From: Qicheng Christopher Li Date: Mon, 24 May 2010 12:44:13 +0000 (+0100) Subject: llvmpipe: Unit test for sin/cos that compares against reference implementation. X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=80ee3a440cd3c0403004cf35e0638fc52088b9ff;p=mesa.git llvmpipe: Unit test for sin/cos that compares against reference implementation. Signed-off-by: José Fonseca --- diff --git a/src/gallium/drivers/llvmpipe/Makefile b/src/gallium/drivers/llvmpipe/Makefile index 526e85c82e1..213e4c4a99a 100644 --- a/src/gallium/drivers/llvmpipe/Makefile +++ b/src/gallium/drivers/llvmpipe/Makefile @@ -50,8 +50,12 @@ CPP_SOURCES = \ PROGS := lp_test_format \ lp_test_blend \ lp_test_conv \ - lp_test_printf + lp_test_printf \ + lp_test_sincos +lp_test_sincos.o : sse_mathfun.h + +PROGS_DEPS := ../../auxiliary/libgallium.a include ../../Makefile.template diff --git a/src/gallium/drivers/llvmpipe/SConscript b/src/gallium/drivers/llvmpipe/SConscript index c155558aa54..2c38dc42b09 100644 --- a/src/gallium/drivers/llvmpipe/SConscript +++ b/src/gallium/drivers/llvmpipe/SConscript @@ -76,6 +76,8 @@ if env['platform'] != 'embedded': 'format', 'blend', 'conv', + 'printf', + 'sincos', ] for test in tests: diff --git a/src/gallium/drivers/llvmpipe/lp_test_sincos.c b/src/gallium/drivers/llvmpipe/lp_test_sincos.c new file mode 100644 index 00000000000..883e15fbb6c --- /dev/null +++ b/src/gallium/drivers/llvmpipe/lp_test_sincos.c @@ -0,0 +1,204 @@ +/************************************************************************** + * + * Copyright 2010 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#include +#include + +#include "gallivm/lp_bld.h" +#include "gallivm/lp_bld_printf.h" +#include "gallivm/lp_bld_arit.h" + +#include +#include +#include +#include + +#include "lp_test.h" + + +struct sincos_test_case { +}; + + +void +write_tsv_header(FILE *fp) +{ + fprintf(fp, + "result\t" + "format\n"); + + fflush(fp); +} + + +#ifdef PIPE_ARCH_SSE + +#define USE_SSE2 +#include "sse_mathfun.h" + +typedef __m128 (*test_sincos_t)(__m128); + +static LLVMValueRef +add_sincos_test(LLVMModuleRef module, bool sin) +{ + LLVMTypeRef v4sf = LLVMVectorType(LLVMFloatType(), 4); + LLVMTypeRef args[1] = { v4sf }; + LLVMValueRef func = LLVMAddFunction(module, "sincos", LLVMFunctionType(v4sf, args, 1, 0)); + LLVMValueRef arg1 = LLVMGetParam(func, 0); + LLVMBuilderRef builder = LLVMCreateBuilder(); + LLVMBasicBlockRef block = LLVMAppendBasicBlock(func, "entry"); + LLVMValueRef ret; + struct lp_build_context bld; + + bld.builder = builder; + bld.type.floating = 1; + bld.type.width = 32; + bld.type.length = 4; + + LLVMSetFunctionCallConv(func, LLVMCCallConv); + + LLVMPositionBuilderAtEnd(builder, block); + ret = sin ? lp_build_sin(&bld, arg1) : lp_build_cos(&bld, arg1); + LLVMBuildRet(builder, ret); + LLVMDisposeBuilder(builder); + return func; +} + +static void +printv(char* string, v4sf value) +{ + v4sf v = value; + uint32_t *p = (uint32_t *) &v; + float *f = (float *)&v; + printf("%s: %f(%x) %f(%x) %f(%x) %f(%x)\n", string, + f[0], p[0], f[1], p[1], f[2], p[2], f[3], p[3]); +} + +PIPE_ALIGN_STACK +static boolean +test_sincos(unsigned verbose, FILE *fp, const struct sincos_test_case *testcase) +{ + LLVMModuleRef module = NULL; + LLVMValueRef test_sin = NULL, test_cos = NULL; + LLVMExecutionEngineRef engine = NULL; + LLVMModuleProviderRef provider = NULL; + LLVMPassManagerRef pass = NULL; + char *error = NULL; + test_sincos_t sin_func; + test_sincos_t cos_func; + float unpacked[4]; + unsigned packed; + boolean success = TRUE; + + module = LLVMModuleCreateWithName("test"); + + test_sin = add_sincos_test(module, TRUE); + test_cos = add_sincos_test(module, FALSE); + + if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) { + printf("LLVMVerifyModule: %s\n", error); + LLVMDumpModule(module); + abort(); + } + LLVMDisposeMessage(error); + + provider = LLVMCreateModuleProviderForExistingModule(module); + if (LLVMCreateJITCompiler(&engine, provider, 1, &error)) { + fprintf(stderr, "%s\n", error); + LLVMDisposeMessage(error); + abort(); + } + +#if 0 + pass = LLVMCreatePassManager(); + LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass); + /* These are the passes currently listed in llvm-c/Transforms/Scalar.h, + * but there are more on SVN. */ + LLVMAddConstantPropagationPass(pass); + LLVMAddInstructionCombiningPass(pass); + LLVMAddPromoteMemoryToRegisterPass(pass); + LLVMAddGVNPass(pass); + LLVMAddCFGSimplificationPass(pass); + LLVMRunPassManager(pass, module); +#else + (void)pass; +#endif + + sin_func = (test_sincos_t)LLVMGetPointerToGlobal(engine, test_sin); + cos_func = (test_sincos_t)LLVMGetPointerToGlobal(engine, test_cos); + + memset(unpacked, 0, sizeof unpacked); + packed = 0; + + + // LLVMDumpModule(module); + { + v4sf src = {3.14159/4.0, -3.14159/4.0, 1.0, -1.0}; + printv("ref ",sin_ps(src)); + printv("llvm", sin_func(src)); + printv("ref ",cos_ps(src)); + printv("llvm",cos_func(src)); + } + + LLVMFreeMachineCodeForFunction(engine, test_sin); + LLVMFreeMachineCodeForFunction(engine, test_cos); + + LLVMDisposeExecutionEngine(engine); + if(pass) + LLVMDisposePassManager(pass); + + return success; +} + +#else /* !PIPE_ARCH_SSE */ + +static boolean +test_sincos(unsigned verbose, FILE *fp, const struct sincos_test_case *testcase) +{ + return TRUE; +} + +#endif /* !PIPE_ARCH_SSE */ + + +boolean +test_all(unsigned verbose, FILE *fp) +{ + bool success = TRUE; + + test_sincos(verbose, fp, NULL); + + return success; +} + + +boolean +test_some(unsigned verbose, FILE *fp, unsigned long n) +{ + return test_all(verbose, fp); +} diff --git a/src/gallium/drivers/llvmpipe/sse_mathfun.h b/src/gallium/drivers/llvmpipe/sse_mathfun.h new file mode 100644 index 00000000000..8ac2064b7bb --- /dev/null +++ b/src/gallium/drivers/llvmpipe/sse_mathfun.h @@ -0,0 +1,773 @@ +/* SIMD (SSE1+MMX or SSE2) implementation of sin, cos, exp and log + + Inspired by Intel Approximate Math library, and based on the + corresponding algorithms of the cephes math library + + The default is to use the SSE1 version. If you define USE_SSE2 the + the SSE2 intrinsics will be used in place of the MMX intrinsics. Do + not expect any significant performance improvement with SSE2. +*/ + +/* Copyright (C) 2007 Julien Pommier + + This software is provided 'as-is', without any express or implied + warranty. In no event will the authors be held liable for any damages + arising from the use of this software. + + Permission is granted to anyone to use this software for any purpose, + including commercial applications, and to alter it and redistribute it + freely, subject to the following restrictions: + + 1. The origin of this software must not be misrepresented; you must not + claim that you wrote the original software. If you use this software + in a product, an acknowledgment in the product documentation would be + appreciated but is not required. + 2. Altered source versions must be plainly marked as such, and must not be + misrepresented as being the original software. + 3. This notice may not be removed or altered from any source distribution. + + (this is the zlib license) +*/ + +#include + +/* yes I know, the top of this file is quite ugly */ + +#ifdef _MSC_VER /* visual c++ */ +# define ALIGN16_BEG __declspec(align(16)) +# define ALIGN16_END +#else /* gcc or icc */ +# define ALIGN16_BEG +# define ALIGN16_END __attribute__((aligned(16))) +#endif + +/* __m128 is ugly to write */ +typedef __m128 v4sf; // vector of 4 float (sse1) + +#ifdef USE_SSE2 +# include +typedef __m128i v4si; // vector of 4 int (sse2) +#else +typedef __m64 v2si; // vector of 2 int (mmx) +#endif + +/* declare some SSE constants -- why can't I figure a better way to do that? */ +#define _PS_CONST(Name, Val) \ + static const ALIGN16_BEG float _ps_##Name[4] ALIGN16_END = { Val, Val, Val, Val } +#define _PI32_CONST(Name, Val) \ + static const ALIGN16_BEG int _pi32_##Name[4] ALIGN16_END = { Val, Val, Val, Val } +#define _PS_CONST_TYPE(Name, Type, Val) \ + static const ALIGN16_BEG Type _ps_##Name[4] ALIGN16_END = { Val, Val, Val, Val } + +_PS_CONST(1 , 1.0f); +_PS_CONST(0p5, 0.5f); +/* the smallest non denormalized float number */ +_PS_CONST_TYPE(min_norm_pos, int, 0x00800000); +_PS_CONST_TYPE(mant_mask, int, 0x7f800000); +_PS_CONST_TYPE(inv_mant_mask, int, ~0x7f800000); + +_PS_CONST_TYPE(sign_mask, int, 0x80000000); +_PS_CONST_TYPE(inv_sign_mask, int, ~0x80000000); + +_PI32_CONST(1, 1); +_PI32_CONST(inv1, ~1); +_PI32_CONST(2, 2); +_PI32_CONST(4, 4); +_PI32_CONST(0x7f, 0x7f); + +_PS_CONST(cephes_SQRTHF, 0.707106781186547524); +_PS_CONST(cephes_log_p0, 7.0376836292E-2); +_PS_CONST(cephes_log_p1, - 1.1514610310E-1); +_PS_CONST(cephes_log_p2, 1.1676998740E-1); +_PS_CONST(cephes_log_p3, - 1.2420140846E-1); +_PS_CONST(cephes_log_p4, + 1.4249322787E-1); +_PS_CONST(cephes_log_p5, - 1.6668057665E-1); +_PS_CONST(cephes_log_p6, + 2.0000714765E-1); +_PS_CONST(cephes_log_p7, - 2.4999993993E-1); +_PS_CONST(cephes_log_p8, + 3.3333331174E-1); +_PS_CONST(cephes_log_q1, -2.12194440e-4); +_PS_CONST(cephes_log_q2, 0.693359375); + +v4sf log_ps(v4sf x); +v4sf exp_ps(v4sf x); +v4sf sin_ps(v4sf x); +v4sf cos_ps(v4sf x); +void sincos_ps(v4sf x, v4sf *s, v4sf *c); + +#if defined (__MINGW32__) + +/* the ugly part below: many versions of gcc used to be completely buggy with respect to some intrinsics + The movehl_ps is fixed in mingw 3.4.5, but I found out that all the _mm_cmp* intrinsics were completely + broken on my mingw gcc 3.4.5 ... + + Note that the bug on _mm_cmp* does occur only at -O0 optimization level +*/ + +inline __m128 my_movehl_ps(__m128 a, const __m128 b) { + asm ( + "movhlps %2,%0\n\t" + : "=x" (a) + : "0" (a), "x"(b) + ); + return a; } +#warning "redefined _mm_movehl_ps (see gcc bug 21179)" +#define _mm_movehl_ps my_movehl_ps + +inline __m128 my_cmplt_ps(__m128 a, const __m128 b) { + asm ( + "cmpltps %2,%0\n\t" + : "=x" (a) + : "0" (a), "x"(b) + ); + return a; + } +inline __m128 my_cmpgt_ps(__m128 a, const __m128 b) { + asm ( + "cmpnleps %2,%0\n\t" + : "=x" (a) + : "0" (a), "x"(b) + ); + return a; +} +inline __m128 my_cmpeq_ps(__m128 a, const __m128 b) { + asm ( + "cmpeqps %2,%0\n\t" + : "=x" (a) + : "0" (a), "x"(b) + ); + return a; +} +#warning "redefined _mm_cmpxx_ps functions..." +#define _mm_cmplt_ps my_cmplt_ps +#define _mm_cmpgt_ps my_cmpgt_ps +#define _mm_cmpeq_ps my_cmpeq_ps +#endif + +#ifndef USE_SSE2 +typedef union xmm_mm_union { + __m128 xmm; + __m64 mm[2]; +} xmm_mm_union; + +#define COPY_XMM_TO_MM(xmm_, mm0_, mm1_) { \ + xmm_mm_union u; u.xmm = xmm_; \ + mm0_ = u.mm[0]; \ + mm1_ = u.mm[1]; \ +} + +#define COPY_MM_TO_XMM(mm0_, mm1_, xmm_) { \ + xmm_mm_union u; u.mm[0]=mm0_; u.mm[1]=mm1_; xmm_ = u.xmm; \ + } + +#endif // USE_SSE2 + +/* natural logarithm computed for 4 simultaneous float + return NaN for x <= 0 +*/ +v4sf log_ps(v4sf x) { +#ifdef USE_SSE2 + v4si emm0; +#else + v2si mm0, mm1; +#endif + v4sf one = *(v4sf*)_ps_1; + + v4sf invalid_mask = _mm_cmple_ps(x, _mm_setzero_ps()); + v4sf e, mask, tmp, z, y; + + x = _mm_max_ps(x, *(v4sf*)_ps_min_norm_pos); /* cut off denormalized stuff */ + +#ifndef USE_SSE2 + /* part 1: x = frexpf(x, &e); */ + COPY_XMM_TO_MM(x, mm0, mm1); + mm0 = _mm_srli_pi32(mm0, 23); + mm1 = _mm_srli_pi32(mm1, 23); +#else + emm0 = _mm_srli_epi32(_mm_castps_si128(x), 23); +#endif + /* keep only the fractional part */ + x = _mm_and_ps(x, *(v4sf*)_ps_inv_mant_mask); + x = _mm_or_ps(x, *(v4sf*)_ps_0p5); + +#ifndef USE_SSE2 + /* now e=mm0:mm1 contain the really base-2 exponent */ + mm0 = _mm_sub_pi32(mm0, *(v2si*)_pi32_0x7f); + mm1 = _mm_sub_pi32(mm1, *(v2si*)_pi32_0x7f); + e = _mm_cvtpi32x2_ps(mm0, mm1); + _mm_empty(); /* bye bye mmx */ +#else + emm0 = _mm_sub_epi32(emm0, *(v4si*)_pi32_0x7f); + e = _mm_cvtepi32_ps(emm0); +#endif + + e = _mm_add_ps(e, one); + + /* part2: + if( x < SQRTHF ) { + e -= 1; + x = x + x - 1.0; + } else { x = x - 1.0; } + */ + + mask = _mm_cmplt_ps(x, *(v4sf*)_ps_cephes_SQRTHF); + tmp = _mm_and_ps(x, mask); + x = _mm_sub_ps(x, one); + e = _mm_sub_ps(e, _mm_and_ps(one, mask)); + x = _mm_add_ps(x, tmp); + + + z = _mm_mul_ps(x,x); + + y = *(v4sf*)_ps_cephes_log_p0; + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p1); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p2); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p3); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p4); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p5); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p6); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p7); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p8); + y = _mm_mul_ps(y, x); + + y = _mm_mul_ps(y, z); + + + tmp = _mm_mul_ps(e, *(v4sf*)_ps_cephes_log_q1); + y = _mm_add_ps(y, tmp); + + + tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5); + y = _mm_sub_ps(y, tmp); + + tmp = _mm_mul_ps(e, *(v4sf*)_ps_cephes_log_q2); + x = _mm_add_ps(x, y); + x = _mm_add_ps(x, tmp); + x = _mm_or_ps(x, invalid_mask); // negative arg will be NAN + return x; +} + +_PS_CONST(exp_hi, 88.3762626647949f); +_PS_CONST(exp_lo, -88.3762626647949f); + +_PS_CONST(cephes_LOG2EF, 1.44269504088896341); +_PS_CONST(cephes_exp_C1, 0.693359375); +_PS_CONST(cephes_exp_C2, -2.12194440e-4); + +_PS_CONST(cephes_exp_p0, 1.9875691500E-4); +_PS_CONST(cephes_exp_p1, 1.3981999507E-3); +_PS_CONST(cephes_exp_p2, 8.3334519073E-3); +_PS_CONST(cephes_exp_p3, 4.1665795894E-2); +_PS_CONST(cephes_exp_p4, 1.6666665459E-1); +_PS_CONST(cephes_exp_p5, 5.0000001201E-1); + +v4sf exp_ps(v4sf x) { + v4sf tmp = _mm_setzero_ps(), fx; +#ifdef USE_SSE2 + v4si emm0; +#else + v2si mm0, mm1; +#endif + v4sf one = *(v4sf*)_ps_1; + v4sf mask, z, y, pow2n; + + x = _mm_min_ps(x, *(v4sf*)_ps_exp_hi); + x = _mm_max_ps(x, *(v4sf*)_ps_exp_lo); + + /* express exp(x) as exp(g + n*log(2)) */ + fx = _mm_mul_ps(x, *(v4sf*)_ps_cephes_LOG2EF); + fx = _mm_add_ps(fx, *(v4sf*)_ps_0p5); + + /* how to perform a floorf with SSE: just below */ +#ifndef USE_SSE2 + /* step 1 : cast to int */ + tmp = _mm_movehl_ps(tmp, fx); + mm0 = _mm_cvttps_pi32(fx); + mm1 = _mm_cvttps_pi32(tmp); + /* step 2 : cast back to float */ + tmp = _mm_cvtpi32x2_ps(mm0, mm1); +#else + emm0 = _mm_cvttps_epi32(fx); + tmp = _mm_cvtepi32_ps(emm0); +#endif + /* if greater, substract 1 */ + mask = _mm_cmpgt_ps(tmp, fx); + mask = _mm_and_ps(mask, one); + fx = _mm_sub_ps(tmp, mask); + + tmp = _mm_mul_ps(fx, *(v4sf*)_ps_cephes_exp_C1); + z = _mm_mul_ps(fx, *(v4sf*)_ps_cephes_exp_C2); + x = _mm_sub_ps(x, tmp); + x = _mm_sub_ps(x, z); + + z = _mm_mul_ps(x,x); + + y = *(v4sf*)_ps_cephes_exp_p0; + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p1); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p2); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p3); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p4); + y = _mm_mul_ps(y, x); + y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p5); + y = _mm_mul_ps(y, z); + y = _mm_add_ps(y, x); + y = _mm_add_ps(y, one); + + /* build 2^n */ +#ifndef USE_SSE2 + z = _mm_movehl_ps(z, fx); + mm0 = _mm_cvttps_pi32(fx); + mm1 = _mm_cvttps_pi32(z); + mm0 = _mm_add_pi32(mm0, *(v2si*)_pi32_0x7f); + mm1 = _mm_add_pi32(mm1, *(v2si*)_pi32_0x7f); + mm0 = _mm_slli_pi32(mm0, 23); + mm1 = _mm_slli_pi32(mm1, 23); + + COPY_MM_TO_XMM(mm0, mm1, pow2n); + _mm_empty(); +#else + emm0 = _mm_cvttps_epi32(fx); + emm0 = _mm_add_epi32(emm0, *(v4si*)_pi32_0x7f); + emm0 = _mm_slli_epi32(emm0, 23); + pow2n = _mm_castsi128_ps(emm0); +#endif + y = _mm_mul_ps(y, pow2n); + return y; +} + +_PS_CONST(minus_cephes_DP1, -0.78515625); +_PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4); +_PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8); +_PS_CONST(sincof_p0, -1.9515295891E-4); +_PS_CONST(sincof_p1, 8.3321608736E-3); +_PS_CONST(sincof_p2, -1.6666654611E-1); +_PS_CONST(coscof_p0, 2.443315711809948E-005); +_PS_CONST(coscof_p1, -1.388731625493765E-003); +_PS_CONST(coscof_p2, 4.166664568298827E-002); +_PS_CONST(cephes_FOPI, 1.27323954473516); // 4 / M_PI + + +/* evaluation of 4 sines at onces, using only SSE1+MMX intrinsics so + it runs also on old athlons XPs and the pentium III of your grand + mother. + + The code is the exact rewriting of the cephes sinf function. + Precision is excellent as long as x < 8192 (I did not bother to + take into account the special handling they have for greater values + -- it does not return garbage for arguments over 8192, though, but + the extra precision is missing). + + Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the + surprising but correct result. + + Performance is also surprisingly good, 1.33 times faster than the + macos vsinf SSE2 function, and 1.5 times faster than the + __vrs4_sinf of amd's ACML (which is only available in 64 bits). Not + too bad for an SSE1 function (with no special tuning) ! + However the latter libraries probably have a much better handling of NaN, + Inf, denormalized and other special arguments.. + + On my core 1 duo, the execution of this function takes approximately 95 cycles. + + From what I have observed on the experiments with Intel AMath lib, switching to an + SSE2 version would improve the perf by only 10%. + + Since it is based on SSE intrinsics, it has to be compiled at -O2 to + deliver full speed. +*/ +v4sf sin_ps(v4sf x) { // any x + v4sf xmm1, xmm2 = _mm_setzero_ps(), xmm3, sign_bit, y; + +#ifdef USE_SSE2 + v4si emm0, emm2; +#else + v2si mm0, mm1, mm2, mm3; +#endif + v4sf swap_sign_bit, poly_mask, z, tmp, y2; + + sign_bit = x; + /* take the absolute value */ + x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask); + /* extract the sign bit (upper one) */ + sign_bit = _mm_and_ps(sign_bit, *(v4sf*)_ps_sign_mask); + + /* scale by 4/Pi */ + y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI); + + //printf("plop:"); print4(y); +#ifdef USE_SSE2 + /* store the integer part of y in mm0 */ + emm2 = _mm_cvttps_epi32(y); + /* j=(j+1) & (~1) (see the cephes sources) */ + emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1); + emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1); + y = _mm_cvtepi32_ps(emm2); + /* get the swap sign flag */ + emm0 = _mm_and_si128(emm2, *(v4si*)_pi32_4); + emm0 = _mm_slli_epi32(emm0, 29); + /* get the polynom selection mask + there is one polynom for 0 <= x <= Pi/4 + and another one for Pi/4