davidberard98

// BUILD COMMAND:
// LD_LIBRARY_PATH=/usr/local/cuda-11.6/extras/CUPTI/lib64:$LD_LIBRARY_PATH nvcc -arch=sm_80 -std=c++17 -o cudagraph cudagraph.cu -lcupti
#include <cstddef>
#include <cuda_runtime_api.h>
#include <cstdio>
#include <sys/time.h>
#include <iostream>
#include <cupti.h>

#define N 500000 // tuned such that kernel takes a few microseconds

davidberard98

import torch
import torch._dynamo
import torch._inductor.inductor_prims

def fn(values, boundaries):
    return torch.bucketize(values, boundaries)

def fn_ind(values, boundaries):
    return torch.ops.prims._inductor_bucketize(values, boundaries)

davidberard98

# $CUDA_HOME/bin/nvcc binary_search_cuda.cu -std=c++17 -o binary_search_cuda -O3 # -Wl,-rpath $CUDA_HOME/lib64
$CUDA_HOME/bin/nvcc dense_to_jagged.cu -std=c++17 -o dense_to_jagged -O3 # -Wl,-rpath $CUDA_HOME/lib64

davidberard98

import torch
import triton
import triton.language as tl

@triton.jit
def dense_to_jagged_triton(
    in_ptr,
    offsets_ptr,
    inverse_offsets_ptr,
    out_ptr,

davidberard98

import torch
import time

profiler_events = []
is_enabled = False
def _start_fn(name, args = None):
    if is_enabled:
        profiler_events.append((name, args, time.time()))

def _stop_fn():

davidberard98

import torch
import triton
import triton.language as tl

@triton.jit
def dense_to_jagged_triton(
    inverse_offsets_ptr, offsets_ptr, dense_ptr, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
    # xnumel = 33106688
    xoffset = tl.program_id(0) * XBLOCK

davidberard98

import torch

def fn(x, y):
    return torch.cat([x + y, y]).sin()

a = torch.ones((1024, 256), dtype=torch.float32)
b = torch.ones((1024, 256), dtype=torch.float32) * 2

with torch.profiler.profile(schedule=torch.profiler.schedule(wait=2, warmup=2, repeat=1, active=2), record_shapes=True) as prof:
    for _ in range(8):

davidberard98

Traceback (most recent call last):
  File "/data/users/dberard/scripts/oncall/112489.py", line 8, in <module>
    fn_opt(*inputs)
  File "/data/users/dberard/pytorch/torch/_dynamo/eval_frame.py", line 411, in _fn
    return fn(*args, **kwargs)
  File "/data/users/dberard/pytorch/torch/_dynamo/eval_frame.py", line 559, in catch_errors
    return callback(frame, cache_entry, hooks, frame_state)
  File "/data/users/dberard/pytorch/torch/_dynamo/convert_frame.py", line 687, in _convert_frame
    result = inner_convert(frame, cache_entry, hooks, frame_state)
  File "/data/users/dberard/pytorch/torch/_dynamo/convert_frame.py", line 148, in _fn

davidberard98

/data/users/dberard/scripts/oncall/112494.py:6: UserWarning: An output with one or more elements was resized since it had shape [10, 9, 8], which does not match the required output shape [1, 9, 8]. This behavior is deprecated, and in a future PyTorch release outputs will not be resized unless they have zero elements. You can explicitly reuse an out tensor t by resizing it, inplace, to zero elements with t.resize_(0). (Triggered internally at ../aten/src/ATen/native/Resize.cpp:28.)
  x = torch.var(correction=4, dim=0, input=x, keepdim=True, out=torch.rand_like(x))
/data/users/dberard/pytorch/torch/_prims_common/wrappers.py:159: UserWarning: An output with one or more elements was resized since it had shape torch.Size([s0, s1, s2]) which does not match the required output shape {str(shape)}. This behavior is deprecated, and in a future PyTorch release outputs will not be resized unless they have zero elements. You can explicitly reuse an out tensor t by resizing it, inplace, to zero elements with t.resize_(0).

davidberard98

/data/users/dberard/scripts/oncall/112502.py:7: UserWarning: An output with one or more elements was resized since it had shape [9, 10], which does not match the required output shape [9]. This behavior is deprecated, and in a future PyTorch release outputs will not be resized unless they have zero elements. You can explicitly reuse an out tensor t by resizing it, inplace, to zero elements with t.resize_(0). (Triggered internally at ../aten/src/ATen/native/Resize.cpp:28.)
  x = torch.diag(input=x, diagonal=0,out=torch.rand([9, 10], dtype=torch.float32).to('cpu'))
build succeded
/data/users/dberard/pytorch/torch/_prims_common/wrappers.py:159: UserWarning: An output with one or more elements was resized since it had shape torch.Size([9, 10]) which does not match the required output shape {str(shape)}. This behavior is deprecated, and in a future PyTorch release outputs will not be resized unless they have zero elements. You can explicitly reuse an out tensor t by resizing it, inplace, to zero elements with t.re