add stencil support

This commit is contained in:
InoriRus
2022-03-21 19:23:27 +10:00
parent 02a77aa81a
commit 838a09c810
30 changed files with 2448 additions and 781 deletions
@@ -12,6 +12,8 @@
namespace Kyty::Libs::Graphics {
class CommandBuffer;
struct VulkanSwapchain
{
VkSwapchainKHR swapchain = nullptr;
@@ -102,6 +104,11 @@ struct VulkanBuffer
VkBufferUsageFlags usage = 0;
};
struct StorageVulkanBuffer: public VulkanBuffer
{
CommandBuffer* cmd_buffer = nullptr;
};
} // namespace Kyty::Libs::Graphics
#endif // KYTY_EMU_ENABLED
@@ -36,16 +36,16 @@ public:
[[nodiscard]] bool IsInvalid() const;
void Allocate();
void Free();
void Begin() const;
void End() const;
void Execute();
void ExecuteWithSemaphore();
VulkanFramebuffer* BeginRenderPass(RenderColorInfo* color, RenderDepthInfo* depth) const;
void EndRenderPass() const;
void WaitForFence();
void WaitForFenceAndReset();
void Allocate();
void Free();
void Begin() const;
void End() const;
void Execute();
void ExecuteWithSemaphore();
void BeginRenderPass(VulkanFramebuffer* framebuffer, RenderColorInfo* color, RenderDepthInfo* depth) const;
void EndRenderPass() const;
void WaitForFence();
void WaitForFenceAndReset();
[[nodiscard]] uint32_t GetIndex() const { return m_index; }
VulkanCommandPool* GetPool() { return m_pool; }
@@ -85,6 +85,7 @@ void GraphicsRenderDispatchDirect(CommandBuffer* buffer, HardwareContext* ctx, u
uint32_t thread_group_z, uint32_t mode);
void GraphicsRenderMemoryBarrier(CommandBuffer* buffer);
void GraphicsRenderRenderTextureBarrier(CommandBuffer* buffer, uint64_t vaddr, uint64_t size);
void GraphicsRenderMemoryFree(uint64_t vaddr, uint64_t size);
void DeleteFramebuffer(VideoOutVulkanImage* image);
void DeleteFramebuffer(DepthStencilVulkanImage* image);
@@ -9,24 +9,31 @@
namespace Kyty::Libs::Graphics {
struct ColorInfo
{
bool fmask_compression_enable = false;
uint32_t fmask_compression_mode = 0;
bool cmask_fast_clear_enable = false;
bool dcc_compression_enable = false;
bool neo_mode = false;
uint32_t cmask_tile_mode = 0;
uint32_t cmask_tile_mode_neo = 0;
uint32_t format = 0;
uint32_t channel_type = 0;
uint32_t channel_order = 0;
};
struct RenderTarget
{
uint64_t base_addr = 0;
uint32_t pitch_div8_minus1 = 0;
uint32_t fmask_pitch_div8_minus1 = 0;
uint32_t slice_div64_minus1 = 0;
uint32_t base_array_slice_index = 0;
uint32_t last_array_slice_index = 0;
bool fmask_compression_enable = false;
uint32_t fmask_compression_mode = 0;
bool cmask_fast_clear_enable = false;
bool dcc_compression_enable = false;
bool neo_mode = false;
uint32_t cmask_tile_mode = 0;
uint32_t cmask_tile_mode_neo = 0;
uint32_t format = 0;
uint32_t channel_type = 0;
uint32_t channel_order = 0;
uint64_t base_addr = 0;
uint32_t pitch_div8_minus1 = 0;
uint32_t fmask_pitch_div8_minus1 = 0;
uint32_t slice_div64_minus1 = 0;
uint32_t base_array_slice_index = 0;
uint32_t last_array_slice_index = 0;
ColorInfo color_info;
bool force_dest_alpha_to_one = false;
uint32_t tile_mode = 0;
uint32_t fmask_tile_mode = 0;
@@ -155,6 +162,28 @@ struct DepthControl
uint8_t stencilfunc_bf = 0;
};
struct StencilControl
{
uint8_t stencil_fail = 0;
uint8_t stencil_zpass = 0;
uint8_t stencil_zfail = 0;
uint8_t stencil_fail_bf = 0;
uint8_t stencil_zpass_bf = 0;
uint8_t stencil_zfail_bf = 0;
};
struct StencilMask
{
uint8_t stencil_testval = 0;
uint8_t stencil_mask = 0;
uint8_t stencil_writemask = 0;
uint8_t stencil_opval = 0;
uint8_t stencil_testval_bf = 0;
uint8_t stencil_mask_bf = 0;
uint8_t stencil_writemask_bf = 0;
uint8_t stencil_opval_bf = 0;
};
struct ModeControl
{
bool cull_front = false;
@@ -343,7 +372,8 @@ public:
void Reset() { *this = HardwareContext(); }
void SetRenderTarget(uint32_t slot, const RenderTarget& target) { m_render_targets[slot] = target; }
void SetRenderTarget(uint32_t slot, const RenderTarget& target) { m_render_targets[slot] = target; }
void SetColorInfo(uint32_t slot, const ColorInfo& color_info) { m_render_targets[slot].color_info = color_info; }
[[nodiscard]] const RenderTarget& GetRenderTargets(uint32_t slot) const { return m_render_targets[slot]; }
void SetBlendControl(uint32_t slot, const BlendControl& control) { m_blend_control[slot] = control; }
@@ -411,7 +441,16 @@ public:
m_vs.vs_embedded = true;
}
void SetPsShader(const PsStageRegisters* ps_regs) { m_ps.ps_regs = *ps_regs; }
void SetPsShader(const PsStageRegisters* ps_regs)
{
m_ps.ps_regs = *ps_regs;
m_ps.ps_embedded = false;
}
void SetPsEmbedded(uint32_t id)
{
m_ps.ps_embedded_id = id;
m_ps.ps_embedded = true;
}
void SetCsShader(const CsStageRegisters* cs_regs, uint32_t shader_modifier)
{
@@ -419,18 +458,22 @@ public:
m_cs.cs_shader_modifier = shader_modifier;
}
[[nodiscard]] const BlendColor& GetBlendColor() const { return m_blend_color; }
void SetBlendColor(const BlendColor& color) { m_blend_color = color; }
[[nodiscard]] const ClipControl& GetClipControl() const { return m_clip_control; }
void SetClipControl(const ClipControl& control) { m_clip_control = control; }
[[nodiscard]] const RenderControl& GetRenderControl() const { return m_render_control; }
void SetRenderControl(const RenderControl& control) { m_render_control = control; }
[[nodiscard]] const DepthControl& GetDepthControl() const { return m_depth_control; }
void SetDepthControl(const DepthControl& control) { m_depth_control = control; }
[[nodiscard]] const ModeControl& GetModeControl() const { return m_mode_control; }
void SetModeControl(const ModeControl& control) { m_mode_control = control; }
[[nodiscard]] const EqaaControl& GetEqaaControl() const { return m_eqaa_control; }
void SetEqaaControl(const EqaaControl& control) { m_eqaa_control = control; }
[[nodiscard]] const BlendColor& GetBlendColor() const { return m_blend_color; }
void SetBlendColor(const BlendColor& color) { m_blend_color = color; }
[[nodiscard]] const ClipControl& GetClipControl() const { return m_clip_control; }
void SetClipControl(const ClipControl& control) { m_clip_control = control; }
[[nodiscard]] const RenderControl& GetRenderControl() const { return m_render_control; }
void SetRenderControl(const RenderControl& control) { m_render_control = control; }
[[nodiscard]] const DepthControl& GetDepthControl() const { return m_depth_control; }
void SetDepthControl(const DepthControl& control) { m_depth_control = control; }
[[nodiscard]] const ModeControl& GetModeControl() const { return m_mode_control; }
void SetModeControl(const ModeControl& control) { m_mode_control = control; }
[[nodiscard]] const EqaaControl& GetEqaaControl() const { return m_eqaa_control; }
void SetEqaaControl(const EqaaControl& control) { m_eqaa_control = control; }
[[nodiscard]] const StencilControl& GetStencilControl() const { return m_stencil_control; }
void SetStencilControl(const StencilControl& control) { m_stencil_control = control; }
[[nodiscard]] const StencilMask& GetStencilMask() const { return m_stencil_mask; }
void SetStencilMask(const StencilMask& mask) { m_stencil_mask = mask; }
void SetVsUserSgpr(uint32_t id, uint32_t value, UserSgprType type)
{
@@ -460,8 +503,10 @@ public:
[[nodiscard]] const VertexShaderInfo& GetVs() const { return m_vs; }
[[nodiscard]] const ComputeShaderInfo& GetCs() const { return m_cs; }
[[nodiscard]] float GetDepthClearValue() const { return m_depth_clear_value; }
void SetDepthClearValue(float clear_value) { m_depth_clear_value = clear_value; }
[[nodiscard]] float GetDepthClearValue() const { return m_depth_clear_value; }
void SetDepthClearValue(float clear_value) { m_depth_clear_value = clear_value; }
[[nodiscard]] uint8_t GetStencilClearValue() const { return m_stencil_clear_value; }
void SetStencilClearValue(uint8_t clear_value) { m_stencil_clear_value = clear_value; }
private:
BlendControl m_blend_control[8];
@@ -479,7 +524,10 @@ private:
DepthRenderTarget m_depth_render_target;
RenderControl m_render_control;
DepthControl m_depth_control;
float m_depth_clear_value = 0.0f;
StencilControl m_stencil_control;
StencilMask m_stencil_mask;
float m_depth_clear_value = 0.0f;
uint8_t m_stencil_clear_value = 0;
ModeControl m_mode_control;
EqaaControl m_eqaa_control;
@@ -10,6 +10,7 @@
namespace Kyty::Libs::Graphics {
class CommandBuffer;
struct GraphicContext;
struct VulkanMemory;
struct VulkanBuffer;
@@ -39,6 +40,12 @@ enum class GpuMemoryObjectType : uint64_t
Max
};
enum class GpuMemoryScenario
{
Common,
GenerateMips,
};
struct GpuMemoryObject
{
GpuMemoryObjectType type = GpuMemoryObjectType::Invalid;
@@ -50,8 +57,8 @@ class GpuObject
public:
using create_func_t = void* (*)(GraphicContext* ctx, const uint64_t* params, const uint64_t* vaddr, const uint64_t* size, int vaddr_num,
VulkanMemory* mem);
using create_from_objects_func_t = void* (*)(GraphicContext* ctx, const uint64_t* params, const Vector<GpuMemoryObject>& objects,
VulkanMemory* mem);
using create_from_objects_func_t = void* (*)(GraphicContext* ctx, CommandBuffer* buffer, const uint64_t* params,
GpuMemoryScenario scenario, const Vector<GpuMemoryObject>& objects, VulkanMemory* mem);
using write_back_func_t = void (*)(GraphicContext* ctx, void* obj, const uint64_t* vaddr, const uint64_t* size, int vaddr_num);
using delete_func_t = void (*)(GraphicContext* ctx, void* obj, VulkanMemory* mem);
using update_func_t = void (*)(GraphicContext* ctx, const uint64_t* params, void* obj, const uint64_t* vaddr, const uint64_t* size,
@@ -81,16 +88,18 @@ public:
void GpuMemoryInit();
void GpuMemorySetAllocatedRange(uint64_t vaddr, uint64_t size);
void GpuMemoryFree(GraphicContext* ctx, uint64_t vaddr, uint64_t size);
void* GpuMemoryCreateObject(GraphicContext* ctx, uint64_t vaddr, uint64_t size, const GpuObject& info);
void* GpuMemoryCreateObject(GraphicContext* ctx, const uint64_t* vaddr, const uint64_t* size, int vaddr_num, const GpuObject& info);
void GpuMemoryFree(GraphicContext* ctx, uint64_t vaddr, uint64_t size, bool unmap);
void* GpuMemoryCreateObject(GraphicContext* ctx, CommandBuffer* buffer, uint64_t vaddr, uint64_t size, const GpuObject& info);
void* GpuMemoryCreateObject(GraphicContext* ctx, CommandBuffer* buffer, const uint64_t* vaddr, const uint64_t* size, int vaddr_num,
const GpuObject& info);
void GpuMemoryResetHash(GraphicContext* ctx, const uint64_t* vaddr, const uint64_t* size, int vaddr_num, GpuMemoryObjectType type);
void GpuMemoryDbgDump();
void GpuMemoryFlush(GraphicContext* ctx);
void GpuMemoryFrameDone();
void GpuMemoryWriteBack(GraphicContext* ctx);
bool GpuMemoryCheckAccessViolation(uint64_t vaddr, uint64_t size);
Vector<GpuMemoryObject> GpuMemoryFindObjects(uint64_t vaddr, uint64_t size);
Vector<GpuMemoryObject> GpuMemoryFindObjects(uint64_t vaddr, uint64_t size, bool exact);
bool VulkanAllocate(GraphicContext* ctx, VulkanMemory* mem);
void VulkanFree(GraphicContext* ctx, VulkanMemory* mem);
@@ -40,7 +40,7 @@ public:
bool Equal(const uint64_t* other) const override;
[[nodiscard]] create_func_t GetCreateFunc() const override;
[[nodiscard]] create_from_objects_func_t GetCreateFromObjectsFunc() const override { return nullptr; };
[[nodiscard]] create_from_objects_func_t GetCreateFromObjectsFunc() const override;
[[nodiscard]] write_back_func_t GetWriteBackFunc() const override { return nullptr; };
[[nodiscard]] delete_func_t GetDeleteFunc() const override;
[[nodiscard]] update_func_t GetUpdateFunc() const override;
@@ -10,6 +10,9 @@
namespace Kyty::Libs::Graphics {
struct StorageVulkanBuffer;
class CommandBuffer;
class StorageBufferGpuObject: public GpuObject
{
public:
@@ -31,6 +34,8 @@ public:
[[nodiscard]] update_func_t GetUpdateFunc() const override;
};
void StorageBufferSet(CommandBuffer* cmd_buffer, StorageVulkanBuffer* buffer);
} // namespace Kyty::Libs::Graphics
#endif // KYTY_EMU_ENABLED
@@ -21,13 +21,13 @@ public:
static constexpr int PARAM_NEO = 5;
static constexpr int PARAM_SWIZZLE = 6;
StorageTextureObject(uint32_t dfmt, uint32_t nfmt, uint32_t width, uint32_t height, uint32_t pitch, uint32_t levels, uint32_t tile,
bool neo, uint32_t swizzle)
StorageTextureObject(uint32_t dfmt, uint32_t nfmt, uint32_t width, uint32_t height, uint32_t pitch, uint32_t base_level,
uint32_t levels, uint32_t tile, bool neo, uint32_t swizzle)
{
params[PARAM_DFMT_NFMT] = (static_cast<uint64_t>(dfmt) << 32u) | nfmt;
params[PARAM_PITCH] = pitch;
params[PARAM_WIDTH_HEIGHT] = (static_cast<uint64_t>(width) << 32u) | height;
params[PARAM_LEVELS] = levels;
params[PARAM_LEVELS] = (static_cast<uint64_t>(base_level) << 32u) | levels;
params[PARAM_TILE] = tile;
params[PARAM_NEO] = neo ? 1 : 0;
params[PARAM_SWIZZLE] = swizzle;
@@ -21,13 +21,13 @@ public:
static constexpr int PARAM_NEO = 5;
static constexpr int PARAM_SWIZZLE = 6;
TextureObject(uint32_t dfmt, uint32_t nfmt, uint32_t width, uint32_t height, uint32_t pitch, uint32_t levels, uint32_t tile, bool neo,
uint32_t swizzle)
TextureObject(uint32_t dfmt, uint32_t nfmt, uint32_t width, uint32_t height, uint32_t pitch, uint32_t base_level, uint32_t levels,
uint32_t tile, bool neo, uint32_t swizzle)
{
params[PARAM_DFMT_NFMT] = (static_cast<uint64_t>(dfmt) << 32u) | nfmt;
params[PARAM_PITCH] = pitch;
params[PARAM_WIDTH_HEIGHT] = (static_cast<uint64_t>(width) << 32u) | height;
params[PARAM_LEVELS] = levels;
params[PARAM_LEVELS] = (static_cast<uint64_t>(base_level) << 32u) | levels;
params[PARAM_TILE] = tile;
params[PARAM_NEO] = neo ? 1 : 0;
params[PARAM_SWIZZLE] = swizzle;
@@ -111,6 +111,10 @@ constexpr uint32_t DB_HTILE_DATA_BASE = 0x5;
constexpr uint32_t DB_HTILE_DATA_BASE_BASE_256B_SHIFT = 0;
constexpr uint32_t DB_HTILE_DATA_BASE_BASE_256B_MASK = 0xFFFFFFFF;
constexpr uint32_t DB_STENCIL_CLEAR = 0xA;
constexpr uint32_t DB_STENCIL_CLEAR_CLEAR_SHIFT = 0;
constexpr uint32_t DB_STENCIL_CLEAR_CLEAR_MASK = 0xFF;
constexpr uint32_t DB_DEPTH_CLEAR = 0xB;
constexpr uint32_t DB_DEPTH_CLEAR_DEPTH_CLEAR_SHIFT = 0;
constexpr uint32_t DB_DEPTH_CLEAR_DEPTH_CLEAR_MASK = 0xFFFFFFFF;
@@ -181,6 +185,40 @@ constexpr uint32_t CB_BLEND_GREEN = 0x106;
constexpr uint32_t CB_BLEND_BLUE = 0x107;
constexpr uint32_t CB_BLEND_ALPHA = 0x108;
constexpr uint32_t DB_STENCIL_CONTROL = 0x10B;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILFAIL_SHIFT = 0;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILFAIL_MASK = 0xF;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILZPASS_SHIFT = 4;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILZPASS_MASK = 0xF;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILZFAIL_SHIFT = 8;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILZFAIL_MASK = 0xF;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILFAIL_BF_SHIFT = 12;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILFAIL_BF_MASK = 0xF;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILZPASS_BF_SHIFT = 16;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILZPASS_BF_MASK = 0xF;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILZFAIL_BF_SHIFT = 20;
constexpr uint32_t DB_STENCIL_CONTROL_STENCILZFAIL_BF_MASK = 0xF;
constexpr uint32_t DB_STENCILREFMASK = 0x10C;
constexpr uint32_t DB_STENCILREFMASK_STENCILTESTVAL_SHIFT = 0;
constexpr uint32_t DB_STENCILREFMASK_STENCILTESTVAL_MASK = 0xFF;
constexpr uint32_t DB_STENCILREFMASK_STENCILMASK_SHIFT = 8;
constexpr uint32_t DB_STENCILREFMASK_STENCILMASK_MASK = 0xFF;
constexpr uint32_t DB_STENCILREFMASK_STENCILWRITEMASK_SHIFT = 16;
constexpr uint32_t DB_STENCILREFMASK_STENCILWRITEMASK_MASK = 0xFF;
constexpr uint32_t DB_STENCILREFMASK_STENCILOPVAL_SHIFT = 24;
constexpr uint32_t DB_STENCILREFMASK_STENCILOPVAL_MASK = 0xFF;
constexpr uint32_t DB_STENCILREFMASK_BF = 0x10D;
constexpr uint32_t DB_STENCILREFMASK_BF_STENCILTESTVAL_BF_SHIFT = 0;
constexpr uint32_t DB_STENCILREFMASK_BF_STENCILTESTVAL_BF_MASK = 0xFF;
constexpr uint32_t DB_STENCILREFMASK_BF_STENCILMASK_BF_SHIFT = 8;
constexpr uint32_t DB_STENCILREFMASK_BF_STENCILMASK_BF_MASK = 0xFF;
constexpr uint32_t DB_STENCILREFMASK_BF_STENCILWRITEMASK_BF_SHIFT = 16;
constexpr uint32_t DB_STENCILREFMASK_BF_STENCILWRITEMASK_BF_MASK = 0xFF;
constexpr uint32_t DB_STENCILREFMASK_BF_STENCILOPVAL_BF_SHIFT = 24;
constexpr uint32_t DB_STENCILREFMASK_BF_STENCILOPVAL_BF_MASK = 0xFF;
constexpr uint32_t PA_CL_VPORT_XSCALE = 0x10F;
constexpr uint32_t SPI_PS_INPUT_CNTL_0 = 0x191;
@@ -282,6 +320,7 @@ constexpr uint32_t DB_HTILE_SURFACE_DST_OUTSIDE_ZERO_TO_ONE_MASK = 0x1;
constexpr uint32_t VGT_SHADER_STAGES_EN = 0x2D5;
constexpr uint32_t CB_COLOR0_BASE = 0x318;
constexpr uint32_t CB_COLOR0_INFO = 0x31C;
constexpr uint32_t SPI_SHADER_PGM_RSRC1_PS = 0xA;
constexpr uint32_t SPI_SHADER_PGM_RSRC1_PS_VGPRS_SHIFT = 0;
@@ -42,6 +42,7 @@ enum class ShaderInstructionType
Exp,
ImageLoad,
ImageSample,
ImageStore,
ImageStoreMip,
SAddcU32,
SAddI32,
@@ -137,6 +138,7 @@ enum class ShaderInstructionType
VCmpxGeU32,
VCmpxGtU32,
VCmpxNeU32,
VCmpxNeqF32,
VCndmaskB32,
VCosF32,
VCvtF32F16,
@@ -394,18 +396,24 @@ public:
{ return m_instructions.Contains(type, [](auto inst, auto type) { return inst.type == type; }); });
}
[[nodiscard]] bool IsEmbedded() const { return m_embedded; }
void SetEmbedded(bool embedded) { this->m_embedded = embedded; }
[[nodiscard]] uint32_t GetEmbeddedId() const { return m_embedded_id; }
void SetEmbeddedId(uint32_t embedded_id) { m_embedded_id = embedded_id; }
[[nodiscard]] bool IsVsEmbedded() const { return m_vs_embedded; }
void SetVsEmbedded(bool embedded) { this->m_vs_embedded = embedded; }
[[nodiscard]] uint32_t GetVsEmbeddedId() const { return m_vs_embedded_id; }
void SetVsEmbeddedId(uint32_t embedded_id) { m_vs_embedded_id = embedded_id; }
[[nodiscard]] bool IsPsEmbedded() const { return m_ps_embedded; }
void SetPsEmbedded(bool embedded) { this->m_ps_embedded = embedded; }
[[nodiscard]] uint32_t GetPsEmbeddedId() const { return m_ps_embedded_id; }
void SetPsEmbeddedId(uint32_t embedded_id) { m_ps_embedded_id = embedded_id; }
private:
Vector<ShaderInstruction> m_instructions;
Vector<ShaderLabel> m_labels;
ShaderType m_type = ShaderType::Unknown;
Vector<ShaderDebugPrintf> m_debug_printfs;
uint32_t m_embedded_id = 0;
bool m_embedded = false;
uint32_t m_vs_embedded_id = 0;
uint32_t m_ps_embedded_id = 0;
bool m_vs_embedded = false;
bool m_ps_embedded = false;
};
struct ShaderId
@@ -18,6 +18,7 @@ struct ShaderComputeInputInfo;
String SpirvGenerateSource(const ShaderCode& code, const ShaderVertexInputInfo* vs_input_info, const ShaderPixelInputInfo* ps_input_info,
const ShaderComputeInputInfo* cs_input_info);
String SpirvGetEmbeddedVs(uint32_t id);
String SpirvGetEmbeddedPs(uint32_t id);
} // namespace Kyty::Libs::Graphics