Add a "max complexity" heuristic to allow unrolling long loops with small
bodies and short loops with large bodies.
The loop unroll limits may need further tweaking...
Loops such as this will be unrolled:
for (i = 0; i < 4; ++i) {
body;
}
where 'body' isn't too large.
This also helps to fix the issue reported in bug #19190. The problem there
is indexing vector types with a variable index. For example:
vec4 v;
v[2] = 1.0; // equivalent to v.z = 1.0
v[i] = 2.0; // variable index into vector!!
Since the for-i loop can be unrolled, we can avoid the problems associated
with variable indexing into a vector (at least in this case).
This fixes cases such as:
vec4 v4;
vec2 v2;
v4.xz.yx = v2;
The last line now correctly compiles into MOV TEMP[1].xz, TEMP[0].yyxw;
Helps to fix the Humus Domino demo. See bug 19189.
This adds all of the `mglu' symbols to the list of symbol exports
for GLU. Without this patch, mangled GLU symbols are considered
`internal' symbols, and calling any results in undefined references.
Now only the samplers that are actually used by texture() functions are
saved in the uniform variable list. Before, we could run out of samplers
if too many were declared while only some of them were actually used.
The max texture coord units is still 8. All the fixed-function paths are
still limited to 8 too. But GLSL shaders can use more samplers now.
Note that some texcoord-related data structures are declared to be 16
elements in size rather than 8. This just simplifies the code in a few
places; the extra elements aren't accessible to the user.
These changes haven't been extensively tested yet, but sanity checking has
been done.
It should be possible to increase the max image units/samplers to 32 without
doing anything special. Beyond that we'll need longer bitfields in a few
places.
This lets us avoid software fallbacks when clients forget to turn some state
off (engine demo) or just do crazy things to test conformance (OGLC).
This should probably be brought into mesa generic code so other drivers can
make use of it.
Bug #19016.
This allows code such as "vec4 a = gl_LightSource[i].ambient;" to work.
When a built-in uniform array is indexed with a variable index we need to
"unroll" the whole array into the parameter list (aka constant buffer) because
we don't know which elements may be accessed at compile-time. In the case of
the gl_LightSource array of size [8], we emit 64 state references into the
parameter array (8 elements times 8 vec4s per gl_LightSourceParameters
struct).
Previously, we only allowed constant-indexed references to uniform arrays
(such as gl_LightSource[2].position) which resulted in a single state reference
being added to the parameter array, not 64. We still optimize this case.
Users should be aware that using "gl_LightSource[i].ambient" in their shaders
is a bit expensive since state validation will involve updating all 64
light source entries in the parameter list.
We'd come up with a negative remainder, while we were looking for the positive
version of it in the loop conditional. And, since the "did we hit our target"
break was disabled for the target_msc == 0 ("Just make the divisor/remainder
work") path, we'd never exit.
Simplify the code by just using int64_t all over instead of trying to do it
in a u32 space.
Previously fog parameter and specular color are packed into the
same dword. Note specular color should be packed in BGRA for device,
so if fog parameter and specular color all are present, fog parameter
will dirty the alpha term of specular color. This fixes rendering
issue when playing 'Yo Frankie' on 915/945.
Previously 3D textures were mipmapped using multiple passed through
the 2D mipmap generation code. This had 3 disadvantages. First, the
extra passes were slow. Second, this required the allocation of a
temporary buffer to hold intermediate data. Third, and most
important, the extra passes caused loss of additional bits due to
integer division / bit-shifting.
With this change, our mipmapgen conformance test passes for
non-compressed texture formats.
