| Andrew Bonventre | 3c7fee3 | 2015-09-22 18:14:18 | [diff] [blame] | 1 | # Vertex Texture Fetch |
| 2 | |
| 3 | This page details the steps necessary to implement vertex texture fetch in ANGLE |
| 4 | and documents some of the pitfalls that may be encountered along the way. |
| 5 | |
| 6 | # Details |
| 7 | |
| 8 | Tasks to implement vertex texture support. |
| 9 | |
| 10 | 1. add/enable vertex shader texture look up functions in compiler & HLSL |
| 11 | translator. |
| 12 | * add texture2DLod, texture2DProjLod (2 variants), textureCubeLod (these |
| 13 | are **only** valid in vertex shaders) |
| 14 | * ensure other (non-bias/non-LOD) texture functions work in vertex shaders |
| 15 | * non-mipmapped textures use the only level available |
| 16 | * mipmapped textures use only the base level (ie level 0). |
| Frank Henigman | c31b741 | 2016-12-09 02:57:19 | [diff] [blame] | 17 | 2. update implementation-dependent constants in Context.h |
| Andrew Bonventre | 3c7fee3 | 2015-09-22 18:14:18 | [diff] [blame] | 18 | * MAX\_VERTEX\_TEXTURE\_IMAGE\_UNITS = 4 |
| 19 | * MAX\_COMBINED\_TEXTURE\_IMAGE\_UNITS = |
| 20 | MAX\_VERTEX\_TEXTURE\_IMAGE\_UNITS + MAX\_TEXTURE\_IMAGE\_UNITS (ie 20). |
| 21 | * these limits have to change based on the d3d device characteristics. For |
| 22 | example we likely don't want to advertise vertex image units on SM2.0 |
| 23 | cards (unless we end up using software vertex processing). |
| 24 | * detection of hardware support for various formats, types, etc. |
| 25 | * As a first pass, use the "hasVertexTextures" check that Aras suggested |
| 26 | to only enable VTF on DX10 NVIDIA and AMD parts, and SM3 Intel parts. |
| 27 | * If this proves insufficient, there are other things we can do, but it |
| 28 | involves using software vertex processing for unsupported formats and |
| 29 | system memory copies of textures -- all stuff which is rather annoying |
| 30 | and likely to hurt performance (see point 4. below). |
| 31 | 3. add support and handling for vertex textures/samplers in the API. |
| 32 | * any textures used in a vertex shader need to get assigned to the special |
| 33 | samplers in d3d9 |
| 34 | * there are only 4 of them (D3DVERTEXTEXTURESAMPLER0.. |
| 35 | D3DVERTEXTEXTURESAMPLER3) |
| 36 | * if a texture is used in both vertex & fragment it counts twice against |
| 37 | the "MAX\_COMBINED" limit (validated in Program::validateSamplers) |
| 38 | * there are a number of places in our code where we have arrays of size, |
| 39 | or iterate over, MAX\_TEXTURE\_IMAGE\_UNITS. These will need to be |
| 40 | changed to operate on MAX\_COMBINED\_TEXTURE\_IMAGE\_UNITS instead. A |
| 41 | (possibly incomplete & outdated) list of areas that need to be updated |
| 42 | is as follows: |
| 43 | * Program.h - increase size of mSamplers |
| 44 | * Context.h - increase size of samplerTexture |
| 45 | * glActiveTexture needs accept values in the range |
| 46 | 0..MAX\_COMBINED\_TEXTURE\_IMAGE\_UNITS-1 |
| 47 | * Context::~Context |
| 48 | * GetIntegerv (2D\_BINDING, CUBE\_BINDING) |
| 49 | * Context::applyTextures |
| 50 | * Context::detachTexture |
| 51 | * Program::getSamplerMapping |
| 52 | * Program::dirtyAllSamplers |
| 53 | * Program::applyUniform1iv |
| 54 | * Program::unlink |
| 55 | * Program::validateSamplers |
| 56 | 4. handling the nasty corner cases: texture formats, filtering and cube |
| 57 | textures. |
| 58 | * OpenGL doesn't provide any restrictions on what formats and/or types of |
| 59 | textures can used for vertex textures, or if filtering can be enabled, |
| 60 | whereas D3D9 does. |
| 61 | * Reference Rasterizer / Software Vertex Processing: all formats & types |
| 62 | supported (including filtering) |
| 63 | * ATI R500 (on Google Code) cards do not support VTF (even though they are |
| 64 | SM 3.0) |
| 65 | * ATI R600 (on Google Code) (and later) and in theory the Intel 965+, |
| 66 | claim to support all texture formats/types we care about and some with |
| 67 | filtering |
| 68 | * NVIDIA cards fall into two camps: |
| 69 | * dx9 SM3 (6&7 series): only R32F & A32B32G32R32F supported for 2D and no |
| 70 | filtering, CUBE or VOL texture support |
| 71 | * dx10 (8+ series): only float texture formats for 2D, CUBE & VOLUME. no |
| 72 | filtering (according to caps) |
| 73 | * further info from Aras P. suggests that all formats are supported on |
| 74 | DX10 hardware, but are just not advertised. |
| 75 | * unsure what they do on these cards under OpenGL. Need to do more |
| 76 | testing, but suspect software fallback. |