Trivial predicate is causing a 30% slowdown for matmul with grid swizzle

[zasdfgbnm]

For the example in FusionAmpereSwizzle_CUDA, the generated code contains trivial predicates:

    #pragma unroll
    for(nvfuser_index_t i653 = 0; i653 < 4; ++i653) {
      int i10749;
      i10749 = 32 * i653;
      #pragma unroll
      for(nvfuser_index_t i654 = 0; i654 < 8; ++i654) {
        if (((nvfuser_index_t)blockIdx.x) < ((ceilDiv(T1.size[1], 128)) * 4)) {
          Ampere::M16N8K16TN<16>(
            reinterpret_cast<Array<float,4,4>*>(&T5[(i10749 + (2 * i654))]),
            &(reinterpret_cast<Array<__half,8,8>*>(&T2)[i653]),
            &(reinterpret_cast<Array<__half,4,4>*>(&T3)[i654]));
        }
      }
    }

where ((nvfuser_index_t)blockIdx.x) < ((ceilDiv(T1.size[1], 128)) * 4) is trivial because the rhs of < is identical to gridDim.x. We should simplify this trivial predicate.

On RTX 3090, the perf with and without that trivial predicate is 20.8374 ms vs 16.1956 ms

Tasks

Beta Give feedback

1. Rewrite ParallelDimensionMap with expr simplifier #86
   +0
2. Prove i1 * i2 >= i1 if i2 >= 1 && i1 >= 0 #94
   +0
3. Some misc expr simplification #105
   +0
4. Cleanup computeLaunchParams and warp-pad handling #98
   +0
5. Remove trivial predicate for matmul #106
   +0

[zasdfgbnm]

cc @mmigdal-nv @drzejan2

[mmigdal-nv]

Thank you a lot for these fixes!