N-archive/citra
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

AccelerateTextureCopy: Better support for contiguous copy

Browse source
This commit is contained in:
Phantom 2018-01-15 10:05:56 +01:00 committed by James Rowe
parent 7f1aec8fbb
commit 88f6521511
2 changed files with 65 additions and 75 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

48
src/video_core/renderer_opengl/gl_rasterizer.cpp
View file

@ -1019,41 +1019,33 @@ bool RasterizerOpenGL::AccelerateDisplayTransfer(const GPU::Regs::DisplayTransfe
bool RasterizerOpenGL::AccelerateTextureCopy(const GPU::Regs::DisplayTransferConfig& config) { bool RasterizerOpenGL::AccelerateTextureCopy(const GPU::Regs::DisplayTransferConfig& config) {
u32 copy_size = Common::AlignDown(config.texture_copy.size, 16); u32 copy_size = Common::AlignDown(config.texture_copy.size, 16);
if (copy_size == 0) {
if (copy_size == 0)
return false; return false;
}
u32 input_gap = config.texture_copy.input_gap * 16; u32 input_gap = config.texture_copy.input_gap * 16;
u32 input_width = config.texture_copy.input_width * 16; u32 input_width = config.texture_copy.input_width * 16;
if (input_width == 0) { if (input_width == 0 && input_gap != 0) {
if (input_gap == 0) {
input_width = copy_size;
} else {
return false; return false;
} }
if (input_gap == 0 || input_width >= copy_size) {
input_width = copy_size;
input_gap = 0;
}
if (copy_size % input_width != 0) {
return false;
} }
u32 output_gap = config.texture_copy.output_gap * 16; u32 output_gap = config.texture_copy.output_gap * 16;
u32 output_width = config.texture_copy.output_width * 16; u32 output_width = config.texture_copy.output_width * 16;
if (output_width == 0) { if (output_width == 0 && output_gap != 0) {
if (output_gap == 0) {
output_width = copy_size;
} else {
return false; return false;
} }
} if (output_gap == 0 || output_width >= copy_size) {
if (input_width >= copy_size) {
input_width = copy_size;
input_gap = 0;
}
if (output_width >= copy_size) {
output_width = copy_size; output_width = copy_size;
output_gap = 0; output_gap = 0;
} }
if (copy_size % output_width != 0) {
if (input_width != output_width || copy_size % input_width != 0) {
return false; return false;
} }
@ -1068,12 +1060,16 @@ bool RasterizerOpenGL::AccelerateTextureCopy(const GPU::Regs::DisplayTransferCon
MathUtil::Rectangle<u32> src_rect; MathUtil::Rectangle<u32> src_rect;
Surface src_surface; Surface src_surface;
std::tie(src_surface, src_rect) = res_cache.GetTexCopySurface(src_params); std::tie(src_surface, src_rect) = res_cache.GetTexCopySurface(src_params);
if (src_surface == nullptr) if (src_surface == nullptr) {
return false; return false;
}
if ((output_gap * 8) % SurfaceParams::GetFormatBpp(src_surface->pixel_format) != 0 || if (output_gap != 0 &&
(src_surface->is_tiled && src_surface->PixelsInBytes(output_gap) % 64 != 0)) (output_width != src_surface->BytesInPixels(src_rect.GetWidth() / src_surface->res_scale) *
(src_surface->is_tiled ? 8 : 1) ||
output_gap % src_surface->BytesInPixels(src_surface->is_tiled ? 64 : 1) != 0)) {
return false; return false;
}
SurfaceParams dst_params = *src_surface; SurfaceParams dst_params = *src_surface;
dst_params.addr = config.GetPhysicalOutputAddress(); dst_params.addr = config.GetPhysicalOutputAddress();
@ -1091,11 +1087,13 @@ bool RasterizerOpenGL::AccelerateTextureCopy(const GPU::Regs::DisplayTransferCon
Surface dst_surface; Surface dst_surface;
std::tie(dst_surface, dst_rect) = std::tie(dst_surface, dst_rect) =
res_cache.GetSurfaceSubRect(dst_params, ScaleMatch::Upscale, load_gap); res_cache.GetSurfaceSubRect(dst_params, ScaleMatch::Upscale, load_gap);
if (src_surface == nullptr) if (src_surface == nullptr) {
return false; return false;
}
if (!res_cache.BlitSurfaces(src_surface, src_rect, dst_surface, dst_rect)) if (!res_cache.BlitSurfaces(src_surface, src_rect, dst_surface, dst_rect)) {
return false; return false;
}
res_cache.InvalidateRegion(dst_params.addr, dst_params.size, dst_surface); res_cache.InvalidateRegion(dst_params.addr, dst_params.size, dst_surface);
return true; return true;

76
src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
View file

@ -390,6 +390,7 @@ SurfaceParams SurfaceParams::FromInterval(SurfaceInterval interval) const {
addr + Common::AlignUp(boost::icl::last_next(interval) - addr, tiled_alignment); addr + Common::AlignUp(boost::icl::last_next(interval) - addr, tiled_alignment);
params.addr = aligned_start; params.addr = aligned_start;
params.width = PixelsInBytes(aligned_end - aligned_start) / (is_tiled ? 8 : 1); params.width = PixelsInBytes(aligned_end - aligned_start) / (is_tiled ? 8 : 1);
params.stride = params.width;
params.height = is_tiled ? 8 : 1; params.height = is_tiled ? 8 : 1;
} }
params.UpdateParams(); params.UpdateParams();
@ -447,57 +448,45 @@ MathUtil::Rectangle<u32> SurfaceParams::GetScaledSubRect(const SurfaceParams& su
} }
bool SurfaceParams::ExactMatch(const SurfaceParams& other_surface) const { bool SurfaceParams::ExactMatch(const SurfaceParams& other_surface) const {
return (other_surface.addr == addr && other_surface.width == width && return other_surface.addr == addr && other_surface.width == width &&
other_surface.height == height && other_surface.stride == stride && other_surface.height == height && other_surface.stride == stride &&
other_surface.pixel_format == pixel_format && pixel_format != PixelFormat::Invalid && other_surface.pixel_format == pixel_format && pixel_format != PixelFormat::Invalid &&
other_surface.is_tiled == is_tiled); other_surface.is_tiled == is_tiled;
} }
bool SurfaceParams::CanSubRect(const SurfaceParams& sub_surface) const { bool SurfaceParams::CanSubRect(const SurfaceParams& sub_surface) const {
return (sub_surface.addr >= addr && sub_surface.end <= end && return sub_surface.addr >= addr && sub_surface.end <= end &&
sub_surface.pixel_format == pixel_format && pixel_format != PixelFormat::Invalid && sub_surface.pixel_format == pixel_format && pixel_format != PixelFormat::Invalid &&
sub_surface.is_tiled == is_tiled && sub_surface.is_tiled == is_tiled &&
(sub_surface.addr - addr) * 8 % GetFormatBpp() == 0 && (sub_surface.addr - addr) % BytesInPixels(is_tiled ? 64 : 1) == 0 &&
(!is_tiled || PixelsInBytes(sub_surface.addr - addr) % 64 == 0) &&
(sub_surface.stride == stride || sub_surface.height <= (is_tiled ? 8u : 1u)) && (sub_surface.stride == stride || sub_surface.height <= (is_tiled ? 8u : 1u)) &&
GetSubRect(sub_surface).left + sub_surface.width <= stride); GetSubRect(sub_surface).left + sub_surface.width <= stride;
} }
bool SurfaceParams::CanExpand(const SurfaceParams& expanded_surface) const { bool SurfaceParams::CanExpand(const SurfaceParams& expanded_surface) const {
if (pixel_format == PixelFormat::Invalid || pixel_format != expanded_surface.pixel_format || return pixel_format != PixelFormat::Invalid && pixel_format == expanded_surface.pixel_format &&
is_tiled != expanded_surface.is_tiled || addr > expanded_surface.end || addr <= expanded_surface.end && expanded_surface.addr <= end &&
expanded_surface.addr > end || stride != expanded_surface.stride) is_tiled == expanded_surface.is_tiled && stride == expanded_surface.stride &&
return false; (std::max(expanded_surface.addr, addr) - std::min(expanded_surface.addr, addr)) %
BytesInPixels(stride * (is_tiled ? 8 : 1)) ==
const u32 byte_offset = 0;
std::max(expanded_surface.addr, addr) - std::min(expanded_surface.addr, addr);
const int x0 = byte_offset % BytesInPixels(stride);
const int y0 = byte_offset / BytesInPixels(stride);
return x0 == 0 && (!is_tiled || y0 % 8 == 0);
} }
bool SurfaceParams::CanTexCopy(const SurfaceParams& texcopy_params) const { bool SurfaceParams::CanTexCopy(const SurfaceParams& texcopy_params) const {
if (pixel_format == PixelFormat::Invalid || addr > texcopy_params.addr || if (pixel_format == PixelFormat::Invalid || addr > texcopy_params.addr ||
end < texcopy_params.end || ((texcopy_params.addr - addr) * 8) % GetFormatBpp() != 0 || end < texcopy_params.end) {
(texcopy_params.width * 8) % GetFormatBpp() != 0 ||
(texcopy_params.stride * 8) % GetFormatBpp() != 0)
return false; return false;
const u32 begin_pixel_index = PixelsInBytes(texcopy_params.addr - addr);
if (!is_tiled) {
const int x0 = begin_pixel_index % stride;
return ((texcopy_params.height == 1 || PixelsInBytes(texcopy_params.stride) == stride) &&
x0 + PixelsInBytes(texcopy_params.width) <= stride);
} }
if (texcopy_params.width != texcopy_params.stride) {
const int x0 = (begin_pixel_index % (stride * 8)) / 8; const u32 tile_stride = BytesInPixels(stride * (is_tiled ? 8 : 1));
return (PixelsInBytes(texcopy_params.addr - addr) % 64 == 0 && return (texcopy_params.addr - addr) % BytesInPixels(is_tiled ? 64 : 1) == 0 &&
PixelsInBytes(texcopy_params.width) % 64 == 0 && texcopy_params.width % BytesInPixels(is_tiled ? 64 : 1) == 0 &&
(texcopy_params.height == 1 || PixelsInBytes(texcopy_params.stride) == stride * 8) && (texcopy_params.height == 1 || texcopy_params.stride == tile_stride) &&
x0 + PixelsInBytes(texcopy_params.width / 8) <= stride); ((texcopy_params.addr - addr) % tile_stride) + texcopy_params.width <= tile_stride;
} else {
return FromInterval(texcopy_params.GetInterval()).GetInterval() ==
texcopy_params.GetInterval();
}
} }
bool CachedSurface::CanFill(const SurfaceParams& dest_surface, bool CachedSurface::CanFill(const SurfaceParams& dest_surface,
@ -1245,14 +1234,17 @@ SurfaceRect_Tuple RasterizerCacheOpenGL::GetTexCopySurface(const SurfaceParams&
if (match_surface != nullptr) { if (match_surface != nullptr) {
ValidateSurface(match_surface, params.addr, params.size); ValidateSurface(match_surface, params.addr, params.size);
SurfaceParams match_subrect = params; SurfaceParams match_subrect;
match_subrect.width = match_surface->PixelsInBytes(params.width); if (params.width != params.stride) {
match_subrect.stride = match_surface->PixelsInBytes(params.stride); match_subrect = params;
match_subrect.width =
if (match_surface->is_tiled) { match_surface->PixelsInBytes(params.width) / (match_surface->is_tiled ? 8 : 1);
match_subrect.width /= 8; match_subrect.stride =
match_subrect.stride /= 8; match_surface->PixelsInBytes(params.stride) / (match_surface->is_tiled ? 8 : 1);
match_subrect.height *= 8; match_subrect.height *= (match_surface->is_tiled ? 8 : 1);
} else {
match_subrect = match_surface->FromInterval(params.GetInterval());
ASSERT(match_subrect.GetInterval() == params.GetInterval());
} }
rect = match_surface->GetScaledSubRect(match_subrect); rect = match_surface->GetScaledSubRect(match_subrect);