From f6e7cae62c3a3df41fe86566d4b786454f90b48a Mon Sep 17 00:00:00 2001 From: FengChen Date: Thu, 27 Oct 2022 13:21:47 +0800 Subject: video_core: Fix drawing trigger mechanism regression --- src/video_core/engines/maxwell_3d.cpp | 129 ++++++++++++++++++---------------- src/video_core/engines/maxwell_3d.h | 2 + 2 files changed, 70 insertions(+), 61 deletions(-) diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp index 5208bea75..f9794dfe4 100644 --- a/src/video_core/engines/maxwell_3d.cpp +++ b/src/video_core/engines/maxwell_3d.cpp @@ -123,9 +123,6 @@ void Maxwell3D::InitializeRegisterDefaults() { draw_command[MAXWELL3D_REG_INDEX(vertex_buffer.count)] = true; draw_command[MAXWELL3D_REG_INDEX(index_buffer.first)] = true; draw_command[MAXWELL3D_REG_INDEX(index_buffer.count)] = true; - draw_command[MAXWELL3D_REG_INDEX(index_buffer32_first)] = true; - draw_command[MAXWELL3D_REG_INDEX(index_buffer16_first)] = true; - draw_command[MAXWELL3D_REG_INDEX(index_buffer8_first)] = true; draw_command[MAXWELL3D_REG_INDEX(draw_inline_index)] = true; draw_command[MAXWELL3D_REG_INDEX(inline_index_2x16.even)] = true; draw_command[MAXWELL3D_REG_INDEX(inline_index_4x8.index0)] = true; @@ -216,6 +213,21 @@ void Maxwell3D::ProcessMethodCall(u32 method, u32 argument, u32 nonshadow_argume return ProcessCBBind(3); case MAXWELL3D_REG_INDEX(bind_groups[4].raw_config): return ProcessCBBind(4); + 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 ProcessDraw(); + 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 ProcessDraw(); + 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; + return ProcessDraw(); case MAXWELL3D_REG_INDEX(topology_override): use_topology_override = true; return; @@ -583,6 +595,31 @@ void Maxwell3D::ProcessClearBuffers() { rasterizer->Clear(); } +void Maxwell3D::ProcessDraw(u32 instance_count) { + LOG_TRACE(HW_GPU, "called, topology={}, count={}", regs.draw.topology.Value(), + regs.vertex_buffer.count); + + 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_id == Maxwell3D::Regs::Draw::InstanceId::First || + regs.draw.instance_id != Maxwell3D::Regs::Draw::InstanceId::Unchanged, + "Illegal combination of instancing parameters"); + + ProcessTopologyOverride(); + + const bool is_indexed = regs.index_buffer.count && !regs.vertex_buffer.count; + if (ShouldExecute()) { + rasterizer->Draw(is_indexed, instance_count); + } + + if (is_indexed) { + regs.index_buffer.count = 0; + } else { + regs.vertex_buffer.count = 0; + } +} + void Maxwell3D::ProcessDeferredDraw() { if (deferred_draw_method.empty()) { return; @@ -596,23 +633,28 @@ void Maxwell3D::ProcessDeferredDraw() { DrawMode draw_mode{DrawMode::Undefined}; u32 instance_count = 1; - auto first_method = deferred_draw_method[0]; - if (MAXWELL3D_REG_INDEX(draw.begin) == first_method) { - // The minimum number of methods for drawing must be greater than or equal to - // 3[draw.begin->vertex(index)count->draw.end] to avoid errors in index mode drawing - if (deferred_draw_method.size() < 3) { - return; - } - draw_mode = (regs.draw.instance_id == Maxwell3D::Regs::Draw::InstanceId::Subsequent) || - (regs.draw.instance_id == Maxwell3D::Regs::Draw::InstanceId::Unchanged) - ? DrawMode::Instance - : DrawMode::General; - } else if (MAXWELL3D_REG_INDEX(index_buffer32_first) == first_method || - MAXWELL3D_REG_INDEX(index_buffer16_first) == first_method || - MAXWELL3D_REG_INDEX(index_buffer8_first) == first_method) { - draw_mode = DrawMode::General; + u32 index = 0; + u32 method = 0; + u32 method_count = static_cast(deferred_draw_method.size()); + for (; index < method_count && + (method = deferred_draw_method[index]) != MAXWELL3D_REG_INDEX(draw.begin); + ++index) + ; + + if (MAXWELL3D_REG_INDEX(draw.begin) != method) { + return; } + // The minimum number of methods for drawing must be greater than or equal to + // 3[draw.begin->vertex(index)count(first)->draw.end] to avoid errors in index mode drawing + if ((method_count - index) < 3) { + return; + } + draw_mode = (regs.draw.instance_id == Maxwell3D::Regs::Draw::InstanceId::Subsequent) || + (regs.draw.instance_id == Maxwell3D::Regs::Draw::InstanceId::Unchanged) + ? DrawMode::Instance + : DrawMode::General; + // Drawing will only begin with draw.begin or index_buffer method, other methods directly // clear if (draw_mode == DrawMode::Undefined) { @@ -622,53 +664,18 @@ void Maxwell3D::ProcessDeferredDraw() { if (draw_mode == DrawMode::Instance) { ASSERT_MSG(deferred_draw_method.size() % 4 == 0, "Instance mode method size error"); - instance_count = static_cast(deferred_draw_method.size()) / 4; + instance_count = static_cast(method_count - index) / 4; } else { - if (MAXWELL3D_REG_INDEX(index_buffer32_first) == first_method) { - regs.index_buffer.count = regs.index_buffer32_first.count; - regs.index_buffer.first = regs.index_buffer32_first.first; - dirty.flags[VideoCommon::Dirty::IndexBuffer] = true; - } else if (MAXWELL3D_REG_INDEX(index_buffer32_first) == first_method) { - regs.index_buffer.count = regs.index_buffer16_first.count; - regs.index_buffer.first = regs.index_buffer16_first.first; - dirty.flags[VideoCommon::Dirty::IndexBuffer] = true; - } else if (MAXWELL3D_REG_INDEX(index_buffer32_first) == first_method) { - regs.index_buffer.count = regs.index_buffer8_first.count; - regs.index_buffer.first = regs.index_buffer8_first.first; - dirty.flags[VideoCommon::Dirty::IndexBuffer] = true; - } else { - auto second_method = deferred_draw_method[1]; - if (MAXWELL3D_REG_INDEX(draw_inline_index) == second_method || - MAXWELL3D_REG_INDEX(inline_index_2x16.even) == second_method || - MAXWELL3D_REG_INDEX(inline_index_4x8.index0) == second_method) { - regs.index_buffer.count = static_cast(inline_index_draw_indexes.size() / 4); - regs.index_buffer.format = Regs::IndexFormat::UnsignedInt; - } + method = deferred_draw_method[index + 1]; + if (MAXWELL3D_REG_INDEX(draw_inline_index) == method || + MAXWELL3D_REG_INDEX(inline_index_2x16.even) == method || + MAXWELL3D_REG_INDEX(inline_index_4x8.index0) == method) { + regs.index_buffer.count = static_cast(inline_index_draw_indexes.size() / 4); + regs.index_buffer.format = Regs::IndexFormat::UnsignedInt; } } - LOG_TRACE(HW_GPU, "called, topology={}, count={}", regs.draw.topology.Value(), - regs.vertex_buffer.count); - - 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_id == Maxwell3D::Regs::Draw::InstanceId::First || - regs.draw.instance_id != Maxwell3D::Regs::Draw::InstanceId::Unchanged, - "Illegal combination of instancing parameters"); - - ProcessTopologyOverride(); - - const bool is_indexed = regs.index_buffer.count && !regs.vertex_buffer.count; - if (ShouldExecute()) { - rasterizer->Draw(is_indexed, instance_count); - } - - if (is_indexed) { - regs.index_buffer.count = 0; - } else { - regs.vertex_buffer.count = 0; - } + ProcessDraw(instance_count); deferred_draw_method.clear(); inline_index_draw_indexes.clear(); diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h index bd23ebc12..a948fcb14 100644 --- a/src/video_core/engines/maxwell_3d.h +++ b/src/video_core/engines/maxwell_3d.h @@ -3143,6 +3143,8 @@ private: /// Handles use of topology overrides (e.g., to avoid using a topology assigned from a macro) void ProcessTopologyOverride(); + void ProcessDraw(u32 instance_count = 1); + void ProcessDeferredDraw(); /// Returns a query's value or an empty object if the value will be deferred through a cache. -- cgit v1.2.3