tenpercent

--- a/dnn-providers/hip-kernel-provider/rocke/platform/cpp/core/lower_llvm/core.cpp
+++ b/dnn-providers/hip-kernel-provider/rocke/platform/cpp/core/lower_llvm/core.cpp
@@ -277,6 +277,7 @@
     return found;
 }
 
+#ifndef _WIN32 /* POSIX-only: used solely by the dirent scans below */
 /* Python runtime_coexistence._version_key, as a comparison: compare the runs of
  * digits in each name as an integer sequence so "rocm-7.10" sorts NEWER than
  * "rocm-7.2" (a plain strcmp gets this backwards, because '1' < '2'). Returns

tenpercent

#!/bin/bash
# Which kernels changed between two PR heads? Instruments the suite's own _lower
# to dump IR, runs it at both heads, and diffs. Catches collateral damage from a
# "small fix" that touches many production files.
#
# Env: OLD_SRC, OLD_BUILD, NEW_SRC, NEW_BUILD, PYDEPS
: "${OLD_SRC:?}"; : "${NEW_SRC:?}"
instrument(){ python3 - "$1" "$2" <<'PY'
import pathlib, sys
d, tag = sys.argv[1], sys.argv[2]

tenpercent

#!/usr/bin/env python3
"""Emit IR from BOTH rocke engines for the affected cases. Writes $ROCKE_WORK/repro2/*.ll

Env: ROCKE_WORK with subdirs  src/ (rocke/platform @ f19a5b09)  build/ (rocke_engine*.so)
"""
import os, sys
W = os.environ["ROCKE_WORK"]
sys.path.insert(0, f"{W}/build"); sys.path.insert(0, f"{W}/src/python")
import rocke_engine
from rocke.core.ir import IRBuilder, PtrType, F16, I32

tenpercent

#!/bin/bash
# Build the rocke C++ engine + pybind extension from PR #9583 head.
# Env: ROCKE_SRC = <checkout>/dnn-providers/hip-kernel-provider/rocke/platform
#      OUT       = build output dir
set -eo pipefail
ROCKE_SRC="${ROCKE_SRC:?set ROCKE_SRC to .../rocke/platform}"
OUT="${OUT:-/tmp/rocke_build}"
PYDEPS="${PYDEPS:-/tmp/rocke_pydeps}"

export PYTHONPATH="$PYDEPS:$PYTHONPATH"

tenpercent

[CKBatchedGemmOperation(a_layout='Row', b_layout='Col', ds_layouts=(), c_layout='Row', a_element_dtype='F8', b_element_dtype='F8', ds_element_dtypes=(), c_element_dtype='BF16', acc_dtype='F32', c_shuffle_dtype='BF16', a_elementwise_op='PassThrough', b_elementwise_op='PassThrough', c_elementwise_op='PassThrough', gemm_specialization='GemmSpecialization::Default', block_size=256, m_per_block=256, n_per_block=256, k_per_block=64, a_k1=16, b_k1=16, m_per_xdl=32, n_per_xdl=32, m_xdl_per_wave=4, n_xdl_per_wave=4, a_block_transfer_thread_cluster_lengths_ak0_m_ak1=(4, 64, 1), a_block_transfer_thread_cluster_arrange_order=(1, 0, 2), a_block_transfer_src_access_order=(1, 0, 2), a_block_transfer_src_vector_dim=2, a_block_transfer_src_scalar_per_vector=16, a_block_transfer_dst_scalar_per_vector_ak1=16, a_block_lds_extra_m=0, b_block_transfer_thread_cluster_lengths_bk0_n_bk1=(4, 64, 1), b_block_transfer_thread_cluster_arrange_order=(1, 0, 2), b_block_transfer_src_access_order=(1, 0, 2), b_block_transfer_src_vector_dim=2, 

tenpercent

# Edit this configuration file to define what should be installed on
# your system.  Help is available in the configuration.nix(5) man page
# and in the NixOS manual (accessible by running ‘nixos-help’).

{ config, pkgs, ... }:

let
  nvidia-offload = pkgs.writeShellScriptBin "nvidia-offload" ''
    export __NV_PRIME_RENDER_OFFLOAD=1
    export __NV_PRIME_RENDER_OFFLOAD_PROVIDER=NVIDIA-G0

tenpercent

name: ait
channels:
  - pytorch
  - nvidia
  - conda-forge
  - defaults
dependencies:
  - _libgcc_mutex=0.1=conda_forge
  - _openmp_mutex=4.5=2_gnu
  - binutils_impl_linux-64=2.38=h2a08ee3_1

tenpercent

{
  "channels": [
    "intel",
    "defaults"
  ],
  "dependencies": [
    "_libgcc_mutex=0.1=main",
    "_pytorch_select=0.2=gpu_0",
    "absl-py=0.8.1=py36_0",
    "asn1crypto=0.24.0=py36_3",

tenpercent

TORCH_VER="20191208"
pip uninstall -y torch torchvision
rm -r wheels
mkdir -p wheels
for f in {torch,torch_xla,torchvision}; do 
  whl="${f}-nightly+${TORCH_VER}-cp36-cp36m-linux_x86_64.whl";
  gsutil cp "gs://tpu-pytorch/wheels/${whl}" wheels/;
  pip install "wheels/${whl}";
done
apt-get install -y libomp5

tenpercent

class MUniquePointerPrinter:
    """ Print a unique_ptr with pointed-to-value
        credits: https://gcc.gnu.org/ml/libstdc++/2013-04/msg00098.html 
    """

    def __init__ (self, typename, val):
        self.val = val

    def to_string (self):
        v = self.val['_M_t']['_M_head_impl']