# AOT ID: ['2_inference']
from ctypes import c_void_p, c_long
import torch
import math
import random
import os
import tempfile
from math import inf, nan
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

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
alloc_from_pool = torch.ops.inductor._alloc_from_pool
reinterpret_tensor = torch.ops.inductor._reinterpret_tensor
async_compile = AsyncCompile()


# kernel path: /tmp/torchinductor_root/o6/co6h7wbmqbtjafyje7vguzp77plihb4ndnxudln5qnxynjlsiu3w.py
# Source Nodes: [q, q_1, relative_emb], Original ATen: [aten._to_copy, aten.convolution, aten.div, aten.view]
# q => div
# q_1 => view_2
# relative_emb => convert_element_type, convert_element_type_1, convolution
triton_poi_fused__to_copy_convolution_div_view_0 = async_compile.triton('triton_', '''
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, instance_descriptor, DeviceProperties

@triton_heuristics.pointwise(
    size_hints=[256, 8192], tile_hint=TileHint.DEFAULT,
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp16', 2: '*fp16', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_convolution_div_view_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
    ynumel = 256
    xnumel = 4624
    yoffset = (tl.program_id(1) + tl.program_id(2) * tl.num_programs(1)) * YBLOCK
    yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
    ymask = yindex < ynumel
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    x1 = xindex
    y0 = yindex
    tmp0 = tl.load(in_ptr0 + (x1 + (4624*y0)), xmask & ymask, eviction_policy='evict_last').to(tl.float32)
    tmp1 = 0.0625
    tmp2 = tmp0 * tmp1
    tl.store(out_ptr0 + (x1 + (4624*y0)), tmp2, xmask & ymask)
    tl.store(out_ptr1 + (y0 + (256*x1)), tmp0, xmask & ymask)
''', device_str='cuda')

import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream


# kernel path: /tmp/torchinductor_root/r5/cr5jbru6ik4j52tm3h4oioj65zsxekjxi24xn5vg6hj2expq5hys.py
# Source Nodes: [qk_mask], Original ATen: [aten.rsub]
# qk_mask => sub
triton_poi_fused_rsub_1 = async_compile.triton('triton_', '''
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, instance_descriptor, DeviceProperties

@triton_heuristics.pointwise(
    size_hints=[1048576], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_rsub_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 0, 'num_reduction': 0, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_(out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 1040400
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex % 4624
    x1 = (xindex // 4624)
    x2 = xindex
    tmp0 = (-7) + (x0 // 68) + (x1 // 15)
    tmp1 = tl.full([1], 0, tl.int64)
    tmp2 = tmp0 >= tmp1
    tmp3 = tl.full([1], 68, tl.int64)
    tmp4 = tmp0 < tmp3
    tmp5 = (-7) + (x0 % 68) + (x1 % 15)
    tmp6 = tmp5 >= tmp1
    tmp7 = tmp5 < tmp3
    tmp8 = tmp2 & tmp4
    tmp9 = tmp8 & tmp6
    tmp10 = tmp9 & tmp7
    tmp11 = 1.0
    tmp12 = tl.full(tmp11.shape, 0.0, tmp11.dtype)
    tmp13 = tl.where(tmp10, tmp11, tmp12)
    tmp14 = tmp11 - tmp13
    tl.store(out_ptr0 + (x2), tmp14, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/ci/ccijl7m2wftc5tqd5nleuyw4dw26w77xs2cvgy6ouilvwke5mb4l.py
# Source Nodes: [relative_emb], Original ATen: [aten._to_copy]
# relative_emb => convert_element_type_1
triton_poi_fused__to_copy_2 = async_compile.triton('triton_', '''
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, instance_descriptor, DeviceProperties

@triton_heuristics.pointwise(
    size_hints=[65536], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '*fp16', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 57600
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex
    tmp0 = tl.load(in_ptr0 + (x0), xmask)
    tmp1 = tmp0.to(tl.float32)
    tl.store(out_ptr0 + (x0), tmp1, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/m3/cm3zqihrjigjaqqnkumbqnrntwqlyf2weue6tphmcjbl6dyg3y7p.py
# Source Nodes: [relative_emb], Original ATen: [aten._to_copy, aten.convolution]
# relative_emb => convert_element_type, convert_element_type_1, convolution
triton_poi_fused__to_copy_convolution_3 = async_compile.triton('triton_', '''
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, instance_descriptor, DeviceProperties

@triton_heuristics.pointwise(
    size_hints=[256, 8192], tile_hint=TileHint.DEFAULT,
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp32', 2: '*fp16', 3: 'i32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 4), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_convolution_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
    ynumel = 225
    xnumel = 4624
    yoffset = (tl.program_id(1) + tl.program_id(2) * tl.num_programs(1)) * YBLOCK
    yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
    ymask = yindex < ynumel
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    x1 = xindex
    y0 = yindex
    tmp0 = tl.load(in_ptr0 + (y0 + (225*x1)), xmask & ymask, eviction_policy='evict_last').to(tl.float32)
    tmp1 = tl.load(in_ptr1 + (y0), ymask, eviction_policy='evict_last')
    tmp2 = tmp1.to(tl.float32)
    tmp3 = tmp0 + tmp2
    tl.store(out_ptr0 + (x1 + (4672*y0)), tmp3, xmask & ymask)
''', device_str='cuda')


async_compile.wait(globals())
del async_compile

def call(args):
    arg0_1, arg1_1, arg2_1, arg3_1, arg4_1 = args
    args.clear()
    assert_size_stride(arg0_1, (1, 256, 68, 68), (1183744, 4624, 68, 1))
    assert_size_stride(arg1_1, (225, 256, 1, 1), (256, 1, 1, 1))
    assert_size_stride(arg2_1, (225, ), (1, ))
    assert_size_stride(arg3_1, (1, 256, 68, 68), (1183744, 4624, 68, 1))
    assert_size_stride(arg4_1, (1, 256, 68, 68), (1183744, 4624, 68, 1))
    with torch.cuda._DeviceGuard(0):
        torch.cuda.set_device(0)
        buf0 = empty_strided_cuda((1, 256, 68, 68), (1183744, 4624, 68, 1), torch.float16)
        buf3 = empty_strided_cuda((1, 256, 68, 68), (1183744, 1, 17408, 256), torch.float16)
        # Source Nodes: [q, q_1, relative_emb], Original ATen: [aten._to_copy, aten.convolution, aten.div, aten.view]
        stream0 = get_raw_stream(0)
        triton_poi_fused__to_copy_convolution_div_view_0.run(arg3_1, buf0, buf3, 256, 4624, grid=grid(256, 4624), stream=stream0)
        del arg3_1
        buf1 = empty_strided_cuda((1, 1, 225, 4624), (1040400, 1040400, 4624, 1), torch.float32)
        # Source Nodes: [qk_mask], Original ATen: [aten.rsub]
        triton_poi_fused_rsub_1.run(buf1, 1040400, grid=grid(1040400), stream=stream0)
        buf2 = empty_strided_cuda((225, 256, 1, 1), (256, 1, 1, 1), torch.float16)
        # Source Nodes: [relative_emb], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_2.run(arg1_1, buf2, 57600, grid=grid(57600), stream=stream0)
        del arg1_1
        # Source Nodes: [relative_emb], Original ATen: [aten._to_copy, aten.convolution]
        buf4 = extern_kernels.convolution(buf3, buf2, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None)
        assert_size_stride(buf4, (1, 225, 68, 68), (1040400, 1, 15300, 225))
        del buf2
        del buf3
        buf5 = empty_strided_cuda((1, 225, 68, 68), (1051200, 4672, 68, 1), torch.float16)
        # Source Nodes: [relative_emb], Original ATen: [aten._to_copy, aten.convolution]
        triton_poi_fused__to_copy_convolution_3.run(buf4, arg2_1, buf5, 225, 4624, grid=grid(225, 4624), stream=stream0)
        del arg2_1
        del buf4
    return (buf0, arg4_1, reinterpret_tensor(arg0_1, (1, 1, 256, 4624), (1183744, 1183744, 4624, 1), 0), buf1, reinterpret_tensor(buf5, (1, 1, 225, 4624), (0, 0, 4672, 1), 0), )


def benchmark_compiled_module(times=10, repeat=10):
    from torch._dynamo.testing import rand_strided
    from torch._inductor.utils import print_performance
    arg0_1 = rand_strided((1, 256, 68, 68), (1183744, 4624, 68, 1), device='cuda:0', dtype=torch.float16)
    arg1_1 = rand_strided((225, 256, 1, 1), (256, 1, 1, 1), device='cuda:0', dtype=torch.float32)
    arg2_1 = rand_strided((225, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg3_1 = rand_strided((1, 256, 68, 68), (1183744, 4624, 68, 1), device='cuda:0', dtype=torch.float16)
    arg4_1 = rand_strided((1, 256, 68, 68), (1183744, 4624, 68, 1), device='cuda:0', dtype=torch.float16)
    fn = lambda: call([arg0_1, arg1_1, arg2_1, arg3_1, arg4_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)