Add dual-GEMM examples for SM90 (Hopper) and SM120 (Blackwell)

[Inodayy]

Summary

Implements dual-GEMM examples for SM90 (Hopper) and SM120 (Blackwell) using CUTLASS 3.x.

The dual-GEMM operation implemented is:

  D0 = epilogue0(X @ B0, C0)
  D1 = epilogue1(X @ B1, C1)
  D2 = element_wise(D0, D1)

Implementation details

• Based on the single-GEMM examples 48_hopper_warp_specialized_gemm.cu
  and 79a_blackwell_geforce_nvfp4_bf16_gemm.cu

• B0 and B1 layouts are not decoupled, but both are passed separately to the builders for potential future flexibility.
  (Blackwell supports only TN layout; Hopper assumes NK layout for make_tma_copy_B_sm90 etc.)

• D2 performs LeftSiLUAndMul similar to example 45_dual_gemm, implemented in collective/sm90_epilogue_tma_warpspecialized_dual.hpp store()

• D0 and D1 are intermediate results only and are not stored.

• Added template<class Op0, class Op1> in fusion/sm90_callbacks… to allow distinct operations for D0 and D1.

Performance (keeping all configurations same as single-GEMM examples)

SM90 (Hopper)

• Problem size: 2048×2048×2048
• Rasterization: Heuristic with max CTA swizzle 2
• Avg runtime: 0.20429 ms
• GFLOPS: 168,191
• ≈5% faster than two single-GEMM baseline

SM120 (Blackwell)

• Problem size: 2048×2048×2048
• Avg runtime: 0.155648 ms
• GFLOPS: 220,753
• ≈30% slower than two single-GEMM baseline (haven’t been able to find the root cause yet)

Notes

• I am relatively new to CUTLASS C++; this work was implemented as a learning exercise. I followed example structure similar to 63_hopper_gemm_with_weight_prefetch.
• The SM120 example was an initial local starting point and can be removed if unnecessary

Closes #1123

[Inodayy]

@hwu36 @mnicely Hi, just checking whether 3.x dual-gemm is still planned, and if there’s any chance this PR might get reviewed later if time allows? I’d appreciate any feedback on whether I’m on the right track. Thanks!