1
0
mirror of https://github.com/RPCS3/rpcs3.git synced 2024-11-25 20:22:30 +01:00

rsx: Fix graphics state foot-gun

This commit is contained in:
kd-11 2023-01-25 15:03:52 +03:00 committed by kd-11
parent 52b67b46fc
commit 6adcabda29
11 changed files with 150 additions and 67 deletions

View File

@ -35,6 +35,7 @@ namespace rsx
}
template <typename T, typename bitmask_type = u32>
requires std::is_integral_v<bitmask_type>&& std::is_enum_v<T>
class atomic_bitmask_t
{
private:
@ -95,4 +96,70 @@ namespace rsx
return m_data.observe() != 0;
}
};
template <typename T, typename bitmask_type = u32>
requires std::is_integral_v<bitmask_type> && std::is_enum_v<T>
class bitmask_t
{
private:
bitmask_type m_data = 0;
public:
bitmask_t() = default;
bool operator & (bitmask_type mask) const
{
return !!(m_data & mask);
}
bitmask_type operator | (bitmask_type mask) const
{
return m_data | mask;
}
void operator &= (bitmask_type mask)
{
m_data &= mask;
}
void operator |= (bitmask_type mask)
{
m_data |= mask;
}
bool test(T mask)
{
return !!(m_data & static_cast<bitmask_type>(mask));
}
void set(T mask)
{
m_data |= static_cast<bitmask_type>(mask);
}
bool test_and_set(T mask)
{
const auto old = m_data;
m_data |= static_cast<bitmask_type>(mask);
return !!(old & mask);
}
template <typename U>
requires std::is_same_v<T, U> || std::is_same_v<bitmask_type, U>
void clear(U mask)
{
const bitmask_type clear_mask = ~(static_cast<bitmask_type>(mask));
m_data &= clear_mask;
}
void clear()
{
m_data = 0;
}
operator bool() const
{
return !!m_data;
}
};
}

View File

@ -320,7 +320,7 @@ void GLGSRender::update_draw_state()
//NV4097_SET_CLIP_ID_TEST_ENABLE
// For OGL Z range is updated every draw as it is separate from viewport config
m_graphics_state &= ~(rsx::pipeline_state::zclip_config_state_dirty);
m_graphics_state.clear(rsx::pipeline_state::zclip_config_state_dirty);
m_frame_stats.setup_time += m_profiler.duration();
}

View File

@ -799,13 +799,13 @@ void GLGSRender::load_program_env()
const u32 fragment_constants_size = current_fp_metadata.program_constants_buffer_length;
const bool update_transform_constants = !!(m_graphics_state & rsx::pipeline_state::transform_constants_dirty);
const bool update_fragment_constants = !!(m_graphics_state & rsx::pipeline_state::fragment_constants_dirty) && fragment_constants_size;
const bool update_vertex_env = !!(m_graphics_state & rsx::pipeline_state::vertex_state_dirty);
const bool update_fragment_env = !!(m_graphics_state & rsx::pipeline_state::fragment_state_dirty);
const bool update_fragment_texture_env = !!(m_graphics_state & rsx::pipeline_state::fragment_texture_state_dirty);
const bool update_instruction_buffers = (!!m_interpreter_state && m_shader_interpreter.is_interpreter(m_program));
const bool update_raster_env = (rsx::method_registers.polygon_stipple_enabled() && !!(m_graphics_state & rsx::pipeline_state::polygon_stipple_pattern_dirty));
const bool update_transform_constants = m_graphics_state & rsx::pipeline_state::transform_constants_dirty;
const bool update_fragment_constants = (m_graphics_state & rsx::pipeline_state::fragment_constants_dirty) && fragment_constants_size;
const bool update_vertex_env = m_graphics_state & rsx::pipeline_state::vertex_state_dirty;
const bool update_fragment_env = m_graphics_state & rsx::pipeline_state::fragment_state_dirty;
const bool update_fragment_texture_env = m_graphics_state & rsx::pipeline_state::fragment_texture_state_dirty;
const bool update_instruction_buffers = !!m_interpreter_state && m_shader_interpreter.is_interpreter(m_program);
const bool update_raster_env = rsx::method_registers.polygon_stipple_enabled() && (m_graphics_state & rsx::pipeline_state::polygon_stipple_pattern_dirty);
if (manually_flush_ring_buffers)
{
@ -892,7 +892,7 @@ void GLGSRender::load_program_env()
std::memcpy(mapping.first, rsx::method_registers.polygon_stipple_pattern(), 128);
m_raster_env_ring_buffer->bind_range(GL_RASTERIZER_STATE_BIND_SLOT, mapping.second, 128);
m_graphics_state &= ~(rsx::pipeline_state::polygon_stipple_pattern_dirty);
m_graphics_state.clear(rsx::pipeline_state::polygon_stipple_pattern_dirty);
}
if (update_instruction_buffers)
@ -954,8 +954,12 @@ void GLGSRender::load_program_env()
}
}
const u32 handled_flags = (rsx::pipeline_state::fragment_state_dirty | rsx::pipeline_state::vertex_state_dirty | rsx::pipeline_state::transform_constants_dirty | rsx::pipeline_state::fragment_constants_dirty | rsx::pipeline_state::fragment_texture_state_dirty);
m_graphics_state &= ~handled_flags;
m_graphics_state.clear(
rsx::pipeline_state::fragment_state_dirty |
rsx::pipeline_state::vertex_state_dirty |
rsx::pipeline_state::transform_constants_dirty |
rsx::pipeline_state::fragment_constants_dirty |
rsx::pipeline_state::fragment_texture_state_dirty);
}
void GLGSRender::update_vertex_env(const gl::vertex_upload_info& upload_info)
@ -1065,7 +1069,7 @@ void GLGSRender::do_local_task(rsx::FIFO::state state)
//This will re-engage locks and break the texture cache if another thread is waiting in access violation handler!
//Only call when there are no waiters
m_gl_texture_cache.do_update();
m_graphics_state &= ~rsx::pipeline_state::framebuffer_reads_dirty;
m_graphics_state.clear(rsx::pipeline_state::framebuffer_reads_dirty);
}
}

View File

@ -78,7 +78,7 @@ class GLGSRender : public GSRender, public ::rsx::reports::ZCULL_control
const GLFragmentProgram *m_fragment_prog = nullptr;
const GLVertexProgram *m_vertex_prog = nullptr;
u32 m_interpreter_state = 0;
rsx::flags32_t m_interpreter_state = 0;
gl::shader_interpreter m_shader_interpreter;
gl_render_targets m_rtts;

View File

@ -409,11 +409,11 @@ void GLGSRender::flip(const rsx::display_flip_info_t& info)
return false;
});
if (m_draw_fbo && !m_rtts_dirty)
if (m_draw_fbo && !m_graphics_state.test(rsx::rtt_config_dirty))
{
// Always restore the active framebuffer
m_draw_fbo->bind();
set_viewport();
set_scissor(!!(m_graphics_state & rsx::pipeline_state::scissor_setup_clipped));
set_scissor(m_graphics_state & rsx::pipeline_state::scissor_setup_clipped);
}
}

View File

@ -103,7 +103,7 @@ u8 rsx::internals::get_pixel_size(rsx::surface_depth_format format)
void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool /*skip_reading*/)
{
const bool clipped_scissor = (context == rsx::framebuffer_creation_context::context_draw);
if (m_current_framebuffer_context == context && !m_rtts_dirty && m_draw_fbo)
if (m_current_framebuffer_context == context && !m_graphics_state.test(rsx::rtt_config_dirty) && m_draw_fbo)
{
// Fast path
// Framebuffer usage has not changed, framebuffer exists and config regs have not changed
@ -111,9 +111,8 @@ void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool /*
return;
}
m_rtts_dirty = false;
m_graphics_state.clear(rsx::rtt_config_dirty | rsx::rtt_config_contested);
framebuffer_status_valid = false;
m_framebuffer_state_contested = false;
get_framebuffer_layout(context, m_framebuffer_layout);
if (!framebuffer_status_valid)

View File

@ -524,11 +524,10 @@ namespace rsx
return on_access_violation(address, is_writing);
};
m_rtts_dirty = true;
m_textures_dirty.fill(true);
m_vertex_textures_dirty.fill(true);
m_graphics_state = pipeline_state::all_dirty;
m_graphics_state |= pipeline_state::all_dirty;
g_user_asked_for_frame_capture = false;
@ -685,7 +684,7 @@ namespace rsx
push_buf.clear();
}
m_graphics_state &= ~rsx::pipeline_state::push_buffer_arrays_dirty;
m_graphics_state.clear(rsx::pipeline_state::push_buffer_arrays_dirty);
}
element_push_buffer.clear();
@ -1359,8 +1358,8 @@ namespace rsx
layout.width = rsx::method_registers.surface_clip_width();
layout.height = rsx::method_registers.surface_clip_height();
m_graphics_state.clear(rsx::rtt_config_contested);
framebuffer_status_valid = false;
m_framebuffer_state_contested = false;
m_current_framebuffer_context = context;
if (layout.width == 0 || layout.height == 0)
@ -1481,7 +1480,10 @@ namespace rsx
}
color_buffer_unused = !color_write_enabled || layout.target == rsx::surface_target::none;
m_framebuffer_state_contested = color_buffer_unused || depth_buffer_unused;
if (color_buffer_unused || depth_buffer_unused)
{
m_graphics_state.set(rsx::rtt_config_contested);
}
break;
default:
fmt::throw_exception("Unknown framebuffer context 0x%x", static_cast<u32>(context));
@ -1686,7 +1688,7 @@ namespace rsx
void thread::on_framebuffer_options_changed(u32 opt)
{
if (m_rtts_dirty)
if (m_graphics_state & rsx::rtt_config_dirty)
{
// Nothing to do
return;
@ -1773,9 +1775,9 @@ namespace rsx
{
evaluate_depth_buffer_state();
if (m_framebuffer_state_contested)
if (m_graphics_state.test(rsx::rtt_config_contested) && evaluate_depth_buffer_contested())
{
m_rtts_dirty |= evaluate_depth_buffer_contested();
m_graphics_state.set(rsx::rtt_config_dirty);
}
break;
}
@ -1798,16 +1800,16 @@ namespace rsx
evaluate_stencil_buffer_state();
}
if (m_framebuffer_state_contested)
if (m_graphics_state.test(rsx::rtt_config_contested) && evaluate_depth_buffer_contested())
{
m_rtts_dirty |= evaluate_depth_buffer_contested();
m_graphics_state.set(rsx::rtt_config_dirty);
}
break;
}
case NV4097_SET_COLOR_MASK:
case NV4097_SET_COLOR_MASK_MRT:
{
if (!m_framebuffer_state_contested) [[likely]]
if (!m_graphics_state.test(rsx::rtt_config_contested)) [[likely]]
{
// Update write masks and continue
evaluate_color_buffer_state();
@ -1824,7 +1826,7 @@ namespace rsx
if (!old_state && new_state)
{
// Color buffers now in use
m_rtts_dirty = true;
m_graphics_state.set(rsx::rtt_config_dirty);
}
}
break;
@ -1836,16 +1838,16 @@ namespace rsx
bool thread::get_scissor(areau& region, bool clip_viewport)
{
if (!(m_graphics_state & rsx::pipeline_state::scissor_config_state_dirty))
if (!m_graphics_state.test(rsx::pipeline_state::scissor_config_state_dirty))
{
if (clip_viewport == !!(m_graphics_state & rsx::pipeline_state::scissor_setup_clipped))
if (clip_viewport == m_graphics_state.test(rsx::pipeline_state::scissor_setup_clipped))
{
// Nothing to do
return false;
}
}
m_graphics_state &= ~(rsx::pipeline_state::scissor_config_state_dirty | rsx::pipeline_state::scissor_setup_clipped);
m_graphics_state.clear(rsx::pipeline_state::scissor_config_state_dirty | rsx::pipeline_state::scissor_setup_clipped);
u16 x1, x2, y1, y2;
@ -1889,7 +1891,7 @@ namespace rsx
if (m_graphics_state & rsx::pipeline_state::scissor_setup_invalid)
{
m_graphics_state &= ~rsx::pipeline_state::scissor_setup_invalid;
m_graphics_state.clear(rsx::pipeline_state::scissor_setup_invalid);
framebuffer_status_valid = true;
}
@ -1901,10 +1903,12 @@ namespace rsx
void thread::prefetch_fragment_program()
{
if (!(m_graphics_state & rsx::pipeline_state::fragment_program_ucode_dirty))
if (!m_graphics_state.test(rsx::pipeline_state::fragment_program_ucode_dirty))
{
return;
}
m_graphics_state &= ~rsx::pipeline_state::fragment_program_ucode_dirty;
m_graphics_state.clear(rsx::pipeline_state::fragment_program_ucode_dirty);
// Request for update of fragment constants if the program block is invalidated
m_graphics_state |= rsx::pipeline_state::fragment_constants_dirty;
@ -1922,7 +1926,7 @@ namespace rsx
current_fragment_program.texture_state.import(current_fp_texture_state, current_fp_metadata.referenced_textures_mask);
current_fragment_program.valid = true;
if (!(m_graphics_state & rsx::pipeline_state::fragment_program_state_dirty))
if (!m_graphics_state.test(rsx::pipeline_state::fragment_program_state_dirty))
{
// Verify current texture state is valid
for (u32 textures_ref = current_fp_metadata.referenced_textures_mask, i = 0; textures_ref; textures_ref >>= 1, ++i)
@ -1937,7 +1941,7 @@ namespace rsx
}
}
if (!(m_graphics_state & rsx::pipeline_state::fragment_program_state_dirty) &&
if (!m_graphics_state.test(rsx::pipeline_state::fragment_program_state_dirty) &&
(prev_textures_reference_mask != current_fp_metadata.referenced_textures_mask))
{
// If different textures are used, upload their coefficients.
@ -1948,10 +1952,12 @@ namespace rsx
void thread::prefetch_vertex_program()
{
if (!(m_graphics_state & rsx::pipeline_state::vertex_program_ucode_dirty))
if (!m_graphics_state.test(rsx::pipeline_state::vertex_program_ucode_dirty))
{
return;
}
m_graphics_state &= ~rsx::pipeline_state::vertex_program_ucode_dirty;
m_graphics_state.clear(rsx::pipeline_state::vertex_program_ucode_dirty);
// Reload transform constants unconditionally for now
m_graphics_state |= rsx::pipeline_state::transform_constants_dirty;
@ -1969,7 +1975,7 @@ namespace rsx
current_vertex_program.texture_state.import(current_vp_texture_state, current_vp_metadata.referenced_textures_mask);
if (!(m_graphics_state & rsx::pipeline_state::vertex_program_state_dirty))
if (!m_graphics_state.test(rsx::pipeline_state::vertex_program_state_dirty))
{
// Verify current texture state is valid
for (u32 textures_ref = current_vp_metadata.referenced_textures_mask, i = 0; textures_ref; textures_ref >>= 1, ++i)
@ -1993,10 +1999,12 @@ namespace rsx
void thread::get_current_vertex_program(const std::array<std::unique_ptr<rsx::sampled_image_descriptor_base>, rsx::limits::vertex_textures_count>& sampler_descriptors)
{
if (!(m_graphics_state & rsx::pipeline_state::vertex_program_dirty))
if (!m_graphics_state.test(rsx::pipeline_state::vertex_program_dirty))
{
return;
}
ensure(!(m_graphics_state & rsx::pipeline_state::vertex_program_ucode_dirty));
ensure(!m_graphics_state.test(rsx::pipeline_state::vertex_program_ucode_dirty));
current_vertex_program.output_mask = rsx::method_registers.vertex_attrib_output_mask();
current_vertex_program.ctrl = 0; // Reserved
@ -2195,12 +2203,14 @@ namespace rsx
void thread::get_current_fragment_program(const std::array<std::unique_ptr<rsx::sampled_image_descriptor_base>, rsx::limits::fragment_textures_count>& sampler_descriptors)
{
if (!(m_graphics_state & rsx::pipeline_state::fragment_program_dirty))
if (!m_graphics_state.test(rsx::pipeline_state::fragment_program_dirty))
{
return;
}
ensure(!(m_graphics_state & rsx::pipeline_state::fragment_program_ucode_dirty));
ensure(!m_graphics_state.test(rsx::pipeline_state::fragment_program_ucode_dirty));
m_graphics_state &= ~(rsx::pipeline_state::fragment_program_dirty);
m_graphics_state.clear(rsx::pipeline_state::fragment_program_dirty);
current_fragment_program.ctrl = rsx::method_registers.shader_control() & (CELL_GCM_SHADER_CONTROL_32_BITS_EXPORTS | CELL_GCM_SHADER_CONTROL_DEPTH_EXPORT);
current_fragment_program.texcoord_control_mask = rsx::method_registers.texcoord_control_mask();
@ -2428,9 +2438,7 @@ namespace rsx
rsx::method_registers.reset();
check_zcull_status(false);
nv4097::set_render_mode(this, 0, method_registers.registers[NV4097_SET_RENDER_ENABLE]);
m_graphics_state = pipeline_state::all_dirty;
m_rtts_dirty = true;
m_framebuffer_state_contested = false;
m_graphics_state |= pipeline_state::all_dirty;
}
void thread::init(u32 ctrlAddress)

View File

@ -82,6 +82,10 @@ namespace rsx
pipeline_config_dirty = 0x100000, // Generic pipeline configuration changes. Shader peek hint.
rtt_config_dirty = 0x200000, // Render target configuration changed
rtt_config_contested = 0x400000, // Render target configuration is indeterminate
texture_cache_state_dirty = 0x800000, // Texture cache state is indeterminate
fragment_program_dirty = fragment_program_ucode_dirty | fragment_program_state_dirty,
vertex_program_dirty = vertex_program_ucode_dirty | vertex_program_state_dirty,
invalidate_pipeline_bits = fragment_program_dirty | vertex_program_dirty,
@ -288,14 +292,12 @@ namespace rsx
bool send_event(u64, u64, u64);
bool m_rtts_dirty = true;
std::array<bool, 16> m_textures_dirty;
std::array<bool, 4> m_vertex_textures_dirty;
bool m_framebuffer_state_contested = false;
rsx::framebuffer_creation_context m_current_framebuffer_context = rsx::framebuffer_creation_context::context_draw;
rsx::atomic_bitmask_t<rsx::eng_interrupt_reason> m_eng_interrupt_mask;
u32 m_graphics_state = 0;
rsx::bitmask_t<rsx::pipeline_state> m_graphics_state;
u64 ROP_sync_timestamp = 0;
program_hash_util::fragment_program_utils::fragment_program_metadata current_fp_metadata = {};

View File

@ -225,7 +225,7 @@ void VKGSRender::update_draw_state()
bind_viewport();
m_current_command_buffer->flags &= ~vk::command_buffer::cb_reload_dynamic_state;
m_graphics_state &= ~(rsx::pipeline_state::polygon_offset_state_dirty | rsx::pipeline_state::depth_bounds_state_dirty);
m_graphics_state.clear(rsx::pipeline_state::polygon_offset_state_dirty | rsx::pipeline_state::depth_bounds_state_dirty);
m_frame_stats.setup_time += m_profiler.duration();
}

View File

@ -1355,7 +1355,7 @@ void VKGSRender::set_viewport()
}
m_current_command_buffer->flags |= vk::command_buffer::cb_reload_dynamic_state;
m_graphics_state &= ~(rsx::pipeline_state::zclip_config_state_dirty);
m_graphics_state.clear(rsx::pipeline_state::zclip_config_state_dirty);
}
void VKGSRender::set_scissor(bool clip_viewport)
@ -1382,7 +1382,7 @@ void VKGSRender::bind_viewport()
m_viewport.maxDepth = rsx::method_registers.clip_max();
}
m_graphics_state &= ~(rsx::pipeline_state::zclip_config_state_dirty);
m_graphics_state.clear(rsx::pipeline_state::zclip_config_state_dirty);
}
vkCmdSetViewport(*m_current_command_buffer, 0, 1, &m_viewport);
@ -1888,7 +1888,7 @@ void VKGSRender::do_local_task(rsx::FIFO::state state)
//This will re-engage locks and break the texture cache if another thread is waiting in access violation handler!
//Only call when there are no waiters
m_texture_cache.do_update();
m_graphics_state &= ~rsx::pipeline_state::framebuffer_reads_dirty;
m_graphics_state.clear(rsx::pipeline_state::framebuffer_reads_dirty);
}
}
@ -1934,7 +1934,7 @@ bool VKGSRender::load_program()
get_current_vertex_program(vs_sampler_state);
m_graphics_state &= ~rsx::pipeline_state::invalidate_pipeline_bits;
m_graphics_state.clear(rsx::pipeline_state::invalidate_pipeline_bits);
}
else if (!(m_graphics_state & rsx::pipeline_state::pipeline_config_dirty) &&
m_program &&
@ -1978,7 +1978,7 @@ bool VKGSRender::load_program()
// Fallthrough
m_pipeline_properties = properties;
m_graphics_state &= ~rsx::pipeline_state::pipeline_config_dirty;
m_graphics_state.clear(rsx::pipeline_state::pipeline_config_dirty);
}
else
{
@ -2152,7 +2152,7 @@ void VKGSRender::load_program_env()
m_raster_env_ring_info.unmap();
m_raster_env_buffer_info = { m_raster_env_ring_info.heap->value, mem, 128 };
m_graphics_state &= ~(rsx::pipeline_state::polygon_stipple_pattern_dirty);
m_graphics_state.clear(rsx::pipeline_state::polygon_stipple_pattern_dirty);
}
if (update_instruction_buffers)
@ -2219,9 +2219,13 @@ void VKGSRender::load_program_env()
m_program->bind_buffer({ predicate, 0, 4 }, binding_table.conditional_render_predicate_slot, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, m_current_frame->descriptor_set);
}
//Clear flags
const u32 handled_flags = (rsx::pipeline_state::fragment_state_dirty | rsx::pipeline_state::vertex_state_dirty | rsx::pipeline_state::transform_constants_dirty | rsx::pipeline_state::fragment_constants_dirty | rsx::pipeline_state::fragment_texture_state_dirty);
m_graphics_state &= ~handled_flags;
// Clear flags
m_graphics_state.clear(
rsx::pipeline_state::fragment_state_dirty |
rsx::pipeline_state::vertex_state_dirty |
rsx::pipeline_state::transform_constants_dirty |
rsx::pipeline_state::fragment_constants_dirty |
rsx::pipeline_state::fragment_texture_state_dirty);
}
void VKGSRender::update_vertex_env(u32 id, const vk::vertex_upload_info& vertex_info)
@ -2394,7 +2398,7 @@ void VKGSRender::close_and_submit_command_buffer(vk::fence* pFence, VkSemaphore
void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
{
const bool clipped_scissor = (context == rsx::framebuffer_creation_context::context_draw);
if (m_current_framebuffer_context == context && !m_rtts_dirty && m_draw_fbo)
if (m_current_framebuffer_context == context && !m_graphics_state.test(rsx::rtt_config_dirty) && m_draw_fbo)
{
// Fast path
// Framebuffer usage has not changed, framebuffer exists and config regs have not changed
@ -2402,9 +2406,8 @@ void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
return;
}
m_rtts_dirty = false;
m_graphics_state.clear(rsx::rtt_config_dirty | rsx::rtt_config_contested);
framebuffer_status_valid = false;
m_framebuffer_state_contested = false;
get_framebuffer_layout(context, m_framebuffer_layout);
if (!framebuffer_status_valid)

View File

@ -801,8 +801,8 @@ namespace rsx
break;
}
rsx->m_rtts_dirty = true;
rsx->m_framebuffer_state_contested = false;
rsx->m_graphics_state.set(rtt_config_dirty);
rsx->m_graphics_state.clear(rtt_config_contested);
}
void set_surface_format(thread* rsx, u32 reg, u32 arg)