submission 97155

#!POPCORN leaderboard nvfp4_gemv

import os
os.environ["TORCH_CUDA_ARCH_LIST"] = "10.0"

import torch
from torch.utils.cpp_extension import load_inline
from task import input_t, output_t

# Kernel configuration parameters
sf_vec_size = 16

gemv_cuda_source = r"""
#include<cuda_fp4.h>
#include<cuda_fp16.h>

#define M_BLOCK 4
#define FP4X2_PER_16B 16
#define FP8X2_PER_16B 8
#define K_BLOCK 32 * FP4X2_PER_16B
#define K_BLOCK_SMOL 32 * FP4X2_PER_16B / 16
#define ceilDiv(x, y) (((x) + (y) - 1) / (y))


template<int TILE_SIZE>
__device__ __forceinline__
void get_tile(int idx, int& tile_id, int& offset) {
    static_assert((TILE_SIZE & (TILE_SIZE - 1)) == 0, "Must be power of 2");

    constexpr int mask = TILE_SIZE - 1;
    constexpr int shift = __builtin_ctz(TILE_SIZE);

    tile_id = idx >> shift;
    offset  = idx & mask;
}


__device__ __forceinline__ __half2 fp4x2_e2m1_to_half2(__nv_fp4x2_e2m1 v) {
    __nv_fp4x2_storage_t raw = v.__x;  // packed 2×fp4
    __half2_raw hraw = __nv_cvt_fp4x2_to_halfraw2(raw, __NV_E2M1);
    return *reinterpret_cast<__half2*>(&hraw);
}

__device__ __forceinline__ __half2 fp8x2_e4m3_to_half2(__nv_fp8x2_e4m3 v) {
    __nv_fp8x2_storage_t raw = v.__x;
    __half2_raw hraw = __nv_cvt_fp8x2_to_halfraw2(raw, __NV_E4M3);
    return *reinterpret_cast<__half2*>(&hraw);
}

__device__ __forceinline__ __half fp8_e4m3_to_half(__nv_fp8_e4m3 v) {
    __nv_fp8_storage_t raw = v.__x;
    __half_raw hraw = __nv_cvt_fp8_to_halfraw(raw, __NV_E4M3);
    return *reinterpret_cast<__half*>(&hraw);
}

template<int M, int K>
__global__ void gemv_kernel(
		const __nv_fp4x2_e2m1* A, 
		const __nv_fp4x2_e2m1* B, 
    const __nv_fp8x2_e4m3* SFA,
    const __nv_fp8x2_e4m3* SFB,
		half* C
		) {
  int threadID = threadIdx.x;
  int warpID, laneID; 
  get_tile<32>(threadID, warpID, laneID);
  int rowID = warpID;

  constexpr int MK = M * K;
  constexpr int N = 128;
  constexpr int NK = N * K;
  constexpr int MK_SF = MK / 16;
  constexpr int NK_SF = NK / 16;
  constexpr int K_SF = K / 16;

  int blockRowIdx = blockIdx.x * M_BLOCK;
  int threadRowIdx = blockRowIdx + rowID;
  int batchBlockIdx = blockIdx.z;

  int batchOffset = MK * batchBlockIdx;
  int bBatchOffset = NK * batchBlockIdx;
  int rowOffset =  K * threadRowIdx;
  int cOffset = (M * batchBlockIdx + blockRowIdx);

  // scale factor offsets
  // Have K//16 fp8 values per row 
  // We are interpreting the pointer as fp8x2 so we have K//32 values per row
  int sfaBatchOffset = MK_SF * batchBlockIdx;
  int sfbBatchOffset = NK_SF * batchBlockIdx;
  int sfaRowOffset = K_SF * threadRowIdx;

  const __nv_fp4x2_e2m1 *gALanePtr = A + batchOffset + rowOffset + FP4X2_PER_16B * laneID;
  const __nv_fp8x2_e4m3 *gSFALanePtr = SFA + sfaBatchOffset + sfaRowOffset + laneID;

  const __nv_fp4x2_e2m1 *gBLanePtr = B + bBatchOffset + FP4X2_PER_16B * laneID;
  const __nv_fp8x2_e4m3 *gSFBLanePtr = SFB + sfbBatchOffset + laneID;

  __nv_fp4x2_e2m1 b_reg_fp4x2[16];
  __nv_fp4x2_e2m1 a_reg_fp4x2[16];
  //float2 a_reg_float2[16];
  //float2 b_reg_float2[16];
  __half2 a_reg_half2[16];
  __half2 b_reg_half2[16];
  uint4 * a_reg_ptr = reinterpret_cast<uint4 *>(&a_reg_fp4x2[0]);
  uint4 * b_reg_ptr = reinterpret_cast<uint4 *>(&b_reg_fp4x2[0]);

  __nv_fp8x2_e4m3 sfa_reg_fp8x2;
  __nv_fp8x2_e4m3 sfb_reg_fp8x2;

  int laneOffset = laneID * FP4X2_PER_16B;
  float final_accum = 0.0f;
  int smol_k = 0;
  for (int k_tile=0; k_tile<K; k_tile+=K_BLOCK) {
    bool in_range = laneOffset < K - k_tile;
    if (in_range) {
      // read 16B from global to reg
      const uint4 *gA_ptr = reinterpret_cast<const uint4 *>(gALanePtr + k_tile);
      const uint4 *gB_ptr = reinterpret_cast<const uint4 *>(gBLanePtr + k_tile);
      const __nv_fp8x2_e4m3 *gSFA_ptr = (gSFALanePtr + smol_k);
      const __nv_fp8x2_e4m3 *gSFB_ptr = (gSFBLanePtr + smol_k);

      // Read vals from global/shared to reg
      *a_reg_ptr = *gA_ptr;
      sfa_reg_fp8x2 = *gSFA_ptr;
      *b_reg_ptr = *gB_ptr;
      sfb_reg_fp8x2 = *gSFB_ptr;

      // Convert all vals to float
/*
#pragma unroll
      for (int j=0; j<16; ++j) {
        a_reg_float2[j] = __half22float2(fp4x2_e2m1_to_half2(a_reg_fp4x2[j]));
        b_reg_float2[j] = __half22float2(fp4x2_e2m1_to_half2(b_reg_fp4x2[j]));
      }
*/

      __half2 sfa_vals_h = (fp8x2_e4m3_to_half2(sfa_reg_fp8x2));
      __half2 sfb_vals_h = (fp8x2_e4m3_to_half2(sfb_reg_fp8x2));
      #pragma unroll
      for (int j=0; j<16; ++j) {
        a_reg_half2[j] = (fp4x2_e2m1_to_half2(a_reg_fp4x2[j]));
        b_reg_half2[j] = (fp4x2_e2m1_to_half2(b_reg_fp4x2[j]));
      }
      __half2 scale = __hmul2(sfa_vals_h, sfb_vals_h);

      /*
      float2 sfa_vals = __half22float2(fp8x2_e4m3_to_half2(sfa_reg_fp8x2));
      float2 sfb_vals = __half22float2(fp8x2_e4m3_to_half2(sfb_reg_fp8x2));
      float scale0 = sfa_vals.x * sfb_vals.x;
      float scale1 = sfa_vals.y * sfb_vals.y;
      float acc0 = 0.0f;
      float acc1 = 0.0f;
      */

      __half2 acc_h0  = __float2half2_rn(0.0f);
      __half2 acc_h1  = __float2half2_rn(0.0f);
      /*
      __half2 acc_h3  = __float2half2_rn(0.0f);
      __half2 acc_h4  = __float2half2_rn(0.0f);
      __half2 acc_h5  = __float2half2_rn(0.0f);
      __half2 acc_h6  = __float2half2_rn(0.0f);
      __half2 acc_h7  = __float2half2_rn(0.0f);
      __half2 acc_h8  = __float2half2_rn(0.0f);
	  acc_h0 = __hfma2(a_reg_half2[i], b_reg_half2[i], acc_h0);
	  acc_h1 = __hfma2(a_reg_half2[i], b_reg_half2[i], acc_h0);
	  acc_h2 = __hfma2(a_reg_half2[i], b_reg_half2[i], acc_h0);
	  acc_h3 = __hfma2(a_reg_half2[i], b_reg_half2[i], acc_h0);
	  acc_h4 = __hfma2(a_reg_half2[i], b_reg_half2[i], acc_h0);
	  acc_h5 = __hfma2(a_reg_half2[i], b_reg_half2[i], acc_h0);
	  acc_h6 = __hfma2(a_reg_half2[i], b_reg_half2[i], acc_h0);
	  acc_h7 = __hfma2(a_reg_half2[i], b_reg_half2[i], acc_h0);
      */

      #pragma unroll
      for (int i = 0; i < 8; ++i) {
        acc_h0 = __hfma2(a_reg_half2[i], b_reg_half2[i], acc_h0);
        acc_h1 = __hfma2(a_reg_half2[i+8], b_reg_half2[i+8], acc_h1);
      }
      __half2 scale0_h = __half2half2(__low2half(scale));
      __half2 scale1_h = __half2half2(__high2half(scale));
      acc_h0 = __hmul2(acc_h0, scale0_h);
      acc_h0 = __hfma2(acc_h1, scale1_h, acc_h0);
      float2 tmp = __half22float2(acc_h0);
      final_accum = final_accum + tmp.x + tmp.y;

      /*
      float2 tmp0 = __half22float2(acc_h0);
      float2 tmp1 = __half22float2(acc_h1);
      acc0 = __fmaf_rn(tmp0.x, scale0, acc0);
      acc0 = __fmaf_rn(tmp0.y, scale0, acc0);
      acc1 = __fmaf_rn(tmp1.x, scale1, acc1);
      acc1 = __fmaf_rn(tmp1.y, scale1, acc1);
      final_accum = final_accum + acc0 + acc1;
      */

      /*
      float acc0 = 0.0f;
      float acc1 = 0.0f;
      float* a_reg_float = reinterpret_cast<float *>(a_reg_float2);
      float* b_reg_float = reinterpret_cast<float *>(b_reg_float2);
      #pragma unroll
      for (int j=0; j<16; ++j) {
        acc0 = __fmaf_rn(a_reg_float[j], b_reg_float[j], acc0);
        //acc0 = __fmaf_rn(a_reg_float2[j].x, b_reg_float2[j].x, acc0);
        //acc0 = __fmaf_rn(a_reg_float2[j].y, b_reg_float2[j].y, acc0);
      }
      #pragma unroll
      for (int j=16; j<32; ++j) {
        acc1 = __fmaf_rn(a_reg_float[j], b_reg_float[j], acc1);
        //acc1 = __fmaf_rn(a_reg_float2[j].x, b_reg_float2[j].x, acc1);
        //acc1 = __fmaf_rn(a_reg_float2[j].y, b_reg_float2[j].y, acc1);
      }
      final_accum = __fmaf_rn(acc0, scale0, final_accum);
      final_accum = __fmaf_rn(acc1, scale1, final_accum);
      */
    }
    smol_k += K_BLOCK_SMOL;
  }
  // at this point each thread contains the sum of it's strided values in the row
  // need to use a warp reduction on each warp to compute final row sum
  constexpr unsigned FULL_MASK = 0xffffffff;
  for (int offset = 16; offset > 0; offset >>= 1) {
    final_accum += __shfl_down_sync(FULL_MASK, final_accum, offset);
  }
  if (laneID == 0) {
    C[cOffset + rowID] = __float2half(final_accum);
  }
}


template<int M, int K>
void launch_gemv(
const __nv_fp4x2_e2m1* A,
const __nv_fp4x2_e2m1* B,
const __nv_fp8x2_e4m3* SFA,
const __nv_fp8x2_e4m3* SFB,
half* C,
dim3 grid,
int threads)
{
    gemv_kernel<M, K><<<grid, threads>>>(A, B, SFA, SFB, C);
}


torch::Tensor gemv_cuda(torch::Tensor A, torch::Tensor B, torch::Tensor SFA, torch::Tensor SFB, torch::Tensor C) {
    //TORCH_CHECK(A.device().is_cuda(), "Tensor A must be a CUDA tensor");
    //TORCH_CHECK(B.device().is_cuda(), "Tensor B must be a CUDA tensor");
    //TORCH_CHECK(SFA.device().is_cuda(), "Tensor SFA must be a CUDA tensor");
    //TORCH_CHECK(SFB.device().is_cuda(), "Tensor SFB must be a CUDA tensor");
    //TORCH_CHECK(C.device().is_cuda(), "Tensor C must be a CUDA tensor");
    
    int M = A.size(0); 
    int K = A.size(1); 
    int L = A.size(2); 

    //dim3 block(M_BLOCK * 32, 1, 1);
    int threads = M_BLOCK * 32;
    dim3 grid(ceilDiv(M, M_BLOCK), 1, L);
    //printf("Problem size M: %d, K: %d, N: %d, L: %d \n", M, K, N, L);
    //printf("Threads per block: %d, Block dims (%d, 1, %d)\n", threads, grid.x, grid.z);

    auto A_ptr = reinterpret_cast<__nv_fp4x2_e2m1*>(A.data_ptr());
    auto B_ptr = reinterpret_cast<__nv_fp4x2_e2m1*>(B.data_ptr());
    auto SFA_ptr = reinterpret_cast<__nv_fp8x2_e4m3*>(SFA.data_ptr());
    auto SFB_ptr = reinterpret_cast<__nv_fp8x2_e4m3*>(SFB.data_ptr());
    auto C_ptr = reinterpret_cast<__half*>(C.data_ptr());
    
    if (M==128 && K==128) {
      launch_gemv<128, 128>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==128 && K==768) {
      launch_gemv<128, 768>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==128 && K==1536) {
      launch_gemv<128, 1536>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==256 && K==3584) {
      launch_gemv<256, 3584>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==2432 && K==2304) {
      launch_gemv<2432, 2304>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==384 && K==3584) {
      launch_gemv<384, 3584>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==512 && K==256) {
      launch_gemv<512, 256>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==512 && K==2048) {
      launch_gemv<512, 2048>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==512 && K==768) {
      launch_gemv<512, 768>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==7168 && K==8192) {
      launch_gemv<7168, 8192>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==4096 && K==3584) {
      launch_gemv<4096, 3584>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else if (M==7168 && K==1024) {
      launch_gemv<7168, 1024>(A_ptr, B_ptr, SFA_ptr, SFB_ptr, C_ptr, grid, threads);
    }
    else {
        throw std::runtime_error("Unsupported (M, K) combination");
    }

    cudaError_t err = cudaGetLastError();
    if (err != cudaSuccess) {
        throw std::runtime_error(cudaGetErrorString(err));
    }
    return C;
}
"""

gemv_cpp_source = """
#include <torch/extension.h>

torch::Tensor gemv_cuda(
  torch::Tensor A, 
  torch::Tensor B, 
  torch::Tensor SFA, 
  torch::Tensor SFB, 
  torch::Tensor C);
"""
extra_cuda_cflags = [
    "-O3",
    "--use_fast_math",
    "--fmad=true",
    "--ftz=true",
    "-Xcompiler", "-fno-strict-aliasing",

    # Aggressive math optimizations
    "-Xptxas=-O3",
    #"-Xptxas=--fastmath",

    # Cache behavior
    "-Xptxas=-dlcm=ca",

    # For debugging performance
    "-Xptxas=--warn-on-spills",
    "-Xptxas=-v",

    # Blackwell target
    "--gpu-architecture=sm_100a",
]

extra_cflags = [
    "-O3",
    "-ffast-math",
    "-fno-strict-aliasing",
]


gemv_module = load_inline(
    name='gemv_cuda',
    cpp_sources=gemv_cpp_source,
    cuda_sources=gemv_cuda_source,
    functions=['gemv_cuda'],
    verbose=True,
    extra_cuda_cflags=extra_cuda_cflags,
    extra_cflags=extra_cflags,
)



def gemv_cuda(A, B, SFA, SFB, C):
    if not A.is_cuda or not B.is_cuda or not SFA.is_cuda or not SFB.is_cuda or not C.is_cuda:
        raise RuntimeError("Both tensors must be on GPU")
    return gemv_module.gemv_cuda(A, B, SFA, SFB, C)


# Helper function for ceiling division
def ceil_div(a, b):
    return (a + b - 1) // b


def custom_kernel(
    data: input_t,
) -> output_t:
    """
    PyTorch reference implementation of NVFP4 block-scaled GEMV.
    """
    a_ref, b_ref, sfa, sfb, _, _, c_ref = data
    m, k, l = a_ref.shape
    n, k, l = b_ref.shape
    """
    print(f"K is {k}, n is {n}")
    print(f"A shape {a_ref.shape}")
    print(f"A shape {a_ref.stride()}")
    print(f"SFA shape {sfa.shape}")
    print(f"SFA shape {sfa.stride()}")
    print(f"B shape {b_ref.shape}")
    print(f"B shape {b_ref.stride()}")
    print(f"SFB shape {sfb.shape}")
    print(f"SFB shape {sfb.stride()}")
    print(f"C shape {c_ref.shape}")
    print(f"C shape {c_ref.stride()}")
    """

    # Get dimensions from MxNxL layout
    _, _, l = c_ref.shape
    #print(sfa.shape, sfa.stride())
    #print(f"SFA[0,0:32,0]: {sfa[0,:32,0].reshape(-1,2)}")
    gemv_cuda(a_ref, b_ref, sfa, sfb, c_ref)
    #torch.cuda.synchronize()
    #print(c_ref)
    return c_ref