math_ptx

Auto generated PTX files for various CUDA Math functions.

This is an alternative to using libdeivce when using NVVM IR. The PTX files can be converted to strings and used as required in your repository.

Contents

• src/arith.cu: Basic arithmetic operations (+,-,*,/).

  • Function handles: add, sub, div, mul
  • Includes operations on complex numbers

• src/logic.cu: Basic logical operations (>, <, >=, <=, ==, !=).

  • Function handles: gt, lt, ge, le, eq, neq
  • Includes operations on complex numbers
  • For complex numbers, eq and neq check for both real and imaginary values.
  • For complex numbers, everything else checks for absolute values

• src/numeric.cu: Numerical operations on numbers

  • Functions included: min, max, ceil, floor, round, rem, mod, tgamma, lgamma
  • Complex numbers supported for min, and max. Both functions compare absolute values.

• src/exp.cu: Exponential and logarithmic functions

  • Functions included: exp, expm1, erf, erfc, log, log10, log1p, sqrt, cbrt, pow
  • Complex numbers not supported

• src/trig.cu: Trigonometric functions

  • Functions included: sin, cos, tan, asin, acos, atan
  • Complex numbers not supported

• src/hyper.cu: Hyperbolic functions

  • Functions included: sinh, cosh, tanh, asinh, acosh, atanh
  • Complex numbers not supported

• PTX64/sm_xx/*.ptx: PTX files for 64 bit systems for compute sm_xx

• PTX32/sm_xx/*.ptx: PTX files for 32 bit systems for compute sm_xx

Naming scheme

The native name mangling from the auto-generated PTX is changed to something friendlier for string manipulation.

The naming scheme consists of three parts.

• three underscores (common everywhere)
• function name
• a letter denoting each of the input types.
  • float is s
  • double is d
  • cuFloatComplex is c (a.k.a float2)
  • cuDoubleComplex is z (a.k.a double2)
  • int is i
  • unsigned int is u (a.k.a uint)
  • char is j
  • unsigned char is v (a.k.a uchar)

Sample names:

• add on two floating point complex numbers would be: ___addcc.
• sin on a double precision number would be ___sind
• max of two unsigned integers would be ___maxuu