submission 190884

#!POPCORN leaderboard nvfp4_dual_gemm
#!POPCORN gpu NVIDIA

import torch
from typing import Optional
from torch.utils.cpp_extension import load_inline


# 编译内核并缓存
_CUDA_SRC_COMMON = r"""
#include <cuda.h>
#include <cudaTypedefs.h>
#include <cuda_fp16.h>
#include <cuda_runtime.h>
#include <math.h>

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

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

constexpr uint64_t EVICT_NORMAL = 0x1000000000000000;
constexpr uint64_t EVICT_FIRST = 0x12F0000000000000;
constexpr uint64_t EVICT_LAST = 0x14F0000000000000;

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

__device__ 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"
    "}"
    : "+r"(pred)
    : "r"(0xFFFFFFFF)
  );
  return pred;
}

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

__device__ 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"
    "}"
    :: "r"(mbar_addr), "r"(phase), "r"(ticks)
  );
}

__device__ inline 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__ inline 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__ inline 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__ inline void tcgen05_mma_nvfp4(
  uint64_t a_desc,
  uint64_t b_desc,
  uint32_t i_desc,
  int scale_A_tmem,
  int scale_B_tmem,
  int enable_input_d
) {
  const int d_tmem = 0;
  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"
    "}"
    :: "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 _32x32b[]  = ".32x32b";
  static constexpr char _16x128b[] = ".16x128b";
  static constexpr char _16x256b[] = ".16x256b";
};

struct NUM {
  static constexpr char x4[]  = ".x4";
  static constexpr char x8[]  = ".x8";
  static constexpr char x16[] = ".x16";
  static constexpr char x32[] = ".x32";
  static constexpr char x64[] = ".x64";
  static constexpr char x128[] = ".x128";
};

template <const char *SHAPE, const char *NUM>
__device__ inline void tcgen05_ld_16regs(float *tmp, int row, int col) {
  asm volatile("tcgen05.ld.sync.aligned%17%18.b32 "
              "{ %0,  %1,  %2,  %3,  %4,  %5,  %6,  %7, "
              "  %8,  %9, %10, %11, %12, %13, %14, %15}, [%16];"
              : "=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])
              : "r"((row << 16) | col), "C"(SHAPE), "C"(NUM));
}

template <const char *SHAPE, const char *NUM>
__device__ inline 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));
}

template <const char *SHAPE, const char *NUM>
__device__ inline void tcgen05_ld_64regs(float *tmp, int row, int col) {
  asm volatile("tcgen05.ld.sync.aligned%65%66.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, %33, %34, %35, %36, %37, %38, %39, "
              " %40, %41, %42, %43, %44, %45, %46, %47, "
              " %48, %49, %50, %51, %52, %53, %54, %55, "
              " %56, %57, %58, %59, %60, %61, %62, %63}, [%64];"
              : "=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]),
                "=f"(tmp[32]), "=f"(tmp[33]), "=f"(tmp[34]), "=f"(tmp[35]), "=f"(tmp[36]), "=f"(tmp[37]), "=f"(tmp[38]), "=f"(tmp[39]),
                "=f"(tmp[40]), "=f"(tmp[41]), "=f"(tmp[42]), "=f"(tmp[43]), "=f"(tmp[44]), "=f"(tmp[45]), "=f"(tmp[46]), "=f"(tmp[47]),
                "=f"(tmp[48]), "=f"(tmp[49]), "=f"(tmp[50]), "=f"(tmp[51]), "=f"(tmp[52]), "=f"(tmp[53]), "=f"(tmp[54]), "=f"(tmp[55]),
                "=f"(tmp[56]), "=f"(tmp[57]), "=f"(tmp[58]), "=f"(tmp[59]), "=f"(tmp[60]), "=f"(tmp[61]), "=f"(tmp[62]), "=f"(tmp[63])
              : "r"((row << 16) | col), "C"(SHAPE), "C"(NUM));
}

template <const char *SHAPE, const char *NUM>
__device__ inline void tcgen05_ld_128regs(float *tmp, int row, int col) {
  asm volatile("tcgen05.ld.sync.aligned%129%130.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, %33, %34, %35, %36, %37, %38, %39, "
              " %40, %41, %42, %43, %44, %45, %46, %47, "
              " %48, %49, %50, %51, %52, %53, %54, %55, "
              " %56, %57, %58, %59, %60, %61, %62, %63, "
              " %64, %65, %66, %67, %68, %69, %70, %71, "
              " %72, %73, %74, %75, %76, %77, %78, %79, "
              " %80, %81, %82, %83, %84, %85, %86, %87, "
              " %88, %89, %90, %91, %92, %93, %94, %95, "
              " %96, %97, %98, %99,%100,%101,%102,%103, "
              "%104,%105,%106,%107,%108,%109,%110,%111, "
              "%112,%113,%114,%115,%116,%117,%118,%119, "
              "%120,%121,%122,%123,%124,%125,%126,%127}, [%128];"
              : "=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]),
                "=f"(tmp[32]), "=f"(tmp[33]), "=f"(tmp[34]), "=f"(tmp[35]), "=f"(tmp[36]), "=f"(tmp[37]), "=f"(tmp[38]), "=f"(tmp[39]),
                "=f"(tmp[40]), "=f"(tmp[41]), "=f"(tmp[42]), "=f"(tmp[43]), "=f"(tmp[44]), "=f"(tmp[45]), "=f"(tmp[46]), "=f"(tmp[47]),
                "=f"(tmp[48]), "=f"(tmp[49]), "=f"(tmp[50]), "=f"(tmp[51]), "=f"(tmp[52]), "=f"(tmp[53]), "=f"(tmp[54]), "=f"(tmp[55]),
                "=f"(tmp[56]), "=f"(tmp[57]), "=f"(tmp[58]), "=f"(tmp[59]), "=f"(tmp[60]), "=f"(tmp[61]), "=f"(tmp[62]), "=f"(tmp[63]),
                "=f"(tmp[64]), "=f"(tmp[65]), "=f"(tmp[66]), "=f"(tmp[67]), "=f"(tmp[68]), "=f"(tmp[69]), "=f"(tmp[70]), "=f"(tmp[71]),
                "=f"(tmp[72]), "=f"(tmp[73]), "=f"(tmp[74]), "=f"(tmp[75]), "=f"(tmp[76]), "=f"(tmp[77]), "=f"(tmp[78]), "=f"(tmp[79]),
                "=f"(tmp[80]), "=f"(tmp[81]), "=f"(tmp[82]), "=f"(tmp[83]), "=f"(tmp[84]), "=f"(tmp[85]), "=f"(tmp[86]), "=f"(tmp[87]),
                "=f"(tmp[88]), "=f"(tmp[89]), "=f"(tmp[90]), "=f"(tmp[91]), "=f"(tmp[92]), "=f"(tmp[93]), "=f"(tmp[94]), "=f"(tmp[95]),
                "=f"(tmp[96]), "=f"(tmp[97]), "=f"(tmp[98]), "=f"(tmp[99]), "=f"(tmp[100]),"=f"(tmp[101]),"=f"(tmp[102]),"=f"(tmp[103]),
                "=f"(tmp[104]),"=f"(tmp[105]),"=f"(tmp[106]),"=f"(tmp[107]),"=f"(tmp[108]),"=f"(tmp[109]),"=f"(tmp[110]),"=f"(tmp[111]),
                "=f"(tmp[112]),"=f"(tmp[113]),"=f"(tmp[114]),"=f"(tmp[115]),"=f"(tmp[116]),"=f"(tmp[117]),"=f"(tmp[118]),"=f"(tmp[119]),
                "=f"(tmp[120]),"=f"(tmp[121]),"=f"(tmp[122]),"=f"(tmp[123]),"=f"(tmp[124]),"=f"(tmp[125]),"=f"(tmp[126]),"=f"(tmp[127])
              : "r"((row << 16) | col), "C"(SHAPE), "C"(NUM));
}

__device__ inline void tcgen05_ld_32x32bx32(float *tmp, int row, int col) { tcgen05_ld_32regs<SHAPE::_32x32b, NUM::x32>(tmp, row, col); }
__device__ inline void tcgen05_ld_32x32bx64(float *tmp, int row, int col) { tcgen05_ld_64regs<SHAPE::_32x32b, NUM::x64>(tmp, row, col); }
__device__ inline void tcgen05_ld_32x32bx128(float *tmp, int row, int col) { tcgen05_ld_128regs<SHAPE::_32x32b, NUM::x128>(tmp, row, col); }

__device__ inline void tcgen05_ld_16x128bx8(float *tmp, int row, int col) { tcgen05_ld_16regs<SHAPE::_16x128b, NUM::x8>(tmp, row, col); }
__device__ inline void tcgen05_ld_16x128bx16(float *tmp, int row, int col) { tcgen05_ld_32regs<SHAPE::_16x128b, NUM::x16>(tmp, row, col); }
__device__ inline void tcgen05_ld_16x128bx32(float *tmp, int row, int col) { tcgen05_ld_64regs<SHAPE::_16x128b, NUM::x32>(tmp, row, col); }

__device__ inline void tcgen05_ld_16x256bx4(float *tmp, int row, int col) { tcgen05_ld_16regs<SHAPE::_16x256b, NUM::x4>(tmp, row, col); }
__device__ inline void tcgen05_ld_16x256bx8(float *tmp, int row, int col) { tcgen05_ld_32regs<SHAPE::_16x256b, NUM::x8>(tmp, row, col); }
__device__ inline void tcgen05_ld_16x256bx16(float *tmp, int row, int col) { tcgen05_ld_64regs<SHAPE::_16x256b, NUM::x16>(tmp, row, col); }

void init_AB_tmap(
  CUtensorMap *tmap,
  const char *ptr,
  uint64_t global_height, uint64_t global_width,
  uint32_t shared_height, uint32_t shared_width
) {
  constexpr uint32_t rank = 3;
  uint64_t globalDim[rank]       = {256, global_height, global_width / 256};
  uint64_t globalStrides[rank-1] = {global_width / 2, 128};
  uint32_t boxDim[rank]          = {256, shared_height, shared_width / 256};
  uint32_t elementStrides[rank]  = {1, 1, 1};

  CUresult status = 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
  );
  (void)status;
}
"""

_CUDA_SRC_V4 = r"""
template <
  int K,
  int BLOCK_M,
  int BLOCK_N,
  int BLOCK_K,
  int SPLIT_K,
  bool C_N_MAJOR,
  int NUM_STAGES
>
__global__ __launch_bounds__(BLOCK_M + 2 * WARP_SIZE)
void kernel(
  const __grid_constant__ CUtensorMap A_tmap,
  const __grid_constant__ CUtensorMap B_tmap,
  const char *SFA_ptr,
  const char *SFB_ptr,
  half *C_ptr,
  float *buf_ptr,
  int M, int N
) {
  const int tid = threadIdx.x;
  const int bid_k = blockIdx.x;
  const int bid = blockIdx.y;

  const int lane_id = tid % WARP_SIZE;
  const int warp_id = tid / WARP_SIZE;

  const int grid_m = M / BLOCK_M;
  const int grid_n = N / BLOCK_N;
  const int bid_m = bid / grid_n;
  const int bid_n = bid % 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 + B_size + SFA_size + 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 SFA_tmem = BLOCK_N;
  constexpr int SFB_tmem = SFA_tmem + 4 * (BLOCK_K / MMA_K);

  if (warp_id == 0 && elect_sync()) {
    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"(BLOCK_N * 2));
  }
  __syncthreads();

  constexpr int num_iters = K / BLOCK_K / SPLIT_K;

  if (warp_id == NUM_WARPS - 2 && elect_sync()) {
    uint64_t cache_A, cache_B;
    if (M > N) {
      cache_A = EVICT_FIRST;
      cache_B = EVICT_LAST;
    } else {
      cache_A = EVICT_LAST;
      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 B_smem = A_smem + A_size;
      const int SFA_smem = B_smem + B_size;
      const int SFB_smem = SFA_smem + SFA_size;

      const int off_k = SPLIT_K == 1 ? iter_k * BLOCK_K : (iter_k * SPLIT_K + bid_k) * BLOCK_K;
      tma_3d_gmem2smem(A_smem, &A_tmap, 0, off_m, off_k / 256, mbar_addr, cache_A);
      tma_3d_gmem2smem(B_smem, &B_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 *SFB_src = SFB_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(SFB_smem, SFB_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");
    };

    for (int iter_k = 0; iter_k < NUM_STAGES; 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) % 2;
      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) % 2;
      mbarrier_wait(tma_mbar_addr + stage_id * 8, tma_phase);

      const int A_smem = smem + stage_id * STAGE_SIZE;
      const int B_smem = A_smem + A_size;
      const int SFA_smem = B_smem + B_size;
      const int SFB_smem = SFA_smem + SFA_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 SFB_desc = SF_desc + ((uint64_t)SFB_smem >> 4ULL);

      for (int k = 0; k < BLOCK_K / MMA_K; k++) {
        uint64_t sfa_desc = SFA_desc + (uint64_t)k * (512ULL >> 4ULL);
        uint64_t sfb_desc = SFB_desc + (uint64_t)k * (512ULL >> 4ULL);
        tcgen05_cp_nvfp4(SFA_tmem + k * 4, sfa_desc);
        tcgen05_cp_nvfp4(SFB_tmem + k * 4, sfb_desc);
      }

      for (int k1 = 0; k1 < BLOCK_K / 256; k1++)
        for (int k2 = 0; k2 < 256 / MMA_K; k2++) {
          uint64_t a_desc = make_desc_AB(A_smem + k1 * BLOCK_M * 128 + k2 * 32);
          uint64_t b_desc = make_desc_AB(B_smem + k1 * BLOCK_N * 128 + k2 * 32);

          int k_sf = k1 * 4 + k2;
          const int scale_A_tmem = SFA_tmem + k_sf * 4 + (bid_m % (128 / BLOCK_M)) * (BLOCK_M / 32);
          const int scale_B_tmem = SFB_tmem + k_sf * 4 + (bid_n % (128 / BLOCK_N)) * (BLOCK_N / 32);

          const int enable_input_d = (k1 == 0 && k2 == 0) ? iter_k : 1;
          tcgen05_mma_nvfp4(a_desc, b_desc, i_desc, scale_A_tmem, scale_B_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 < BLOCK_M) {
    mbarrier_wait(mainloop_mbar_addr, 0);
    asm volatile("tcgen05.fence::after_thread_sync;");

    auto epilogue_M_major = [&]() {
      constexpr int WIDTH = (BLOCK_N < 64 ? BLOCK_N : 64);

      for (int n = 0; n < BLOCK_N / WIDTH; n++) {
        float tmp[WIDTH];
        if constexpr (WIDTH == 128) tcgen05_ld_32x32bx128(tmp, warp_id * 32, n * WIDTH);
        if constexpr (WIDTH == 64) tcgen05_ld_32x32bx64(tmp, warp_id * 32, n * WIDTH);
        if constexpr (WIDTH == 32) tcgen05_ld_32x32bx32(tmp, warp_id * 32, n * WIDTH);
        asm volatile("tcgen05.wait::ld.sync.aligned;");

        for (int i = 0; i < WIDTH; i++) {
          const int row = off_n + n * WIDTH + i;
          const int col = off_m + tid;

          if constexpr (SPLIT_K == 1)
            C_ptr[row * M + col] = __float2half(tmp[i]);
          else
            atomicAdd(buf_ptr + row * M + col, tmp[i]);
        }
      }
    };
    auto epilogue_N_major = [&]() {
      for (int m = 0; m < 32 / 16; m++) {
        float tmp[BLOCK_N / 2];
        if constexpr (BLOCK_N == 128) tcgen05_ld_16x256bx16(tmp, warp_id * 32 + m * 16, 0);
        if constexpr (BLOCK_N == 64) tcgen05_ld_16x256bx8(tmp, warp_id * 32 + m * 16, 0);
        if constexpr (BLOCK_N == 32) tcgen05_ld_16x256bx4(tmp, warp_id * 32 + m * 16, 0);
        asm volatile("tcgen05.wait::ld.sync.aligned;");

        for (int i = 0; i < BLOCK_N / 8; i++) {
          const int row = off_m + warp_id * 32 + m * 16 + lane_id / 4;
          const int col = off_n + i * 8 + (lane_id % 4) * 2;

          if constexpr (SPLIT_K == 1) {
            reinterpret_cast<half2 *>(C_ptr + (row + 0) * N + col)[0] = __float22half2_rn({tmp[i * 4 + 0], tmp[i * 4 + 1]});
            reinterpret_cast<half2 *>(C_ptr + (row + 8) * N + col)[0] = __float22half2_rn({tmp[i * 4 + 2], tmp[i * 4 + 3]});
          } else {
            atomicAdd(reinterpret_cast<float2 *>(buf_ptr + (row + 0) * N + col), float2({tmp[i * 4 + 0], tmp[i * 4 + 1]}));
            atomicAdd(reinterpret_cast<float2 *>(buf_ptr + (row + 8) * N + col), float2({tmp[i * 4 + 2], tmp[i * 4 + 3]}));
          }
        }
      }
    };

    if constexpr (C_N_MAJOR)
      epilogue_N_major();
    else
      epilogue_M_major();

    asm volatile("bar.sync 1, %0;" :: "r"(BLOCK_M) : "memory");
    if (warp_id == 0)
      asm volatile("tcgen05.dealloc.cta_group::1.sync.aligned.b32 %0, %1;" :: "r"(0), "r"(BLOCK_N * 2));
  }
}

template <
  int K,
  int BLOCK_M,
  int BLOCK_N,
  int BLOCK_K,
  int SPLIT_K,
  bool SWAP_AB,
  bool C_N_MAJOR,
  int NUM_STAGES
>
at::Tensor gemm_launch(
  const at::Tensor& A,
  const at::Tensor& B,
  const at::Tensor& SFA,
  const at::Tensor& SFB,
        at::Tensor& C,
        at::Tensor& buf
) {
  static_assert(BLOCK_K % 256 == 0);

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

  auto A_ptr   = reinterpret_cast<const char *>(A.data_ptr());
  auto B_ptr   = reinterpret_cast<const char *>(B.data_ptr());
  auto SFA_ptr = reinterpret_cast<const char *>(SFA.data_ptr());
  auto SFB_ptr = reinterpret_cast<const char *>(SFB.data_ptr());
  auto C_ptr   = reinterpret_cast<half *>(C.data_ptr());
  auto buf_ptr = buf.data_ptr<float>();

  int new_M = M;
  int new_N = N;
  if constexpr (SWAP_AB) {
    std::swap(A_ptr, B_ptr);
    std::swap(SFA_ptr, SFB_ptr);
    std::swap(new_M, new_N);
  }

  CUtensorMap A_tmap, B_tmap;
  init_AB_tmap(&A_tmap, A_ptr, new_M, K, BLOCK_M, BLOCK_K);
  init_AB_tmap(&B_tmap, B_ptr, new_N, K, BLOCK_N, BLOCK_K);

  dim3 grid(SPLIT_K, (new_M / BLOCK_M) * (new_N / BLOCK_N));
  int tb_size = BLOCK_M + 2 * WARP_SIZE;
  int AB_size = (BLOCK_M + BLOCK_N) * (BLOCK_K / 2);
  int SFAB_size = 128 * (BLOCK_K / 16) * 2;
  int smem_size = (AB_size + SFAB_size) * NUM_STAGES;

  auto this_kernel = kernel<K, BLOCK_M, BLOCK_N, BLOCK_K, SPLIT_K, C_N_MAJOR != SWAP_AB, NUM_STAGES>;
  if (smem_size > 48000)
    cudaFuncSetAttribute(this_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
  this_kernel<<<grid, tb_size, smem_size>>>(A_tmap, B_tmap, SFA_ptr, SFB_ptr, C_ptr, buf_ptr, new_M, new_N);

  if constexpr (SPLIT_K == 1)
    return C_N_MAJOR ? C : C.view({N, M, 1}).transpose(0, 1);
  else
    return C_N_MAJOR ? buf : buf.view({N, M, 1}).transpose(0, 1);
}

at::Tensor gemm(
  const at::Tensor& A,
  const at::Tensor& B,
  const at::Tensor& SFA,
  const at::Tensor& SFB,
        at::Tensor& C,
        at::Tensor& buf
) {
  const int K = A.size(1) * 2;
  const int M = A.size(0);

#define LAUNCH(K_, BLOCK_M, BLOCK_N, BLOCK_K, SPLIT_K, SWAP_AB, C_N_MAJOR, NUM_STAGES) \
  else if (K == K_) C = gemm_launch<K_, BLOCK_M, BLOCK_N, BLOCK_K, SPLIT_K, SWAP_AB, C_N_MAJOR, NUM_STAGES>(A, B, SFA, SFB, C, buf);

  if (false) {}
  else if (K == 7168) {
    if (M >= 512)
      C = gemm_launch<7168, 128, 128, 256, 1, true, true, 5>(A, B, SFA, SFB, C, buf);
    else
      C = gemm_launch<7168, 128, 64, 256, 1, true, true, 6>(A, B, SFA, SFB, C, buf);
  }
  else if (K == 4096) {
    if (M >= 512)
      C = gemm_launch<4096, 128, 128, 256, 1, true, true, 6>(A, B, SFA, SFB, C, buf);
    else
      C = gemm_launch<4096, 128, 64, 256, 1, true, true, 6>(A, B, SFA, SFB, C, buf);
  }
  LAUNCH(2048, 128, 64, 256, 1, true, true, 8)
  LAUNCH(2304, 128, 64, 256, 1, true, true, 6)
  LAUNCH(1536, 128, 64, 256, 1, true, true, 6)
  LAUNCH(512, 128, 64, 256, 1, true, true, 6)
  LAUNCH(256, 128, 64, 256, 1, true, true, 6)

#undef LAUNCH

  return C;
}

__global__ void silu_mul_kernel(const half2* g1, const half2* g2, half2* out, int64_t count2) {
  int64_t idx = static_cast<int64_t>(blockIdx.x) * blockDim.x + threadIdx.x;
  if (idx >= count2) {
    return;
  }
  half2 x2 = g1[idx];
  half2 y2 = g2[idx];
  float2 xf = __half22float2(x2);
  float2 yf = __half22float2(y2);
  float2 out_f;
  out_f.x = xf.x / (1.0f + expf(-xf.x));
  out_f.y = xf.y / (1.0f + expf(-xf.y));
  out[idx] = __floats2half2_rn(out_f.x * yf.x, out_f.y * yf.y);
}

at::Tensor silu_mul(
  const at::Tensor& g1,
  const at::Tensor& g2,
        at::Tensor& out
) {
  int64_t count = g1.numel();
  TORCH_CHECK((count & 1) == 0, "silu_mul 需要偶数元素");
  int64_t count2 = count / 2;
  int threads = 256;
  int blocks = static_cast<int>((count2 + threads - 1) / threads);
  silu_mul_kernel<<<blocks, threads>>>(
    reinterpret_cast<const half2 *>(g1.data_ptr()),
    reinterpret_cast<const half2 *>(g2.data_ptr()),
    reinterpret_cast<half2 *>(out.data_ptr()),
    count2
  );
  return out;
}

TORCH_LIBRARY(nvfp4_ops, m) {
  m.def("gemm(Tensor A, Tensor B, Tensor SFA, Tensor SFB, Tensor(a!) C, Tensor(b!) buf) -> Tensor");
  m.def("silu_mul(Tensor g1, Tensor g2, Tensor(a!) out) -> Tensor");
  m.impl("gemm", &gemm);
  m.impl("silu_mul", &silu_mul);
}
"""


_EXT_READY = False
_G2_CACHE = {}
_BUF_CACHE = {}
_SCALE_CACHE = {}
_SCALE_CACHE_KEYS = []
_SCALE_CACHE_MAX = 8


def _scale_cache_key(scale_p: torch.Tensor):
    return (
        scale_p.device,
        scale_p.dtype,
        tuple(scale_p.shape),
        tuple(scale_p.stride()),
        scale_p.data_ptr(),
    )


def _scale_cache_get(scale_p: torch.Tensor):
    key = _scale_cache_key(scale_p)
    cached = _SCALE_CACHE.get(key)
    if cached is None:
        return None
    if cached.device != scale_p.device:
        return None
    return cached


def _scale_cache_put(scale_p: torch.Tensor, linear: torch.Tensor):
    key = _scale_cache_key(scale_p)
    if key in _SCALE_CACHE:
        return linear
    if len(_SCALE_CACHE_KEYS) >= _SCALE_CACHE_MAX:
        old_key = _SCALE_CACHE_KEYS.pop(0)
        _SCALE_CACHE.pop(old_key, None)
    _SCALE_CACHE[key] = linear
    _SCALE_CACHE_KEYS.append(key)
    return linear


def _get_buf(device: torch.device) -> torch.Tensor:
    buf = _BUF_CACHE.get(device)
    if buf is None or buf.device != device:
        buf = torch.empty((1,), device=device, dtype=torch.float32)
        _BUF_CACHE[device] = buf
    return buf


def _get_g2(device: torch.device, m: int, n: int) -> torch.Tensor:
    key = (device, m, n)
    g2 = _G2_CACHE.get(key)
    if g2 is None or g2.device != device:
        g2 = torch.empty((m, n, 1), device=device, dtype=torch.float16)
        _G2_CACHE[key] = g2
    return g2


def _load_ext():
    global _EXT_READY
    if _EXT_READY:
        return
    if hasattr(torch.ops, "nvfp4_ops") and hasattr(torch.ops.nvfp4_ops, "gemm"):
        _EXT_READY = True
        return
    load_inline(
        name="nvfp4_dual_gemm_ext",
        cpp_sources="",
        cuda_sources=_CUDA_SRC_COMMON + _CUDA_SRC_V4,
        functions=None,
        with_cuda=True,
        verbose=False,
        is_python_module=False,
        extra_cflags=["-O3", "-std=c++17"],
        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"],
    )
    _EXT_READY = True


def _from_permuted_scale(scale_p: torch.Tensor) -> torch.Tensor:
    # 使用 permuted scale 生成 kernel 期望的线性布局，并缓存线性化结果
    if scale_p.dim() != 6:
        raise RuntimeError("scale_p 维度不符合预期")
    cached = _scale_cache_get(scale_p)
    if cached is not None:
        return cached
    scale_l = scale_p.select(5, 0)
    permuted = scale_l.permute(2, 4, 0, 1, 3).contiguous()
    linear = permuted.reshape(-1)
    return _scale_cache_put(scale_p, linear)


def _gemm(
    a: torch.Tensor,
    b: torch.Tensor,
    sfa: torch.Tensor,
    sfb: torch.Tensor,
    out: Optional[torch.Tensor] = None,
    buf: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    m = a.size(0)
    n = b.size(0)
    if out is None:
        out = torch.empty((m, n, 1), device=a.device, dtype=torch.float16)
    elif not out.is_contiguous():
        out = out.contiguous()
    if buf is None:
        buf = torch.empty((1,), device=a.device, dtype=torch.float32)
    return torch.ops.nvfp4_ops.gemm(a, b, sfa, sfb, out, buf)


def custom_kernel(data):
    a, b1, b2, _sfa, _sfb1, _sfb2, sfa_p, sfb1_p, sfb2_p, c = data
    _load_ext()
    out = c if c.is_contiguous() else c.contiguous()

    m = a.size(0)
    n = b1.size(0)
    k = a.size(1) * 2
    l = out.size(2)
    if l != 1:
        raise RuntimeError("仅支持 L=1")
    if k not in (7168, 4096, 2304, 2048, 1536, 512, 256):
        raise RuntimeError("仅支持预设 K 集合")

    a_l = a.select(2, 0)
    b1_l = b1.select(2, 0)
    b2_l = b2.select(2, 0)
    scale_a = _from_permuted_scale(sfa_p)
    scale_b1 = _from_permuted_scale(sfb1_p)
    scale_b2 = _from_permuted_scale(sfb2_p)

    # 缓存临时缓冲减少重复分配
    buf = _get_buf(a.device)
    # 复用 out 作为 g1 缓冲，减少一次分配与拷贝
    _gemm(a_l, b1_l, scale_a, scale_b1, out=out, buf=buf)
    g2 = _get_g2(a.device, m, n)
    g2 = _gemm(a_l, b2_l, scale_a, scale_b2, out=g2, buf=buf)

    out_slice = out.view(m, n)
    g2_view = g2.view(m, n)
    torch.ops.nvfp4_ops.silu_mul(out_slice, g2_view, out_slice)
    return out


__all__ = ["custom_kernel"]