artemist/yuzu
1
0
Fork 0
Code Issues Pull requests Packages Projects Releases Wiki Activity

Merge pull request #4242 from ReinUsesLisp/maxwell-dma

Browse source 
maxwell_dma: Match official doc and support pitch->voxel copies
This commit is contained in:
bunnei 2020-07-14 14:04:16 -04:00 committed by GitHub
parent e2730372b8 c574ab5aa1
commit 666b37ad56
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
6 changed files with 492 additions and 383 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

328
src/video_core/engines/maxwell_dma.cpp
View file

@ -14,50 +14,45 @@
namespace Tegra::Engines {
using namespace Texture;
MaxwellDMA::MaxwellDMA(Core::System& system, MemoryManager& memory_manager)
: system{system}, memory_manager{memory_manager} {}
void MaxwellDMA::CallMethod(u32 method, u32 method_argument, bool is_last_call) {
ASSERT_MSG(method < Regs::NUM_REGS,
"Invalid MaxwellDMA register, increase the size of the Regs structure");
ASSERT_MSG(method < NUM_REGS, "Invalid MaxwellDMA register");
regs.reg_array[method] = method_argument;
#define MAXWELLDMA_REG_INDEX(field_name) \
(offsetof(Tegra::Engines::MaxwellDMA::Regs, field_name) / sizeof(u32))
switch (method) {
case MAXWELLDMA_REG_INDEX(exec): {
HandleCopy();
break;
if (method == offsetof(Regs, launch_dma) / sizeof(u32)) {
Launch();
}
}
#undef MAXWELLDMA_REG_INDEX
}
void MaxwellDMA::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
u32 methods_pending) {
for (std::size_t i = 0; i < amount; i++) {
for (size_t i = 0; i < amount; ++i) {
CallMethod(method, base_start[i], methods_pending - static_cast<u32>(i) <= 1);
}
}
void MaxwellDMA::HandleCopy() {
LOG_TRACE(HW_GPU, "Requested a DMA copy");
const GPUVAddr source = regs.src_address.Address();
const GPUVAddr dest = regs.dst_address.Address();
void MaxwellDMA::Launch() {
LOG_TRACE(Render_OpenGL, "DMA copy 0x{:x} -> 0x{:x}", static_cast<GPUVAddr>(regs.offset_in),
static_cast<GPUVAddr>(regs.offset_out));
// TODO(Subv): Perform more research and implement all features of this engine.
ASSERT(regs.exec.enable_swizzle == 0);
ASSERT(regs.exec.query_mode == Regs::QueryMode::None);
ASSERT(regs.exec.query_intr == Regs::QueryIntr::None);
ASSERT(regs.exec.copy_mode == Regs::CopyMode::Unk2);
ASSERT(regs.dst_params.pos_x == 0);
ASSERT(regs.dst_params.pos_y == 0);
const LaunchDMA& launch = regs.launch_dma;
ASSERT(launch.remap_enable == 0);
ASSERT(launch.semaphore_type == LaunchDMA::SemaphoreType::NONE);
ASSERT(launch.interrupt_type == LaunchDMA::InterruptType::NONE);
ASSERT(launch.data_transfer_type == LaunchDMA::DataTransferType::NON_PIPELINED);
ASSERT(regs.dst_params.origin.x == 0);
ASSERT(regs.dst_params.origin.y == 0);
if (!regs.exec.is_dst_linear && !regs.exec.is_src_linear) {
const bool is_src_pitch = launch.src_memory_layout == LaunchDMA::MemoryLayout::PITCH;
const bool is_dst_pitch = launch.dst_memory_layout == LaunchDMA::MemoryLayout::PITCH;
if (!is_src_pitch && !is_dst_pitch) {
// If both the source and the destination are in block layout, assert.
UNREACHABLE_MSG("Tiled->Tiled DMA transfers are not yet implemented");
return;
@ -66,144 +61,161 @@ void MaxwellDMA::HandleCopy() {
// All copies here update the main memory, so mark all rasterizer states as invalid.
system.GPU().Maxwell3D().OnMemoryWrite();
if (regs.exec.is_dst_linear && regs.exec.is_src_linear) {
// When the enable_2d bit is disabled, the copy is performed as if we were copying a 1D
// buffer of length `x_count`, otherwise we copy a 2D image of dimensions (x_count,
// y_count).
if (!regs.exec.enable_2d) {
memory_manager.CopyBlock(dest, source, regs.x_count);
return;
}
// If both the source and the destination are in linear layout, perform a line-by-line
// copy. We're going to take a subrect of size (x_count, y_count) from the source
// rectangle. There is no need to manually flush/invalidate the regions because
// CopyBlock does that for us.
for (u32 line = 0; line < regs.y_count; ++line) {
const GPUVAddr source_line = source + line * regs.src_pitch;
const GPUVAddr dest_line = dest + line * regs.dst_pitch;
memory_manager.CopyBlock(dest_line, source_line, regs.x_count);
}
return;
}
ASSERT(regs.exec.enable_2d == 1);
if (regs.exec.is_dst_linear && !regs.exec.is_src_linear) {
ASSERT(regs.src_params.BlockDepth() == 0);
// Optimized path for micro copies.
if (regs.dst_pitch * regs.y_count < Texture::GetGOBSize() && regs.dst_pitch <= 64) {
const u32 bytes_per_pixel = regs.dst_pitch / regs.x_count;
const std::size_t src_size = Texture::GetGOBSize();
const std::size_t dst_size = regs.dst_pitch * regs.y_count;
u32 pos_x = regs.src_params.pos_x;
u32 pos_y = regs.src_params.pos_y;
const u64 offset =
Texture::GetGOBOffset(regs.src_params.size_x, regs.src_params.size_y, pos_x, pos_y,
regs.src_params.BlockDepth(), bytes_per_pixel);
const u32 x_in_gob = 64 / bytes_per_pixel;
pos_x = pos_x % x_in_gob;
pos_y = pos_y % 8;
if (read_buffer.size() < src_size) {
read_buffer.resize(src_size);
}
if (write_buffer.size() < dst_size) {
write_buffer.resize(dst_size);
}
if (Settings::IsGPULevelExtreme()) {
memory_manager.ReadBlock(source + offset, read_buffer.data(), src_size);
memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
} else {
memory_manager.ReadBlockUnsafe(source + offset, read_buffer.data(), src_size);
memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
}
Texture::UnswizzleSubrect(regs.x_count, regs.y_count, regs.dst_pitch,
regs.src_params.size_x, bytes_per_pixel, read_buffer.data(),
write_buffer.data(), regs.src_params.BlockHeight(), pos_x,
pos_y);
memory_manager.WriteBlock(dest, write_buffer.data(), dst_size);
return;
}
// If the input is tiled and the output is linear, deswizzle the input and copy it over.
const u32 bytes_per_pixel = regs.dst_pitch / regs.x_count;
const std::size_t src_size = Texture::CalculateSize(
true, bytes_per_pixel, regs.src_params.size_x, regs.src_params.size_y,
regs.src_params.size_z, regs.src_params.BlockHeight(), regs.src_params.BlockDepth());
const std::size_t src_layer_size = Texture::CalculateSize(
true, bytes_per_pixel, regs.src_params.size_x, regs.src_params.size_y, 1,
regs.src_params.BlockHeight(), regs.src_params.BlockDepth());
const std::size_t dst_size = regs.dst_pitch * regs.y_count;
if (read_buffer.size() < src_size) {
read_buffer.resize(src_size);
}
if (write_buffer.size() < dst_size) {
write_buffer.resize(dst_size);
}
if (Settings::IsGPULevelExtreme()) {
memory_manager.ReadBlock(source, read_buffer.data(), src_size);
memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
} else {
memory_manager.ReadBlockUnsafe(source, read_buffer.data(), src_size);
memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
}
Texture::UnswizzleSubrect(
regs.x_count, regs.y_count, regs.dst_pitch, regs.src_params.size_x, bytes_per_pixel,
read_buffer.data() + src_layer_size * regs.src_params.pos_z, write_buffer.data(),
regs.src_params.BlockHeight(), regs.src_params.pos_x, regs.src_params.pos_y);
memory_manager.WriteBlock(dest, write_buffer.data(), dst_size);
if (is_src_pitch && is_dst_pitch) {
CopyPitchToPitch();
} else {
ASSERT(regs.dst_params.BlockDepth() == 0);
ASSERT(launch.multi_line_enable == 1);
const u32 bytes_per_pixel = regs.src_pitch / regs.x_count;
const std::size_t dst_size = Texture::CalculateSize(
true, bytes_per_pixel, regs.dst_params.size_x, regs.dst_params.size_y,
regs.dst_params.size_z, regs.dst_params.BlockHeight(), regs.dst_params.BlockDepth());
const std::size_t dst_layer_size = Texture::CalculateSize(
true, bytes_per_pixel, regs.dst_params.size_x, regs.dst_params.size_y, 1,
regs.dst_params.BlockHeight(), regs.dst_params.BlockDepth());
const std::size_t src_size = regs.src_pitch * regs.y_count;
if (read_buffer.size() < src_size) {
read_buffer.resize(src_size);
}
if (write_buffer.size() < dst_size) {
write_buffer.resize(dst_size);
}
if (Settings::IsGPULevelExtreme()) {
memory_manager.ReadBlock(source, read_buffer.data(), src_size);
memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
if (!is_src_pitch && is_dst_pitch) {
CopyBlockLinearToPitch();
} else {
memory_manager.ReadBlockUnsafe(source, read_buffer.data(), src_size);
memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
CopyPitchToBlockLinear();
}
// If the input is linear and the output is tiled, swizzle the input and copy it over.
Texture::SwizzleSubrect(
regs.x_count, regs.y_count, regs.src_pitch, regs.dst_params.size_x, bytes_per_pixel,
write_buffer.data() + dst_layer_size * regs.dst_params.pos_z, read_buffer.data(),
regs.dst_params.BlockHeight(), regs.dst_params.pos_x, regs.dst_params.pos_y);
memory_manager.WriteBlock(dest, write_buffer.data(), dst_size);
}
}
void MaxwellDMA::CopyPitchToPitch() {
// When `multi_line_enable` bit is disabled the copy is performed as if we were copying a 1D
// buffer of length `line_length_in`.
// Otherwise we copy a 2D image of dimensions (line_length_in, line_count).
if (!regs.launch_dma.multi_line_enable) {
memory_manager.CopyBlock(regs.offset_out, regs.offset_in, regs.line_length_in);
return;
}
// Perform a line-by-line copy.
// We're going to take a subrect of size (line_length_in, line_count) from the source rectangle.
// There is no need to manually flush/invalidate the regions because CopyBlock does that for us.
for (u32 line = 0; line < regs.line_count; ++line) {
const GPUVAddr source_line = regs.offset_in + static_cast<size_t>(line) * regs.pitch_in;
const GPUVAddr dest_line = regs.offset_out + static_cast<size_t>(line) * regs.pitch_out;
memory_manager.CopyBlock(dest_line, source_line, regs.line_length_in);
}
}
void MaxwellDMA::CopyBlockLinearToPitch() {
ASSERT(regs.src_params.block_size.depth == 0);
// Optimized path for micro copies.
const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
if (dst_size < GOB_SIZE && regs.pitch_out <= GOB_SIZE_X) {
FastCopyBlockLinearToPitch();
return;
}
// Deswizzle the input and copy it over.
const u32 bytes_per_pixel = regs.pitch_out / regs.line_length_in;
const Parameters& src_params = regs.src_params;
const u32 width = src_params.width;
const u32 height = src_params.height;
const u32 depth = src_params.depth;
const u32 block_height = src_params.block_size.height;
const u32 block_depth = src_params.block_size.depth;
const size_t src_size =
CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
const size_t src_layer_size =
CalculateSize(true, bytes_per_pixel, width, height, 1, block_height, block_depth);
if (read_buffer.size() < src_size) {
read_buffer.resize(src_size);
}
if (write_buffer.size() < dst_size) {
write_buffer.resize(dst_size);
}
if (Settings::IsGPULevelExtreme()) {
memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
} else {
memory_manager.ReadBlockUnsafe(regs.offset_in, read_buffer.data(), src_size);
memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
}
UnswizzleSubrect(regs.line_length_in, regs.line_count, regs.pitch_out, width, bytes_per_pixel,
read_buffer.data() + src_layer_size * src_params.layer, write_buffer.data(),
block_height, src_params.origin.x, src_params.origin.y);
memory_manager.WriteBlock(regs.offset_out, write_buffer.data(), dst_size);
}
void MaxwellDMA::CopyPitchToBlockLinear() {
const auto& dst_params = regs.dst_params;
const u32 bytes_per_pixel = regs.pitch_in / regs.line_length_in;
const u32 width = dst_params.width;
const u32 height = dst_params.height;
const u32 depth = dst_params.depth;
const u32 block_height = dst_params.block_size.height;
const u32 block_depth = dst_params.block_size.depth;
const size_t dst_size =
CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
const size_t dst_layer_size =
CalculateSize(true, bytes_per_pixel, width, height, 1, block_height, block_depth);
const size_t src_size = static_cast<size_t>(regs.pitch_in) * regs.line_count;
if (read_buffer.size() < src_size) {
read_buffer.resize(src_size);
}
if (write_buffer.size() < dst_size) {
write_buffer.resize(dst_size);
}
if (Settings::IsGPULevelExtreme()) {
memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
} else {
memory_manager.ReadBlockUnsafe(regs.offset_in, read_buffer.data(), src_size);
memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
}
// If the input is linear and the output is tiled, swizzle the input and copy it over.
if (regs.dst_params.block_size.depth > 0) {
ASSERT(dst_params.layer == 0);
SwizzleSliceToVoxel(regs.line_length_in, regs.line_count, regs.pitch_in, width, height,
bytes_per_pixel, block_height, block_depth, dst_params.origin.x,
dst_params.origin.y, write_buffer.data(), read_buffer.data());
} else {
SwizzleSubrect(regs.line_length_in, regs.line_count, regs.pitch_in, width, bytes_per_pixel,
write_buffer.data() + dst_layer_size * dst_params.layer, read_buffer.data(),
block_height, dst_params.origin.x, dst_params.origin.y);
}
memory_manager.WriteBlock(regs.offset_out, write_buffer.data(), dst_size);
}
void MaxwellDMA::FastCopyBlockLinearToPitch() {
const u32 bytes_per_pixel = regs.pitch_out / regs.line_length_in;
const size_t src_size = GOB_SIZE;
const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
u32 pos_x = regs.src_params.origin.x;
u32 pos_y = regs.src_params.origin.y;
const u64 offset = GetGOBOffset(regs.src_params.width, regs.src_params.height, pos_x, pos_y,
regs.src_params.block_size.height, bytes_per_pixel);
const u32 x_in_gob = 64 / bytes_per_pixel;
pos_x = pos_x % x_in_gob;
pos_y = pos_y % 8;
if (read_buffer.size() < src_size) {
read_buffer.resize(src_size);
}
if (write_buffer.size() < dst_size) {
write_buffer.resize(dst_size);
}
if (Settings::IsGPULevelExtreme()) {
memory_manager.ReadBlock(regs.offset_in + offset, read_buffer.data(), src_size);
memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
} else {
memory_manager.ReadBlockUnsafe(regs.offset_in + offset, read_buffer.data(), src_size);
memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
}
UnswizzleSubrect(regs.line_length_in, regs.line_count, regs.pitch_out, regs.src_params.width,
bytes_per_pixel, read_buffer.data(), write_buffer.data(),
regs.src_params.block_size.height, pos_x, pos_y);
memory_manager.WriteBlock(regs.offset_out, write_buffer.data(), dst_size);
}
} // namespace Tegra::Engines

368
src/video_core/engines/maxwell_dma.h
View file

@ -24,12 +24,167 @@ class MemoryManager;
namespace Tegra::Engines {
/**
* This Engine is known as GK104_Copy. Documentation can be found in:
* This engine is known as gk104_copy. Documentation can be found in:
* https://github.com/NVIDIA/open-gpu-doc/blob/master/classes/dma-copy/clb0b5.h
* https://github.com/envytools/envytools/blob/master/rnndb/fifo/gk104_copy.xml
*/
class MaxwellDMA final : public EngineInterface {
public:
struct PackedGPUVAddr {
u32 upper;
u32 lower;
constexpr operator GPUVAddr() const noexcept {
return (static_cast<GPUVAddr>(upper & 0xff) << 32) | lower;
}
};
union BlockSize {
BitField<0, 4, u32> width;
BitField<4, 4, u32> height;
BitField<8, 4, u32> depth;
BitField<12, 4, u32> gob_height;
};
static_assert(sizeof(BlockSize) == 4);
union Origin {
BitField<0, 16, u32> x;
BitField<16, 16, u32> y;
};
static_assert(sizeof(Origin) == 4);
struct Parameters {
BlockSize block_size;
u32 width;
u32 height;
u32 depth;
u32 layer;
Origin origin;
};
static_assert(sizeof(Parameters) == 24);
struct Semaphore {
PackedGPUVAddr address;
u32 payload;
};
static_assert(sizeof(Semaphore) == 12);
struct RenderEnable {
enum class Mode : u32 {
FALSE = 0,
TRUE = 1,
CONDITIONAL = 2,
RENDER_IF_EQUAL = 3,
RENDER_IF_NOT_EQUAL = 4,
};
PackedGPUVAddr address;
BitField<0, 3, Mode> mode;
};
static_assert(sizeof(RenderEnable) == 12);
enum class PhysModeTarget : u32 {
LOCAL_FB = 0,
COHERENT_SYSMEM = 1,
NONCOHERENT_SYSMEM = 2,
};
using PhysMode = BitField<0, 2, PhysModeTarget>;
union LaunchDMA {
enum class DataTransferType : u32 {
NONE = 0,
PIPELINED = 1,
NON_PIPELINED = 2,
};
enum class SemaphoreType : u32 {
NONE = 0,
RELEASE_ONE_WORD_SEMAPHORE = 1,
RELEASE_FOUR_WORD_SEMAPHORE = 2,
};
enum class InterruptType : u32 {
NONE = 0,
BLOCKING = 1,
NON_BLOCKING = 2,
};
enum class MemoryLayout : u32 {
BLOCKLINEAR = 0,
PITCH = 1,
};
enum class Type : u32 {
VIRTUAL = 0,
PHYSICAL = 1,
};
enum class SemaphoreReduction : u32 {
IMIN = 0,
IMAX = 1,
IXOR = 2,
IAND = 3,
IOR = 4,
IADD = 5,
INC = 6,
DEC = 7,
FADD = 0xA,
};
enum class SemaphoreReductionSign : u32 {
SIGNED = 0,
UNSIGNED = 1,
};
enum class BypassL2 : u32 {
USE_PTE_SETTING = 0,
FORCE_VOLATILE = 1,
};
BitField<0, 2, DataTransferType> data_transfer_type;
BitField<2, 1, u32> flush_enable;
BitField<3, 2, SemaphoreType> semaphore_type;
BitField<5, 2, InterruptType> interrupt_type;
BitField<7, 1, MemoryLayout> src_memory_layout;
BitField<8, 1, MemoryLayout> dst_memory_layout;
BitField<9, 1, u32> multi_line_enable;
BitField<10, 1, u32> remap_enable;
BitField<11, 1, u32> rmwdisable;
BitField<12, 1, Type> src_type;
BitField<13, 1, Type> dst_type;
BitField<14, 4, SemaphoreReduction> semaphore_reduction;
BitField<18, 1, SemaphoreReductionSign> semaphore_reduction_sign;
BitField<19, 1, u32> reduction_enable;
BitField<20, 1, BypassL2> bypass_l2;
};
static_assert(sizeof(LaunchDMA) == 4);
struct RemapConst {
enum Swizzle : u32 {
SRC_X = 0,
SRC_Y = 1,
SRC_Z = 2,
SRC_W = 3,
CONST_A = 4,
CONST_B = 5,
NO_WRITE = 6,
};
PackedGPUVAddr address;
union {
BitField<0, 3, Swizzle> dst_x;
BitField<4, 3, Swizzle> dst_y;
BitField<8, 3, Swizzle> dst_z;
BitField<12, 3, Swizzle> dst_w;
BitField<16, 2, u32> component_size_minus_one;
BitField<20, 2, u32> num_src_components_minus_one;
BitField<24, 2, u32> num_dst_components_minus_one;
};
};
static_assert(sizeof(RemapConst) == 12);
explicit MaxwellDMA(Core::System& system, MemoryManager& memory_manager);
~MaxwellDMA() = default;
@ -40,144 +195,19 @@ public:
void CallMultiMethod(u32 method, const u32* base_start, u32 amount,
u32 methods_pending) override;
struct Regs {
static constexpr std::size_t NUM_REGS = 0x1D6;
struct Parameters {
union {
BitField<0, 4, u32> block_depth;
BitField<4, 4, u32> block_height;
BitField<8, 4, u32> block_width;
};
u32 size_x;
u32 size_y;
u32 size_z;
u32 pos_z;
union {
BitField<0, 16, u32> pos_x;
BitField<16, 16, u32> pos_y;
};
u32 BlockHeight() const {
return block_height.Value();
}
u32 BlockDepth() const {
return block_depth.Value();
}
};
static_assert(sizeof(Parameters) == 24, "Parameters has wrong size");
enum class ComponentMode : u32 {
Src0 = 0,
Src1 = 1,
Src2 = 2,
Src3 = 3,
Const0 = 4,
Const1 = 5,
Zero = 6,
};
enum class CopyMode : u32 {
None = 0,
Unk1 = 1,
Unk2 = 2,
};
enum class QueryMode : u32 {
None = 0,
Short = 1,
Long = 2,
};
enum class QueryIntr : u32 {
None = 0,
Block = 1,
NonBlock = 2,
};
union {
struct {
INSERT_UNION_PADDING_WORDS(0xC0);
struct {
union {
BitField<0, 2, CopyMode> copy_mode;
BitField<2, 1, u32> flush;
BitField<3, 2, QueryMode> query_mode;
BitField<5, 2, QueryIntr> query_intr;
BitField<7, 1, u32> is_src_linear;
BitField<8, 1, u32> is_dst_linear;
BitField<9, 1, u32> enable_2d;
BitField<10, 1, u32> enable_swizzle;
};
} exec;
INSERT_UNION_PADDING_WORDS(0x3F);
struct {
u32 address_high;
u32 address_low;
GPUVAddr Address() const {
return static_cast<GPUVAddr>((static_cast<GPUVAddr>(address_high) << 32) |
address_low);
}
} src_address;
struct {
u32 address_high;
u32 address_low;
GPUVAddr Address() const {
return static_cast<GPUVAddr>((static_cast<GPUVAddr>(address_high) << 32) |
address_low);
}
} dst_address;
u32 src_pitch;
u32 dst_pitch;
u32 x_count;
u32 y_count;
INSERT_UNION_PADDING_WORDS(0xB8);
u32 const0;
u32 const1;
union {
BitField<0, 4, ComponentMode> component0;
BitField<4, 4, ComponentMode> component1;
BitField<8, 4, ComponentMode> component2;
BitField<12, 4, ComponentMode> component3;
BitField<16, 2, u32> component_size;
BitField<20, 3, u32> src_num_components;
BitField<24, 3, u32> dst_num_components;
u32 SrcBytePerPixel() const {
return src_num_components.Value() * component_size.Value();
}
u32 DstBytePerPixel() const {
return dst_num_components.Value() * component_size.Value();
}
} swizzle_config;
Parameters dst_params;
INSERT_UNION_PADDING_WORDS(1);
Parameters src_params;
INSERT_UNION_PADDING_WORDS(0x13);
};
std::array<u32, NUM_REGS> reg_array;
};
} regs{};
private:
/// Performs the copy from the source buffer to the destination buffer as configured in the
/// registers.
void Launch();
void CopyPitchToPitch();
void CopyBlockLinearToPitch();
void CopyPitchToBlockLinear();
void FastCopyBlockLinearToPitch();
Core::System& system;
MemoryManager& memory_manager;
@ -185,28 +215,58 @@ private:
std::vector<u8> read_buffer;
std::vector<u8> write_buffer;
/// Performs the copy from the source buffer to the destination buffer as configured in the
/// registers.
void HandleCopy();
};
static constexpr std::size_t NUM_REGS = 0x800;
struct Regs {
union {
struct {
u32 reserved[0x40];
u32 nop;
u32 reserved01[0xf];
u32 pm_trigger;
u32 reserved02[0x3f];
Semaphore semaphore;
u32 reserved03[0x2];
RenderEnable render_enable;
PhysMode src_phys_mode;
PhysMode dst_phys_mode;
u32 reserved04[0x26];
LaunchDMA launch_dma;
u32 reserved05[0x3f];
PackedGPUVAddr offset_in;
PackedGPUVAddr offset_out;
u32 pitch_in;
u32 pitch_out;
u32 line_length_in;
u32 line_count;
u32 reserved06[0xb8];
RemapConst remap_const;
Parameters dst_params;
u32 reserved07[0x1];
Parameters src_params;
u32 reserved08[0x275];
u32 pm_trigger_end;
u32 reserved09[0x3ba];
};
std::array<u32, NUM_REGS> reg_array;
};
} regs{};
#define ASSERT_REG_POSITION(field_name, position) \
static_assert(offsetof(MaxwellDMA::Regs, field_name) == position * 4, \
"Field " #field_name " has invalid position")
ASSERT_REG_POSITION(exec, 0xC0);
ASSERT_REG_POSITION(src_address, 0x100);
ASSERT_REG_POSITION(dst_address, 0x102);
ASSERT_REG_POSITION(src_pitch, 0x104);
ASSERT_REG_POSITION(dst_pitch, 0x105);
ASSERT_REG_POSITION(x_count, 0x106);
ASSERT_REG_POSITION(y_count, 0x107);
ASSERT_REG_POSITION(const0, 0x1C0);
ASSERT_REG_POSITION(const1, 0x1C1);
ASSERT_REG_POSITION(swizzle_config, 0x1C2);
ASSERT_REG_POSITION(dst_params, 0x1C3);
ASSERT_REG_POSITION(src_params, 0x1CA);
ASSERT_REG_POSITION(launch_dma, 0xC0);
ASSERT_REG_POSITION(offset_in, 0x100);
ASSERT_REG_POSITION(offset_out, 0x102);
ASSERT_REG_POSITION(pitch_in, 0x104);
ASSERT_REG_POSITION(pitch_out, 0x105);
ASSERT_REG_POSITION(line_length_in, 0x106);
ASSERT_REG_POSITION(line_count, 0x107);
ASSERT_REG_POSITION(remap_const, 0x1C0);
ASSERT_REG_POSITION(dst_params, 0x1C3);
ASSERT_REG_POSITION(src_params, 0x1CA);
#undef ASSERT_REG_POSITION
};
} // namespace Tegra::Engines

5
src/video_core/texture_cache/surface_params.cpp
View file

@ -343,8 +343,7 @@ std::size_t SurfaceParams::GetLayerSize(bool as_host_size, bool uncompressed) co
size += GetInnerMipmapMemorySize(level, as_host_size, uncompressed);
}
if (is_tiled && is_layered) {
return Common::AlignBits(size,
Tegra::Texture::GetGOBSizeShift() + block_height + block_depth);
return Common::AlignBits(size, Tegra::Texture::GOB_SIZE_SHIFT + block_height + block_depth);
}
return size;
}
@ -418,7 +417,7 @@ std::tuple<u32, u32, u32> SurfaceParams::GetBlockOffsetXYZ(u32 offset) const {
const u32 block_size = GetBlockSize();
const u32 block_index = offset / block_size;
const u32 gob_offset = offset % block_size;
const u32 gob_index = gob_offset / static_cast<u32>(Tegra::Texture::GetGOBSize());
const u32 gob_index = gob_offset / static_cast<u32>(Tegra::Texture::GOB_SIZE);
const u32 x_gob_pixels = 64U / GetBytesPerPixel();
const u32 x_block_pixels = x_gob_pixels << block_width;
const u32 y_block_pixels = 8U << block_height;

2
src/video_core/texture_cache/surface_params.h
View file

@ -204,7 +204,7 @@ public:
static std::size_t AlignLayered(const std::size_t out_size, const u32 block_height,
const u32 block_depth) {
return Common::AlignBits(out_size,
Tegra::Texture::GetGOBSizeShift() + block_height + block_depth);
Tegra::Texture::GOB_SIZE_SHIFT + block_height + block_depth);
}
/// Converts a width from a type of surface into another. This helps represent the

134
src/video_core/textures/decoders.cpp
View file

@ -6,6 +6,7 @@
#include <cstring>
#include "common/alignment.h"
#include "common/assert.h"
#include "common/bit_util.h"
#include "video_core/gpu.h"
#include "video_core/textures/decoders.h"
#include "video_core/textures/texture.h"
@ -37,20 +38,10 @@ struct alignas(64) SwizzleTable {
std::array<std::array<u16, M>, N> values{};
};
constexpr u32 gob_size_x_shift = 6;
constexpr u32 gob_size_y_shift = 3;
constexpr u32 gob_size_z_shift = 0;
constexpr u32 gob_size_shift = gob_size_x_shift + gob_size_y_shift + gob_size_z_shift;
constexpr u32 FAST_SWIZZLE_ALIGN = 16;
constexpr u32 gob_size_x = 1U << gob_size_x_shift;
constexpr u32 gob_size_y = 1U << gob_size_y_shift;
constexpr u32 gob_size_z = 1U << gob_size_z_shift;
constexpr u32 gob_size = 1U << gob_size_shift;
constexpr u32 fast_swizzle_align = 16;
constexpr auto legacy_swizzle_table = SwizzleTable<gob_size_y, gob_size_x, gob_size_z>();
constexpr auto fast_swizzle_table = SwizzleTable<gob_size_y, 4, fast_swizzle_align>();
constexpr auto LEGACY_SWIZZLE_TABLE = SwizzleTable<GOB_SIZE_X, GOB_SIZE_X, GOB_SIZE_Z>();
constexpr auto FAST_SWIZZLE_TABLE = SwizzleTable<GOB_SIZE_Y, 4, FAST_SWIZZLE_ALIGN>();
/**
* This function manages ALL the GOBs(Group of Bytes) Inside a single block.
@ -69,17 +60,17 @@ void PreciseProcessBlock(u8* const swizzled_data, u8* const unswizzled_data, con
u32 y_address = z_address;
u32 pixel_base = layer_z * z + y_start * stride_x;
for (u32 y = y_start; y < y_end; y++) {
const auto& table = legacy_swizzle_table[y % gob_size_y];
const auto& table = LEGACY_SWIZZLE_TABLE[y % GOB_SIZE_Y];
for (u32 x = x_start; x < x_end; x++) {
const u32 swizzle_offset{y_address + table[x * bytes_per_pixel % gob_size_x]};
const u32 swizzle_offset{y_address + table[x * bytes_per_pixel % GOB_SIZE_X]};
const u32 pixel_index{x * out_bytes_per_pixel + pixel_base};
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);
}
pixel_base += stride_x;
if ((y + 1) % gob_size_y == 0)
y_address += gob_size;
if ((y + 1) % GOB_SIZE_Y == 0)
y_address += GOB_SIZE;
}
z_address += xy_block_size;
}
@ -104,18 +95,18 @@ void FastProcessBlock(u8* const swizzled_data, u8* const unswizzled_data, const
u32 y_address = z_address;
u32 pixel_base = layer_z * z + y_start * stride_x;
for (u32 y = y_start; y < y_end; y++) {
const auto& table = fast_swizzle_table[y % gob_size_y];
for (u32 xb = x_startb; xb < x_endb; xb += fast_swizzle_align) {
const u32 swizzle_offset{y_address + table[(xb / fast_swizzle_align) % 4]};
const auto& table = FAST_SWIZZLE_TABLE[y % GOB_SIZE_Y];
for (u32 xb = x_startb; xb < x_endb; xb += FAST_SWIZZLE_ALIGN) {
const u32 swizzle_offset{y_address + table[(xb / FAST_SWIZZLE_ALIGN) % 4]};
const u32 out_x = xb * out_bytes_per_pixel / bytes_per_pixel;
const u32 pixel_index{out_x + pixel_base};
data_ptrs[unswizzle ? 1 : 0] = swizzled_data + swizzle_offset;
data_ptrs[unswizzle ? 0 : 1] = unswizzled_data + pixel_index;
std::memcpy(data_ptrs[0], data_ptrs[1], fast_swizzle_align);
std::memcpy(data_ptrs[0], data_ptrs[1], FAST_SWIZZLE_ALIGN);
}
pixel_base += stride_x;
if ((y + 1) % gob_size_y == 0)
y_address += gob_size;
if ((y + 1) % GOB_SIZE_Y == 0)
y_address += GOB_SIZE;
}
z_address += xy_block_size;
}
@ -138,9 +129,9 @@ void SwizzledData(u8* const swizzled_data, u8* const unswizzled_data, const bool
auto div_ceil = [](const u32 x, const u32 y) { return ((x + y - 1) / y); };
const u32 stride_x = width * out_bytes_per_pixel;
const u32 layer_z = height * stride_x;
const u32 gob_elements_x = gob_size_x / bytes_per_pixel;
constexpr u32 gob_elements_y = gob_size_y;
constexpr u32 gob_elements_z = gob_size_z;
const u32 gob_elements_x = GOB_SIZE_X / bytes_per_pixel;
constexpr u32 gob_elements_y = GOB_SIZE_Y;
constexpr u32 gob_elements_z = GOB_SIZE_Z;
const u32 block_x_elements = gob_elements_x;
const u32 block_y_elements = gob_elements_y * block_height;
const u32 block_z_elements = gob_elements_z * block_depth;
@ -148,7 +139,7 @@ void SwizzledData(u8* const swizzled_data, u8* const unswizzled_data, const bool
const u32 blocks_on_x = div_ceil(aligned_width, block_x_elements);
const u32 blocks_on_y = div_ceil(height, block_y_elements);
const u32 blocks_on_z = div_ceil(depth, block_z_elements);
const u32 xy_block_size = gob_size * block_height;
const u32 xy_block_size = GOB_SIZE * block_height;
const u32 block_size = xy_block_size * block_depth;
u32 tile_offset = 0;
for (u32 zb = 0; zb < blocks_on_z; zb++) {
@ -182,7 +173,7 @@ void CopySwizzledData(u32 width, u32 height, u32 depth, u32 bytes_per_pixel,
bool unswizzle, u32 block_height, u32 block_depth, u32 width_spacing) {
const u32 block_height_size{1U << block_height};
const u32 block_depth_size{1U << block_depth};
if (bytes_per_pixel % 3 != 0 && (width * bytes_per_pixel) % fast_swizzle_align == 0) {
if (bytes_per_pixel % 3 != 0 && (width * bytes_per_pixel) % FAST_SWIZZLE_ALIGN == 0) {
SwizzledData<true>(swizzled_data, unswizzled_data, unswizzle, width, height, depth,
bytes_per_pixel, out_bytes_per_pixel, block_height_size,
block_depth_size, width_spacing);
@ -259,25 +250,26 @@ std::vector<u8> UnswizzleTexture(u8* address, u32 tile_size_x, u32 tile_size_y,
}
void SwizzleSubrect(u32 subrect_width, u32 subrect_height, u32 source_pitch, u32 swizzled_width,
u32 bytes_per_pixel, u8* swizzled_data, u8* unswizzled_data,
u32 bytes_per_pixel, u8* swizzled_data, const u8* unswizzled_data,
u32 block_height_bit, u32 offset_x, u32 offset_y) {
const u32 block_height = 1U << block_height_bit;
const u32 image_width_in_gobs{(swizzled_width * bytes_per_pixel + (gob_size_x - 1)) /
gob_size_x};
const u32 image_width_in_gobs =
(swizzled_width * bytes_per_pixel + (GOB_SIZE_X - 1)) / GOB_SIZE_X;
for (u32 line = 0; line < subrect_height; ++line) {
const u32 dst_y = line + offset_y;
const u32 gob_address_y =
(dst_y / (gob_size_y * block_height)) * gob_size * block_height * image_width_in_gobs +
((dst_y % (gob_size_y * block_height)) / gob_size_y) * gob_size;
const auto& table = legacy_swizzle_table[dst_y % gob_size_y];
(dst_y / (GOB_SIZE_Y * block_height)) * GOB_SIZE * block_height * image_width_in_gobs +
((dst_y % (GOB_SIZE_Y * block_height)) / GOB_SIZE_Y) * GOB_SIZE;
const auto& table = LEGACY_SWIZZLE_TABLE[dst_y % GOB_SIZE_Y];
for (u32 x = 0; x < subrect_width; ++x) {
const u32 dst_x = x + offset_x;
const u32 gob_address =
gob_address_y + (dst_x * bytes_per_pixel / gob_size_x) * gob_size * block_height;
const u32 swizzled_offset = gob_address + table[(dst_x * bytes_per_pixel) % gob_size_x];
u8* source_line = unswizzled_data + line * source_pitch + x * bytes_per_pixel;
u8* dest_addr = swizzled_data + swizzled_offset;
gob_address_y + (dst_x * bytes_per_pixel / GOB_SIZE_X) * GOB_SIZE * block_height;
const u32 swizzled_offset = gob_address + table[(dst_x * bytes_per_pixel) % GOB_SIZE_X];
const u32 unswizzled_offset = line * source_pitch + x * bytes_per_pixel;
const u8* const source_line = unswizzled_data + unswizzled_offset;
u8* const dest_addr = swizzled_data + swizzled_offset;
std::memcpy(dest_addr, source_line, bytes_per_pixel);
}
}
@ -289,14 +281,15 @@ void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32
const u32 block_height = 1U << block_height_bit;
for (u32 line = 0; line < subrect_height; ++line) {
const u32 y2 = line + offset_y;
const u32 gob_address_y = (y2 / (gob_size_y * block_height)) * gob_size * block_height +
((y2 % (gob_size_y * block_height)) / gob_size_y) * gob_size;
const auto& table = legacy_swizzle_table[y2 % gob_size_y];
const u32 gob_address_y = (y2 / (GOB_SIZE_Y * block_height)) * GOB_SIZE * block_height +
((y2 % (GOB_SIZE_Y * block_height)) / GOB_SIZE_Y) * GOB_SIZE;
const auto& table = LEGACY_SWIZZLE_TABLE[y2 % GOB_SIZE_Y];
for (u32 x = 0; x < subrect_width; ++x) {
const u32 x2 = (x + offset_x) * bytes_per_pixel;
const u32 gob_address = gob_address_y + (x2 / gob_size_x) * gob_size * block_height;
const u32 swizzled_offset = gob_address + table[x2 % gob_size_x];
u8* dest_line = unswizzled_data + line * dest_pitch + x * bytes_per_pixel;
const u32 gob_address = gob_address_y + (x2 / GOB_SIZE_X) * GOB_SIZE * block_height;
const u32 swizzled_offset = gob_address + table[x2 % GOB_SIZE_X];
const u32 unswizzled_offset = line * dest_pitch + x * bytes_per_pixel;
u8* dest_line = unswizzled_data + unswizzled_offset;
u8* source_addr = swizzled_data + swizzled_offset;
std::memcpy(dest_line, source_addr, bytes_per_pixel);
@ -304,21 +297,48 @@ void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32
}
}
void SwizzleSliceToVoxel(u32 line_length_in, u32 line_count, u32 pitch, u32 width, u32 height,
u32 bytes_per_pixel, u32 block_height, u32 block_depth, u32 origin_x,
u32 origin_y, u8* output, const u8* input) {
UNIMPLEMENTED_IF(origin_x > 0);
UNIMPLEMENTED_IF(origin_y > 0);
const u32 stride = width * bytes_per_pixel;
const u32 gobs_in_x = (stride + GOB_SIZE_X - 1) / GOB_SIZE_X;
const u32 block_size = gobs_in_x << (GOB_SIZE_SHIFT + block_height + block_depth);
const u32 block_height_mask = (1U << block_height) - 1;
const u32 x_shift = Common::CountTrailingZeroes32(GOB_SIZE << (block_height + block_depth));
for (u32 line = 0; line < line_count; ++line) {
const auto& table = LEGACY_SWIZZLE_TABLE[line % GOB_SIZE_Y];
const u32 block_y = line / GOB_SIZE_Y;
const u32 dst_offset_y =
(block_y >> block_height) * block_size + (block_y & block_height_mask) * GOB_SIZE;
for (u32 x = 0; x < line_length_in; ++x) {
const u32 dst_offset =
((x / GOB_SIZE_X) << x_shift) + dst_offset_y + table[x % GOB_SIZE_X];
const u32 src_offset = x * bytes_per_pixel + line * pitch;
std::memcpy(output + dst_offset, input + src_offset, bytes_per_pixel);
}
}
}
void SwizzleKepler(const u32 width, const u32 height, const u32 dst_x, const u32 dst_y,
const u32 block_height_bit, const std::size_t copy_size, const u8* source_data,
u8* swizzle_data) {
const u32 block_height = 1U << block_height_bit;
const u32 image_width_in_gobs{(width + gob_size_x - 1) / gob_size_x};
const u32 image_width_in_gobs{(width + GOB_SIZE_X - 1) / GOB_SIZE_X};
std::size_t count = 0;
for (std::size_t y = dst_y; y < height && count < copy_size; ++y) {
const std::size_t gob_address_y =
(y / (gob_size_y * block_height)) * gob_size * block_height * image_width_in_gobs +
((y % (gob_size_y * block_height)) / gob_size_y) * gob_size;
const auto& table = legacy_swizzle_table[y % gob_size_y];
(y / (GOB_SIZE_Y * block_height)) * GOB_SIZE * block_height * image_width_in_gobs +
((y % (GOB_SIZE_Y * block_height)) / GOB_SIZE_Y) * GOB_SIZE;
const auto& table = LEGACY_SWIZZLE_TABLE[y % GOB_SIZE_Y];
for (std::size_t x = dst_x; x < width && count < copy_size; ++x) {
const std::size_t gob_address =
gob_address_y + (x / gob_size_x) * gob_size * block_height;
const std::size_t swizzled_offset = gob_address + table[x % gob_size_x];
gob_address_y + (x / GOB_SIZE_X) * GOB_SIZE * block_height;
const std::size_t swizzled_offset = gob_address + table[x % GOB_SIZE_X];
const u8* source_line = source_data + count;
u8* dest_addr = swizzle_data + swizzled_offset;
count++;
@ -373,9 +393,9 @@ std::vector<u8> DecodeTexture(const std::vector<u8>& texture_data, TextureFormat
std::size_t CalculateSize(bool tiled, u32 bytes_per_pixel, u32 width, u32 height, u32 depth,
u32 block_height, u32 block_depth) {
if (tiled) {
const u32 aligned_width = Common::AlignBits(width * bytes_per_pixel, gob_size_x_shift);
const u32 aligned_height = Common::AlignBits(height, gob_size_y_shift + block_height);
const u32 aligned_depth = Common::AlignBits(depth, gob_size_z_shift + block_depth);
const u32 aligned_width = Common::AlignBits(width * bytes_per_pixel, GOB_SIZE_X_SHIFT);
const u32 aligned_height = Common::AlignBits(height, GOB_SIZE_Y_SHIFT + block_height);
const u32 aligned_depth = Common::AlignBits(depth, GOB_SIZE_Z_SHIFT + block_depth);
return aligned_width * aligned_height * aligned_depth;
} else {
return width * height * depth * bytes_per_pixel;
@ -386,14 +406,14 @@ 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 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 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;
return base + (relative_y / GOB_SIZE_Y) * GOB_SIZE;
}
} // namespace Tegra::Texture

38
src/video_core/textures/decoders.h
View file

@ -10,15 +10,15 @@
namespace Tegra::Texture {
// GOBSize constant. Calculated by 64 bytes in x multiplied by 8 y coords, represents
// an small rect of (64/bytes_per_pixel)X8.
inline std::size_t GetGOBSize() {
return 512;
}
constexpr u32 GOB_SIZE_X = 64;
constexpr u32 GOB_SIZE_Y = 8;
constexpr u32 GOB_SIZE_Z = 1;
constexpr u32 GOB_SIZE = GOB_SIZE_X * GOB_SIZE_Y * GOB_SIZE_Z;
inline std::size_t GetGOBSizeShift() {
return 9;
}
constexpr std::size_t GOB_SIZE_X_SHIFT = 6;
constexpr std::size_t GOB_SIZE_Y_SHIFT = 3;
constexpr std::size_t GOB_SIZE_Z_SHIFT = 0;
constexpr std::size_t GOB_SIZE_SHIFT = GOB_SIZE_X_SHIFT + GOB_SIZE_Y_SHIFT + GOB_SIZE_Z_SHIFT;
/// Unswizzles a swizzled texture without changing its format.
void UnswizzleTexture(u8* unswizzled_data, u8* address, u32 tile_size_x, u32 tile_size_y,
@ -48,14 +48,32 @@ std::size_t CalculateSize(bool tiled, u32 bytes_per_pixel, u32 width, u32 height
/// Copies an untiled subrectangle into a tiled surface.
void SwizzleSubrect(u32 subrect_width, u32 subrect_height, u32 source_pitch, u32 swizzled_width,
u32 bytes_per_pixel, u8* swizzled_data, u8* unswizzled_data, u32 block_height,
u32 offset_x, u32 offset_y);
u32 bytes_per_pixel, u8* swizzled_data, const u8* unswizzled_data,
u32 block_height_bit, u32 offset_x, u32 offset_y);
/// Copies a tiled subrectangle into a linear surface.
void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32 swizzled_width,
u32 bytes_per_pixel, u8* swizzled_data, u8* unswizzled_data, u32 block_height,
u32 offset_x, u32 offset_y);
/// @brief Swizzles a 2D array of pixels into a 3D texture
/// @param line_length_in Number of pixels per line
/// @param line_count Number of lines
/// @param pitch Number of bytes per line
/// @param width Width of the swizzled texture
/// @param height Height of the swizzled texture
/// @param bytes_per_pixel Number of bytes used per pixel
/// @param block_height Block height shift
/// @param block_depth Block depth shift
/// @param origin_x Column offset in pixels of the swizzled texture
/// @param origin_y Row offset in pixels of the swizzled texture
/// @param output Pointer to the pixels of the swizzled texture
/// @param input Pointer to the 2D array of pixels used as input
/// @pre input and output points to an array large enough to hold the number of bytes used
void SwizzleSliceToVoxel(u32 line_length_in, u32 line_count, u32 pitch, u32 width, u32 height,
u32 bytes_per_pixel, u32 block_height, u32 block_depth, u32 origin_x,
u32 origin_y, u8* output, const u8* input);
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);