# AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from cmath import nanj
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._C import _cuda_getCurrentRawStream as get_raw_stream

aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p


# kernel path: /tmp/torchinductor_paulzhan/tmpoh6x4mxl/du/cduc6gfuinf54m6dt2zocma7cftsn6qphgpnc76sxg3aipc24qe2.py
# Topologically Sorted Source Nodes: [], Original ATen: []
# Source node to ATen node mapping:
# Graph fragment:
#   %convert_element_type_2 : [num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%bmm, torch.float32), kwargs = {})
#   %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%convert_element_type_2, [0]), kwargs = {})
#   %convert_element_type_3 : [num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%sum_1, torch.float16), kwargs = {})
triton_per_fused_0 = async_compile.triton('triton_per_fused_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
triton_helpers.set_driver_to_gpu()

@triton_heuristics.persistent_reduction(
    size_hints={'x': 1024, 'r0_': 16},
    reduction_hint=ReductionHint.DEFAULT,
    filename=__file__,
    triton_meta={'signature': {'in_ptr0': '*fp16', 'out_ptr1': '*fp16', 'xnumel': 'i32', 'r0_numel': 'i32'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [AttrsDescriptor.from_dict({'arg_properties': {'tt.divisibility': (0, 1, 2, 3), 'tt.equal_to': ()}, 'cls': 'AttrsDescriptor'})]},
    inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_per_fused_0', 'mutated_arg_names': [], 'optimize_mem': False, 'no_x_dim': False, 'num_load': 1, 'num_reduction': 1, 'backend_hash': 'E5A2A8A2DA3940D7982306EB4DB6969673A9C6972732D60489D0F131EE13D9BF', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_per_fused_0(in_ptr0, out_ptr1, xnumel, r0_numel, XBLOCK : tl.constexpr):
    xnumel = 1024
    r0_numel = 16
    R0_BLOCK: tl.constexpr = 16
    rnumel = r0_numel
    RBLOCK: tl.constexpr = R0_BLOCK
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    r0_index = tl.arange(0, R0_BLOCK)[None, :]
    r0_offset = 0
    r0_mask = tl.full([XBLOCK, R0_BLOCK], True, tl.int1)
    roffset = r0_offset
    rindex = r0_index
    r0_1 = r0_index
    x0 = xindex
    tmp0 = tl.load(in_ptr0 + (x0 + 1024*r0_1), xmask, other=0.0).to(tl.float32)
    tmp1 = tmp0.to(tl.float32)
    tmp2 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK])
    tmp4 = tl.where(xmask, tmp2, 0)
    tmp5 = tl.sum(tmp4, 1)[:, None]
    tmp6 = tmp5.to(tl.float32)
    tl.store(out_ptr1 + (x0), tmp6, xmask)
''', device_str='cuda')


async_compile.wait(globals())
del async_compile

def call(args):
    benchmark_decompose_k_mm_arg0_1, benchmark_decompose_k_mm_arg1_1 = args
    args.clear()
    assert_size_stride(benchmark_decompose_k_mm_arg0_1, (32, 4096), (4096, 1))
    assert_size_stride(benchmark_decompose_k_mm_arg1_1, (4096, 32), (32, 1))
    with torch.cuda._DeviceGuard(0):
        torch.cuda.set_device(0)
        benchmark_decompose_k_mm_buf0 = empty_strided_cuda((16, 32, 32), (1024, 32, 1), torch.float16)
        # Topologically Sorted Source Nodes: [], Original ATen: []
        extern_kernels.bmm(reinterpret_tensor(benchmark_decompose_k_mm_arg0_1, (16, 32, 256), (256, 4096, 1), 0), reinterpret_tensor(benchmark_decompose_k_mm_arg1_1, (16, 256, 32), (8192, 32, 1), 0), out=benchmark_decompose_k_mm_buf0)
        del benchmark_decompose_k_mm_arg0_1
        del benchmark_decompose_k_mm_arg1_1
        benchmark_decompose_k_mm_buf2 = empty_strided_cuda((32, 32), (32, 1), torch.float16)
        # Topologically Sorted Source Nodes: [], Original ATen: []
        stream0 = get_raw_stream(0)
        triton_per_fused_0.run(benchmark_decompose_k_mm_buf0, benchmark_decompose_k_mm_buf2, 1024, 16, stream=stream0)
        del benchmark_decompose_k_mm_buf0
    return (benchmark_decompose_k_mm_buf2, )


def benchmark_compiled_module(times=10, repeat=10):
    from torch._dynamo.testing import rand_strided
    from torch._inductor.utils import print_performance
    benchmark_decompose_k_mm_arg0_1 = rand_strided((32, 4096), (4096, 1), device='cuda:0', dtype=torch.float16)
    benchmark_decompose_k_mm_arg1_1 = rand_strided((4096, 32), (32, 1), device='cuda:0', dtype=torch.float16)
    fn = lambda: call([benchmark_decompose_k_mm_arg0_1, benchmark_decompose_k_mm_arg1_1])
    return print_performance(fn, times=times, repeat=repeat)


if __name__ == "__main__":
    from torch._inductor.wrapper_benchmark import compiled_module_main
    compiled_module_main('None', benchmark_compiled_module)