[AIE2] Combine 2D/3D post-increments for 512-bit load/stores

[gbossu]

QoR results are mixed, but I believe this is a change in the right direction as this is simplifying the MIR. Those combines are also needed if we want to post-pipeline benchmarks.

| Core_Compute_Cycle_Count      | Mul2D_0      | Mul2D_1      | Conv2D_2x8_0 | Mul2d_bf16_0 | ReduceSumAxis_7_aie2_int8 | ReduceMeanAxis_7_aie2_int8 | Mul2d_bf16_1 | Add2D_bf16_1 | Add2D_bf16_0 | ReduceSumAxis_6_aie2_int8 | ReduceSumAxis_3_aie2_int8 | ReduceSumAxis_5_aie2_int8 | ReduceMeanAxis_6_aie2_int8 | ReduceMeanAxis_3_aie2_int8 | ReduceMeanAxis_5_aie2_int8 | InstanceNormPart1_aie2_bf16_0 | LayerNormC8Part1_aie2_bf16_0 | ReduceSumAxis_7_aie2_bf16 | InstanceNormPart2_aie2_bf16_0 | ReduceMeanAxis_7_aie2_bf16 | InstanceNormPart1_aie2_int8_0 | ReduceSumAxis_6_aie2_bf16 | ReduceSumAxis_3_aie2_bf16 | ReduceSumAxis_5_aie2_bf16 | ReduceMeanAxis_5_aie2_bf16 | ReduceMeanAxis_6_aie2_bf16 | ReduceProdAxis_6_aie2_bf16 | ReduceProdAxis_5_aie2_bf16 | ReduceProdAxis_3_aie2_bf16 | LayerNormC8Part2_aie2_bf16_0 | ReduceProdAxis_7_aie2_bf16 | ReduceSumAxis_4_aie2_int8 | ReduceSumAxis_1_aie2_int8 | ReduceMeanAxis_1_aie2_int8 | ReduceSumAxis_2_aie2_int8 | ReduceMeanAxis_4_aie2_int8 | Mish_aie2_bfloat16 | ReduceMeanAxis_2_aie2_int8 | LayerNormC8Part1_aie2_int8_0 | ReduceSumAxis_1_aie2_bf16 | ReduceSumAxis_2_aie2_bf16 | ReduceSumAxis_4_aie2_bf16 | ReduceMeanAxis_1_aie2_bf16 | ReduceMeanAxis_4_aie2_bf16 | ReduceMeanAxis_2_aie2_bf16 | Conv2D_bf16_0 | ReduceProdAxis_2_aie2_bf16 | ReduceProdAxis_1_aie2_bf16 | ReduceProdAxis_4_aie2_bf16 | Conv2D_bf16_1 | Conv2D_Transpose_AIE2_0 | Conv2D_2x8_1 |    ...    | Mish_aie2_int8 | BatchNorm2D_0 | Conv2D_ReLU_1 | Conv2D_FC_0  | DilatedConv2D_1 | Conv2D_FC_1  | Conv2D_0     | Conv2D_mixed_batch_1 | BilinearInterpolation_1 | GEMM_int8_0  | GEMM_int8_1  | BilinearInterpolation_0 | Average diff |
| ----------------------------- | ------------ | ------------ | ------------ | ------------ | ------------------------- | -------------------------- | ------------ | ------------ | ------------ | ------------------------- | ------------------------- | ------------------------- | -------------------------- | -------------------------- | -------------------------- | ----------------------------- | ---------------------------- | ------------------------- | ----------------------------- | -------------------------- | ----------------------------- | ------------------------- | ------------------------- | ------------------------- | -------------------------- | -------------------------- | -------------------------- | -------------------------- | -------------------------- | ---------------------------- | -------------------------- | ------------------------- | ------------------------- | -------------------------- | ------------------------- | -------------------------- | ------------------ | -------------------------- | ---------------------------- | ------------------------- | ------------------------- | ------------------------- | -------------------------- | -------------------------- | -------------------------- | ------------- | -------------------------- | -------------------------- | -------------------------- | ------------- | ----------------------- | ------------ |    ...    | -------------- | ------------- | ------------- | ------------ | --------------- | ------------ | ------------ | -------------------- | ----------------------- | ------------ | ------------ | ----------------------- | ------------ |
| Baseline                      | 451          | 451          | 1684         | 416          | 2071                      | 2111                       | 272          | 311          | 257          | 2855                      | 2879                      | 2882                      | 2928                       | 2932                       | 2949                       | 2838                          | 8780                         | 6686                      | 13040                         | 6737                       | 11108                         | 7450                      | 7464                      | 7467                      | 7613                       | 7620                       | 10509                      | 10526                      | 10527                      | 11710                        | 2252                       | 7211                      | 7149                      | 7438                       | 7190                      | 7465                       | 5661               | 7500                       | 7712                         | 12349                     | 12372                     | 12387                     | 13438                      | 13444                      | 13474                      | 25071         | 18747                      | 37320                      | 37349                      | 39569         | 64836                   | 3811         |    ...    | 9418           | 386           | 27951         | 2692         | 5510            | 1177         | 7967         | 22344                | 378                     | 3019         | 36178        | 724                     |              |
| This PR                       | 496          | 496          | 1819         | 448          | 2215                      | 2255                       | 288          | 329          | 269          | 2945                      | 2969                      | 2972                      | 3018                       | 3022                       | 3039                       | 2906                          | 8952                         | 6814                      | 13288                         | 6865                       | 11303                         | 7578                      | 7592                      | 7595                      | 7743                       | 7750                       | 10677                      | 10694                      | 10695                      | 11894                        | 2284                       | 7302                      | 7235                      | 7527                       | 7275                      | 7553                       | 5726               | 7586                       | 7798                         | 12477                     | 12500                     | 12515                     | 13566                      | 13572                      | 13602                      | 25197         | 18835                      | 37488                      | 37517                      | 39655         | 64976                   | 3819         |    ...    | 9417           | 385           | 27503         | 2644         | 5382            | 1140         | 7687         | 21504                | 361                     | 2815         | 33499        | 667                     | +0.25%       |
| Total diff                    | REGR(+9.98%) | REGR(+9.98%) | REGR(+8.02%) | REGR(+7.69%) | REGR(+6.95%)              | REGR(+6.82%)               | REGR(+5.88%) | REGR(+5.79%) | REGR(+4.67%) | REGR(+3.15%)              | REGR(+3.13%)              | REGR(+3.12%)              | REGR(+3.07%)               | REGR(+3.07%)               | REGR(+3.05%)               | REGR(+2.40%)                  | REGR(+1.96%)                 | REGR(+1.91%)              | REGR(+1.90%)                  | REGR(+1.90%)               | REGR(+1.76%)                  | REGR(+1.72%)              | REGR(+1.71%)              | REGR(+1.71%)              | REGR(+1.71%)               | REGR(+1.71%)               | REGR(+1.60%)               | REGR(+1.60%)               | REGR(+1.60%)               | REGR(+1.57%)                 | REGR(+1.42%)               | REGR(+1.26%)              | REGR(+1.20%)              | REGR(+1.20%)               | REGR(+1.18%)              | REGR(+1.18%)               | REGR(+1.15%)       | REGR(+1.15%)               | REGR(+1.12%)                 | REGR(+1.04%)              | REGR(+1.03%)              | REGR(+1.03%)              | REGR(+0.95%)               | REGR(+0.95%)               | REGR(+0.95%)               | REGR(+0.50%)  | REGR(+0.47%)               | REGR(+0.45%)               | REGR(+0.45%)               | REGR(+0.22%)  | REGR(+0.22%)            | REGR(+0.21%) |    ...    | SAME(-0.01%)   | IMPR(-0.26%)  | IMPR(-1.60%)  | IMPR(-1.78%) | IMPR(-2.32%)    | IMPR(-3.14%) | IMPR(-3.51%) | IMPR(-3.76%)         | IMPR(-4.50%)            | IMPR(-6.76%) | IMPR(-7.41%) | IMPR(-7.87%)            | +0.25%       |

Some notes on the regressions:

• Mul2D goes from II=9 to II=10. We are still missing VST.SRS combining. Once this is there, this should be a candidate for the newer pipeliner
• Add2D/Mul2D bf16 variants: we now introduce lots of pointer copies instead of re-using the same pointer and enforcing ordering. This should be investigated separately
• ReduceSum/Mean int8 variants: should be handled in the post-pipeliner to target II=4
• Conv2D_2x8_0 The regression is in the outer loop. We are changing the live ranges of our 2D/3D iterators and forcing a spill

[andcarminati]

LGTM. We would certainly miss these combiners, the sooner they are enabled the better we can deal with the effects.