This document compares the D3D10/D3D11 device driver interface with Gallium. It is written from the perspective of a developer implementing a D3D10/D3D11 driver as a Gallium state tracker. Note that naming and other cosmetic differences are not noted, since they don't really matter and would severely clutter the document. Gallium/OpenGL terminology is used in preference to D3D terminology. NOTE: this document tries to be complete but most likely isn't fully complete and also not fully correct: please submit patches if you spot anything incorrect Also note that this is specifically for the DirectX 10/11 Windows Vista/7 DDI interfaces. DirectX 9 has both user-mode (for Vista) and kernel mode (pre-Vista) interfaces, but they are significantly different from Gallium due to the presence of a lot of fixed function functionality. The user-visible DirectX 10/11 interfaces are distinct from the kernel DDI, but they match very closely. * Accessing Microsoft documentation See http://msdn.microsoft.com/en-us/library/dd445501.aspx ("D3D11DDI_DEVICEFUNCS") for D3D documentation. Also see http://download.microsoft.com/download/f/2/d/f2d5ee2c-b7ba-4cd0-9686-b6508b5479a1/direct3d10_web.pdf ("The Direct3D 10 System" by David Blythe) for an introduction to Direct3D 10 and the rationale for its design. The Windows Driver Kit contains the actual headers, as well as shader bytecode documentation. To get the headers from Linux, run the following, in a dedicated directory: wget http://download.microsoft.com/download/4/A/2/4A25C7D5-EFBE-4182-B6A9-AE6850409A78/GRMWDK_EN_7600_1.ISO sudo mount -o loop GRMWDK_EN_7600_1.ISO /mnt/tmp cabextract -x /mnt/tmp/wdk/headers_cab001.cab rename 's/^_(.*)_[0-9]*$/$1/' * sudo umount /mnt/tmp d3d10umddi.h contains the DDI interface analyzed in this document: note that it is much easier to read this online on MSDN. d3d{10,11}TokenizedProgramFormat.hpp contains the shader bytecode definitions: this is not available on MSDN. d3d9types.h contains DX9 shader bytecode, and DX9 types d3dumddi.h contains the DirectX 9 DDI interface * Glossary BC1: DXT1 BC2: DXT3 BC3: DXT5 BC5: RGTC BC6H: BPTC float BC7: BPTC CS = compute shader: OpenCL-like shader DS = domain shader: tessellation evaluation shader HS = hull shader: tessellation control shader IA = input assembler: primitive assembly Input layout: vertex elements OM = output merger: blender PS = pixel shader: fragment shader Primitive topology: primitive type Resource: buffer or texture Shader resource (view): sampler view SO = stream out: transform feedback Unordered access view: view supporting random read/write access (usually from compute shaders) * Legend -: features D3D11 has and Gallium lacks +: features Gallium has and D3D11 lacks !: differences between D3D11 and Gallium *: possible improvements to Gallium >: references to comparisons of special enumerations #: comment * Gallium functions with no direct D3D10/D3D11 equivalent clear + Gallium supports clearing both render targets and depth/stencil with a single call draw_range_elements + Gallium supports indexed draw with explicit range fence_signalled fence_finish + D3D10/D3D11 don't appear to support explicit fencing; queries can often substitute though, and flushing is supported set_clip_state + Gallium supports fixed function user clip planes, D3D10/D3D11 only support using the vertex shader for them set_polygon_stipple + Gallium supports polygon stipple surface_fill + Gallium supports subrectangle fills of surfaces, D3D10 only supports full clears of views * DirectX 10/11 DDI functions and Gallium equivalents AbandonCommandList (D3D11 only) - Gallium does not support deferred contexts CalcPrivateBlendStateSize CalcPrivateDepthStencilStateSize CalcPrivateDepthStencilViewSize CalcPrivateElementLayoutSize CalcPrivateGeometryShaderWithStreamOutput CalcPrivateOpenedResourceSize CalcPrivateQuerySize CalcPrivateRasterizerStateSize CalcPrivateRenderTargetViewSize CalcPrivateResourceSize CalcPrivateSamplerSize CalcPrivateShaderResourceViewSize CalcPrivateShaderSize CalcDeferredContextHandleSize (D3D11 only) CalcPrivateCommandListSize (D3D11 only) CalcPrivateDeferredContextSize (D3D11 only) CalcPrivateTessellationShaderSize (D3D11 only) CalcPrivateUnorderedAccessViewSize (D3D11 only) ! D3D11 allocates private objects itself, using the size computed here * Gallium could do something similar to be able to put the private data inline into state tracker objects: this would allow them to fit in the same cacheline and improve performance CheckDeferredContextHandleSizes (D3D11 only) - Gallium does not support deferred contexts CheckFormatSupport -> screen->is_format_supported ! Gallium passes usages to this function, D3D11 returns them - Gallium does not differentiate between blendable and non-blendable render targets - Gallium lacks multisampled-texture and multisampled-render-target usages CheckMultisampleQualityLevels * could merge this with is_format_supported - Gallium lacks multisampling support CommandListExecute (D3D11 only) - Gallium does not support command lists CopyStructureCount (D3D11 only) - Gallium does not support unordered access views (views that can be written to arbitrarily from compute shaders) ClearDepthStencilView -> clear ClearRenderTargetView -> clear # D3D11 is not totally clear about whether this applies to any view or only a "currently-bound view" + Gallium allows to clear both depth/stencil and render target(s) in a single operation + Gallium supports double-precision depth values (but not rgba values!) * May want to also support double-precision rgba or use "float" for "depth" ClearUnorderedAccessViewFloat (D3D11 only) ClearUnorderedAccessViewUint (D3D11 only) - Gallium does not support unordered access views (views that can be written to arbitrarily from compute shaders) CreateBlendState (extended in D3D10.1) -> create_blend_state # D3D10 does not support per-RT blending, only D3D10.1 does - Gallium lacks alpha-to-coverage + Gallium supports logic ops + Gallium supports dithering + Gallium supports using the broadcast alpha component of the blend constant color CreateCommandList (D3D11 only) - Gallium does not support command lists CreateComputeShader (D3D11 only) - Gallium does not support compute shaders CreateDeferredContext (D3D11 only) - Gallium does not support deferred contexts CreateDomainShader (D3D11 only) - Gallium does not support domain shaders CreateHullShader (D3D11 only) - Gallium does not support hull shaders CreateUnorderedAccessView (D3D11 only) - Gallium does not support unordered access views CreateDepthStencilState -> create_depth_stencil_alpha_state ! D3D11 has both a global stencil enable, and front/back enables; Gallium has only front/back enables + Gallium has per-face writemask/valuemasks, D3D11 uses the same value for back and front + Gallium supports the alpha test, which D3D11 lacks CreateDepthStencilView -> get_tex_surface CreateRenderTargetView -> get_tex_surface ! Gallium merges depthstencil and rendertarget views into pipe_surface, which also doubles as a 2D surface abstraction - lack of texture array support - lack of render-to-buffer support + Gallium supports using 3D texture zslices as a depth/stencil buffer (in theory) CreateElementLayout -> create_vertex_elements_state ! D3D11 allows sparse vertex elements (via InputRegister); in Gallium they must be specified sequentially ! D3D11 has an extra flag (InputSlotClass) that is the same as instance_divisor == 0 CreateGeometryShader -> create_gs_state CreateGeometryShaderWithStreamOutput -> create_gs_state CreatePixelShader -> create_fs_state CreateVertexShader -> create_vs_state > bytecode is different (see D3d10tokenizedprogramformat.hpp) ! D3D11 describes input/outputs separately from bytecode; Gallium has the tgsi_scan.c module to extract it from TGSI @ TODO: look into DirectX 10/11 semantics specification and bytecode CheckCounter CheckCounterInfo CreateQuery -> create_query - Gallium only supports occlusion, primitives generated and primitives emitted queries ! D3D11 implements fences with "event" queries * TIMESTAMP could be implemented as an additional fields for other queries: some cards have hardware support for exactly this * OCCLUSIONPREDICATE is required for the OpenGL v2 occlusion query functionality * others are performance counters, we may want them but they are not critical CreateRasterizerState - Gallium lacks clamping of polygon offset depth biases - Gallium lacks support to disable depth clipping - Gallium lacks multisampling + Gallium, like OpenGL, supports PIPE_POLYGON_MODE_POINT + Gallium, like OpenGL, supports per-face polygon fill modes + Gallium, like OpenGL, supports culling everything + Gallium, like OpenGL, supports two-side lighting; D3D11 only has the facing attribute + Gallium, like OpenGL, supports per-fill-mode polygon offset enables + Gallium, like OpenGL, supports polygon smoothing + Gallium, like OpenGL, supports polygon stipple + Gallium, like OpenGL, supports point smoothing + Gallium, like OpenGL, supports point sprites + Gallium supports specifying point quad rasterization + Gallium, like OpenGL, supports per-point point size + Gallium, like OpenGL, supports line smoothing + Gallium, like OpenGL, supports line stipple + Gallium supports line last pixel rule specification + Gallium, like OpenGL, supports provoking vertex convention + Gallium supports D3D9 rasterization rules + Gallium supports fixed line width + Gallium supports fixed point size CreateResource -> texture_create or buffer_create ! D3D11 passes the dimensions of all mipmap levels to the create call, while Gallium has an implicit floor(x/2) rule # Note that hardware often has the implicit rule, so the D3D11 interface seems to make little sense # Also, the D3D11 API does not allow the user to specify mipmap sizes, so this really seems a dubious decision on Microsoft's part - D3D11 supports specifying initial data to write in the resource - Gallium lacks support for stream output buffer usage - Gallium does not support unordered access buffers ! D3D11 specifies mapping flags (i.e. read/write/discard);:it's unclear what they are used for here - D3D11 supports odd things in the D3D10_DDI_RESOURCE_MISC_FLAG enum (D3D10_DDI_RESOURCE_MISC_DISCARD_ON_PRESENT, D3D11_DDI_RESOURCE_MISC_BUFFER_ALLOW_RAW_VIEWS, D3D11_DDI_RESOURCE_MISC_BUFFER_STRUCTURED) - Gallium does not support indirect draw call parameter buffers - Gallium lacks multisampling - Gallium lacks array textures ! D3D11 supports specifying hardware modes and other stuff here for scanout resources + Gallium allows specifying minimum buffer alignment ! D3D11 implements cube maps as 2D array textures CreateSampler - D3D11 supports a monochrome convolution filter for "text filtering" + Gallium supports non-normalized coordinates + Gallium supports CLAMP, MIRROR_CLAMP and MIRROR_CLAMP_TO_BORDER + Gallium supports setting min/max/mip filters and anisotropy independently CreateShaderResourceView (extended in D3D10.1) -> create_sampler_view - Gallium lacks sampler views over buffers - Gallium lacks texture arrays, and cube map views over texture arrays + Gallium supports specifying a swizzle ! D3D11 implements "cube views" as views into a 2D array texture CsSetConstantBuffers (D3D11 only) CsSetSamplers (D3D11 only) CsSetShader (D3D11 only) CsSetShaderResources (D3D11 only) CsSetShaderWithIfaces (D3D11 only) CsSetUnorderedAccessViews (D3D11 only) - Gallium does not support compute shaders DestroyBlendState DestroyCommandList (D3D11 only) DestroyDepthStencilState DestroyDepthStencilView DestroyDevice DestroyElementLayout DestroyQuery DestroyRasterizerState DestroyRenderTargetView DestroyResource DestroySampler DestroyShader DestroyShaderResourceView DestroyUnorderedAccessView (D3D11 only) # these are trivial Dispatch (D3D11 only) - Gallium does not support compute shaders DispatchIndirect (D3D11 only) - Gallium does not support compute shaders Draw -> draw_arrays ! D3D11 sets primitive modes separately with IaSetTopology: it's not obvious which is better DrawAuto - Gallium lacks stream out and DrawAuto DrawIndexed -> draw_elements ! D3D11 sets primitive modes separately with IaSetTopology: it's not obvious which is better * may want to add a separate set_index_buffer - Gallium lacks base vertex for indexed draw calls + D3D11 lacks draw_range_elements functionality, which is required for OpenGL DrawIndexedInstanced -> draw_elements_instanced ! D3D11 sets primitive modes separately with IaSetTopology: it's not obvious which is better * may want to add a separate set_index_buffer - Gallium lacks base vertex for indexed draw calls DrawIndexedInstancedIndirect (D3D11 only) -> call draw_elements_instanced multiple times in software # this allows to use an hardware buffer to specify the parameters for multiple draw_elements_instanced calls - Gallium does not support draw call parameter buffers and indirect draw DrawInstanced -> draw_arrays_instanced ! D3D11 sets primitive modes separately with IaSetTopology: it's not obvious which is better DrawInstancedIndirect (D3D11 only) -> call draw_arrays_instanced multiple times in software # this allows to use an hardware buffer to specify the parameters for multiple draw_arrays_instanced calls - Gallium does not support draw call parameter buffers and indirect draws DsSetConstantBuffers (D3D11 only) DsSetSamplers (D3D11 only) DsSetShader (D3D11 only) DsSetShaderResources (D3D11 only) DsSetShaderWithIfaces (D3D11 only) - Gallium does not support domain shaders Flush -> flush ! Gallium supports fencing and several kinds of flushing here, D3D11 just has a dumb glFlush-like function GenMips - Gallium lacks a mipmap generation interface, and does this manually with the 3D engine * it may be useful to add a mipmap generation interface, since the hardware (especially older cards) may have a better way than using the 3D engine GsSetConstantBuffers -> for(i = StartBuffer; i < NumBuffers; ++i) set_constant_buffer(PIPE_SHADER_GEOMETRY, i, phBuffers[i]) GsSetSamplers - Gallium does not support sampling in geometry shaders GsSetShader -> bind_gs_state GsSetShaderWithIfaces (D3D11 only) - Gallium does not support shader interfaces GsSetShaderResources - Gallium does not support sampling in geometry shaders HsSetConstantBuffers (D3D11 only) HsSetSamplers (D3D11 only) HsSetShader (D3D11 only) HsSetShaderResources (D3D11 only) HsSetShaderWithIfaces (D3D11 only) - Gallium does not support hull shaders IaSetIndexBuffer ! Gallium passes this to the draw_elements or draw_elements_instanced calls + Gallium supports 8-bit indices ! the D3D11 interface allows index-size-unaligned byte offsets into index buffers; it's not clear whether they actually work IaSetInputLayout -> bind_vertex_elements_state IaSetTopology ! Gallium passes the topology = primitive type to the draw calls * may want to add an interface for this - Gallium lacks support for DirectX 11 tessellated primitives + Gallium supports line loops, triangle fans, quads, quad strips and polygons IaSetVertexBuffers -> set_vertex_buffers + Gallium allows to specify a max_index here - Gallium only allows setting all vertex buffers at once, while D3D11 supports setting a subset OpenResource -> texture_from_handle PsSetConstantBuffers -> for(i = StartBuffer; i < NumBuffers; ++i) set_constant_buffer(PIPE_SHADER_FRAGMENT, i, phBuffers[i]) * may want to split into fragment/vertex-specific versions PsSetSamplers -> bind_fragment_sampler_states * may want to allow binding subsets instead of all at once PsSetShader -> bind_fs_state PsSetShaderWithIfaces (D3D11 only) - Gallium does not support shader interfaces PsSetShaderResources -> set_fragment_sampler_views * may want to allow binding subsets instead of all at once QueryBegin -> begin_query QueryEnd -> end_query QueryGetData -> get_query_result - D3D11 supports reading an arbitrary data chunk for query results, Gallium only supports reading a 64-bit integer + D3D11 doesn't seem to support actually waiting for the query result (?!) - D3D11 supports optionally not flushing command buffers here and instead returning DXGI_DDI_ERR_WASSTILLDRAWING RecycleCommandList (D3D11 only) RecycleCreateCommandList (D3D11 only) RecycleDestroyCommandList (D3D11 only) - Gallium does not support command lists RecycleCreateDeferredContext (D3D11 only) - Gallium does not support deferred contexts RelocateDeviceFuncs - Gallium does not support moving pipe_context, while D3D11 seems to, using this ResetPrimitiveID (D3D10.1+ only, #ifdef D3D10PSGP) # used to do vertex processing on the GPU on Intel G45 chipsets when it is faster this way (see www.intel.com/Assets/PDF/whitepaper/322931.pdf) # presumably this resets the primitive id system value - Gallium does not support vertex pipeline bypass anymore ResourceCopy ResourceCopyRegion ResourceConvert (D3D10.1+ only) ResourceConvertRegion (D3D10.1+ only) -> surface_copy - Gallium does not support hardware buffer copies - Gallium does not support copying 3D texture subregions in a single call ResourceIsStagingBusy -> is_texture_referenced, is_buffer_referenced - Gallium does not support checking reference for a whole texture, but only a specific surface ResourceReadAfterWriteHazard ! Gallium specifies hides this, except for the render and texture caches ResourceResolveSubresource - Gallium does not support multisample sample resolution ResourceMap ResourceUnmap DynamicConstantBufferMapDiscard DynamicConstantBufferUnmap DynamicIABufferMapDiscard DynamicIABufferMapNoOverwrite DynamicIABufferUnmap DynamicResourceMapDiscard DynamicResourceUnmap StagingResourceMap StagingResourceUnmap -> buffer_map / buffer_unmap -> transfer functions ! Gallium and D3D have different semantics for transfers * D3D separates vertex/index buffers from constant buffers ! D3D separates some buffer flags into specialized calls ResourceUpdateSubresourceUP -> transfer functionality, transfer_inline_write in gallium-resources DefaultConstantBufferUpdateSubresourceUP -> transfer functionality, transfer_inline_write in gallium-resources SetBlendState -> bind_blend_state and set_blend_color ! D3D11 fuses bind_blend_state and set_blend_color in a single function - Gallium lacks the sample mask SetDepthStencilState -> bind_depth_stencil_alpha_state and set_stencil_ref ! D3D11 fuses bind_depth_stencil_alpha_state and set_stencil_ref in a single function SetPredication -> render_condition # here both D3D11 and Gallium seem very limited (hardware is too, probably though) # ideally, we should support nested conditional rendering, as well as more complex tests (checking for an arbitrary range, after an AND with arbitrary mask ) # of couse, hardware support is probably as limited as OpenGL/D3D11 + Gallium, like NV_conditional_render, supports by-region and wait flags - D3D11 supports predication conditional on being equal any value (along with occlusion predicates); Gallium only supports on non-zero SetRasterizerState -> bind_rasterizer_state SetRenderTargets (extended in D3D11) -> set_framebuffer_state ! Gallium passed a width/height here, D3D11 does not ! Gallium lacks ClearTargets (but this is redundant and the driver can trivially compute this if desired) - Gallium does not support unordered access views - Gallium does not support geometry shader selection of texture array image / 3D texture zslice SetResourceMinLOD (D3D11 only) - Gallium does not support min lod directly on textures SetScissorRects - Gallium lacks support for multiple geometry-shader-selectable scissor rectangles D3D11 has SetTextFilterSize - Gallium lacks support for text filters SetVertexPipelineOutput (D3D10.1+ only) # used to do vertex processing on the GPU on Intel G45 chipsets when it is faster this way (see www.intel.com/Assets/PDF/whitepaper/322931.pdf) - Gallium does not support vertex pipeline bypass anymore SetViewports - Gallium lacks support for multiple geometry-shader-selectable viewports D3D11 has ShaderResourceViewReadAfterWriteHazard -> flush(PIPE_FLUSH_RENDER_CACHE) - Gallium does not support specifying this per-render-target/view SoSetTargets - Gallium does not support stream out VsSetConstantBuffers -> for(i = StartBuffer; i < NumBuffers; ++i) set_constant_buffer(PIPE_SHADER_VERTEX, i, phBuffers[i]) * may want to split into fragment/vertex-specific versions VsSetSamplers -> bind_vertex_sampler_states * may want to allow binding subsets instead of all at once VsSetShader -> bind_vs_state VsSetShaderWithIfaces (D3D11 only) - Gallium does not support shader interfaces VsSetShaderResources -> set_fragment_sampler_views * may want to allow binding subsets instead of all at once