{lib}[GCCcore/12.2.0] UCC-CUDA v1.1.0 + add patch for UCC 1.1.0 for multiple component paths

easybuild/easyconfigs/u/UCC-CUDA/UCC-CUDA-1.1.0-GCCcore-12.2.0-CUDA-12.0.0.eb

@@ -0,0 +1,57 @@
+easyblock = 'ConfigureMake'
+
+name = 'UCC-CUDA'
+version = '1.1.0'
+versionsuffix = '-CUDA-%(cudaver)s'
+
+homepage = 'https://www.openucx.org/'
+description = """UCC (Unified Collective Communication) is a collective
+communication operations API and library that is flexible, complete, and 
+feature-rich for current and emerging programming models and runtimes.
+
+This module adds the UCC CUDA support.
+"""
+
+toolchain = {'name': 'GCCcore', 'version': '12.2.0'}
+toolchainopts = {'pic': True}
+
+source_urls = ['https://github.com/openucx/ucc/archive/refs/tags']
+sources = ['v%(version)s.tar.gz']
+patches = [
+    '%(name)s-1.0.0_link_against_existing_UCC_libs.patch',
+    '%(name)s-%(version)s_cuda_12_mem_ops.patch',
+]
+checksums = [
+    {'v1.1.0.tar.gz': '74c8ba75037b5bd88cb703e8c8ae55639af3fecfd4428912a433c010c97b4df7'},
+    {'UCC-CUDA-1.0.0_link_against_existing_UCC_libs.patch':
+     '9fa11cf6779174f4e9048df5812096e4261e1769d465cc7f34a6354398876856'},
+    {'UCC-CUDA-1.1.0_cuda_12_mem_ops.patch': 'fc3ea1487d29dc626db2363ef5a79e7f0906f6a7507a363fa6167a812b143eb6'},
+]
+
+builddependencies = [
+    ('binutils', '2.39'),
+    ('Autotools', '20220317'),
+]
+
+dependencies = [
+    ('UCC', '1.1.0'),
+    ('CUDA',  '12.0.0', '', SYSTEM),
+    ('UCX-CUDA', '1.13.1', '-CUDA-%(cudaver)s'),
+    ('NCCL', '2.16.2', '-CUDA-%(cudaver)s'),
+]
+
+preconfigopts = "./autogen.sh && "
+
+buildopts = '-C src/components/mc/cuda V=1 && make -C src/components/tl/nccl V=1'
+installopts = '-C src/components/mc/cuda && make -C src/components/tl/nccl install'
+
+sanity_check_paths = {
+    'files': ['lib/ucc/libucc_mc_cuda.%s' % SHLIB_EXT, 'lib/ucc/libucc_tl_nccl.%s' % SHLIB_EXT],
+    'dirs': ['lib']
+}
+
+sanity_check_commands = ["ucc_info -c"]
+
+modextrapaths = {'EB_UCC_EXTRA_COMPONENT_PATH': 'lib/ucc'}
+
+moduleclass = 'lib'

easybuild/easyconfigs/u/UCC-CUDA/UCC-CUDA-1.1.0_cuda_12_mem_ops.patch

@@ -0,0 +1,99 @@
+Backported fix for CUDA 12 https://github.com/openucx/ucc/pull/700
+Essentially just removes the deprecated checks for CUDA MEM OPS as they are required in CUDA 12
+author: [email protected]
+--- src/components/ec/cuda/ec_cuda.c.orig	2023-02-02 18:44:36.085221084 +0000
++++ src/components/ec/cuda/ec_cuda.c	2023-02-02 18:47:23.726819030 +0000
+@@ -205,11 +205,10 @@
+ {
+     ucc_ec_cuda_config_t *cfg = EC_CUDA_CONFIG;
+     ucc_status_t          status;
+-    int                   device, num_devices, attr;
++    int                   device, num_devices;
+     CUdevice              cu_dev;
+     CUresult              cu_st;
+     cudaError_t           cuda_st;
+-    const char           *cu_err_st_str;
+
+     ucc_ec_cuda.stream                   = NULL;
+     ucc_ec_cuda.stream_initialized       = 0;
+@@ -272,9 +271,14 @@
+     } else {
+         ucc_ec_cuda.strm_task_mode = UCC_EC_CUDA_TASK_MEM_OPS;
+         ucc_ec_cuda.post_strm_task = ucc_ec_cuda_post_driver_stream_task;
++#if CUDA_VERSION < 12000
++        CUresult cu_st;
++        CUdevice cu_dev;
++        int attr;
+
+         cu_st = cuCtxGetDevice(&cu_dev);
+         if (cu_st != CUDA_SUCCESS){
++            const char *cu_err_st_str;
+             cuGetErrorString(cu_st, &cu_err_st_str);
+             ec_debug(&ucc_ec_cuda.super, "cuCtxGetDevice() failed: %s",
+                      cu_err_st_str);
+@@ -297,6 +301,7 @@
+                      "CUDA MEM OPS are not supported or disabled");
+             return UCC_ERR_NOT_SUPPORTED;
+         }
++#endif
+     }
+     ucc_ec_cuda.task_strm_type = cfg->task_strm_type;
+     ucc_spinlock_init(&ucc_ec_cuda.init_spinlock, 0);
+--- src/components/tl/nccl/tl_nccl_context.c.orig	2023-02-03 15:17:09.358881676 +0000
++++ src/components/tl/nccl/tl_nccl_context.c	2023-02-03 17:04:31.680185749 +0000
+@@ -101,13 +101,14 @@
+         ucc_derived_of(config, ucc_tl_nccl_context_config_t);
+     int mem_ops_attr = 0;
+     ucc_status_t status;
+-    CUresult cu_st;
+-    CUdevice cu_dev;
+
+     UCC_CLASS_CALL_SUPER_INIT(ucc_tl_context_t, &tl_nccl_config->super,
+                               params->context);
+     memcpy(&self->cfg, tl_nccl_config, sizeof(*tl_nccl_config));
+     if (self->cfg.sync_type != UCC_TL_NCCL_COMPLETION_SYNC_TYPE_EVENT) {
++#if CUDA_VERSION < 12000
++        CUresult cu_st;
++        CUdevice cu_dev;
+         cu_st = cuCtxGetDevice(&cu_dev);
+         if (cu_st == CUDA_SUCCESS) {
+             cu_st = cuDeviceGetAttribute(&mem_ops_attr,
+@@ -116,6 +117,9 @@
+         } else {
+             tl_info(self->super.super.lib, "failed to get cuda device");
+         }
++#else
++        mem_ops_attr = 1;
++#endif
+         if (mem_ops_attr == 0) {
+             if (self->cfg.sync_type == UCC_TL_NCCL_COMPLETION_SYNC_TYPE_MEMOPS) {
+                 tl_error(self->super.super.lib, "memops not supported");
+--- config/m4/cuda.m4.orig	2023-02-03 17:04:44.367155175 +0000
++++ config/m4/cuda.m4	2023-02-03 17:06:26.110909987 +0000
+@@ -15,6 +15,11 @@
+ ARCH10="-gencode=arch=compute_75,code=sm_75"
+ ARCH11="-gencode=arch=compute_80,code=sm_80 \
+ -gencode=arch=compute_80,code=compute_80"
++ARCH111="-gencode=arch=compute_86,code=sm_86 \
++-gencode=arch=compute_86,code=compute_86"
++ARCH120="-gencode=arch=compute_90,code=sm_90 \
++-gencode=arch=compute_90,code=compute_90"
++
+
+ AC_DEFUN([CHECK_CUDA],[
+ AS_IF([test "x$cuda_checked" != "xyes"],
+@@ -104,8 +109,12 @@
+                [NVCC_CFLAGS="$NVCC_CFLAGS -O3 -g -DNDEBUG"])
+          AS_IF([test "x$cuda_happy" = "xyes"],
+                [AS_IF([test "x$with_nvcc_gencode" = "xdefault"],
+-                      [AS_IF([test $CUDA_MAJOR_VERSION -eq 11],
+-                             [NVCC_ARCH="${ARCH8} ${ARCH9} ${ARCH10} ${ARCH11}"])],
++                      [AS_IF([test $CUDA_MAJOR_VERSION -eq 12],
++                             [NVCC_ARCH="${ARCH7} ${ARCH8} ${ARCH9} ${ARCH10} ${ARCH110} ${ARCH111} ${ARCH120}"],
++                             [AS_IF([test $CUDA_MAJOR_VERSION -eq 11],
++                                   [AS_IF([test $CUDA_MINOR_VERSION -lt 1],
++                                           [NVCC_ARCH="${ARCH7} ${ARCH8} ${ARCH9} ${ARCH10} ${ARCH110}"],
++                                           [NVCC_ARCH="${ARCH7} ${ARCH8} ${ARCH9} ${ARCH10} ${ARCH110} ${ARCH111}"])])])],
+                       [NVCC_ARCH="$with_nvcc_gencode"])
+                 AC_SUBST([NVCC_ARCH], ["$NVCC_ARCH"])])
+          LDFLAGS="$save_LDFLAGS"

easybuild/easyconfigs/u/UCC/UCC-1.1.0-GCCcore-12.2.0.eb

@@ -14,7 +14,11 @@ toolchainopts = {'pic': True}
 
 source_urls = ['https://github.com/openucx/ucc/archive/refs/tags']
 sources = ['v%(version)s.tar.gz']
-checksums = ['74c8ba75037b5bd88cb703e8c8ae55639af3fecfd4428912a433c010c97b4df7']
+patches = ['UCC-%(version)s-multiple_component_paths.patch']
+checksums = [
+    {'v1.1.0.tar.gz': '74c8ba75037b5bd88cb703e8c8ae55639af3fecfd4428912a433c010c97b4df7'},
+    {'UCC-1.1.0-multiple_component_paths.patch': '3081d0f694331daa4a88a0fa3fb54b9a918015248ae5eb7b3157b924abd31bee'},
+]
 
 builddependencies = [
     ('binutils', '2.39'),

easybuild/easyconfigs/u/UCC/UCC-1.1.0-multiple_component_paths.patch

@@ -0,0 +1,46 @@
+Adds support for multiple extra components paths to UCC
+Was necessary to add, as they broke the logic we used for UCC_COMPONEN_PATH in 1.1.0,
+then removed it completely in 1.2.0.
+This patch also removes the need to copy over all the other plugins, thus supporting multiple external components.
+
+author: [email protected]
+
+diff -ru ucc-1.1.0.orig/src/utils/ucc_component.c ucc-1.1.0.fixed/src/utils/ucc_component.c
+--- ucc-1.1.0.orig/src/utils/ucc_component.c	2023-04-06 15:16:18.341729973 +0200
++++ ucc-1.1.0.fixed/src/utils/ucc_component.c	2023-04-08 13:42:28.448319486 +0200
+@@ -128,6 +128,27 @@
+         return UCC_ERR_INVALID_PARAM;
+     }
+
++    char *extra_component_path = getenv("EB_UCC_EXTRA_COMPONENT_PATH");
++    if (extra_component_path) {
++        // Add extra room for extra paths, braces and comma (+3):
++        pattern_size =
++            strlen(ucc_global_config.component_path) + strlen(framework_name) + 16 + strlen(extra_component_path) + 3;
++        full_pattern = (char *)ucc_malloc(pattern_size, "full_pattern");
++        if (!full_pattern) {
++            ucc_error("failed to allocate %zd bytes for full_pattern",
++                      pattern_size);
++            return UCC_ERR_NO_MEMORY;
++        }
++        ucc_snprintf_safe(full_pattern, pattern_size, "{%s,%s}/libucc_%s_*.so",
++                          extra_component_path, ucc_global_config.component_path, framework_name);
++        // Replace ; typically used in PATHs variables with , for glob
++        char *current_pos = strchr(full_pattern, ';');
++        while (current_pos) {
++            *current_pos = ',';
++            current_pos = strchr(current_pos, ';');
++        }
++        glob(full_pattern, GLOB_BRACE, NULL, &globbuf);
++    } else {
+     pattern_size =
+         strlen(ucc_global_config.component_path) + strlen(framework_name) + 16;
+     full_pattern = (char *)ucc_malloc(pattern_size, "full_pattern");
+@@ -139,6 +160,7 @@
+     ucc_snprintf_safe(full_pattern, pattern_size, "%s/libucc_%s_*.so",
+                       ucc_global_config.component_path, framework_name);
+     glob(full_pattern, 0, NULL, &globbuf);
++    }
+     ucc_free(full_pattern);
+     n_loaded          = 0;
+