submission 249898

#!POPCORN leaderboard nvfp4_dual_gemm
#!POPCORN gpu NVIDIA

import torch
from torch.utils.cpp_extension import load_inline


_CUDA_SRC = r"""
#include <cuda.h>
#include <cudaTypedefs.h>
#include <cuda_fp16.h>
#include <cuda_runtime.h>

#include <torch/library.h>
#include <ATen/core/Tensor.h>

constexpr int WARP_SIZE = 32;
constexpr int MMA_K = 64;

constexpr uint64_t EVICT_FIRST = 0x12F0000000000000ULL;
constexpr uint64_t EVICT_LAST  = 0x14F0000000000000ULL;

__device__ __forceinline__ constexpr uint64_t desc_encode(uint64_t x) { return (x & 0x3'FFFFULL) >> 4ULL; }

__device__ __forceinline__ uint32_t elect_sync() {
  uint32_t pred = 0;
  asm volatile(
    "{\n\t"
    ".reg .pred %%px;\n\t"
    "elect.sync _|%%px, %1;\n\t"
    "@%%px mov.s32 %0, 1;\n\t"
    "}\n\t"
    : "+r"(pred)
    : "r"(0xFFFFFFFF)
  );
  return pred;
}

__device__ __forceinline__ void mbarrier_init(int mbar_addr, int count) {
  asm volatile("mbarrier.init.shared::cta.b64 [%0], %1;" :: "r"(mbar_addr), "r"(count));
}

__device__ __forceinline__ void mbarrier_wait(int mbar_addr, int phase) {
  uint32_t ticks = 0x989680;
  asm volatile(
    "{\n\t"
    ".reg .pred P1;\n\t"
    "LAB_WAIT:\n\t"
    "mbarrier.try_wait.parity.acquire.cta.shared::cta.b64 P1, [%0], %1, %2;\n\t"
    "@P1 bra.uni DONE;\n\t"
    "bra.uni LAB_WAIT;\n\t"
    "DONE:\n\t"
    "}\n\t"
    :: "r"(mbar_addr), "r"(phase), "r"(ticks)
  );
}

__device__ __forceinline__ void tma_gmem2smem(int dst, const void *src, int size, int mbar_addr, uint64_t cache_policy) {
  asm volatile(
    "cp.async.bulk.shared::cta.global.mbarrier::complete_tx::bytes.L2::cache_hint "
    "[%0], [%1], %2, [%3], %4;"
    :: "r"(dst), "l"(src), "r"(size), "r"(mbar_addr), "l"(cache_policy)
  );
}

__device__ __forceinline__ void tma_3d_gmem2smem(int dst, const void *tmap_ptr, int x, int y, int z, int mbar_addr, uint64_t cache_policy) {
  asm volatile(
    "cp.async.bulk.tensor.3d.shared::cta.global.mbarrier::complete_tx::bytes.cta_group::1.L2::cache_hint "
    "[%0], [%1, {%2, %3, %4}], [%5], %6;"
    :: "r"(dst), "l"(tmap_ptr), "r"(x), "r"(y), "r"(z), "r"(mbar_addr), "l"(cache_policy)
    : "memory"
  );
}

__device__ __forceinline__ void tcgen05_cp_nvfp4(int taddr, uint64_t s_desc) {
  asm volatile("tcgen05.cp.cta_group::1.32x128b.warpx4 [%0], %1;" :: "r"(taddr), "l"(s_desc));
}

__device__ __forceinline__ void tcgen05_mma_nvfp4(
  int d_tmem,
  uint64_t a_desc,
  uint64_t b_desc,
  uint32_t i_desc,
  int scale_A_tmem,
  int scale_B_tmem,
  int enable_input_d
) {
  asm volatile(
    "{\n\t"
    ".reg .pred p;\n\t"
    "setp.ne.b32 p, %6, 0;\n\t"
    "tcgen05.mma.cta_group::1.kind::mxf4nvf4.block_scale.block16 [%0], %1, %2, %3, [%4], [%5], p;\n\t"
    "}\n\t"
    :: "r"(d_tmem), "l"(a_desc), "l"(b_desc), "r"(i_desc),
       "r"(scale_A_tmem), "r"(scale_B_tmem), "r"(enable_input_d)
  );
}

struct SHAPE { static constexpr char _16x256b[] = ".16x256b"; };
struct NUM { static constexpr char x8[] = ".x8"; };

template <const char *SHAPE_, const char *NUM_>
__device__ __forceinline__ void tcgen05_ld_32regs(float *tmp, int row, int col) {
  asm volatile(
    "tcgen05.ld.sync.aligned%33%34.b32 "
    "{ %0,  %1,  %2,  %3,  %4,  %5,  %6,  %7, "
    "  %8,  %9, %10, %11, %12, %13, %14, %15, "
    " %16, %17, %18, %19, %20, %21, %22, %23, "
    " %24, %25, %26, %27, %28, %29, %30, %31}, [%32];"
    : "=f"(tmp[ 0]), "=f"(tmp[ 1]), "=f"(tmp[ 2]), "=f"(tmp[ 3]), "=f"(tmp[ 4]), "=f"(tmp[ 5]), "=f"(tmp[ 6]), "=f"(tmp[ 7]),
      "=f"(tmp[ 8]), "=f"(tmp[ 9]), "=f"(tmp[10]), "=f"(tmp[11]), "=f"(tmp[12]), "=f"(tmp[13]), "=f"(tmp[14]), "=f"(tmp[15]),
      "=f"(tmp[16]), "=f"(tmp[17]), "=f"(tmp[18]), "=f"(tmp[19]), "=f"(tmp[20]), "=f"(tmp[21]), "=f"(tmp[22]), "=f"(tmp[23]),
      "=f"(tmp[24]), "=f"(tmp[25]), "=f"(tmp[26]), "=f"(tmp[27]), "=f"(tmp[28]), "=f"(tmp[29]), "=f"(tmp[30]), "=f"(tmp[31])
    : "r"((row << 16) | col), "C"(SHAPE_), "C"(NUM_)
  );
}

__device__ __forceinline__ void tcgen05_ld_16x256bx8(float *tmp, int row, int col) {
  tcgen05_ld_32regs<SHAPE::_16x256b, NUM::x8>(tmp, row, col);
}

static inline void ck_cu(CUresult err) {
  if (err == CUDA_SUCCESS) return;
  const char *msg = nullptr;
  if (cuGetErrorString(err, &msg) != CUDA_SUCCESS) msg = "cu err";
  TORCH_CHECK(false, msg);
}

static inline void init_AB_tmap(
  CUtensorMap *tmap,
  const char *ptr,
  uint64_t global_h, uint64_t global_w,
  uint32_t shared_h, uint32_t shared_w
) {
  constexpr uint32_t rank = 3;
  uint64_t globalDim[rank]       = {256, global_h, global_w / 256};
  uint64_t globalStrides[rank-1] = {global_w / 2, 128};
  uint32_t boxDim[rank]          = {256, shared_h, shared_w / 256};
  uint32_t elementStrides[rank]  = {1, 1, 1};

  auto err = cuTensorMapEncodeTiled(
    tmap,
    CUtensorMapDataType::CU_TENSOR_MAP_DATA_TYPE_16U4_ALIGN8B,
    rank,
    (void *)ptr,
    globalDim,
    globalStrides,
    boxDim,
    elementStrides,
    CUtensorMapInterleave::CU_TENSOR_MAP_INTERLEAVE_NONE,
    CUtensorMapSwizzle::CU_TENSOR_MAP_SWIZZLE_128B,
    CUtensorMapL2promotion::CU_TENSOR_MAP_L2_PROMOTION_NONE,
    CUtensorMapFloatOOBfill::CU_TENSOR_MAP_FLOAT_OOB_FILL_NONE
  );
  ck_cu(err);
}

template <int K>
__global__ __launch_bounds__(128 + 2 * WARP_SIZE)
void dual_gemm_silu_mul_bn64_s5(
  const __grid_constant__ CUtensorMap A_tmap,
  const __grid_constant__ CUtensorMap B1_tmap,
  const __grid_constant__ CUtensorMap B2_tmap,
  const char *SFA_ptr,
  const char *SFB1_ptr,
  const char *SFB2_ptr,
  half *Out_ptr,
  int M, int N
) {
  constexpr int BLOCK_M = 128;
  constexpr int BLOCK_N = 64;
  constexpr int BLOCK_K = 256;
  constexpr int NUM_STAGES = 4;

  const int tid = threadIdx.x;
  const int bid = blockIdx.y;

  const int lane_id = tid & 31;
  const int warp_id = tid >> 5;

  const int grid_n = N / BLOCK_N;
  const int bid_m = bid / grid_n;
  const int bid_n = bid - bid_m * grid_n;

  const int off_m = bid_m * BLOCK_M;
  const int off_n = bid_n * BLOCK_N;

  constexpr int NUM_WARPS = BLOCK_M / WARP_SIZE + 2;

  extern __shared__ __align__(1024) char smem_ptr[];
  const int smem = static_cast<int>(__cvta_generic_to_shared(smem_ptr));
  constexpr int A_size = BLOCK_M * BLOCK_K / 2;
  constexpr int B_size = BLOCK_N * BLOCK_K / 2;
  constexpr int SFA_size = 128 * BLOCK_K / 16;
  constexpr int SFB_size = 128 * BLOCK_K / 16;
  constexpr int STAGE_SIZE = A_size + (2 * B_size) + SFA_size + (2 * SFB_size);

  #pragma nv_diag_suppress static_var_with_dynamic_init
  __shared__ int64_t mbars[NUM_STAGES * 2 + 1];
  const int tma_mbar_addr = static_cast<int>(__cvta_generic_to_shared(mbars));
  const int mma_mbar_addr = tma_mbar_addr + NUM_STAGES * 8;
  const int mainloop_mbar_addr = mma_mbar_addr + NUM_STAGES * 8;

  constexpr int OUT1_tmem = 0;
  constexpr int OUT2_tmem = BLOCK_N;
  constexpr int SFA_tmem = 2 * BLOCK_N;
  constexpr int SFB1_tmem = SFA_tmem + 4 * (BLOCK_K / MMA_K);
  constexpr int SFB2_tmem = SFB1_tmem + 4 * (BLOCK_K / MMA_K);
  constexpr int TMEM_ALLOC = BLOCK_N * 4;

  if (warp_id == 0 && elect_sync()) {
    #pragma unroll
    for (int i = 0; i < NUM_STAGES * 2 + 1; i++) mbarrier_init(tma_mbar_addr + i * 8, 1);
    asm volatile("fence.mbarrier_init.release.cluster;");
  } else if (warp_id == 1) {
    asm volatile(
      "tcgen05.alloc.cta_group::1.sync.aligned.shared::cta.b32 [%0], %1;"
      :: "r"(smem), "r"(TMEM_ALLOC)
    );
  }
  __syncthreads();

  constexpr int num_iters = K / BLOCK_K;

  if (warp_id == NUM_WARPS - 2 && elect_sync()) {
    const uint64_t cache_A = EVICT_LAST;
    const uint64_t cache_B = EVICT_FIRST;

    auto issue_tma = [&](int iter_k, int stage_id) {
      const int mbar_addr = tma_mbar_addr + stage_id * 8;
      const int A_smem = smem + stage_id * STAGE_SIZE;
      const int B1_smem = A_smem + A_size;
      const int B2_smem = B1_smem + B_size;
      const int SFA_smem = B2_smem + B_size;
      const int SFB1_smem = SFA_smem + SFA_size;
      const int SFB2_smem = SFB1_smem + SFB_size;

      const int off_k = iter_k * BLOCK_K;
      tma_3d_gmem2smem(A_smem, &A_tmap, 0, off_m, off_k / 256, mbar_addr, cache_A);
      tma_3d_gmem2smem(B1_smem, &B1_tmap, 0, off_n, off_k / 256, mbar_addr, cache_B);
      tma_3d_gmem2smem(B2_smem, &B2_tmap, 0, off_n, off_k / 256, mbar_addr, cache_B);

      const int rest_k = K / 16 / 4;
      const char *SFA_src = SFA_ptr + ((off_m / 128) * rest_k + off_k / (16 * 4)) * 512;
      const char *SFB1_src = SFB1_ptr + ((off_n / 128) * rest_k + off_k / (16 * 4)) * 512;
      const char *SFB2_src = SFB2_ptr + ((off_n / 128) * rest_k + off_k / (16 * 4)) * 512;
      tma_gmem2smem(SFA_smem, SFA_src, SFA_size, mbar_addr, cache_A);
      tma_gmem2smem(SFB1_smem, SFB1_src, SFB_size, mbar_addr, cache_B);
      tma_gmem2smem(SFB2_smem, SFB2_src, SFB_size, mbar_addr, cache_B);

      asm volatile(
        "mbarrier.arrive.expect_tx.release.cta.shared::cta.b64 _, [%0], %1;"
        :: "r"(mbar_addr), "r"(STAGE_SIZE)
        : "memory"
      );
    };

    constexpr int PRELOAD = (num_iters < NUM_STAGES) ? num_iters : NUM_STAGES;
    #pragma unroll
    for (int iter_k = 0; iter_k < PRELOAD; iter_k++) issue_tma(iter_k, iter_k);
    for (int iter_k = NUM_STAGES; iter_k < num_iters; iter_k++) {
      const int stage_id = iter_k % NUM_STAGES;
      const int mma_phase = (iter_k / NUM_STAGES - 1) & 1;
      mbarrier_wait(mma_mbar_addr + stage_id * 8, mma_phase);
      issue_tma(iter_k, stage_id);
    }
  } else if (warp_id == NUM_WARPS - 1 && elect_sync()) {
    constexpr int MMA_N = BLOCK_N;
    constexpr int MMA_M = 128;
    constexpr uint32_t i_desc = (1U << 7U) | (1U << 10U) | ((uint32_t)MMA_N >> 3U << 17U) | ((uint32_t)MMA_M >> 7U << 27U);

    for (int iter_k = 0; iter_k < num_iters; iter_k++) {
      const int stage_id = iter_k % NUM_STAGES;
      const int tma_phase = (iter_k / NUM_STAGES) & 1;
      mbarrier_wait(tma_mbar_addr + stage_id * 8, tma_phase);

      const int A_smem = smem + stage_id * STAGE_SIZE;
      const int B1_smem = A_smem + A_size;
      const int B2_smem = B1_smem + B_size;
      const int SFA_smem = B2_smem + B_size;
      const int SFB1_smem = SFA_smem + SFA_size;
      const int SFB2_smem = SFB1_smem + SFB_size;

      auto make_desc_AB = [](int addr) -> uint64_t {
        const int SBO = 8 * 128;
        return desc_encode(addr) | (desc_encode(SBO) << 32ULL) | (1ULL << 46ULL) | (2ULL << 61ULL);
      };
      auto make_desc_SF = [](int addr) -> uint64_t {
        const int SBO = 8 * 16;
        return desc_encode(addr) | (desc_encode(SBO) << 32ULL) | (1ULL << 46ULL);
      };

      constexpr uint64_t SF_desc = make_desc_SF(0);
      const uint64_t SFA_desc = SF_desc + ((uint64_t)SFA_smem >> 4ULL);
      const uint64_t SFB1_desc = SF_desc + ((uint64_t)SFB1_smem >> 4ULL);
      const uint64_t SFB2_desc = SF_desc + ((uint64_t)SFB2_smem >> 4ULL);

      #pragma unroll
      for (int k = 0; k < BLOCK_K / MMA_K; k++) {
        const uint64_t sfa_desc = SFA_desc + (uint64_t)k * (512ULL >> 4ULL);
        const uint64_t sfb1_desc = SFB1_desc + (uint64_t)k * (512ULL >> 4ULL);
        const uint64_t sfb2_desc = SFB2_desc + (uint64_t)k * (512ULL >> 4ULL);
        tcgen05_cp_nvfp4(SFA_tmem + k * 4, sfa_desc);
        tcgen05_cp_nvfp4(SFB1_tmem + k * 4, sfb1_desc);
        tcgen05_cp_nvfp4(SFB2_tmem + k * 4, sfb2_desc);
      }

      #pragma unroll
      for (int k2 = 0; k2 < 256 / MMA_K; k2++) {
        const uint64_t a_desc = make_desc_AB(A_smem + k2 * 32);
        const uint64_t b1_desc = make_desc_AB(B1_smem + k2 * 32);
        const uint64_t b2_desc = make_desc_AB(B2_smem + k2 * 32);

        const int k_sf = k2;
        const int scale_A_tmem = SFA_tmem + k_sf * 4;
        const int sel_n = (bid_n & 1) * (BLOCK_N / 32);
        const int scale_B1_tmem = SFB1_tmem + k_sf * 4 + sel_n;
        const int scale_B2_tmem = SFB2_tmem + k_sf * 4 + sel_n;

        const int enable_input_d = (k2 == 0) ? iter_k : 1;
        tcgen05_mma_nvfp4(OUT1_tmem, a_desc, b1_desc, i_desc, scale_A_tmem, scale_B1_tmem, enable_input_d);
        tcgen05_mma_nvfp4(OUT2_tmem, a_desc, b2_desc, i_desc, scale_A_tmem, scale_B2_tmem, enable_input_d);
      }

      asm volatile(
        "tcgen05.commit.cta_group::1.mbarrier::arrive::one.shared::cluster.b64 [%0];"
        :: "r"(mma_mbar_addr + stage_id * 8)
        : "memory"
      );
    }

    asm volatile(
      "tcgen05.commit.cta_group::1.mbarrier::arrive::one.shared::cluster.b64 [%0];"
      :: "r"(mainloop_mbar_addr)
      : "memory"
    );
  } else if (tid < 128) {
    mbarrier_wait(mainloop_mbar_addr, 0);
    asm volatile("tcgen05.fence::after_thread_sync;");

    #pragma unroll
    for (int mm = 0; mm < 2; mm++) {
      float x[BLOCK_N / 2];
      float y[BLOCK_N / 2];
      tcgen05_ld_16x256bx8(x, warp_id * 32 + mm * 16, 0);
      tcgen05_ld_16x256bx8(y, warp_id * 32 + mm * 16, OUT2_tmem);
      asm volatile("tcgen05.wait::ld.sync.aligned;");

      #pragma unroll
      for (int i = 0; i < BLOCK_N / 8; i++) {
        const int row = off_m + warp_id * 32 + mm * 16 + lane_id / 4;
        const int col = off_n + i * 8 + (lane_id & 3) * 2;

        const float2 x0 = float2{x[i * 4 + 0], x[i * 4 + 1]};
        const float2 x8 = float2{x[i * 4 + 2], x[i * 4 + 3]};
        const float2 y0 = float2{y[i * 4 + 0], y[i * 4 + 1]};
        const float2 y8 = float2{y[i * 4 + 2], y[i * 4 + 3]};

        float2 o0;
        float2 o8;
        const float s00 = 1.0f / (1.0f + __expf(-x0.x));
        const float s01 = 1.0f / (1.0f + __expf(-x0.y));
        const float s80 = 1.0f / (1.0f + __expf(-x8.x));
        const float s81 = 1.0f / (1.0f + __expf(-x8.y));
        o0.x = (x0.x * s00) * y0.x;
        o0.y = (x0.y * s01) * y0.y;
        o8.x = (x8.x * s80) * y8.x;
        o8.y = (x8.y * s81) * y8.y;

        reinterpret_cast<half2 *>(Out_ptr + (row + 0) * N + col)[0] = __float22half2_rn(o0);
        reinterpret_cast<half2 *>(Out_ptr + (row + 8) * N + col)[0] = __float22half2_rn(o8);
      }
    }

    asm volatile("bar.sync 1, %0;" :: "r"(128) : "memory");
    if (warp_id == 0) asm volatile("tcgen05.dealloc.cta_group::1.sync.aligned.b32 %0, %1;" :: "r"(0), "r"(TMEM_ALLOC));
  }
}

template <int K>
__global__ __launch_bounds__(128 + 2 * WARP_SIZE)
void dual_gemm_silu_mul_bn128_s4(
  const __grid_constant__ CUtensorMap A_tmap,
  const __grid_constant__ CUtensorMap B1_tmap,
  const __grid_constant__ CUtensorMap B2_tmap,
  const char *SFA_ptr,
  const char *SFB1_ptr,
  const char *SFB2_ptr,
  half *Out_ptr,
  int M, int N
) {
  constexpr int BLOCK_M = 128;
  constexpr int BLOCK_N = 128;
  constexpr int BLOCK_K = 256;
  constexpr int NUM_STAGES = 4;

  const int tid = threadIdx.x;
  const int bid = blockIdx.y;

  const int lane_id = tid & 31;
  const int warp_id = tid >> 5;

  const int grid_n = N / BLOCK_N;
  const int bid_m = bid / grid_n;
  const int bid_n = bid - bid_m * grid_n;

  const int off_m = bid_m * BLOCK_M;
  const int off_n = bid_n * BLOCK_N;

  constexpr int NUM_WARPS = BLOCK_M / WARP_SIZE + 2;

  extern __shared__ __align__(1024) char smem_ptr[];
  const int smem = static_cast<int>(__cvta_generic_to_shared(smem_ptr));
  constexpr int A_size = BLOCK_M * BLOCK_K / 2;
  constexpr int B_size = BLOCK_N * BLOCK_K / 2;
  constexpr int SFA_size = 128 * BLOCK_K / 16;
  constexpr int SFB_size = 128 * BLOCK_K / 16;
  constexpr int STAGE_SIZE = A_size + (2 * B_size) + SFA_size + (2 * SFB_size);

  #pragma nv_diag_suppress static_var_with_dynamic_init
  __shared__ int64_t mbars[NUM_STAGES * 2 + 1];
  const int tma_mbar_addr = static_cast<int>(__cvta_generic_to_shared(mbars));
  const int mma_mbar_addr = tma_mbar_addr + NUM_STAGES * 8;
  const int mainloop_mbar_addr = mma_mbar_addr + NUM_STAGES * 8;

  constexpr int OUT1_tmem = 0;
  constexpr int OUT2_tmem = BLOCK_N;
  constexpr int SFA_tmem = 2 * BLOCK_N;
  constexpr int SFB1_tmem = SFA_tmem + 4 * (BLOCK_K / MMA_K);
  constexpr int SFB2_tmem = SFB1_tmem + 4 * (BLOCK_K / MMA_K);
  constexpr int TMEM_ALLOC = 512;

  if (warp_id == 0 && elect_sync()) {
    #pragma unroll
    for (int i = 0; i < NUM_STAGES * 2 + 1; i++) mbarrier_init(tma_mbar_addr + i * 8, 1);
    asm volatile("fence.mbarrier_init.release.cluster;");
  } else if (warp_id == 1) {
    asm volatile(
      "tcgen05.alloc.cta_group::1.sync.aligned.shared::cta.b32 [%0], %1;"
      :: "r"(smem), "r"(TMEM_ALLOC)
    );
  }
  __syncthreads();

  constexpr int num_iters = K / BLOCK_K;

  if (warp_id == NUM_WARPS - 2 && elect_sync()) {
    const uint64_t cache_A = EVICT_LAST;
    const uint64_t cache_B = EVICT_FIRST;

    auto issue_tma = [&](int iter_k, int stage_id) {
      const int mbar_addr = tma_mbar_addr + stage_id * 8;
      const int A_smem = smem + stage_id * STAGE_SIZE;
      const int B1_smem = A_smem + A_size;
      const int B2_smem = B1_smem + B_size;
      const int SFA_smem = B2_smem + B_size;
      const int SFB1_smem = SFA_smem + SFA_size;
      const int SFB2_smem = SFB1_smem + SFB_size;

      const int off_k = iter_k * BLOCK_K;
      tma_3d_gmem2smem(A_smem, &A_tmap, 0, off_m, off_k / 256, mbar_addr, cache_A);
      tma_3d_gmem2smem(B1_smem, &B1_tmap, 0, off_n, off_k / 256, mbar_addr, cache_B);
      tma_3d_gmem2smem(B2_smem, &B2_tmap, 0, off_n, off_k / 256, mbar_addr, cache_B);

      const int rest_k = K / 16 / 4;
      const char *SFA_src = SFA_ptr + ((off_m / 128) * rest_k + off_k / (16 * 4)) * 512;
      const char *SFB1_src = SFB1_ptr + ((off_n / 128) * rest_k + off_k / (16 * 4)) * 512;
      const char *SFB2_src = SFB2_ptr + ((off_n / 128) * rest_k + off_k / (16 * 4)) * 512;
      tma_gmem2smem(SFA_smem, SFA_src, SFA_size, mbar_addr, cache_A);
      tma_gmem2smem(SFB1_smem, SFB1_src, SFB_size, mbar_addr, cache_B);
      tma_gmem2smem(SFB2_smem, SFB2_src, SFB_size, mbar_addr, cache_B);

      asm volatile(
        "mbarrier.arrive.expect_tx.release.cta.shared::cta.b64 _, [%0], %1;"
        :: "r"(mbar_addr), "r"(STAGE_SIZE)
        : "memory"
      );
    };

    constexpr int PRELOAD = (num_iters < NUM_STAGES) ? num_iters : NUM_STAGES;
    #pragma unroll
    for (int iter_k = 0; iter_k < PRELOAD; iter_k++) issue_tma(iter_k, iter_k);
    for (int iter_k = NUM_STAGES; iter_k < num_iters; iter_k++) {
      const int stage_id = iter_k % NUM_STAGES;
      const int mma_phase = (iter_k / NUM_STAGES - 1) & 1;
      mbarrier_wait(mma_mbar_addr + stage_id * 8, mma_phase);
      issue_tma(iter_k, stage_id);
    }
  } else if (warp_id == NUM_WARPS - 1 && elect_sync()) {
    constexpr int MMA_N = BLOCK_N;
    constexpr int MMA_M = 128;
    constexpr uint32_t i_desc = (1U << 7U) | (1U << 10U) | ((uint32_t)MMA_N >> 3U << 17U) | ((uint32_t)MMA_M >> 7U << 27U);

    for (int iter_k = 0; iter_k < num_iters; iter_k++) {
      const int stage_id = iter_k % NUM_STAGES;
      const int tma_phase = (iter_k / NUM_STAGES) & 1;
      mbarrier_wait(tma_mbar_addr + stage_id * 8, tma_phase);

      const int A_smem = smem + stage_id * STAGE_SIZE;
      const int B1_smem = A_smem + A_size;
      const int B2_smem = B1_smem + B_size;
      const int SFA_smem = B2_smem + B_size;
      const int SFB1_smem = SFA_smem + SFA_size;
      const int SFB2_smem = SFB1_smem + SFB_size;

      auto make_desc_AB = [](int addr) -> uint64_t {
        const int SBO = 8 * 128;
        return desc_encode(addr) | (desc_encode(SBO) << 32ULL) | (1ULL << 46ULL) | (2ULL << 61ULL);
      };
      auto make_desc_SF = [](int addr) -> uint64_t {
        const int SBO = 8 * 16;
        return desc_encode(addr) | (desc_encode(SBO) << 32ULL) | (1ULL << 46ULL);
      };

      constexpr uint64_t SF_desc = make_desc_SF(0);
      const uint64_t SFA_desc = SF_desc + ((uint64_t)SFA_smem >> 4ULL);
      const uint64_t SFB1_desc = SF_desc + ((uint64_t)SFB1_smem >> 4ULL);
      const uint64_t SFB2_desc = SF_desc + ((uint64_t)SFB2_smem >> 4ULL);

      #pragma unroll
      for (int k = 0; k < BLOCK_K / MMA_K; k++) {
        const uint64_t sfa_desc = SFA_desc + (uint64_t)k * (512ULL >> 4ULL);
        const uint64_t sfb1_desc = SFB1_desc + (uint64_t)k * (512ULL >> 4ULL);
        const uint64_t sfb2_desc = SFB2_desc + (uint64_t)k * (512ULL >> 4ULL);
        tcgen05_cp_nvfp4(SFA_tmem + k * 4, sfa_desc);
        tcgen05_cp_nvfp4(SFB1_tmem + k * 4, sfb1_desc);
        tcgen05_cp_nvfp4(SFB2_tmem + k * 4, sfb2_desc);
      }

      #pragma unroll
      for (int k2 = 0; k2 < 256 / MMA_K; k2++) {
        const uint64_t a_desc = make_desc_AB(A_smem + k2 * 32);
        const uint64_t b1_desc = make_desc_AB(B1_smem + k2 * 32);
        const uint64_t b2_desc = make_desc_AB(B2_smem + k2 * 32);

        const int k_sf = k2;
        const int scale_A_tmem = SFA_tmem + k_sf * 4;
        const int scale_B1_tmem = SFB1_tmem + k_sf * 4;
        const int scale_B2_tmem = SFB2_tmem + k_sf * 4;

        const int enable_input_d = (k2 == 0) ? iter_k : 1;
        tcgen05_mma_nvfp4(OUT1_tmem, a_desc, b1_desc, i_desc, scale_A_tmem, scale_B1_tmem, enable_input_d);
        tcgen05_mma_nvfp4(OUT2_tmem, a_desc, b2_desc, i_desc, scale_A_tmem, scale_B2_tmem, enable_input_d);
      }

      asm volatile(
        "tcgen05.commit.cta_group::1.mbarrier::arrive::one.shared::cluster.b64 [%0];"
        :: "r"(mma_mbar_addr + stage_id * 8)
        : "memory"
      );
    }

    asm volatile(
      "tcgen05.commit.cta_group::1.mbarrier::arrive::one.shared::cluster.b64 [%0];"
      :: "r"(mainloop_mbar_addr)
      : "memory"
    );
  } else if (tid < 128) {
    mbarrier_wait(mainloop_mbar_addr, 0);
    asm volatile("tcgen05.fence::after_thread_sync;");

    #pragma unroll
    for (int mm = 0; mm < 2; mm++) {
      float x[BLOCK_N / 2];
      float y[BLOCK_N / 2];
      tcgen05_ld_16x256bx8(x, warp_id * 32 + mm * 16, 0);
      tcgen05_ld_16x256bx8(y, warp_id * 32 + mm * 16, OUT2_tmem);
      tcgen05_ld_16x256bx8(x + 32, warp_id * 32 + mm * 16, 64);
      tcgen05_ld_16x256bx8(y + 32, warp_id * 32 + mm * 16, OUT2_tmem + 64);
      asm volatile("tcgen05.wait::ld.sync.aligned;");

      #pragma unroll
      for (int i = 0; i < BLOCK_N / 8; i++) {
        const int row = off_m + warp_id * 32 + mm * 16 + lane_id / 4;
        const int col = off_n + i * 8 + (lane_id & 3) * 2;

        const float2 x0 = float2{x[i * 4 + 0], x[i * 4 + 1]};
        const float2 x8 = float2{x[i * 4 + 2], x[i * 4 + 3]};
        const float2 y0 = float2{y[i * 4 + 0], y[i * 4 + 1]};
        const float2 y8 = float2{y[i * 4 + 2], y[i * 4 + 3]};

        float2 o0;
        float2 o8;
        const float s00 = 1.0f / (1.0f + __expf(-x0.x));
        const float s01 = 1.0f / (1.0f + __expf(-x0.y));
        const float s80 = 1.0f / (1.0f + __expf(-x8.x));
        const float s81 = 1.0f / (1.0f + __expf(-x8.y));
        o0.x = (x0.x * s00) * y0.x;
        o0.y = (x0.y * s01) * y0.y;
        o8.x = (x8.x * s80) * y8.x;
        o8.y = (x8.y * s81) * y8.y;

        reinterpret_cast<half2 *>(Out_ptr + (row + 0) * N + col)[0] = __float22half2_rn(o0);
        reinterpret_cast<half2 *>(Out_ptr + (row + 8) * N + col)[0] = __float22half2_rn(o8);
      }
    }

    asm volatile("bar.sync 1, %0;" :: "r"(128) : "memory");
    if (warp_id == 0) asm volatile("tcgen05.dealloc.cta_group::1.sync.aligned.b32 %0, %1;" :: "r"(0), "r"(TMEM_ALLOC));
  }
}

template <int K>
static inline void launch_dual_bn64_s5(
  const at::Tensor& A,
  const at::Tensor& B1,
  const at::Tensor& B2,
  const at::Tensor& SFA,
  const at::Tensor& SFB1,
  const at::Tensor& SFB2,
  at::Tensor& out
) {
  constexpr int BLOCK_M = 128;
  constexpr int BLOCK_N = 64;
  constexpr int BLOCK_K = 256;
  constexpr int NUM_STAGES = 4;

  const int M = (int)A.size(0);
  const int N = (int)B1.size(0);

  auto A_ptr = reinterpret_cast<const char *>(A.data_ptr());
  auto B1_ptr = reinterpret_cast<const char *>(B1.data_ptr());
  auto B2_ptr = reinterpret_cast<const char *>(B2.data_ptr());
  auto SFA_ptr = reinterpret_cast<const char *>(SFA.data_ptr());
  auto SFB1_ptr = reinterpret_cast<const char *>(SFB1.data_ptr());
  auto SFB2_ptr = reinterpret_cast<const char *>(SFB2.data_ptr());
  auto Out_ptr = reinterpret_cast<half *>(out.data_ptr());

  CUtensorMap A_tmap, B1_tmap, B2_tmap;
  init_AB_tmap(&A_tmap, A_ptr, (uint64_t)M, (uint64_t)K, (uint32_t)BLOCK_M, (uint32_t)BLOCK_K);
  init_AB_tmap(&B1_tmap, B1_ptr, (uint64_t)N, (uint64_t)K, (uint32_t)BLOCK_N, (uint32_t)BLOCK_K);
  init_AB_tmap(&B2_tmap, B2_ptr, (uint64_t)N, (uint64_t)K, (uint32_t)BLOCK_N, (uint32_t)BLOCK_K);

  dim3 grid(1, (unsigned)((M / BLOCK_M) * (N / BLOCK_N)));
  const int tb_size = BLOCK_M + 2 * WARP_SIZE;
  const int AB_size = (BLOCK_M + 2 * BLOCK_N) * (BLOCK_K / 2);
  const int SF_size = 128 * (BLOCK_K / 16) * 3;
  const int smem_size = (AB_size + SF_size) * NUM_STAGES;

  auto kptr = dual_gemm_silu_mul_bn64_s5<K>;
  if (smem_size > 48'000) cudaFuncSetAttribute(kptr, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
  kptr<<<grid, tb_size, smem_size>>>(A_tmap, B1_tmap, B2_tmap, SFA_ptr, SFB1_ptr, SFB2_ptr, Out_ptr, M, N);
}

template <int K>
static inline void launch_dual_bn128_s4(
  const at::Tensor& A,
  const at::Tensor& B1,
  const at::Tensor& B2,
  const at::Tensor& SFA,
  const at::Tensor& SFB1,
  const at::Tensor& SFB2,
  at::Tensor& out
) {
  constexpr int BLOCK_M = 128;
  constexpr int BLOCK_N = 128;
  constexpr int BLOCK_K = 256;
  constexpr int NUM_STAGES = 4;

  const int M = (int)A.size(0);
  const int N = (int)B1.size(0);

  auto A_ptr = reinterpret_cast<const char *>(A.data_ptr());
  auto B1_ptr = reinterpret_cast<const char *>(B1.data_ptr());
  auto B2_ptr = reinterpret_cast<const char *>(B2.data_ptr());
  auto SFA_ptr = reinterpret_cast<const char *>(SFA.data_ptr());
  auto SFB1_ptr = reinterpret_cast<const char *>(SFB1.data_ptr());
  auto SFB2_ptr = reinterpret_cast<const char *>(SFB2.data_ptr());
  auto Out_ptr = reinterpret_cast<half *>(out.data_ptr());

  CUtensorMap A_tmap, B1_tmap, B2_tmap;
  init_AB_tmap(&A_tmap, A_ptr, (uint64_t)M, (uint64_t)K, (uint32_t)BLOCK_M, (uint32_t)BLOCK_K);
  init_AB_tmap(&B1_tmap, B1_ptr, (uint64_t)N, (uint64_t)K, (uint32_t)BLOCK_N, (uint32_t)BLOCK_K);
  init_AB_tmap(&B2_tmap, B2_ptr, (uint64_t)N, (uint64_t)K, (uint32_t)BLOCK_N, (uint32_t)BLOCK_K);

  dim3 grid(1, (unsigned)((M / BLOCK_M) * (N / BLOCK_N)));
  const int tb_size = BLOCK_M + 2 * WARP_SIZE;
  const int AB_size = (BLOCK_M + 2 * BLOCK_N) * (BLOCK_K / 2);
  const int SF_size = 128 * (BLOCK_K / 16) * 3;
  const int smem_size = (AB_size + SF_size) * NUM_STAGES;

  auto kptr = dual_gemm_silu_mul_bn128_s4<K>;
  if (smem_size > 48'000) cudaFuncSetAttribute(kptr, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
  kptr<<<grid, tb_size, smem_size>>>(A_tmap, B1_tmap, B2_tmap, SFA_ptr, SFB1_ptr, SFB2_ptr, Out_ptr, M, N);
}

at::Tensor fused(
  const at::Tensor& A,
  const at::Tensor& B1,
  const at::Tensor& B2,
  const at::Tensor& SFA,
  const at::Tensor& SFB1,
  const at::Tensor& SFB2,
  at::Tensor& out
) {
  TORCH_CHECK(A.is_cuda() && B1.is_cuda() && B2.is_cuda(), "cuda");
  TORCH_CHECK(SFA.is_cuda() && SFB1.is_cuda() && SFB2.is_cuda(), "cuda");
  TORCH_CHECK(out.is_cuda(), "cuda");
  TORCH_CHECK(A.dim() == 3 && B1.dim() == 3 && B2.dim() == 3, "dim");
  TORCH_CHECK(out.dim() == 3, "dim");

  const int64_t M = A.size(0);
  const int64_t Kp = A.size(1);
  const int64_t L = A.size(2);
  const int64_t N = B1.size(0);
  TORCH_CHECK(L == 1, "l");
  TORCH_CHECK(B1.size(1) == Kp && B1.size(2) == L, "b1");
  TORCH_CHECK(B2.size(1) == Kp && B2.size(2) == L, "b2");
  TORCH_CHECK(out.size(0) == M && out.size(1) == N && out.size(2) == L, "out");

  TORCH_CHECK((M % 128) == 0, "m");
  TORCH_CHECK((N % 64) == 0, "n");

  const int K = (int)(Kp * 2);

  if (K == 7168) {
    if ((M >= 512) && ((N & 127) == 0)) {
      launch_dual_bn128_s4<7168>(A, B1, B2, SFA, SFB1, SFB2, out);
    } else {
      launch_dual_bn64_s5<7168>(A, B1, B2, SFA, SFB1, SFB2, out);
    }
  } else if (K == 4096) {
    if ((M >= 512) && ((N & 127) == 0)) {
      launch_dual_bn128_s4<4096>(A, B1, B2, SFA, SFB1, SFB2, out);
    } else {
      launch_dual_bn64_s5<4096>(A, B1, B2, SFA, SFB1, SFB2, out);
    }
  } else if (K == 2304) {
    launch_dual_bn64_s5<2304>(A, B1, B2, SFA, SFB1, SFB2, out);
  } else if (K == 2048) {
    launch_dual_bn64_s5<2048>(A, B1, B2, SFA, SFB1, SFB2, out);
  } else if (K == 1536) {
    launch_dual_bn64_s5<1536>(A, B1, B2, SFA, SFB1, SFB2, out);
  } else if (K == 512) {
    launch_dual_bn64_s5<512>(A, B1, B2, SFA, SFB1, SFB2, out);
  } else if (K == 256) {
    launch_dual_bn64_s5<256>(A, B1, B2, SFA, SFB1, SFB2, out);
  } else {
    TORCH_CHECK(false, "k ", K);
  }

  return out;
}

TORCH_LIBRARY(nvfp4_dual_lib, m) {
  m.def("fused(Tensor A, Tensor B1, Tensor B2, Tensor SFA, Tensor SFB1, Tensor SFB2, Tensor(a!) out) -> Tensor");
  m.impl("fused", &fused);
}
"""


_loaded = False


def _load():
    global _loaded
    if _loaded:
        return
    load_inline(
        name="nvfp4_dual_ext_tc_dual_bn64_s5",
        cpp_sources="",
        cuda_sources=_CUDA_SRC,
        functions=None,
        with_cuda=True,
        extra_cuda_cflags=[
            "-O3",
            "-gencode=arch=compute_100a,code=sm_100a",
            "--use_fast_math",
            "--expt-relaxed-constexpr",
            "--relocatable-device-code=false",
            "-lineinfo",
        ],
        extra_ldflags=["-lcuda"],
        verbose=False,
        is_python_module=False,
        no_implicit_headers=True,
    )
    _loaded = True


def custom_kernel(data):
    _load()
    a, b1, b2, _sfa, _sfb1, _sfb2, sfa_p, sfb1_p, sfb2_p, c = data
    return torch.ops.nvfp4_dual_lib.fused(a, b1, b2, sfa_p, sfb1_p, sfb2_p, c)


__all__ = ["custom_kernel"]