src/common/RTMath.h

/***************************************************************************
 *                                                                         *
 *   LinuxSampler - modular, streaming capable sampler                     *
 *                                                                         *
 *   Copyright (C) 2003, 2004 by Benno Senoner and Christian Schoenebeck   *
 *   Copyright (C) 2005 - 2016 Christian Schoenebeck                       *
 *                                                                         *
 *   This program is free software; you can redistribute it and/or modify  *
 *   it under the terms of the GNU General Public License as published by  *
 *   the Free Software Foundation; either version 2 of the License, or     *
 *   (at your option) any later version.                                   *
 *                                                                         *
 *   This program is distributed in the hope that it will be useful,       *
 *   but WITHOUT ANY WARRANTY; without even the implied warranty of        *
 *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the         *
 *   GNU General Public License for more details.                          *
 *                                                                         *
 *   You should have received a copy of the GNU General Public License     *
 *   along with this program; if not, write to the Free Software           *
 *   Foundation, Inc., 59 Temple Place, Suite 330, Boston,                 *
 *   MA  02111-1307  USA                                                   *
 ***************************************************************************/

#ifndef __RT_MATH_H__
#define __RT_MATH_H__

#include <math.h>
#include <stdint.h>
#include "global_private.h"

/// Needed for calculating frequency ratio used to pitch a sample
#define TWELVEHUNDREDTH_ROOT_OF_TWO     1.000577789506555

enum implementation_t {
    CPP,
    ASM_X86_MMX_SSE
};

/** @brief Real Time Math Base Class
 *
 * Math functions for real time operation. This base class contains all
 * non-template methods.
 */
class RTMathBase {
    public:
        /**
         * Highly accurate time stamp.
         */
        typedef uint32_t time_stamp_t;

        /**
         * We read the processor's cycle count register as a reference
         * for the real time. These are of course only abstract values
         * with arbitrary time entity, but that's not a problem as long
         * as we calculate relatively.
         */
        static time_stamp_t CreateTimeStamp();

        /**
         * Calculates the frequency ratio for a pitch value given in cents
         * (assuming equal tempered scale of course, divided into 12
         * semitones per octave and 100 cents per semitone).
         *
         * Note: CONFIG_MAX_PITCH (defined in config.h) has to be defined to an
         * appropriate value, otherwise the behavior of this function is
         * undefined, but most probably if CONFIG_MAX_PITCH is too small, the
         * application will crash due to segmentation fault here.
         *
         * @param cents - pitch value in cents (+1200 cents means +1 octave)
         * @returns  frequency ratio (e.g. +2.0 for +1 octave)
         */
        inline static double CentsToFreqRatio(double Cents) {
            int   index_int   = (int) (Cents);      // integer index
            float index_fract = Cents - index_int;  // fractional part of index
            return pCentsToFreqTable[index_int] + index_fract * (pCentsToFreqTable[index_int+1] - pCentsToFreqTable[index_int]);
        }

        /**
         * Slower version of CentsToFreqRatio, for big values.
         *
         * @param cents - pitch value in cents (+1200 cents means +1 octave)
         * @returns  frequency ratio (e.g. +2.0 for +1 octave)
         */
        static double CentsToFreqRatioUnlimited(double Cents) {
            int octaves = int(Cents / 1200);
            double x = CentsToFreqRatio(Cents - octaves * 1200);
            return  octaves < 0 ? x / (1 << -octaves) : x * (1 << octaves);
        }

        /**
         * Inverse function to CentsToFreqRatio(). This function is a bit
         * slow, so it should not be called too frequently.
         */
        static double FreqRatioToCents(double FreqRatio) {
            return log(FreqRatio) / log(TWELVEHUNDREDTH_ROOT_OF_TWO);
        }

        /**
         * Calculates the line ratio value representation (linear scale)
         * of the @a decibel value provided (exponential scale).
         *
         * The context of audio acoustic sound pressure levels is assumed, and
         * hence the field version of the dB unit is used here (which uses a
         * linear factor of 20). This function is a bit slow, so it should
         * not be called too frequently.
         *
         * @param decibel - sound pressure level in dB
         * @returns linear ratio of the supplied dB value
         */
        static float DecibelToLinRatio(float decibel) {
            return powf(10.f, decibel / 20.f);
        }

        /**
         * Calculates the relatively summed average of a set of values.
         *
         * @param current - the current avaerage value of all previously summed values
         * @param sample - new value to be applied as summed average to the existing values
         * @param n - amount of sample values applied so far
         * @returns new average value of all summed values (including the new @a sample)
         */
        inline static float RelativeSummedAvg(float current, float sample, int n) {
            return current + (sample - current) / float(n);
        }

    private:
        static float* pCentsToFreqTable;

        static float* InitCentsToFreqTable();
};

/** @brief Real Time Math
 *
 * This is a template which provides customized methods for the desired low
 * level implementation. The ASM_X86_MMX_SSE implementation of each method
 * for example doesn't use 387 FPU instruction. This is needed for MMX
 * algorithms which do not allow mixed MMX and 387 instructions.
 */
template<implementation_t IMPL = CPP>
class __RTMath : public RTMathBase {
    public:
        // conversion using truncate
        inline static int Int(const float a) {
            switch (IMPL) {
                #if CONFIG_ASM && ARCH_X86
                case ASM_X86_MMX_SSE: {
                    int ret;
                    asm (
                        "cvttss2si %1, %0  # convert to int\n\t"
                        : "=r" (ret)
                        : "m" (a)
                    );
                    return ret;
                }
                #endif // CONFIG_ASM && ARCH_X86
                default: {
                    return (int) a;
                }
            }
        }

        //for doubles and everything else except floats
        template<class T_a> inline static int Int(const T_a a) {
            return (int) a;
        }

        inline static float Float(const int a) {
            switch (IMPL) {
                #if CONFIG_ASM && ARCH_X86
                case ASM_X86_MMX_SSE: {
                    float ret;
                    asm (
                        "cvtsi2ss %1, %%xmm0  # convert to float\n\t"
                        "movss    %%xmm0,%0   # output\n\t"
                        : "=m" (ret)
                        : "r" (a)
                    );
                    return ret;
                }
                #endif // CONFIG_ASM && ARCH_X86
                default: {
                    return (float) a;
                }
            }
        }

#if 0
        //for everything except ints
        template<class T_a> inline static float Float(T_a a) {
            return (float) a;
        }
#endif

        inline static float Sum(const float& a, const float& b) {
            switch (IMPL) {
                #if CONFIG_ASM && ARCH_X86
                case ASM_X86_MMX_SSE: {
                    float ret;
                    asm (
                        "movss    %1, %%xmm0  # load a\n\t"
                        "addss    %2, %%xmm0  # a + b\n\t"
                        "movss    %%xmm0, %0  # output\n\t"
                        : "=m" (ret)
                        : "m" (a), "m" (b)
                    );
                    return ret;
                }
                #endif // CONFIG_ASM && ARCH_X86
                default: {
                    return (a + b);
                }
            }
        }

        template<class T_a, class T_b> inline static T_a Sum(const T_a a, const T_b b) {
            return (a + b);
        }

        inline static float Sub(const float& a, const float& b) {
            switch (IMPL) {
                #if CONFIG_ASM && ARCH_X86
                case ASM_X86_MMX_SSE: {
                    float ret;
                    asm (
                        "movss    %1, %%xmm0  # load a\n\t"
                        "subss    %2, %%xmm0  # a - b\n\t"
                        "movss    %%xmm0, %0  # output\n\t"
                        : "=m" (ret)
                        : "m" (a), "m" (b)
                    );
                    return ret;
                }
                #endif // CONFIG_ASM && ARCH_X86
                default: {
                    return (a - b);
                }
            }
        }

        template<class T_a, class T_b> inline static T_a Sub(const T_a a, const T_b b) {
            return (a - b);
        }

        inline static float Mul(const float a, const float b) {
            switch (IMPL) {
                #if CONFIG_ASM && ARCH_X86
                case ASM_X86_MMX_SSE: {
                    float ret;
                    asm (
                        "movss    %1, %%xmm0  # load a\n\t"
                        "mulss    %2, %%xmm0  # a * b\n\t"
                        "movss    %%xmm0, %0  # output\n\t"
                        : "=m" (ret)
                        : "m" (a), "m" (b)
                    );
                    return ret;
                }
                #endif // CONFIG_ASM && ARCH_X86
                default: {
                    return (a * b);
                }
            }
        }

        template<class T_a, class T_b> inline static T_a Mul(const T_a a, const T_b b) {
            return (a * b);
        }

        inline static float Div(const float a, const float b) {
            switch (IMPL) {
                #if CONFIG_ASM && ARCH_X86
                case ASM_X86_MMX_SSE: {
                    float ret;
                    asm (
                        "movss    %1, %%xmm0  # load a\n\t"
                        "divss    %2, %%xmm0  # a / b\n\t"
                        "movss    %%xmm0, %0  # output\n\t"
                        : "=m" (ret)
                        : "m" (a), "m" (b)
                    );
                    return ret;
                }
                #endif // CONFIG_ASM && ARCH_X86
                default: {
                    return (a / b);
                }
            }
        }

        template<class T_a, class T_b> inline static T_a Div(const T_a a, const T_b b) {
            return (a / b);
        }

        inline static float Min(const float a, const float b) {
            switch (IMPL) {
                #if CONFIG_ASM && ARCH_X86
                case ASM_X86_MMX_SSE: {
                    float ret;
                    asm (
                        "movss    %1, %%xmm0  # load a\n\t"
                        "minss    %2, %%xmm0  # Minimum(a, b)\n\t"
                        "movss    %%xmm0, %0  # output\n\t"
                        : "=m" (ret)
                        : "m" (a), "m" (b)
                    );
                    return ret;
                }
                #endif // CONFIG_ASM && ARCH_X86
                default: {
                    return std::min(a, b);
                }
            }
        }

        template<class T_a, class T_b> inline static T_a Min(const T_a a, const T_b b) {
            return (b < a) ? b : a;
        }

        inline static float Max(const float a, const float b) {
            switch (IMPL) {
                #if CONFIG_ASM && ARCH_X86
                case ASM_X86_MMX_SSE: {
                    float ret;
                    asm (
                        "movss    %1, %%xmm0  # load a\n\t"
                        "maxss    %2, %%xmm0  # Maximum(a, b)\n\t"
                        "movss    %%xmm0, %0  # output\n\t"
                        : "=m" (ret)
                        : "m" (a), "m" (b)
                    );
                    return ret;
                }
                #endif // CONFIG_ASM && ARCH_X86
                default: {
                    return std::max(a, b);
                }
            }
        }

        template<class T_a, class T_b> inline static T_a Max(const T_a a, const T_b b) {
            return (b > a) ? b : a;
        }

        inline static float Fmodf(const float &a, const float &b) {
            switch (IMPL) {
                #if CONFIG_ASM && ARCH_X86
                case ASM_X86_MMX_SSE: {
                    float ret;
                    asm (
                        "movss    %1, %%xmm0  # load a\n\t"
                        "movss    %2, %%xmm1  # load b\n\t"
                        "movss    %%xmm0,%%xmm2\n\t"
                        "divss    %%xmm1, %%xmm2  # xmm2 = a / b\n\t"
                        "cvttss2si %%xmm2, %%ecx  #convert to int\n\t"
                        "cvtsi2ss %%ecx, %%xmm2  #convert back to float\n\t"
                        "mulss    %%xmm1, %%xmm2  # xmm2 = b * int(a/b)\n\t"
                        "subss    %%xmm2, %%xmm0  #sub a\n\t"
                        "movss    %%xmm0, %0  # output\n\t"
                        : "=m" (ret)
                        : "m" (a), "m" (b)
                        : "%ecx"
                    );
                    return ret;
                }
                #endif // CONFIG_ASM && ARCH_X86
                default: {
                    return fmodf(a, b);
                }
            }
        }
};

/// convenience typedef for using the default implementation (which is CPP)
typedef __RTMath<> RTMath;

#endif // __RT_MATH_H__
1	schoenebeck	53	/***************************************************************************
2			* *
3			* LinuxSampler - modular, streaming capable sampler *
4			* *
5	schoenebeck	56	* Copyright (C) 2003, 2004 by Benno Senoner and Christian Schoenebeck *
6	schoenebeck	2931	* Copyright (C) 2005 - 2016 Christian Schoenebeck *
7	schoenebeck	53	* *
8			* This program is free software; you can redistribute it and/or modify *
9			* it under the terms of the GNU General Public License as published by *
10			* the Free Software Foundation; either version 2 of the License, or *
11			* (at your option) any later version. *
12			* *
13			* This program is distributed in the hope that it will be useful, *
14			* but WITHOUT ANY WARRANTY; without even the implied warranty of *
15			* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the *
16			* GNU General Public License for more details. *
17			* *
18			* You should have received a copy of the GNU General Public License *
19			* along with this program; if not, write to the Free Software *
20			* Foundation, Inc., 59 Temple Place, Suite 330, Boston, *
21			* MA 02111-1307 USA *
22			***************************************************************************/
23
24			#ifndef __RT_MATH_H__
25			#define __RT_MATH_H__
26
27			#include <math.h>
28	schoenebeck	328	#include <stdint.h>
29	schoenebeck	1424	#include "global_private.h"
30	schoenebeck	53
31			/// Needed for calculating frequency ratio used to pitch a sample
32			#define TWELVEHUNDREDTH_ROOT_OF_TWO 1.000577789506555
33
34	schoenebeck	319	enum implementation_t {
35	schoenebeck	361	CPP,
36			ASM_X86_MMX_SSE
37	schoenebeck	319	};
38
39	schoenebeck	563	/** @brief Real Time Math Base Class
40			*
41			* Math functions for real time operation. This base class contains all
42			* non-template methods.
43			*/
44	schoenebeck	319	class RTMathBase {
45	schoenebeck	53	public:
46			/**
47	schoenebeck	328	* Highly accurate time stamp.
48			*/
49			typedef uint32_t time_stamp_t;
50
51			/**
52			* We read the processor's cycle count register as a reference
53			* for the real time. These are of course only abstract values
54			* with arbitrary time entity, but that's not a problem as long
55			* as we calculate relatively.
56			*/
57			static time_stamp_t CreateTimeStamp();
58
59			/**
60	schoenebeck	53	* Calculates the frequency ratio for a pitch value given in cents
61			* (assuming equal tempered scale of course, divided into 12
62			* semitones per octave and 100 cents per semitone).
63			*
64	schoenebeck	554	* Note: CONFIG_MAX_PITCH (defined in config.h) has to be defined to an
65	schoenebeck	53	* appropriate value, otherwise the behavior of this function is
66	schoenebeck	554	* undefined, but most probably if CONFIG_MAX_PITCH is too small, the
67	schoenebeck	53	* application will crash due to segmentation fault here.
68			*
69			* @param cents - pitch value in cents (+1200 cents means +1 octave)
70			* @returns frequency ratio (e.g. +2.0 for +1 octave)
71			*/
72			inline static double CentsToFreqRatio(double Cents) {
73	schoenebeck	319	int index_int = (int) (Cents); // integer index
74	schoenebeck	53	float index_fract = Cents - index_int; // fractional part of index
75			return pCentsToFreqTable[index_int] + index_fract * (pCentsToFreqTable[index_int+1] - pCentsToFreqTable[index_int]);
76			}
77
78	schoenebeck	829	/**
79	persson	1862	* Slower version of CentsToFreqRatio, for big values.
80			*
81			* @param cents - pitch value in cents (+1200 cents means +1 octave)
82			* @returns frequency ratio (e.g. +2.0 for +1 octave)
83			*/
84			static double CentsToFreqRatioUnlimited(double Cents) {
85			int octaves = int(Cents / 1200);
86			double x = CentsToFreqRatio(Cents - octaves * 1200);
87			return octaves < 0 ? x / (1 << -octaves) : x * (1 << octaves);
88			}
89
90			/**
91	schoenebeck	829	* Inverse function to CentsToFreqRatio(). This function is a bit
92			* slow, so it should not be called too frequently.
93			*/
94			static double FreqRatioToCents(double FreqRatio) {
95			return log(FreqRatio) / log(TWELVEHUNDREDTH_ROOT_OF_TWO);
96			}
97
98	schoenebeck	2931	/**
99			* Calculates the line ratio value representation (linear scale)
100			* of the @a decibel value provided (exponential scale).
101			*
102			* The context of audio acoustic sound pressure levels is assumed, and
103			* hence the field version of the dB unit is used here (which uses a
104			* linear factor of 20). This function is a bit slow, so it should
105			* not be called too frequently.
106			*
107			* @param decibel - sound pressure level in dB
108			* @returns linear ratio of the supplied dB value
109			*/
110			static float DecibelToLinRatio(float decibel) {
111			return powf(10.f, decibel / 20.f);
112			}
113
114			/**
115			* Calculates the relatively summed average of a set of values.
116			*
117			* @param current - the current avaerage value of all previously summed values
118			* @param sample - new value to be applied as summed average to the existing values
119			* @param n - amount of sample values applied so far
120			* @returns new average value of all summed values (including the new @a sample)
121			*/
122			inline static float RelativeSummedAvg(float current, float sample, int n) {
123			return current + (sample - current) / float(n);
124			}
125
126	schoenebeck	319	private:
127			static float* pCentsToFreqTable;
128
129			static float* InitCentsToFreqTable();
130			};
131
132	schoenebeck	563	/** @brief Real Time Math
133	schoenebeck	319	*
134	schoenebeck	563	* This is a template which provides customized methods for the desired low
135			* level implementation. The ASM_X86_MMX_SSE implementation of each method
136			* for example doesn't use 387 FPU instruction. This is needed for MMX
137			* algorithms which do not allow mixed MMX and 387 instructions.
138	schoenebeck	319	*/
139			template<implementation_t IMPL = CPP>
140			class __RTMath : public RTMathBase {
141			public:
142			// conversion using truncate
143			inline static int Int(const float a) {
144			switch (IMPL) {
145	schoenebeck	617	#if CONFIG_ASM && ARCH_X86
146	schoenebeck	319	case ASM_X86_MMX_SSE: {
147			int ret;
148			asm (
149			"cvttss2si %1, %0 # convert to int\n\t"
150			: "=r" (ret)
151			: "m" (a)
152			);
153			return ret;
154			}
155	schoenebeck	617	#endif // CONFIG_ASM && ARCH_X86
156	persson	685	default: {
157			return (int) a;
158			}
159	schoenebeck	319	}
160			}
161
162			//for doubles and everything else except floats
163			template<class T_a> inline static int Int(const T_a a) {
164			return (int) a;
165			}
166
167			inline static float Float(const int a) {
168			switch (IMPL) {
169	schoenebeck	617	#if CONFIG_ASM && ARCH_X86
170	schoenebeck	319	case ASM_X86_MMX_SSE: {
171			float ret;
172			asm (
173			"cvtsi2ss %1, %%xmm0 # convert to float\n\t"
174			"movss %%xmm0,%0 # output\n\t"
175			: "=m" (ret)
176			: "r" (a)
177			);
178			return ret;
179			}
180	schoenebeck	617	#endif // CONFIG_ASM && ARCH_X86
181	persson	685	default: {
182			return (float) a;
183			}
184	schoenebeck	319	}
185			}
186
187			#if 0
188			//for everything except ints
189			template<class T_a> inline static float Float(T_a a) {
190			return (float) a;
191			}
192			#endif
193
194			inline static float Sum(const float& a, const float& b) {
195			switch (IMPL) {
196	schoenebeck	617	#if CONFIG_ASM && ARCH_X86
197	schoenebeck	319	case ASM_X86_MMX_SSE: {
198			float ret;
199			asm (
200			"movss %1, %%xmm0 # load a\n\t"
201			"addss %2, %%xmm0 # a + b\n\t"
202			"movss %%xmm0, %0 # output\n\t"
203			: "=m" (ret)
204			: "m" (a), "m" (b)
205			);
206			return ret;
207			}
208	schoenebeck	617	#endif // CONFIG_ASM && ARCH_X86
209	persson	685	default: {
210			return (a + b);
211			}
212	schoenebeck	319	}
213			}
214
215			template<class T_a, class T_b> inline static T_a Sum(const T_a a, const T_b b) {
216			return (a + b);
217			}
218
219			inline static float Sub(const float& a, const float& b) {
220			switch (IMPL) {
221	schoenebeck	617	#if CONFIG_ASM && ARCH_X86
222	schoenebeck	319	case ASM_X86_MMX_SSE: {
223			float ret;
224			asm (
225			"movss %1, %%xmm0 # load a\n\t"
226			"subss %2, %%xmm0 # a - b\n\t"
227			"movss %%xmm0, %0 # output\n\t"
228			: "=m" (ret)
229			: "m" (a), "m" (b)
230			);
231			return ret;
232			}
233	schoenebeck	617	#endif // CONFIG_ASM && ARCH_X86
234	persson	685	default: {
235			return (a - b);
236			}
237	schoenebeck	319	}
238			}
239
240			template<class T_a, class T_b> inline static T_a Sub(const T_a a, const T_b b) {
241			return (a - b);
242			}
243
244			inline static float Mul(const float a, const float b) {
245			switch (IMPL) {
246	schoenebeck	617	#if CONFIG_ASM && ARCH_X86
247	schoenebeck	319	case ASM_X86_MMX_SSE: {
248			float ret;
249			asm (
250			"movss %1, %%xmm0 # load a\n\t"
251			"mulss %2, %%xmm0 # a * b\n\t"
252			"movss %%xmm0, %0 # output\n\t"
253			: "=m" (ret)
254			: "m" (a), "m" (b)
255			);
256			return ret;
257			}
258	schoenebeck	617	#endif // CONFIG_ASM && ARCH_X86
259	persson	685	default: {
260			return (a * b);
261			}
262	schoenebeck	319	}
263			}
264
265			template<class T_a, class T_b> inline static T_a Mul(const T_a a, const T_b b) {
266			return (a * b);
267			}
268
269			inline static float Div(const float a, const float b) {
270			switch (IMPL) {
271	schoenebeck	617	#if CONFIG_ASM && ARCH_X86
272	schoenebeck	319	case ASM_X86_MMX_SSE: {
273			float ret;
274			asm (
275			"movss %1, %%xmm0 # load a\n\t"
276			"divss %2, %%xmm0 # a / b\n\t"
277			"movss %%xmm0, %0 # output\n\t"
278			: "=m" (ret)
279			: "m" (a), "m" (b)
280			);
281			return ret;
282			}
283	schoenebeck	617	#endif // CONFIG_ASM && ARCH_X86
284	persson	685	default: {
285			return (a / b);
286			}
287	schoenebeck	319	}
288			}
289
290			template<class T_a, class T_b> inline static T_a Div(const T_a a, const T_b b) {
291			return (a / b);
292			}
293
294			inline static float Min(const float a, const float b) {
295			switch (IMPL) {
296	schoenebeck	617	#if CONFIG_ASM && ARCH_X86
297	schoenebeck	319	case ASM_X86_MMX_SSE: {
298			float ret;
299			asm (
300			"movss %1, %%xmm0 # load a\n\t"
301			"minss %2, %%xmm0 # Minimum(a, b)\n\t"
302			"movss %%xmm0, %0 # output\n\t"
303			: "=m" (ret)
304			: "m" (a), "m" (b)
305			);
306			return ret;
307			}
308	schoenebeck	617	#endif // CONFIG_ASM && ARCH_X86
309	persson	685	default: {
310	wylder	818	return std::min(a, b);
311	persson	685	}
312	schoenebeck	319	}
313			}
314
315			template<class T_a, class T_b> inline static T_a Min(const T_a a, const T_b b) {
316	schoenebeck	53	return (b < a) ? b : a;
317			}
318
319	schoenebeck	319	inline static float Max(const float a, const float b) {
320			switch (IMPL) {
321	schoenebeck	617	#if CONFIG_ASM && ARCH_X86
322	schoenebeck	319	case ASM_X86_MMX_SSE: {
323			float ret;
324			asm (
325			"movss %1, %%xmm0 # load a\n\t"
326			"maxss %2, %%xmm0 # Maximum(a, b)\n\t"
327			"movss %%xmm0, %0 # output\n\t"
328			: "=m" (ret)
329			: "m" (a), "m" (b)
330			);
331			return ret;
332			}
333	schoenebeck	617	#endif // CONFIG_ASM && ARCH_X86
334	persson	685	default: {
335	wylder	818	return std::max(a, b);
336	persson	685	}
337	schoenebeck	319	}
338			}
339
340			template<class T_a, class T_b> inline static T_a Max(const T_a a, const T_b b) {
341	schoenebeck	53	return (b > a) ? b : a;
342			}
343
344	schoenebeck	319	inline static float Fmodf(const float &a, const float &b) {
345			switch (IMPL) {
346	schoenebeck	617	#if CONFIG_ASM && ARCH_X86
347	schoenebeck	319	case ASM_X86_MMX_SSE: {
348			float ret;
349			asm (
350			"movss %1, %%xmm0 # load a\n\t"
351			"movss %2, %%xmm1 # load b\n\t"
352			"movss %%xmm0,%%xmm2\n\t"
353			"divss %%xmm1, %%xmm2 # xmm2 = a / b\n\t"
354			"cvttss2si %%xmm2, %%ecx #convert to int\n\t"
355			"cvtsi2ss %%ecx, %%xmm2 #convert back to float\n\t"
356			"mulss %%xmm1, %%xmm2 # xmm2 = b * int(a/b)\n\t"
357			"subss %%xmm2, %%xmm0 #sub a\n\t"
358			"movss %%xmm0, %0 # output\n\t"
359			: "=m" (ret)
360			: "m" (a), "m" (b)
361			: "%ecx"
362			);
363			return ret;
364			}
365	schoenebeck	617	#endif // CONFIG_ASM && ARCH_X86
366	persson	685	default: {
367			return fmodf(a, b);
368			}
369	schoenebeck	319	}
370			}
371	schoenebeck	53	};
372
373	schoenebeck	319	/// convenience typedef for using the default implementation (which is CPP)
374			typedef __RTMath<> RTMath;
375
376	schoenebeck	53	#endif // __RT_MATH_H__