Merge pull request #1377 from FernandoS27/faster-swizzle
Improved Fast Swizzle and Legacy Swizzle
This commit is contained in:
commit
62048edc15
|
@ -13,47 +13,20 @@
|
||||||
namespace Tegra::Texture {
|
namespace Tegra::Texture {
|
||||||
|
|
||||||
/**
|
/**
|
||||||
|
* This table represents the internal swizzle of a gob,
|
||||||
|
* in format 16 bytes x 2 sector packing.
|
||||||
* Calculates the offset of an (x, y) position within a swizzled texture.
|
* Calculates the offset of an (x, y) position within a swizzled texture.
|
||||||
* Taken from the Tegra X1 TRM.
|
* Taken from the Tegra X1 Technical Reference Manual. pages 1187-1188
|
||||||
*/
|
*/
|
||||||
static u32 GetSwizzleOffset(u32 x, u32 y, u32 image_width, u32 bytes_per_pixel, u32 block_height) {
|
template <std::size_t N, std::size_t M, u32 Align>
|
||||||
// Round up to the next gob
|
|
||||||
const u32 image_width_in_gobs{(image_width * bytes_per_pixel + 63) / 64};
|
|
||||||
|
|
||||||
u32 GOB_address = 0 + (y / (8 * block_height)) * 512 * block_height * image_width_in_gobs +
|
|
||||||
(x * bytes_per_pixel / 64) * 512 * block_height +
|
|
||||||
(y % (8 * block_height) / 8) * 512;
|
|
||||||
x *= bytes_per_pixel;
|
|
||||||
u32 address = GOB_address + ((x % 64) / 32) * 256 + ((y % 8) / 2) * 64 + ((x % 32) / 16) * 32 +
|
|
||||||
(y % 2) * 16 + (x % 16);
|
|
||||||
|
|
||||||
return address;
|
|
||||||
}
|
|
||||||
|
|
||||||
void CopySwizzledData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel,
|
|
||||||
u8* swizzled_data, u8* unswizzled_data, bool unswizzle, u32 block_height) {
|
|
||||||
u8* data_ptrs[2];
|
|
||||||
for (unsigned y = 0; y < height; ++y) {
|
|
||||||
for (unsigned x = 0; x < width; ++x) {
|
|
||||||
u32 swizzle_offset = GetSwizzleOffset(x, y, width, bytes_per_pixel, block_height);
|
|
||||||
u32 pixel_index = (x + y * width) * out_bytes_per_pixel;
|
|
||||||
|
|
||||||
data_ptrs[unswizzle] = swizzled_data + swizzle_offset;
|
|
||||||
data_ptrs[!unswizzle] = &unswizzled_data[pixel_index];
|
|
||||||
|
|
||||||
std::memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
template <std::size_t N, std::size_t M>
|
|
||||||
struct alignas(64) SwizzleTable {
|
struct alignas(64) SwizzleTable {
|
||||||
|
static_assert(M * Align == 64, "Swizzle Table does not align to GOB");
|
||||||
constexpr SwizzleTable() {
|
constexpr SwizzleTable() {
|
||||||
for (u32 y = 0; y < N; ++y) {
|
for (u32 y = 0; y < N; ++y) {
|
||||||
for (u32 x = 0; x < M; ++x) {
|
for (u32 x = 0; x < M; ++x) {
|
||||||
const u32 x2 = x * 16;
|
const u32 x2 = x * Align;
|
||||||
values[y][x] = static_cast<u16>(((x2 % 64) / 32) * 256 + ((y % 8) / 2) * 64 +
|
values[y][x] = static_cast<u16>(((x2 % 64) / 32) * 256 + ((y % 8) / 2) * 64 +
|
||||||
((x2 % 32) / 16) * 32 + (y % 2) * 16);
|
((x2 % 32) / 16) * 32 + (y % 2) * 16 + (x2 % 16));
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
@ -63,24 +36,60 @@ struct alignas(64) SwizzleTable {
|
||||||
std::array<std::array<u16, M>, N> values{};
|
std::array<std::array<u16, M>, N> values{};
|
||||||
};
|
};
|
||||||
|
|
||||||
constexpr auto swizzle_table = SwizzleTable<8, 4>();
|
constexpr auto legacy_swizzle_table = SwizzleTable<8, 64, 1>();
|
||||||
|
constexpr auto fast_swizzle_table = SwizzleTable<8, 4, 16>();
|
||||||
|
|
||||||
void FastSwizzleData(u32 width, u32 height, u32 bytes_per_pixel, u8* swizzled_data,
|
static void LegacySwizzleData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel,
|
||||||
u8* unswizzled_data, bool unswizzle, u32 block_height) {
|
u8* swizzled_data, u8* unswizzled_data, bool unswizzle,
|
||||||
|
u32 block_height) {
|
||||||
|
std::array<u8*, 2> data_ptrs;
|
||||||
|
const std::size_t stride = width * bytes_per_pixel;
|
||||||
|
const std::size_t gobs_in_x = 64;
|
||||||
|
const std::size_t gobs_in_y = 8;
|
||||||
|
const std::size_t gobs_size = gobs_in_x * gobs_in_y;
|
||||||
|
const std::size_t image_width_in_gobs{(stride + gobs_in_x - 1) / gobs_in_x};
|
||||||
|
for (std::size_t y = 0; y < height; ++y) {
|
||||||
|
const std::size_t gob_y_address =
|
||||||
|
(y / (gobs_in_y * block_height)) * gobs_size * block_height * image_width_in_gobs +
|
||||||
|
(y % (gobs_in_y * block_height) / gobs_in_y) * gobs_size;
|
||||||
|
const auto& table = legacy_swizzle_table[y % gobs_in_y];
|
||||||
|
for (std::size_t x = 0; x < width; ++x) {
|
||||||
|
const std::size_t gob_address =
|
||||||
|
gob_y_address + (x * bytes_per_pixel / gobs_in_x) * gobs_size * block_height;
|
||||||
|
const std::size_t x2 = x * bytes_per_pixel;
|
||||||
|
const std::size_t swizzle_offset = gob_address + table[x2 % gobs_in_x];
|
||||||
|
const std::size_t pixel_index = (x + y * width) * out_bytes_per_pixel;
|
||||||
|
|
||||||
|
data_ptrs[unswizzle] = swizzled_data + swizzle_offset;
|
||||||
|
data_ptrs[!unswizzle] = unswizzled_data + pixel_index;
|
||||||
|
|
||||||
|
std::memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
static void FastSwizzleData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel,
|
||||||
|
u8* swizzled_data, u8* unswizzled_data, bool unswizzle,
|
||||||
|
u32 block_height) {
|
||||||
std::array<u8*, 2> data_ptrs;
|
std::array<u8*, 2> data_ptrs;
|
||||||
const std::size_t stride{width * bytes_per_pixel};
|
const std::size_t stride{width * bytes_per_pixel};
|
||||||
const std::size_t image_width_in_gobs{(stride + 63) / 64};
|
const std::size_t gobs_in_x = 64;
|
||||||
|
const std::size_t gobs_in_y = 8;
|
||||||
|
const std::size_t gobs_size = gobs_in_x * gobs_in_y;
|
||||||
|
const std::size_t image_width_in_gobs{(stride + gobs_in_x - 1) / gobs_in_x};
|
||||||
const std::size_t copy_size{16};
|
const std::size_t copy_size{16};
|
||||||
for (std::size_t y = 0; y < height; ++y) {
|
for (std::size_t y = 0; y < height; ++y) {
|
||||||
const std::size_t initial_gob =
|
const std::size_t initial_gob =
|
||||||
(y / (8 * block_height)) * 512 * block_height * image_width_in_gobs +
|
(y / (gobs_in_y * block_height)) * gobs_size * block_height * image_width_in_gobs +
|
||||||
(y % (8 * block_height) / 8) * 512;
|
(y % (gobs_in_y * block_height) / gobs_in_y) * gobs_size;
|
||||||
const std::size_t pixel_base{y * width * bytes_per_pixel};
|
const std::size_t pixel_base{y * width * out_bytes_per_pixel};
|
||||||
const auto& table = swizzle_table[y % 8];
|
const auto& table = fast_swizzle_table[y % gobs_in_y];
|
||||||
for (std::size_t xb = 0; xb < stride; xb += copy_size) {
|
for (std::size_t xb = 0; xb < stride; xb += copy_size) {
|
||||||
const std::size_t gob_address{initial_gob + (xb / 64) * 512 * block_height};
|
const std::size_t gob_address{initial_gob +
|
||||||
|
(xb / gobs_in_x) * gobs_size * block_height};
|
||||||
const std::size_t swizzle_offset{gob_address + table[(xb / 16) % 4]};
|
const std::size_t swizzle_offset{gob_address + table[(xb / 16) % 4]};
|
||||||
const std::size_t pixel_index{xb + pixel_base};
|
const std::size_t out_x = xb * out_bytes_per_pixel / bytes_per_pixel;
|
||||||
|
const std::size_t pixel_index{out_x + pixel_base};
|
||||||
data_ptrs[unswizzle] = swizzled_data + swizzle_offset;
|
data_ptrs[unswizzle] = swizzled_data + swizzle_offset;
|
||||||
data_ptrs[!unswizzle] = unswizzled_data + pixel_index;
|
data_ptrs[!unswizzle] = unswizzled_data + pixel_index;
|
||||||
std::memcpy(data_ptrs[0], data_ptrs[1], copy_size);
|
std::memcpy(data_ptrs[0], data_ptrs[1], copy_size);
|
||||||
|
@ -88,6 +97,17 @@ void FastSwizzleData(u32 width, u32 height, u32 bytes_per_pixel, u8* swizzled_da
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
void CopySwizzledData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel,
|
||||||
|
u8* swizzled_data, u8* unswizzled_data, bool unswizzle, u32 block_height) {
|
||||||
|
if (bytes_per_pixel % 3 != 0 && (width * bytes_per_pixel) % 16 == 0) {
|
||||||
|
FastSwizzleData(width, height, bytes_per_pixel, out_bytes_per_pixel, swizzled_data,
|
||||||
|
unswizzled_data, unswizzle, block_height);
|
||||||
|
} else {
|
||||||
|
LegacySwizzleData(width, height, bytes_per_pixel, out_bytes_per_pixel, swizzled_data,
|
||||||
|
unswizzled_data, unswizzle, block_height);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
u32 BytesPerPixel(TextureFormat format) {
|
u32 BytesPerPixel(TextureFormat format) {
|
||||||
switch (format) {
|
switch (format) {
|
||||||
case TextureFormat::DXT1:
|
case TextureFormat::DXT1:
|
||||||
|
@ -134,13 +154,8 @@ u32 BytesPerPixel(TextureFormat format) {
|
||||||
std::vector<u8> UnswizzleTexture(VAddr address, u32 tile_size, u32 bytes_per_pixel, u32 width,
|
std::vector<u8> UnswizzleTexture(VAddr address, u32 tile_size, u32 bytes_per_pixel, u32 width,
|
||||||
u32 height, u32 block_height) {
|
u32 height, u32 block_height) {
|
||||||
std::vector<u8> unswizzled_data(width * height * bytes_per_pixel);
|
std::vector<u8> unswizzled_data(width * height * bytes_per_pixel);
|
||||||
if (bytes_per_pixel % 3 != 0 && (width * bytes_per_pixel) % 16 == 0) {
|
|
||||||
FastSwizzleData(width / tile_size, height / tile_size, bytes_per_pixel,
|
|
||||||
Memory::GetPointer(address), unswizzled_data.data(), true, block_height);
|
|
||||||
} else {
|
|
||||||
CopySwizzledData(width / tile_size, height / tile_size, bytes_per_pixel, bytes_per_pixel,
|
CopySwizzledData(width / tile_size, height / tile_size, bytes_per_pixel, bytes_per_pixel,
|
||||||
Memory::GetPointer(address), unswizzled_data.data(), true, block_height);
|
Memory::GetPointer(address), unswizzled_data.data(), true, block_height);
|
||||||
}
|
|
||||||
return unswizzled_data;
|
return unswizzled_data;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
Loading…
Reference in a new issue