submission 614359

import os
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"

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

cuda_source = r"""
#include <torch/extension.h>
#include <cuda_runtime.h>
#include <cstdint>

__global__ __launch_bounds__(512)
void histogram_custom(
    const uint8_t* __restrict__ data,
    int64_t* __restrict__ output,
    const int64_t n
) {
    __shared__ uint32_t smem[256];

    // Fused zero: block 0 zeros output using vectorized int4 writes
    if (blockIdx.x == 0 && threadIdx.x < 128) {
        ((int4*)output)[threadIdx.x] = make_int4(0, 0, 0, 0);
    }

    // Zero shared memory (256 uint32s, only first 256 threads)
    if (threadIdx.x < 256) {
        smem[threadIdx.x] = 0;
    }
    __syncthreads();

    const int64_t tid = blockIdx.x * 512 + threadIdx.x;
    const int64_t grid_stride = 512LL * gridDim.x;
    const int64_t n16 = n >> 4;
    const uint4* data_vec = (const uint4*)data;

    // Software pipelining: load next while processing current
    int64_t i = tid;
    if (i < n16) {
        uint4 vals = __ldg(&data_vec[i]);
        for (; i + grid_stride < n16; i += grid_stride) {
            uint4 next = __ldg(&data_vec[i + grid_stride]);

            atomicAdd(&smem[(vals.x      ) & 0xFF], 1u);
            atomicAdd(&smem[(vals.x >>  8) & 0xFF], 1u);
            atomicAdd(&smem[(vals.x >> 16) & 0xFF], 1u);
            atomicAdd(&smem[(vals.x >> 24)       ], 1u);
            atomicAdd(&smem[(vals.y      ) & 0xFF], 1u);
            atomicAdd(&smem[(vals.y >>  8) & 0xFF], 1u);
            atomicAdd(&smem[(vals.y >> 16) & 0xFF], 1u);
            atomicAdd(&smem[(vals.y >> 24)       ], 1u);
            atomicAdd(&smem[(vals.z      ) & 0xFF], 1u);
            atomicAdd(&smem[(vals.z >>  8) & 0xFF], 1u);
            atomicAdd(&smem[(vals.z >> 16) & 0xFF], 1u);
            atomicAdd(&smem[(vals.z >> 24)       ], 1u);
            atomicAdd(&smem[(vals.w      ) & 0xFF], 1u);
            atomicAdd(&smem[(vals.w >>  8) & 0xFF], 1u);
            atomicAdd(&smem[(vals.w >> 16) & 0xFF], 1u);
            atomicAdd(&smem[(vals.w >> 24)       ], 1u);

            vals = next;
        }
        // Process last chunk
        atomicAdd(&smem[(vals.x      ) & 0xFF], 1u);
        atomicAdd(&smem[(vals.x >>  8) & 0xFF], 1u);
        atomicAdd(&smem[(vals.x >> 16) & 0xFF], 1u);
        atomicAdd(&smem[(vals.x >> 24)       ], 1u);
        atomicAdd(&smem[(vals.y      ) & 0xFF], 1u);
        atomicAdd(&smem[(vals.y >>  8) & 0xFF], 1u);
        atomicAdd(&smem[(vals.y >> 16) & 0xFF], 1u);
        atomicAdd(&smem[(vals.y >> 24)       ], 1u);
        atomicAdd(&smem[(vals.z      ) & 0xFF], 1u);
        atomicAdd(&smem[(vals.z >>  8) & 0xFF], 1u);
        atomicAdd(&smem[(vals.z >> 16) & 0xFF], 1u);
        atomicAdd(&smem[(vals.z >> 24)       ], 1u);
        atomicAdd(&smem[(vals.w      ) & 0xFF], 1u);
        atomicAdd(&smem[(vals.w >>  8) & 0xFF], 1u);
        atomicAdd(&smem[(vals.w >> 16) & 0xFF], 1u);
        atomicAdd(&smem[(vals.w >> 24)       ], 1u);
    }

    {
        int64_t tail_start = n16 * 16;
        for (int64_t j = tail_start + tid; j < n; j += grid_stride) {
            atomicAdd(&smem[__ldg(&data[j])], 1u);
        }
    }

    __syncthreads();

    if (threadIdx.x < 256) {
        uint32_t v = smem[threadIdx.x];
        if (v > 0) {
            atomicAdd((unsigned long long*)&output[threadIdx.x], (unsigned long long)v);
        }
    }
}

static int _sm_count = -1;

int get_sm_count() {
    if (_sm_count < 0) {
        cudaDeviceProp prop;
        cudaGetDeviceProperties(&prop, 0);
        _sm_count = prop.multiProcessorCount;
    }
    return _sm_count;
}

torch::Tensor histogram_cuda(torch::Tensor data, torch::Tensor output) {
    const int64_t n = data.numel();

    histogram_custom<<<get_sm_count(), 512>>>(
        data.data_ptr<uint8_t>(),
        output.data_ptr<int64_t>(),
        n
    );

    return output;
}
""";

cpp_source = r"""
torch::Tensor histogram_cuda(torch::Tensor data, torch::Tensor output);
""";

module = load_inline(
    name="histogram_submit_v4",
    cpp_sources=[cpp_source],
    cuda_sources=[cuda_source],
    functions=["histogram_cuda"],
    verbose=False,
    extra_cuda_cflags=["-O3", "--use_fast_math"],
)

_wd = torch.randint(0, 256, (1024,), device="cuda", dtype=torch.uint8)
_wo = torch.zeros(256, device="cuda", dtype=torch.int64)
module.histogram_cuda(_wd, _wo)
torch.cuda.synchronize()
del _wd, _wo


def custom_kernel(data: input_t) -> output_t:
    data_tensor, output = data
    module.histogram_cuda(data_tensor, output)
    return output