[Mesa-dev] RFC: ctx->Driver.Map/UnmapTextureImage() hooks

[Brian Paul]

On 07/15/2011 10:07 AM, Dave Airlie wrote:
> On Fri, Jul 15, 2011 at 4:10 AM, Brian Paul<brian.e.paul at gmail.com>  wrote:
>> On Thu, Jun 23, 2011 at 7:08 PM, Brian Paul<brianp at vmware.com>  wrote:
>>>
>>> I'd like to overhaul the part of Mesa related to texture memory
>>> reading/writing.
>>>
>>> The basic idea would be to add two new driver functions:
>>>
>>> /**
>>>   * Map a 2D slice of a texture image into user space.
>>>   * (x,y,w,h) defines a region of interest (ROI).  Reading/writing
>>>   * texels outside of the ROI is undefined.
>>>   *
>>>   * \param texObj  the texture object
>>>   * \param level  the mipmap level
>>>   * \param faceSlice  the cube face or 3D/array image slice
>>>   * \param x, y, w, h  region of interest
>>>   * \param mode  bitmask of GL_MAP_READ_BIT, GL_MAP_WRITE_BIT
>>>   * \param mapOut  returns start of mapping of ROI
>>>   * \param rowStrideOut  returns row stride (in bytes)
>>>   */
>>> void (*MapTextureImage)(struct gl_context *ctx,
>>>                         struct gl_texture_object *texObj,
>>>                         GLuint level, GLuint faceSlice,
>>>                         GLuint x, GLuint y, GLuint w, GLuint h,
>>>                         GLbitfield mode,
>>>                         GLubyte **mapOut, GLint *rowStrideOut);
>>>
>>> void (*UnmapTextureImage)(struct gl_context *ctx,
>>>                           struct gl_texture_object *texObj,
>>>                           GLuint level, GLuint faceSlice);
>>>
>>>
>>> glTexImage() would use these when loading texture data.  Similarly when
>>> glGetTexImage() returns texture data.  swrast would also use these
>>> before/after rendering.
>>>
>>> We'd get rid of these fields:
>>>
>>> struct gl_texture_image
>>> {
>>>    ...
>>>    FetchTexelFuncC FetchTexelc;
>>>    FetchTexelFuncF FetchTexelf;
>>>    GLuint RowStride;
>>>    GLuint *ImageOffsets;
>>>    GLvoid *Data;
>>> ...
>>> }
>>>
>>> The texel fetch/store code would get pushed into swrast.
>>>
>>> The next step would be to do the same thing for renderbuffers and get rid of
>>> all the Read/WriteSpan() stuff.
>>>
>>> After that, maybe unify texture images and renderbuffers.  I think I'd like
>>> to do these things step by step to avoid too much disruption.
>>
>>
>> The map-texture-image-v4 branch that I just pushed implements this
>> change.  It really cleaned things up for the better and will lead to a
>> few more follow-on improvements.
>>
>> There's no obvious regressions with swrast or the gallium drivers.  I
>> updated the intel driver code and tested i915 and it seems OK too, but
>> I didn't do a full piglit run on it.  I also updated the legacy r600
>> driver in case anyone cares but didn't do any testing of it.
>>
>> I didn't update any of the other legacy DRI drivers.  Would anyone
>> object if we simply stop building mach64, r128, unichrome, sis, etc?
>> I'd be happy to remove those drivers altogether for that matter.
>
> we could EOL those in 7.11, and if anyone wants to ship them, they can
> just build 7.11 for them.

Sounds good to me.  I think we'd only keep the swrast, intel and maybe 
r300/r600 drivers.  Opinions?

-Brian