Artem-B

sccache stats: N/A No new compilation requests
+ PRESET=thrust-cpp20
+ test_preset Thrust thrust-cpp20
+ local BUILD_NAME=Thrust
+ local PRESET=thrust-cpp20
+ pushd ..
+ ctest --preset=thrust-cpp20
Test project /usr/local/google/home/tra/work/cccl/build/thrust-cpp20
        Start   1: thrust.cpp.cuda.cpp20.test.adjacent_difference

Artem-B

from itertools import product
from string import Template
from itertools import product, count
from string import Template
from absl import app
from typing import Sequence
types = [
    "char", "signed char", "char1", "char2", "char4", "unsigned char", "uchar1",
    "uchar2", "uchar4", "short", "short1", "short2", "short4", "ushort",

Artem-B

Do cudnn execution plan with plan tag: ConvFwd_Add_Add_ReluFwd_eng15_k5=1_k6=0_k7=1_k10=1
Workspace size in bytes: 409856
VariantPack: CUDNN_BACKEND_VARIANT_PACK_DESCRIPTOR : has 5 data pointers
I0509 11:27:06.030540 1504356 cuda_dnn.cc:4319]
Tensor_x: CUDNN_BACKEND_TENSOR_DESCRIPTOR : Datatype: CUDNN_DATA_FLOAT Id: 120 Alignment: 32 nDims 4 VectorCount: 1 vectorDimension -1 Dim [ 1,64,7,9 ] Str [ 4032,63,9,1 ] isVirtual: 0 isByValue: 0 reorder_type: CUDNN_TENSOR_REORDERING_NONE
Tensor_y: CUDNN_BACKEND_TENSOR_DESCRIPTOR : Datatype: CUDNN_DATA_FLOAT Id: 121 Alignment: 32 nDims 4 VectorCount: 1 vectorDimension -1 Dim [ 1,64,3,4 ] Str [ 768,12,4,1 ] isVirtual: 0 isByValue: 0 reorder_type: CUDNN_TENSOR_REORDERING_NONE
Tensor_z: CUDNN_BACKEND_TENSOR_DESCRIPTOR : Datatype: CUDNN_DATA_FLOAT Id: 122 Alignment: 32 nDims 4 VectorCount: 1 vectorDimension -1 Dim [ 1,64,3,4 ] Str [ 768,12,4,1 ] isVirtual: 0 isByValue: 0 reorder_type: CUDNN_TENSOR_REORDERING_NONE
Tensor_w: CUDNN_BACKEND_TENSOR_DESCRIPTOR : Datatype: CUDNN_

Artem-B

1687-Tag: ConvFwd_Add_Add_
1688-
1689-[cudnn_frontend] CUDNN_BACKEND_ENGINE_DESCRIPTOR : ID: 15 Has 4 knobs
1690-[cudnn_frontend] CUDNN_BACKEND_ENGINECFG_DESCRIPTOR : Number of knobs: 4
1691:[cudnn_frontend] CUDNN_BACKEND_EXECUTION_PLAN_DESCRIPTOR : ConvFwd_Add_Add_eng15_k5=1_k6=0_k7=1_k10=1, numeric_notes:[CUDNN_NUMERICAL_NOTE_WINOGRAD,CUDNN_NUMERICAL_NOTE_WINOGRAD_TILE_4x4,] behavior_notes:[] workSpaceSize: 895504
1692-[cudnn_frontend] CUDNN_BACKEND_VARIANT_PACK_DESCRIPTOR : has 5 data pointers
1693-[cudnn_frontend] CUDNN_BACKEND_TENSOR_DESCRIPTOR : Datatype: CUDNN_DATA_HALF Id: 120 Alignment: 32 nDims 4 VectorCount: 1 vectorDimension -1 Dim [ 1,96,44,60 ] Str [ 253440,2640,60,1 ] isVirtual: 0 isByValue: 0 reorder_type: CUDNN_TENSOR_REORDERING_NONE
1694-[cudnn_frontend] CUDNN_BACKEND_TENSOR_DESCRIPTOR : Datatype: CUDNN_DATA_HALF Id: 121 Alignment: 32 nDims 4 VectorCount: 1 vectorDimension -1 Dim [ 1,32,44,60 ] Str [ 84480,2640,60,1 ] isVirtual: 0 isByValue: 0 reorder_type: CUDNN_TENSOR_REORDERING_NONE
1695-

Artem-B

#! /usr/bin/bash

GOOD=0
BAD=0
N=100

for ((i=0; i<N; i++)); do
    glxinfo -B > /dev/null 2>/dev/null |:
    if [[ ${PIPESTATUS[0]} = 0 ]]; then
        ((GOOD++))

Artem-B

test

Artem-B

// ABI compatibility shims for  CUDA-11.7.
// Patch affected libraries with:
// objcopy \
// --redefine-sym cudaCreateTextureObject=cudaCreateTextureObject_v115 \
// --redefine-sym cudaGetTextureObjectTextureDesc=cudaGetTextureObjectTextureDesc_v115 \
// --redefine-sym cublasGetVersion_v2=cublasGetVersion_v2_v115 \
// --redefine-sym cublasLtGetVersion=cublasLtGetVersion_v115 \
// libnvinfer_static.a libcudnn_static.a
//

Artem-B

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>

#include "cuda_runtime.h"
#include "helper_cuda.h" // from cuda_samples

// Copy of texture descriptor from CUDA-11.7, so we can build the sample 
// in a way that simulates compilation with an older CUDA version.

Artem-B

Now and then video/sound stutters for about a second or two. I do have Hyper-V enabled and do use WSL2 (and docker, configured to use it), but in my case the issue does seem to happen when none of WSL2 VMs are running, so it's possible that it may be more of a Hyper-V issue than WSL2 itself.

I've managed to capture it with ETW trace. As far as I can tell, during this time, everything stalls for about 50ms, then resumes for 10ms and this cycle continues.

Absolutely no events get captured during the quiet periods. Here's one example:

Artem-B

; Reproducer for a bad performance regression triggered by switch to the new PM.
; `barney` ended up with the local variables not being optimized away and that
; had rather dramatic effect on some GPU code. See
; https://bugs.llvm.org/show_bug.cgi?id=52037 for the gory details.
;
; NOTE that opt -O3 produces different IR.
;
; RUN: opt -mtriple=nvptx64-nvidia-cuda -passes='default<O3>' -S %s -o - \
; RUN:  | llc -mtriple=nvptx64-nvidia-cuda -mcpu=sm_70 -O3  -o - \
; RUN:  | FileCheck %s