Runtime code dispatch support

Let CineForm SDK choose available SIMD optimization if processor has such extensions.

added Doing label

assigned to @enpinion

changed title from Runtime optimization support to Runtime code dispatch support

changed the description

Processor Dispatch

There are two ways to achieve this:

1. Use "thunks".
2. Let encoder or decoder decide it.

Common things:

1. Encoder and decoder object (or struct, context...) will detect whether processor has SIMD extensions on its initialization.

Thunks method

1. Register all function tables.
2. Connect to corresponding function with function pointer. e.g.

// Register functions
void init_function(void* table, int extension)
{
  void* wavelet = table;
  switch(extension)
  {
// ...
    case BUILD_WITH_SIMD_X86_AVX2:
    // wavelet has function tables of wavelet.c/h
    // void (*TransformForwardSpatialYUV)(...)
    wavelet->TransformForwardSpatialYUV = &TransformForwardSpatialYUV_AVX2;
    break;
// ... and so on
  }
}

//

{
  // ...
  init_function(table->wavelet, CFHD_PROCESSOR_AVX2);
  table->wavelet->TransformForwardSpatialYUV(...);;;
}

Pros: Compare only once

Cons: Use (small but) more memory usage.

encoder/decoder dispatch method:

1. switch->case e.g.

switch (encoder->deviceExtension)
{
#  ifdef BUILD_WITH_SIMD_X86_MMX
        case CFHD_PROCESSOR_MMX:
          TransformForwardSpatialYUV_mmx();
          break;
#  endif // BUILD_WITH_SIMD_X86_MMX
#  ifdef BUILD_WITH_SIMD_X86_SSE2
        case CFHD_PROCESSOR_SSE2:
          TransformForwardSpatialYUV_sse2();
          break;
#  endif // BUILD_WITH_SIMD_X86_SSE2
#  ifdef BUILD_WITH_SIMD_X86_AVX2
        case CFHD_PROCESSOR_AVX2:
          TransformForwardSpatialYUV_avx2();
          break;
#  endif // BUILD_WITH_SIMD_X86_AVX2
#  ifdef BUILD_WITH_SIMD_ARM_NEON
        case CFHD_PROCESSOR_NEON:
          TransformForwardSpatialYUV_neon();
          break;
#  endif // BUILD_WITH_SIMD_ARM_NEON
#  ifdef BUILD_WITH_SIMD_PPC_ALTIVEC
        case CFHD_PROCESSOR_ALTIVEC:
          TransformForwardSpatialYUV_altivec();
          break;
#  endif // BUILD_WITH_SIMD_PPC_ALTIVEC
        default: // Generic scalar operation
          TransformForwardSpatialYUV();
}

Pros: Small code changes.

Cons: Loooong codes. Code may runs slower than "thunks"

Benchmark

57.5fps switch->case method.

207.6fps function table method.

208.0fps direct call method. (No dispatch)

Conclusion

switch->case method is 3-4 times slower. Use function tables.

Things to do

Add table pointer to all related functions

So the other sub functions can track a proper code path such as when SSE2 code calls SSE4 codes when SSE4 flag is set.

Such as:

void func1_avx(void *function_table, int *param, int *param2)
{
  // Do something
  function_table->subfunction(function_table, param, param2);
}