Sunday, 22 December 2013

Building FFmpeg on Windows with in-line asm and the Intel compiler.

Building FFmpeg for Windows is generally done using MinGW compiler in either a MSYS or Cygwin environment. But the resulting libraries don't always play all that well when you try and use them in a native Windows program (i.e. something built with the Microsoft compiler -msvc). For those who work on Windows it is often slightly more convenient to be able to use a native build chain when creating FFmpeg binaries.

Unfortunately FFmpeg (much like most opensource projects) uses C99 code which will not build using the msvc compiler found in Visual Studio. This is because Microsoft in all there wisdom don't support anything past C89. With the standardisation of C++11 newer versions of msvc are slowly starting to support C99 features as the C++11 specification mandates that they must be supported. So supporting C++11 generally also means supporting C99. However this support started to feature in Visual Studio 2013 and it is still far from complete and can not compile many C99 based projects (at time of writing).

Things are made worse as FFmpeg doesnt just use C99 but also has large amounts of in-line assembly. This assembly is written in the AT&T assembly syntax (or more specifically GAS - GNU Assembler) as opposed to Intel assembly syntax. Since msvc only supports Intel syntax then even with C99 support it will never be possible to compile FFmpeg with in-line assembly natively. Of-course FFmpeg has a nice config option that can be used to disable all in-line asm which would mean that it would compile on a C99 compliant compiler without needing AT&T assembly support. But then we don't get all the nicely hand tuned assembly optimisations that were written specifically to maximise performance.

Luckily there is the Intel compiler to the rescue. The Intel compiler for Windows will generate a native windows library that can be easily linked and debugged like any other msvc library. It also has the added benefit of supporting AT&T assembly compilation on Windows.

Unfortunately Intels AT&T support is not fully featured on Windows. On Linux the Intel compiler just uses GAS so it supports everything GAS does. But on Windows Intel had to implement the support themselves and as a result it is not as feature complete as GAS. The main issue is that the Intel compiler on Windows does not support 'Direct symbol references' which are used in various places within FFmpeg. In order to compile on Windows these lines need to be updated in order to remove the symbol references.

An example of the problem and the fix can be found in libavcodec/x86/dsputil_mmx.c:

--- a/libavcodec/x86/dsputil_mmx.c
+++ b/libavcodec/x86/dsputil_mmx.c
@@ -115,13 +115,13 @@ void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
     x86_reg line_skip3;
     __asm__ volatile (
-        "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
+        "movq %4, %%mm0                     \n\t"
         "lea         (%3, %3, 2), %1        \n\t"
         "lea         (%0, %3, 4), %0        \n\t"
         : "+&r"(pixels), "=&r"(line_skip3)
-        : "r"(block), "r"(line_skip)
+        : "r"(block), "r"(line_skip), "m"(ff_pb_80)
         : "memory");

Here the symbol reference is replaced with an asm-interface.

There are many occurrences of direct symbol references in FFmpeg (to many to list here) so I will instead direct people to check out the attached patch.

The attached patch includes updates to the in-line asm that allow most of it to compile under the Intel compiler. Unfortunately there are 2 places where the asm uses slightly more complicated features than just symbol references so these cant be easily converted. These instances are in libavcodec/x86/mlpdsp.c where the use of labels (combined with a jump list) is not supported by Intel on Windows and is not easily modified. The second instance is in the BRANCHLESS_GET_CABAC macro in libavcodec/x86/cabac.h. This contains slightly more complex use of symbol references that if just swapped out like previous occurrences still doesn't compile without some additional work.

However despite a couple of problem functions the attached patch will allow for FFmpeg compilation on Windows with the Intel compiler and in-line asm optimisations enabled.

Update 2: New and improved patches have been created and these are now available directly in FFmpeg master (no patching required). See my post for more information (Building FFmpeg on Windows with in-line asm and the Intel compiler (Part 3)).
Update: A new patch is available in my newer post (Building FFmpeg on Windows with in-line asm and the Intel compiler (Part 2)).

Download patch file:

Lastly for those who don't want to apply the modifications themselves and would like a nicer build method than using MSYS/Cygwin then you can check out my git repository which has working Visual Studio build projects. These projects are in the 'SMP' directory and of course require a valid install of the Intel Windows compiler.

No comments:

Post a comment