mctaylors/suyu
1
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

Merge pull request #8766 from Kelebek1/regs

Browse source 
[video_core] Update 3D registers
This commit is contained in:
Fernando S 2022-10-09 07:04:03 +02:00 committed by GitHub
commit 55e6d0dae0
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
29 changed files with 4036 additions and 2105 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

34
src/video_core/buffer_cache/buffer_cache.h
View file

@ -994,13 +994,13 @@ void BufferCache<P>::BindHostIndexBuffer() {
const u32 size = index_buffer.size;
SynchronizeBuffer(buffer, index_buffer.cpu_addr, size);
if constexpr (HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT) {
const u32 new_offset = offset + maxwell3d->regs.index_array.first *
maxwell3d->regs.index_array.FormatSizeInBytes();
const u32 new_offset = offset + maxwell3d->regs.index_buffer.first *
maxwell3d->regs.index_buffer.FormatSizeInBytes();
runtime.BindIndexBuffer(buffer, new_offset, size);
} else {
runtime.BindIndexBuffer(maxwell3d->regs.draw.topology, maxwell3d->regs.index_array.format,
maxwell3d->regs.index_array.first,
maxwell3d->regs.index_array.count, buffer, offset, size);
runtime.BindIndexBuffer(maxwell3d->regs.draw.topology, maxwell3d->regs.index_buffer.format,
maxwell3d->regs.index_buffer.first,
maxwell3d->regs.index_buffer.count, buffer, offset, size);
}
}
@ -1017,7 +1017,7 @@ void BufferCache<P>::BindHostVertexBuffers() {
}
flags[Dirty::VertexBuffer0 + index] = false;
const u32 stride = maxwell3d->regs.vertex_array[index].stride;
const u32 stride = maxwell3d->regs.vertex_streams[index].stride;
const u32 offset = buffer.Offset(binding.cpu_addr);
runtime.BindVertexBuffer(index, buffer, offset, binding.size, stride);
}
@ -1157,7 +1157,7 @@ void BufferCache<P>::BindHostGraphicsTextureBuffers(size_t stage) {
template <class P>
void BufferCache<P>::BindHostTransformFeedbackBuffers() {
if (maxwell3d->regs.tfb_enabled == 0) {
if (maxwell3d->regs.transform_feedback_enabled == 0) {
return;
}
for (u32 index = 0; index < NUM_TRANSFORM_FEEDBACK_BUFFERS; ++index) {
@ -1268,7 +1268,7 @@ template <class P>
void BufferCache<P>::UpdateIndexBuffer() {
// We have to check for the dirty flags and index count
// The index count is currently changed without updating the dirty flags
const auto& index_array = maxwell3d->regs.index_array;
const auto& index_array = maxwell3d->regs.index_buffer;
auto& flags = maxwell3d->dirty.flags;
if (!flags[Dirty::IndexBuffer] && last_index_count == index_array.count) {
return;
@ -1311,10 +1311,10 @@ void BufferCache<P>::UpdateVertexBuffer(u32 index) {
if (!maxwell3d->dirty.flags[Dirty::VertexBuffer0 + index]) {
return;
}
const auto& array = maxwell3d->regs.vertex_array[index];
const auto& limit = maxwell3d->regs.vertex_array_limit[index];
const GPUVAddr gpu_addr_begin = array.StartAddress();
const GPUVAddr gpu_addr_end = limit.LimitAddress() + 1;
const auto& array = maxwell3d->regs.vertex_streams[index];
const auto& limit = maxwell3d->regs.vertex_stream_limits[index];
const GPUVAddr gpu_addr_begin = array.Address();
const GPUVAddr gpu_addr_end = limit.Address() + 1;
const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr_begin);
u32 address_size = static_cast<u32>(
std::min(gpu_addr_end - gpu_addr_begin, static_cast<u64>(std::numeric_limits<u32>::max())));
@ -1380,7 +1380,7 @@ void BufferCache<P>::UpdateTextureBuffers(size_t stage) {
template <class P>
void BufferCache<P>::UpdateTransformFeedbackBuffers() {
if (maxwell3d->regs.tfb_enabled == 0) {
if (maxwell3d->regs.transform_feedback_enabled == 0) {
return;
}
for (u32 index = 0; index < NUM_TRANSFORM_FEEDBACK_BUFFERS; ++index) {
@ -1390,11 +1390,11 @@ void BufferCache<P>::UpdateTransformFeedbackBuffers() {
template <class P>
void BufferCache<P>::UpdateTransformFeedbackBuffer(u32 index) {
const auto& binding = maxwell3d->regs.tfb_bindings[index];
const GPUVAddr gpu_addr = binding.Address() + binding.buffer_offset;
const u32 size = binding.buffer_size;
const auto& binding = maxwell3d->regs.transform_feedback.buffers[index];
const GPUVAddr gpu_addr = binding.Address() + binding.start_offset;
const u32 size = binding.size;
const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
if (binding.buffer_enable == 0 || size == 0 || !cpu_addr) {
if (binding.enable == 0 || size == 0 || !cpu_addr) {
transform_feedback_buffers[index] = NULL_BINDING;
return;
}

24
src/video_core/dirty_flags.cpp
View file

@ -17,21 +17,23 @@ using Tegra::Engines::Maxwell3D;
void SetupDirtyVertexBuffers(Maxwell3D::DirtyState::Tables& tables) {
static constexpr std::size_t num_array = 3;
for (std::size_t i = 0; i < Maxwell3D::Regs::NumVertexArrays; ++i) {
const std::size_t array_offset = OFF(vertex_array) + i * NUM(vertex_array[0]);
const std::size_t limit_offset = OFF(vertex_array_limit) + i * NUM(vertex_array_limit[0]);
const std::size_t array_offset = OFF(vertex_streams) + i * NUM(vertex_streams[0]);
const std::size_t limit_offset =
OFF(vertex_stream_limits) + i * NUM(vertex_stream_limits[0]);
FillBlock(tables, array_offset, num_array, VertexBuffer0 + i, VertexBuffers);
FillBlock(tables, limit_offset, NUM(vertex_array_limit), VertexBuffer0 + i, VertexBuffers);
FillBlock(tables, limit_offset, NUM(vertex_stream_limits), VertexBuffer0 + i,
VertexBuffers);
}
}
void SetupIndexBuffer(Maxwell3D::DirtyState::Tables& tables) {
FillBlock(tables[0], OFF(index_array), NUM(index_array), IndexBuffer);
FillBlock(tables[0], OFF(index_buffer), NUM(index_buffer), IndexBuffer);
}
void SetupDirtyDescriptors(Maxwell3D::DirtyState::Tables& tables) {
FillBlock(tables[0], OFF(tic), NUM(tic), Descriptors);
FillBlock(tables[0], OFF(tsc), NUM(tsc), Descriptors);
FillBlock(tables[0], OFF(tex_header), NUM(tex_header), Descriptors);
FillBlock(tables[0], OFF(tex_sampler), NUM(tex_sampler), Descriptors);
}
void SetupDirtyRenderTargets(Maxwell3D::DirtyState::Tables& tables) {
@ -42,7 +44,7 @@ void SetupDirtyRenderTargets(Maxwell3D::DirtyState::Tables& tables) {
FillBlock(tables[0], begin + rt * num_per_rt, num_per_rt, ColorBuffer0 + rt);
}
FillBlock(tables[1], begin, num, RenderTargets);
FillBlock(tables[0], OFF(render_area), NUM(render_area), RenderTargets);
FillBlock(tables[0], OFF(surface_clip), NUM(surface_clip), RenderTargets);
tables[0][OFF(rt_control)] = RenderTargets;
tables[1][OFF(rt_control)] = RenderTargetControl;
@ -52,15 +54,15 @@ void SetupDirtyRenderTargets(Maxwell3D::DirtyState::Tables& tables) {
const u8 flag = zeta_flags[i];
auto& table = tables[i];
table[OFF(zeta_enable)] = flag;
table[OFF(zeta_width)] = flag;
table[OFF(zeta_height)] = flag;
table[OFF(zeta_size.width)] = flag;
table[OFF(zeta_size.height)] = flag;
FillBlock(table, OFF(zeta), NUM(zeta), flag);
}
}
void SetupDirtyShaders(Maxwell3D::DirtyState::Tables& tables) {
FillBlock(tables[0], OFF(shader_config[0]),
NUM(shader_config[0]) * Maxwell3D::Regs::MaxShaderProgram, Shaders);
FillBlock(tables[0], OFF(pipelines), NUM(pipelines) * Maxwell3D::Regs::MaxShaderProgram,
Shaders);
}
} // Anonymous namespace

334
src/video_core/engines/maxwell_3d.cpp
View file

@ -56,37 +56,37 @@ void Maxwell3D::InitializeRegisterDefaults() {
// Doom and Bomberman seems to use the uninitialized registers and just enable blend
// so initialize blend registers with sane values
regs.blend.equation_rgb = Regs::Blend::Equation::Add;
regs.blend.factor_source_rgb = Regs::Blend::Factor::One;
regs.blend.factor_dest_rgb = Regs::Blend::Factor::Zero;
regs.blend.equation_a = Regs::Blend::Equation::Add;
regs.blend.factor_source_a = Regs::Blend::Factor::One;
regs.blend.factor_dest_a = Regs::Blend::Factor::Zero;
for (auto& blend : regs.independent_blend) {
blend.equation_rgb = Regs::Blend::Equation::Add;
blend.factor_source_rgb = Regs::Blend::Factor::One;
blend.factor_dest_rgb = Regs::Blend::Factor::Zero;
blend.equation_a = Regs::Blend::Equation::Add;
blend.factor_source_a = Regs::Blend::Factor::One;
blend.factor_dest_a = Regs::Blend::Factor::Zero;
regs.blend.color_op = Regs::Blend::Equation::Add_D3D;
regs.blend.color_source = Regs::Blend::Factor::One_D3D;
regs.blend.color_dest = Regs::Blend::Factor::Zero_D3D;
regs.blend.alpha_op = Regs::Blend::Equation::Add_D3D;
regs.blend.alpha_source = Regs::Blend::Factor::One_D3D;
regs.blend.alpha_dest = Regs::Blend::Factor::Zero_D3D;
for (auto& blend : regs.blend_per_target) {
blend.color_op = Regs::Blend::Equation::Add_D3D;
blend.color_source = Regs::Blend::Factor::One_D3D;
blend.color_dest = Regs::Blend::Factor::Zero_D3D;
blend.alpha_op = Regs::Blend::Equation::Add_D3D;
blend.alpha_source = Regs::Blend::Factor::One_D3D;
blend.alpha_dest = Regs::Blend::Factor::Zero_D3D;
}
regs.stencil_front_op_fail = Regs::StencilOp::Keep;
regs.stencil_front_op_zfail = Regs::StencilOp::Keep;
regs.stencil_front_op_zpass = Regs::StencilOp::Keep;
regs.stencil_front_func_func = Regs::ComparisonOp::Always;
regs.stencil_front_func_mask = 0xFFFFFFFF;
regs.stencil_front_mask = 0xFFFFFFFF;
regs.stencil_front_op.fail = Regs::StencilOp::Op::Keep_D3D;
regs.stencil_front_op.zfail = Regs::StencilOp::Op::Keep_D3D;
regs.stencil_front_op.zpass = Regs::StencilOp::Op::Keep_D3D;
regs.stencil_front_op.func = Regs::ComparisonOp::Always_GL;
regs.stencil_front_func.func_mask = 0xFFFFFFFF;
regs.stencil_front_func.mask = 0xFFFFFFFF;
regs.stencil_two_side_enable = 1;
regs.stencil_back_op_fail = Regs::StencilOp::Keep;
regs.stencil_back_op_zfail = Regs::StencilOp::Keep;
regs.stencil_back_op_zpass = Regs::StencilOp::Keep;
regs.stencil_back_func_func = Regs::ComparisonOp::Always;
regs.stencil_back_func_mask = 0xFFFFFFFF;
regs.stencil_back_mask = 0xFFFFFFFF;
regs.stencil_back_op.fail = Regs::StencilOp::Op::Keep_D3D;
regs.stencil_back_op.zfail = Regs::StencilOp::Op::Keep_D3D;
regs.stencil_back_op.zpass = Regs::StencilOp::Op::Keep_D3D;
regs.stencil_back_op.func = Regs::ComparisonOp::Always_GL;
regs.stencil_back_func.func_mask = 0xFFFFFFFF;
regs.stencil_back_func.mask = 0xFFFFFFFF;
regs.depth_test_func = Regs::ComparisonOp::Always;
regs.front_face = Regs::FrontFace::CounterClockWise;
regs.cull_face = Regs::CullFace::Back;
regs.depth_test_func = Regs::ComparisonOp::Always_GL;
regs.gl_front_face = Regs::FrontFace::CounterClockWise;
regs.gl_cull_face = Regs::CullFace::Back;
// TODO(Rodrigo): Most games do not set a point size. I think this is a case of a
// register carrying a default value. Assume it's OpenGL's default (1).
@ -107,20 +107,20 @@ void Maxwell3D::InitializeRegisterDefaults() {
// NVN games expect these values to be enabled at boot
regs.rasterize_enable = 1;
regs.rt_separate_frag_data = 1;
regs.color_target_mrt_enable = 1;
regs.framebuffer_srgb = 1;
regs.line_width_aliased = 1.0f;
regs.line_width_smooth = 1.0f;
regs.front_face = Maxwell3D::Regs::FrontFace::ClockWise;
regs.gl_front_face = Maxwell3D::Regs::FrontFace::ClockWise;
regs.polygon_mode_back = Maxwell3D::Regs::PolygonMode::Fill;
regs.polygon_mode_front = Maxwell3D::Regs::PolygonMode::Fill;
shadow_state = regs;
mme_inline[MAXWELL3D_REG_INDEX(draw.vertex_end_gl)] = true;
mme_inline[MAXWELL3D_REG_INDEX(draw.vertex_begin_gl)] = true;
mme_inline[MAXWELL3D_REG_INDEX(draw.end)] = true;
mme_inline[MAXWELL3D_REG_INDEX(draw.begin)] = true;
mme_inline[MAXWELL3D_REG_INDEX(vertex_buffer.count)] = true;
mme_inline[MAXWELL3D_REG_INDEX(index_array.count)] = true;
mme_inline[MAXWELL3D_REG_INDEX(index_buffer.count)] = true;
}
void Maxwell3D::ProcessMacro(u32 method, const u32* base_start, u32 amount, bool is_last_call) {
@ -173,51 +173,56 @@ void Maxwell3D::ProcessMethodCall(u32 method, u32 argument, u32 nonshadow_argume
case MAXWELL3D_REG_INDEX(shadow_ram_control):
shadow_state.shadow_ram_control = static_cast<Regs::ShadowRamControl>(nonshadow_argument);
return;
case MAXWELL3D_REG_INDEX(macros.upload_address):
return macro_engine->ClearCode(regs.macros.upload_address);
case MAXWELL3D_REG_INDEX(macros.data):
return macro_engine->AddCode(regs.macros.upload_address, argument);
case MAXWELL3D_REG_INDEX(macros.bind):
case MAXWELL3D_REG_INDEX(load_mme.instruction_ptr):
return macro_engine->ClearCode(regs.load_mme.instruction_ptr);
case MAXWELL3D_REG_INDEX(load_mme.instruction):
return macro_engine->AddCode(regs.load_mme.instruction_ptr, argument);
case MAXWELL3D_REG_INDEX(load_mme.start_address):
return ProcessMacroBind(argument);
case MAXWELL3D_REG_INDEX(firmware[4]):
case MAXWELL3D_REG_INDEX(falcon[4]):
return ProcessFirmwareCall4();
case MAXWELL3D_REG_INDEX(const_buffer.cb_data):
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 1:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 2:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 3:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 4:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 5:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 6:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 7:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 8:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 9:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 10:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 11:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 12:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 13:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 14:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 15:
case MAXWELL3D_REG_INDEX(const_buffer.buffer):
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 1:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 2:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 3:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 4:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 5:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 6:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 7:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 8:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 9:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 10:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 11:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 12:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 13:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 14:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 15:
return ProcessCBData(argument);
case MAXWELL3D_REG_INDEX(cb_bind[0]):
case MAXWELL3D_REG_INDEX(bind_groups[0].raw_config):
return ProcessCBBind(0);
case MAXWELL3D_REG_INDEX(cb_bind[1]):
case MAXWELL3D_REG_INDEX(bind_groups[1].raw_config):
return ProcessCBBind(1);
case MAXWELL3D_REG_INDEX(cb_bind[2]):
case MAXWELL3D_REG_INDEX(bind_groups[2].raw_config):
return ProcessCBBind(2);
case MAXWELL3D_REG_INDEX(cb_bind[3]):
case MAXWELL3D_REG_INDEX(bind_groups[3].raw_config):
return ProcessCBBind(3);
case MAXWELL3D_REG_INDEX(cb_bind[4]):
case MAXWELL3D_REG_INDEX(bind_groups[4].raw_config):
return ProcessCBBind(4);
case MAXWELL3D_REG_INDEX(draw.vertex_end_gl):
case MAXWELL3D_REG_INDEX(draw.end):
return DrawArrays();
case MAXWELL3D_REG_INDEX(small_index):
regs.index_array.count = regs.small_index.count;
regs.index_array.first = regs.small_index.first;
case MAXWELL3D_REG_INDEX(index_buffer32_first):
regs.index_buffer.count = regs.index_buffer32_first.count;
regs.index_buffer.first = regs.index_buffer32_first.first;
dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
return DrawArrays();
case MAXWELL3D_REG_INDEX(small_index_2):
regs.index_array.count = regs.small_index_2.count;
regs.index_array.first = regs.small_index_2.first;
case MAXWELL3D_REG_INDEX(index_buffer16_first):
regs.index_buffer.count = regs.index_buffer16_first.count;
regs.index_buffer.first = regs.index_buffer16_first.first;
dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
return DrawArrays();
case MAXWELL3D_REG_INDEX(index_buffer8_first):
regs.index_buffer.count = regs.index_buffer8_first.count;
regs.index_buffer.first = regs.index_buffer8_first.first;
dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
// a macro calls this one over and over, should it increase instancing?
// Used by Hades and likely other Vulkan games.
@ -225,28 +230,24 @@ void Maxwell3D::ProcessMethodCall(u32 method, u32 argument, u32 nonshadow_argume
case MAXWELL3D_REG_INDEX(topology_override):
use_topology_override = true;
return;
case MAXWELL3D_REG_INDEX(clear_buffers):
case MAXWELL3D_REG_INDEX(clear_surface):
return ProcessClearBuffers();
case MAXWELL3D_REG_INDEX(query.query_get):
case MAXWELL3D_REG_INDEX(report_semaphore.query):
return ProcessQueryGet();
case MAXWELL3D_REG_INDEX(condition.mode):
case MAXWELL3D_REG_INDEX(render_enable.mode):
return ProcessQueryCondition();
case MAXWELL3D_REG_INDEX(counter_reset):
case MAXWELL3D_REG_INDEX(clear_report_value):
return ProcessCounterReset();
case MAXWELL3D_REG_INDEX(sync_info):
return ProcessSyncPoint();
case MAXWELL3D_REG_INDEX(exec_upload):
return upload_state.ProcessExec(regs.exec_upload.linear != 0);
case MAXWELL3D_REG_INDEX(data_upload):
case MAXWELL3D_REG_INDEX(launch_dma):
return upload_state.ProcessExec(regs.launch_dma.memory_layout.Value() ==
Regs::LaunchDMA::Layout::Pitch);
case MAXWELL3D_REG_INDEX(inline_data):
upload_state.ProcessData(argument, is_last_call);
return;
case MAXWELL3D_REG_INDEX(fragment_barrier):
return rasterizer->FragmentBarrier();
case MAXWELL3D_REG_INDEX(invalidate_texture_data_cache):
rasterizer->InvalidateGPUCache();
return rasterizer->WaitForIdle();
case MAXWELL3D_REG_INDEX(tiled_cache_barrier):
return rasterizer->TiledCacheBarrier();
}
}
@ -296,25 +297,25 @@ void Maxwell3D::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
return;
}
switch (method) {
case MAXWELL3D_REG_INDEX(const_buffer.cb_data):
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 1:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 2:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 3:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 4:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 5:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 6:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 7:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 8:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 9:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 10:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 11:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 12:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 13:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 14:
case MAXWELL3D_REG_INDEX(const_buffer.cb_data) + 15:
case MAXWELL3D_REG_INDEX(const_buffer.buffer):
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 1:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 2:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 3:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 4:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 5:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 6:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 7:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 8:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 9:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 10:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 11:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 12:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 13:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 14:
case MAXWELL3D_REG_INDEX(const_buffer.buffer) + 15:
ProcessCBMultiData(base_start, amount);
break;
case MAXWELL3D_REG_INDEX(data_upload):
case MAXWELL3D_REG_INDEX(inline_data):
upload_state.ProcessData(base_start, static_cast<size_t>(amount));
return;
default:
@ -353,14 +354,15 @@ void Maxwell3D::CallMethodFromMME(u32 method, u32 method_argument) {
if (mme_inline[method]) {
regs.reg_array[method] = method_argument;
if (method == MAXWELL3D_REG_INDEX(vertex_buffer.count) ||
method == MAXWELL3D_REG_INDEX(index_array.count)) {
method == MAXWELL3D_REG_INDEX(index_buffer.count)) {
const MMEDrawMode expected_mode = method == MAXWELL3D_REG_INDEX(vertex_buffer.count)
? MMEDrawMode::Array
: MMEDrawMode::Indexed;
StepInstance(expected_mode, method_argument);
} else if (method == MAXWELL3D_REG_INDEX(draw.vertex_begin_gl)) {
} else if (method == MAXWELL3D_REG_INDEX(draw.begin)) {
mme_draw.instance_mode =
(regs.draw.instance_next != 0) || (regs.draw.instance_cont != 0);
(regs.draw.instance_id == Maxwell3D::Regs::Draw::InstanceId::Subsequent) ||
(regs.draw.instance_id == Maxwell3D::Regs::Draw::InstanceId::Unchanged);
mme_draw.gl_begin_consume = true;
} else {
mme_draw.gl_end_count++;
@ -405,11 +407,12 @@ void Maxwell3D::ProcessTopologyOverride() {
void Maxwell3D::FlushMMEInlineDraw() {
LOG_TRACE(HW_GPU, "called, topology={}, count={}", regs.draw.topology.Value(),
regs.vertex_buffer.count);
ASSERT_MSG(!(regs.index_array.count && regs.vertex_buffer.count), "Both indexed and direct?");
ASSERT_MSG(!(regs.index_buffer.count && regs.vertex_buffer.count), "Both indexed and direct?");
ASSERT(mme_draw.instance_count == mme_draw.gl_end_count);
// Both instance configuration registers can not be set at the same time.
ASSERT_MSG(!regs.draw.instance_next || !regs.draw.instance_cont,
ASSERT_MSG(regs.draw.instance_id == Maxwell3D::Regs::Draw::InstanceId::First ||
regs.draw.instance_id != Maxwell3D::Regs::Draw::InstanceId::Unchanged,
"Illegal combination of instancing parameters");
ProcessTopologyOverride();
@ -424,7 +427,7 @@ void Maxwell3D::FlushMMEInlineDraw() {
// it's possible that it is incorrect and that there is some other register used to specify the
// drawing mode.
if (is_indexed) {
regs.index_array.count = 0;
regs.index_buffer.count = 0;
} else {
regs.vertex_buffer.count = 0;
}
@ -437,11 +440,11 @@ void Maxwell3D::FlushMMEInlineDraw() {
}
void Maxwell3D::ProcessMacroUpload(u32 data) {
macro_engine->AddCode(regs.macros.upload_address++, data);
macro_engine->AddCode(regs.load_mme.instruction_ptr++, data);
}
void Maxwell3D::ProcessMacroBind(u32 data) {
macro_positions[regs.macros.entry++] = data;
macro_positions[regs.load_mme.start_address_ptr++] = data;
}
void Maxwell3D::ProcessFirmwareCall4() {
@ -449,11 +452,11 @@ void Maxwell3D::ProcessFirmwareCall4() {
// Firmware call 4 is a blob that changes some registers depending on its parameters.
// These registers don't affect emulation and so are stubbed by setting 0xd00 to 1.
regs.reg_array[0xd00] = 1;
regs.shadow_scratch[0] = 1;
}
void Maxwell3D::StampQueryResult(u64 payload, bool long_query) {
const GPUVAddr sequence_address{regs.query.QueryAddress()};
const GPUVAddr sequence_address{regs.report_semaphore.Address()};
if (long_query) {
memory_manager.Write<u64>(sequence_address + sizeof(u64), system.GPU().GetTicks());
memory_manager.Write<u64>(sequence_address, payload);
@ -464,15 +467,17 @@ void Maxwell3D::StampQueryResult(u64 payload, bool long_query) {
void Maxwell3D::ProcessQueryGet() {
// TODO(Subv): Support the other query units.
if (regs.query.query_get.unit != Regs::QueryUnit::Crop) {
LOG_DEBUG(HW_GPU, "Units other than CROP are unimplemented");
if (regs.report_semaphore.query.location != Regs::ReportSemaphore::Location::All) {
LOG_DEBUG(HW_GPU, "Locations other than ALL are unimplemented");
}
switch (regs.query.query_get.operation) {
case Regs::QueryOperation::Release:
if (regs.query.query_get.fence == 1 || regs.query.query_get.short_query != 0) {
const GPUVAddr sequence_address{regs.query.QueryAddress()};
const u32 payload = regs.query.query_sequence;
switch (regs.report_semaphore.query.operation) {
case Regs::ReportSemaphore::Operation::Release:
if (regs.report_semaphore.query.release ==
Regs::ReportSemaphore::Release::AfterAllPreceedingWrites ||
regs.report_semaphore.query.short_query != 0) {
const GPUVAddr sequence_address{regs.report_semaphore.Address()};
const u32 payload = regs.report_semaphore.payload;
std::function<void()> operation([this, sequence_address, payload] {
memory_manager.Write<u32>(sequence_address, payload);
});
@ -482,8 +487,8 @@ void Maxwell3D::ProcessQueryGet() {
u64_le value;
u64_le timestamp;
};
const GPUVAddr sequence_address{regs.query.QueryAddress()};
const u32 payload = regs.query.query_sequence;
const GPUVAddr sequence_address{regs.report_semaphore.Address()};
const u32 payload = regs.report_semaphore.payload;
std::function<void()> operation([this, sequence_address, payload] {
memory_manager.Write<u64>(sequence_address + sizeof(u64), system.GPU().GetTicks());
memory_manager.Write<u64>(sequence_address, payload);
@ -491,19 +496,19 @@ void Maxwell3D::ProcessQueryGet() {
rasterizer->SyncOperation(std::move(operation));
}
break;
case Regs::QueryOperation::Acquire:
case Regs::ReportSemaphore::Operation::Acquire:
// TODO(Blinkhawk): Under this operation, the GPU waits for the CPU to write a value that
// matches the current payload.
UNIMPLEMENTED_MSG("Unimplemented query operation ACQUIRE");
break;
case Regs::QueryOperation::Counter:
case Regs::ReportSemaphore::Operation::ReportOnly:
if (const std::optional<u64> result = GetQueryResult()) {
// If the query returns an empty optional it means it's cached and deferred.
// In this case we have a non-empty result, so we stamp it immediately.
StampQueryResult(*result, regs.query.query_get.short_query == 0);
StampQueryResult(*result, regs.report_semaphore.query.short_query == 0);
}
break;
case Regs::QueryOperation::Trap:
case Regs::ReportSemaphore::Operation::Trap:
UNIMPLEMENTED_MSG("Unimplemented query operation TRAP");
break;
default:
@ -513,31 +518,31 @@ void Maxwell3D::ProcessQueryGet() {
}
void Maxwell3D::ProcessQueryCondition() {
const GPUVAddr condition_address{regs.condition.Address()};
switch (regs.condition.mode) {
case Regs::ConditionMode::Always: {
const GPUVAddr condition_address{regs.render_enable.Address()};
switch (regs.render_enable.mode) {
case Regs::RenderEnable::Mode::True: {
execute_on = true;
break;
}
case Regs::ConditionMode::Never: {
case Regs::RenderEnable::Mode::False: {
execute_on = false;
break;
}
case Regs::ConditionMode::ResNonZero: {
Regs::QueryCompare cmp;
case Regs::RenderEnable::Mode::Conditional: {
Regs::ReportSemaphore::Compare cmp;
memory_manager.ReadBlock(condition_address, &cmp, sizeof(cmp));
execute_on = cmp.initial_sequence != 0U && cmp.initial_mode != 0U;
break;
}
case Regs::ConditionMode::Equal: {
Regs::QueryCompare cmp;
case Regs::RenderEnable::Mode::IfEqual: {
Regs::ReportSemaphore::Compare cmp;
memory_manager.ReadBlock(condition_address, &cmp, sizeof(cmp));
execute_on =
cmp.initial_sequence == cmp.current_sequence && cmp.initial_mode == cmp.current_mode;
break;
}
case Regs::ConditionMode::NotEqual: {
Regs::QueryCompare cmp;
case Regs::RenderEnable::Mode::IfNotEqual: {
Regs::ReportSemaphore::Compare cmp;
memory_manager.ReadBlock(condition_address, &cmp, sizeof(cmp));
execute_on =
cmp.initial_sequence != cmp.current_sequence || cmp.initial_mode != cmp.current_mode;
@ -552,21 +557,21 @@ void Maxwell3D::ProcessQueryCondition() {
}
void Maxwell3D::ProcessCounterReset() {
switch (regs.counter_reset) {
case Regs::CounterReset::SampleCnt:
switch (regs.clear_report_value) {
case Regs::ClearReport::ZPassPixelCount:
rasterizer->ResetCounter(QueryType::SamplesPassed);
break;
default:
LOG_DEBUG(Render_OpenGL, "Unimplemented counter reset={}", regs.counter_reset);
LOG_DEBUG(Render_OpenGL, "Unimplemented counter reset={}", regs.clear_report_value);
break;
}
}
void Maxwell3D::ProcessSyncPoint() {
const u32 sync_point = regs.sync_info.sync_point.Value();
const u32 increment = regs.sync_info.increment.Value();
[[maybe_unused]] const u32 cache_flush = regs.sync_info.unknown.Value();
if (increment) {
const auto condition = regs.sync_info.condition.Value();
[[maybe_unused]] const u32 cache_flush = regs.sync_info.clean_l2.Value();
if (condition == Regs::SyncInfo::Condition::RopWritesDone) {
rasterizer->SignalSyncPoint(sync_point);
}
}
@ -574,23 +579,24 @@ void Maxwell3D::ProcessSyncPoint() {
void Maxwell3D::DrawArrays() {
LOG_TRACE(HW_GPU, "called, topology={}, count={}", regs.draw.topology.Value(),
regs.vertex_buffer.count);
ASSERT_MSG(!(regs.index_array.count && regs.vertex_buffer.count), "Both indexed and direct?");
ASSERT_MSG(!(regs.index_buffer.count && regs.vertex_buffer.count), "Both indexed and direct?");
// Both instance configuration registers can not be set at the same time.
ASSERT_MSG(!regs.draw.instance_next || !regs.draw.instance_cont,
ASSERT_MSG(regs.draw.instance_id == Maxwell3D::Regs::Draw::InstanceId::First ||
regs.draw.instance_id != Maxwell3D::Regs::Draw::InstanceId::Unchanged,
"Illegal combination of instancing parameters");
ProcessTopologyOverride();
if (regs.draw.instance_next) {
if (regs.draw.instance_id == Maxwell3D::Regs::Draw::InstanceId::Subsequent) {
// Increment the current instance *before* drawing.
state.current_instance += 1;
} else if (!regs.draw.instance_cont) {
state.current_instance++;
} else if (regs.draw.instance_id != Maxwell3D::Regs::Draw::InstanceId::Unchanged) {
// Reset the current instance to 0.
state.current_instance = 0;
}
const bool is_indexed{regs.index_array.count && !regs.vertex_buffer.count};
const bool is_indexed{regs.index_buffer.count && !regs.vertex_buffer.count};
if (ShouldExecute()) {
rasterizer->Draw(is_indexed, false);
}
@ -600,60 +606,60 @@ void Maxwell3D::DrawArrays() {
// it's possible that it is incorrect and that there is some other register used to specify the
// drawing mode.
if (is_indexed) {
regs.index_array.count = 0;
regs.index_buffer.count = 0;
} else {
regs.vertex_buffer.count = 0;
}
}
std::optional<u64> Maxwell3D::GetQueryResult() {
switch (regs.query.query_get.select) {
case Regs::QuerySelect::Payload:
return regs.query.query_sequence;
case Regs::QuerySelect::SamplesPassed:
switch (regs.report_semaphore.query.report) {
case Regs::ReportSemaphore::Report::Payload:
return regs.report_semaphore.payload;
case Regs::ReportSemaphore::Report::ZPassPixelCount64:
// Deferred.
rasterizer->Query(regs.query.QueryAddress(), QueryType::SamplesPassed,
rasterizer->Query(regs.report_semaphore.Address(), QueryType::SamplesPassed,
system.GPU().GetTicks());
return std::nullopt;
default:
LOG_DEBUG(HW_GPU, "Unimplemented query select type {}",
regs.query.query_get.select.Value());
LOG_DEBUG(HW_GPU, "Unimplemented query report type {}",
regs.report_semaphore.query.report.Value());
return 1;
}
}
void Maxwell3D::ProcessCBBind(size_t stage_index) {
// Bind the buffer currently in CB_ADDRESS to the specified index in the desired shader stage.
const auto& bind_data = regs.cb_bind[stage_index];
auto& buffer = state.shader_stages[stage_index].const_buffers[bind_data.index];
const auto& bind_data = regs.bind_groups[stage_index];
auto& buffer = state.shader_stages[stage_index].const_buffers[bind_data.shader_slot];
buffer.enabled = bind_data.valid.Value() != 0;
buffer.address = regs.const_buffer.BufferAddress();
buffer.size = regs.const_buffer.cb_size;
buffer.address = regs.const_buffer.Address();
buffer.size = regs.const_buffer.size;
const bool is_enabled = bind_data.valid.Value() != 0;
if (!is_enabled) {
rasterizer->DisableGraphicsUniformBuffer(stage_index, bind_data.index);
rasterizer->DisableGraphicsUniformBuffer(stage_index, bind_data.shader_slot);
return;
}
const GPUVAddr gpu_addr = regs.const_buffer.BufferAddress();
const u32 size = regs.const_buffer.cb_size;
rasterizer->BindGraphicsUniformBuffer(stage_index, bind_data.index, gpu_addr, size);
const GPUVAddr gpu_addr = regs.const_buffer.Address();
const u32 size = regs.const_buffer.size;
rasterizer->BindGraphicsUniformBuffer(stage_index, bind_data.shader_slot, gpu_addr, size);
}
void Maxwell3D::ProcessCBMultiData(const u32* start_base, u32 amount) {
// Write the input value to the current const buffer at the current position.
const GPUVAddr buffer_address = regs.const_buffer.BufferAddress();
const GPUVAddr buffer_address = regs.const_buffer.Address();
ASSERT(buffer_address != 0);
// Don't allow writing past the end of the buffer.
ASSERT(regs.const_buffer.cb_pos <= regs.const_buffer.cb_size);
ASSERT(regs.const_buffer.offset <= regs.const_buffer.size);
const GPUVAddr address{buffer_address + regs.const_buffer.cb_pos};
const GPUVAddr address{buffer_address + regs.const_buffer.offset};
const size_t copy_size = amount * sizeof(u32);
memory_manager.WriteBlock(address, start_base, copy_size);
// Increment the current buffer position.
regs.const_buffer.cb_pos += static_cast<u32>(copy_size);
regs.const_buffer.offset += static_cast<u32>(copy_size);
}
void Maxwell3D::ProcessCBData(u32 value) {
@ -661,7 +667,8 @@ void Maxwell3D::ProcessCBData(u32 value) {
}
Texture::TICEntry Maxwell3D::GetTICEntry(u32 tic_index) const {
const GPUVAddr tic_address_gpu{regs.tic.Address() + tic_index * sizeof(Texture::TICEntry)};
const GPUVAddr tic_address_gpu{regs.tex_header.Address() +
tic_index * sizeof(Texture::TICEntry)};
Texture::TICEntry tic_entry;
memory_manager.ReadBlockUnsafe(tic_address_gpu, &tic_entry, sizeof(Texture::TICEntry));
@ -670,7 +677,8 @@ Texture::TICEntry Maxwell3D::GetTICEntry(u32 tic_index) const {
}
Texture::TSCEntry Maxwell3D::GetTSCEntry(u32 tsc_index) const {
const GPUVAddr tsc_address_gpu{regs.tsc.Address() + tsc_index * sizeof(Texture::TSCEntry)};
const GPUVAddr tsc_address_gpu{regs.tex_sampler.Address() +
tsc_index * sizeof(Texture::TSCEntry)};
Texture::TSCEntry tsc_entry;
memory_manager.ReadBlockUnsafe(tsc_address_gpu, &tsc_entry, sizeof(Texture::TSCEntry));

3965
src/video_core/engines/maxwell_3d.h
View file

File diff suppressed because it is too large Load diff

50
src/video_core/gpu.h
View file

@ -24,11 +24,15 @@ namespace Tegra {
class DmaPusher;
struct CommandList;
// TODO: Implement the commented ones
enum class RenderTargetFormat : u32 {
NONE = 0x0,
R32B32G32A32_FLOAT = 0xC0,
R32G32B32A32_SINT = 0xC1,
R32G32B32A32_UINT = 0xC2,
// R32G32B32X32_FLOAT = 0xC3,
// R32G32B32X32_SINT = 0xC4,
// R32G32B32X32_UINT = 0xC5,
R16G16B16A16_UNORM = 0xC6,
R16G16B16A16_SNORM = 0xC7,
R16G16B16A16_SINT = 0xC8,
@ -38,8 +42,8 @@ enum class RenderTargetFormat : u32 {
R32G32_SINT = 0xCC,
R32G32_UINT = 0xCD,
R16G16B16X16_FLOAT = 0xCE,
B8G8R8A8_UNORM = 0xCF,
B8G8R8A8_SRGB = 0xD0,
A8R8G8B8_UNORM = 0xCF,
A8R8G8B8_SRGB = 0xD0,
A2B10G10R10_UNORM = 0xD1,
A2B10G10R10_UINT = 0xD2,
A8B8G8R8_UNORM = 0xD5,
@ -52,10 +56,13 @@ enum class RenderTargetFormat : u32 {
R16G16_SINT = 0xDC,
R16G16_UINT = 0xDD,
R16G16_FLOAT = 0xDE,
// A2R10G10B10_UNORM = 0xDF,
B10G11R11_FLOAT = 0xE0,
R32_SINT = 0xE3,
R32_UINT = 0xE4,
R32_FLOAT = 0xE5,
// X8R8G8B8_UNORM = 0xE6,
// X8R8G8B8_SRGB = 0xE7,
R5G6B5_UNORM = 0xE8,
A1R5G5B5_UNORM = 0xE9,
R8G8_UNORM = 0xEA,
@ -71,17 +78,42 @@ enum class RenderTargetFormat : u32 {
R8_SNORM = 0xF4,
R8_SINT = 0xF5,
R8_UINT = 0xF6,
/*
A8_UNORM = 0xF7,
X1R5G5B5_UNORM = 0xF8,
X8B8G8R8_UNORM = 0xF9,
X8B8G8R8_SRGB = 0xFA,
Z1R5G5B5_UNORM = 0xFB,
O1R5G5B5_UNORM = 0xFC,
Z8R8G8B8_UNORM = 0xFD,
O8R8G8B8_UNORM = 0xFE,
R32_UNORM = 0xFF,
A16_UNORM = 0x40,
A16_FLOAT = 0x41,
A32_FLOAT = 0x42,
A8R8_UNORM = 0x43,
R16A16_UNORM = 0x44,
R16A16_FLOAT = 0x45,
R32A32_FLOAT = 0x46,
B8G8R8A8_UNORM = 0x47,
*/
};
enum class DepthFormat : u32 {
D32_FLOAT = 0xA,
D16_UNORM = 0x13,
S8_UINT_Z24_UNORM = 0x14,
D24X8_UNORM = 0x15,
D24S8_UNORM = 0x16,
Z32_FLOAT = 0xA,
Z16_UNORM = 0x13,
Z24_UNORM_S8_UINT = 0x14,
X8Z24_UNORM = 0x15,
S8Z24_UNORM = 0x16,
S8_UINT = 0x17,
D24C8_UNORM = 0x18,
D32_FLOAT_S8X24_UINT = 0x19,
V8Z24_UNORM = 0x18,
Z32_FLOAT_X24S8_UINT = 0x19,
/*
X8Z24_UNORM_X16V8S8_UINT = 0x1D,
Z32_FLOAT_X16V8X8_UINT = 0x1E,
Z32_FLOAT_X16V8S8_UINT = 0x1F,
*/
};
namespace Engines {

48
src/video_core/macro/macro_hle.cpp
View file

@ -21,16 +21,16 @@ void HLE_771BB18C62444DA0(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
maxwell3d.regs.draw.topology.Assign(
static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[0] & 0x3ffffff));
maxwell3d.regs.vb_base_instance = parameters[5];
maxwell3d.regs.global_base_instance_index = parameters[5];
maxwell3d.mme_draw.instance_count = instance_count;
maxwell3d.regs.vb_element_base = parameters[3];
maxwell3d.regs.index_array.count = parameters[1];
maxwell3d.regs.index_array.first = parameters[4];
maxwell3d.regs.global_base_vertex_index = parameters[3];
maxwell3d.regs.index_buffer.count = parameters[1];
maxwell3d.regs.index_buffer.first = parameters[4];
if (maxwell3d.ShouldExecute()) {
maxwell3d.Rasterizer().Draw(true, true);
}
maxwell3d.regs.index_array.count = 0;
maxwell3d.regs.index_buffer.count = 0;
maxwell3d.mme_draw.instance_count = 0;
maxwell3d.mme_draw.current_mode = Engines::Maxwell3D::MMEDrawMode::Undefined;
}
@ -40,7 +40,7 @@ void HLE_0D61FC9FAAC9FCAD(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
maxwell3d.regs.vertex_buffer.first = parameters[3];
maxwell3d.regs.vertex_buffer.count = parameters[1];
maxwell3d.regs.vb_base_instance = parameters[4];
maxwell3d.regs.global_base_instance_index = parameters[4];
maxwell3d.regs.draw.topology.Assign(
static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[0]));
maxwell3d.mme_draw.instance_count = count;
@ -57,12 +57,12 @@ void HLE_0217920100488FF7(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
const u32 instance_count = (maxwell3d.GetRegisterValue(0xD1B) & parameters[2]);
const u32 element_base = parameters[4];
const u32 base_instance = parameters[5];
maxwell3d.regs.index_array.first = parameters[3];
maxwell3d.regs.reg_array[0x446] = element_base; // vertex id base?
maxwell3d.regs.index_array.count = parameters[1];
maxwell3d.regs.index_buffer.first = parameters[3];
maxwell3d.regs.vertex_id_base = element_base;
maxwell3d.regs.index_buffer.count = parameters[1];
maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
maxwell3d.regs.vb_element_base = element_base;
maxwell3d.regs.vb_base_instance = base_instance;
maxwell3d.regs.global_base_vertex_index = element_base;
maxwell3d.regs.global_base_instance_index = base_instance;
maxwell3d.mme_draw.instance_count = instance_count;
maxwell3d.CallMethodFromMME(0x8e3, 0x640);
maxwell3d.CallMethodFromMME(0x8e4, element_base);
@ -72,10 +72,10 @@ void HLE_0217920100488FF7(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
if (maxwell3d.ShouldExecute()) {
maxwell3d.Rasterizer().Draw(true, true);
}
maxwell3d.regs.reg_array[0x446] = 0x0; // vertex id base?
maxwell3d.regs.index_array.count = 0;
maxwell3d.regs.vb_element_base = 0x0;
maxwell3d.regs.vb_base_instance = 0x0;
maxwell3d.regs.vertex_id_base = 0x0;
maxwell3d.regs.index_buffer.count = 0;
maxwell3d.regs.global_base_vertex_index = 0x0;
maxwell3d.regs.global_base_instance_index = 0x0;
maxwell3d.mme_draw.instance_count = 0;
maxwell3d.CallMethodFromMME(0x8e3, 0x640);
maxwell3d.CallMethodFromMME(0x8e4, 0x0);
@ -87,10 +87,10 @@ void HLE_0217920100488FF7(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
void HLE_3F5E74B9C9A50164(Engines::Maxwell3D& maxwell3d, const std::vector<u32>& parameters) {
SCOPE_EXIT({
// Clean everything.
maxwell3d.regs.reg_array[0x446] = 0x0; // vertex id base?
maxwell3d.regs.index_array.count = 0;
maxwell3d.regs.vb_element_base = 0x0;
maxwell3d.regs.vb_base_instance = 0x0;
maxwell3d.regs.vertex_id_base = 0x0;
maxwell3d.regs.index_buffer.count = 0;
maxwell3d.regs.global_base_vertex_index = 0x0;
maxwell3d.regs.global_base_instance_index = 0x0;
maxwell3d.mme_draw.instance_count = 0;
maxwell3d.CallMethodFromMME(0x8e3, 0x640);
maxwell3d.CallMethodFromMME(0x8e4, 0x0);
@ -122,11 +122,11 @@ void HLE_3F5E74B9C9A50164(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
const u32 first_index = parameters[base + 2];
const u32 base_vertex = parameters[base + 3];
const u32 base_instance = parameters[base + 4];
maxwell3d.regs.index_array.first = first_index;
maxwell3d.regs.reg_array[0x446] = base_vertex;
maxwell3d.regs.index_array.count = num_vertices;
maxwell3d.regs.vb_element_base = base_vertex;
maxwell3d.regs.vb_base_instance = base_instance;
maxwell3d.regs.index_buffer.first = first_index;
maxwell3d.regs.vertex_id_base = base_vertex;
maxwell3d.regs.index_buffer.count = num_vertices;
maxwell3d.regs.global_base_vertex_index = base_vertex;
maxwell3d.regs.global_base_instance_index = base_instance;
maxwell3d.mme_draw.instance_count = instance_count;
maxwell3d.CallMethodFromMME(0x8e3, 0x640);
maxwell3d.CallMethodFromMME(0x8e4, base_vertex);

2
src/video_core/query_cache.h
View file

@ -137,7 +137,7 @@ public:
std::unique_lock lock{mutex};
if (maxwell3d) {
const auto& regs = maxwell3d->regs;
Stream(VideoCore::QueryType::SamplesPassed).Update(regs.samplecnt_enable);
Stream(VideoCore::QueryType::SamplesPassed).Update(regs.zpass_pixel_count_enable);
}
}

29
src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
View file

@ -73,8 +73,8 @@ GLenum AssemblyStage(size_t stage_index) {
/// @param location Hardware location
/// @return Pair of ARB_transform_feedback3 token stream first and third arguments
/// @note Read https://www.khronos.org/registry/OpenGL/extensions/ARB/ARB_transform_feedback3.txt
std::pair<GLint, GLint> TransformFeedbackEnum(u8 location) {
const u8 index = location / 4;
std::pair<GLint, GLint> TransformFeedbackEnum(u32 location) {
const auto index = location / 4;
if (index >= 8 && index <= 39) {
return {GL_GENERIC_ATTRIB_NV, index - 8};
}
@ -286,7 +286,7 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
buffer_cache.runtime.SetEnableStorageBuffers(use_storage_buffers);
const auto& regs{maxwell3d->regs};
const bool via_header_index{regs.sampler_index == Maxwell::SamplerIndex::ViaHeaderIndex};
const bool via_header_index{regs.sampler_binding == Maxwell::SamplerBinding::ViaHeaderBinding};
const auto config_stage{[&](size_t stage) LAMBDA_FORCEINLINE {
const Shader::Info& info{stage_infos[stage]};
buffer_cache.UnbindGraphicsStorageBuffers(stage);
@ -557,10 +557,25 @@ void GraphicsPipeline::GenerateTransformFeedbackState() {
++current_stream;
const auto& locations = key.xfb_state.varyings[feedback];
std::optional<u8> current_index;
std::optional<u32> current_index;
for (u32 offset = 0; offset < layout.varying_count; ++offset) {
const u8 location = locations[offset];
const u8 index = location / 4;
const auto get_attribute = [&locations](u32 index) -> u32 {
switch (index % 4) {
case 0:
return locations[index / 4].attribute0.Value();
case 1:
return locations[index / 4].attribute1.Value();
case 2:
return locations[index / 4].attribute2.Value();
case 3:
return locations[index / 4].attribute3.Value();
}
UNREACHABLE();
return 0;
};
const auto attribute{get_attribute(offset)};
const auto index = attribute / 4U;
if (current_index == index) {
// Increase number of components of the previous attachment
@ -569,7 +584,7 @@ void GraphicsPipeline::GenerateTransformFeedbackState() {
}
current_index = index;
std::tie(cursor[0], cursor[2]) = TransformFeedbackEnum(location);
std::tie(cursor[0], cursor[2]) = TransformFeedbackEnum(attribute);
cursor[1] = 1;
cursor += XFB_ENTRY_STRIDE;
}

4
src/video_core/renderer_opengl/gl_graphics_pipeline.h
View file

@ -37,8 +37,8 @@ struct GraphicsPipelineKey {
BitField<0, 1, u32> xfb_enabled;
BitField<1, 1, u32> early_z;
BitField<2, 4, Maxwell::PrimitiveTopology> gs_input_topology;
BitField<6, 2, Maxwell::TessellationPrimitive> tessellation_primitive;
BitField<8, 2, Maxwell::TessellationSpacing> tessellation_spacing;
BitField<6, 2, Maxwell::Tessellation::DomainType> tessellation_primitive;
BitField<8, 2, Maxwell::Tessellation::Spacing> tessellation_spacing;
BitField<10, 1, u32> tessellation_clockwise;
};
std::array<u32, 3> padding;

140
src/video_core/renderer_opengl/gl_rasterizer.cpp
View file

@ -87,7 +87,7 @@ void RasterizerOpenGL::SyncVertexFormats() {
}
flags[Dirty::VertexFormat0 + index] = false;
const auto attrib = maxwell3d->regs.vertex_attrib_format[index];
const auto& attrib = maxwell3d->regs.vertex_attrib_format[index];
const auto gl_index = static_cast<GLuint>(index);
// Disable constant attributes.
@ -97,8 +97,8 @@ void RasterizerOpenGL::SyncVertexFormats() {
}
glEnableVertexAttribArray(gl_index);
if (attrib.type == Maxwell::VertexAttribute::Type::SignedInt ||
attrib.type == Maxwell::VertexAttribute::Type::UnsignedInt) {
if (attrib.type == Maxwell::VertexAttribute::Type::SInt ||
attrib.type == Maxwell::VertexAttribute::Type::UInt) {
glVertexAttribIFormat(gl_index, attrib.ComponentCount(),
MaxwellToGL::VertexFormat(attrib), attrib.offset);
} else {
@ -125,8 +125,8 @@ void RasterizerOpenGL::SyncVertexInstances() {
flags[Dirty::VertexInstance0 + index] = false;
const auto gl_index = static_cast<GLuint>(index);
const bool instancing_enabled = regs.instanced_arrays.IsInstancingEnabled(gl_index);
const GLuint divisor = instancing_enabled ? regs.vertex_array[index].divisor : 0;
const bool instancing_enabled = regs.vertex_stream_instances.IsInstancingEnabled(gl_index);
const GLuint divisor = instancing_enabled ? regs.vertex_streams[index].frequency : 0;
glVertexBindingDivisor(gl_index, divisor);
}
}
@ -147,27 +147,27 @@ void RasterizerOpenGL::Clear() {
bool use_depth{};
bool use_stencil{};
if (regs.clear_buffers.R || regs.clear_buffers.G || regs.clear_buffers.B ||
regs.clear_buffers.A) {
if (regs.clear_surface.R || regs.clear_surface.G || regs.clear_surface.B ||
regs.clear_surface.A) {
use_color = true;
const GLuint index = regs.clear_buffers.RT;
const GLuint index = regs.clear_surface.RT;
state_tracker.NotifyColorMask(index);
glColorMaski(index, regs.clear_buffers.R != 0, regs.clear_buffers.G != 0,
regs.clear_buffers.B != 0, regs.clear_buffers.A != 0);
glColorMaski(index, regs.clear_surface.R != 0, regs.clear_surface.G != 0,
regs.clear_surface.B != 0, regs.clear_surface.A != 0);
// TODO(Rodrigo): Determine if clamping is used on clears
SyncFragmentColorClampState();
SyncFramebufferSRGB();
}
if (regs.clear_buffers.Z) {
if (regs.clear_surface.Z) {
ASSERT_MSG(regs.zeta_enable != 0, "Tried to clear Z but buffer is not enabled!");
use_depth = true;
state_tracker.NotifyDepthMask();
glDepthMask(GL_TRUE);
}
if (regs.clear_buffers.S) {
if (regs.clear_surface.S) {
ASSERT_MSG(regs.zeta_enable, "Tried to clear stencil but buffer is not enabled!");
use_stencil = true;
}
@ -184,16 +184,16 @@ void RasterizerOpenGL::Clear() {
texture_cache.UpdateRenderTargets(true);
state_tracker.BindFramebuffer(texture_cache.GetFramebuffer()->Handle());
SyncViewport();
if (regs.clear_flags.scissor) {
if (regs.clear_control.use_scissor) {
SyncScissorTest();
} else {
state_tracker.NotifyScissor0();
glDisablei(GL_SCISSOR_TEST, 0);
}
UNIMPLEMENTED_IF(regs.clear_flags.viewport);
UNIMPLEMENTED_IF(regs.clear_control.use_viewport_clip0);
if (use_color) {
glClearBufferfv(GL_COLOR, regs.clear_buffers.RT, regs.clear_color);
glClearBufferfv(GL_COLOR, regs.clear_surface.RT, regs.clear_color.data());
}
if (use_depth && use_stencil) {
glClearBufferfi(GL_DEPTH_STENCIL, 0, regs.clear_depth, regs.clear_stencil);
@ -227,14 +227,14 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
const GLenum primitive_mode = MaxwellToGL::PrimitiveTopology(maxwell3d->regs.draw.topology);
BeginTransformFeedback(pipeline, primitive_mode);
const GLuint base_instance = static_cast<GLuint>(maxwell3d->regs.vb_base_instance);
const GLuint base_instance = static_cast<GLuint>(maxwell3d->regs.global_base_instance_index);
const GLsizei num_instances =
static_cast<GLsizei>(is_instanced ? maxwell3d->mme_draw.instance_count : 1);
if (is_indexed) {
const GLint base_vertex = static_cast<GLint>(maxwell3d->regs.vb_element_base);
const GLsizei num_vertices = static_cast<GLsizei>(maxwell3d->regs.index_array.count);
const GLint base_vertex = static_cast<GLint>(maxwell3d->regs.global_base_vertex_index);
const GLsizei num_vertices = static_cast<GLsizei>(maxwell3d->regs.index_buffer.count);
const GLvoid* const offset = buffer_cache_runtime.IndexOffset();
const GLenum format = MaxwellToGL::IndexFormat(maxwell3d->regs.index_array.format);
const GLenum format = MaxwellToGL::IndexFormat(maxwell3d->regs.index_buffer.format);
if (num_instances == 1 && base_instance == 0 && base_vertex == 0) {
glDrawElements(primitive_mode, num_vertices, format, offset);
} else if (num_instances == 1 && base_instance == 0) {
@ -555,9 +555,9 @@ void RasterizerOpenGL::SyncViewport() {
if (dirty_viewport || dirty_clip_control || flags[Dirty::FrontFace]) {
flags[Dirty::FrontFace] = false;
GLenum mode = MaxwellToGL::FrontFace(regs.front_face);
GLenum mode = MaxwellToGL::FrontFace(regs.gl_front_face);
bool flip_faces = true;
if (regs.screen_y_control.triangle_rast_flip != 0) {
if (regs.window_origin.flip_y != 0) {
flip_faces = !flip_faces;
}
if (regs.viewport_transform[0].scale_y < 0.0f) {
@ -582,14 +582,15 @@ void RasterizerOpenGL::SyncViewport() {
if (regs.viewport_transform[0].scale_y < 0.0f) {
flip_y = !flip_y;
}
if (regs.screen_y_control.y_negate != 0) {
const bool lower_left{regs.window_origin.mode != Maxwell::WindowOrigin::Mode::UpperLeft};
if (lower_left) {
flip_y = !flip_y;
}
const bool is_zero_to_one = regs.depth_mode == Maxwell::DepthMode::ZeroToOne;
const GLenum origin = flip_y ? GL_UPPER_LEFT : GL_LOWER_LEFT;
const GLenum depth = is_zero_to_one ? GL_ZERO_TO_ONE : GL_NEGATIVE_ONE_TO_ONE;
state_tracker.ClipControl(origin, depth);
state_tracker.SetYNegate(regs.screen_y_control.y_negate != 0);
state_tracker.SetYNegate(lower_left);
}
const bool is_rescaling{texture_cache.IsRescaling()};
const float scale = is_rescaling ? Settings::values.resolution_info.up_factor : 1.0f;
@ -657,7 +658,8 @@ void RasterizerOpenGL::SyncDepthClamp() {
}
flags[Dirty::DepthClampEnabled] = false;
oglEnable(GL_DEPTH_CLAMP, maxwell3d->regs.view_volume_clip_control.depth_clamp_disabled == 0);
oglEnable(GL_DEPTH_CLAMP, maxwell3d->regs.viewport_clip_control.geometry_clip !=
Maxwell::ViewportClipControl::GeometryClip::Passthrough);
}
void RasterizerOpenGL::SyncClipEnabled(u32 clip_mask) {
@ -667,7 +669,7 @@ void RasterizerOpenGL::SyncClipEnabled(u32 clip_mask) {
}
flags[Dirty::ClipDistances] = false;
clip_mask &= maxwell3d->regs.clip_distance_enabled;
clip_mask &= maxwell3d->regs.user_clip_enable.raw;
if (clip_mask == last_clip_distance_mask) {
return;
}
@ -689,9 +691,9 @@ void RasterizerOpenGL::SyncCullMode() {
if (flags[Dirty::CullTest]) {
flags[Dirty::CullTest] = false;
if (regs.cull_test_enabled) {
if (regs.gl_cull_test_enabled) {
glEnable(GL_CULL_FACE);
glCullFace(MaxwellToGL::CullFace(regs.cull_face));
glCullFace(MaxwellToGL::CullFace(regs.gl_cull_face));
} else {
glDisable(GL_CULL_FACE);
}
@ -743,20 +745,20 @@ void RasterizerOpenGL::SyncStencilTestState() {
const auto& regs = maxwell3d->regs;
oglEnable(GL_STENCIL_TEST, regs.stencil_enable);
glStencilFuncSeparate(GL_FRONT, MaxwellToGL::ComparisonOp(regs.stencil_front_func_func),
regs.stencil_front_func_ref, regs.stencil_front_func_mask);
glStencilOpSeparate(GL_FRONT, MaxwellToGL::StencilOp(regs.stencil_front_op_fail),
MaxwellToGL::StencilOp(regs.stencil_front_op_zfail),
MaxwellToGL::StencilOp(regs.stencil_front_op_zpass));
glStencilMaskSeparate(GL_FRONT, regs.stencil_front_mask);
glStencilFuncSeparate(GL_FRONT, MaxwellToGL::ComparisonOp(regs.stencil_front_op.func),
regs.stencil_front_func.ref, regs.stencil_front_func.func_mask);
glStencilOpSeparate(GL_FRONT, MaxwellToGL::StencilOp(regs.stencil_front_op.fail),
MaxwellToGL::StencilOp(regs.stencil_front_op.zfail),
MaxwellToGL::StencilOp(regs.stencil_front_op.zpass));
glStencilMaskSeparate(GL_FRONT, regs.stencil_front_func.mask);
if (regs.stencil_two_side_enable) {
glStencilFuncSeparate(GL_BACK, MaxwellToGL::ComparisonOp(regs.stencil_back_func_func),
regs.stencil_back_func_ref, regs.stencil_back_func_mask);
glStencilOpSeparate(GL_BACK, MaxwellToGL::StencilOp(regs.stencil_back_op_fail),
MaxwellToGL::StencilOp(regs.stencil_back_op_zfail),
MaxwellToGL::StencilOp(regs.stencil_back_op_zpass));
glStencilMaskSeparate(GL_BACK, regs.stencil_back_mask);
glStencilFuncSeparate(GL_BACK, MaxwellToGL::ComparisonOp(regs.stencil_back_op.func),
regs.stencil_back_func.ref, regs.stencil_back_func.mask);
glStencilOpSeparate(GL_BACK, MaxwellToGL::StencilOp(regs.stencil_back_op.fail),
MaxwellToGL::StencilOp(regs.stencil_back_op.zfail),
MaxwellToGL::StencilOp(regs.stencil_back_op.zpass));
glStencilMaskSeparate(GL_BACK, regs.stencil_back_func.mask);
} else {
glStencilFuncSeparate(GL_BACK, GL_ALWAYS, 0, 0xFFFFFFFF);
glStencilOpSeparate(GL_BACK, GL_KEEP, GL_KEEP, GL_KEEP);
@ -782,7 +784,7 @@ void RasterizerOpenGL::SyncPolygonModes() {
flags[Dirty::PolygonModes] = false;
const auto& regs = maxwell3d->regs;
if (regs.fill_rectangle) {
if (regs.fill_via_triangle_mode != Maxwell::FillViaTriangleMode::Disabled) {
if (!GLAD_GL_NV_fill_rectangle) {
LOG_ERROR(Render_OpenGL, "GL_NV_fill_rectangle used and not supported");
glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
@ -855,8 +857,8 @@ void RasterizerOpenGL::SyncMultiSampleState() {
flags[Dirty::MultisampleControl] = false;
const auto& regs = maxwell3d->regs;
oglEnable(GL_SAMPLE_ALPHA_TO_COVERAGE, regs.multisample_control.alpha_to_coverage);
oglEnable(GL_SAMPLE_ALPHA_TO_ONE, regs.multisample_control.alpha_to_one);
oglEnable(GL_SAMPLE_ALPHA_TO_COVERAGE, regs.anti_alias_alpha_control.alpha_to_coverage);
oglEnable(GL_SAMPLE_ALPHA_TO_ONE, regs.anti_alias_alpha_control.alpha_to_one);
}
void RasterizerOpenGL::SyncFragmentColorClampState() {
@ -866,7 +868,8 @@ void RasterizerOpenGL::SyncFragmentColorClampState() {
}
flags[Dirty::FragmentClampColor] = false;
glClampColor(GL_CLAMP_FRAGMENT_COLOR, maxwell3d->regs.frag_color_clamp ? GL_TRUE : GL_FALSE);
glClampColor(GL_CLAMP_FRAGMENT_COLOR,
maxwell3d->regs.frag_color_clamp.AnyEnabled() ? GL_TRUE : GL_FALSE);
}
void RasterizerOpenGL::SyncBlendState() {
@ -886,18 +889,18 @@ void RasterizerOpenGL::SyncBlendState() {
}
flags[Dirty::BlendStates] = false;
if (!regs.independent_blend_enable) {
if (!regs.blend_per_target_enabled) {
if (!regs.blend.enable[0]) {
glDisable(GL_BLEND);
return;
}
glEnable(GL_BLEND);
glBlendFuncSeparate(MaxwellToGL::BlendFunc(regs.blend.factor_source_rgb),
MaxwellToGL::BlendFunc(regs.blend.factor_dest_rgb),
MaxwellToGL::BlendFunc(regs.blend.factor_source_a),
MaxwellToGL::BlendFunc(regs.blend.factor_dest_a));
glBlendEquationSeparate(MaxwellToGL::BlendEquation(regs.blend.equation_rgb),
MaxwellToGL::BlendEquation(regs.blend.equation_a));
glBlendFuncSeparate(MaxwellToGL::BlendFunc(regs.blend.color_source),
MaxwellToGL::BlendFunc(regs.blend.color_dest),
MaxwellToGL::BlendFunc(regs.blend.alpha_source),
MaxwellToGL::BlendFunc(regs.blend.alpha_dest));
glBlendEquationSeparate(MaxwellToGL::BlendEquation(regs.blend.color_op),
MaxwellToGL::BlendEquation(regs.blend.alpha_op));
return;
}
@ -916,14 +919,13 @@ void RasterizerOpenGL::SyncBlendState() {
}
glEnablei(GL_BLEND, static_cast<GLuint>(i));
const auto& src = regs.independent_blend[i];
glBlendFuncSeparatei(static_cast<GLuint>(i), MaxwellToGL::BlendFunc(src.factor_source_rgb),
MaxwellToGL::BlendFunc(src.factor_dest_rgb),
MaxwellToGL::BlendFunc(src.factor_source_a),
MaxwellToGL::BlendFunc(src.factor_dest_a));
glBlendEquationSeparatei(static_cast<GLuint>(i),
MaxwellToGL::BlendEquation(src.equation_rgb),
MaxwellToGL::BlendEquation(src.equation_a));
const auto& src = regs.blend_per_target[i];
glBlendFuncSeparatei(static_cast<GLuint>(i), MaxwellToGL::BlendFunc(src.color_source),
MaxwellToGL::BlendFunc(src.color_dest),
MaxwellToGL::BlendFunc(src.alpha_source),
MaxwellToGL::BlendFunc(src.alpha_dest));
glBlendEquationSeparatei(static_cast<GLuint>(i), MaxwellToGL::BlendEquation(src.color_op),
MaxwellToGL::BlendEquation(src.alpha_op));
}
}
@ -937,7 +939,7 @@ void RasterizerOpenGL::SyncLogicOpState() {
const auto& regs = maxwell3d->regs;
if (regs.logic_op.enable) {
glEnable(GL_COLOR_LOGIC_OP);
glLogicOp(MaxwellToGL::LogicOp(regs.logic_op.operation));
glLogicOp(MaxwellToGL::LogicOp(regs.logic_op.op));
} else {
glDisable(GL_COLOR_LOGIC_OP);
}
@ -996,7 +998,7 @@ void RasterizerOpenGL::SyncPointState() {
flags[Dirty::PointSize] = false;
oglEnable(GL_POINT_SPRITE, maxwell3d->regs.point_sprite_enable);
oglEnable(GL_PROGRAM_POINT_SIZE, maxwell3d->regs.vp_point_size.enable);
oglEnable(GL_PROGRAM_POINT_SIZE, maxwell3d->regs.point_size_attribute.enabled);
const bool is_rescaling{texture_cache.IsRescaling()};
const float scale = is_rescaling ? Settings::values.resolution_info.up_factor : 1.0f;
glPointSize(std::max(1.0f, maxwell3d->regs.point_size * scale));
@ -1010,8 +1012,8 @@ void RasterizerOpenGL::SyncLineState() {
flags[Dirty::LineWidth] = false;
const auto& regs = maxwell3d->regs;
oglEnable(GL_LINE_SMOOTH, regs.line_smooth_enable);
glLineWidth(regs.line_smooth_enable ? regs.line_width_smooth : regs.line_width_aliased);
oglEnable(GL_LINE_SMOOTH, regs.line_anti_alias_enable);
glLineWidth(regs.line_anti_alias_enable ? regs.line_width_smooth : regs.line_width_aliased);
}
void RasterizerOpenGL::SyncPolygonOffset() {
@ -1029,8 +1031,8 @@ void RasterizerOpenGL::SyncPolygonOffset() {
if (regs.polygon_offset_fill_enable || regs.polygon_offset_line_enable ||
regs.polygon_offset_point_enable) {
// Hardware divides polygon offset units by two
glPolygonOffsetClamp(regs.polygon_offset_factor, regs.polygon_offset_units / 2.0f,
regs.polygon_offset_clamp);
glPolygonOffsetClamp(regs.slope_scale_depth_bias, regs.depth_bias / 2.0f,
regs.depth_bias_clamp);
}
}
@ -1062,14 +1064,14 @@ void RasterizerOpenGL::SyncFramebufferSRGB() {
void RasterizerOpenGL::BeginTransformFeedback(GraphicsPipeline* program, GLenum primitive_mode) {
const auto& regs = maxwell3d->regs;
if (regs.tfb_enabled == 0) {
if (regs.transform_feedback_enabled == 0) {
return;
}
program->ConfigureTransformFeedback();
UNIMPLEMENTED_IF(regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationControl) ||
regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationEval) ||
regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::Geometry));
UNIMPLEMENTED_IF(regs.IsShaderConfigEnabled(Maxwell::ShaderType::TessellationInit) ||
regs.IsShaderConfigEnabled(Maxwell::ShaderType::Tessellation) ||
regs.IsShaderConfigEnabled(Maxwell::ShaderType::Geometry));
UNIMPLEMENTED_IF(primitive_mode != GL_POINTS);
// We may have to call BeginTransformFeedbackNV here since they seem to call different
@ -1080,7 +1082,7 @@ void RasterizerOpenGL::BeginTransformFeedback(GraphicsPipeline* program, GLenum
}
void RasterizerOpenGL::EndTransformFeedback() {
if (maxwell3d->regs.tfb_enabled != 0) {
if (maxwell3d->regs.transform_feedback_enabled != 0) {
glEndTransformFeedback();
}
}

43
src/video_core/renderer_opengl/gl_shader_cache.cpp
View file

@ -78,11 +78,11 @@ Shader::RuntimeInfo MakeRuntimeInfo(const GraphicsPipelineKey& key,
info.tess_clockwise = key.tessellation_clockwise != 0;
info.tess_primitive = [&key] {
switch (key.tessellation_primitive) {
case Maxwell::TessellationPrimitive::Isolines:
case Maxwell::Tessellation::DomainType::Isolines:
return Shader::TessPrimitive::Isolines;
case Maxwell::TessellationPrimitive::Triangles:
case Maxwell::Tessellation::DomainType::Triangles:
return Shader::TessPrimitive::Triangles;
case Maxwell::TessellationPrimitive::Quads:
case Maxwell::Tessellation::DomainType::Quads:
return Shader::TessPrimitive::Quads;
}
ASSERT(false);
@ -90,11 +90,11 @@ Shader::RuntimeInfo MakeRuntimeInfo(const GraphicsPipelineKey& key,
}();
info.tess_spacing = [&] {
switch (key.tessellation_spacing) {
case Maxwell::TessellationSpacing::Equal:
case Maxwell::Tessellation::Spacing::Integer:
return Shader::TessSpacing::Equal;
case Maxwell::TessellationSpacing::FractionalOdd:
case Maxwell::Tessellation::Spacing::FractionalOdd:
return Shader::TessSpacing::FractionalOdd;
case Maxwell::TessellationSpacing::FractionalEven:
case Maxwell::Tessellation::Spacing::FractionalEven:
return Shader::TessSpacing::FractionalEven;
}
ASSERT(false);
@ -139,14 +139,15 @@ Shader::RuntimeInfo MakeRuntimeInfo(const GraphicsPipelineKey& key,
}
void SetXfbState(VideoCommon::TransformFeedbackState& state, const Maxwell& regs) {
std::ranges::transform(regs.tfb_layouts, state.layouts.begin(), [](const auto& layout) {
return VideoCommon::TransformFeedbackState::Layout{
.stream = layout.stream,
.varying_count = layout.varying_count,
.stride = layout.stride,
};
});
state.varyings = regs.tfb_varying_locs;
std::ranges::transform(regs.transform_feedback.controls, state.layouts.begin(),
[](const auto& layout) {
return VideoCommon::TransformFeedbackState::Layout{
.stream = layout.stream,
.varying_count = layout.varying_count,
.stride = layout.stride,
};
});
state.varyings = regs.stream_out_layout;
}
} // Anonymous namespace
@ -309,14 +310,16 @@ GraphicsPipeline* ShaderCache::CurrentGraphicsPipeline() {
}
const auto& regs{maxwell3d->regs};
graphics_key.raw = 0;
graphics_key.early_z.Assign(regs.force_early_fragment_tests != 0 ? 1 : 0);
graphics_key.early_z.Assign(regs.mandated_early_z != 0 ? 1 : 0);
graphics_key.gs_input_topology.Assign(graphics_key.unique_hashes[4] != 0
? regs.draw.topology.Value()
: Maxwell::PrimitiveTopology{});
graphics_key.tessellation_primitive.Assign(regs.tess_mode.prim.Value());
graphics_key.tessellation_spacing.Assign(regs.tess_mode.spacing.Value());
graphics_key.tessellation_clockwise.Assign(regs.tess_mode.cw.Value());
graphics_key.xfb_enabled.Assign(regs.tfb_enabled != 0 ? 1 : 0);
graphics_key.tessellation_primitive.Assign(regs.tessellation.params.domain_type.Value());
graphics_key.tessellation_spacing.Assign(regs.tessellation.params.spacing.Value());
graphics_key.tessellation_clockwise.Assign(
regs.tessellation.params.output_primitives.Value() !=
Maxwell::Tessellation::OutputPrimitves::Triangles_CCW);
graphics_key.xfb_enabled.Assign(regs.transform_feedback_enabled != 0 ? 1 : 0);
if (graphics_key.xfb_enabled) {
SetXfbState(graphics_key.xfb_state, regs);
}
@ -354,7 +357,7 @@ GraphicsPipeline* ShaderCache::BuiltPipeline(GraphicsPipeline* pipeline) const n
// If games are using a small index count, we can assume these are full screen quads.
// Usually these shaders are only used once for building textures so we can assume they
// can't be built async
if (maxwell3d->regs.index_array.count <= 6 || maxwell3d->regs.vertex_buffer.count <= 6) {
if (maxwell3d->regs.index_buffer.count <= 6 || maxwell3d->regs.vertex_buffer.count <= 6) {
return pipeline;
}
return nullptr;

57
src/video_core/renderer_opengl/gl_state_tracker.cpp
View file

@ -38,12 +38,12 @@ void SetupDirtyColorMasks(Tables& tables) {
void SetupDirtyVertexInstances(Tables& tables) {
static constexpr std::size_t instance_base_offset = 3;
for (std::size_t i = 0; i < Regs::NumVertexArrays; ++i) {
const std::size_t array_offset = OFF(vertex_array) + i * NUM(vertex_array[0]);
const std::size_t array_offset = OFF(vertex_streams) + i * NUM(vertex_streams[0]);
const std::size_t instance_array_offset = array_offset + instance_base_offset;
tables[0][instance_array_offset] = static_cast<u8>(VertexInstance0 + i);
tables[1][instance_array_offset] = VertexInstances;
const std::size_t instance_offset = OFF(instanced_arrays) + i;
const std::size_t instance_offset = OFF(vertex_stream_instances) + i;
tables[0][instance_offset] = static_cast<u8>(VertexInstance0 + i);
tables[1][instance_offset] = VertexInstances;
}
@ -70,8 +70,8 @@ void SetupDirtyViewports(Tables& tables) {
FillBlock(tables[1], OFF(viewport_transform), NUM(viewport_transform), Viewports);
FillBlock(tables[1], OFF(viewports), NUM(viewports), Viewports);
tables[0][OFF(viewport_transform_enabled)] = ViewportTransform;
tables[1][OFF(viewport_transform_enabled)] = Viewports;
tables[0][OFF(viewport_scale_offset_enbled)] = ViewportTransform;
tables[1][OFF(viewport_scale_offset_enbled)] = Viewports;
}
void SetupDirtyScissors(Tables& tables) {
@ -88,7 +88,7 @@ void SetupDirtyPolygonModes(Tables& tables) {
tables[1][OFF(polygon_mode_front)] = PolygonModes;
tables[1][OFF(polygon_mode_back)] = PolygonModes;
tables[0][OFF(fill_rectangle)] = PolygonModes;
tables[0][OFF(fill_via_triangle_mode)] = PolygonModes;
}
void SetupDirtyDepthTest(Tables& tables) {
@ -100,12 +100,14 @@ void SetupDirtyDepthTest(Tables& tables) {
void SetupDirtyStencilTest(Tables& tables) {
static constexpr std::array offsets = {
OFF(stencil_enable), OFF(stencil_front_func_func), OFF(stencil_front_func_ref),
OFF(stencil_front_func_mask), OFF(stencil_front_op_fail), OFF(stencil_front_op_zfail),
OFF(stencil_front_op_zpass), OFF(stencil_front_mask), OFF(stencil_two_side_enable),
OFF(stencil_back_func_func), OFF(stencil_back_func_ref), OFF(stencil_back_func_mask),
OFF(stencil_back_op_fail), OFF(stencil_back_op_zfail), OFF(stencil_back_op_zpass),
OFF(stencil_back_mask)};
OFF(stencil_enable), OFF(stencil_front_op.func),
OFF(stencil_front_func.ref), OFF(stencil_front_func.func_mask),
OFF(stencil_front_op.fail), OFF(stencil_front_op.zfail),
OFF(stencil_front_op.zpass), OFF(stencil_front_func.mask),
OFF(stencil_two_side_enable), OFF(stencil_back_op.func),
OFF(stencil_back_func.ref), OFF(stencil_back_func.func_mask),
OFF(stencil_back_op.fail), OFF(stencil_back_op.zfail),
OFF(stencil_back_op.zpass), OFF(stencil_back_func.mask)};
for (const auto offset : offsets) {
tables[0][offset] = StencilTest;
}
@ -121,15 +123,15 @@ void SetupDirtyAlphaTest(Tables& tables) {
void SetupDirtyBlend(Tables& tables) {
FillBlock(tables[0], OFF(blend_color), NUM(blend_color), BlendColor);
tables[0][OFF(independent_blend_enable)] = BlendIndependentEnabled;
tables[0][OFF(blend_per_target_enabled)] = BlendIndependentEnabled;
for (std::size_t i = 0; i < Regs::NumRenderTargets; ++i) {
const std::size_t offset = OFF(independent_blend) + i * NUM(independent_blend[0]);
FillBlock(tables[0], offset, NUM(independent_blend[0]), BlendState0 + i);
const std::size_t offset = OFF(blend_per_target) + i * NUM(blend_per_target[0]);
FillBlock(tables[0], offset, NUM(blend_per_target[0]), BlendState0 + i);
tables[0][OFF(blend.enable) + i] = static_cast<u8>(BlendState0 + i);
}
FillBlock(tables[1], OFF(independent_blend), NUM(independent_blend), BlendStates);
FillBlock(tables[1], OFF(blend_per_target), NUM(blend_per_target), BlendStates);
FillBlock(tables[1], OFF(blend), NUM(blend), BlendStates);
}
@ -142,13 +144,14 @@ void SetupDirtyPolygonOffset(Tables& tables) {
table[OFF(polygon_offset_fill_enable)] = PolygonOffset;
table[OFF(polygon_offset_line_enable)] = PolygonOffset;
table[OFF(polygon_offset_point_enable)] = PolygonOffset;
table[OFF(polygon_offset_factor)] = PolygonOffset;
table[OFF(polygon_offset_units)] = PolygonOffset;
table[OFF(polygon_offset_clamp)] = PolygonOffset;
table[OFF(slope_scale_depth_bias)] = PolygonOffset;
table[OFF(depth_bias)] = PolygonOffset;
table[OFF(depth_bias_clamp)] = PolygonOffset;
}
void SetupDirtyMultisampleControl(Tables& tables) {
FillBlock(tables[0], OFF(multisample_control), NUM(multisample_control), MultisampleControl);
FillBlock(tables[0], OFF(anti_alias_alpha_control), NUM(anti_alias_alpha_control),
MultisampleControl);
}
void SetupDirtyRasterizeEnable(Tables& tables) {
@ -168,7 +171,7 @@ void SetupDirtyFragmentClampColor(Tables& tables) {
}
void SetupDirtyPointSize(Tables& tables) {
tables[0][OFF(vp_point_size)] = PointSize;
tables[0][OFF(point_size_attribute)] = PointSize;
tables[0][OFF(point_size)] = PointSize;
tables[0][OFF(point_sprite_enable)] = PointSize;
}
@ -176,28 +179,28 @@ void SetupDirtyPointSize(Tables& tables) {
void SetupDirtyLineWidth(Tables& tables) {
tables[0][OFF(line_width_smooth)] = LineWidth;
tables[0][OFF(line_width_aliased)] = LineWidth;
tables[0][OFF(line_smooth_enable)] = LineWidth;
tables[0][OFF(line_anti_alias_enable)] = LineWidth;
}
void SetupDirtyClipControl(Tables& tables) {
auto& table = tables[0];
table[OFF(screen_y_control)] = ClipControl;
table[OFF(window_origin)] = ClipControl;
table[OFF(depth_mode)] = ClipControl;
}
void SetupDirtyDepthClampEnabled(Tables& tables) {
tables[0][OFF(view_volume_clip_control)] = DepthClampEnabled;
tables[0][OFF(viewport_clip_control)] = DepthClampEnabled;
}
void SetupDirtyMisc(Tables& tables) {
auto& table = tables[0];
table[OFF(clip_distance_enabled)] = ClipDistances;
table[OFF(user_clip_enable)] = ClipDistances;
table[OFF(front_face)] = FrontFace;
table[OFF(gl_front_face)] = FrontFace;
table[OFF(cull_test_enabled)] = CullTest;
table[OFF(cull_face)] = CullTest;
table[OFF(gl_cull_test_enabled)] = CullTest;
table[OFF(gl_cull_face)] = CullTest;
}
} // Anonymous namespace

287
src/video_core/renderer_opengl/maxwell_to_gl.h
View file

@ -126,51 +126,60 @@ inline const FormatTuple& GetFormatTuple(VideoCore::Surface::PixelFormat pixel_f
inline GLenum VertexFormat(Maxwell::VertexAttribute attrib) {
switch (attrib.type) {
case Maxwell::VertexAttribute::Type::UnsignedNorm:
case Maxwell::VertexAttribute::Type::UnsignedScaled:
case Maxwell::VertexAttribute::Type::UnsignedInt:
case Maxwell::VertexAttribute::Type::UnusedEnumDoNotUseBecauseItWillGoAway:
ASSERT_MSG(false, "Invalid vertex attribute type!");
break;
case Maxwell::VertexAttribute::Type::UNorm:
case Maxwell::VertexAttribute::Type::UScaled:
case Maxwell::VertexAttribute::Type::UInt:
switch (attrib.size) {
case Maxwell::VertexAttribute::Size::Size_8:
case Maxwell::VertexAttribute::Size::Size_8_8:
case Maxwell::VertexAttribute::Size::Size_8_8_8:
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8:
case Maxwell::VertexAttribute::Size::Size_A8:
case Maxwell::VertexAttribute::Size::Size_R8_G8:
case Maxwell::VertexAttribute::Size::Size_G8_R8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8_A8:
case Maxwell::VertexAttribute::Size::Size_X8_B8_G8_R8:
return GL_UNSIGNED_BYTE;
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return GL_UNSIGNED_SHORT;
case Maxwell::VertexAttribute::Size::Size_32:
case Maxwell::VertexAttribute::Size::Size_32_32:
case Maxwell::VertexAttribute::Size::Size_32_32_32:
case Maxwell::VertexAttribute::Size::Size_32_32_32_32:
case Maxwell::VertexAttribute::Size::Size_R32:
case Maxwell::VertexAttribute::Size::Size_R32_G32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32_A32:
return GL_UNSIGNED_INT;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
case Maxwell::VertexAttribute::Size::Size_A2_B10_G10_R10:
return GL_UNSIGNED_INT_2_10_10_10_REV;
default:
break;
}
break;
case Maxwell::VertexAttribute::Type::SignedNorm:
case Maxwell::VertexAttribute::Type::SignedScaled:
case Maxwell::VertexAttribute::Type::SignedInt:
case Maxwell::VertexAttribute::Type::SNorm:
case Maxwell::VertexAttribute::Type::SScaled:
case Maxwell::VertexAttribute::Type::SInt:
switch (attrib.size) {
case Maxwell::VertexAttribute::Size::Size_8:
case Maxwell::VertexAttribute::Size::Size_8_8:
case Maxwell::VertexAttribute::Size::Size_8_8_8:
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8:
case Maxwell::VertexAttribute::Size::Size_A8:
case Maxwell::VertexAttribute::Size::Size_R8_G8:
case Maxwell::VertexAttribute::Size::Size_G8_R8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8_A8:
case Maxwell::VertexAttribute::Size::Size_X8_B8_G8_R8:
return GL_BYTE;
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return GL_SHORT;
case Maxwell::VertexAttribute::Size::Size_32:
case Maxwell::VertexAttribute::Size::Size_32_32:
case Maxwell::VertexAttribute::Size::Size_32_32_32:
case Maxwell::VertexAttribute::Size::Size_32_32_32_32:
case Maxwell::VertexAttribute::Size::Size_R32:
case Maxwell::VertexAttribute::Size::Size_R32_G32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32_A32:
return GL_INT;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
case Maxwell::VertexAttribute::Size::Size_A2_B10_G10_R10:
return GL_INT_2_10_10_10_REV;
default:
break;
@ -178,17 +187,17 @@ inline GLenum VertexFormat(Maxwell::VertexAttribute attrib) {
break;
case Maxwell::VertexAttribute::Type::Float:
switch (attrib.size) {
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return GL_HALF_FLOAT;
case Maxwell::VertexAttribute::Size::Size_32:
case Maxwell::VertexAttribute::Size::Size_32_32:
case Maxwell::VertexAttribute::Size::Size_32_32_32:
case Maxwell::VertexAttribute::Size::Size_32_32_32_32:
case Maxwell::VertexAttribute::Size::Size_R32:
case Maxwell::VertexAttribute::Size::Size_R32_G32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32_A32:
return GL_FLOAT;
case Maxwell::VertexAttribute::Size::Size_11_11_10:
case Maxwell::VertexAttribute::Size::Size_B10_G11_R11:
return GL_UNSIGNED_INT_10F_11F_11F_REV;
default:
break;
@ -335,20 +344,20 @@ inline GLenum DepthCompareFunc(Tegra::Texture::DepthCompareFunc func) {
inline GLenum BlendEquation(Maxwell::Blend::Equation equation) {
switch (equation) {
case Maxwell::Blend::Equation::Add:
case Maxwell::Blend::Equation::AddGL:
case Maxwell::Blend::Equation::Add_D3D:
case Maxwell::Blend::Equation::Add_GL:
return GL_FUNC_ADD;
case Maxwell::Blend::Equation::Subtract:
case Maxwell::Blend::Equation::SubtractGL:
case Maxwell::Blend::Equation::Subtract_D3D:
case Maxwell::Blend::Equation::Subtract_GL:
return GL_FUNC_SUBTRACT;
case Maxwell::Blend::Equation::ReverseSubtract:
case Maxwell::Blend::Equation::ReverseSubtractGL:
case Maxwell::Blend::Equation::ReverseSubtract_D3D:
case Maxwell::Blend::Equation::ReverseSubtract_GL:
return GL_FUNC_REVERSE_SUBTRACT;
case Maxwell::Blend::Equation::Min:
case Maxwell::Blend::Equation::MinGL:
case Maxwell::Blend::Equation::Min_D3D:
case Maxwell::Blend::Equation::Min_GL:
return GL_MIN;
case Maxwell::Blend::Equation::Max:
case Maxwell::Blend::Equation::MaxGL:
case Maxwell::Blend::Equation::Max_D3D:
case Maxwell::Blend::Equation::Max_GL:
return GL_MAX;
}
UNIMPLEMENTED_MSG("Unimplemented blend equation={}", equation);
@ -357,62 +366,62 @@ inline GLenum BlendEquation(Maxwell::Blend::Equation equation) {
inline GLenum BlendFunc(Maxwell::Blend::Factor factor) {
switch (factor) {
case Maxwell::Blend::Factor::Zero:
case Maxwell::Blend::Factor::ZeroGL:
case Maxwell::Blend::Factor::Zero_D3D:
case Maxwell::Blend::Factor::Zero_GL:
return GL_ZERO;
case Maxwell::Blend::Factor::One:
case Maxwell::Blend::Factor::OneGL:
case Maxwell::Blend::Factor::One_D3D:
case Maxwell::Blend::Factor::One_GL:
return GL_ONE;
case Maxwell::Blend::Factor::SourceColor:
case Maxwell::Blend::Factor::SourceColorGL:
case Maxwell::Blend::Factor::SourceColor_D3D:
case Maxwell::Blend::Factor::SourceColor_GL:
return GL_SRC_COLOR;
case Maxwell::Blend::Factor::OneMinusSourceColor:
case Maxwell::Blend::Factor::OneMinusSourceColorGL:
case Maxwell::Blend::Factor::OneMinusSourceColor_D3D:
case Maxwell::Blend::Factor::OneMinusSourceColor_GL:
return GL_ONE_MINUS_SRC_COLOR;
case Maxwell::Blend::Factor::SourceAlpha:
case Maxwell::Blend::Factor::SourceAlphaGL:
case Maxwell::Blend::Factor::SourceAlpha_D3D:
case Maxwell::Blend::Factor::SourceAlpha_GL:
return GL_SRC_ALPHA;
case Maxwell::Blend::Factor::OneMinusSourceAlpha:
case Maxwell::Blend::Factor::OneMinusSourceAlphaGL:
case Maxwell::Blend::Factor::OneMinusSourceAlpha_D3D:
case Maxwell::Blend::Factor::OneMinusSourceAlpha_GL:
return GL_ONE_MINUS_SRC_ALPHA;
case Maxwell::Blend::Factor::DestAlpha:
case Maxwell::Blend::Factor::DestAlphaGL:
case Maxwell::Blend::Factor::DestAlpha_D3D:
case Maxwell::Blend::Factor::DestAlpha_GL:
return GL_DST_ALPHA;
case Maxwell::Blend::Factor::OneMinusDestAlpha:
case Maxwell::Blend::Factor::OneMinusDestAlphaGL:
case Maxwell::Blend::Factor::OneMinusDestAlpha_D3D:
case Maxwell::Blend::Factor::OneMinusDestAlpha_GL:
return GL_ONE_MINUS_DST_ALPHA;
case Maxwell::Blend::Factor::DestColor:
case Maxwell::Blend::Factor::DestColorGL:
case Maxwell::Blend::Factor::DestColor_D3D:
case Maxwell::Blend::Factor::DestColor_GL:
return GL_DST_COLOR;
case Maxwell::Blend::Factor::OneMinusDestColor:
case Maxwell::Blend::Factor::OneMinusDestColorGL:
case Maxwell::Blend::Factor::OneMinusDestColor_D3D:
case Maxwell::Blend::Factor::OneMinusDestColor_GL:
return GL_ONE_MINUS_DST_COLOR;
case Maxwell::Blend::Factor::SourceAlphaSaturate:
case Maxwell::Blend::Factor::SourceAlphaSaturateGL:
case Maxwell::Blend::Factor::SourceAlphaSaturate_D3D:
case Maxwell::Blend::Factor::SourceAlphaSaturate_GL:
return GL_SRC_ALPHA_SATURATE;
case Maxwell::Blend::Factor::Source1Color:
case Maxwell::Blend::Factor::Source1ColorGL:
case Maxwell::Blend::Factor::Source1Color_D3D:
case Maxwell::Blend::Factor::Source1Color_GL:
return GL_SRC1_COLOR;
case Maxwell::Blend::Factor::OneMinusSource1Color:
case Maxwell::Blend::Factor::OneMinusSource1ColorGL:
case Maxwell::Blend::Factor::OneMinusSource1Color_D3D:
case Maxwell::Blend::Factor::OneMinusSource1Color_GL:
return GL_ONE_MINUS_SRC1_COLOR;
case Maxwell::Blend::Factor::Source1Alpha:
case Maxwell::Blend::Factor::Source1AlphaGL:
case Maxwell::Blend::Factor::Source1Alpha_D3D:
case Maxwell::Blend::Factor::Source1Alpha_GL:
return GL_SRC1_ALPHA;
case Maxwell::Blend::Factor::OneMinusSource1Alpha:
case Maxwell::Blend::Factor::OneMinusSource1AlphaGL:
case Maxwell::Blend::Factor::OneMinusSource1Alpha_D3D:
case Maxwell::Blend::Factor::OneMinusSource1Alpha_GL:
return GL_ONE_MINUS_SRC1_ALPHA;
case Maxwell::Blend::Factor::ConstantColor:
case Maxwell::Blend::Factor::ConstantColorGL:
case Maxwell::Blend::Factor::BlendFactor_D3D:
case Maxwell::Blend::Factor::ConstantColor_GL:
return GL_CONSTANT_COLOR;
case Maxwell::Blend::Factor::OneMinusConstantColor:
case Maxwell::Blend::Factor::OneMinusConstantColorGL:
case Maxwell::Blend::Factor::OneMinusBlendFactor_D3D:
case Maxwell::Blend::Factor::OneMinusConstantColor_GL:
return GL_ONE_MINUS_CONSTANT_COLOR;
case Maxwell::Blend::Factor::ConstantAlpha:
case Maxwell::Blend::Factor::ConstantAlphaGL:
case Maxwell::Blend::Factor::BothSourceAlpha_D3D:
case Maxwell::Blend::Factor::ConstantAlpha_GL:
return GL_CONSTANT_ALPHA;
case Maxwell::Blend::Factor::OneMinusConstantAlpha:
case Maxwell::Blend::Factor::OneMinusConstantAlphaGL:
case Maxwell::Blend::Factor::OneMinusBothSourceAlpha_D3D:
case Maxwell::Blend::Factor::OneMinusConstantAlpha_GL:
return GL_ONE_MINUS_CONSTANT_ALPHA;
}
UNIMPLEMENTED_MSG("Unimplemented blend factor={}", factor);
@ -421,60 +430,60 @@ inline GLenum BlendFunc(Maxwell::Blend::Factor factor) {
inline GLenum ComparisonOp(Maxwell::ComparisonOp comparison) {
switch (comparison) {
case Maxwell::ComparisonOp::Never:
case Maxwell::ComparisonOp::NeverOld:
case Maxwell::ComparisonOp::Never_D3D:
case Maxwell::ComparisonOp::Never_GL:
return GL_NEVER;
case Maxwell::ComparisonOp::Less:
case Maxwell::ComparisonOp::LessOld:
case Maxwell::ComparisonOp::Less_D3D:
case Maxwell::ComparisonOp::Less_GL:
return GL_LESS;
case Maxwell::ComparisonOp::Equal:
case Maxwell::ComparisonOp::EqualOld:
case Maxwell::ComparisonOp::Equal_D3D:
case Maxwell::ComparisonOp::Equal_GL:
return GL_EQUAL;
case Maxwell::ComparisonOp::LessEqual:
case Maxwell::ComparisonOp::LessEqualOld:
case Maxwell::ComparisonOp::LessEqual_D3D:
case Maxwell::ComparisonOp::LessEqual_GL:
return GL_LEQUAL;
case Maxwell::ComparisonOp::Greater:
case Maxwell::ComparisonOp::GreaterOld:
case Maxwell::ComparisonOp::Greater_D3D:
case Maxwell::ComparisonOp::Greater_GL:
return GL_GREATER;
case Maxwell::ComparisonOp::NotEqual:
case Maxwell::ComparisonOp::NotEqualOld:
case Maxwell::ComparisonOp::NotEqual_D3D:
case Maxwell::ComparisonOp::NotEqual_GL:
return GL_NOTEQUAL;
case Maxwell::ComparisonOp::GreaterEqual:
case Maxwell::ComparisonOp::GreaterEqualOld:
case Maxwell::ComparisonOp::GreaterEqual_D3D:
case Maxwell::ComparisonOp::GreaterEqual_GL:
return GL_GEQUAL;
case Maxwell::ComparisonOp::Always:
case Maxwell::ComparisonOp::AlwaysOld:
case Maxwell::ComparisonOp::Always_D3D:
case Maxwell::ComparisonOp::Always_GL:
return GL_ALWAYS;
}
UNIMPLEMENTED_MSG("Unimplemented comparison op={}", comparison);
return GL_ALWAYS;
}
inline GLenum StencilOp(Maxwell::StencilOp stencil) {
inline GLenum StencilOp(Maxwell::StencilOp::Op stencil) {
switch (stencil) {
case Maxwell::StencilOp::Keep:
case Maxwell::StencilOp::KeepOGL:
case Maxwell::StencilOp::Op::Keep_D3D:
case Maxwell::StencilOp::Op::Keep_GL:
return GL_KEEP;
case Maxwell::StencilOp::Zero:
case Maxwell::StencilOp::ZeroOGL:
case Maxwell::StencilOp::Op::Zero_D3D:
case Maxwell::StencilOp::Op::Zero_GL:
return GL_ZERO;
case Maxwell::StencilOp::Replace:
case Maxwell::StencilOp::ReplaceOGL:
case Maxwell::StencilOp::Op::Replace_D3D:
case Maxwell::StencilOp::Op::Replace_GL:
return GL_REPLACE;
case Maxwell::StencilOp::Incr:
case Maxwell::StencilOp::IncrOGL:
case Maxwell::StencilOp::Op::IncrSaturate_D3D:
case Maxwell::StencilOp::Op::IncrSaturate_GL:
return GL_INCR;
case Maxwell::StencilOp::Decr:
case Maxwell::StencilOp::DecrOGL:
case Maxwell::StencilOp::Op::DecrSaturate_D3D:
case Maxwell::StencilOp::Op::DecrSaturate_GL:
return GL_DECR;
case Maxwell::StencilOp::Invert:
case Maxwell::StencilOp::InvertOGL:
case Maxwell::StencilOp::Op::Invert_D3D:
case Maxwell::StencilOp::Op::Invert_GL:
return GL_INVERT;
case Maxwell::StencilOp::IncrWrap:
case Maxwell::StencilOp::IncrWrapOGL:
case Maxwell::StencilOp::Op::Incr_D3D:
case Maxwell::StencilOp::Op::Incr_GL:
return GL_INCR_WRAP;
case Maxwell::StencilOp::DecrWrap:
case Maxwell::StencilOp::DecrWrapOGL:
case Maxwell::StencilOp::Op::Decr_D3D:
case Maxwell::StencilOp::Op::Decr_GL:
return GL_DECR_WRAP;
}
UNIMPLEMENTED_MSG("Unimplemented stencil op={}", stencil);
@ -505,39 +514,39 @@ inline GLenum CullFace(Maxwell::CullFace cull_face) {
return GL_BACK;
}
inline GLenum LogicOp(Maxwell::LogicOperation operation) {
inline GLenum LogicOp(Maxwell::LogicOp::Op operation) {
switch (operation) {
case Maxwell::LogicOperation::Clear:
case Maxwell::LogicOp::Op::Clear:
return GL_CLEAR;
case Maxwell::LogicOperation::And:
case Maxwell::LogicOp::Op::And:
return GL_AND;
case Maxwell::LogicOperation::AndReverse:
case Maxwell::LogicOp::Op::AndReverse:
return GL_AND_REVERSE;
case Maxwell::LogicOperation::Copy:
case Maxwell::LogicOp::Op::Copy:
return GL_COPY;
case Maxwell::LogicOperation::AndInverted:
case Maxwell::LogicOp::Op::AndInverted:
return GL_AND_INVERTED;
case Maxwell::LogicOperation::NoOp:
case Maxwell::LogicOp::Op::NoOp:
return GL_NOOP;
case Maxwell::LogicOperation::Xor:
case Maxwell::LogicOp::Op::Xor:
return GL_XOR;
case Maxwell::LogicOperation::Or:
case Maxwell::LogicOp::Op::Or:
return GL_OR;
case Maxwell::LogicOperation::Nor:
case Maxwell::LogicOp::Op::Nor:
return GL_NOR;
case Maxwell::LogicOperation::Equiv:
case Maxwell::LogicOp::Op::Equiv:
return GL_EQUIV;
case Maxwell::LogicOperation::Invert:
case Maxwell::LogicOp::Op::Invert:
return GL_INVERT;
case Maxwell::LogicOperation::OrReverse:
case Maxwell::LogicOp::Op::OrReverse:
return GL_OR_REVERSE;
case Maxwell::LogicOperation::CopyInverted:
case Maxwell::LogicOp::Op::CopyInverted:
return GL_COPY_INVERTED;
case Maxwell::LogicOperation::OrInverted:
case Maxwell::LogicOp::Op::OrInverted:
return GL_OR_INVERTED;
case Maxwell::LogicOperation::Nand:
case Maxwell::LogicOp::Op::Nand:
return GL_NAND;
case Maxwell::LogicOperation::Set:
case Maxwell::LogicOp::Op::Set:
return GL_SET;
}
UNIMPLEMENTED_MSG("Unimplemented logic operation={}", operation);

299
src/video_core/renderer_vulkan/fixed_pipeline_state.cpp
View file

@ -34,14 +34,15 @@ constexpr std::array POLYGON_OFFSET_ENABLE_LUT = {
};
void RefreshXfbState(VideoCommon::TransformFeedbackState& state, const Maxwell& regs) {
std::ranges::transform(regs.tfb_layouts, state.layouts.begin(), [](const auto& layout) {
return VideoCommon::TransformFeedbackState::Layout{
.stream = layout.stream,
.varying_count = layout.varying_count,
.stride = layout.stride,
};
});
state.varyings = regs.tfb_varying_locs;
std::ranges::transform(regs.transform_feedback.controls, state.layouts.begin(),
[](const auto& layout) {
return VideoCommon::TransformFeedbackState::Layout{
.stream = layout.stream,
.varying_count = layout.varying_count,
.stride = layout.stride,
};
});
state.varyings = regs.stream_out_layout;
}
} // Anonymous namespace
@ -58,33 +59,34 @@ void FixedPipelineState::Refresh(Tegra::Engines::Maxwell3D& maxwell3d,
raw1 = 0;
extended_dynamic_state.Assign(has_extended_dynamic_state ? 1 : 0);
dynamic_vertex_input.Assign(has_dynamic_vertex_input ? 1 : 0);
xfb_enabled.Assign(regs.tfb_enabled != 0);
xfb_enabled.Assign(regs.transform_feedback_enabled != 0);
primitive_restart_enable.Assign(regs.primitive_restart.enabled != 0 ? 1 : 0);
depth_bias_enable.Assign(enabled_lut[POLYGON_OFFSET_ENABLE_LUT[topology_index]] != 0 ? 1 : 0);
depth_clamp_disabled.Assign(regs.view_volume_clip_control.depth_clamp_disabled.Value());
depth_clamp_disabled.Assign(regs.viewport_clip_control.geometry_clip ==
Maxwell::ViewportClipControl::GeometryClip::Passthrough);
ndc_minus_one_to_one.Assign(regs.depth_mode == Maxwell::DepthMode::MinusOneToOne ? 1 : 0);
polygon_mode.Assign(PackPolygonMode(regs.polygon_mode_front));
patch_control_points_minus_one.Assign(regs.patch_vertices - 1);
tessellation_primitive.Assign(static_cast<u32>(regs.tess_mode.prim.Value()));
tessellation_spacing.Assign(static_cast<u32>(regs.tess_mode.spacing.Value()));
tessellation_clockwise.Assign(regs.tess_mode.cw.Value());
tessellation_primitive.Assign(static_cast<u32>(regs.tessellation.params.domain_type.Value()));
tessellation_spacing.Assign(static_cast<u32>(regs.tessellation.params.spacing.Value()));
tessellation_clockwise.Assign(regs.tessellation.params.output_primitives.Value() !=
Maxwell::Tessellation::OutputPrimitves::Triangles_CCW);
logic_op_enable.Assign(regs.logic_op.enable != 0 ? 1 : 0);
logic_op.Assign(PackLogicOp(regs.logic_op.operation));
logic_op.Assign(PackLogicOp(regs.logic_op.op));
topology.Assign(regs.draw.topology);
msaa_mode.Assign(regs.multisample_mode);
msaa_mode.Assign(regs.anti_alias_samples_mode);
raw2 = 0;
rasterize_enable.Assign(regs.rasterize_enable != 0 ? 1 : 0);
const auto test_func =
regs.alpha_test_enabled != 0 ? regs.alpha_test_func : Maxwell::ComparisonOp::Always;
regs.alpha_test_enabled != 0 ? regs.alpha_test_func : Maxwell::ComparisonOp::Always_GL;
alpha_test_func.Assign(PackComparisonOp(test_func));
early_z.Assign(regs.force_early_fragment_tests != 0 ? 1 : 0);
early_z.Assign(regs.mandated_early_z != 0 ? 1 : 0);
depth_enabled.Assign(regs.zeta_enable != 0 ? 1 : 0);
depth_format.Assign(static_cast<u32>(regs.zeta.format));
y_negate.Assign(regs.screen_y_control.y_negate != 0 ? 1 : 0);
provoking_vertex_last.Assign(regs.provoking_vertex_last != 0 ? 1 : 0);
conservative_raster_enable.Assign(regs.conservative_raster_enable != 0 ? 1 : 0);
smooth_lines.Assign(regs.line_smooth_enable != 0 ? 1 : 0);
y_negate.Assign(regs.window_origin.mode != Maxwell::WindowOrigin::Mode::UpperLeft ? 1 : 0);
provoking_vertex_last.Assign(regs.provoking_vertex == Maxwell::ProvokingVertex::Last ? 1 : 0);
smooth_lines.Assign(regs.line_anti_alias_enable != 0 ? 1 : 0);
for (size_t i = 0; i < regs.rt.size(); ++i) {
color_formats[i] = static_cast<u8>(regs.rt[i].format);
@ -116,8 +118,8 @@ void FixedPipelineState::Refresh(Tegra::Engines::Maxwell3D& maxwell3d,
maxwell3d.dirty.flags[Dirty::VertexInput] = false;
enabled_divisors = 0;
for (size_t index = 0; index < Maxwell::NumVertexArrays; ++index) {
const bool is_enabled = regs.instanced_arrays.IsInstancingEnabled(index);
binding_divisors[index] = is_enabled ? regs.vertex_array[index].divisor : 0;
const bool is_enabled = regs.vertex_stream_instances.IsInstancingEnabled(index);
binding_divisors[index] = is_enabled ? regs.vertex_streams[index].frequency : 0;
enabled_divisors |= (is_enabled ? u64{1} : 0) << index;
}
for (size_t index = 0; index < Maxwell::NumVertexAttributes; ++index) {
@ -164,17 +166,17 @@ void FixedPipelineState::BlendingAttachment::Refresh(const Maxwell& regs, size_t
// TODO: C++20 Use templated lambda to deduplicate code
if (!regs.independent_blend_enable) {
const auto& src = regs.blend;
if (!src.enable[index]) {
if (!regs.blend_per_target_enabled) {
if (!regs.blend.enable[index]) {
return;
}
equation_rgb.Assign(PackBlendEquation(src.equation_rgb));
equation_a.Assign(PackBlendEquation(src.equation_a));
factor_source_rgb.Assign(PackBlendFactor(src.factor_source_rgb));
factor_dest_rgb.Assign(PackBlendFactor(src.factor_dest_rgb));
factor_source_a.Assign(PackBlendFactor(src.factor_source_a));
factor_dest_a.Assign(PackBlendFactor(src.factor_dest_a));
const auto& src = regs.blend;
equation_rgb.Assign(PackBlendEquation(src.color_op));
equation_a.Assign(PackBlendEquation(src.alpha_op));
factor_source_rgb.Assign(PackBlendFactor(src.color_source));
factor_dest_rgb.Assign(PackBlendFactor(src.color_dest));
factor_source_a.Assign(PackBlendFactor(src.alpha_source));
factor_dest_a.Assign(PackBlendFactor(src.alpha_dest));
enable.Assign(1);
return;
}
@ -182,34 +184,34 @@ void FixedPipelineState::BlendingAttachment::Refresh(const Maxwell& regs, size_t
if (!regs.blend.enable[index]) {
return;
}
const auto& src = regs.independent_blend[index];
equation_rgb.Assign(PackBlendEquation(src.equation_rgb));
equation_a.Assign(PackBlendEquation(src.equation_a));
factor_source_rgb.Assign(PackBlendFactor(src.factor_source_rgb));
factor_dest_rgb.Assign(PackBlendFactor(src.factor_dest_rgb));
factor_source_a.Assign(PackBlendFactor(src.factor_source_a));
factor_dest_a.Assign(PackBlendFactor(src.factor_dest_a));
const auto& src = regs.blend_per_target[index];
equation_rgb.Assign(PackBlendEquation(src.color_op));
equation_a.Assign(PackBlendEquation(src.alpha_op));
factor_source_rgb.Assign(PackBlendFactor(src.color_source));
factor_dest_rgb.Assign(PackBlendFactor(src.color_dest));
factor_source_a.Assign(PackBlendFactor(src.alpha_source));
factor_dest_a.Assign(PackBlendFactor(src.alpha_dest));
enable.Assign(1);
}
void FixedPipelineState::DynamicState::Refresh(const Maxwell& regs) {
u32 packed_front_face = PackFrontFace(regs.front_face);
if (regs.screen_y_control.triangle_rast_flip != 0) {
u32 packed_front_face = PackFrontFace(regs.gl_front_face);
if (regs.window_origin.flip_y != 0) {
// Flip front face
packed_front_face = 1 - packed_front_face;
}
raw1 = 0;
raw2 = 0;
front.action_stencil_fail.Assign(PackStencilOp(regs.stencil_front_op_fail));
front.action_depth_fail.Assign(PackStencilOp(regs.stencil_front_op_zfail));
front.action_depth_pass.Assign(PackStencilOp(regs.stencil_front_op_zpass));
front.test_func.Assign(PackComparisonOp(regs.stencil_front_func_func));
front.action_stencil_fail.Assign(PackStencilOp(regs.stencil_front_op.fail));
front.action_depth_fail.Assign(PackStencilOp(regs.stencil_front_op.zfail));
front.action_depth_pass.Assign(PackStencilOp(regs.stencil_front_op.zpass));
front.test_func.Assign(PackComparisonOp(regs.stencil_front_op.func));
if (regs.stencil_two_side_enable) {
back.action_stencil_fail.Assign(PackStencilOp(regs.stencil_back_op_fail));
back.action_depth_fail.Assign(PackStencilOp(regs.stencil_back_op_zfail));
back.action_depth_pass.Assign(PackStencilOp(regs.stencil_back_op_zpass));
back.test_func.Assign(PackComparisonOp(regs.stencil_back_func_func));
back.action_stencil_fail.Assign(PackStencilOp(regs.stencil_back_op.fail));
back.action_depth_fail.Assign(PackStencilOp(regs.stencil_back_op.zfail));
back.action_depth_pass.Assign(PackStencilOp(regs.stencil_back_op.zpass));
back.test_func.Assign(PackComparisonOp(regs.stencil_back_op.func));
} else {
back.action_stencil_fail.Assign(front.action_stencil_fail);
back.action_depth_fail.Assign(front.action_depth_fail);
@ -222,9 +224,9 @@ void FixedPipelineState::DynamicState::Refresh(const Maxwell& regs) {
depth_test_enable.Assign(regs.depth_test_enable);
front_face.Assign(packed_front_face);
depth_test_func.Assign(PackComparisonOp(regs.depth_test_func));
cull_face.Assign(PackCullFace(regs.cull_face));
cull_enable.Assign(regs.cull_test_enabled != 0 ? 1 : 0);
std::ranges::transform(regs.vertex_array, vertex_strides.begin(), [](const auto& array) {
cull_face.Assign(PackCullFace(regs.gl_cull_face));
cull_enable.Assign(regs.gl_cull_test_enabled != 0 ? 1 : 0);
std::ranges::transform(regs.vertex_streams, vertex_strides.begin(), [](const auto& array) {
return static_cast<u16>(array.stride.Value());
});
}
@ -251,41 +253,42 @@ Maxwell::ComparisonOp FixedPipelineState::UnpackComparisonOp(u32 packed) noexcep
return static_cast<Maxwell::ComparisonOp>(packed + 1);
}
u32 FixedPipelineState::PackStencilOp(Maxwell::StencilOp op) noexcept {
u32 FixedPipelineState::PackStencilOp(Maxwell::StencilOp::Op op) noexcept {
switch (op) {
case Maxwell::StencilOp::Keep:
case Maxwell::StencilOp::KeepOGL:
case Maxwell::StencilOp::Op::Keep_D3D:
case Maxwell::StencilOp::Op::Keep_GL:
return 0;
case Maxwell::StencilOp::Zero:
case Maxwell::StencilOp::ZeroOGL:
case Maxwell::StencilOp::Op::Zero_D3D:
case Maxwell::StencilOp::Op::Zero_GL:
return 1;
case Maxwell::StencilOp::Replace:
case Maxwell::StencilOp::ReplaceOGL:
case Maxwell::StencilOp::Op::Replace_D3D:
case Maxwell::StencilOp::Op::Replace_GL:
return 2;
case Maxwell::StencilOp::Incr:
case Maxwell::StencilOp::IncrOGL:
case Maxwell::StencilOp::Op::IncrSaturate_D3D:
case Maxwell::StencilOp::Op::IncrSaturate_GL:
return 3;
case Maxwell::StencilOp::Decr:
case Maxwell::StencilOp::DecrOGL:
case Maxwell::StencilOp::Op::DecrSaturate_D3D:
case Maxwell::StencilOp::Op::DecrSaturate_GL:
return 4;
case Maxwell::StencilOp::Invert:
case Maxwell::StencilOp::InvertOGL:
case Maxwell::StencilOp::Op::Invert_D3D:
case Maxwell::StencilOp::Op::Invert_GL:
return 5;
case Maxwell::StencilOp::IncrWrap:
case Maxwell::StencilOp::IncrWrapOGL:
case Maxwell::StencilOp::Op::Incr_D3D:
case Maxwell::StencilOp::Op::Incr_GL:
return 6;
case Maxwell::StencilOp::DecrWrap:
case Maxwell::StencilOp::DecrWrapOGL:
case Maxwell::StencilOp::Op::Decr_D3D:
case Maxwell::StencilOp::Op::Decr_GL:
return 7;
}
return 0;
}
Maxwell::StencilOp FixedPipelineState::UnpackStencilOp(u32 packed) noexcept {
static constexpr std::array LUT = {Maxwell::StencilOp::Keep, Maxwell::StencilOp::Zero,
Maxwell::StencilOp::Replace, Maxwell::StencilOp::Incr,
Maxwell::StencilOp::Decr, Maxwell::StencilOp::Invert,
Maxwell::StencilOp::IncrWrap, Maxwell::StencilOp::DecrWrap};
Maxwell::StencilOp::Op FixedPipelineState::UnpackStencilOp(u32 packed) noexcept {
static constexpr std::array LUT = {
Maxwell::StencilOp::Op::Keep_D3D, Maxwell::StencilOp::Op::Zero_D3D,
Maxwell::StencilOp::Op::Replace_D3D, Maxwell::StencilOp::Op::IncrSaturate_D3D,
Maxwell::StencilOp::Op::DecrSaturate_D3D, Maxwell::StencilOp::Op::Invert_D3D,
Maxwell::StencilOp::Op::Incr_D3D, Maxwell::StencilOp::Op::Decr_D3D};
return LUT[packed];
}
@ -318,30 +321,30 @@ Maxwell::PolygonMode FixedPipelineState::UnpackPolygonMode(u32 packed) noexcept
return static_cast<Maxwell::PolygonMode>(packed + 0x1B00);
}
u32 FixedPipelineState::PackLogicOp(Maxwell::LogicOperation op) noexcept {
u32 FixedPipelineState::PackLogicOp(Maxwell::LogicOp::Op op) noexcept {
return static_cast<u32>(op) - 0x1500;
}
Maxwell::LogicOperation FixedPipelineState::UnpackLogicOp(u32 packed) noexcept {
return static_cast<Maxwell::LogicOperation>(packed + 0x1500);
Maxwell::LogicOp::Op FixedPipelineState::UnpackLogicOp(u32 packed) noexcept {
return static_cast<Maxwell::LogicOp::Op>(packed + 0x1500);
}
u32 FixedPipelineState::PackBlendEquation(Maxwell::Blend::Equation equation) noexcept {
switch (equation) {
case Maxwell::Blend::Equation::Add:
case Maxwell::Blend::Equation::AddGL:
case Maxwell::Blend::Equation::Add_D3D:
case Maxwell::Blend::Equation::Add_GL:
return 0;
case Maxwell::Blend::Equation::Subtract:
case Maxwell::Blend::Equation::SubtractGL:
case Maxwell::Blend::Equation::Subtract_D3D:
case Maxwell::Blend::Equation::Subtract_GL:
return 1;
case Maxwell::Blend::Equation::ReverseSubtract:
case Maxwell::Blend::Equation::ReverseSubtractGL:
case Maxwell::Blend::Equation::ReverseSubtract_D3D:
case Maxwell::Blend::Equation::ReverseSubtract_GL:
return 2;
case Maxwell::Blend::Equation::Min:
case Maxwell::Blend::Equation::MinGL:
case Maxwell::Blend::Equation::Min_D3D:
case Maxwell::Blend::Equation::Min_GL:
return 3;
case Maxwell::Blend::Equation::Max:
case Maxwell::Blend::Equation::MaxGL:
case Maxwell::Blend::Equation::Max_D3D:
case Maxwell::Blend::Equation::Max_GL:
return 4;
}
return 0;
@ -349,97 +352,99 @@ u32 FixedPipelineState::PackBlendEquation(Maxwell::Blend::Equation equation) noe
Maxwell::Blend::Equation FixedPipelineState::UnpackBlendEquation(u32 packed) noexcept {
static constexpr std::array LUT = {
Maxwell::Blend::Equation::Add, Maxwell::Blend::Equation::Subtract,
Maxwell::Blend::Equation::ReverseSubtract, Maxwell::Blend::Equation::Min,
Maxwell::Blend::Equation::Max};
Maxwell::Blend::Equation::Add_D3D, Maxwell::Blend::Equation::Subtract_D3D,
Maxwell::Blend::Equation::ReverseSubtract_D3D, Maxwell::Blend::Equation::Min_D3D,
Maxwell::Blend::Equation::Max_D3D};
return LUT[packed];
}
u32 FixedPipelineState::PackBlendFactor(Maxwell::Blend::Factor factor) noexcept {
switch (factor) {
case Maxwell::Blend::Factor::Zero:
case Maxwell::Blend::Factor::ZeroGL:
case Maxwell::Blend::Factor::Zero_D3D:
case Maxwell::Blend::Factor::Zero_GL:
return 0;
case Maxwell::Blend::Factor::One:
case Maxwell::Blend::Factor::OneGL:
case Maxwell::Blend::Factor::One_D3D:
case Maxwell::Blend::Factor::One_GL:
return 1;
case Maxwell::Blend::Factor::SourceColor:
case Maxwell::Blend::Factor::SourceColorGL:
case Maxwell::Blend::Factor::SourceColor_D3D:
case Maxwell::Blend::Factor::SourceColor_GL:
return 2;
case Maxwell::Blend::Factor::OneMinusSourceColor:
case Maxwell::Blend::Factor::OneMinusSourceColorGL:
case Maxwell::Blend::Factor::OneMinusSourceColor_D3D:
case Maxwell::Blend::Factor::OneMinusSourceColor_GL:
return 3;
case Maxwell::Blend::Factor::SourceAlpha:
case Maxwell::Blend::Factor::SourceAlphaGL:
case Maxwell::Blend::Factor::SourceAlpha_D3D:
case Maxwell::Blend::Factor::SourceAlpha_GL:
return 4;
case Maxwell::Blend::Factor::OneMinusSourceAlpha:
case Maxwell::Blend::Factor::OneMinusSourceAlphaGL:
case Maxwell::Blend::Factor::OneMinusSourceAlpha_D3D:
case Maxwell::Blend::Factor::OneMinusSourceAlpha_GL:
return 5;
case Maxwell::Blend::Factor::DestAlpha:
case Maxwell::Blend::Factor::DestAlphaGL:
case Maxwell::Blend::Factor::DestAlpha_D3D:
case Maxwell::Blend::Factor::DestAlpha_GL:
return 6;
case Maxwell::Blend::Factor::OneMinusDestAlpha:
case Maxwell::Blend::Factor::OneMinusDestAlphaGL:
case Maxwell::Blend::Factor::OneMinusDestAlpha_D3D:
case Maxwell::Blend::Factor::OneMinusDestAlpha_GL:
return 7;
case Maxwell::Blend::Factor::DestColor:
case Maxwell::Blend::Factor::DestColorGL:
case Maxwell::Blend::Factor::DestColor_D3D:
case Maxwell::Blend::Factor::DestColor_GL:
return 8;
case Maxwell::Blend::Factor::OneMinusDestColor:
case Maxwell::Blend::Factor::OneMinusDestColorGL:
case Maxwell::Blend::Factor::OneMinusDestColor_D3D:
case Maxwell::Blend::Factor::OneMinusDestColor_GL:
return 9;
case Maxwell::Blend::Factor::SourceAlphaSaturate:
case Maxwell::Blend::Factor::SourceAlphaSaturateGL:
case Maxwell::Blend::Factor::SourceAlphaSaturate_D3D:
case Maxwell::Blend::Factor::SourceAlphaSaturate_GL:
return 10;
case Maxwell::Blend::Factor::Source1Color:
case Maxwell::Blend::Factor::Source1ColorGL:
case Maxwell::Blend::Factor::Source1Color_D3D:
case Maxwell::Blend::Factor::Source1Color_GL:
return 11;
case Maxwell::Blend::Factor::OneMinusSource1Color:
case Maxwell::Blend::Factor::OneMinusSource1ColorGL:
case Maxwell::Blend::Factor::OneMinusSource1Color_D3D:
case Maxwell::Blend::Factor::OneMinusSource1Color_GL:
return 12;
case Maxwell::Blend::Factor::Source1Alpha:
case Maxwell::Blend::Factor::Source1AlphaGL:
case Maxwell::Blend::Factor::Source1Alpha_D3D:
case Maxwell::Blend::Factor::Source1Alpha_GL:
return 13;
case Maxwell::Blend::Factor::OneMinusSource1Alpha:
case Maxwell::Blend::Factor::OneMinusSource1AlphaGL:
case Maxwell::Blend::Factor::OneMinusSource1Alpha_D3D:
case Maxwell::Blend::Factor::OneMinusSource1Alpha_GL:
return 14;
case Maxwell::Blend::Factor::ConstantColor:
case Maxwell::Blend::Factor::ConstantColorGL:
case Maxwell::Blend::Factor::BlendFactor_D3D:
case Maxwell::Blend::Factor::ConstantColor_GL:
return 15;
case Maxwell::Blend::Factor::OneMinusConstantColor:
case Maxwell::Blend::Factor::OneMinusConstantColorGL:
case Maxwell::Blend::Factor::OneMinusBlendFactor_D3D:
case Maxwell::Blend::Factor::OneMinusConstantColor_GL:
return 16;
case Maxwell::Blend::Factor::ConstantAlpha:
case Maxwell::Blend::Factor::ConstantAlphaGL:
case Maxwell::Blend::Factor::BothSourceAlpha_D3D:
case Maxwell::Blend::Factor::ConstantAlpha_GL:
return 17;
case Maxwell::Blend::Factor::OneMinusConstantAlpha:
case Maxwell::Blend::Factor::OneMinusConstantAlphaGL:
case Maxwell::Blend::Factor::OneMinusBothSourceAlpha_D3D:
case Maxwell::Blend::Factor::OneMinusConstantAlpha_GL:
return 18;
}
UNIMPLEMENTED_MSG("Unknown blend factor {}", static_cast<u32>(factor));
return 0;
}
Maxwell::Blend::Factor FixedPipelineState::UnpackBlendFactor(u32 packed) noexcept {
static constexpr std::array LUT = {
Maxwell::Blend::Factor::Zero,
Maxwell::Blend::Factor::One,
Maxwell::Blend::Factor::SourceColor,
Maxwell::Blend::Factor::OneMinusSourceColor,
Maxwell::Blend::Factor::SourceAlpha,
Maxwell::Blend::Factor::OneMinusSourceAlpha,
Maxwell::Blend::Factor::DestAlpha,
Maxwell::Blend::Factor::OneMinusDestAlpha,
Maxwell::Blend::Factor::DestColor,
Maxwell::Blend::Factor::OneMinusDestColor,
Maxwell::Blend::Factor::SourceAlphaSaturate,
Maxwell::Blend::Factor::Source1Color,
Maxwell::Blend::Factor::OneMinusSource1Color,
Maxwell::Blend::Factor::Source1Alpha,
Maxwell::Blend::Factor::OneMinusSource1Alpha,
Maxwell::Blend::Factor::ConstantColor,
Maxwell::Blend::Factor::OneMinusConstantColor,
Maxwell::Blend::Factor::ConstantAlpha,
Maxwell::Blend::Factor::OneMinusConstantAlpha,
Maxwell::Blend::Factor::Zero_D3D,
Maxwell::Blend::Factor::One_D3D,
Maxwell::Blend::Factor::SourceColor_D3D,
Maxwell::Blend::Factor::OneMinusSourceColor_D3D,
Maxwell::Blend::Factor::SourceAlpha_D3D,
Maxwell::Blend::Factor::OneMinusSourceAlpha_D3D,
Maxwell::Blend::Factor::DestAlpha_D3D,
Maxwell::Blend::Factor::OneMinusDestAlpha_D3D,
Maxwell::Blend::Factor::DestColor_D3D,
Maxwell::Blend::Factor::OneMinusDestColor_D3D,
Maxwell::Blend::Factor::SourceAlphaSaturate_D3D,
Maxwell::Blend::Factor::Source1Color_D3D,
Maxwell::Blend::Factor::OneMinusSource1Color_D3D,
Maxwell::Blend::Factor::Source1Alpha_D3D,
Maxwell::Blend::Factor::OneMinusSource1Alpha_D3D,
Maxwell::Blend::Factor::BlendFactor_D3D,
Maxwell::Blend::Factor::OneMinusBlendFactor_D3D,
Maxwell::Blend::Factor::BothSourceAlpha_D3D,
Maxwell::Blend::Factor::OneMinusBothSourceAlpha_D3D,
};
ASSERT(packed < LUT.size());
return LUT[packed];
}

15
src/video_core/renderer_vulkan/fixed_pipeline_state.h
View file

@ -21,8 +21,8 @@ struct FixedPipelineState {
static u32 PackComparisonOp(Maxwell::ComparisonOp op) noexcept;
static Maxwell::ComparisonOp UnpackComparisonOp(u32 packed) noexcept;
static u32 PackStencilOp(Maxwell::StencilOp op) noexcept;
static Maxwell::StencilOp UnpackStencilOp(u32 packed) noexcept;
static u32 PackStencilOp(Maxwell::StencilOp::Op op) noexcept;
static Maxwell::StencilOp::Op UnpackStencilOp(u32 packed) noexcept;
static u32 PackCullFace(Maxwell::CullFace cull) noexcept;
static Maxwell::CullFace UnpackCullFace(u32 packed) noexcept;
@ -33,8 +33,8 @@ struct FixedPipelineState {
static u32 PackPolygonMode(Maxwell::PolygonMode mode) noexcept;
static Maxwell::PolygonMode UnpackPolygonMode(u32 packed) noexcept;
static u32 PackLogicOp(Maxwell::LogicOperation op) noexcept;
static Maxwell::LogicOperation UnpackLogicOp(u32 packed) noexcept;
static u32 PackLogicOp(Maxwell::LogicOp::Op op) noexcept;
static Maxwell::LogicOp::Op UnpackLogicOp(u32 packed) noexcept;
static u32 PackBlendEquation(Maxwell::Blend::Equation equation) noexcept;
static Maxwell::Blend::Equation UnpackBlendEquation(u32 packed) noexcept;
@ -113,15 +113,15 @@ struct FixedPipelineState {
BitField<Position + 6, 3, u32> action_depth_pass;
BitField<Position + 9, 3, u32> test_func;
Maxwell::StencilOp ActionStencilFail() const noexcept {
Maxwell::StencilOp::Op ActionStencilFail() const noexcept {
return UnpackStencilOp(action_stencil_fail);
}
Maxwell::StencilOp ActionDepthFail() const noexcept {
Maxwell::StencilOp::Op ActionDepthFail() const noexcept {
return UnpackStencilOp(action_depth_fail);
}
Maxwell::StencilOp ActionDepthPass() const noexcept {
Maxwell::StencilOp::Op ActionDepthPass() const noexcept {
return UnpackStencilOp(action_depth_pass);
}
@ -193,7 +193,6 @@ struct FixedPipelineState {
BitField<6, 5, u32> depth_format;
BitField<11, 1, u32> y_negate;
BitField<12, 1, u32> provoking_vertex_last;
BitField<13, 1, u32> conservative_raster_enable;
BitField<14, 1, u32> smooth_lines;
};
std::array<u8, Maxwell::NumRenderTargets> color_formats;

337
src/video_core/renderer_vulkan/maxwell_to_vk.cpp
View file

@ -323,161 +323,182 @@ VkFormat VertexFormat(const Device& device, Maxwell::VertexAttribute::Type type,
Maxwell::VertexAttribute::Size size) {
const VkFormat format{([&]() {
switch (type) {
case Maxwell::VertexAttribute::Type::UnsignedNorm:
case Maxwell::VertexAttribute::Type::UnusedEnumDoNotUseBecauseItWillGoAway:
ASSERT_MSG(false, "Invalid vertex attribute type!");
break;
case Maxwell::VertexAttribute::Type::UNorm:
switch (size) {
case Maxwell::VertexAttribute::Size::Size_8:
case Maxwell::VertexAttribute::Size::Size_R8:
case Maxwell::VertexAttribute::Size::Size_A8:
return VK_FORMAT_R8_UNORM;
case Maxwell::VertexAttribute::Size::Size_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8:
case Maxwell::VertexAttribute::Size::Size_G8_R8:
return VK_FORMAT_R8G8_UNORM;
case Maxwell::VertexAttribute::Size::Size_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8:
return VK_FORMAT_R8G8B8_UNORM;
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8_A8:
case Maxwell::VertexAttribute::Size::Size_X8_B8_G8_R8:
return VK_FORMAT_R8G8B8A8_UNORM;
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_R16:
return VK_FORMAT_R16_UNORM;
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
return VK_FORMAT_R16G16_UNORM;
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
return VK_FORMAT_R16G16B16_UNORM;
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return VK_FORMAT_R16G16B16A16_UNORM;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
case Maxwell::VertexAttribute::Size::Size_A2_B10_G10_R10:
return VK_FORMAT_A2B10G10R10_UNORM_PACK32;
default:
break;
}
break;
case Maxwell::VertexAttribute::Type::SignedNorm:
case Maxwell::VertexAttribute::Type::SNorm:
switch (size) {
case Maxwell::VertexAttribute::Size::Size_8:
case Maxwell::VertexAttribute::Size::Size_R8:
case Maxwell::VertexAttribute::Size::Size_A8:
return VK_FORMAT_R8_SNORM;
case Maxwell::VertexAttribute::Size::Size_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8:
case Maxwell::VertexAttribute::Size::Size_G8_R8:
return VK_FORMAT_R8G8_SNORM;
case Maxwell::VertexAttribute::Size::Size_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8:
return VK_FORMAT_R8G8B8_SNORM;
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8_A8:
case Maxwell::VertexAttribute::Size::Size_X8_B8_G8_R8:
return VK_FORMAT_R8G8B8A8_SNORM;
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_R16:
return VK_FORMAT_R16_SNORM;
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
return VK_FORMAT_R16G16_SNORM;
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
return VK_FORMAT_R16G16B16_SNORM;
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return VK_FORMAT_R16G16B16A16_SNORM;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
case Maxwell::VertexAttribute::Size::Size_A2_B10_G10_R10:
return VK_FORMAT_A2B10G10R10_SNORM_PACK32;
default:
break;
}
break;
case Maxwell::VertexAttribute::Type::UnsignedScaled:
case Maxwell::VertexAttribute::Type::UScaled:
switch (size) {
case Maxwell::VertexAttribute::Size::Size_8:
case Maxwell::VertexAttribute::Size::Size_R8:
case Maxwell::VertexAttribute::Size::Size_A8:
return VK_FORMAT_R8_USCALED;
case Maxwell::VertexAttribute::Size::Size_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8:
case Maxwell::VertexAttribute::Size::Size_G8_R8:
return VK_FORMAT_R8G8_USCALED;
case Maxwell::VertexAttribute::Size::Size_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8:
return VK_FORMAT_R8G8B8_USCALED;
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8_A8:
case Maxwell::VertexAttribute::Size::Size_X8_B8_G8_R8:
return VK_FORMAT_R8G8B8A8_USCALED;
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_R16:
return VK_FORMAT_R16_USCALED;
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
return VK_FORMAT_R16G16_USCALED;
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
return VK_FORMAT_R16G16B16_USCALED;
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return VK_FORMAT_R16G16B16A16_USCALED;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
case Maxwell::VertexAttribute::Size::Size_A2_B10_G10_R10:
return VK_FORMAT_A2B10G10R10_USCALED_PACK32;
default:
break;
}
break;
case Maxwell::VertexAttribute::Type::SignedScaled:
case Maxwell::VertexAttribute::Type::SScaled:
switch (size) {
case Maxwell::VertexAttribute::Size::Size_8:
case Maxwell::VertexAttribute::Size::Size_R8:
case Maxwell::VertexAttribute::Size::Size_A8:
return VK_FORMAT_R8_SSCALED;
case Maxwell::VertexAttribute::Size::Size_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8:
case Maxwell::VertexAttribute::Size::Size_G8_R8:
return VK_FORMAT_R8G8_SSCALED;
case Maxwell::VertexAttribute::Size::Size_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8:
return VK_FORMAT_R8G8B8_SSCALED;
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8_A8:
case Maxwell::VertexAttribute::Size::Size_X8_B8_G8_R8:
return VK_FORMAT_R8G8B8A8_SSCALED;
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_R16:
return VK_FORMAT_R16_SSCALED;
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
return VK_FORMAT_R16G16_SSCALED;
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
return VK_FORMAT_R16G16B16_SSCALED;
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return VK_FORMAT_R16G16B16A16_SSCALED;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
case Maxwell::VertexAttribute::Size::Size_A2_B10_G10_R10:
return VK_FORMAT_A2B10G10R10_SSCALED_PACK32;
default:
break;
}
break;
case Maxwell::VertexAttribute::Type::UnsignedInt:
case Maxwell::VertexAttribute::Type::UInt:
switch (size) {
case Maxwell::VertexAttribute::Size::Size_8:
case Maxwell::VertexAttribute::Size::Size_R8:
case Maxwell::VertexAttribute::Size::Size_A8:
return VK_FORMAT_R8_UINT;
case Maxwell::VertexAttribute::Size::Size_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8:
case Maxwell::VertexAttribute::Size::Size_G8_R8:
return VK_FORMAT_R8G8_UINT;
case Maxwell::VertexAttribute::Size::Size_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8:
return VK_FORMAT_R8G8B8_UINT;
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8_A8:
case Maxwell::VertexAttribute::Size::Size_X8_B8_G8_R8:
return VK_FORMAT_R8G8B8A8_UINT;
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_R16:
return VK_FORMAT_R16_UINT;
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
return VK_FORMAT_R16G16_UINT;
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
return VK_FORMAT_R16G16B16_UINT;
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return VK_FORMAT_R16G16B16A16_UINT;
case Maxwell::VertexAttribute::Size::Size_32:
case Maxwell::VertexAttribute::Size::Size_R32:
return VK_FORMAT_R32_UINT;
case Maxwell::VertexAttribute::Size::Size_32_32:
case Maxwell::VertexAttribute::Size::Size_R32_G32:
return VK_FORMAT_R32G32_UINT;
case Maxwell::VertexAttribute::Size::Size_32_32_32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32:
return VK_FORMAT_R32G32B32_UINT;
case Maxwell::VertexAttribute::Size::Size_32_32_32_32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32_A32:
return VK_FORMAT_R32G32B32A32_UINT;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
case Maxwell::VertexAttribute::Size::Size_A2_B10_G10_R10:
return VK_FORMAT_A2B10G10R10_UINT_PACK32;
default:
break;
}
break;
case Maxwell::VertexAttribute::Type::SignedInt:
case Maxwell::VertexAttribute::Type::SInt:
switch (size) {
case Maxwell::VertexAttribute::Size::Size_8:
case Maxwell::VertexAttribute::Size::Size_R8:
case Maxwell::VertexAttribute::Size::Size_A8:
return VK_FORMAT_R8_SINT;
case Maxwell::VertexAttribute::Size::Size_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8:
case Maxwell::VertexAttribute::Size::Size_G8_R8:
return VK_FORMAT_R8G8_SINT;
case Maxwell::VertexAttribute::Size::Size_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8:
return VK_FORMAT_R8G8B8_SINT;
case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
case Maxwell::VertexAttribute::Size::Size_R8_G8_B8_A8:
case Maxwell::VertexAttribute::Size::Size_X8_B8_G8_R8:
return VK_FORMAT_R8G8B8A8_SINT;
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_R16:
return VK_FORMAT_R16_SINT;
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
return VK_FORMAT_R16G16_SINT;
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
return VK_FORMAT_R16G16B16_SINT;
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return VK_FORMAT_R16G16B16A16_SINT;
case Maxwell::VertexAttribute::Size::Size_32:
case Maxwell::VertexAttribute::Size::Size_R32:
return VK_FORMAT_R32_SINT;
case Maxwell::VertexAttribute::Size::Size_32_32:
case Maxwell::VertexAttribute::Size::Size_R32_G32:
return VK_FORMAT_R32G32_SINT;
case Maxwell::VertexAttribute::Size::Size_32_32_32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32:
return VK_FORMAT_R32G32B32_SINT;
case Maxwell::VertexAttribute::Size::Size_32_32_32_32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32_A32:
return VK_FORMAT_R32G32B32A32_SINT;
case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
case Maxwell::VertexAttribute::Size::Size_A2_B10_G10_R10:
return VK_FORMAT_A2B10G10R10_SINT_PACK32;
default:
break;
@ -485,23 +506,23 @@ VkFormat VertexFormat(const Device& device, Maxwell::VertexAttribute::Type type,
break;
case Maxwell::VertexAttribute::Type::Float:
switch (size) {
case Maxwell::VertexAttribute::Size::Size_16:
case Maxwell::VertexAttribute::Size::Size_R16:
return VK_FORMAT_R16_SFLOAT;
case Maxwell::VertexAttribute::Size::Size_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16:
return VK_FORMAT_R16G16_SFLOAT;
case Maxwell::VertexAttribute::Size::Size_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16:
return VK_FORMAT_R16G16B16_SFLOAT;
case Maxwell::VertexAttribute::Size::Size_16_16_16_16:
case Maxwell::VertexAttribute::Size::Size_R16_G16_B16_A16:
return VK_FORMAT_R16G16B16A16_SFLOAT;
case Maxwell::VertexAttribute::Size::Size_32:
case Maxwell::VertexAttribute::Size::Size_R32:
return VK_FORMAT_R32_SFLOAT;
case Maxwell::VertexAttribute::Size::Size_32_32:
case Maxwell::VertexAttribute::Size::Size_R32_G32:
return VK_FORMAT_R32G32_SFLOAT;
case Maxwell::VertexAttribute::Size::Size_32_32_32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32:
return VK_FORMAT_R32G32B32_SFLOAT;
case Maxwell::VertexAttribute::Size::Size_32_32_32_32:
case Maxwell::VertexAttribute::Size::Size_R32_G32_B32_A32:
return VK_FORMAT_R32G32B32A32_SFLOAT;
case Maxwell::VertexAttribute::Size::Size_11_11_10:
case Maxwell::VertexAttribute::Size::Size_B10_G11_R11:
return VK_FORMAT_B10G11R11_UFLOAT_PACK32;
default:
break;
@ -521,29 +542,29 @@ VkFormat VertexFormat(const Device& device, Maxwell::VertexAttribute::Type type,
VkCompareOp ComparisonOp(Maxwell::ComparisonOp comparison) {
switch (comparison) {
case Maxwell::ComparisonOp::Never:
case Maxwell::ComparisonOp::NeverOld:
case Maxwell::ComparisonOp::Never_D3D:
case Maxwell::ComparisonOp::Never_GL:
return VK_COMPARE_OP_NEVER;
case Maxwell::ComparisonOp::Less:
case Maxwell::ComparisonOp::LessOld:
case Maxwell::ComparisonOp::Less_D3D:
case Maxwell::ComparisonOp::Less_GL:
return VK_COMPARE_OP_LESS;
case Maxwell::ComparisonOp::Equal:
case Maxwell::ComparisonOp::EqualOld:
case Maxwell::ComparisonOp::Equal_D3D:
case Maxwell::ComparisonOp::Equal_GL:
return VK_COMPARE_OP_EQUAL;
case Maxwell::ComparisonOp::LessEqual:
case Maxwell::ComparisonOp::LessEqualOld:
case Maxwell::ComparisonOp::LessEqual_D3D:
case Maxwell::ComparisonOp::LessEqual_GL:
return VK_COMPARE_OP_LESS_OR_EQUAL;
case Maxwell::ComparisonOp::Greater:
case Maxwell::ComparisonOp::GreaterOld:
case Maxwell::ComparisonOp::Greater_D3D:
case Maxwell::ComparisonOp::Greater_GL:
return VK_COMPARE_OP_GREATER;
case Maxwell::ComparisonOp::NotEqual:
case Maxwell::ComparisonOp::NotEqualOld:
case Maxwell::ComparisonOp::NotEqual_D3D:
case Maxwell::ComparisonOp::NotEqual_GL:
return VK_COMPARE_OP_NOT_EQUAL;
case Maxwell::ComparisonOp::GreaterEqual:
case Maxwell::ComparisonOp::GreaterEqualOld:
case Maxwell::ComparisonOp::GreaterEqual_D3D:
case Maxwell::ComparisonOp::GreaterEqual_GL:
return VK_COMPARE_OP_GREATER_OR_EQUAL;
case Maxwell::ComparisonOp::Always:
case Maxwell::ComparisonOp::AlwaysOld:
case Maxwell::ComparisonOp::Always_D3D:
case Maxwell::ComparisonOp::Always_GL:
return VK_COMPARE_OP_ALWAYS;
}
UNIMPLEMENTED_MSG("Unimplemented comparison op={}", comparison);
@ -563,31 +584,31 @@ VkIndexType IndexFormat(Maxwell::IndexFormat index_format) {
return {};
}
VkStencilOp StencilOp(Maxwell::StencilOp stencil_op) {
VkStencilOp StencilOp(Maxwell::StencilOp::Op stencil_op) {
switch (stencil_op) {
case Maxwell::StencilOp::Keep:
case Maxwell::StencilOp::KeepOGL:
case Maxwell::StencilOp::Op::Keep_D3D:
case Maxwell::StencilOp::Op::Keep_GL:
return VK_STENCIL_OP_KEEP;
case Maxwell::StencilOp::Zero:
case Maxwell::StencilOp::ZeroOGL:
case Maxwell::StencilOp::Op::Zero_D3D:
case Maxwell::StencilOp::Op::Zero_GL:
return VK_STENCIL_OP_ZERO;
case Maxwell::StencilOp::Replace:
case Maxwell::StencilOp::ReplaceOGL:
case Maxwell::StencilOp::Op::Replace_D3D:
case Maxwell::StencilOp::Op::Replace_GL:
return VK_STENCIL_OP_REPLACE;
case Maxwell::StencilOp::Incr:
case Maxwell::StencilOp::IncrOGL:
case Maxwell::StencilOp::Op::IncrSaturate_D3D:
case Maxwell::StencilOp::Op::IncrSaturate_GL:
return VK_STENCIL_OP_INCREMENT_AND_CLAMP;
case Maxwell::StencilOp::Decr:
case Maxwell::StencilOp::DecrOGL:
case Maxwell::StencilOp::Op::DecrSaturate_D3D:
case Maxwell::StencilOp::Op::DecrSaturate_GL:
return VK_STENCIL_OP_DECREMENT_AND_CLAMP;
case Maxwell::StencilOp::Invert:
case Maxwell::StencilOp::InvertOGL:
case Maxwell::StencilOp::Op::Invert_D3D:
case Maxwell::StencilOp::Op::Invert_GL:
return VK_STENCIL_OP_INVERT;
case Maxwell::StencilOp::IncrWrap:
case Maxwell::StencilOp::IncrWrapOGL:
case Maxwell::StencilOp::Op::Incr_D3D:
case Maxwell::StencilOp::Op::Incr_GL:
return VK_STENCIL_OP_INCREMENT_AND_WRAP;
case Maxwell::StencilOp::DecrWrap:
case Maxwell::StencilOp::DecrWrapOGL:
case Maxwell::StencilOp::Op::Decr_D3D:
case Maxwell::StencilOp::Op::Decr_GL:
return VK_STENCIL_OP_DECREMENT_AND_WRAP;
}
UNIMPLEMENTED_MSG("Unimplemented stencil op={}", stencil_op);
@ -596,20 +617,20 @@ VkStencilOp StencilOp(Maxwell::StencilOp stencil_op) {
VkBlendOp BlendEquation(Maxwell::Blend::Equation equation) {
switch (equation) {
case Maxwell::Blend::Equation::Add:
case Maxwell::Blend::Equation::AddGL:
case Maxwell::Blend::Equation::Add_D3D:
case Maxwell::Blend::Equation::Add_GL:
return VK_BLEND_OP_ADD;
case Maxwell::Blend::Equation::Subtract:
case Maxwell::Blend::Equation::SubtractGL:
case Maxwell::Blend::Equation::Subtract_D3D:
case Maxwell::Blend::Equation::Subtract_GL:
return VK_BLEND_OP_SUBTRACT;
case Maxwell::Blend::Equation::ReverseSubtract:
case Maxwell::Blend::Equation::ReverseSubtractGL:
case Maxwell::Blend::Equation::ReverseSubtract_D3D:
case Maxwell::Blend::Equation::ReverseSubtract_GL:
return VK_BLEND_OP_REVERSE_SUBTRACT;
case Maxwell::Blend::Equation::Min:
case Maxwell::Blend::Equation::MinGL:
case Maxwell::Blend::Equation::Min_D3D:
case Maxwell::Blend::Equation::Min_GL:
return VK_BLEND_OP_MIN;
case Maxwell::Blend::Equation::Max:
case Maxwell::Blend::Equation::MaxGL:
case Maxwell::Blend::Equation::Max_D3D:
case Maxwell::Blend::Equation::Max_GL:
return VK_BLEND_OP_MAX;
}
UNIMPLEMENTED_MSG("Unimplemented blend equation={}", equation);
@ -618,62 +639,62 @@ VkBlendOp BlendEquation(Maxwell::Blend::Equation equation) {
VkBlendFactor BlendFactor(Maxwell::Blend::Factor factor) {
switch (factor) {
case Maxwell::Blend::Factor::Zero:
case Maxwell::Blend::Factor::ZeroGL:
case Maxwell::Blend::Factor::Zero_D3D:
case Maxwell::Blend::Factor::Zero_GL:
return VK_BLEND_FACTOR_ZERO;
case Maxwell::Blend::Factor::One:
case Maxwell::Blend::Factor::OneGL:
case Maxwell::Blend::Factor::One_D3D:
case Maxwell::Blend::Factor::One_GL:
return VK_BLEND_FACTOR_ONE;
case Maxwell::Blend::Factor::SourceColor:
case Maxwell::Blend::Factor::SourceColorGL:
case Maxwell::Blend::Factor::SourceColor_D3D:
case Maxwell::Blend::Factor::SourceColor_GL:
return VK_BLEND_FACTOR_SRC_COLOR;
case Maxwell::Blend::Factor::OneMinusSourceColor:
case Maxwell::Blend::Factor::OneMinusSourceColorGL:
case Maxwell::Blend::Factor::OneMinusSourceColor_D3D:
case Maxwell::Blend::Factor::OneMinusSourceColor_GL:
return VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR;
case Maxwell::Blend::Factor::SourceAlpha:
case Maxwell::Blend::Factor::SourceAlphaGL:
case Maxwell::Blend::Factor::SourceAlpha_D3D:
case Maxwell::Blend::Factor::SourceAlpha_GL:
return VK_BLEND_FACTOR_SRC_ALPHA;
case Maxwell::Blend::Factor::OneMinusSourceAlpha:
case Maxwell::Blend::Factor::OneMinusSourceAlphaGL:
case Maxwell::Blend::Factor::OneMinusSourceAlpha_D3D:
case Maxwell::Blend::Factor::OneMinusSourceAlpha_GL:
return VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA;
case Maxwell::Blend::Factor::DestAlpha:
case Maxwell::Blend::Factor::DestAlphaGL:
case Maxwell::Blend::Factor::DestAlpha_D3D:
case Maxwell::Blend::Factor::DestAlpha_GL:
return VK_BLEND_FACTOR_DST_ALPHA;
case Maxwell::Blend::Factor::OneMinusDestAlpha:
case Maxwell::Blend::Factor::OneMinusDestAlphaGL:
case Maxwell::Blend::Factor::OneMinusDestAlpha_D3D:
case Maxwell::Blend::Factor::OneMinusDestAlpha_GL:
return VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA;
case Maxwell::Blend::Factor::DestColor:
case Maxwell::Blend::Factor::DestColorGL:
case Maxwell::Blend::Factor::DestColor_D3D:
case Maxwell::Blend::Factor::DestColor_GL:
return VK_BLEND_FACTOR_DST_COLOR;
case Maxwell::Blend::Factor::OneMinusDestColor:
case Maxwell::Blend::Factor::OneMinusDestColorGL:
case Maxwell::Blend::Factor::OneMinusDestColor_D3D:
case Maxwell::Blend::Factor::OneMinusDestColor_GL:
return VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR;
case Maxwell::Blend::Factor::SourceAlphaSaturate:
case Maxwell::Blend::Factor::SourceAlphaSaturateGL:
case Maxwell::Blend::Factor::SourceAlphaSaturate_D3D:
case Maxwell::Blend::Factor::SourceAlphaSaturate_GL:
return VK_BLEND_FACTOR_SRC_ALPHA_SATURATE;
case Maxwell::Blend::Factor::Source1Color:
case Maxwell::Blend::Factor::Source1ColorGL:
case Maxwell::Blend::Factor::Source1Color_D3D:
case Maxwell::Blend::Factor::Source1Color_GL:
return VK_BLEND_FACTOR_SRC1_COLOR;
case Maxwell::Blend::Factor::OneMinusSource1Color:
case Maxwell::Blend::Factor::OneMinusSource1ColorGL:
case Maxwell::Blend::Factor::OneMinusSource1Color_D3D:
case Maxwell::Blend::Factor::OneMinusSource1Color_GL:
return VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR;
case Maxwell::Blend::Factor::Source1Alpha:
case Maxwell::Blend::Factor::Source1AlphaGL:
case Maxwell::Blend::Factor::Source1Alpha_D3D:
case Maxwell::Blend::Factor::Source1Alpha_GL:
return VK_BLEND_FACTOR_SRC1_ALPHA;
case Maxwell::Blend::Factor::OneMinusSource1Alpha:
case Maxwell::Blend::Factor::OneMinusSource1AlphaGL:
case Maxwell::Blend::Factor::OneMinusSource1Alpha_D3D:
case Maxwell::Blend::Factor::OneMinusSource1Alpha_GL:
return VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA;
case Maxwell::Blend::Factor::ConstantColor:
case Maxwell::Blend::Factor::ConstantColorGL:
case Maxwell::Blend::Factor::BlendFactor_D3D:
case Maxwell::Blend::Factor::ConstantColor_GL:
return VK_BLEND_FACTOR_CONSTANT_COLOR;
case Maxwell::Blend::Factor::OneMinusConstantColor:
case Maxwell::Blend::Factor::OneMinusConstantColorGL:
case Maxwell::Blend::Factor::OneMinusBlendFactor_D3D:
case Maxwell::Blend::Factor::OneMinusConstantColor_GL:
return VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR;
case Maxwell::Blend::Factor::ConstantAlpha:
case Maxwell::Blend::Factor::ConstantAlphaGL:
case Maxwell::Blend::Factor::BothSourceAlpha_D3D:
case Maxwell::Blend::Factor::ConstantAlpha_GL:
return VK_BLEND_FACTOR_CONSTANT_ALPHA;
case Maxwell::Blend::Factor::OneMinusConstantAlpha:
case Maxwell::Blend::Factor::OneMinusConstantAlphaGL:
case Maxwell::Blend::Factor::OneMinusBothSourceAlpha_D3D:
case Maxwell::Blend::Factor::OneMinusConstantAlpha_GL:
return VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA;
}
UNIMPLEMENTED_MSG("Unimplemented blend factor={}", factor);

2
src/video_core/renderer_vulkan/maxwell_to_vk.h
View file

@ -55,7 +55,7 @@ VkCompareOp ComparisonOp(Maxwell::ComparisonOp comparison);
VkIndexType IndexFormat(Maxwell::IndexFormat index_format);
VkStencilOp StencilOp(Maxwell::StencilOp stencil_op);
VkStencilOp StencilOp(Maxwell::StencilOp::Op stencil_op);
VkBlendOp BlendEquation(Maxwell::Blend::Equation equation);

14
src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
View file

@ -288,7 +288,7 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
buffer_cache.SetUniformBuffersState(enabled_uniform_buffer_masks, &uniform_buffer_sizes);
const auto& regs{maxwell3d->regs};
const bool via_header_index{regs.sampler_index == Maxwell::SamplerIndex::ViaHeaderIndex};
const bool via_header_index{regs.sampler_binding == Maxwell::SamplerBinding::ViaHeaderBinding};
const auto config_stage{[&](size_t stage) LAMBDA_FORCEINLINE {
const Shader::Info& info{stage_infos[stage]};
buffer_cache.UnbindGraphicsStorageBuffers(stage);
@ -664,15 +664,6 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
.lineStippleFactor = 0,
.lineStipplePattern = 0,
};
VkPipelineRasterizationConservativeStateCreateInfoEXT conservative_raster{
.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_CONSERVATIVE_STATE_CREATE_INFO_EXT,
.pNext = nullptr,
.flags = 0,
.conservativeRasterizationMode = key.state.conservative_raster_enable != 0
? VK_CONSERVATIVE_RASTERIZATION_MODE_OVERESTIMATE_EXT
: VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT,
.extraPrimitiveOverestimationSize = 0.0f,
};
VkPipelineRasterizationProvokingVertexStateCreateInfoEXT provoking_vertex{
.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_PROVOKING_VERTEX_STATE_CREATE_INFO_EXT,
.pNext = nullptr,
@ -683,9 +674,6 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
if (IsLine(input_assembly_topology) && device.IsExtLineRasterizationSupported()) {
line_state.pNext = std::exchange(rasterization_ci.pNext, &line_state);
}
if (device.IsExtConservativeRasterizationSupported()) {
conservative_raster.pNext = std::exchange(rasterization_ci.pNext, &conservative_raster);
}
if (device.IsExtProvokingVertexSupported()) {
provoking_vertex.pNext = std::exchange(rasterization_ci.pNext, &provoking_vertex);
}

67
src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
View file

@ -62,29 +62,29 @@ auto MakeSpan(Container& container) {
Shader::CompareFunction MaxwellToCompareFunction(Maxwell::ComparisonOp comparison) {
switch (comparison) {
case Maxwell::ComparisonOp::Never:
case Maxwell::ComparisonOp::NeverOld:
case Maxwell::ComparisonOp::Never_D3D:
case Maxwell::ComparisonOp::Never_GL:
return Shader::CompareFunction::Never;
case Maxwell::ComparisonOp::Less:
case Maxwell::ComparisonOp::LessOld:
case Maxwell::ComparisonOp::Less_D3D:
case Maxwell::ComparisonOp::Less_GL:
return Shader::CompareFunction::Less;
case Maxwell::ComparisonOp::Equal:
case Maxwell::ComparisonOp::EqualOld:
case Maxwell::ComparisonOp::Equal_D3D:
case Maxwell::ComparisonOp::Equal_GL:
return Shader::CompareFunction::Equal;
case Maxwell::ComparisonOp::LessEqual:
case Maxwell::ComparisonOp::LessEqualOld:
case Maxwell::ComparisonOp::LessEqual_D3D:
case Maxwell::ComparisonOp::LessEqual_GL:
return Shader::CompareFunction::LessThanEqual;
case Maxwell::ComparisonOp::Greater:
case Maxwell::ComparisonOp::GreaterOld:
case Maxwell::ComparisonOp::Greater_D3D:
case Maxwell::ComparisonOp::Greater_GL:
return Shader::CompareFunction::Greater;
case Maxwell::ComparisonOp::NotEqual:
case Maxwell::ComparisonOp::NotEqualOld:
case Maxwell::ComparisonOp::NotEqual_D3D:
case Maxwell::ComparisonOp::NotEqual_GL:
return Shader::CompareFunction::NotEqual;
case Maxwell::ComparisonOp::GreaterEqual:
case Maxwell::ComparisonOp::GreaterEqualOld:
case Maxwell::ComparisonOp::GreaterEqual_D3D:
case Maxwell::ComparisonOp::GreaterEqual_GL:
return Shader::CompareFunction::GreaterThanEqual;
case Maxwell::ComparisonOp::Always:
case Maxwell::ComparisonOp::AlwaysOld:
case Maxwell::ComparisonOp::Always_D3D:
case Maxwell::ComparisonOp::Always_GL:
return Shader::CompareFunction::Always;
}
UNIMPLEMENTED_MSG("Unimplemented comparison op={}", comparison);
@ -96,15 +96,18 @@ Shader::AttributeType CastAttributeType(const FixedPipelineState::VertexAttribut
return Shader::AttributeType::Disabled;
}
switch (attr.Type()) {
case Maxwell::VertexAttribute::Type::SignedNorm:
case Maxwell::VertexAttribute::Type::UnsignedNorm:
case Maxwell::VertexAttribute::Type::UnsignedScaled:
case Maxwell::VertexAttribute::Type::SignedScaled:
case Maxwell::VertexAttribute::Type::UnusedEnumDoNotUseBecauseItWillGoAway:
ASSERT_MSG(false, "Invalid vertex attribute type!");
return Shader::AttributeType::Disabled;
case Maxwell::VertexAttribute::Type::SNorm:
case Maxwell::VertexAttribute::Type::UNorm:
case Maxwell::VertexAttribute::Type::UScaled:
case Maxwell::VertexAttribute::Type::SScaled:
case Maxwell::VertexAttribute::Type::Float:
return Shader::AttributeType::Float;
case Maxwell::VertexAttribute::Type::SignedInt:
case Maxwell::VertexAttribute::Type::SInt:
return Shader::AttributeType::SignedInt;
case Maxwell::VertexAttribute::Type::UnsignedInt:
case Maxwell::VertexAttribute::Type::UInt:
return Shader::AttributeType::UnsignedInt;
}
return Shader::AttributeType::Float;
@ -162,16 +165,14 @@ Shader::RuntimeInfo MakeRuntimeInfo(std::span<const Shader::IR::Program> program
}
break;
case Shader::Stage::TessellationEval:
// We have to flip tessellation clockwise for some reason...
info.tess_clockwise = key.state.tessellation_clockwise == 0;
info.tess_primitive = [&key] {
const u32 raw{key.state.tessellation_primitive.Value()};
switch (static_cast<Maxwell::TessellationPrimitive>(raw)) {
case Maxwell::TessellationPrimitive::Isolines:
switch (static_cast<Maxwell::Tessellation::DomainType>(raw)) {
case Maxwell::Tessellation::DomainType::Isolines:
return Shader::TessPrimitive::Isolines;
case Maxwell::TessellationPrimitive::Triangles:
case Maxwell::Tessellation::DomainType::Triangles:
return Shader::TessPrimitive::Triangles;
case Maxwell::TessellationPrimitive::Quads:
case Maxwell::Tessellation::DomainType::Quads:
return Shader::TessPrimitive::Quads;
}
ASSERT(false);
@ -179,12 +180,12 @@ Shader::RuntimeInfo MakeRuntimeInfo(std::span<const Shader::IR::Program> program
}();
info.tess_spacing = [&] {
const u32 raw{key.state.tessellation_spacing};
switch (static_cast<Maxwell::TessellationSpacing>(raw)) {
case Maxwell::TessellationSpacing::Equal:
switch (static_cast<Maxwell::Tessellation::Spacing>(raw)) {
case Maxwell::Tessellation::Spacing::Integer:
return Shader::TessSpacing::Equal;
case Maxwell::TessellationSpacing::FractionalOdd:
case Maxwell::Tessellation::Spacing::FractionalOdd:
return Shader::TessSpacing::FractionalOdd;
case Maxwell::TessellationSpacing::FractionalEven:
case Maxwell::Tessellation::Spacing::FractionalEven:
return Shader::TessSpacing::FractionalEven;
}
ASSERT(false);
@ -490,7 +491,7 @@ GraphicsPipeline* PipelineCache::BuiltPipeline(GraphicsPipeline* pipeline) const
// If games are using a small index count, we can assume these are full screen quads.
// Usually these shaders are only used once for building textures so we can assume they
// can't be built async
if (maxwell3d->regs.index_array.count <= 6 || maxwell3d->regs.vertex_buffer.count <= 6) {
if (maxwell3d->regs.index_buffer.count <= 6 || maxwell3d->regs.vertex_buffer.count <= 6) {
return pipeline;
}
return nullptr;

104
src/video_core/renderer_vulkan/vk_rasterizer.cpp
View file

@ -70,7 +70,7 @@ VkViewport GetViewportState(const Device& device, const Maxwell& regs, size_t in
const float width = conv(src.scale_x * 2.0f);
float y = conv(src.translate_y - src.scale_y);
float height = conv(src.scale_y * 2.0f);
bool y_negate = regs.screen_y_control.y_negate;
bool y_negate = regs.window_origin.mode != Maxwell::WindowOrigin::Mode::UpperLeft;
if (!device.IsNvViewportSwizzleSupported()) {
y_negate = y_negate != (src.swizzle.y == Maxwell::ViewportSwizzle::NegativeY);
@ -130,11 +130,11 @@ VkRect2D GetScissorState(const Maxwell& regs, size_t index, u32 up_scale = 1, u3
DrawParams MakeDrawParams(const Maxwell& regs, u32 num_instances, bool is_instanced,
bool is_indexed) {
DrawParams params{
.base_instance = regs.vb_base_instance,
.base_instance = regs.global_base_instance_index,
.num_instances = is_instanced ? num_instances : 1,
.base_vertex = is_indexed ? regs.vb_element_base : regs.vertex_buffer.first,
.num_vertices = is_indexed ? regs.index_array.count : regs.vertex_buffer.count,
.first_index = is_indexed ? regs.index_array.first : 0,
.base_vertex = is_indexed ? regs.global_base_vertex_index : regs.vertex_buffer.first,
.num_vertices = is_indexed ? regs.index_buffer.count : regs.vertex_buffer.count,
.first_index = is_indexed ? regs.index_buffer.first : 0,
.is_indexed = is_indexed,
};
if (regs.draw.topology == Maxwell::PrimitiveTopology::Quads) {
@ -225,10 +225,10 @@ void RasterizerVulkan::Clear() {
query_cache.UpdateCounters();
auto& regs = maxwell3d->regs;
const bool use_color = regs.clear_buffers.R || regs.clear_buffers.G || regs.clear_buffers.B ||
regs.clear_buffers.A;
const bool use_depth = regs.clear_buffers.Z;
const bool use_stencil = regs.clear_buffers.S;
const bool use_color = regs.clear_surface.R || regs.clear_surface.G || regs.clear_surface.B ||
regs.clear_surface.A;
const bool use_depth = regs.clear_surface.Z;
const bool use_stencil = regs.clear_surface.S;
if (!use_color && !use_depth && !use_stencil) {
return;
}
@ -254,9 +254,9 @@ void RasterizerVulkan::Clear() {
default_scissor.extent.height = std::numeric_limits<s32>::max();
VkClearRect clear_rect{
.rect = regs.clear_flags.scissor ? GetScissorState(regs, 0, up_scale, down_shift)
: default_scissor,
.baseArrayLayer = regs.clear_buffers.layer,
.rect = regs.clear_control.use_scissor ? GetScissorState(regs, 0, up_scale, down_shift)
: default_scissor,
.baseArrayLayer = regs.clear_surface.layer,
.layerCount = 1,
};
if (clear_rect.rect.extent.width == 0 || clear_rect.rect.extent.height == 0) {
@ -267,7 +267,7 @@ void RasterizerVulkan::Clear() {
.height = std::min(clear_rect.rect.extent.height, render_area.height),
};
const u32 color_attachment = regs.clear_buffers.RT;
const u32 color_attachment = regs.clear_surface.RT;
if (use_color && framebuffer->HasAspectColorBit(color_attachment)) {
VkClearValue clear_value;
bool is_integer = false;
@ -289,7 +289,8 @@ void RasterizerVulkan::Clear() {
break;
}
if (!is_integer) {
std::memcpy(clear_value.color.float32, regs.clear_color, sizeof(regs.clear_color));
std::memcpy(clear_value.color.float32, regs.clear_color.data(),
regs.clear_color.size() * sizeof(f32));
} else if (!is_signed) {
for (size_t i = 0; i < 4; i++) {
clear_value.color.uint32[i] = static_cast<u32>(
@ -648,23 +649,23 @@ void RasterizerVulkan::UpdateDynamicStates() {
void RasterizerVulkan::BeginTransformFeedback() {
const auto& regs = maxwell3d->regs;
if (regs.tfb_enabled == 0) {
if (regs.transform_feedback_enabled == 0) {
return;
}
if (!device.IsExtTransformFeedbackSupported()) {
LOG_ERROR(Render_Vulkan, "Transform feedbacks used but not supported");
return;
}
UNIMPLEMENTED_IF(regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationControl) ||
regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationEval) ||
regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::Geometry));
UNIMPLEMENTED_IF(regs.IsShaderConfigEnabled(Maxwell::ShaderType::TessellationInit) ||
regs.IsShaderConfigEnabled(Maxwell::ShaderType::Tessellation) ||
regs.IsShaderConfigEnabled(Maxwell::ShaderType::Geometry));
scheduler.Record(
[](vk::CommandBuffer cmdbuf) { cmdbuf.BeginTransformFeedbackEXT(0, 0, nullptr, nullptr); });
}
void RasterizerVulkan::EndTransformFeedback() {
const auto& regs = maxwell3d->regs;
if (regs.tfb_enabled == 0) {
if (regs.transform_feedback_enabled == 0) {
return;
}
if (!device.IsExtTransformFeedbackSupported()) {
@ -728,11 +729,11 @@ void RasterizerVulkan::UpdateDepthBias(Tegra::Engines::Maxwell3D::Regs& regs) {
if (!state_tracker.TouchDepthBias()) {
return;
}
float units = regs.polygon_offset_units / 2.0f;
const bool is_d24 = regs.zeta.format == Tegra::DepthFormat::S8_UINT_Z24_UNORM ||
regs.zeta.format == Tegra::DepthFormat::D24X8_UNORM ||
regs.zeta.format == Tegra::DepthFormat::D24S8_UNORM ||
regs.zeta.format == Tegra::DepthFormat::D24C8_UNORM;
float units = regs.depth_bias / 2.0f;
const bool is_d24 = regs.zeta.format == Tegra::DepthFormat::Z24_UNORM_S8_UINT ||
regs.zeta.format == Tegra::DepthFormat::X8Z24_UNORM ||
regs.zeta.format == Tegra::DepthFormat::S8Z24_UNORM ||
regs.zeta.format == Tegra::DepthFormat::V8Z24_UNORM;
if (is_d24 && !device.SupportsD24DepthBuffer()) {
// the base formulas can be obtained from here:
// https://docs.microsoft.com/en-us/windows/win32/direct3d11/d3d10-graphics-programming-guide-output-merger-stage-depth-bias
@ -740,8 +741,8 @@ void RasterizerVulkan::UpdateDepthBias(Tegra::Engines::Maxwell3D::Regs& regs) {
static_cast<double>(1ULL << (32 - 24)) / (static_cast<double>(0x1.ep+127));
units = static_cast<float>(static_cast<double>(units) * rescale_factor);
}
scheduler.Record([constant = units, clamp = regs.polygon_offset_clamp,
factor = regs.polygon_offset_factor](vk::CommandBuffer cmdbuf) {
scheduler.Record([constant = units, clamp = regs.depth_bias_clamp,
factor = regs.slope_scale_depth_bias](vk::CommandBuffer cmdbuf) {
cmdbuf.SetDepthBias(constant, clamp, factor);
});
}
@ -771,10 +772,11 @@ void RasterizerVulkan::UpdateStencilFaces(Tegra::Engines::Maxwell3D::Regs& regs)
if (regs.stencil_two_side_enable) {
// Separate values per face
scheduler.Record(
[front_ref = regs.stencil_front_func_ref, front_write_mask = regs.stencil_front_mask,
front_test_mask = regs.stencil_front_func_mask, back_ref = regs.stencil_back_func_ref,
back_write_mask = regs.stencil_back_mask,
back_test_mask = regs.stencil_back_func_mask](vk::CommandBuffer cmdbuf) {
[front_ref = regs.stencil_front_func.ref,
front_write_mask = regs.stencil_front_func.mask,
front_test_mask = regs.stencil_front_func.func_mask,
back_ref = regs.stencil_back_func.ref, back_write_mask = regs.stencil_back_func.mask,
back_test_mask = regs.stencil_back_func.func_mask](vk::CommandBuffer cmdbuf) {
// Front face
cmdbuf.SetStencilReference(VK_STENCIL_FACE_FRONT_BIT, front_ref);
cmdbuf.SetStencilWriteMask(VK_STENCIL_FACE_FRONT_BIT, front_write_mask);
@ -787,8 +789,9 @@ void RasterizerVulkan::UpdateStencilFaces(Tegra::Engines::Maxwell3D::Regs& regs)
});
} else {
// Front face defines both faces
scheduler.Record([ref = regs.stencil_front_func_ref, write_mask = regs.stencil_front_mask,
test_mask = regs.stencil_front_func_mask](vk::CommandBuffer cmdbuf) {
scheduler.Record([ref = regs.stencil_front_func.ref,
write_mask = regs.stencil_front_func.mask,
test_mask = regs.stencil_front_func.func_mask](vk::CommandBuffer cmdbuf) {
cmdbuf.SetStencilReference(VK_STENCIL_FACE_FRONT_AND_BACK, ref);
cmdbuf.SetStencilWriteMask(VK_STENCIL_FACE_FRONT_AND_BACK, write_mask);
cmdbuf.SetStencilCompareMask(VK_STENCIL_FACE_FRONT_AND_BACK, test_mask);
@ -800,7 +803,8 @@ void RasterizerVulkan::UpdateLineWidth(Tegra::Engines::Maxwell3D::Regs& regs) {
if (!state_tracker.TouchLineWidth()) {
return;
}
const float width = regs.line_smooth_enable ? regs.line_width_smooth : regs.line_width_aliased;
const float width =
regs.line_anti_alias_enable ? regs.line_width_smooth : regs.line_width_aliased;
scheduler.Record([width](vk::CommandBuffer cmdbuf) { cmdbuf.SetLineWidth(width); });
}
@ -808,10 +812,10 @@ void RasterizerVulkan::UpdateCullMode(Tegra::Engines::Maxwell3D::Regs& regs) {
if (!state_tracker.TouchCullMode()) {
return;
}
scheduler.Record(
[enabled = regs.cull_test_enabled, cull_face = regs.cull_face](vk::CommandBuffer cmdbuf) {
cmdbuf.SetCullModeEXT(enabled ? MaxwellToVK::CullFace(cull_face) : VK_CULL_MODE_NONE);
});
scheduler.Record([enabled = regs.gl_cull_test_enabled,
cull_face = regs.gl_cull_face](vk::CommandBuffer cmdbuf) {
cmdbuf.SetCullModeEXT(enabled ? MaxwellToVK::CullFace(cull_face) : VK_CULL_MODE_NONE);
});
}
void RasterizerVulkan::UpdateDepthBoundsTestEnable(Tegra::Engines::Maxwell3D::Regs& regs) {
@ -860,8 +864,8 @@ void RasterizerVulkan::UpdateFrontFace(Tegra::Engines::Maxwell3D::Regs& regs) {
return;
}
VkFrontFace front_face = MaxwellToVK::FrontFace(regs.front_face);
if (regs.screen_y_control.triangle_rast_flip != 0) {
VkFrontFace front_face = MaxwellToVK::FrontFace(regs.gl_front_face);
if (regs.window_origin.flip_y != 0) {
front_face = front_face == VK_FRONT_FACE_CLOCKWISE ? VK_FRONT_FACE_COUNTER_CLOCKWISE
: VK_FRONT_FACE_CLOCKWISE;
}
@ -873,16 +877,16 @@ void RasterizerVulkan::UpdateStencilOp(Tegra::Engines::Maxwell3D::Regs& regs) {
if (!state_tracker.TouchStencilOp()) {
return;
}
const Maxwell::StencilOp fail = regs.stencil_front_op_fail;
const Maxwell::StencilOp zfail = regs.stencil_front_op_zfail;
const Maxwell::StencilOp zpass = regs.stencil_front_op_zpass;
const Maxwell::ComparisonOp compare = regs.stencil_front_func_func;
const Maxwell::StencilOp::Op fail = regs.stencil_front_op.fail;
const Maxwell::StencilOp::Op zfail = regs.stencil_front_op.zfail;
const Maxwell::StencilOp::Op zpass = regs.stencil_front_op.zpass;
const Maxwell::ComparisonOp compare = regs.stencil_front_op.func;
if (regs.stencil_two_side_enable) {
// Separate stencil op per face
const Maxwell::StencilOp back_fail = regs.stencil_back_op_fail;
const Maxwell::StencilOp back_zfail = regs.stencil_back_op_zfail;
const Maxwell::StencilOp back_zpass = regs.stencil_back_op_zpass;
const Maxwell::ComparisonOp back_compare = regs.stencil_back_func_func;
const Maxwell::StencilOp::Op back_fail = regs.stencil_back_op.fail;
const Maxwell::StencilOp::Op back_zfail = regs.stencil_back_op.zfail;
const Maxwell::StencilOp::Op back_zpass = regs.stencil_back_op.zpass;
const Maxwell::ComparisonOp back_compare = regs.stencil_back_op.func;
scheduler.Record([fail, zfail, zpass, compare, back_fail, back_zfail, back_zpass,
back_compare](vk::CommandBuffer cmdbuf) {
cmdbuf.SetStencilOpEXT(VK_STENCIL_FACE_FRONT_BIT, MaxwellToVK::StencilOp(fail),
@ -954,15 +958,15 @@ void RasterizerVulkan::UpdateVertexInput(Tegra::Engines::Maxwell3D::Regs& regs)
dirty[Dirty::VertexBinding0 + index] = false;
const u32 binding{static_cast<u32>(index)};
const auto& input_binding{regs.vertex_array[binding]};
const bool is_instanced{regs.instanced_arrays.IsInstancingEnabled(binding)};
const auto& input_binding{regs.vertex_streams[binding]};
const bool is_instanced{regs.vertex_stream_instances.IsInstancingEnabled(binding)};
bindings.push_back({
.sType = VK_STRUCTURE_TYPE_VERTEX_INPUT_BINDING_DESCRIPTION_2_EXT,
.pNext = nullptr,
.binding = binding,
.stride = input_binding.stride,
.inputRate = is_instanced ? VK_VERTEX_INPUT_RATE_INSTANCE : VK_VERTEX_INPUT_RATE_VERTEX,
.divisor = is_instanced ? input_binding.divisor : 1,
.divisor = is_instanced ? input_binding.frequency : 1,
});
}
scheduler.Record([bindings, attributes](vk::CommandBuffer cmdbuf) {

58
src/video_core/renderer_vulkan/vk_state_tracker.cpp
View file

@ -51,8 +51,8 @@ Flags MakeInvalidationFlags() {
void SetupDirtyViewports(Tables& tables) {
FillBlock(tables[0], OFF(viewport_transform), NUM(viewport_transform), Viewports);
FillBlock(tables[0], OFF(viewports), NUM(viewports), Viewports);
tables[0][OFF(viewport_transform_enabled)] = Viewports;
tables[1][OFF(screen_y_control)] = Viewports;
tables[0][OFF(viewport_scale_offset_enbled)] = Viewports;
tables[1][OFF(window_origin)] = Viewports;
}
void SetupDirtyScissors(Tables& tables) {
@ -61,9 +61,9 @@ void SetupDirtyScissors(Tables& tables) {
void SetupDirtyDepthBias(Tables& tables) {
auto& table = tables[0];
table[OFF(polygon_offset_units)] = DepthBias;
table[OFF(polygon_offset_clamp)] = DepthBias;
table[OFF(polygon_offset_factor)] = DepthBias;
table[OFF(depth_bias)] = DepthBias;
table[OFF(depth_bias_clamp)] = DepthBias;
table[OFF(slope_scale_depth_bias)] = DepthBias;
}
void SetupDirtyBlendConstants(Tables& tables) {
@ -77,12 +77,12 @@ void SetupDirtyDepthBounds(Tables& tables) {
void SetupDirtyStencilProperties(Tables& tables) {
auto& table = tables[0];
table[OFF(stencil_two_side_enable)] = StencilProperties;
table[OFF(stencil_front_func_ref)] = StencilProperties;
table[OFF(stencil_front_mask)] = StencilProperties;
table[OFF(stencil_front_func_mask)] = StencilProperties;
table[OFF(stencil_back_func_ref)] = StencilProperties;
table[OFF(stencil_back_mask)] = StencilProperties;
table[OFF(stencil_back_func_mask)] = StencilProperties;
table[OFF(stencil_front_func.ref)] = StencilProperties;
table[OFF(stencil_front_func.mask)] = StencilProperties;
table[OFF(stencil_front_func.func_mask)] = StencilProperties;
table[OFF(stencil_back_func.ref)] = StencilProperties;
table[OFF(stencil_back_func.mask)] = StencilProperties;
table[OFF(stencil_back_func.func_mask)] = StencilProperties;
}
void SetupDirtyLineWidth(Tables& tables) {
@ -92,8 +92,8 @@ void SetupDirtyLineWidth(Tables& tables) {
void SetupDirtyCullMode(Tables& tables) {
auto& table = tables[0];
table[OFF(cull_face)] = CullMode;
table[OFF(cull_test_enabled)] = CullMode;
table[OFF(gl_cull_face)] = CullMode;
table[OFF(gl_cull_test_enabled)] = CullMode;
}
void SetupDirtyDepthBoundsEnable(Tables& tables) {
@ -114,20 +114,20 @@ void SetupDirtyDepthCompareOp(Tables& tables) {
void SetupDirtyFrontFace(Tables& tables) {
auto& table = tables[0];
table[OFF(front_face)] = FrontFace;
table[OFF(screen_y_control)] = FrontFace;
table[OFF(gl_front_face)] = FrontFace;
table[OFF(window_origin)] = FrontFace;
}
void SetupDirtyStencilOp(Tables& tables) {
auto& table = tables[0];
table[OFF(stencil_front_op_fail)] = StencilOp;
table[OFF(stencil_front_op_zfail)] = StencilOp;
table[OFF(stencil_front_op_zpass)] = StencilOp;
table[OFF(stencil_front_func_func)] = StencilOp;
table[OFF(stencil_back_op_fail)] = StencilOp;
table[OFF(stencil_back_op_zfail)] = StencilOp;
table[OFF(stencil_back_op_zpass)] = StencilOp;
table[OFF(stencil_back_func_func)] = StencilOp;
table[OFF(stencil_front_op.fail)] = StencilOp;
table[OFF(stencil_front_op.zfail)] = StencilOp;
table[OFF(stencil_front_op.zpass)] = StencilOp;
table[OFF(stencil_front_op.func)] = StencilOp;
table[OFF(stencil_back_op.fail)] = StencilOp;
table[OFF(stencil_back_op.zfail)] = StencilOp;
table[OFF(stencil_back_op.zpass)] = StencilOp;
table[OFF(stencil_back_op.func)] = StencilOp;
// Table 0 is used by StencilProperties
tables[1][OFF(stencil_two_side_enable)] = StencilOp;
@ -139,10 +139,10 @@ void SetupDirtyStencilTestEnable(Tables& tables) {
void SetupDirtyBlending(Tables& tables) {
tables[0][OFF(color_mask_common)] = Blending;
tables[0][OFF(independent_blend_enable)] = Blending;
tables[0][OFF(blend_per_target_enabled)] = Blending;
FillBlock(tables[0], OFF(color_mask), NUM(color_mask), Blending);
FillBlock(tables[0], OFF(blend), NUM(blend), Blending);
FillBlock(tables[0], OFF(independent_blend), NUM(independent_blend), Blending);
FillBlock(tables[0], OFF(blend_per_target), NUM(blend_per_target), Blending);
}
void SetupDirtyViewportSwizzles(Tables& tables) {
@ -166,10 +166,10 @@ void SetupDirtyVertexBindings(Tables& tables) {
static constexpr size_t divisor_offset = 3;
for (size_t i = 0; i < Regs::NumVertexArrays; ++i) {
const u8 flag = static_cast<u8>(VertexBinding0 + i);
tables[0][OFF(instanced_arrays) + i] = VertexInput;
tables[1][OFF(instanced_arrays) + i] = flag;
tables[0][OFF(vertex_array) + i * NUM(vertex_array[0]) + divisor_offset] = VertexInput;
tables[1][OFF(vertex_array) + i * NUM(vertex_array[0]) + divisor_offset] = flag;
tables[0][OFF(vertex_stream_instances) + i] = VertexInput;
tables[1][OFF(vertex_stream_instances) + i] = flag;
tables[0][OFF(vertex_streams) + i * NUM(vertex_streams[0]) + divisor_offset] = VertexInput;
tables[1][OFF(vertex_streams) + i * NUM(vertex_streams[0]) + divisor_offset] = flag;
}
}
} // Anonymous namespace

12
src/video_core/shader_cache.cpp
View file

@ -43,14 +43,14 @@ bool ShaderCache::RefreshStages(std::array<u64, 6>& unique_hashes) {
}
dirty[VideoCommon::Dirty::Shaders] = false;
const GPUVAddr base_addr{maxwell3d->regs.code_address.CodeAddress()};
const GPUVAddr base_addr{maxwell3d->regs.program_region.Address()};
for (size_t index = 0; index < Tegra::Engines::Maxwell3D::Regs::MaxShaderProgram; ++index) {
if (!maxwell3d->regs.IsShaderConfigEnabled(index)) {
unique_hashes[index] = 0;
continue;
}
const auto& shader_config{maxwell3d->regs.shader_config[index]};
const auto program{static_cast<Tegra::Engines::Maxwell3D::Regs::ShaderProgram>(index)};
const auto& shader_config{maxwell3d->regs.pipelines[index]};
const auto program{static_cast<Tegra::Engines::Maxwell3D::Regs::ShaderType>(index)};
const GPUVAddr shader_addr{base_addr + shader_config.offset};
const std::optional<VAddr> cpu_shader_addr{gpu_memory->GpuToCpuAddress(shader_addr)};
if (!cpu_shader_addr) {
@ -90,14 +90,14 @@ const ShaderInfo* ShaderCache::ComputeShader() {
void ShaderCache::GetGraphicsEnvironments(GraphicsEnvironments& result,
const std::array<u64, NUM_PROGRAMS>& unique_hashes) {
size_t env_index{};
const GPUVAddr base_addr{maxwell3d->regs.code_address.CodeAddress()};
const GPUVAddr base_addr{maxwell3d->regs.program_region.Address()};
for (size_t index = 0; index < NUM_PROGRAMS; ++index) {
if (unique_hashes[index] == 0) {
continue;
}
const auto program{static_cast<Tegra::Engines::Maxwell3D::Regs::ShaderProgram>(index)};
const auto program{static_cast<Tegra::Engines::Maxwell3D::Regs::ShaderType>(index)};
auto& env{result.envs[index]};
const u32 start_address{maxwell3d->regs.shader_config[index].offset};
const u32 start_address{maxwell3d->regs.pipelines[index].offset};
env = GraphicsEnvironment{*maxwell3d, *gpu_memory, program, base_addr, start_address};
env.SetCachedSize(shader_infos[index]->size_bytes);
result.env_ptrs[env_index++] = &env;

23
src/video_core/shader_environment.cpp
View file

@ -250,34 +250,34 @@ Shader::TextureType GenericEnvironment::ReadTextureTypeImpl(GPUVAddr tic_addr, u
GraphicsEnvironment::GraphicsEnvironment(Tegra::Engines::Maxwell3D& maxwell3d_,
Tegra::MemoryManager& gpu_memory_,
Maxwell::ShaderProgram program, GPUVAddr program_base_,
Maxwell::ShaderType program, GPUVAddr program_base_,
u32 start_address_)
: GenericEnvironment{gpu_memory_, program_base_, start_address_}, maxwell3d{&maxwell3d_} {
gpu_memory->ReadBlock(program_base + start_address, &sph, sizeof(sph));
initial_offset = sizeof(sph);
gp_passthrough_mask = maxwell3d->regs.gp_passthrough_mask;
gp_passthrough_mask = maxwell3d->regs.post_vtg_shader_attrib_skip_mask;
switch (program) {
case Maxwell::ShaderProgram::VertexA:
case Maxwell::ShaderType::VertexA:
stage = Shader::Stage::VertexA;
stage_index = 0;
break;
case Maxwell::ShaderProgram::VertexB:
case Maxwell::ShaderType::VertexB:
stage = Shader::Stage::VertexB;
stage_index = 0;
break;
case Maxwell::ShaderProgram::TesselationControl:
case Maxwell::ShaderType::TessellationInit:
stage = Shader::Stage::TessellationControl;
stage_index = 1;
break;
case Maxwell::ShaderProgram::TesselationEval:
case Maxwell::ShaderType::Tessellation:
stage = Shader::Stage::TessellationEval;
stage_index = 2;
break;
case Maxwell::ShaderProgram::Geometry:
case Maxwell::ShaderType::Geometry:
stage = Shader::Stage::Geometry;
stage_index = 3;
break;
case Maxwell::ShaderProgram::Fragment:
case Maxwell::ShaderType::Pixel:
stage = Shader::Stage::Fragment;
stage_index = 4;
break;
@ -288,7 +288,7 @@ GraphicsEnvironment::GraphicsEnvironment(Tegra::Engines::Maxwell3D& maxwell3d_,
const u64 local_size{sph.LocalMemorySize()};
ASSERT(local_size <= std::numeric_limits<u32>::max());
local_memory_size = static_cast<u32>(local_size) + sph.common3.shader_local_memory_crs_size;
texture_bound = maxwell3d->regs.tex_cb_index;
texture_bound = maxwell3d->regs.bindless_texture_const_buffer_slot;
}
u32 GraphicsEnvironment::ReadCbufValue(u32 cbuf_index, u32 cbuf_offset) {
@ -304,8 +304,9 @@ u32 GraphicsEnvironment::ReadCbufValue(u32 cbuf_index, u32 cbuf_offset) {
Shader::TextureType GraphicsEnvironment::ReadTextureType(u32 handle) {
const auto& regs{maxwell3d->regs};
const bool via_header_index{regs.sampler_index == Maxwell::SamplerIndex::ViaHeaderIndex};
return ReadTextureTypeImpl(regs.tic.Address(), regs.tic.limit, via_header_index, handle);
const bool via_header_index{regs.sampler_binding == Maxwell::SamplerBinding::ViaHeaderBinding};
return ReadTextureTypeImpl(regs.tex_header.Address(), regs.tex_header.limit, via_header_index,
handle);
}
ComputeEnvironment::ComputeEnvironment(Tegra::Engines::KeplerCompute& kepler_compute_,

2
src/video_core/shader_environment.h
View file

@ -93,7 +93,7 @@ public:
explicit GraphicsEnvironment() = default;
explicit GraphicsEnvironment(Tegra::Engines::Maxwell3D& maxwell3d_,
Tegra::MemoryManager& gpu_memory_,
Tegra::Engines::Maxwell3D::Regs::ShaderProgram program,
Tegra::Engines::Maxwell3D::Regs::ShaderType program,
GPUVAddr program_base_, u32 start_address_);
~GraphicsEnvironment() override = default;

14
src/video_core/surface.cpp
View file

@ -73,17 +73,17 @@ bool SurfaceTargetIsArray(SurfaceTarget target) {
PixelFormat PixelFormatFromDepthFormat(Tegra::DepthFormat format) {
switch (format) {
case Tegra::DepthFormat::S8_UINT_Z24_UNORM:
case Tegra::DepthFormat::Z24_UNORM_S8_UINT:
return PixelFormat::S8_UINT_D24_UNORM;
case Tegra::DepthFormat::D24S8_UNORM:
case Tegra::DepthFormat::S8Z24_UNORM:
return PixelFormat::D24_UNORM_S8_UINT;
case Tegra::DepthFormat::D32_FLOAT:
case Tegra::DepthFormat::Z32_FLOAT:
return PixelFormat::D32_FLOAT;
case Tegra::DepthFormat::D16_UNORM:
case Tegra::DepthFormat::Z16_UNORM:
return PixelFormat::D16_UNORM;
case Tegra::DepthFormat::S8_UINT:
return PixelFormat::S8_UINT;
case Tegra::DepthFormat::D32_FLOAT_S8X24_UINT:
case Tegra::DepthFormat::Z32_FLOAT_X24S8_UINT:
return PixelFormat::D32_FLOAT_S8_UINT;
default:
UNIMPLEMENTED_MSG("Unimplemented format={}", format);
@ -117,9 +117,9 @@ PixelFormat PixelFormatFromRenderTargetFormat(Tegra::RenderTargetFormat format)
return PixelFormat::R32G32_UINT;
case Tegra::RenderTargetFormat::R16G16B16X16_FLOAT:
return PixelFormat::R16G16B16X16_FLOAT;
case Tegra::RenderTargetFormat::B8G8R8A8_UNORM:
case Tegra::RenderTargetFormat::A8R8G8B8_UNORM:
return PixelFormat::B8G8R8A8_UNORM;
case Tegra::RenderTargetFormat::B8G8R8A8_SRGB:
case Tegra::RenderTargetFormat::A8R8G8B8_SRGB:
return PixelFormat::B8G8R8A8_SRGB;
case Tegra::RenderTargetFormat::A2B10G10R10_UNORM:
return PixelFormat::A2B10G10R10_UNORM;

37
src/video_core/texture_cache/image_info.cpp
View file

@ -1,6 +1,8 @@
// SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include <fmt/format.h>
#include "common/assert.h"
#include "video_core/surface.h"
#include "video_core/texture_cache/format_lookup_table.h"
@ -12,6 +14,7 @@
namespace VideoCommon {
using Tegra::Engines::Maxwell3D;
using Tegra::Texture::TextureType;
using Tegra::Texture::TICEntry;
using VideoCore::Surface::PixelFormat;
@ -107,12 +110,13 @@ ImageInfo::ImageInfo(const TICEntry& config) noexcept {
}
}
ImageInfo::ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs, size_t index) noexcept {
ImageInfo::ImageInfo(const Maxwell3D::Regs& regs, size_t index) noexcept {
const auto& rt = regs.rt[index];
format = VideoCore::Surface::PixelFormatFromRenderTargetFormat(rt.format);
rescaleable = false;
if (rt.tile_mode.is_pitch_linear) {
ASSERT(rt.tile_mode.is_3d == 0);
ASSERT(rt.tile_mode.dim_control ==
Maxwell3D::Regs::TileMode::DimensionControl::DepthDefinesArray);
type = ImageType::Linear;
pitch = rt.width;
size = Extent3D{
@ -124,15 +128,16 @@ ImageInfo::ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs, size_t index)
}
size.width = rt.width;
size.height = rt.height;
layer_stride = rt.layer_stride * 4;
layer_stride = rt.array_pitch * 4;
maybe_unaligned_layer_stride = layer_stride;
num_samples = NumSamples(regs.multisample_mode);
num_samples = NumSamples(regs.anti_alias_samples_mode);
block = Extent3D{
.width = rt.tile_mode.block_width,
.height = rt.tile_mode.block_height,
.depth = rt.tile_mode.block_depth,
};
if (rt.tile_mode.is_3d) {
if (rt.tile_mode.dim_control ==
Maxwell3D::Regs::TileMode::DimensionControl::DepthDefinesDepth) {
type = ImageType::e3D;
size.depth = rt.depth;
} else {
@ -146,31 +151,37 @@ ImageInfo::ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs, size_t index)
ImageInfo::ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs) noexcept {
format = VideoCore::Surface::PixelFormatFromDepthFormat(regs.zeta.format);
size.width = regs.zeta_width;
size.height = regs.zeta_height;
size.width = regs.zeta_size.width;
size.height = regs.zeta_size.height;
rescaleable = false;
resources.levels = 1;
layer_stride = regs.zeta.layer_stride * 4;
layer_stride = regs.zeta.array_pitch * 4;
maybe_unaligned_layer_stride = layer_stride;
num_samples = NumSamples(regs.multisample_mode);
num_samples = NumSamples(regs.anti_alias_samples_mode);
block = Extent3D{
.width = regs.zeta.tile_mode.block_width,
.height = regs.zeta.tile_mode.block_height,
.depth = regs.zeta.tile_mode.block_depth,
};
if (regs.zeta.tile_mode.is_pitch_linear) {
ASSERT(regs.zeta.tile_mode.is_3d == 0);
ASSERT(regs.zeta.tile_mode.dim_control ==
Maxwell3D::Regs::TileMode::DimensionControl::DepthDefinesArray);
type = ImageType::Linear;
pitch = size.width * BytesPerBlock(format);
} else if (regs.zeta.tile_mode.is_3d) {
} else if (regs.zeta.tile_mode.dim_control ==
Maxwell3D::Regs::TileMode::DimensionControl::DepthDefinesDepth) {
ASSERT(regs.zeta.tile_mode.is_pitch_linear == 0);
ASSERT(regs.zeta_size.dim_control ==
Maxwell3D::Regs::ZetaSize::DimensionControl::ArraySizeOne);
type = ImageType::e3D;
size.depth = regs.zeta_depth;
size.depth = regs.zeta_size.depth;
} else {
ASSERT(regs.zeta_size.dim_control ==
Maxwell3D::Regs::ZetaSize::DimensionControl::DepthDefinesArray);
rescaleable = block.depth == 0;
downscaleable = size.height > 512;
type = ImageType::e2D;
resources.layers = regs.zeta_depth;
resources.layers = regs.zeta_size.depth;
}
}

19
src/video_core/texture_cache/texture_cache.h
View file

@ -189,15 +189,16 @@ typename P::Sampler* TextureCache<P>::GetComputeSampler(u32 index) {
template <class P>
void TextureCache<P>::SynchronizeGraphicsDescriptors() {
using SamplerIndex = Tegra::Engines::Maxwell3D::Regs::SamplerIndex;
const bool linked_tsc = maxwell3d->regs.sampler_index == SamplerIndex::ViaHeaderIndex;
const u32 tic_limit = maxwell3d->regs.tic.limit;
const u32 tsc_limit = linked_tsc ? tic_limit : maxwell3d->regs.tsc.limit;
if (channel_state->graphics_sampler_table.Synchornize(maxwell3d->regs.tsc.Address(),
using SamplerBinding = Tegra::Engines::Maxwell3D::Regs::SamplerBinding;
const bool linked_tsc = maxwell3d->regs.sampler_binding == SamplerBinding::ViaHeaderBinding;
const u32 tic_limit = maxwell3d->regs.tex_header.limit;
const u32 tsc_limit = linked_tsc ? tic_limit : maxwell3d->regs.tex_sampler.limit;
if (channel_state->graphics_sampler_table.Synchornize(maxwell3d->regs.tex_sampler.Address(),
tsc_limit)) {
channel_state->graphics_sampler_ids.resize(tsc_limit + 1, CORRUPT_ID);
}
if (channel_state->graphics_image_table.Synchornize(maxwell3d->regs.tic.Address(), tic_limit)) {
if (channel_state->graphics_image_table.Synchornize(maxwell3d->regs.tex_header.Address(),
tic_limit)) {
channel_state->graphics_image_view_ids.resize(tic_limit + 1, CORRUPT_ID);
}
}
@ -352,8 +353,8 @@ void TextureCache<P>::UpdateRenderTargets(bool is_clear) {
down_shift = Settings::values.resolution_info.down_shift;
}
render_targets.size = Extent2D{
(maxwell3d->regs.render_area.width * up_scale) >> down_shift,
(maxwell3d->regs.render_area.height * up_scale) >> down_shift,
(maxwell3d->regs.surface_clip.width * up_scale) >> down_shift,
(maxwell3d->regs.surface_clip.height * up_scale) >> down_shift,
};
render_targets.is_rescaled = is_rescaling;
@ -1980,7 +1981,7 @@ bool TextureCache<P>::IsFullClear(ImageViewId id) {
// Images with multiple resources can't be cleared in a single call
return false;
}
if (regs.clear_flags.scissor == 0) {
if (regs.clear_control.use_scissor == 0) {
// If scissor testing is disabled, the clear is always full
return true;
}

118
src/video_core/transform_feedback.cpp
View file

@ -15,51 +15,51 @@ namespace VideoCommon {
std::vector<Shader::TransformFeedbackVarying> MakeTransformFeedbackVaryings(
const TransformFeedbackState& state) {
static constexpr std::array VECTORS{
28, // gl_Position
32, // Generic 0
36, // Generic 1
40, // Generic 2
44, // Generic 3
48, // Generic 4
52, // Generic 5
56, // Generic 6
60, // Generic 7
64, // Generic 8
68, // Generic 9
72, // Generic 10
76, // Generic 11
80, // Generic 12
84, // Generic 13
88, // Generic 14
92, // Generic 15
96, // Generic 16
100, // Generic 17
104, // Generic 18
108, // Generic 19
112, // Generic 20
116, // Generic 21
120, // Generic 22
124, // Generic 23
128, // Generic 24
132, // Generic 25
136, // Generic 26
140, // Generic 27
144, // Generic 28
148, // Generic 29
152, // Generic 30
156, // Generic 31
160, // gl_FrontColor
164, // gl_FrontSecondaryColor
160, // gl_BackColor
164, // gl_BackSecondaryColor
192, // gl_TexCoord[0]
196, // gl_TexCoord[1]
200, // gl_TexCoord[2]
204, // gl_TexCoord[3]
208, // gl_TexCoord[4]
212, // gl_TexCoord[5]
216, // gl_TexCoord[6]
220, // gl_TexCoord[7]
28U, // gl_Position
32U, // Generic 0
36U, // Generic 1
40U, // Generic 2
44U, // Generic 3
48U, // Generic 4
52U, // Generic 5
56U, // Generic 6
60U, // Generic 7
64U, // Generic 8
68U, // Generic 9
72U, // Generic 10
76U, // Generic 11
80U, // Generic 12
84U, // Generic 13
88U, // Generic 14
92U, // Generic 15
96U, // Generic 16
100U, // Generic 17
104U, // Generic 18
108U, // Generic 19
112U, // Generic 20
116U, // Generic 21
120U, // Generic 22
124U, // Generic 23
128U, // Generic 24
132U, // Generic 25
136U, // Generic 26
140U, // Generic 27
144U, // Generic 28
148U, // Generic 29
152U, // Generic 30
156U, // Generic 31
160U, // gl_FrontColor
164U, // gl_FrontSecondaryColor
160U, // gl_BackColor
164U, // gl_BackSecondaryColor
192U, // gl_TexCoord[0]
196U, // gl_TexCoord[1]
200U, // gl_TexCoord[2]
204U, // gl_TexCoord[3]
208U, // gl_TexCoord[4]
212U, // gl_TexCoord[5]
216U, // gl_TexCoord[6]
220U, // gl_TexCoord[7]
};
std::vector<Shader::TransformFeedbackVarying> xfb(256);
for (size_t buffer = 0; buffer < state.layouts.size(); ++buffer) {
@ -68,8 +68,20 @@ std::vector<Shader::TransformFeedbackVarying> MakeTransformFeedbackVaryings(
const u32 varying_count = layout.varying_count;
u32 highest = 0;
for (u32 offset = 0; offset < varying_count; ++offset) {
const u32 base_offset = offset;
const u8 location = locations[offset];
const auto get_attribute = [&locations](u32 index) -> u32 {
switch (index % 4) {
case 0:
return locations[index / 4].attribute0.Value();
case 1:
return locations[index / 4].attribute1.Value();
case 2:
return locations[index / 4].attribute2.Value();
case 3:
return locations[index / 4].attribute3.Value();
}
UNREACHABLE();
return 0;
};
UNIMPLEMENTED_IF_MSG(layout.stream != 0, "Stream is not zero: {}", layout.stream);
Shader::TransformFeedbackVarying varying{
@ -78,16 +90,18 @@ std::vector<Shader::TransformFeedbackVarying> MakeTransformFeedbackVaryings(
.offset = offset * 4,
.components = 1,
};
if (std::ranges::find(VECTORS, Common::AlignDown(location, 4)) != VECTORS.end()) {
UNIMPLEMENTED_IF_MSG(location % 4 != 0, "Unaligned TFB");
const u32 base_offset = offset;
const auto attribute{get_attribute(offset)};
if (std::ranges::find(VECTORS, Common::AlignDown(attribute, 4)) != VECTORS.end()) {
UNIMPLEMENTED_IF_MSG(attribute % 4 != 0, "Unaligned TFB {}", attribute);
const u8 base_index = location / 4;
while (offset + 1 < varying_count && base_index == locations[offset + 1] / 4) {
const auto base_index = attribute / 4;
while (offset + 1 < varying_count && base_index == get_attribute(offset + 1) / 4) {
++offset;
++varying.components;
}
}
xfb[location] = varying;
xfb[attribute] = varying;
highest = std::max(highest, (base_offset + varying.components) * 4);
}
UNIMPLEMENTED_IF(highest != layout.stride);

3
src/video_core/transform_feedback.h
View file

@ -19,7 +19,8 @@ struct TransformFeedbackState {
u32 stride;
};
std::array<Layout, Tegra::Engines::Maxwell3D::Regs::NumTransformFeedbackBuffers> layouts;
std::array<std::array<u8, 128>, Tegra::Engines::Maxwell3D::Regs::NumTransformFeedbackBuffers>
std::array<std::array<Tegra::Engines::Maxwell3D::Regs::StreamOutLayout, 32>,
Tegra::Engines::Maxwell3D::Regs::NumTransformFeedbackBuffers>
varyings;
};