summaryrefslogtreecommitdiffstats
path: root/src/video_core/engines
diff options
context:
space:
mode:
Diffstat (limited to 'src/video_core/engines')
-rw-r--r--src/video_core/engines/maxwell_3d.cpp53
-rw-r--r--src/video_core/engines/maxwell_3d.h212
-rw-r--r--src/video_core/engines/shader_bytecode.h8
3 files changed, 197 insertions, 76 deletions
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index d79c50919..2bc534be3 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -34,8 +34,49 @@ void Maxwell3D::InitializeRegisterDefaults() {
// Depth range near/far is not always set, but is expected to be the default 0.0f, 1.0f. This is
// needed for ARMS.
for (std::size_t viewport{}; viewport < Regs::NumViewports; ++viewport) {
- regs.viewport[viewport].depth_range_near = 0.0f;
- regs.viewport[viewport].depth_range_far = 1.0f;
+ regs.viewports[viewport].depth_range_near = 0.0f;
+ regs.viewports[viewport].depth_range_far = 1.0f;
+ }
+ // 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 (std::size_t blend_index = 0; blend_index < Regs::NumRenderTargets; blend_index++) {
+ regs.independent_blend[blend_index].equation_rgb = Regs::Blend::Equation::Add;
+ regs.independent_blend[blend_index].factor_source_rgb = Regs::Blend::Factor::One;
+ regs.independent_blend[blend_index].factor_dest_rgb = Regs::Blend::Factor::Zero;
+ regs.independent_blend[blend_index].equation_a = Regs::Blend::Equation::Add;
+ regs.independent_blend[blend_index].factor_source_a = Regs::Blend::Factor::One;
+ regs.independent_blend[blend_index].factor_dest_a = Regs::Blend::Factor::Zero;
+ }
+ 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_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;
+ // 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).
+ regs.point_size = 1.0f;
+
+ // TODO(bunnei): Some games do not initialize the color masks (e.g. Sonic Mania). Assuming a
+ // default of enabled fixes rendering here.
+ for (std::size_t color_mask = 0; color_mask < Regs::NumRenderTargets; color_mask++) {
+ regs.color_mask[color_mask].R.Assign(1);
+ regs.color_mask[color_mask].G.Assign(1);
+ regs.color_mask[color_mask].B.Assign(1);
+ regs.color_mask[color_mask].A.Assign(1);
}
}
@@ -92,7 +133,13 @@ void Maxwell3D::WriteReg(u32 method, u32 value, u32 remaining_params) {
debug_context->OnEvent(Tegra::DebugContext::Event::MaxwellCommandLoaded, nullptr);
}
- regs.reg_array[method] = value;
+ if (regs.reg_array[method] != value) {
+ regs.reg_array[method] = value;
+ if (method >= MAXWELL3D_REG_INDEX(vertex_attrib_format) &&
+ method < MAXWELL3D_REG_INDEX(vertex_attrib_format) + regs.vertex_attrib_format.size()) {
+ dirty_flags.vertex_attrib_format = true;
+ }
+ }
switch (method) {
case MAXWELL3D_REG_INDEX(macros.data): {
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 50873813e..4f137e693 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -345,6 +345,14 @@ public:
Invert = 6,
IncrWrap = 7,
DecrWrap = 8,
+ KeepOGL = 0x1E00,
+ ZeroOGL = 0,
+ ReplaceOGL = 0x1E01,
+ IncrOGL = 0x1E02,
+ DecrOGL = 0x1E03,
+ InvertOGL = 0x150A,
+ IncrWrapOGL = 0x8507,
+ DecrWrapOGL = 0x8508,
};
enum class MemoryLayout : u32 {
@@ -381,6 +389,13 @@ public:
ReverseSubtract = 3,
Min = 4,
Max = 5,
+
+ // These values are used by Nouveau and some games.
+ AddGL = 0x8006,
+ SubtractGL = 0x8007,
+ ReverseSubtractGL = 0x8008,
+ MinGL = 0x800a,
+ MaxGL = 0x800b
};
enum class Factor : u32 {
@@ -462,6 +477,77 @@ public:
}
};
+ struct ColorMask {
+ union {
+ u32 raw;
+ BitField<0, 4, u32> R;
+ BitField<4, 4, u32> G;
+ BitField<8, 4, u32> B;
+ BitField<12, 4, u32> A;
+ };
+ };
+
+ struct ViewportTransform {
+ f32 scale_x;
+ f32 scale_y;
+ f32 scale_z;
+ f32 translate_x;
+ f32 translate_y;
+ f32 translate_z;
+ INSERT_PADDING_WORDS(2);
+
+ MathUtil::Rectangle<s32> GetRect() const {
+ return {
+ GetX(), // left
+ GetY() + GetHeight(), // top
+ GetX() + GetWidth(), // right
+ GetY() // bottom
+ };
+ };
+
+ s32 GetX() const {
+ return static_cast<s32>(std::max(0.0f, translate_x - std::fabs(scale_x)));
+ }
+
+ s32 GetY() const {
+ return static_cast<s32>(std::max(0.0f, translate_y - std::fabs(scale_y)));
+ }
+
+ s32 GetWidth() const {
+ return static_cast<s32>(translate_x + std::fabs(scale_x)) - GetX();
+ }
+
+ s32 GetHeight() const {
+ return static_cast<s32>(translate_y + std::fabs(scale_y)) - GetY();
+ }
+ };
+
+ struct ScissorTest {
+ u32 enable;
+ union {
+ BitField<0, 16, u32> min_x;
+ BitField<16, 16, u32> max_x;
+ };
+ union {
+ BitField<0, 16, u32> min_y;
+ BitField<16, 16, u32> max_y;
+ };
+ u32 fill;
+ };
+
+ struct ViewPort {
+ union {
+ BitField<0, 16, u32> x;
+ BitField<16, 16, u32> width;
+ };
+ union {
+ BitField<0, 16, u32> y;
+ BitField<16, 16, u32> height;
+ };
+ float depth_range_near;
+ float depth_range_far;
+ };
+
bool IsShaderConfigEnabled(std::size_t index) const {
// The VertexB is always enabled.
if (index == static_cast<std::size_t>(Regs::ShaderProgram::VertexB)) {
@@ -487,55 +573,11 @@ public:
INSERT_PADDING_WORDS(0x2E);
- RenderTargetConfig rt[NumRenderTargets];
+ std::array<RenderTargetConfig, NumRenderTargets> rt;
- struct {
- f32 scale_x;
- f32 scale_y;
- f32 scale_z;
- f32 translate_x;
- f32 translate_y;
- f32 translate_z;
- INSERT_PADDING_WORDS(2);
-
- MathUtil::Rectangle<s32> GetRect() const {
- return {
- GetX(), // left
- GetY() + GetHeight(), // top
- GetX() + GetWidth(), // right
- GetY() // bottom
- };
- };
+ std::array<ViewportTransform, NumViewports> viewport_transform;
- s32 GetX() const {
- return static_cast<s32>(std::max(0.0f, translate_x - std::fabs(scale_x)));
- }
-
- s32 GetY() const {
- return static_cast<s32>(std::max(0.0f, translate_y - std::fabs(scale_y)));
- }
-
- s32 GetWidth() const {
- return static_cast<s32>(translate_x + std::fabs(scale_x)) - GetX();
- }
-
- s32 GetHeight() const {
- return static_cast<s32>(translate_y + std::fabs(scale_y)) - GetY();
- }
- } viewport_transform[NumViewports];
-
- struct {
- union {
- BitField<0, 16, u32> x;
- BitField<16, 16, u32> width;
- };
- union {
- BitField<0, 16, u32> y;
- BitField<16, 16, u32> height;
- };
- float depth_range_near;
- float depth_range_far;
- } viewport[NumViewports];
+ std::array<ViewPort, NumViewports> viewports;
INSERT_PADDING_WORDS(0x1D);
@@ -553,25 +595,19 @@ public:
INSERT_PADDING_WORDS(0x17);
- struct {
- u32 enable;
- union {
- BitField<0, 16, u32> min_x;
- BitField<16, 16, u32> max_x;
- };
- union {
- BitField<0, 16, u32> min_y;
- BitField<16, 16, u32> max_y;
- };
- } scissor_test;
+ std::array<ScissorTest, NumViewports> scissor_test;
- INSERT_PADDING_WORDS(0x52);
+ INSERT_PADDING_WORDS(0x15);
s32 stencil_back_func_ref;
u32 stencil_back_mask;
u32 stencil_back_func_mask;
- INSERT_PADDING_WORDS(0x13);
+ INSERT_PADDING_WORDS(0xC);
+
+ u32 color_mask_common;
+
+ INSERT_PADDING_WORDS(0x6);
u32 rt_separate_frag_data;
@@ -595,7 +631,16 @@ public:
}
} zeta;
- INSERT_PADDING_WORDS(0x5B);
+ INSERT_PADDING_WORDS(0x41);
+
+ union {
+ BitField<0, 4, u32> stencil;
+ BitField<4, 4, u32> unknown;
+ BitField<8, 4, u32> scissor;
+ BitField<12, 4, u32> viewport;
+ } clear_flags;
+
+ INSERT_PADDING_WORDS(0x19);
std::array<VertexAttribute, NumVertexAttributes> vertex_attrib_format;
@@ -646,8 +691,14 @@ public:
ComparisonOp depth_test_func;
float alpha_test_ref;
ComparisonOp alpha_test_func;
-
- INSERT_PADDING_WORDS(0x9);
+ u32 draw_tfb_stride;
+ struct {
+ float r;
+ float g;
+ float b;
+ float a;
+ } blend_color;
+ INSERT_PADDING_WORDS(0x4);
struct {
u32 separate_alpha;
@@ -672,7 +723,9 @@ public:
u32 stencil_front_func_mask;
u32 stencil_front_mask;
- INSERT_PADDING_WORDS(0x3);
+ INSERT_PADDING_WORDS(0x2);
+
+ u32 frag_color_clamp;
union {
BitField<4, 1, u32> triangle_rast_flip;
@@ -690,7 +743,12 @@ public:
u32 zeta_enable;
- INSERT_PADDING_WORDS(0x8);
+ union {
+ BitField<0, 1, u32> alpha_to_coverage;
+ BitField<4, 1, u32> alpha_to_one;
+ } multisample_control;
+
+ INSERT_PADDING_WORDS(0x7);
struct {
u32 tsc_address_high;
@@ -841,8 +899,9 @@ public:
BitField<6, 4, u32> RT;
BitField<10, 11, u32> layer;
} clear_buffers;
-
- INSERT_PADDING_WORDS(0x4B);
+ INSERT_PADDING_WORDS(0xB);
+ std::array<ColorMask, NumRenderTargets> color_mask;
+ INSERT_PADDING_WORDS(0x38);
struct {
u32 query_address_high;
@@ -983,6 +1042,12 @@ public:
State state{};
MemoryManager& memory_manager;
+ struct DirtyFlags {
+ bool vertex_attrib_format = true;
+ };
+
+ DirtyFlags dirty_flags;
+
/// Reads a register value located at the input method address
u32 GetRegisterValue(u32 method) const;
@@ -1065,8 +1130,8 @@ private:
ASSERT_REG_POSITION(macros, 0x45);
ASSERT_REG_POSITION(tfb_enabled, 0x1D1);
ASSERT_REG_POSITION(rt, 0x200);
-ASSERT_REG_POSITION(viewport_transform[0], 0x280);
-ASSERT_REG_POSITION(viewport, 0x300);
+ASSERT_REG_POSITION(viewport_transform, 0x280);
+ASSERT_REG_POSITION(viewports, 0x300);
ASSERT_REG_POSITION(vertex_buffer, 0x35D);
ASSERT_REG_POSITION(clear_color[0], 0x360);
ASSERT_REG_POSITION(clear_depth, 0x364);
@@ -1075,8 +1140,10 @@ ASSERT_REG_POSITION(scissor_test, 0x380);
ASSERT_REG_POSITION(stencil_back_func_ref, 0x3D5);
ASSERT_REG_POSITION(stencil_back_mask, 0x3D6);
ASSERT_REG_POSITION(stencil_back_func_mask, 0x3D7);
+ASSERT_REG_POSITION(color_mask_common, 0x3E4);
ASSERT_REG_POSITION(rt_separate_frag_data, 0x3EB);
ASSERT_REG_POSITION(zeta, 0x3F8);
+ASSERT_REG_POSITION(clear_flags, 0x43E);
ASSERT_REG_POSITION(vertex_attrib_format, 0x458);
ASSERT_REG_POSITION(rt_control, 0x487);
ASSERT_REG_POSITION(zeta_width, 0x48a);
@@ -1087,6 +1154,10 @@ ASSERT_REG_POSITION(depth_write_enabled, 0x4BA);
ASSERT_REG_POSITION(alpha_test_enabled, 0x4BB);
ASSERT_REG_POSITION(d3d_cull_mode, 0x4C2);
ASSERT_REG_POSITION(depth_test_func, 0x4C3);
+ASSERT_REG_POSITION(alpha_test_ref, 0x4C4);
+ASSERT_REG_POSITION(alpha_test_func, 0x4C5);
+ASSERT_REG_POSITION(draw_tfb_stride, 0x4C6);
+ASSERT_REG_POSITION(blend_color, 0x4C7);
ASSERT_REG_POSITION(blend, 0x4CF);
ASSERT_REG_POSITION(stencil_enable, 0x4E0);
ASSERT_REG_POSITION(stencil_front_op_fail, 0x4E1);
@@ -1096,10 +1167,12 @@ ASSERT_REG_POSITION(stencil_front_func_func, 0x4E4);
ASSERT_REG_POSITION(stencil_front_func_ref, 0x4E5);
ASSERT_REG_POSITION(stencil_front_func_mask, 0x4E6);
ASSERT_REG_POSITION(stencil_front_mask, 0x4E7);
+ASSERT_REG_POSITION(frag_color_clamp, 0x4EA);
ASSERT_REG_POSITION(screen_y_control, 0x4EB);
ASSERT_REG_POSITION(vb_element_base, 0x50D);
ASSERT_REG_POSITION(point_size, 0x546);
ASSERT_REG_POSITION(zeta_enable, 0x54E);
+ASSERT_REG_POSITION(multisample_control, 0x54F);
ASSERT_REG_POSITION(tsc, 0x557);
ASSERT_REG_POSITION(tic, 0x55D);
ASSERT_REG_POSITION(stencil_two_side_enable, 0x565);
@@ -1117,6 +1190,7 @@ ASSERT_REG_POSITION(instanced_arrays, 0x620);
ASSERT_REG_POSITION(cull, 0x646);
ASSERT_REG_POSITION(logic_op, 0x671);
ASSERT_REG_POSITION(clear_buffers, 0x674);
+ASSERT_REG_POSITION(color_mask, 0x680);
ASSERT_REG_POSITION(query, 0x6C0);
ASSERT_REG_POSITION(vertex_array[0], 0x700);
ASSERT_REG_POSITION(independent_blend, 0x780);
diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h
index 83a6fd875..7e8449bc4 100644
--- a/src/video_core/engines/shader_bytecode.h
+++ b/src/video_core/engines/shader_bytecode.h
@@ -153,6 +153,7 @@ enum class PredCondition : u64 {
NotEqual = 5,
GreaterEqual = 6,
LessThanWithNan = 9,
+ LessEqualWithNan = 11,
GreaterThanWithNan = 12,
NotEqualWithNan = 13,
GreaterEqualWithNan = 14,
@@ -261,7 +262,7 @@ enum class FlowCondition : u64 {
Fcsm_Tr = 0x1C, // TODO(bunnei): What is this used for?
};
-enum class ControlCode : u64 {
+enum class ConditionCode : u64 {
F = 0,
LT = 1,
EQ = 2,
@@ -569,7 +570,6 @@ union Instruction {
BitField<39, 2, u64> tab5cb8_2;
BitField<41, 3, u64> tab5c68_1;
BitField<44, 2, u64> tab5c68_0;
- BitField<47, 1, u64> cc;
BitField<48, 1, u64> negate_b;
} fmul;
@@ -831,7 +831,7 @@ union Instruction {
union {
BitField<0, 3, u64> pred0;
BitField<3, 3, u64> pred3;
- BitField<8, 5, ControlCode> cc; // flag in cc
+ BitField<8, 5, ConditionCode> cc; // flag in cc
BitField<39, 3, u64> pred39;
BitField<42, 1, u64> neg_pred39;
BitField<45, 4, PredOperation> op; // op with pred39
@@ -1235,7 +1235,7 @@ union Instruction {
BitField<60, 1, u64> is_b_gpr;
BitField<59, 1, u64> is_c_gpr;
BitField<20, 24, s64> smem_imm;
- BitField<0, 5, ControlCode> flow_control_code;
+ BitField<0, 5, ConditionCode> flow_condition_code;
Attribute attribute;
Sampler sampler;