[Mesa-dev] [RFC 0/2] nir and ttn support for indirect/arrays
Jason Ekstrand
jason at jlekstrand.net
Tue Apr 7 09:36:20 PDT 2015
On Tue, Apr 7, 2015 at 8:52 AM, Rob Clark <robdclark at gmail.com> wrote:
> From: Rob Clark <robclark at freedesktop.org>
>
> Introduce intrinsics to load/store global vars (since I'm not sure what
> the point is to have global as a var type if there is no way to access
> it, but maybe I'm missing something), and update ttn to generate global
> variables for arrays.
By and large, variables are supposed to be accessed with
nir_intrinsic_load/store_var. We then (optionally) lower to explicit
index+offset intrinsics for shader inputs/outputs to make things
easier on the backends. However, globals and locals are expected to
be lowered to registers not intrinsics. With TGSI, I think they have
an input/output register file, so it was easier for Eric to simply use
the index+offset intrinsics right from the start.
--Jason
> So, for example:
>
> FRAG
> PROPERTY FS_COLOR0_WRITES_ALL_CBUFS 1
> DCL OUT[0], COLOR
> DCL CONST[0..1]
> DCL TEMP[0..2], ARRAY(1), LOCAL
> DCL TEMP[3..4], LOCAL
> DCL ADDR[0]
> IMM[0] FLT32 { 1.0000, 2.0000, 3.0000, 0.0000}
> IMM[1] FLT32 { 4.0000, 5.0000, 6.0000, 7.0000}
> IMM[2] FLT32 { 7.0000, 8.0000, 9.0000, 0.0000}
> 0: MOV TEMP[0], IMM[0].xyzx
> 1: MOV TEMP[1], IMM[1].xyzx
> 2: MOV TEMP[2], IMM[2].xyzx
> 3: UARL ADDR[0].x, CONST[0].xxxx
> 4: FSEQ TEMP[3].xyz, TEMP[ADDR[0].x](1).xyzz, CONST[1].xyzz
> 5: AND TEMP[3].y, TEMP[3].yyyy, TEMP[3].zzzz
> 6: AND TEMP[3].x, TEMP[3].xxxx, TEMP[3].yyyy
> 7: UCMP TEMP[4], TEMP[3].xxxx, IMM[0].wxwx, TEMP[4]
> 8: NOT TEMP[3].x, TEMP[3].xxxx
> 9: UCMP TEMP[4], TEMP[3].xxxx, IMM[0].xwwx, TEMP[4]
> 10: MOV OUT[0], TEMP[4]
> 11: END
>
> becomes:
>
> decl_var uniform vec4[2] uniform_0 (0, 0)
> decl_var shader_out vec4 out_0 (1, 0)
> decl_overload main returning void
>
> impl main {
> block block_0:
> /* preds: */
> vec4 ssa_0 = load_const (0x3f800000 /* 1.000000 */, 0x40000000 /* 2.000000 */, 0x40400000 /* 3.000000 */, 0x00000000 /* 0.000000 */
> vec4 ssa_233 = load_const (0x3f800000 /* 1.000000 */, 0x40000000 /* 2.000000 */, 0x40400000 /* 3.000000 */, 0x3f800000 /* 1.000000 */
> intrinsic store_global (ssa_233) () (0, 1)
> vec4 ssa_234 = load_const (0x40800000 /* 4.000000 */, 0x40a00000 /* 5.000000 */, 0x40c00000 /* 6.000000 */, 0x40800000 /* 4.000000 */
> intrinsic store_global (ssa_234) () (1, 1)
> vec4 ssa_235 = load_const (0x40e00000 /* 7.000000 */, 0x41000000 /* 8.000000 */, 0x41100000 /* 9.000000 */, 0x40e00000 /* 7.000000 */
> intrinsic store_global (ssa_235) () (2, 1)
> vec4 ssa_18 = intrinsic load_uniform () () (0, 1)
> vec4 ssa_25 = intrinsic load_global_indirect (ssa_18) () (0, 1)
> vec4 ssa_27 = intrinsic load_uniform () () (1, 1)
> vec1 ssa_132 = feq ssa_25, ssa_27
> vec1 ssa_133 = feq ssa_25.y, ssa_27.y
> vec1 ssa_134 = feq ssa_25.z, ssa_27.z
> vec1 ssa_77 = iand ssa_133, ssa_134
> vec1 ssa_79 = iand ssa_132, ssa_77
> vec1 ssa_244 = bcsel ssa_79, ssa_0.w, ssa_0
> vec1 ssa_246 = bcsel ssa_79, ssa_0, ssa_0.w
> vec1 ssa_254 = load_const (0x00000000 /* 0.000000 */
> vec1 ssa_255 = load_const (0x3f800000 /* 1.000000 */
> vec4 ssa_230 = vec4 ssa_244, ssa_246, ssa_254, ssa_255
> intrinsic store_output (ssa_230) () (0, 1)
> /* succs: block_1 */
> block block_1:
> }
>
> note, in one of the opt passes the 'decl_var vec4[3] arr_1' is getting lost
> but I haven't debugged that yet
>
> Rob Clark (2):
> nir: add intrinsics for load/store global
> gallium/ttn: add support for temp arrays
>
> src/gallium/auxiliary/nir/tgsi_to_nir.c | 116 ++++++++++++++++++++++++++------
> src/glsl/nir/nir_intrinsics.h | 7 +-
> 2 files changed, 100 insertions(+), 23 deletions(-)
>
> --
> 2.1.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list