type(8)));
 150 typedef short short16 __attribute__((ext_vector_type(16)));
 151 typedef ushort ushort2 __attribute__((ext_vector_type(2)));
 152 typedef ushort ushort3 __attribute__((ext_vector_type(3)));
 153 typedef ushort ushort4 __attribute__((ext_vector_type(4)));
 154 typedef ushort ushort8 __attribute__((ext_vector_type(8)));
 155 typedef ushort ushort16 __attribute__((ext_vector_type(16)));
 156 typedef int int2 __attribute__((ext_vector_type(2)));
 157 typedef int int3 __attribute__((ext_vector_type(3)));
 158 typedef int int4 __attribute__((ext_vector_type(4)));
 159 typedef int int8 __attribute__((ext_vector_type(8)));
 160 typedef int int16 __attribute__((ext_vector_type(16)));
 161 typedef uint uint2 __attribute__((ext_vector_type(2)));
 162 typedef uint uint3 __attribute__((ext_vector_type(3)));
 163 typedef uint uint4 __attribute__((ext_vector_type(4)));
 164 typedef uint uint8 __attribute__((ext_vector_type(8)));
 165 typedef uint uint16 __attribute__((ext_vector_type(16)));
 166 typedef long long2 __attribute__((ext_vector_type(2)));
 167 typedef long long3 __attribute__((ext_vector_type(3)));
 168 typedef long long4 __attribute__((ext_vector_type(4)));
 169 typedef long long8 __attribute__((ext_vector_type(8)));
 170 typedef long long16 __attribute__((ext_vector_type(16)));
 171 typedef ulong ulong2 __attribute__((ext_vector_type(2)));
 172 typedef ulong ulong3 __attribute__((ext_vector_type(3)));
 173 typedef ulong ulong4 __attribute__((ext_vector_type(4)));
 174 typedef ulong ulong8 __attribute__((ext_vector_type(8)));
 175 typedef ulong ulong16 __attribute__((ext_vector_type(16)));
 176 typedef float float2 __attribute__((ext_vector_type(2)));
 177 typedef float float3 __attribute__((ext_vector_type(3)));
 178 typedef float float4 __attribute__((ext_vector_type(4)));
 179 typedef float float8 __attribute__((ext_vector_type(8)));
 180 typedef float float16 __attribute__((ext_vector_type(16)));
 181 #ifdef cl_khr_fp16
 182 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 183 typedef half half2 __attribute__((ext_vector_type(2)));
 184 typedef half half3 __attribute__((ext_vector_type(3)));
 185 typedef half half4 __attribute__((ext_vector_type(4)));
 186 typedef half half8 __attribute__((ext_vector_type(8)));
 187 typedef half half16 __attribute__((ext_vector_type(16)));
 188 #endif
 189 #ifdef cl_khr_fp64
 190 #if __OPENCL_C_VERSION__ < CL_VERSION_1_2
 191 #pragma OPENCL EXTENSION cl_khr_fp64 : enable
 192 #endif
 193 typedef double double2 __attribute__((ext_vector_type(2)));
 194 typedef double double3 __attribute__((ext_vector_type(3)));
 195 typedef double double4 __attribute__((ext_vector_type(4)));
 196 typedef double double8 __attribute__((ext_vector_type(8)));
 197 typedef double double16 __attribute__((ext_vector_type(16)));
 198 #endif
 199
 200 #if defined(__OPENCL_CPP_VERSION__)
 201 #define NULL nullptr
 202 #elif defined(__OPENCL_C_VERSION__)
 203 #define NULL ((void*)0)
 204 #endif
 205
 206 /**
 207  * Value of maximum non-infinite single-precision floating-point
 208  * number.
 209  */
 210 #define MAXFLOAT 0x1.fffffep127f
 211
 212 /**
 213  * A positive float constant expression. HUGE_VALF evaluates
 214  * to +infinity. Used as an error value returned by the built-in
 215  * math functions.
 216  */
 217 #define HUGE_VALF (__builtin_huge_valf())
 218
 219 /**
 220  * A positive double constant expression. HUGE_VAL evaluates
 221  * to +infinity. Used as an error value returned by the built-in
 222  * math functions.
 223  */
 224 #define HUGE_VAL (__builtin_huge_val())
 225
 226 /**
 227  * A constant expression of type float representing positive or
 228  * unsigned infinity.
 229  */
 230 #define INFINITY (__builtin_inff())
 231
 232 /**
 233  * A constant expression of type float representing a quiet NaN.
 234  */
 235 #define NAN as_float(INT_MAX)
 236
 237 #define FP_ILOGB0    INT_MIN
 238 #define FP_ILOGBNAN  INT_MAX
 239
 240 #define FLT_DIG 6
 241 #define FLT_MANT_DIG 24
 242 #define FLT_MAX_10_EXP +38
 243 #define FLT_MAX_EXP +128
 244 #define FLT_MIN_10_EXP -37
 245 #define FLT_MIN_EXP -125
 246 #define FLT_RADIX 2
 247 #define FLT_MAX 0x1.fffffep127f
 248 #define FLT_MIN 0x1.0p-126f
 249 #define FLT_EPSILON 0x1.0p-23f
 250
 251 #define M_E_F         2.71828182845904523536028747135266250f
 252 #define M_LOG2E_F     1.44269504088896340735992468100189214f
 253 #define M_LOG10E_F    0.434294481903251827651128918916605082f
 254 #define M_LN2_F       0.693147180559945309417232121458176568f
 255 #define M_LN10_F      2.30258509299404568401799145468436421f
 256 #define M_PI_F        3.14159265358979323846264338327950288f
 257 #define M_PI_2_F      1.57079632679489661923132169163975144f
 258 #define M_PI_4_F      0.785398163397448309615660845819875721f
 259 #define M_1_PI_F      0.318309886183790671537767526745028724f
 260 #define M_2_PI_F      0.636619772367581343075535053490057448f
 261 #define M_2_SQRTPI_F  1.12837916709551257389615890312154517f
 262 #define M_SQRT2_F     1.41421356237309504880168872420969808f
 263 #define M_SQRT1_2_F   0.707106781186547524400844362104849039f
 264
 265 #define DBL_DIG 15
 266 #define DBL_MANT_DIG 53
 267 #define DBL_MAX_10_EXP +308
 268 #define DBL_MAX_EXP +1024
 269 #define DBL_MIN_10_EXP -307
 270 #define DBL_MIN_EXP -1021
 271 #define DBL_RADIX 2
 272 #define DBL_MAX 0x1.fffffffffffffp1023
 273 #define DBL_MIN 0x1.0p-1022
 274 #define DBL_EPSILON 0x1.0p-52
 275
 276 #define M_E           0x1.5bf0a8b145769p+1
 277 #define M_LOG2E       0x1.71547652b82fep+0
 278 #define M_LOG10E      0x1.bcb7b1526e50ep-2
 279 #define M_LN2         0x1.62e42fefa39efp-1
 280 #define M_LN10        0x1.26bb1bbb55516p+1
 281 #define M_PI          0x1.921fb54442d18p+1
 282 #define M_PI_2        0x1.921fb54442d18p+0
 283 #define M_PI_4        0x1.921fb54442d18p-1
 284 #define M_1_PI        0x1.45f306dc9c883p-2
 285 #define M_2_PI        0x1.45f306dc9c883p-1
 286 #define M_2_SQRTPI    0x1.20dd750429b6dp+0
 287 #define M_SQRT2       0x1.6a09e667f3bcdp+0
 288 #define M_SQRT1_2     0x1.6a09e667f3bcdp-1
 289
 290 #ifdef cl_khr_fp16
 291
 292 #define HALF_DIG 3
 293 #define HALF_MANT_DIG 11
 294 #define HALF_MAX_10_EXP +4
 295 #define HALF_MAX_EXP +16
 296 #define HALF_MIN_10_EXP -4
 297 #define HALF_MIN_EXP -13
 298 #define HALF_RADIX 2
 299 #define HALF_MAX ((0x1.ffcp15h))
 300 #define HALF_MIN ((0x1.0p-14h))
 301 #define HALF_EPSILON ((0x1.0p-10h))
 302
 303 #define M_E_H         2.71828182845904523536028747135266250h
 304 #define M_LOG2E_H     1.44269504088896340735992468100189214h
 305 #define M_LOG10E_H    0.434294481903251827651128918916605082h
 306 #define M_LN2_H       0.693147180559945309417232121458176568h
 307 #define M_LN10_H      2.30258509299404568401799145468436421h
 308 #define M_PI_H        3.14159265358979323846264338327950288h
 309 #define M_PI_2_H      1.57079632679489661923132169163975144h
 310 #define M_PI_4_H      0.785398163397448309615660845819875721h
 311 #define M_1_PI_H      0.318309886183790671537767526745028724h
 312 #define M_2_PI_H      0.636619772367581343075535053490057448h
 313 #define M_2_SQRTPI_H  1.12837916709551257389615890312154517h
 314 #define M_SQRT2_H     1.41421356237309504880168872420969808h
 315 #define M_SQRT1_2_H   0.707106781186547524400844362104849039h
 316
 317 #endif //cl_khr_fp16
 318
 319 #define CHAR_BIT  8
 320 #define SCHAR_MAX 127
 321 #define SCHAR_MIN (-128)
 322 #define UCHAR_MAX 255
 323 #define CHAR_MAX  SCHAR_MAX
 324 #define CHAR_MIN  SCHAR_MIN
 325 #define USHRT_MAX 65535
 326 #define SHRT_MAX  32767
 327 #define SHRT_MIN  (-32768)
 328 #define UINT_MAX  0xffffffff
 329 #define INT_MAX   2147483647
 330 #define INT_MIN   (-2147483647-1)
 331 #define ULONG_MAX 0xffffffffffffffffUL
 332 #define LONG_MAX  0x7fffffffffffffffL
 333 #define LONG_MIN  (-0x7fffffffffffffffL-1)
 334
 335 // OpenCL v1.1 s6.11.8, v1.2 s6.12.8, v2.0 s6.13.8 - Synchronization Functions
 336
 337 // Flag type and values for barrier, mem_fence, read_mem_fence, write_mem_fence
 338 typedef uint cl_mem_fence_flags;
 339
 340 /**
 341  * Queue a memory fence to ensure correct
 342  * ordering of memory operations to local memory
 343  */
 344 #define CLK_LOCAL_MEM_FENCE    0x01
 345
 346 /**
 347  * Queue a memory fence to ensure correct
 348  * ordering of memory operations to global memory
 349  */
 350 #define CLK_GLOBAL_MEM_FENCE   0x02
 351
 352 #if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
 353
 354 typedef enum memory_scope {
 355   memory_scope_work_item = __OPENCL_MEMORY_SCOPE_WORK_ITEM,
 356   memory_scope_work_group = __OPENCL_MEMORY_SCOPE_WORK_GROUP,
 357   memory_scope_device = __OPENCL_MEMORY_SCOPE_DEVICE,
 358 #if defined(__opencl_c_atomic_scope_all_devices)
 359   memory_scope_all_svm_devices = __OPENCL_MEMORY_SCOPE_ALL_SVM_DEVICES,
 360 #if (__OPENCL_C_VERSION__ >= CL_VERSION_3_0 || __OPENCL_CPP_VERSION__ >= 202100)
 361   memory_scope_all_devices = memory_scope_all_svm_devices,
 362 #endif // (__OPENCL_C_VERSION__ >= CL_VERSION_3_0 || __OPENCL_CPP_VERSION__ >= 202100)
 363 #endif // defined(__opencl_c_atomic_scope_all_devices)
 364 /**
 365  * Subgroups have different requirements on forward progress, so just test
 366  * all the relevant macros.
 367  * CL 3.0 sub-groups "they are not guaranteed to make independent forward progress"
 368  * KHR subgroups "Subgroups within a workgroup are independent, make forward progress with respect to each other"
 369  */
 370 #if defined(cl_intel_subgroups) || defined(cl_khr_subgroups) || defined(__opencl_c_subgroups)
 371   memory_scope_sub_group = __OPENCL_MEMORY_SCOPE_SUB_GROUP
 372 #endif
 373 } memory_scope;
 374
 375 /**
 376  * Queue a memory fence to ensure correct ordering of memory
 377  * operations between work-items of a work-group to
 378  * image memory.
 379  */
 380 #define CLK_IMAGE_MEM_FENCE  0x04
 381
 382 #ifndef ATOMIC_VAR_INIT
 383 #define ATOMIC_VAR_INIT(x) (x)
 384 #endif //ATOMIC_VAR_INIT
 385 #define ATOMIC_FLAG_INIT 0
 386
 387 // enum values aligned with what clang uses in EmitAtomicExpr()
 388 typedef enum memory_order
 389 {
 390   memory_order_relaxed = __ATOMIC_RELAXED,
 391   memory_order_acquire = __ATOMIC_ACQUIRE,
 392   memory_order_release = __ATOMIC_RELEASE,
 393   memory_order_acq_rel = __ATOMIC_ACQ_REL,
 394 #if defined(__opencl_c_atomic_order_seq_cst)
 395   memory_order_seq_cst = __ATOMIC_SEQ_CST
 396 #endif
 397 } memory_order;
 398
 399 #endif // defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
 400
 401 // OpenCL v1.1 s6.11.3, v1.2 s6.12.14, v2.0 s6.13.14 - Image Read and Write Functions
 402
 403 // These values need to match the runtime equivalent
 404 //
 405 // Addressing Mode.
 406 //
 407 #define CLK_ADDRESS_NONE                0
 408 #define CLK_ADDRESS_CLAMP_TO_EDGE       2
 409 #define CLK_ADDRESS_CLAMP               4
 410 #define CLK_ADDRESS_REPEAT              6
 411 #define CLK_ADDRESS_MIRRORED_REPEAT     8
 412
 413 //
 414 // Coordination Normalization
 415 //
 416 #define CLK_NORMALIZED_COORDS_FALSE     0
 417 #define CLK_NORMALIZED_COORDS_TRUE      1
 418
 419 //
 420 // Filtering Mode.
 421 //
 422 #define CLK_FILTER_NEAREST              0x10
 423 #define CLK_FILTER_LINEAR               0x20
 424
 425 #ifdef cl_khr_gl_msaa_sharing
 426 #pragma OPENCL EXTENSION cl_khr_gl_msaa_sharing : enable
 427 #endif //cl_khr_gl_msaa_sharing
 428
 429 //
 430 // Channel Datatype.
 431 //
 432 #define CLK_SNORM_INT8        0x10D0
 433 #define CLK_SNORM_INT16       0x10D1
 434 #define CLK_UNORM_INT8        0x10D2
 435 #define CLK_UNORM_INT16       0x10D3
 436 #define CLK_UNORM_SHORT_565   0x10D4
 437 #define CLK_UNORM_SHORT_555   0x10D5
 438 #define CLK_UNORM_INT_101010  0x10D6
 439 #define CLK_SIGNED_INT8       0x10D7
 440 #define CLK_SIGNED_INT16      0x10D8
 441 #define CLK_SIGNED_INT32      0x10D9
 442 #define CLK_UNSIGNED_INT8     0x10DA
 443 #define CLK_UNSIGNED_INT16    0x10DB
 444 #define CLK_UNSIGNED_INT32    0x10DC
 445 #define CLK_HALF_FLOAT        0x10DD
 446 #define CLK_FLOAT             0x10DE
 447 #define CLK_UNORM_INT24       0x10DF
 448
 449 // Channel order, numbering must be aligned with cl_channel_order in cl.h
 450 //
 451 #define CLK_R         0x10B0
 452 #define CLK_A         0x10B1
 453 #define CLK_RG        0x10B2
 454 #define CLK_RA        0x10B3
 455 #define CLK_RGB       0x10B4
 456 #define CLK_RGBA      0x10B5
 457 #define CLK_BGRA      0x10B6
 458 #define CLK_ARGB      0x10B7
 459 #define CLK_INTENSITY 0x10B8
 460 #define CLK_LUMINANCE 0x10B9
 461 #define CLK_Rx                0x10BA
 462 #define CLK_RGx               0x10BB
 463 #define CLK_RGBx              0x10BC
 464 #define CLK_DEPTH             0x10BD
 465 #define CLK_DEPTH_STENCIL     0x10BE
 466 #if __OPENCL_C_VERSION__ >= CL_VERSION_2_0
 467 #define CLK_sRGB              0x10BF
 468 #define CLK_sRGBx             0x10C0
 469 #define CLK_sRGBA             0x10C1
 470 #define CLK_sBGRA             0x10C2
 471 #define CLK_ABGR              0x10C3
 472 #endif //__OPENCL_C_VERSION__ >= CL_VERSION_2_0
 473
 474 // OpenCL v2.0 s6.13.16 - Pipe Functions
 475 #if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
 476 #define CLK_NULL_RESERVE_ID (__builtin_astype(((void*)(__SIZE_MAX__)), reserve_id_t))
 477
 478 // OpenCL v2.0 s6.13.17 - Enqueue Kernels
 479 #define CL_COMPLETE                                 0x0
 480 #define CL_RUNNING                                  0x1
 481 #define CL_SUBMITTED                                0x2
 482 #define CL_QUEUED                                   0x3
 483
 484 #define CLK_SUCCESS                                 0
 485 #define CLK_ENQUEUE_FAILURE                         -101
 486 #define CLK_INVALID_QUEUE                           -102
 487 #define CLK_INVALID_NDRANGE                         -160
 488 #define CLK_INVALID_EVENT_WAIT_LIST                 -57
 489 #define CLK_DEVICE_QUEUE_FULL                       -161
 490 #define CLK_INVALID_ARG_SIZE                        -51
 491 #define CLK_EVENT_ALLOCATION_FAILURE                -100
 492 #define CLK_OUT_OF_RESOURCES                        -5
 493
 494 #define CLK_NULL_QUEUE                              0
 495 #define CLK_NULL_EVENT (__builtin_astype(((__SIZE_MAX__)), clk_event_t))
 496
 497 // execution model related definitions
 498 #define CLK_ENQUEUE_FLAGS_NO_WAIT                   0x0
 499 #define CLK_ENQUEUE_FLAGS_WAIT_KERNEL               0x1
 500 #define CLK_ENQUEUE_FLAGS_WAIT_WORK_GROUP           0x2
 501
 502 typedef int kernel_enqueue_flags_t;
 503 typedef int clk_profiling_info;
 504
 505 // Profiling info name (see capture_event_profiling_info)
 506 #define CLK_PROFILING_COMMAND_EXEC_TIME 0x1
 507
 508 #define MAX_WORK_DIM 3
 509
 510 #ifdef __opencl_c_device_enqueue
 511 typedef struct {
 512   unsigned int workDimension;
 513   size_t globalWorkOffset[MAX_WORK_DIM];
 514   size_t globalWorkSize[MAX_WORK_DIM];
 515   size_t localWorkSize[MAX_WORK_DIM];
 516 } ndrange_t;
 517 #endif // __opencl_c_device_enqueue
 518
 519 #endif // defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
 520
 521 /**
 522  * OpenCL v1.1/1.2/2.0 s6.2.4.2 - as_type operators
 523  * Reinterprets a data type as another data type of the same size
 524  */
 525 #define as_char(x) __builtin_astype((x), char)
 526 #define as_char2(x) __builtin_astype((x), char2)
 527 #define as_char3(x) __builtin_astype((x), char3)
 528 #define as_char4(x) __builtin_astype((x), char4)
 529 #define as_char8(x) __builtin_astype((x), char8)
 530 #define as_char16(x) __builtin_astype((x), char16)
 531
 532 #define as_uchar(x) __builtin_astype((x), uchar)
 533 #define as_uchar2(x) __builtin_astype((x), uchar2)
 534 #define as_uchar3(x) __builtin_astype((x), uchar3)
 535 #define as_uchar4(x) __builtin_astype((x), uchar4)
 536 #define as_uchar8(x) __builtin_astype((x), uchar8)
 537 #define as_uchar16(x) __builtin_astype((x), uchar16)
 538
 539 #define as_short(x) __builtin_astype((x), short)
 540 #define as_short2(x) __builtin_astype((x), short2)
 541 #define as_short3(x) __builtin_astype((x), short3)
 542 #define as_short4(x) __builtin_astype((x), short4)
 543 #define as_short8(x) __builtin_astype((x), short8)
 544 #define as_short16(x) __builtin_astype((x), short16)
 545
 546 #define as_ushort(x) __builtin_astype((x), ushort)
 547 #define as_ushort2(x) __builtin_astype((x), ushort2)
 548 #define as_ushort3(x) __builtin_astype((x), ushort3)
 549 #define as_ushort4(x) __builtin_astype((x), ushort4)
 550 #define as_ushort8(x) __builtin_astype((x), ushort8)
 551 #define as_ushort16(x) __builtin_astype((x), ushort16)
 552
 553 #define as_int(x) __builtin_astype((x), int)
 554 #define as_int2(x) __builtin_astype((x), int2)
 555 #define as_int3(x) __builtin_astype((x), int3)
 556 #define as_int4(x) __builtin_astype((x), int4)
 557 #define as_int8(x) __builtin_astype((x), int8)
 558 #define as_int16(x) __builtin_astype((x), int16)
 559
 560 #define as_uint(x) __builtin_astype((x), uint)
 561 #define as_uint2(x) __builtin_astype((x), uint2)
 562 #define as_uint3(x) __builtin_astype((x), uint3)
 563 #define as_uint4(x) __builtin_astype((x), uint4)
 564 #define as_uint8(x) __builtin_astype((x), uint8)
 565 #define as_uint16(x) __builtin_astype((x), uint16)
 566
 567 #define as_long(x) __builtin_astype((x), long)
 568 #define as_long2(x) __builtin_astype((x), long2)
 569 #define as_long3(x) __builtin_astype((x), long3)
 570 #define as_long4(x) __builtin_astype((x), long4)
 571 #define as_long8(x) __builtin_astype((x), long8)
 572 #define as_long16(x) __builtin_astype((x), long16)
 573
 574 #define as_ulong(x) __builtin_astype((x), ulong)
 575 #define as_ulong2(x) __builtin_astype((x), ulong2)
 576 #define as_ulong3(x) __builtin_astype((x), ulong3)
 577 #define as_ulong4(x) __builtin_astype((x), ulong4)
 578 #define as_ulong8(x) __builtin_astype((x), ulong8)
 579 #define as_ulong16(x) __builtin_astype((x), ulong16)
 580
 581 #define as_float(x) __builtin_astype((x), float)
 582 #define as_float2(x) __builtin_astype((x), float2)
 583 #define as_float3(x) __builtin_astype((x), float3)
 584 #define as_float4(x) __builtin_astype((x), float4)
 585 #define as_float8(x) __builtin_astype((x), float8)
 586 #define as_float16(x) __builtin_astype((x), float16)
 587
 588 #ifdef cl_khr_fp64
 589 #define as_double(x) __builtin_astype((x), double)
 590 #define as_double2(x) __builtin_astype((x), double2)
 591 #define as_double3(x) __builtin_astype((x), double3)
 592 #define as_double4(x) __builtin_astype((x), double4)
 593 #define as_double8(x) __builtin_astype((x), double8)
 594 #define as_double16(x) __builtin_astype((x), double16)
 595 #endif // cl_khr_fp64
 596
 597 #ifdef cl_khr_fp16
 598 #define as_half(x) __builtin_astype((x), half)
 599 #define as_half2(x) __builtin_astype((x), half2)
 600 #define as_half3(x) __builtin_astype((x), half3)
 601 #define as_half4(x) __builtin_astype((x), half4)
 602 #define as_half8(x) __builtin_astype((x), half8)
 603 #define as_half16(x) __builtin_astype((x), half16)
 604 #endif // cl_khr_fp16
 605
 606 #define as_size_t(x) __builtin_astype((x), size_t)
 607 #define as_ptrdiff_t(x) __builtin_astype((x), ptrdiff_t)
 608 #define as_intptr_t(x) __builtin_astype((x), intptr_t)
 609 #define as_uintptr_t(x) __builtin_astype((x), uintptr_t)
 610
 611 // C++ for OpenCL - __remove_address_space
 612 #if defined(__OPENCL_CPP_VERSION__)
 613 template <typename _Tp> struct __remove_address_space { using type = _Tp; };
 614 #if defined(__opencl_c_generic_address_space)
 615 template <typename _Tp> struct __remove_address_space<__generic _Tp> {
 616   using type = _Tp;
 617 };
 618 #endif
 619 template <typename _Tp> struct __remove_address_space<__global _Tp> {
 620   using type = _Tp;
 621 };
 622 template <typename _Tp> struct __remove_address_space<__private _Tp> {
 623   using type = _Tp;
 624 };
 625 template <typename _Tp> struct __remove_address_space<__local _Tp> {
 626   using type = _Tp;
 627 };
 628 template <typename _Tp> struct __remove_address_space<__constant _Tp> {
 629   using type = _Tp;
 630 };
 631 #endif
 632
 633 // OpenCL v1.1 s6.9, v1.2/2.0 s6.10 - Function qualifiers
 634
 635 #define __kernel_exec(X, typen) __kernel \
 636         __attribute__((work_group_size_hint(X, 1, 1))) \
 637         __attribute__((vec_type_hint(typen)))
 638
 639 #define kernel_exec(X, typen) __kernel \
 640         __attribute__((work_group_size_hint(X, 1, 1))) \
 641         __attribute__((vec_type_hint(typen)))
 642
 643 #if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_1_2)
 644 // OpenCL v1.2 s6.12.13, v2.0 s6.13.13 - printf
 645
 646 int printf(__constant const char* st, ...) __attribute__((format(printf, 1, 2)));
 647 #endif
 648
 649 #ifdef cl_intel_device_side_avc_motion_estimation
 650
 651 #define CLK_AVC_ME_MAJOR_16x16_INTEL 0x0
 652 #define CLK_AVC_ME_MAJOR_16x8_INTEL 0x1
 653 #define CLK_AVC_ME_MAJOR_8x16_INTEL 0x2
 654 #define CLK_AVC_ME_MAJOR_8x8_INTEL 0x3
 655
 656 #define CLK_AVC_ME_MINOR_8x8_INTEL 0x0
 657 #define CLK_AVC_ME_MINOR_8x4_INTEL 0x1
 658 #define CLK_AVC_ME_MINOR_4x8_INTEL 0x2
 659 #define CLK_AVC_ME_MINOR_4x4_INTEL 0x3
 660
 661 #define CLK_AVC_ME_MAJOR_FORWARD_INTEL 0x0
 662 #define CLK_AVC_ME_MAJOR_BACKWARD_INTEL 0x1
 663 #define CLK_AVC_ME_MAJOR_BIDIRECTIONAL_INTEL 0x2
 664
 665 #define CLK_AVC_ME_PARTITION_MASK_ALL_INTEL 0x0
 666 #define CLK_AVC_ME_PARTITION_MASK_16x16_INTEL 0x7E
 667 #define CLK_AVC_ME_PARTITION_MASK_16x8_INTEL 0x7D
 668 #define CLK_AVC_ME_PARTITION_MASK_8x16_INTEL 0x7B
 669 #define CLK_AVC_ME_PARTITION_MASK_8x8_INTEL 0x77
 670 #define CLK_AVC_ME_PARTITION_MASK_8x4_INTEL 0x6F
 671 #define CLK_AVC_ME_PARTITION_MASK_4x8_INTEL 0x5F
 672 #define CLK_AVC_ME_PARTITION_MASK_4x4_INTEL 0x3F
 673
 674 #define CLK_AVC_ME_SLICE_TYPE_PRED_INTEL 0x0
 675 #define CLK_AVC_ME_SLICE_TYPE_BPRED_INTEL 0x1
 676 #define CLK_AVC_ME_SLICE_TYPE_INTRA_INTEL 0x2
 677
 678 #define CLK_AVC_ME_SEARCH_WINDOW_EXHAUSTIVE_INTEL 0x0
 679 #define CLK_AVC_ME_SEARCH_WINDOW_SMALL_INTEL 0x1
 680 #define CLK_AVC_ME_SEARCH_WINDOW_TINY_INTEL 0x2
 681 #define CLK_AVC_ME_SEARCH_WINDOW_EXTRA_TINY_INTEL 0x3
 682 #define CLK_AVC_ME_SEARCH_WINDOW_DIAMOND_INTEL 0x4
 683 #define CLK_AVC_ME_SEARCH_WINDOW_LARGE_DIAMOND_INTEL 0x5
 684 #define CLK_AVC_ME_SEARCH_WINDOW_RESERVED0_INTEL 0x6
 685 #define CLK_AVC_ME_SEARCH_WINDOW_RESERVED1_INTEL 0x7
 686 #define CLK_AVC_ME_SEARCH_WINDOW_CUSTOM_INTEL 0x8
 687
 688 #define CLK_AVC_ME_SAD_ADJUST_MODE_NONE_INTEL 0x0
 689 #define CLK_AVC_ME_SAD_ADJUST_MODE_HAAR_INTEL 0x2
 690
 691 #define CLK_AVC_ME_SUBPIXEL_MODE_INTEGER_INTEL 0x0
 692 #define CLK_AVC_ME_SUBPIXEL_MODE_HPEL_INTEL 0x1
 693 #define CLK_AVC_ME_SUBPIXEL_MODE_QPEL_INTEL 0x3
 694
 695 #define CLK_AVC_ME_COST_PRECISION_QPEL_INTEL 0x0
 696 #define CLK_AVC_ME_COST_PRECISION_HPEL_INTEL 0x1
 697 #define CLK_AVC_ME_COST_PRECISION_PEL_INTEL 0x2
 698 #define CLK_AVC_ME_COST_PRECISION_DPEL_INTEL 0x3
 699
 700 #define CLK_AVC_ME_BIDIR_WEIGHT_QUARTER_INTEL 0x10
 701 #define CLK_AVC_ME_BIDIR_WEIGHT_THIRD_INTEL 0x15
 702 #define CLK_AVC_ME_BIDIR_WEIGHT_HALF_INTEL 0x20
 703 #define CLK_AVC_ME_BIDIR_WEIGHT_TWO_THIRD_INTEL 0x2B
 704 #define CLK_AVC_ME_BIDIR_WEIGHT_THREE_QUARTER_INTEL 0x30
 705
 706 #define CLK_AVC_ME_BORDER_REACHED_LEFT_INTEL 0x0
 707 #define CLK_AVC_ME_BORDER_REACHED_RIGHT_INTEL 0x2
 708 #define CLK_AVC_ME_BORDER_REACHED_TOP_INTEL 0x4
 709 #define CLK_AVC_ME_BORDER_REACHED_BOTTOM_INTEL 0x8
 710
 711 #define CLK_AVC_ME_INTRA_16x16_INTEL 0x0
 712 #define CLK_AVC_ME_INTRA_8x8_INTEL 0x1
 713 #define CLK_AVC_ME_INTRA_4x4_INTEL 0x2
 714
 715 #define CLK_AVC_ME_SKIP_BLOCK_PARTITION_16x16_INTEL 0x0
 716 #define CLK_AVC_ME_SKIP_BLOCK_PARTITION_8x8_INTEL 0x4000
 717
 718 #define CLK_AVC_ME_SKIP_BLOCK_16x16_FORWARD_ENABLE_INTEL (0x1 << 24)
 719 #define CLK_AVC_ME_SKIP_BLOCK_16x16_BACKWARD_ENABLE_INTEL (0x2 << 24)
 720 #define CLK_AVC_ME_SKIP_BLOCK_16x16_DUAL_ENABLE_INTEL (0x3 << 24)
 721 #define CLK_AVC_ME_SKIP_BLOCK_8x8_FORWARD_ENABLE_INTEL (0x55 << 24)
 722 #define CLK_AVC_ME_SKIP_BLOCK_8x8_BACKWARD_ENABLE_INTEL (0xAA << 24)
 723 #define CLK_AVC_ME_SKIP_BLOCK_8x8_DUAL_ENABLE_INTEL (0xFF << 24)
 724 #define CLK_AVC_ME_SKIP_BLOCK_8x8_0_FORWARD_ENABLE_INTEL (0x1 << 24)
 725 #define CLK_AVC_ME_SKIP_BLOCK_8x8_0_BACKWARD_ENABLE_INTEL (0x2 << 24)
 726 #define CLK_AVC_ME_SKIP_BLOCK_8x8_1_FORWARD_ENABLE_INTEL (0x1 << 26)
 727 #define CLK_AVC_ME_SKIP_BLOCK_8x8_1_BACKWARD_ENABLE_INTEL (0x2 << 26)
 728 #define CLK_AVC_ME_SKIP_BLOCK_8x8_2_FORWARD_ENABLE_INTEL (0x1 << 28)
 729 #define CLK_AVC_ME_SKIP_BLOCK_8x8_2_BACKWARD_ENABLE_INTEL (0x2 << 28)
 730 #define CLK_AVC_ME_SKIP_BLOCK_8x8_3_FORWARD_ENABLE_INTEL (0x1 << 30)
 731 #define CLK_AVC_ME_SKIP_BLOCK_8x8_3_BACKWARD_ENABLE_INTEL (0x2 << 30)
 732
 733 #define CLK_AVC_ME_BLOCK_BASED_SKIP_4x4_INTEL 0x00
 734 #define CLK_AVC_ME_BLOCK_BASED_SKIP_8x8_INTEL 0x80
 735
 736 #define CLK_AVC_ME_INTRA_LUMA_PARTITION_MASK_ALL_INTEL 0x0
 737 #define CLK_AVC_ME_INTRA_LUMA_PARTITION_MASK_16x16_INTEL 0x6
 738 #define CLK_AVC_ME_INTRA_LUMA_PARTITION_MASK_8x8_INTEL 0x5
 739 #define CLK_AVC_ME_INTRA_LUMA_PARTITION_MASK_4x4_INTEL 0x3
 740
 741 #define CLK_AVC_ME_INTRA_NEIGHBOR_LEFT_MASK_ENABLE_INTEL 0x60
 742 #define CLK_AVC_ME_INTRA_NEIGHBOR_UPPER_MASK_ENABLE_INTEL 0x10
 743 #define CLK_AVC_ME_INTRA_NEIGHBOR_UPPER_RIGHT_MASK_ENABLE_INTEL 0x8
 744 #define CLK_AVC_ME_INTRA_NEIGHBOR_UPPER_LEFT_MASK_ENABLE_INTEL 0x4
 745
 746 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_VERTICAL_INTEL 0x0
 747 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_HORIZONTAL_INTEL 0x1
 748 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_DC_INTEL 0x2
 749 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_DIAGONAL_DOWN_LEFT_INTEL 0x3
 750 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_DIAGONAL_DOWN_RIGHT_INTEL 0x4
 751 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_PLANE_INTEL 0x4
 752 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_VERTICAL_RIGHT_INTEL 0x5
 753 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_HORIZONTAL_DOWN_INTEL 0x6
 754 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_VERTICAL_LEFT_INTEL 0x7
 755 #define CLK_AVC_ME_LUMA_PREDICTOR_MODE_HORIZONTAL_UP_INTEL 0x8
 756 #define CLK_AVC_ME_CHROMA_PREDICTOR_MODE_DC_INTEL 0x0
 757 #define CLK_AVC_ME_CHROMA_PREDICTOR_MODE_HORIZONTAL_INTEL 0x1
 758 #define CLK_AVC_ME_CHROMA_PREDICTOR_MODE_VERTICAL_INTEL 0x2
 759 #define CLK_AVC_ME_CHROMA_PREDICTOR_MODE_PLANE_INTEL 0x3
 760
 761 #define CLK_AVC_ME_FRAME_FORWARD_INTEL 0x1
 762 #define CLK_AVC_ME_FRAME_BACKWARD_INTEL 0x2
 763 #define CLK_AVC_ME_FRAME_DUAL_INTEL 0x3
 764
 765 #define CLK_AVC_ME_INTERLACED_SCAN_TOP_FIELD_INTEL 0x0
 766 #define CLK_AVC_ME_INTERLACED_SCAN_BOTTOM_FIELD_INTEL 0x1
 767
 768 #define CLK_AVC_ME_INITIALIZE_INTEL 0x0
 769
 770 #define CLK_AVC_IME_PAYLOAD_INITIALIZE_INTEL 0x0
 771 #define CLK_AVC_REF_PAYLOAD_INITIALIZE_INTEL 0x0
 772 #define CLK_AVC_SIC_PAYLOAD_INITIALIZE_INTEL 0x0
 773
 774 #define CLK_AVC_IME_RESULT_INITIALIZE_INTEL 0x0
 775 #define CLK_AVC_REF_RESULT_INITIALIZE_INTEL 0x0
 776 #define CLK_AVC_SIC_RESULT_INITIALIZE_INTEL 0x0
 777
 778 #define CLK_AVC_IME_RESULT_SINGLE_REFERENCE_STREAMOUT_INITIALIZE_INTEL 0x0
 779 #define CLK_AVC_IME_RESULT_SINGLE_REFERENCE_STREAMIN_INITIALIZE_INTEL 0x0
 780 #define CLK_AVC_IME_RESULT_DUAL_REFERENCE_STREAMOUT_INITIALIZE_INTEL 0x0
 781 #define CLK_AVC_IME_RESULT_DUAL_REFERENCE_STREAMIN_INITIALIZE_INTEL 0x0
 782
 783 #endif // cl_intel_device_side_avc_motion_estimation
 784
 785 // Disable any extensions we may have enabled previously.
 786 #pragma OPENCL EXTENSION all : disable
 787
 788 #endif //_OPENCL_BASE_H_