summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--src/video_core/textures/decoders.cpp36
1 files changed, 21 insertions, 15 deletions
diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp
index 225c999dc..89d218bc7 100644
--- a/src/video_core/textures/decoders.cpp
+++ b/src/video_core/textures/decoders.cpp
@@ -14,17 +14,12 @@ namespace Tegra::Texture {
14 14
15/** 15/**
16 * Calculates the offset of an (x, y) position within a swizzled texture. 16 * Calculates the offset of an (x, y) position within a swizzled texture.
17 * Taken from the Tegra X1 TRM. 17 * Taken from the Tegra X1 Technical Reference Manual. pages 1187-1188
18 */ 18 */
19static u32 GetSwizzleOffset(u32 x, u32 y, u32 image_width, u32 bytes_per_pixel, u32 block_height) { 19static u32 GetSwizzleOffset(u32 x, u32 y, u32 bytes_per_pixel, u32 gob_address) {
20 // Round up to the next gob 20 // Round up to the next gob
21 const u32 image_width_in_gobs{(image_width * bytes_per_pixel + 63) / 64};
22
23 u32 GOB_address = 0 + (y / (8 * block_height)) * 512 * block_height * image_width_in_gobs +
24 (x * bytes_per_pixel / 64) * 512 * block_height +
25 (y % (8 * block_height) / 8) * 512;
26 x *= bytes_per_pixel; 21 x *= bytes_per_pixel;
27 u32 address = GOB_address + ((x % 64) / 32) * 256 + ((y % 8) / 2) * 64 + ((x % 32) / 16) * 32 + 22 u32 address = gob_address + ((x % 64) / 32) * 256 + ((y % 8) / 2) * 64 + ((x % 32) / 16) * 32 +
28 (y % 2) * 16 + (x % 16); 23 (y % 2) * 16 + (x % 16);
29 24
30 return address; 25 return address;
@@ -32,21 +27,31 @@ static u32 GetSwizzleOffset(u32 x, u32 y, u32 image_width, u32 bytes_per_pixel,
32 27
33void CopySwizzledData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel, 28void CopySwizzledData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel,
34 u8* swizzled_data, u8* unswizzled_data, bool unswizzle, u32 block_height) { 29 u8* swizzled_data, u8* unswizzled_data, bool unswizzle, u32 block_height) {
35 u8* data_ptrs[2]; 30 std::array<u8*, 2> data_ptrs;
31 const u32 stride = width * bytes_per_pixel;
32 const u32 gobs_in_x = 64;
33 const u32 gobs_in_y = 8;
34 const u32 gobs_size = gobs_in_x * gobs_in_y;
35 const u32 image_width_in_gobs{(stride + gobs_in_x - 1) / gobs_in_x};
36 for (unsigned y = 0; y < height; ++y) { 36 for (unsigned y = 0; y < height; ++y) {
37 const u32 gob_y_address =
38 (y / (gobs_in_y * block_height)) * gobs_size * block_height * image_width_in_gobs +
39 (y % (gobs_in_y * block_height) / gobs_in_y) * gobs_size;
37 for (unsigned x = 0; x < width; ++x) { 40 for (unsigned x = 0; x < width; ++x) {
38 u32 swizzle_offset = GetSwizzleOffset(x, y, width, bytes_per_pixel, block_height); 41 const u32 gob_address =
39 u32 pixel_index = (x + y * width) * out_bytes_per_pixel; 42 gob_y_address + (x * bytes_per_pixel / gobs_in_x) * gobs_size * block_height;
43 const u32 swizzle_offset = GetSwizzleOffset(x, y, bytes_per_pixel, gob_address);
44 const u32 pixel_index = (x + y * width) * out_bytes_per_pixel;
40 45
41 data_ptrs[unswizzle] = swizzled_data + swizzle_offset; 46 data_ptrs[unswizzle] = swizzled_data + swizzle_offset;
42 data_ptrs[!unswizzle] = &unswizzled_data[pixel_index]; 47 data_ptrs[!unswizzle] = unswizzled_data + pixel_index;
43 48
44 std::memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel); 49 std::memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel);
45 } 50 }
46 } 51 }
47} 52}
48 53
49//This table represents the internal swizzle of a gob. 54// This table represents the internal swizzle of a gob.
50template <std::size_t N, std::size_t M> 55template <std::size_t N, std::size_t M>
51struct alignas(64) SwizzleTable { 56struct alignas(64) SwizzleTable {
52 constexpr SwizzleTable() { 57 constexpr SwizzleTable() {
@@ -72,7 +77,7 @@ void FastSwizzleData(u32 width, u32 height, u32 bytes_per_pixel, u8* swizzled_da
72 const std::size_t stride{width * bytes_per_pixel}; 77 const std::size_t stride{width * bytes_per_pixel};
73 const std::size_t gobs_in_x = 64; 78 const std::size_t gobs_in_x = 64;
74 const std::size_t gobs_in_y = 8; 79 const std::size_t gobs_in_y = 8;
75 const std::size_t gobs_size = gobs_in_x*gobs_in_y; 80 const std::size_t gobs_size = gobs_in_x * gobs_in_y;
76 const std::size_t image_width_in_gobs{(stride + gobs_in_x - 1) / gobs_in_x}; 81 const std::size_t image_width_in_gobs{(stride + gobs_in_x - 1) / gobs_in_x};
77 const std::size_t copy_size{16}; 82 const std::size_t copy_size{16};
78 for (std::size_t y = 0; y < height; ++y) { 83 for (std::size_t y = 0; y < height; ++y) {
@@ -83,7 +88,8 @@ void FastSwizzleData(u32 width, u32 height, u32 bytes_per_pixel, u8* swizzled_da
83 const auto& table = swizzle_table[y % gobs_in_y]; 88 const auto& table = swizzle_table[y % gobs_in_y];
84 for (std::size_t xb = 0; xb < stride; xb += copy_size) { 89 for (std::size_t xb = 0; xb < stride; xb += copy_size) {
85 const std::size_t truncated_copy = std::min(copy_size, stride - xb); 90 const std::size_t truncated_copy = std::min(copy_size, stride - xb);
86 const std::size_t gob_address{initial_gob + (xb / gobs_in_x) * gobs_size * block_height}; 91 const std::size_t gob_address{initial_gob +
92 (xb / gobs_in_x) * gobs_size * block_height};
87 const std::size_t swizzle_offset{gob_address + table[(xb / 16) % 4]}; 93 const std::size_t swizzle_offset{gob_address + table[(xb / 16) % 4]};
88 const std::size_t pixel_index{xb + pixel_base}; 94 const std::size_t pixel_index{xb + pixel_base};
89 data_ptrs[unswizzle] = swizzled_data + swizzle_offset; 95 data_ptrs[unswizzle] = swizzled_data + swizzle_offset;