427 lines
17 KiB
C
427 lines
17 KiB
C
////////////////////////////////////////////////////////////
|
|
//~ Opaque types
|
|
|
|
Struct(GPU_Arena);
|
|
Struct(GPU_CommandList);
|
|
Struct(GPU_Swapchain);
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ Queue types
|
|
|
|
#define GPU_MultiQueueEnabled !ProfilingIsEnabled
|
|
Enum(GPU_QueueKind)
|
|
{
|
|
#if GPU_MultiQueueEnabled
|
|
GPU_QueueKind_Direct = 0,
|
|
GPU_QueueKind_Compute = 1,
|
|
GPU_QueueKind_Copy = 2,
|
|
GPU_QueueKind_BackgroundCopy = 3,
|
|
GPU_NumQueues = 4
|
|
#else
|
|
GPU_QueueKind_Direct = 0,
|
|
GPU_QueueKind_Compute = 0,
|
|
GPU_QueueKind_Copy = 0,
|
|
GPU_QueueKind_BackgroundCopy = 0,
|
|
GPU_NumQueues = 1
|
|
#endif
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ Format types
|
|
|
|
/* NOTE: Matches DirectX DXGI_FORMAT */
|
|
Enum(GPU_Format)
|
|
{
|
|
GPU_Format_Unknown = 0,
|
|
GPU_Format_R32G32B32A32_Typeless = 1,
|
|
GPU_Format_R32G32B32A32_Float = 2,
|
|
GPU_Format_R32G32B32A32_Uint = 3,
|
|
GPU_Format_R32G32B32A32_Sint = 4,
|
|
GPU_Format_R32G32B32_Typeless = 5,
|
|
GPU_Format_R32G32B32_Float = 6,
|
|
GPU_Format_R32G32B32_Uint = 7,
|
|
GPU_Format_R32G32B32_Sint = 8,
|
|
GPU_Format_R16G16B16A16_Typeless = 9,
|
|
GPU_Format_R16G16B16A16_Float = 10,
|
|
GPU_Format_R16G16B16A16_Unorm = 11,
|
|
GPU_Format_R16G16B16A16_Uint = 12,
|
|
GPU_Format_R16G16B16A16_Snorm = 13,
|
|
GPU_Format_R16G16B16A16_Sint = 14,
|
|
GPU_Format_R32G32_Typeless = 15,
|
|
GPU_Format_R32G32_Float = 16,
|
|
GPU_Format_R32G32_Uint = 17,
|
|
GPU_Format_R32G32_Sint = 18,
|
|
GPU_Format_R32G8X24_Typeless = 19,
|
|
GPU_Format_D32_Float_S8X24_Uint = 20,
|
|
GPU_Format_R32_Float_X8X24_Typeless = 21,
|
|
GPU_Format_X32_Typeless_G8X24_Uint = 22,
|
|
GPU_Format_R10G10B10A2_Typeless = 23,
|
|
GPU_Format_R10G10B10A2_Unorm = 24,
|
|
GPU_Format_R10G10B10A2_Uint = 25,
|
|
GPU_Format_R11G11B10_Float = 26,
|
|
GPU_Format_R8G8B8A8_Typeless = 27,
|
|
GPU_Format_R8G8B8A8_Unorm = 28,
|
|
GPU_Format_R8G8B8A8_Unorm_Srgb = 29,
|
|
GPU_Format_R8G8B8A8_Uint = 30,
|
|
GPU_Format_R8G8B8A8_Snorm = 31,
|
|
GPU_Format_R8G8B8A8_Sint = 32,
|
|
GPU_Format_R16G16_Typeless = 33,
|
|
GPU_Format_R16G16_Float = 34,
|
|
GPU_Format_R16G16_Unorm = 35,
|
|
GPU_Format_R16G16_Uint = 36,
|
|
GPU_Format_R16G16_Snorm = 37,
|
|
GPU_Format_R16G16_Sint = 38,
|
|
GPU_Format_R32_Typeless = 39,
|
|
GPU_Format_D32_Float = 40,
|
|
GPU_Format_R32_Float = 41,
|
|
GPU_Format_R32_Uint = 42,
|
|
GPU_Format_R32_Sint = 43,
|
|
GPU_Format_R24G8_Typeless = 44,
|
|
GPU_Format_D24_Unorm_S8_Uint = 45,
|
|
GPU_Format_R24_Unorm_X8_Typeless = 46,
|
|
GPU_Format_X24_Typeless_G8_Uint = 47,
|
|
GPU_Format_R8G8_Typeless = 48,
|
|
GPU_Format_R8G8_Unorm = 49,
|
|
GPU_Format_R8G8_Uint = 50,
|
|
GPU_Format_R8G8_Snorm = 51,
|
|
GPU_Format_R8G8_Sint = 52,
|
|
GPU_Format_R16_Typeless = 53,
|
|
GPU_Format_R16_Float = 54,
|
|
GPU_Format_D16_Unorm = 55,
|
|
GPU_Format_R16_Unorm = 56,
|
|
GPU_Format_R16_Uint = 57,
|
|
GPU_Format_R16_Snorm = 58,
|
|
GPU_Format_R16_Sint = 59,
|
|
GPU_Format_R8_Typeless = 60,
|
|
GPU_Format_R8_Unorm = 61,
|
|
GPU_Format_R8_Uint = 62,
|
|
GPU_Format_R8_Snorm = 63,
|
|
GPU_Format_R8_Sint = 64,
|
|
GPU_Format_A8_Unorm = 65,
|
|
GPU_Format_R1_Unorm = 66,
|
|
GPU_Format_R9G9B9E5_SharedXP = 67,
|
|
GPU_Format_R8G8_B8G8_Unorm = 68,
|
|
GPU_Format_G8R8_G8B8_Unorm = 69,
|
|
GPU_Format_BC1_Typeless = 70,
|
|
GPU_Format_BC1_Unorm = 71,
|
|
GPU_Format_BC1_Unorm_Srgb = 72,
|
|
GPU_Format_BC2_Typeless = 73,
|
|
GPU_Format_BC2_Unorm = 74,
|
|
GPU_Format_BC2_Unorm_Srgb = 75,
|
|
GPU_Format_BC3_Typeless = 76,
|
|
GPU_Format_BC3_Unorm = 77,
|
|
GPU_Format_BC3_Unorm_Srgb = 78,
|
|
GPU_Format_BC4_Typeless = 79,
|
|
GPU_Format_BC4_Unorm = 80,
|
|
GPU_Format_BC4_Snorm = 81,
|
|
GPU_Format_BC5_Typeless = 82,
|
|
GPU_Format_BC5_Unorm = 83,
|
|
GPU_Format_BC5_Snorm = 84,
|
|
GPU_Format_B5G6R5_Unorm = 85,
|
|
GPU_Format_B5G5R5A1_Unorm = 86,
|
|
GPU_Format_B8G8R8A8_Unorm = 87,
|
|
GPU_Format_B8G8R8X8_Unorm = 88,
|
|
GPU_Format_R10G10B10_XR_BIAS_A2_Unorm = 89,
|
|
GPU_Format_B8G8R8A8_Typeless = 90,
|
|
GPU_Format_B8G8R8A8_Unorm_Srgb = 91,
|
|
GPU_Format_B8G8R8X8_Typeless = 92,
|
|
GPU_Format_B8G8R8X8_Unorm_Srgb = 93,
|
|
GPU_Format_BC6H_Typeless = 94,
|
|
GPU_Format_BC6H_UF16 = 95,
|
|
GPU_Format_BC6H_SF16 = 96,
|
|
GPU_Format_BC7_Typeless = 97,
|
|
GPU_Format_BC7_Unorm = 98,
|
|
GPU_Format_BC7_Unorm_Srgb = 99,
|
|
GPU_Format_AYUV = 100,
|
|
GPU_Format_Y410 = 101,
|
|
GPU_Format_Y416 = 102,
|
|
GPU_Format_NV12 = 103,
|
|
GPU_Format_P010 = 104,
|
|
GPU_Format_P016 = 105,
|
|
GPU_Format_420_Opaque = 106,
|
|
GPU_Format_YUY2 = 107,
|
|
GPU_Format_Y210 = 108,
|
|
GPU_Format_Y216 = 109,
|
|
GPU_Format_NV11 = 110,
|
|
GPU_Format_AI44 = 111,
|
|
GPU_Format_IA44 = 112,
|
|
GPU_Format_P8 = 113,
|
|
GPU_Format_A8P8 = 114,
|
|
GPU_Format_B4G4R4A4_Unorm = 115,
|
|
GPU_Format_P208 = 130,
|
|
GPU_Format_V208 = 131,
|
|
GPU_Format_V408 = 132,
|
|
GPU_Format_SamplerFeedbackMinMipOpaque = 189,
|
|
GPU_Format_SamplerFeedbackMipRegionUsedOpaque = 190,
|
|
GPU_Format_A4B4G4R4_Unorm = 191,
|
|
GPU_Format_Count = 192
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ Shader access types
|
|
|
|
Enum(GPU_ShaderAccessKind)
|
|
{
|
|
GPU_ShaderAccessKind_Readonly, /* Default state for all resources */
|
|
GPU_ShaderAccessKind_ReadWrite,
|
|
GPU_ShaderAccessKind_RasterTarget,
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ Arena types
|
|
|
|
Struct(GPU_TempArena)
|
|
{
|
|
GPU_Arena *arena;
|
|
u64 start_pos;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ Buffer types
|
|
|
|
Enum(GPU_BufferFlag)
|
|
{
|
|
GPU_BufferFlag_None = 0,
|
|
GPU_BufferFlag_Writable = (1 << 0),
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ Texture types
|
|
|
|
#define GPU_MaxRasterTargets 8
|
|
|
|
Enum(GPU_TextureFlag)
|
|
{
|
|
GPU_TextureFlag_None = 0,
|
|
GPU_TextureFlag_Writable = (1 << 0),
|
|
GPU_TextureFlag_Rasterizable = (1 << 1),
|
|
};
|
|
|
|
Enum(GPU_TextureKind)
|
|
{
|
|
GPU_TextureKind_1D,
|
|
GPU_TextureKind_2D,
|
|
GPU_TextureKind_3D,
|
|
};
|
|
|
|
Struct(GPU_TextureDesc)
|
|
{
|
|
GPU_TextureFlag flags;
|
|
GPU_Format format;
|
|
Vec3I32 size;
|
|
Vec4 clear_color;
|
|
i32 mip_levels;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ Sampler types
|
|
|
|
/* NOTE: Matches DirectX D3D12_FILTER */
|
|
Enum(GPU_Filter)
|
|
{
|
|
/* Standard filter */
|
|
GPU_Filter_MinMagMipPoint = 0,
|
|
GPU_Filter_MinMagPointMipLinear = 0x1,
|
|
GPU_Filter_MinPointMagLinearMipPoint = 0x4,
|
|
GPU_Filter_MinPointMagMipLinear = 0x5,
|
|
GPU_Filter_MinLinearMagMipPoint = 0x10,
|
|
GPU_Filter_MinLinearMagPointMipLinear = 0x11,
|
|
GPU_Filter_MinMagLinearMipPoint = 0x14,
|
|
GPU_Filter_MinMagMipLinear = 0x15,
|
|
GPU_Filter_MinMagAnisotropicMipPoint = 0x54,
|
|
GPU_Filter_Anisotropic = 0x55,
|
|
|
|
/* Comparison filter */
|
|
GPU_Filter_Comparison_MinMagMipPoint = 0x80,
|
|
GPU_Filter_Comparison_MinMagPointMipLinear = 0x81,
|
|
GPU_Filter_Comparison_MinPointMagLinearMipPoint = 0x84,
|
|
GPU_Filter_Comparison_MinPointMagMipLinear = 0x85,
|
|
GPU_Filter_Comparison_MinLinearMagMipPoint = 0x90,
|
|
GPU_Filter_Comparison_MinLinearMagPointMipLinear = 0x91,
|
|
GPU_Filter_Comparison_MinMagLinearMipPoint = 0x94,
|
|
GPU_Filter_Comparison_MinMagMipLinear = 0x95,
|
|
GPU_Filter_Comparison_MinMagAnisotropicMipPoint = 0xd4,
|
|
GPU_Filter_Comparison_Anisotropic = 0xd5,
|
|
|
|
/* Minimum filter */
|
|
GPU_Filter_Minimum_MinMagMipPoint = 0x100,
|
|
GPU_Filter_Minimum_MinMagPointMipLinear = 0x101,
|
|
GPU_Filter_Minimum_MinPointMagLinearMipPoint = 0x104,
|
|
GPU_Filter_Minimum_MinPointMagMipLinear = 0x105,
|
|
GPU_Filter_Minimum_MinLinearMagMipPoint = 0x110,
|
|
GPU_Filter_Minimum_MinLinearMagPointMipLinear = 0x111,
|
|
GPU_Filter_Minimum_MinMagLinearMipPoint = 0x114,
|
|
GPU_Filter_Minimum_MinMagMipLinear = 0x115,
|
|
GPU_Filter_Minimum_MinMagAnisotropicMipPoint = 0x155,
|
|
GPU_Filter_Minimum_Anisotropic = 0x155,
|
|
|
|
/* Maximum filter */
|
|
GPU_Filter_Maximum_MinMagMipPoint = 0x180,
|
|
GPU_Filter_Maximum_MinMagPointMipLinear = 0x181,
|
|
GPU_Filter_Maximum_MinPointMagLinearMipPoint = 0x184,
|
|
GPU_Filter_Maximum_MinPointMagMipLinear = 0x185,
|
|
GPU_Filter_Maximum_MinLinearMagMipPoint = 0x190,
|
|
GPU_Filter_Maximum_MinLinearMagPointMipLinear = 0x191,
|
|
GPU_Filter_Maximum_MinMagLinearMipPoint = 0x194,
|
|
GPU_Filter_Maximum_MinMagMipLinear = 0x195,
|
|
GPU_Filter_Maximum_MinMagAnisotropicMipPoint = 0x1d4,
|
|
GPU_Filter_Maximum_Anisotropic = 0x1d5
|
|
};
|
|
|
|
/* NOTE: Matches DirectX D3D12_TEXTURE_ADDRESS_MODE */
|
|
Enum(GPU_AddressMode)
|
|
{
|
|
GPU_AddressMode_Wrap = 1,
|
|
GPU_AddressMode_Mirror = 2,
|
|
GPU_AddressMode_Clamp = 3, /* Default */
|
|
GPU_AddressMode_Border = 4,
|
|
GPU_AddressMode_MirrorOnce = 5
|
|
};
|
|
|
|
/* NOTE: Matches DirectX D3D12_COMPARISON_FUNC */
|
|
Enum(GPU_ComparisonFunc)
|
|
{
|
|
GPU_ComparisonFunc_None = 0,
|
|
GPU_ComparisonFunc_Never = 1,
|
|
GPU_ComparisonFunc_Less = 2,
|
|
GPU_ComparisonFunc_Equal = 3,
|
|
GPU_ComparisonFunc_LessEqual = 4,
|
|
GPU_ComparisonFunc_Greater = 5,
|
|
GPU_ComparisonFunc_NotEqual = 6,
|
|
GPU_ComparisonFunc_GreaterEqual = 7,
|
|
GPU_ComparisonFunc_Always = 8
|
|
};
|
|
|
|
Struct(GPU_SamplerDesc)
|
|
{
|
|
GPU_Filter filter;
|
|
GPU_AddressMode x;
|
|
GPU_AddressMode y;
|
|
GPU_AddressMode z;
|
|
f32 mip_lod_bias;
|
|
u32 max_anisotropy;
|
|
GPU_ComparisonFunc comparison;
|
|
Vec4 border_color;
|
|
f32 min_lod;
|
|
f32 max_lod;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ Rasterization types
|
|
|
|
Enum(GPU_RasterMode)
|
|
{
|
|
GPU_RasterMode_None,
|
|
GPU_RasterMode_PointList,
|
|
GPU_RasterMode_LineList,
|
|
GPU_RasterMode_LineStrip,
|
|
GPU_RasterMode_TriangleList,
|
|
GPU_RasterMode_WireTriangleList,
|
|
GPU_RasterMode_TriangleStrip,
|
|
GPU_RasterMode_WireTriangleStrip,
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ Statistic types
|
|
|
|
Struct(GPU_Stats)
|
|
{
|
|
/* Memory usage */
|
|
u64 local_committed;
|
|
u64 local_budget;
|
|
u64 non_local_committed;
|
|
u64 non_local_budget;
|
|
|
|
/* Resources */
|
|
u64 driver_resources_allocated;
|
|
u64 driver_descriptors_allocated;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ @hookdecl Startup
|
|
|
|
void GPU_Startup(void);
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ @hookdecl Arenas
|
|
|
|
GPU_Arena *GPU_AcquireArena(void);
|
|
void GPU_ReleaseArena(GPU_Arena *arena);
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ @hookdecl Resource creation
|
|
|
|
GpuPointer GPU_PushBufferEx(GPU_Arena *arena, i32 element_size, i32 element_align, i32 element_count, GPU_BufferFlag flags);
|
|
#define GPU_PushBuffer(arena, type, count, flags) GPU_PushBufferEx((arena), sizeof(type), alignof(type), (count), (flags))
|
|
|
|
GpuPointer GPU_PushTextureEx(GPU_Arena *arena, GPU_TextureDesc desc);
|
|
GpuPointer GPU_PushTexture(GPU_Arena *arena, GPU_TextureKind kind, GPU_Format format, Vec3I32 size, GPU_TextureFlag flags);
|
|
|
|
GpuPointer GPU_PushSampler(GPU_Arena *arena, GPU_SamplerDesc desc);
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ @hookdecl Commands
|
|
|
|
//- Command list creation
|
|
GPU_CommandList *GPU_OpenCommandList(GPU_QueueKind queue);
|
|
void GPU_CloseCommandList(GPU_CommandList *cl);
|
|
|
|
//- Cpu -> Gpu
|
|
void GPU_CopyBytesFromCpu(GPU_CommandList *cl, GpuPointer dst, RngU64 dst_range, void *src);
|
|
void GPU_CopyTexelsFromCpu(GPU_CommandList *cl, GpuPointer dst, Rng3U64 dst_range, void *src);
|
|
void GPU_CopyFromCpu(GPU_CommandList *cl, GpuPointer dst, String src);
|
|
|
|
//- Gpu -> Cpu
|
|
void GPU_AddCpuFence(GPU_CommandList *cl, Fence *fence, i64 v);
|
|
void GPU_SetCpuFence(GPU_CommandList *cl, Fence *fence, i64 v);
|
|
|
|
//- Implicit state
|
|
void GPU_SetShaderAccess(GPU_CommandList *cl, GpuPointer ptr, GPU_ShaderAccessKind access_kind);
|
|
void GPU_SetRasterizeMode(GPU_CommandList *cl, GPU_RasterMode mode);
|
|
void GPU_SetConstantU32(GPU_CommandList *cl, i32 slot, u32 v);
|
|
void GPU_SetConstantF32(GPU_CommandList *cl, i32 slot, f32 v);
|
|
void GPU_SetConstantPtr(GPU_CommandList *cl, i32 slot, GpuPointer v);
|
|
|
|
//- Clear
|
|
void GPU_ClearRasterTarget(GPU_CommandList *cl, GpuPointer target);
|
|
|
|
//- Compute
|
|
void GPU_Compute(GPU_CommandList *cl, ComputeShader cs, Vec3U32 threads);
|
|
|
|
//- Rasterize
|
|
void GPU_RasterizeEx(GPU_CommandList *cl,
|
|
VertexShader vs, PixelShader ps,
|
|
u32 instances_count,
|
|
GpuPointer idx_buff, RngU64 idx_buff_range,
|
|
u32 raster_targets_count, GpuPointer *raster_targets,
|
|
Rng3 viewport, Rng2 scissor);
|
|
|
|
void GPU_Rasterize(GPU_CommandList *cl,
|
|
VertexShader vs, PixelShader ps,
|
|
u32 instances_count, GpuPointer idx_buff,
|
|
u32 raster_targets_count, GpuPointer *raster_targets);
|
|
|
|
//- Profiling
|
|
void GPU_ProfN(GPU_CommandList *cl, String name);
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ @hookdecl Statistics
|
|
|
|
GPU_Stats GPU_QueryStats(void);
|
|
|
|
////////////////////////////////////////////////////////////
|
|
//~ @hookdecl Swapchain
|
|
|
|
GPU_Swapchain *GPU_AcquireSwapchain(WND_Handle window, GPU_Format format, Vec2I32 size);
|
|
void GPU_ReleaseSwapchain(GPU_Swapchain *swapchain);
|
|
|
|
/* Waits until a new backbuffer is ready to be written to.
|
|
* This should be called before rendering for minimum latency. */
|
|
void GPU_YieldOnSwapchain(GPU_Swapchain *swapchain);
|
|
|
|
/* 1. Recreates backbuffer at desired size if necessary
|
|
* 2. Blits the source texture into the backbuffer
|
|
* 3. Presents the backbuffer
|
|
*/
|
|
void GPU_PresentSwapchain(GPU_Swapchain *swapchain, Vec2I32 dst_size, Rng2I32 dst_range, GpuPointer src, Vec2I32 src_pos, i32 vsync, Vec4 clear_color);
|