.file "erff.s" // Copyright (c) 2001 - 2005, Intel Corporation // All rights reserved. // // Contributed 2001 by the Intel Numerics Group, Intel Corporation // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are // met: // // * Redistributions of source code must retain the above copyright // notice, this list of conditions and the following disclaimer. // // * Redistributions in binary form must reproduce the above copyright // notice, this list of conditions and the following disclaimer in the // documentation and/or other materials provided with the distribution. // // * The name of Intel Corporation may not be used to endorse or promote // products derived from this software without specific prior written // permission. // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. // // Intel Corporation is the author of this code, and requests that all // problem reports or change requests be submitted to it directly at // http://www.intel.com/software/products/opensource/libraries/num.htm. // // History //============================================================== // 08/14/01 Initial version // 05/20/02 Cleaned up namespace and sf0 syntax // 02/06/03 Reordered header: .section, .global, .proc, .align // 03/31/05 Reformatted delimiters between data tables // // API //============================================================== // float erff(float) // // Overview of operation //============================================================== // Background // // // There are 8 paths: // 1. x = +/-0.0 // Return erff(x) = +/-0.0 // // 2. 0.0 < |x| < 0.125 // Return erff(x) = x *Pol3(x^2), // where Pol3(x^2) = C3*x^6 + C2*x^4 + C1*x^2 + C0 // // 3. 0.125 <= |x| < 4.0 // Return erff(x) = sign(x)*PolD(x)*PolC(|x|) + sign(x)*PolA(|x|), // where sign(x)*PolD(x) = sign(x)*(|x|^7 + D2*x^6 + D1*|x|^5 + D0*x^4), // PolC(|x|) = B0*x^4 + C3*|x|^3 + C2*|x|^2 + C1*|x| + C0, // PolA(|x|) = A3|x|^3 + A2*x^2 + A1*|x| + A0 // // Actually range 0.125<=|x|< 4.0 is splitted to 5 subranges. // For each subrange there is particular set of coefficients. // Below is the list of subranges: // 3.1 0.125 <= |x| < 0.25 // 3.2 0.25 <= |x| < 0.5 // 3.3 0.5 <= |x| < 1.0 // 3.4 1.0 <= |x| < 2.0 // 3.5 2.0 <= |x| < 4.0 // // 4. 4.0 <= |x| < +INF // Return erff(x) = sign(x)*(1.0d - 2^(-52)) // // 5. |x| = INF // Return erff(x) = sign(x) * 1.0 // // 6. x = [S,Q]NaN // Return erff(x) = QNaN // // 7. x is positive denormal // Return erff(x) = C0*x - x^2, // where C0 = 2.0/sqrt(Pi) // // 8. x is negative denormal // Return erff(x) = C0*x + x^2, // where C0 = 2.0/sqrt(Pi) // // Registers used //============================================================== // Floating Point registers used: // f8, input // f32 -> f59 // General registers used: // r32 -> r45, r2, r3 // Predicate registers used: // p0, p6 -> p12, p14, p15 // p6 to filter out case when x = [Q,S]NaN or +/-0 // p7 to filter out case when x = denormal // p8 set if |x| >= 0.3125, used also to process denormal input // p9 to filter out case when |x| = inf // p10 to filter out case when |x| < 0.125 // p11 to filter out case when 0.125 <= |x| < 4.0 // p12 to filter out case when |x| >= 4.0 // p14 set to 1 for positive x // p15 set to 1 for negative x // Assembly macros //============================================================== rDataPtr = r2 rDataPtr1 = r3 rBias = r33 rCoeffAddr3 = r34 rCoeffAddr1 = r35 rCoeffAddr2 = r36 rOffset2 = r37 rBias2 = r38 rMask = r39 rArg = r40 rBound = r41 rSignBit = r42 rAbsArg = r43 rDataPtr2 = r44 rSaturation = r45 //============================================================== fA0 = f32 fA1 = f33 fA2 = f34 fA3 = f35 fC0 = f36 fC1 = f37 fC2 = f38 fC3 = f39 fD0 = f40 fD1 = f41 fD2 = f42 fB0 = f43 fArgSqr = f44 fAbsArg = f45 fSignumX = f46 fArg4 = f47 fArg4Sgn = f48 fArg3 = f49 fArg3Sgn = f50 fArg7Sgn = f51 fArg6Sgn = f52 fPolC = f53 fPolCTmp = f54 fPolA = f55 fPolATmp = f56 fPolD = f57 fPolDTmp = f58 fArgSqrSgn = f59 // Data tables //============================================================== RODATA .align 16 LOCAL_OBJECT_START(erff_data) // Polynomial coefficients for the erf(x), 0.125 <= |x| < 0.25 data8 0xBE4218BB56B49E66 // C0 data8 0x3F7AFB8315DA322B // C1 data8 0x3F615D6EBEE0CA32 // C2 data8 0xBF468D71CF4F0918 // C3 data8 0x40312115B0932F24 // D0 data8 0xC0160D6CD0991EA3 // D1 data8 0xBFE04A567A6DBE4A // D2 data8 0xBF4207BC640D1509 // B0 // Polynomial coefficients for the erf(x), 0.25 <= |x| < 0.5 data8 0x3F90849356383F58 // C0 data8 0x3F830BD5BA240F09 // C1 data8 0xBF3FA4970E2BCE23 // C2 data8 0xBF6061798E58D0FD // C3 data8 0xBF68C0D83DD22E02 // D0 data8 0x401C0A9EE4108F94 // D1 data8 0xC01056F9B5E387F5 // D2 data8 0x3F1C9744E36A5706 // B0 // Polynomial coefficients for the erf(x), 0.5 <= |x| < 1.0 data8 0x3F85F7D419A13DE3 // C0 data8 0x3F791A13FF66D45A // C1 data8 0x3F46B17B16B5929F // C2 data8 0xBF5124947A8BF45E // C3 data8 0x3FA1B3FD95EA9564 // D0 data8 0x40250CECD79A020A // D1 data8 0xC0190DC96FF66CCD // D2 data8 0x3F4401AE28BA4DD5 // B0 // Polynomial coefficients for the erf(x), 1.0 <= |x| < 2.0 data8 0xBF49E07E3584C3AE // C0 data8 0x3F3166621131445C // C1 data8 0xBF65B7FC1EAC2099 // C2 data8 0x3F508C6BD211D736 // C3 data8 0xC053FABD70601067 // D0 data8 0x404A06640EE87808 // D1 data8 0xC0283F30817A3F08 // D2 data8 0xBF2F6DBBF4D6257F // B0 // Polynomial coefficients for the erf(x), 2.0 <= |x| < 4.0 data8 0xBF849855D67E9407 // C0 data8 0x3F5ECA5FEC01C70C // C1 data8 0xBF483110C30FABA4 // C2 data8 0x3F1618DA72860403 // C3 data8 0xC08A5C9D5FE8B9F6 // D0 data8 0x406EFF5F088CEC4B // D1 data8 0xC03A5743DF38FDE0 // D2 data8 0xBEE397A9FA5686A2 // B0 // Polynomial coefficients for the erf(x), -0.125 < x < 0.125 data8 0x3FF20DD7504270CB // C0 data8 0xBFD8127465AFE719 // C1 data8 0x3FBCE2D77791DD77 // C2 data8 0xBF9B582755CDF345 // C3 // Polynomial coefficients for the erf(x), 0.125 <= |x| < 0.25 data8 0xBD54E7E451AF0E36 // A0 data8 0x3FF20DD75043FE20 // A1 data8 0xBE05680ACF8280E4 // A2 data8 0xBFD812745E92C3D3 // A3 // Polynomial coefficients for the erf(x), 0.25 <= |x| < 0.5 data8 0xBE1ACEC2859CB55F // A0 data8 0x3FF20DD75E8D2B64 // A1 data8 0xBEABC6A83208FCFC // A2 data8 0xBFD81253E42E7B99 // A3 // Polynomial coefficients for the erf(x), 0.5 <= |x| < 1.0 data8 0x3EABD5A2482B4979 // A0 data8 0x3FF20DCAA52085D5 // A1 data8 0x3F13A994A348795B // A2 data8 0xBFD8167B2DFCDE44 // A3 // Polynomial coefficients for the erf(x), 1.0 <= |x| < 2.0 data8 0xBF5BA377DDAB4E17 // A0 data8 0x3FF2397F1D8FC0ED // A1 data8 0xBF9945BFC1915C21 // A2 data8 0xBFD747AAABB690D8 // A3 // Polynomial coefficients for the erf(x), 2.0 <= |x| < 4.0 data8 0x3FF0E2920E0391AF // A0 data8 0xC00D249D1A95A5AE // A1 data8 0x40233905061C3803 // A2 data8 0xC027560B851F7690 // A3 // data8 0x3FEFFFFFFFFFFFFF // 1.0 - epsilon data8 0x3FF20DD750429B6D // C0 = 2.0/sqrt(Pi) LOCAL_OBJECT_END(erff_data) .section .text GLOBAL_LIBM_ENTRY(erff) { .mfi alloc r32 = ar.pfs, 0, 14, 0, 0 fmerge.s fAbsArg = f1, f8 // |x| addl rMask = 0x806, r0 } { .mfi addl rDataPtr = @ltoff(erff_data), gp fma.s1 fArgSqr = f8, f8, f0 // x^2 adds rSignBit = 0x1, r0 } ;; { .mfi getf.s rArg = f8 // x in GR fclass.m p7,p0 = f8, 0x0b // is x denormal ? // sign bit and 2 most bits in significand shl rMask = rMask, 20 } { .mfi ld8 rDataPtr = [rDataPtr] nop.f 0 adds rBias2 = 0x1F0, r0 } ;; { .mfi nop.m 0 fmerge.s fSignumX = f8, f1 // signum(x) shl rSignBit = rSignBit, 31 // mask for sign bit } { .mfi adds rBound = 0x3E0, r0 nop.f 0 adds rSaturation = 0x408, r0 } ;; { .mfi andcm rOffset2 = rArg, rMask fclass.m p6,p0 = f8, 0xc7 // is x [S,Q]NaN or +/-0 ? shl rBound = rBound, 20 // 0.125f in GR } { .mfb andcm rAbsArg = rArg, rSignBit // |x| in GR nop.f 0 (p7) br.cond.spnt erff_denormal // branch out if x is denormal } ;; { .mfi adds rCoeffAddr2 = 352, rDataPtr fclass.m p9,p0 = f8, 0x23 // is x +/- inf? shr rOffset2 = rOffset2, 21 } { .mfi cmp.lt p10, p8 = rAbsArg, rBound // |x| < 0.125? nop.f 0 adds rCoeffAddr3 = 16, rDataPtr } ;; { .mfi (p8) sub rBias = rOffset2, rBias2 fma.s1 fArg4 = fArgSqr, fArgSqr, f0 // x^4 shl rSaturation = rSaturation, 20// 4.0 in GR (saturation bound) } { .mfb (p10) adds rBias = 0x14, r0 (p6) fma.s.s0 f8 = f8,f1,f8 // NaN or +/-0 (p6) br.ret.spnt b0 // exit for x = NaN or +/-0 } ;; { .mfi shladd rCoeffAddr1 = rBias, 4, rDataPtr fma.s1 fArg3Sgn = fArgSqr, f8, f0 // sign(x)*|x|^3 // is |x| < 4.0? cmp.lt p11, p12 = rAbsArg, rSaturation } { .mfi shladd rCoeffAddr3 = rBias, 4, rCoeffAddr3 fma.s1 fArg3 = fArgSqr, fAbsArg, f0 // |x|^3 shladd rCoeffAddr2 = rBias, 3, rCoeffAddr2 } ;; { .mfi (p11) ldfpd fC0, fC1 = [rCoeffAddr1] (p9) fmerge.s f8 = f8,f1 // +/- inf (p12) adds rDataPtr = 512, rDataPtr } { .mfb (p11) ldfpd fC2, fC3 = [rCoeffAddr3], 16 nop.f 0 (p9) br.ret.spnt b0 // exit for x = +/- inf } ;; { .mfi (p11) ldfpd fA0, fA1 = [rCoeffAddr2], 16 nop.f 0 nop.i 0 } { .mfi add rCoeffAddr1 = 48, rCoeffAddr1 nop.f 0 nop.i 0 } ;; { .mfi (p11) ldfpd fD0, fD1 = [rCoeffAddr3] nop.f 0 nop.i 0 } { .mfb (p11) ldfpd fD2, fB0 = [rCoeffAddr1] // sign(x)*|x|^2 fma.s1 fArgSqrSgn = fArgSqr, fSignumX, f0 (p10) br.cond.spnt erff_near_zero } ;; { .mfi (p11) ldfpd fA2, fA3 = [rCoeffAddr2], 16 fcmp.lt.s1 p15, p14 = f8,f0 nop.i 0 } { .mfb (p12) ldfd fA0 = [rDataPtr] fma.s1 fArg4Sgn = fArg4, fSignumX, f0 // sign(x)*|x|^4 (p12) br.cond.spnt erff_saturation } ;; { .mfi nop.m 0 fma.s1 fArg7Sgn = fArg4, fArg3Sgn, f0 // sign(x)*|x|^7 nop.i 0 } { .mfi nop.m 0 fma.s1 fArg6Sgn = fArg3, fArg3Sgn, f0 // sign(x)*|x|^6 nop.i 0 } ;; { .mfi nop.m 0 fma.s1 fPolC = fC3, fAbsArg, fC2 // C3*|x| + C2 nop.i 0 } { .mfi nop.m 0 fma.s1 fPolCTmp = fC1, fAbsArg, fC0 // C1*|x| + C0 nop.i 0 };; { .mfi nop.m 0 fma.s1 fPolA = fA1, fAbsArg, fA0 // A1*|x| + A0 nop.i 0 } ;; { .mfi nop.m 0 fma.s1 fPolD = fD1, fAbsArg, fD0 // D1*|x| + D0 nop.i 0 } { .mfi nop.m 0 // sign(x)*(|x|^7 + D2*x^6) fma.s1 fPolDTmp = fArg6Sgn, fD2, fArg7Sgn nop.i 0 };; { .mfi nop.m 0 fma.s1 fPolATmp = fA3, fAbsArg, fA2 // A3*|x| + A2 nop.i 0 } { .mfi nop.m 0 fma.s1 fB0 = fB0, fArg4, f0 // B0*x^4 nop.i 0 };; { .mfi nop.m 0 // C3*|x|^3 + C2*x^2 + C1*|x| + C0 fma.s1 fPolC = fPolC, fArgSqr, fPolCTmp nop.i 0 } ;; { .mfi nop.m 0 // PolD = sign(x)*(|x|^7 + D2*x^6 + D1*|x|^5 + D0*x^4) fma.d.s1 fPolD = fPolD, fArg4Sgn, fPolDTmp nop.i 0 } ;; { .mfi nop.m 0 // PolA = A3|x|^3 + A2*x^2 + A1*|x| + A0 fma.d.s1 fPolA = fPolATmp, fArgSqr, fPolA nop.i 0 } ;; { .mfi nop.m 0 // PolC = B0*x^4 + C3*|x|^3 + C2*|x|^2 + C1*|x| + C0 fma.d.s1 fPolC = fPolC, f1, fB0 nop.i 0 } ;; { .mfi nop.m 0 (p14) fma.s.s0 f8 = fPolC, fPolD, fPolA // for positive x nop.i 0 } { .mfb nop.m 0 (p15) fms.s.s0 f8 = fPolC, fPolD, fPolA // for negative x br.ret.sptk b0 // Exit for 0.125 <=|x|< 4.0 };; // Here if |x| < 0.125 erff_near_zero: { .mfi nop.m 0 fma.s1 fPolC = fC3, fArgSqr, fC2 // C3*x^2 + C2 nop.i 0 } { .mfi nop.m 0 fma.s1 fPolCTmp = fC1, fArgSqr, fC0 // C1*x^2 + C0 nop.i 0 };; { .mfi nop.m 0 fma.s1 fPolC = fPolC, fArg4, fPolCTmp // C3*x^6 + C2*x^4 + C1*x^2 + C0 nop.i 0 };; { .mfb nop.m 0 // x*(C3*x^6 + C2*x^4 + C1*x^2 + C0) fma.s.s0 f8 = fPolC, f8, f0 br.ret.sptk b0 // Exit for |x| < 0.125 };; // Here if 4.0 <= |x| < +inf erff_saturation: { .mfb nop.m 0 fma.s.s0 f8 = fA0, fSignumX, f0 // sign(x)*(1.0d - 2^(-52)) // Exit for 4.0 <= |x| < +inf br.ret.sptk b0 // Exit for 4.0 <=|x|< +inf } ;; // Here if x is single precision denormal erff_denormal: { .mfi adds rDataPtr = 520, rDataPtr // address of C0 fclass.m p7,p8 = f8, 0x0a // is x -denormal ? nop.i 0 } ;; { .mfi ldfd fC0 = [rDataPtr] // C0 nop.f 0 nop.i 0 } ;; { .mfi nop.m 0 fma.s1 fC0 = fC0,f8,f0 // C0*x nop.i 0 } ;; { .mfi nop.m 0 (p7) fma.s.s0 f8 = f8,f8,fC0 // -denormal nop.i 0 } { .mfb nop.m 0 (p8) fnma.s.s0 f8 = f8,f8,fC0 // +denormal br.ret.sptk b0 // Exit for denormal } ;; GLOBAL_LIBM_END(erff)