From bc4186261da6e4eb53692fe7ac66c6b90776c7ab Mon Sep 17 00:00:00 2001 From: Bernd Kuhls Date: Wed, 23 Jul 2014 06:40:19 +0200 Subject: [PATCH] xbmc: Allow compilation with uClibc - Add dependencies needed by xbmc on BR2_LARGEFILE, BR2_INET_IPV6 & BR2_USE_WCHAR after the removal of BR2_TOOLCHAIN_USES_GLIBC - Add xbmc-0002-mathutil.patch to fix ARM compilation [Thomas: factorize architecture dependency in a new BR2_PACKAGE_XBMC_ARCH_SUPPORTS option.] Signed-off-by: Bernd Kuhls Signed-off-by: Thomas Petazzoni --- package/xbmc/Config.in | 19 ++- package/xbmc/xbmc-0002-mathutil.patch | 213 ++++++++++++++++++++++++++ 2 files changed, 226 insertions(+), 6 deletions(-) create mode 100644 package/xbmc/xbmc-0002-mathutil.patch diff --git a/package/xbmc/Config.in b/package/xbmc/Config.in index 592fdc1593..05d28ee236 100644 --- a/package/xbmc/Config.in +++ b/package/xbmc/Config.in @@ -1,11 +1,14 @@ -comment "xbmc needs an (e)glibc toolchain w/ C++, threads" - depends on !BR2_TOOLCHAIN_USES_GLIBC || !BR2_INSTALL_LIBSTDCPP || !BR2_TOOLCHAIN_HAS_THREADS +config BR2_PACKAGE_XBMC_ARCH_SUPPORTS + bool + default y if BR2_arm || BR2_i386 || BR2_x86_64 + +comment "xbmc needs a toolchain w/ C++, IPv6, largefile, threads, wchar" + depends on BR2_PACKAGE_XBMC_ARCH_SUPPORTS + depends on !BR2_INET_IPV6 || !BR2_INSTALL_LIBSTDCPP || !BR2_LARGEFILE || !BR2_TOOLCHAIN_HAS_THREADS || !BR2_USE_WCHAR comment "xbmc requires an OpenGL ES and EGL backend" + depends on BR2_PACKAGE_XBMC_ARCH_SUPPORTS depends on !BR2_PACKAGE_HAS_LIBEGL || !BR2_PACKAGE_HAS_LIBGLES - depends on BR2_TOOLCHAIN_USES_GLIBC - depends on BR2_INSTALL_LIBSTDCPP - depends on BR2_TOOLCHAIN_HAS_THREADS menuconfig BR2_PACKAGE_XBMC bool "xbmc" @@ -58,10 +61,14 @@ menuconfig BR2_PACKAGE_XBMC select BR2_PACKAGE_TINYXML select BR2_PACKAGE_YAJL select BR2_PACKAGE_ZLIB - depends on BR2_TOOLCHAIN_USES_GLIBC + depends on BR2_INET_IPV6 depends on BR2_INSTALL_LIBSTDCPP + depends on BR2_LARGEFILE depends on BR2_TOOLCHAIN_HAS_THREADS depends on BR2_PACKAGE_HAS_LIBEGL && BR2_PACKAGE_HAS_LIBGLES + depends on BR2_USE_MMU # python + depends on BR2_USE_WCHAR + depends on BR2_PACKAGE_XBMC_ARCH_SUPPORTS help XBMC is an award-winning free and open source (GPL) software media player and entertainment hub for digital media. diff --git a/package/xbmc/xbmc-0002-mathutil.patch b/package/xbmc/xbmc-0002-mathutil.patch new file mode 100644 index 0000000000..33f91eb38a --- /dev/null +++ b/package/xbmc/xbmc-0002-mathutil.patch @@ -0,0 +1,213 @@ +Taken from upstream PR: https://github.com/xbmc/xbmc/pull/3760 + +Signed-off-by: Bernd Kuhls + + +From 7388e8be7cd5e78100532ebf0dba15dccb7b03f8 Mon Sep 17 00:00:00 2001 +From: Ben Avison +Date: Tue, 3 Dec 2013 15:51:39 +0000 +Subject: [PATCH] Faster and simpler portable implementation of + MathUtils::round_int(). + +Much as I like a bit of inline assembler, I have also removed the ARM versions +of MathUtils::truncate_int() and MathUtils::round_int(). The former was just +how any sane compiler should have assembled a cast from double to signed int +anyway. The latter was a much too complicated way to achieve the desired +effect, and was switched out in most ARM builds anyway in favour of the old +portable implementation that used floor(). + +Verified that MathUtils::test() still passes, and that GCC is now able to +inline MathUtils::round_int(), where it didn't previously. + +I tested on a Raspberry Pi with the default theme, displaying the front page +with the RSS ticker enabled. This saturates the CPU, so I'm measuring the +improvement using the debug window's FPS figure. This patch improves this from +~50.8 FPS to ~52.6 FPS. +--- + xbmc/utils/MathUtils.h | 129 +++++++++++++++++++++++-------------------------- + 1 file changed, 61 insertions(+), 68 deletions(-) + +diff --git a/xbmc/utils/MathUtils.h b/xbmc/utils/MathUtils.h +index 96af9f4..0dae77d 100644 +--- a/xbmc/utils/MathUtils.h ++++ b/xbmc/utils/MathUtils.h +@@ -34,17 +34,13 @@ + + #if defined(__ppc__) || \ + defined(__powerpc__) || \ +- (defined(TARGET_DARWIN_IOS) && defined(__llvm__)) || \ +- (defined(TARGET_ANDROID) && defined(__arm__)) || \ +- defined(TARGET_RASPBERRY_PI) ++ defined(__arm__) + #define DISABLE_MATHUTILS_ASM_ROUND_INT + #endif + + #if defined(__ppc__) || \ + defined(__powerpc__) || \ +- (defined(TARGET_DARWIN) && defined(__llvm__)) || \ +- (defined(TARGET_ANDROID) && defined(__arm__)) || \ +- defined(TARGET_RASPBERRY_PI) ++ defined(__arm__) + #define DISABLE_MATHUTILS_ASM_TRUNCATE_INT + #endif + +@@ -73,60 +69,63 @@ + { + assert(x > static_cast(INT_MIN / 2) - 1.0); + assert(x < static_cast(INT_MAX / 2) + 1.0); +- const float round_to_nearest = 0.5f; +- int i; + + #if defined(DISABLE_MATHUTILS_ASM_ROUND_INT) +- i = floor(x + round_to_nearest); +- +-#elif defined(__arm__) +- // From 'ARM-v7-M Architecture Reference Manual' page A7-569: +- // "The floating-point to integer operation (vcvt) [normally] uses the Round towards Zero rounding mode" +- // Because of this...we must use some less-than-straightforward logic to perform this operation without +- // changing the rounding mode flags +- +- /* The assembly below implements the following logic: +- if (x < 0) +- inc = -0.5f +- else +- inc = 0.5f +- int_val = trunc(x+inc); +- err = x - int_val; +- if (err == 0.5f) +- int_val++; +- return int_val; +- */ ++ /* This implementation warrants some further explanation. ++ * ++ * First, a couple of notes on rounding: ++ * 1) C casts from float/double to integer round towards zero. ++ * 2) Float/double additions are rounded according to the normal rules, ++ * in other words: on some architectures, it's fixed at compile-time, ++ * and on others it can be set using fesetround()). The following ++ * analysis assumes round-to-nearest with ties rounding to even. This ++ * is a fairly sensible choice, and is the default with ARM VFP. ++ * ++ * What this function wants is round-to-nearest with ties rounding to ++ * +infinity. This isn't an IEEE rounding mode, even if we could guarantee ++ * that all architectures supported fesetround(), which they don't. Instead, ++ * this adds an offset of 2147483648.5 (= 0x80000000.8p0), then casts to ++ * an unsigned int (crucially, all possible inputs are now in a range where ++ * round to zero acts the same as round to -infinity) and then subtracts ++ * 0x80000000 in the integer domain. The 0.5 component of the offset ++ * converts what is effectively a round down into a round to nearest, with ++ * ties rounding up, as desired. ++ * ++ * There is a catch, that because there is a double rounding, there is a ++ * small region where the input falls just *below* a tie, where the addition ++ * of the offset causes a round *up* to an exact integer, due to the finite ++ * level of precision available in floating point. You need to be aware of ++ * this when calling this function, although at present it is not believed ++ * that XBMC ever attempts to round numbers in this window. ++ * ++ * It is worth proving the size of the affected window. Recall that double ++ * precision employs a mantissa of 52 bits. ++ * 1) For all inputs -0.5 <= x <= INT_MAX ++ * Once the offset is applied, the most significant binary digit in the ++ * floating-point representation is +2^31. ++ * At this magnitude, the smallest step representable in double precision ++ * is 2^31 / 2^52 = 0.000000476837158203125 ++ * So the size of the range which is rounded up due to the addition is ++ * half the size of this step, or 0.0000002384185791015625 ++ * ++ * 2) For all inputs INT_MIN/2 < x < -0.5 ++ * Once the offset is applied, the most significant binary digit in the ++ * floating-point representation is +2^30. ++ * At this magnitude, the smallest step representable in double precision ++ * is 2^30 / 2^52 = 0.0000002384185791015625 ++ * So the size of the range which is rounded up due to the addition is ++ * half the size of this step, or 0.00000011920928955078125 ++ * ++ * 3) For all inputs INT_MIN <= x <= INT_MIN/2 ++ * The representation once the offset is applied has equal or greater ++ * precision than the input, so the addition does not cause rounding. ++ */ ++ return ((unsigned int) (x + 0x80000000.8p0)) - 0x80000000; + +- __asm__ __volatile__ ( +-#if defined(__ARM_PCS_VFP) +- "fconstd d1,#%G[rnd_val] \n\t" // Copy round_to_nearest into a working register (d1 = 0.5) + #else +- "vmov.F64 d1,%[rnd_val] \n\t" +-#endif +- "fcmpezd %P[value] \n\t" // Check value against zero (value == 0?) +- "fmstat \n\t" // Copy the floating-point status flags into the general-purpose status flags +- "it mi \n\t" +- "vnegmi.F64 d1, d1 \n\t" // if N-flag is set, negate round_to_nearest (if (value < 0) d1 = -1 * d1) +- "vadd.F64 d1,%P[value],d1 \n\t" // Add round_to_nearest to value, store result in working register (d1 += value) +- "vcvt.S32.F64 s3,d1 \n\t" // Truncate(round towards zero) (s3 = (int)d1) +- "vmov %[result],s3 \n\t" // Store the integer result in a general-purpose register (result = s3) +- "vcvt.F64.S32 d1,s3 \n\t" // Convert back to floating-point (d1 = (double)s3) +- "vsub.F64 d1,%P[value],d1 \n\t" // Calculate the error (d1 = value - d1) +-#if defined(__ARM_PCS_VFP) +- "fconstd d2,#%G[rnd_val] \n\t" // d2 = 0.5; +-#else +- "vmov.F64 d2,%[rnd_val] \n\t" +-#endif +- "fcmped d1, d2 \n\t" // (d1 == 0.5?) +- "fmstat \n\t" // Copy the floating-point status flags into the general-purpose status flags +- "it eq \n\t" +- "addeq %[result],#1 \n\t" // (if (d1 == d2) result++;) +- : [result] "=r"(i) // Outputs +- : [rnd_val] "Dv" (round_to_nearest), [value] "w"(x) // Inputs +- : "d1", "d2", "s3" // Clobbers +- ); +- +-#elif defined(__SSE2__) ++ const float round_to_nearest = 0.5f; ++ int i; ++#if defined(__SSE2__) + const float round_dn_to_nearest = 0.4999999f; + i = (x > 0) ? _mm_cvttsd_si32(_mm_set_sd(x + round_to_nearest)) : _mm_cvttsd_si32(_mm_set_sd(x - round_dn_to_nearest)); + +@@ -150,8 +149,8 @@ + ); + + #endif +- + return i; ++#endif + } + + /*! \brief Truncate to nearest integer. +@@ -165,20 +164,13 @@ + { + assert(x > static_cast(INT_MIN / 2) - 1.0); + assert(x < static_cast(INT_MAX / 2) + 1.0); +- int i; + + #if defined(DISABLE_MATHUTILS_ASM_TRUNCATE_INT) +- return i = (int)x; +- +-#elif defined(__arm__) +- __asm__ __volatile__ ( +- "vcvt.S32.F64 %[result],%P[value] \n\t" // Truncate(round towards zero) and store the result +- : [result] "=w"(i) // Outputs +- : [value] "w"(x) // Inputs +- ); +- return i; ++ return x; + +-#elif defined(TARGET_WINDOWS) ++#else ++ int i; ++#if defined(TARGET_WINDOWS) + const float round_towards_m_i = -0.5f; + __asm + { +@@ -204,6 +196,7 @@ + if (x < 0) + i = -i; + return (i); ++#endif + } + + inline int64_t abs(int64_t a) +-- +1.9.1 +