UCP/API: device api

.gitignore

@@ -119,3 +119,4 @@ go.work
 go.work.sum
 bindings/go/pkg/
 bindings/go/go.sum
+*.d

config/m4/compiler.m4

@@ -59,6 +59,34 @@ AC_ARG_ENABLE([gcov],
         [enable_gcov=no])
 AM_CONDITIONAL([HAVE_GCOV],[test "x$enable_gcov" = xyes])
 
+#
+# Enable NVCC build
+#
+AC_ARG_WITH([nvcc],
+    AS_HELP_STRING([--with-nvcc], [Enable NVCC compiler support]))
+
+AS_IF([test "x$with_nvcc" != "xno"],
+    [
+    # Look for NVCC compiler
+    AC_PATH_PROG([NVCC], [nvcc], [no])
+    AS_IF([test "x$NVCC" != "xno"],
+        [
+        AC_DEFINE([HAVE_NVCC], [1], [Enable NVCC compiler support])
+        NVCC_FLAGS="-g -lineinfo"
+        AS_IF([test "x$enable_debug" = xyes],
+              [NVCC_FLAGS="$NVCC_FLAGS -G"])
+        AC_DEFINE_UNQUOTED([NVCC_PATH], ["$NVCC"], [Path to NVCC compiler])
+        AC_DEFINE_UNQUOTED([NVCC_FLAGS], ["$NVCC_FLAGS"], [NVCC compiler flags])
+        nvcc_happy=yes
+        ],
+        [
+        nvcc_happy=no
+        AS_IF([test "x$with_nvcc" = "xyes"],
+            [AC_MSG_ERROR([NVCC compiler support requested but nvcc not found])])
+        ])
+    ])
+AM_CONDITIONAL([HAVE_NVCC], [test "x$NVCC" != "xno"])
+
 
 #
 # Optimization level
@@ -71,9 +99,11 @@ AS_IF([test "x$enable_compiler_opt" = "xyes"], [BASE_CFLAGS="-O3 $BASE_CFLAGS"],
       [test "x$enable_compiler_opt" = "xnone"],
           [AS_IF([test "x$enable_debug" = xyes -o "x$enable_gcov" = xyes],
                  [BASE_CFLAGS="-O0 $BASE_CFLAGS"
-                  BASE_CXXFLAGS="-O0 $BASE_CXXFLAGS"],
+                  BASE_CXXFLAGS="-O0 $BASE_CXXFLAGS"
+                  NVCC_FLAGS="-O0 $NVCC_FLAGS"],
                  [BASE_CFLAGS="-O3 $BASE_CFLAGS"
-                  BASE_CXXFLAGS="-O0 $BASE_CXXFLAGS"])],
+                  BASE_CXXFLAGS="-O0 $BASE_CXXFLAGS"
+                  NVCC_FLAGS="-O3 $NVCC_FLAGS"])],
       [test "x$enable_compiler_opt" = "xno"], [],
       [BASE_CFLAGS="-O$enable_compiler_opt $BASE_CFLAGS"])
 
@@ -633,4 +663,7 @@ BASE_CPPFLAGS="-DCPU_FLAGS=\"$OPT_CFLAGS\""
 BASE_CPPFLAGS="$BASE_CPPFLAGS -I\${abs_top_srcdir}/src"
 BASE_CPPFLAGS="$BASE_CPPFLAGS -I\${abs_top_builddir}"
 BASE_CPPFLAGS="$BASE_CPPFLAGS -I\${abs_top_builddir}/src"
+NVCC_WRAP="\$(abs_top_srcdir)/contrib/nvcc_wrap.sh"
 AC_SUBST([BASE_CPPFLAGS], [$BASE_CPPFLAGS])
+AC_SUBST([NVCC_WRAP], [$NVCC_WRAP])
+AC_SUBST([NVCC_FLAGS], [$NVCC_FLAGS])

configure.ac

@@ -386,6 +386,7 @@ AC_CONFIG_FILES([
                  cmake/ucx-config.cmake
                  cmake/ucx-targets.cmake
                  config/Makefile
+                 src/tools/perf/gdaki/Makefile
                  ])
 
 AC_CONFIG_FILES([test/mpi/run_mpi.sh], [chmod a+x test/mpi/run_mpi.sh])
@@ -426,6 +427,10 @@ AC_MSG_NOTICE([        ROCM modules:   <$(echo ${uct_rocm_modules}|tr ':' ' ') >
 AC_MSG_NOTICE([          IB modules:   <$(echo ${uct_ib_modules}|tr ':' ' ') >])
 AC_MSG_NOTICE([         UCM modules:   <$(echo ${ucm_modules}|tr ':' ' ') >])
 AC_MSG_NOTICE([        Perf modules:   <$(echo ${ucx_perftest_modules}|tr ':' ' ') >])
+AS_IF([test "x$NVCC" != "xno"], [
+    AC_MSG_NOTICE([           NVCC path:   ${NVCC}])
+    AC_MSG_NOTICE([          NVCC flags:   ${NVCC_FLAGS}])
+])
 AS_IF([test "x$enable_ucg" != "xno"], [
     AC_MSG_NOTICE([       UCG modules:   <$(echo ${ucg_modules}|tr ':' ' ') >])])
 ])

contrib/configure-devel

@@ -18,6 +18,7 @@ $basedir/../configure \
 	--enable-profiling \
 	--enable-frame-pointer \
 	--enable-stats \
+	--enable-debug \
 	--enable-debug-data \
 	--enable-mt \
 	"$@"

contrib/gdaki/README.md

@@ -0,0 +1,26 @@
+## Usage
+### Initial
+
+```
+cp -r ./ucx-fork/contrib/gdaki ./
+
+cd gdaki
+source ./env.sh
+./build.sh
+
+sudo su
+source ./env.sh
+cd ucx
+./test/gtest/gtest --gtest_filter=*gdaki*:*ucp_batch*
+
+```
+### Incremental
+```
+ssh rock0X
+cd gdaki
+source ./env.sh
+sudo su
+rock01$ ./run_sample.sh -c 1.1.60.10
+rock10$ ./run_sample.sh
+
+```

contrib/gdaki/build.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+# Source ./env.sh
+
+set -e
+./clone_rdma.sh
+./clone_ucx.sh
+./clone_doca.sh
+
+# rdma already built
+
+./build_doca.sh
+./build_ucx.sh

contrib/gdaki/build_doca.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+
+set -e
+
+source env.sh
+mkdir doca/build
+cd doca/build
+
+meson setup --prefix=$(pwd)/install \
+    -Damalgamation_build=true -Denable_gpu_support=true ..
+
+ninja && ninja install

contrib/gdaki/build_ucx.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+
+# Must do: source ./env.sh
+
+set -e
+
+cd ucx-fork
+./autogen.sh
+./contrib/configure-devel --prefix=$(pwd)/rfs \
+    --with-verbs=$(pwd)/../rdma-core/build \
+    --with-cuda=$CUDA_HOME \
+    --with-gdrcopy=/opt/mellanox/gdrcopy
+
+make -j
+make install

contrib/gdaki/clone_doca.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+
+[ -d doca ] && { echo "doca directory already exists"; exit 1; }
+
+set -e
+
+git clone ssh://git-nbu.nvidia.com:12023/doca/doca
+
+#First commit, later reverted
+#git checkout -B verbs-gpunetio ea9a643c2f27fd790d6d2197bc6e10af055caf45
+
+cd doca
+# git fetch origin refs/changes/75/1253175/2
+git checkout -B master origin/master
+git am ../doca-00*.patch
+cd -

contrib/gdaki/clone_rdma.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+
+set -eu
+
+[ -d rdma-core ] && { echo "rdma-core already exists"; exit 1; }
+git clone https://github.com/linux-rdma/rdma-core.git
+cd rdma-core
+git checkout v57.0
+
+mkdir build
+cd build
+cmake -DNO_MAN_PAGES=1 -DCMAKE_INSTALL_PREFIX=$(pwd)/../rfs ../
+make -j
+make install
+cd ../..

contrib/gdaki/clone_ucx.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+git clone ssh://[email protected]:12051/ofarjon/ucx-fork.git
+cd ucx-fork
+git checkout origin/ak_wip
+git am <../ucx-0*.patch
+cd -

contrib/gdaki/doca-0001-Trim-dependencies.patch

@@ -0,0 +1,46 @@
+From bfa6205caef2e5500211142480e8804403ad90f5 Mon Sep 17 00:00:00 2001
+From: Thomas Vegas <[email protected]>
+Date: Wed, 11 Jun 2025 07:50:12 +0000
+Subject: [PATCH] Trim dependencies
+
+---
+ samples/doca_gpunetio/dependencies/meson.build | 3 ---
+ samples/doca_gpunetio/meson.build              | 8 --------
+ 2 files changed, 11 deletions(-)
+
+diff --git a/samples/doca_gpunetio/dependencies/meson.build b/samples/doca_gpunetio/dependencies/meson.build
+index fa18e01ed7..79d9854dae 100644
+--- a/samples/doca_gpunetio/dependencies/meson.build
++++ b/samples/doca_gpunetio/dependencies/meson.build
+@@ -26,8 +26,5 @@
+ # Needed by one of the samples (and by the lib itself) so it is OK to have it here
+ sample_lib_doca_depends += ['rdma']
+ sample_lib_doca_depends += ['rdma_verbs']
+-sample_lib_doca_depends += ['dpdk_bridge']
+ # Needed by some of the samples, and is roughly in the same components group, so it is fine
+-sample_lib_doca_depends += ['flow']
+-sample_lib_doca_depends += ['eth']
+ sample_lib_doca_depends += ['dma']
+diff --git a/samples/doca_gpunetio/meson.build b/samples/doca_gpunetio/meson.build
+index 76552dcdec..474723aa4f 100644
+--- a/samples/doca_gpunetio/meson.build
++++ b/samples/doca_gpunetio/meson.build
+@@ -23,15 +23,7 @@
+ #
+
+ samples = [
+-	'gpunetio_send_wait_time',
+-	'gpunetio_simple_receive',
+-	'gpunetio_simple_send',
+-	'gpunetio_rdma_client_server_write',
+-	'gpunetio_rdma_verbs_onesided_bw',
+-	'gpunetio_rdma_verbs_twosided',
+ 	'gpunetio_rdma_verbs_write_bw',
+-	'gpunetio_rdma_verbs_write_lat',
+-	'gpunetio_dma_memcpy',
+ ]
+
+ sample_lib_srcs = files([
+-- 
+2.31.1
+

contrib/gdaki/env.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+
+module load dev/cuda12.9.0
+module load dev/gdrcopy2.5_cuda12.9.0
+#module load hpcx-gcc
+
+sudo yum install -y glib2-devel libzip-devel json-c-devel \
+    libpcap-devel jsoncpp-devel ninja-build
+sudo yum install -y dpdk-devel
+
+if [ $(id -u) -ne 0 ]; then
+    sudo pip3 install meson>=0.61.2
+fi
+
+gdr=$(echo "$INCLUDE" | cut -d: -f 1)
+gdr_lib=$(echo "$LIBRARY_PATH" | cut -d: -f1)
+ucx=$(pwd)/ucx/rfs
+rdma=$(pwd)/rdma-core/rfs
+doca=$(pwd)/doca/build/install
+sudo ln -sf "$(dirname $gdr)" /opt/mellanox/gdrcopy
+
+export CFLAGS="-I$ucx/include -I$rdma/include -I$gdr -I$doca/include"
+export CPPFLAGS="-I$ucx/include -I$rdma/include -I$gdr -I$doca/include"
+export CXXFLAGS="-I$ucx/include -I$rdma/include -I$gdr -I$doca/include"
+export LDFLAGS="-L$ucx/lib -L$rdma/lib -L$gdr_lib -L$doca/lib64"
+export NVCC_FLAGS="-I$doca/include -I$rdma/include"
+export LD_LIBRARY_PATH="$rdma/lib:$doca/lib64:$LD_LIBRARY_PATH"

contrib/gdaki/run_sample.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+
+# Build DOCA with '0001-Trim-dependencies.patch' applied
+
+# Run on Rock as ROOT
+#./doca/build/samples/doca_gpunetio_rdma_verbs_write_bw -g 0000:a3:00 -d mlx5_0 -l 70 -c 1.1.60.1
+
+source ./env.sh
+./doca/build/samples/doca_gpunetio_rdma_verbs_write_bw -g 0000:a3:00 -d mlx5_0 -l 70 $*

contrib/gdaki/ucx-0001-GDAKI-build.patch

@@ -0,0 +1,69 @@
+From 6844172e352b9ccc78fd265ef5694d3f185b83fa Mon Sep 17 00:00:00 2001
+From: root <[email protected]>
+Date: Wed, 11 Jun 2025 15:24:04 +0000
+Subject: [PATCH] GDAKI build
+
+---
+ src/ucs/datastruct/mpool.inl        | 4 ++++
+ src/uct/cuda/gdaki/Makefile.am      | 2 +-
+ src/uct/cuda/gdaki/configure.m4     | 2 +-
+ src/uct/ib/mlx5/rc/rc_mlx5_common.h | 2 +-
+ 4 files changed, 7 insertions(+), 3 deletions(-)
+
+diff --git a/src/ucs/datastruct/mpool.inl b/src/ucs/datastruct/mpool.inl
+index 25404ec..054a518 100644
+--- a/src/ucs/datastruct/mpool.inl
++++ b/src/ucs/datastruct/mpool.inl
+@@ -35,6 +35,10 @@ static inline void *ucs_mpool_get_inline(ucs_mpool_t *mp)
+     return obj;
+ }
+
++#ifndef ENABLE_DEBUG_DATA
++#define ENABLE_DEBUG_DATA 0
++#endif
++
+ static inline void
+ ucs_mpool_add_to_freelist(ucs_mpool_t *mp, ucs_mpool_elem_t *elem)
+ {
+diff --git a/src/uct/cuda/gdaki/Makefile.am b/src/uct/cuda/gdaki/Makefile.am
+index fffb6ff..50445c8 100644
+--- a/src/uct/cuda/gdaki/Makefile.am
++++ b/src/uct/cuda/gdaki/Makefile.am
+@@ -14,7 +14,7 @@ libuct_cuda_gdaki_la_LIBADD   = $(top_builddir)/src/ucs/libucs.la \
+                                 $(top_builddir)/src/uct/ib/libuct_ib.la \
+                                 $(top_builddir)/src/uct/ib/mlx5/libuct_ib_mlx5.la \
+                                 $(top_builddir)/src/uct/cuda/libuct_cuda.la \
+-                                $(CUDA_LIBS) -ldoca_gpunetio
++                                $(CUDA_LIBS) -ldoca_gpunetio -ldoca_rdma_verbs
+
+ libuct_cuda_gdaki_la_SOURCES = \
+ 	gdaki_iface.c \
+diff --git a/src/uct/cuda/gdaki/configure.m4 b/src/uct/cuda/gdaki/configure.m4
+index b552d5f..b7dd6a5 100644
+--- a/src/uct/cuda/gdaki/configure.m4
++++ b/src/uct/cuda/gdaki/configure.m4
+@@ -4,7 +4,7 @@
+ #
+
+ AC_CHECK_HEADERS([doca_gpunetio.h], [
+-AC_CHECK_LIB([doca_gpunetio], [doca_gpu_create], [have_gpunetio=yes])])
++AC_CHECK_LIB([doca_gpunetio], [doca_gpu_create], [have_gpunetio=yes], [], [-ldoca_rdma_verbs -ldoca_common])])
+
+ AS_IF([test $have_gpunetio = yes], [
+     uct_cuda_modules="${uct_cuda_modules}:gdaki"
+diff --git a/src/uct/ib/mlx5/rc/rc_mlx5_common.h b/src/uct/ib/mlx5/rc/rc_mlx5_common.h
+index 22eed92..719cf6e 100644
+--- a/src/uct/ib/mlx5/rc/rc_mlx5_common.h
++++ b/src/uct/ib/mlx5/rc/rc_mlx5_common.h
+@@ -409,7 +409,7 @@ typedef struct uct_rc_mlx5_iface_common {
+         uint8_t                        atomic_fence_flag;
+         uct_rc_mlx5_srq_topo_t         srq_topo;
+         //uint8_t                        log_ack_req_freq;
+-       // uint8_t                        dp_ordering;
++        uint8_t                        dp_ordering;
+         //uint8_t                        dp_ordering_force;
+     } config;
+     UCS_STATS_NODE_DECLARE(stats)
+-- 
+2.31.1
+