Engines: Implement Accelerate DMA Texture.

This commit is contained in:
Fernando Sahmkow
2022-03-27 05:05:57 +02:00
parent ce8f4da638
commit 8a3411b417
15 changed files with 656 additions and 95 deletions

View File

@ -14,7 +14,13 @@
#include "video_core/textures/decoders.h"
MICROPROFILE_DECLARE(GPU_DMAEngine);
MICROPROFILE_DECLARE(GPU_DMAEngineBL);
MICROPROFILE_DECLARE(GPU_DMAEngineLB);
MICROPROFILE_DECLARE(GPU_DMAEngineBB);
MICROPROFILE_DEFINE(GPU_DMAEngine, "GPU", "DMA Engine", MP_RGB(224, 224, 128));
MICROPROFILE_DEFINE(GPU_DMAEngineBL, "GPU", "DMA Engine Block - Linear", MP_RGB(224, 224, 128));
MICROPROFILE_DEFINE(GPU_DMAEngineLB, "GPU", "DMA Engine Linear - Block", MP_RGB(224, 224, 128));
MICROPROFILE_DEFINE(GPU_DMAEngineBB, "GPU", "DMA Engine Block - Block", MP_RGB(224, 224, 128));
namespace Tegra::Engines {
@ -72,6 +78,7 @@ void MaxwellDMA::Launch() {
memory_manager.FlushCaching();
if (!is_src_pitch && !is_dst_pitch) {
// If both the source and the destination are in block layout, assert.
MICROPROFILE_SCOPE(GPU_DMAEngineBB);
CopyBlockLinearToBlockLinear();
ReleaseSemaphore();
return;
@ -87,8 +94,10 @@ void MaxwellDMA::Launch() {
}
} else {
if (!is_src_pitch && is_dst_pitch) {
MICROPROFILE_SCOPE(GPU_DMAEngineBL);
CopyBlockLinearToPitch();
} else {
MICROPROFILE_SCOPE(GPU_DMAEngineLB);
CopyPitchToBlockLinear();
}
}
@ -153,21 +162,35 @@ void MaxwellDMA::Launch() {
}
void MaxwellDMA::CopyBlockLinearToPitch() {
UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0);
UNIMPLEMENTED_IF(regs.src_params.layer != 0);
UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
const bool is_remapping = regs.launch_dma.remap_enable != 0;
u32 bytes_per_pixel = 1;
DMA::ImageOperand src_operand;
src_operand.bytes_per_pixel = bytes_per_pixel;
src_operand.params = regs.src_params;
src_operand.address = regs.offset_in;
// Optimized path for micro copies.
const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
if (!is_remapping && dst_size < GOB_SIZE && regs.pitch_out <= GOB_SIZE_X &&
regs.src_params.height > GOB_SIZE_Y) {
FastCopyBlockLinearToPitch();
DMA::BufferOperand dst_operand;
dst_operand.pitch = regs.pitch_out;
dst_operand.width = regs.line_length_in;
dst_operand.height = regs.line_count;
dst_operand.address = regs.offset_out;
DMA::ImageCopy copy_info{};
copy_info.length_x = regs.line_length_in;
copy_info.length_y = regs.line_count;
auto& accelerate = rasterizer->AccessAccelerateDMA();
if (accelerate.ImageToBuffer(copy_info, src_operand, dst_operand)) {
return;
}
UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0);
UNIMPLEMENTED_IF(regs.src_params.block_size.depth != 0);
UNIMPLEMENTED_IF(regs.src_params.block_size.depth == 0 && regs.src_params.depth != 1);
// Deswizzle the input and copy it over.
const Parameters& src_params = regs.src_params;
const DMA::Parameters& src_params = regs.src_params;
const bool is_remapping = regs.launch_dma.remap_enable != 0;
const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1;
const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;
@ -187,7 +210,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
x_offset >>= bpp_shift;
}
const u32 bytes_per_pixel = base_bpp << bpp_shift;
bytes_per_pixel = base_bpp << bpp_shift;
const u32 height = src_params.height;
const u32 depth = src_params.depth;
const u32 block_height = src_params.block_size.height;
@ -195,11 +218,12 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
const size_t src_size =
CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
read_buffer.resize_destructive(src_size);
write_buffer.resize_destructive(dst_size);
memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
memory_manager.ReadBlock(src_operand.address, read_buffer.data(), src_size);
memory_manager.ReadBlockUnsafe(dst_operand.address, write_buffer.data(), dst_size);
UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
src_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
@ -216,6 +240,24 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1;
const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;
u32 bytes_per_pixel = 1;
DMA::ImageOperand dst_operand;
dst_operand.bytes_per_pixel = bytes_per_pixel;
dst_operand.params = regs.dst_params;
dst_operand.address = regs.offset_out;
DMA::BufferOperand src_operand;
src_operand.pitch = regs.pitch_in;
src_operand.width = regs.line_length_in;
src_operand.height = regs.line_count;
src_operand.address = regs.offset_in;
DMA::ImageCopy copy_info{};
copy_info.length_x = regs.line_length_in;
copy_info.length_y = regs.line_count;
auto& accelerate = rasterizer->AccessAccelerateDMA();
if (accelerate.BufferToImage(copy_info, src_operand, dst_operand)) {
return;
}
const auto& dst_params = regs.dst_params;
const u32 base_bpp = !is_remapping ? 1U : num_remap_components * remap_components_size;
@ -233,7 +275,7 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
x_offset >>= bpp_shift;
}
const u32 bytes_per_pixel = base_bpp << bpp_shift;
bytes_per_pixel = base_bpp << bpp_shift;
const u32 height = dst_params.height;
const u32 depth = dst_params.depth;
const u32 block_height = dst_params.block_size.height;
@ -260,45 +302,14 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
}
void MaxwellDMA::FastCopyBlockLinearToPitch() {
const u32 bytes_per_pixel = 1U;
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;
read_buffer.resize_destructive(src_size);
write_buffer.resize_destructive(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(write_buffer, read_buffer, bytes_per_pixel, regs.src_params.width,
regs.src_params.height, 1, pos_x, pos_y, regs.line_length_in, regs.line_count,
regs.src_params.block_size.height, regs.src_params.block_size.depth,
regs.pitch_out);
memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
}
void MaxwellDMA::CopyBlockLinearToBlockLinear() {
UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0);
const bool is_remapping = regs.launch_dma.remap_enable != 0;
// Deswizzle the input and copy it over.
const Parameters& src = regs.src_params;
const Parameters& dst = regs.dst_params;
const DMA::Parameters& src = regs.src_params;
const DMA::Parameters& dst = regs.dst_params;
const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1;
const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;