Merge pull request #6249 from stenzek/texture-dump
Texture dumping regression fixes
This commit is contained in:
commit
a70c03a1c0
|
@ -427,12 +427,10 @@ void OGLStagingTexture::CopyFromTexture(const AbstractTexture* src,
|
|||
}
|
||||
else
|
||||
{
|
||||
glActiveTexture(GL_TEXTURE9);
|
||||
glBindTexture(GL_TEXTURE_2D_ARRAY, gltex->GetRawTexIdentifier());
|
||||
if (g_ogl_config.bSupportsTextureSubImage)
|
||||
{
|
||||
glGetTextureSubImage(
|
||||
GL_TEXTURE_2D_ARRAY, src_level, src_rect.left, src_rect.top, src_layer,
|
||||
gltex->GetRawTexIdentifier(), src_level, src_rect.left, src_rect.top, src_layer,
|
||||
src_rect.GetWidth(), src_rect.GetHeight(), 1,
|
||||
GetGLFormatForTextureFormat(m_config.format), GetGLTypeForTextureFormat(m_config.format),
|
||||
static_cast<GLsizei>(m_buffer_size - dst_offset), reinterpret_cast<void*>(dst_offset));
|
||||
|
@ -442,11 +440,12 @@ void OGLStagingTexture::CopyFromTexture(const AbstractTexture* src,
|
|||
// TODO: Investigate whether it's faster to use glReadPixels() with a framebuffer, since we're
|
||||
// copying the whole texture, which may waste bandwidth. So we're trading CPU work in creating
|
||||
// the framebuffer for GPU work in copying potentially redundant texels.
|
||||
glActiveTexture(GL_TEXTURE9);
|
||||
glBindTexture(GL_TEXTURE_2D_ARRAY, gltex->GetRawTexIdentifier());
|
||||
glGetTexImage(GL_TEXTURE_2D_ARRAY, src_level, GetGLFormatForTextureFormat(m_config.format),
|
||||
GetGLTypeForTextureFormat(m_config.format), nullptr);
|
||||
OGLTexture::SetStage();
|
||||
}
|
||||
|
||||
OGLTexture::SetStage();
|
||||
}
|
||||
|
||||
glPixelStorei(GL_PACK_ROW_LENGTH, 0);
|
||||
|
|
|
@ -1090,7 +1090,7 @@ TextureCacheBase::GetTexture(u32 address, u32 width, u32 height, const TextureFo
|
|||
|
||||
entry->has_arbitrary_mips = arbitrary_mip_detector.HasArbitraryMipmaps(dst_buffer);
|
||||
|
||||
if (g_ActiveConfig.bDumpTextures)
|
||||
if (g_ActiveConfig.bDumpTextures && !hires_tex)
|
||||
{
|
||||
for (u32 level = 0; level < texLevels; ++level)
|
||||
{
|
||||
|
|
Loading…
Reference in New Issue