From 37c690576f8b7b9b4207973807b826411452f0c3 Mon Sep 17 00:00:00 2001 From: Fernando Sahmkow Date: Mon, 27 Apr 2020 23:57:24 -0400 Subject: MaxwellDMA: Optimize micro copies. --- src/video_core/textures/decoders.cpp | 14 ++++++++++++++ src/video_core/textures/decoders.h | 3 +++ 2 files changed, 17 insertions(+) (limited to 'src/video_core/textures') diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp index fae8638ec..548e4c3fe 100644 --- a/src/video_core/textures/decoders.cpp +++ b/src/video_core/textures/decoders.cpp @@ -382,4 +382,18 @@ std::size_t CalculateSize(bool tiled, u32 bytes_per_pixel, u32 width, u32 height } } +u64 GetGOBOffset(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height, + u32 bytes_per_pixel) { + auto div_ceil = [](const u32 x, const u32 y) { return ((x + y - 1) / y); }; + const u32 gobs_in_block = 1 << block_height; + const u32 y_blocks = gob_size_y << block_height; + const u32 x_per_gob = gob_size_x / bytes_per_pixel; + const u32 x_blocks = div_ceil(width, x_per_gob); + const u32 block_size = gob_size * gobs_in_block; + const u32 stride = block_size * x_blocks; + const u32 base = (dst_y / y_blocks) * stride + (dst_x / x_per_gob) * block_size; + const u32 relative_y = dst_y % y_blocks; + return base + (relative_y / gob_size_y) * gob_size; +} + } // namespace Tegra::Texture diff --git a/src/video_core/textures/decoders.h b/src/video_core/textures/decoders.h index 9f2d6d308..e0ff83754 100644 --- a/src/video_core/textures/decoders.h +++ b/src/video_core/textures/decoders.h @@ -59,4 +59,7 @@ void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32 void SwizzleKepler(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height, std::size_t copy_size, const u8* source_data, u8* swizzle_data); +u64 GetGOBOffset(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height, + u32 bytes_per_pixel); + } // namespace Tegra::Texture -- cgit v1.2.3 From 9df67b20956c8d0ecaf9fea4a3c82858a66adcca Mon Sep 17 00:00:00 2001 From: Fernando Sahmkow Date: Tue, 28 Apr 2020 13:53:47 -0400 Subject: Clang Format and Documentation. --- src/video_core/textures/decoders.h | 1 + 1 file changed, 1 insertion(+) (limited to 'src/video_core/textures') diff --git a/src/video_core/textures/decoders.h b/src/video_core/textures/decoders.h index e0ff83754..06f3ebf87 100644 --- a/src/video_core/textures/decoders.h +++ b/src/video_core/textures/decoders.h @@ -59,6 +59,7 @@ void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32 void SwizzleKepler(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height, std::size_t copy_size, const u8* source_data, u8* swizzle_data); +/// Obtains the offset of the gob for positions 'dst_x' & 'dst_y' u64 GetGOBOffset(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height, u32 bytes_per_pixel); -- cgit v1.2.3