diff options
Diffstat (limited to 'mesalib/src/mesa/tnl/t_pipeline.c')
| -rw-r--r-- | mesalib/src/mesa/tnl/t_pipeline.c | 78 |
1 files changed, 78 insertions, 0 deletions
diff --git a/mesalib/src/mesa/tnl/t_pipeline.c b/mesalib/src/mesa/tnl/t_pipeline.c index 78508aecf..0adbef012 100644 --- a/mesalib/src/mesa/tnl/t_pipeline.c +++ b/mesalib/src/mesa/tnl/t_pipeline.c @@ -112,6 +112,84 @@ static GLuint check_output_changes( struct gl_context *ctx ) #endif } +/** + * START/END_FAST_MATH macros: + * + * START_FAST_MATH: Set x86 FPU to faster, 32-bit precision mode (and save + * original mode to a temporary). + * END_FAST_MATH: Restore x86 FPU to original mode. + */ +#if defined(__GNUC__) && defined(__i386__) +/* + * Set the x86 FPU control word to guarentee only 32 bits of precision + * are stored in registers. Allowing the FPU to store more introduces + * differences between situations where numbers are pulled out of memory + * vs. situations where the compiler is able to optimize register usage. + * + * In the worst case, we force the compiler to use a memory access to + * truncate the float, by specifying the 'volatile' keyword. + */ +/* Hardware default: All exceptions masked, extended double precision, + * round to nearest (IEEE compliant): + */ +#define DEFAULT_X86_FPU 0x037f +/* All exceptions masked, single precision, round to nearest: + */ +#define FAST_X86_FPU 0x003f +/* The fldcw instruction will cause any pending FP exceptions to be + * raised prior to entering the block, and we clear any pending + * exceptions before exiting the block. Hence, asm code has free + * reign over the FPU while in the fast math block. + */ +#if defined(NO_FAST_MATH) +#define START_FAST_MATH(x) \ +do { \ + static GLuint mask = DEFAULT_X86_FPU; \ + __asm__ ( "fnstcw %0" : "=m" (*&(x)) ); \ + __asm__ ( "fldcw %0" : : "m" (mask) ); \ +} while (0) +#else +#define START_FAST_MATH(x) \ +do { \ + static GLuint mask = FAST_X86_FPU; \ + __asm__ ( "fnstcw %0" : "=m" (*&(x)) ); \ + __asm__ ( "fldcw %0" : : "m" (mask) ); \ +} while (0) +#endif +/* Restore original FPU mode, and clear any exceptions that may have + * occurred in the FAST_MATH block. + */ +#define END_FAST_MATH(x) \ +do { \ + __asm__ ( "fnclex ; fldcw %0" : : "m" (*&(x)) ); \ +} while (0) + +#elif defined(_MSC_VER) && defined(_M_IX86) +#define DEFAULT_X86_FPU 0x037f /* See GCC comments above */ +#define FAST_X86_FPU 0x003f /* See GCC comments above */ +#if defined(NO_FAST_MATH) +#define START_FAST_MATH(x) do {\ + static GLuint mask = DEFAULT_X86_FPU;\ + __asm fnstcw word ptr [x]\ + __asm fldcw word ptr [mask]\ +} while(0) +#else +#define START_FAST_MATH(x) do {\ + static GLuint mask = FAST_X86_FPU;\ + __asm fnstcw word ptr [x]\ + __asm fldcw word ptr [mask]\ +} while(0) +#endif +#define END_FAST_MATH(x) do {\ + __asm fnclex\ + __asm fldcw word ptr [x]\ +} while(0) + +#else +#define START_FAST_MATH(x) x = 0 +#define END_FAST_MATH(x) (void)(x) +#endif + void _tnl_run_pipeline( struct gl_context *ctx ) { |