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.cpp16
-rw-r--r--src/video_core/engines/maxwell_3d.h192
-rw-r--r--src/video_core/engines/shader_bytecode.h27
-rw-r--r--src/video_core/engines/shader_header.h11
4 files changed, 171 insertions, 75 deletions
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 1772882b2..f0a5470b9 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -34,8 +34,8 @@ 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
@@ -66,6 +66,18 @@ void Maxwell3D::InitializeRegisterDefaults() {
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);
+ }
}
void Maxwell3D::CallMacroMethod(u32 method, std::vector<u32> parameters) {
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 0848b7121..e44a23135 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -389,6 +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 {
@@ -480,6 +487,67 @@ public:
};
};
+ 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)) {
@@ -505,55 +573,11 @@ public:
INSERT_PADDING_WORDS(0x2E);
- RenderTargetConfig rt[NumRenderTargets];
-
- 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
- };
- };
-
- 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();
- }
+ std::array<RenderTargetConfig, NumRenderTargets> rt;
- s32 GetHeight() const {
- return static_cast<s32>(translate_y + std::fabs(scale_y)) - GetY();
- }
- } viewport_transform[NumViewports];
+ std::array<ViewportTransform, NumViewports> viewport_transform;
- 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);
@@ -566,24 +590,22 @@ public:
float clear_color[4];
float clear_depth;
+
INSERT_PADDING_WORDS(0x3);
+
s32 clear_stencil;
- INSERT_PADDING_WORDS(0x17);
+ INSERT_PADDING_WORDS(0x7);
- 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;
+ u32 polygon_offset_point_enable;
+ u32 polygon_offset_line_enable;
+ u32 polygon_offset_fill_enable;
+
+ INSERT_PADDING_WORDS(0xD);
+
+ std::array<ScissorTest, NumViewports> scissor_test;
- INSERT_PADDING_WORDS(0x52);
+ INSERT_PADDING_WORDS(0x15);
s32 stencil_back_func_ref;
u32 stencil_back_mask;
@@ -617,7 +639,16 @@ 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;
@@ -700,9 +731,12 @@ public:
u32 stencil_front_func_mask;
u32 stencil_front_mask;
- INSERT_PADDING_WORDS(0x3);
+ INSERT_PADDING_WORDS(0x2);
+
+ u32 frag_color_clamp;
union {
+ BitField<0, 1, u32> y_negate;
BitField<4, 1, u32> triangle_rast_flip;
} screen_y_control;
@@ -718,7 +752,12 @@ 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;
@@ -731,7 +770,11 @@ public:
}
} tsc;
- INSERT_PADDING_WORDS(0x3);
+ INSERT_PADDING_WORDS(0x1);
+
+ float polygon_offset_factor;
+
+ INSERT_PADDING_WORDS(0x1);
struct {
u32 tic_address_high;
@@ -756,7 +799,9 @@ public:
u32 framebuffer_srgb;
- INSERT_PADDING_WORDS(0x12);
+ float polygon_offset_units;
+
+ INSERT_PADDING_WORDS(0x11);
union {
BitField<2, 1, u32> coord_origin;
@@ -833,7 +878,9 @@ public:
INSERT_PADDING_WORDS(0x7);
- INSERT_PADDING_WORDS(0x20);
+ INSERT_PADDING_WORDS(0x1F);
+
+ float polygon_offset_clamp;
struct {
u32 is_instanced[NumVertexArrays];
@@ -1105,12 +1152,15 @@ 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);
ASSERT_REG_POSITION(clear_stencil, 0x368);
+ASSERT_REG_POSITION(polygon_offset_point_enable, 0x370);
+ASSERT_REG_POSITION(polygon_offset_line_enable, 0x371);
+ASSERT_REG_POSITION(polygon_offset_fill_enable, 0x372);
ASSERT_REG_POSITION(scissor_test, 0x380);
ASSERT_REG_POSITION(stencil_back_func_ref, 0x3D5);
ASSERT_REG_POSITION(stencil_back_mask, 0x3D6);
@@ -1118,6 +1168,7 @@ 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);
@@ -1141,11 +1192,14 @@ 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(polygon_offset_factor, 0x55b);
ASSERT_REG_POSITION(tic, 0x55D);
ASSERT_REG_POSITION(stencil_two_side_enable, 0x565);
ASSERT_REG_POSITION(stencil_back_op_fail, 0x566);
@@ -1153,11 +1207,13 @@ ASSERT_REG_POSITION(stencil_back_op_zfail, 0x567);
ASSERT_REG_POSITION(stencil_back_op_zpass, 0x568);
ASSERT_REG_POSITION(stencil_back_func_func, 0x569);
ASSERT_REG_POSITION(framebuffer_srgb, 0x56E);
+ASSERT_REG_POSITION(polygon_offset_units, 0x56F);
ASSERT_REG_POSITION(point_coord_replace, 0x581);
ASSERT_REG_POSITION(code_address, 0x582);
ASSERT_REG_POSITION(draw, 0x585);
ASSERT_REG_POSITION(primitive_restart, 0x591);
ASSERT_REG_POSITION(index_array, 0x5F2);
+ASSERT_REG_POSITION(polygon_offset_clamp, 0x61F);
ASSERT_REG_POSITION(instanced_arrays, 0x620);
ASSERT_REG_POSITION(cull, 0x646);
ASSERT_REG_POSITION(logic_op, 0x671);
diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h
index 83a6fd875..b9faaf8e0 100644
--- a/src/video_core/engines/shader_bytecode.h
+++ b/src/video_core/engines/shader_bytecode.h
@@ -82,6 +82,8 @@ union Attribute {
Position = 7,
Attribute_0 = 8,
Attribute_31 = 39,
+ ClipDistances0123 = 44,
+ ClipDistances4567 = 45,
PointCoord = 46,
// This attribute contains a tuple of (~, ~, InstanceId, VertexId) when inside a vertex
// shader, and a tuple of (TessCoord.x, TessCoord.y, TessCoord.z, ~) when inside a Tess Eval
@@ -153,6 +155,7 @@ enum class PredCondition : u64 {
NotEqual = 5,
GreaterEqual = 6,
LessThanWithNan = 9,
+ LessEqualWithNan = 11,
GreaterThanWithNan = 12,
NotEqualWithNan = 13,
GreaterEqualWithNan = 14,
@@ -261,7 +264,7 @@ 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,
@@ -365,6 +368,11 @@ enum class HalfPrecision : u64 {
FMZ = 2,
};
+enum class R2pMode : u64 {
+ Pr = 0,
+ Cc = 1,
+};
+
enum class IpaInterpMode : u64 {
Linear = 0,
Perspective = 1,
@@ -569,7 +577,6 @@ 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 +838,7 @@ 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
@@ -855,6 +862,12 @@ union Instruction {
} hsetp2;
union {
+ BitField<40, 1, R2pMode> mode;
+ BitField<41, 2, u64> byte;
+ BitField<20, 7, u64> immediate_mask;
+ } r2p;
+
+ union {
BitField<39, 3, u64> pred39;
BitField<42, 1, u64> neg_pred;
BitField<43, 1, u64> neg_a;
@@ -1235,7 +1248,7 @@ 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;
@@ -1256,6 +1269,7 @@ public:
BFE_C,
BFE_R,
BFE_IMM,
+ BFI_IMM_R,
BRA,
PBK,
LD_A,
@@ -1381,6 +1395,7 @@ public:
PSETP,
PSET,
CSETP,
+ R2P_IMM,
XMAD_IMM,
XMAD_CR,
XMAD_RC,
@@ -1396,6 +1411,7 @@ public:
ArithmeticHalf,
ArithmeticHalfImmediate,
Bfe,
+ Bfi,
Shift,
Ffma,
Hfma2,
@@ -1410,6 +1426,7 @@ public:
HalfSetPredicate,
PredicateSetPredicate,
PredicateSetRegister,
+ RegisterSetPredicate,
Conversion,
Xmad,
Unknown,
@@ -1613,6 +1630,7 @@ private:
INST("0100110000000---", Id::BFE_C, Type::Bfe, "BFE_C"),
INST("0101110000000---", Id::BFE_R, Type::Bfe, "BFE_R"),
INST("0011100-00000---", Id::BFE_IMM, Type::Bfe, "BFE_IMM"),
+ INST("0011011-11110---", Id::BFI_IMM_R, Type::Bfi, "BFI_IMM_R"),
INST("0100110001000---", Id::LOP_C, Type::ArithmeticInteger, "LOP_C"),
INST("0101110001000---", Id::LOP_R, Type::ArithmeticInteger, "LOP_R"),
INST("0011100001000---", Id::LOP_IMM, Type::ArithmeticInteger, "LOP_IMM"),
@@ -1647,6 +1665,7 @@ private:
INST("0101000010001---", Id::PSET, Type::PredicateSetRegister, "PSET"),
INST("0101000010010---", Id::PSETP, Type::PredicateSetPredicate, "PSETP"),
INST("010100001010----", Id::CSETP, Type::PredicateSetPredicate, "CSETP"),
+ INST("0011100-11110---", Id::R2P_IMM, Type::RegisterSetPredicate, "R2P_IMM"),
INST("0011011-00------", Id::XMAD_IMM, Type::Xmad, "XMAD_IMM"),
INST("0100111---------", Id::XMAD_CR, Type::Xmad, "XMAD_CR"),
INST("010100010-------", Id::XMAD_RC, Type::Xmad, "XMAD_RC"),
diff --git a/src/video_core/engines/shader_header.h b/src/video_core/engines/shader_header.h
index a0e015c4b..99c34649f 100644
--- a/src/video_core/engines/shader_header.h
+++ b/src/video_core/engines/shader_header.h
@@ -62,7 +62,16 @@ struct Header {
INSERT_PADDING_BYTES(1); // ImapSystemValuesB
INSERT_PADDING_BYTES(16); // ImapGenericVector[32]
INSERT_PADDING_BYTES(2); // ImapColor
- INSERT_PADDING_BYTES(2); // ImapSystemValuesC
+ union {
+ BitField<0, 8, u16> clip_distances;
+ BitField<8, 1, u16> point_sprite_s;
+ BitField<9, 1, u16> point_sprite_t;
+ BitField<10, 1, u16> fog_coordinate;
+ BitField<12, 1, u16> tessellation_eval_point_u;
+ BitField<13, 1, u16> tessellation_eval_point_v;
+ BitField<14, 1, u16> instance_id;
+ BitField<15, 1, u16> vertex_id;
+ };
INSERT_PADDING_BYTES(5); // ImapFixedFncTexture[10]
INSERT_PADDING_BYTES(1); // ImapReserved
INSERT_PADDING_BYTES(3); // OmapSystemValuesA
generated by cgit v1.2.3 (git 2.25.1) at 2024-08-18 22:04:32 +0200