TransformVertices_ref.cpp

#include <iostream>
#include "Utilities/TransformVertices.h"
#include "Context.h"
#include "Services/Features.h"
 
//#ifdef COGS_ARCH_X86
//#include "TransformVerticesSSE.h"
//#ifdef COGS_EXTENSIONS_AVX
//#include "../../Extensions/UtilitiesAVX/TransformVerticesAVX.h"
//#endif
//#endif
 
#ifndef EMSCRIPTEN
namespace Cogs::Core
{
 
  void transformVertex3ToVertex4AVX(uint8_t* dst,
                                    const size_t dst_stride,
                                    const size_t dst_bytes,
                                    const glm::mat4& matrix,
                                    const uint8_t* src,
                                    const size_t src_stride,
                                    const size_t src_bytes,
                                    const size_t src_count);
 
  void transformVertex3ToVertex4FastPathAVX(uint8_t* dst,
                                            const size_t dst_stride,
                                            const size_t dst_bytes,
                                            const glm::mat4& matrix,
                                            const uint8_t* src,
                                            const size_t src_stride,
                                            const size_t src_bytes,
                                            const size_t src_count);
 
  void transformVertex3ToVertex4SSE1(uint8_t* dst, const size_t dst_stride, const size_t dst_bytes,
                                     const glm::mat4& matrix,
                                     const uint8_t* src,
                                     const size_t src_stride,
                                     const size_t src_bytes,
                                     const size_t src_count);
 
 
  void transformVertex3ToVertex4FastPathSSE1(uint8_t* dst, const size_t dst_stride, const size_t dst_bytes,
                                             const glm::mat4& matrix,
                                             const uint8_t* src,
                                             const size_t src_stride,
                                             const size_t src_bytes,
                                             const size_t src_count);
 
}
#endif
 
 
void Cogs::Core::transformVertex3ToVertex4(Context* context,
                                           uint8_t* dst,
                                           const size_t dst_stride,
                                           const size_t dst_bytes,
                                           const glm::mat4& matrix,
                                           const uint8_t* src,
                                           const size_t src_stride,
                                           const size_t src_bytes,
                                           const size_t src_count)
{
  size_t i = 0;
 
#if !defined(EMSCRIPTEN) && !defined(__APPLE__)
 
  if (context->features->supported(CPUFeature::AVX)) {
    if ((((size_t)dst & 0x1f) == 0) && (dst_stride == 16) && (((size_t)src & 0xf) == 0) && ((src_stride & 0xf) == 0)) {
      transformVertex3ToVertex4FastPathAVX(dst, dst_stride, dst_bytes, matrix, src, src_stride, src_bytes, src_count);
    }
    else {
      transformVertex3ToVertex4AVX(dst, dst_stride, dst_bytes, matrix, src, src_stride, src_bytes, src_count);
    }
    return;
  }
 
  if (context->features->supported(CPUFeature::SSE)) {
    if ((((size_t)dst & 0xf) == 0) && ((dst_stride & 0xf) == 0) && (((size_t)src & 0xf) == 0) && ((src_stride & 0xf) == 0)) {
      transformVertex3ToVertex4FastPathSSE1(dst, dst_stride, dst_bytes, matrix, src, src_stride, src_bytes, src_count);
    }
    else {
      transformVertex3ToVertex4SSE1(dst, dst_stride, dst_bytes, matrix, src, src_stride, src_bytes, src_count);
    }
    return; // No remainder
  }
#else
  (void)context;
  (void)src_bytes;
  (void)dst_bytes;
#endif
 
  for (; i < src_count; i++) {
    glm::vec4& d = *reinterpret_cast<glm::vec4*>(dst + dst_stride*i);
    const glm::vec3& s = *reinterpret_cast<const glm::vec3*>(src + src_stride*i);
    d = matrix * glm::vec4(s[0], s[1], s[2], 1.f);
  }
}