Floating Point
Simpler and More Efficient
1 throughput / cycle / core
- Fixed low latencies for each function (<10 cycles for most functions)
- 1.0 GFFPS (Giga Floating-point Functions Per Second) / core @ 1GHz
- Most functions have exact precision with 0 error; a few have 1 maximum ulp
65 OpenCL math
functions ready
We support 65 OpenCL floating-point single-precision math functions / core and 11 floating-point double-precision math functions / core.
Using our Floating-Point Math technology, you got faster computation
and better performance with lower power usage.
Our hardware OpenCL math library contains the following:
exp10, exp2, acos, acosh, acospi, asin, asinh, asinpi, atan, atan2, atan2pi, atanh, atanpi, cbrt, ceil, copysign, cos, cosh, cospi, erf, erfc, ex, exm1, fabs, fadd, fclass_s, fcvt_s_w, fcvt_s_wu, fcvt_w_s, fcvt_wu_s, fdim, fdiv, feq_s, fle_s, floor, flt_s, fmadd, fmax, fmin, fmod, fmsub, fmul, fmv_s_w, fmv_w_s, fnmadd, fnmsub, fract, frexp, fsgnj_s, fsgnjn_s, fsgnjx_s, fsub, hypot, ilogb, ldexp, lgamma, lgammar, log, log10, log1p, log2, logb, maxmag, minmag, modf, nan, nextafter, nop, pow, pown, powr, rcp, remainder, remquo, rint, rootn, round, rsqrt, sigmoid, sin, sinh, sinpi, sqrt, tan, tanh, tanpi, tgamma, and trunc.
fadd
fsub
fmul
sqrt
fnmadd
&
fnmsub
- All single-precision single-operand functions are exhaustively verified against Sollya 168-bit test, except lgamma and tgamma, which Sollya does not cover.
- Single-precision multi-operand and double-precision functions are also verified against 168-bit tests powered by the Sollya math library.
- Single-precision lgamma and tgamma are verified against 64-bit Intel math library.