[SYCL][CUDA] Add basic sub-group functionality

libclc/generic/include/spirv/spirv.h

@@ -37,13 +37,18 @@
 #include <macros.h>
 
 /* 6.11.1 Work-Item Functions */
-#include <spirv/workitem/get_global_size.h>
 #include <spirv/workitem/get_global_id.h>
-#include <spirv/workitem/get_local_size.h>
+#include <spirv/workitem/get_global_offset.h>
+#include <spirv/workitem/get_global_size.h>
+#include <spirv/workitem/get_group_id.h>
 #include <spirv/workitem/get_local_id.h>
+#include <spirv/workitem/get_local_size.h>
+#include <spirv/workitem/get_max_sub_group_size.h>
 #include <spirv/workitem/get_num_groups.h>
-#include <spirv/workitem/get_group_id.h>
-#include <spirv/workitem/get_global_offset.h>
+#include <spirv/workitem/get_num_sub_groups.h>
+#include <spirv/workitem/get_sub_group_id.h>
+#include <spirv/workitem/get_sub_group_local_id.h>
+#include <spirv/workitem/get_sub_group_size.h>
 #include <spirv/workitem/get_work_dim.h>
 
 /* 6.11.2.1 Floating-point macros */

libclc/generic/include/spirv/workitem/get_max_sub_group_size.h

@@ -0,0 +1,9 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_SubgroupMaxSize();

libclc/generic/include/spirv/workitem/get_num_sub_groups.h

@@ -0,0 +1,9 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_NumSubgroups();

libclc/generic/include/spirv/workitem/get_sub_group_id.h

@@ -0,0 +1,9 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_SubgroupId();

libclc/generic/include/spirv/workitem/get_sub_group_local_id.h

@@ -0,0 +1,9 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_SubgroupLocalInvocationId();

libclc/generic/include/spirv/workitem/get_sub_group_size.h

@@ -0,0 +1,9 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_SubgroupSize();

libclc/ptx-nvidiacl/libspirv/SOURCES

@@ -75,6 +75,11 @@ workitem/get_global_size.cl
 workitem/get_group_id.cl
 workitem/get_local_id.cl
 workitem/get_local_size.cl
+workitem/get_max_sub_group_size.cl
 workitem/get_num_groups.cl
+workitem/get_num_sub_groups.cl
+workitem/get_sub_group_id.cl
+workitem/get_sub_group_local_id.cl
+workitem/get_sub_group_size.cl
 images/image_helpers.ll
 images/image.cl

libclc/ptx-nvidiacl/libspirv/workitem/get_max_sub_group_size.cl

@@ -0,0 +1,15 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include <spirv/spirv.h>
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_SubgroupMaxSize() {
+  return 32;
+  // FIXME: warpsize is defined by NVVM IR but doesn't compile if used here
+  // return __nvvm_read_ptx_sreg_warpsize();
+}

libclc/ptx-nvidiacl/libspirv/workitem/get_num_sub_groups.cl

@@ -0,0 +1,20 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include <spirv/spirv.h>
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_NumSubgroups() {
+  // sreg.nwarpid returns number of warp identifiers, not number of warps
+  // see https://docs.nvidia.com/cuda/parallel-thread-execution/index.html
+  size_t size_x = __spirv_WorkgroupSize_x();
+  size_t size_y = __spirv_WorkgroupSize_y();
+  size_t size_z = __spirv_WorkgroupSize_z();
+  uint sg_size = __spirv_SubgroupMaxSize();
+  uint linear_size = size_z * size_y * size_x;
+  return (linear_size + sg_size - 1) / sg_size;
+}

libclc/ptx-nvidiacl/libspirv/workitem/get_sub_group_id.cl

@@ -0,0 +1,22 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include <spirv/spirv.h>
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_SubgroupId() {
+  // sreg.warpid is volatile and doesn't represent virtual warp index
+  // see https://docs.nvidia.com/cuda/parallel-thread-execution/index.html
+  size_t id_x = __spirv_LocalInvocationId_x();
+  size_t id_y = __spirv_LocalInvocationId_y();
+  size_t id_z = __spirv_LocalInvocationId_z();
+  size_t size_x = __spirv_WorkgroupSize_x();
+  size_t size_y = __spirv_WorkgroupSize_y();
+  size_t size_z = __spirv_WorkgroupSize_z();
+  uint sg_size = __spirv_SubgroupMaxSize();
+  return (id_z * size_y * size_x + id_y * size_x + id_x) / sg_size;
+}

libclc/ptx-nvidiacl/libspirv/workitem/get_sub_group_local_id.cl

@@ -0,0 +1,13 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include <spirv/spirv.h>
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_SubgroupLocalInvocationId() {
+  return __nvvm_read_ptx_sreg_laneid();
+}

libclc/ptx-nvidiacl/libspirv/workitem/get_sub_group_size.cl

@@ -0,0 +1,22 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include <spirv/spirv.h>
+
+_CLC_DEF _CLC_OVERLOAD uint __spirv_SubgroupSize() {
+  if (__spirv_SubgroupId() != __spirv_NumSubgroups() - 1) {
+    return __spirv_SubgroupMaxSize();
+  }
+  size_t size_x = __spirv_WorkgroupSize_x();
+  size_t size_y = __spirv_WorkgroupSize_y();
+  size_t size_z = __spirv_WorkgroupSize_z();
+  uint linear_size = size_z * size_y * size_x;
+  uint uniform_groups = __spirv_NumSubgroups() - 1;
+  uint uniform_size = __spirv_SubgroupMaxSize() * uniform_groups;
+  return linear_size - uniform_size;
+}

sycl/include/CL/__spirv/spirv_vars.hpp

@@ -45,6 +45,12 @@ SYCL_EXTERNAL size_t __spirv_LocalInvocationId_x();
 SYCL_EXTERNAL size_t __spirv_LocalInvocationId_y();
 SYCL_EXTERNAL size_t __spirv_LocalInvocationId_z();
 
+SYCL_EXTERNAL uint32_t __spirv_SubgroupSize();
+SYCL_EXTERNAL uint32_t __spirv_SubgroupMaxSize();
+SYCL_EXTERNAL uint32_t __spirv_NumSubgroups();
+SYCL_EXTERNAL uint32_t __spirv_SubgroupId();
+SYCL_EXTERNAL uint32_t __spirv_SubgroupLocalInvocationId();
+
 #else // __SYCL_NVPTX__
 
 typedef size_t size_t_vec __attribute__((ext_vector_type(3)));
@@ -56,6 +62,12 @@ __SPIRV_VAR_QUALIFIERS size_t_vec __spirv_BuiltInLocalInvocationId;
 __SPIRV_VAR_QUALIFIERS size_t_vec __spirv_BuiltInWorkgroupId;
 __SPIRV_VAR_QUALIFIERS size_t_vec __spirv_BuiltInGlobalOffset;
 
+__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInSubgroupSize;
+__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInSubgroupMaxSize;
+__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInNumSubgroups;
+__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInSubgroupId;
+__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInSubgroupLocalInvocationId;
+
 SYCL_EXTERNAL inline size_t __spirv_GlobalInvocationId_x() {
   return __spirv_BuiltInGlobalInvocationId.x;
 }
@@ -126,14 +138,23 @@ SYCL_EXTERNAL inline size_t __spirv_LocalInvocationId_z() {
   return __spirv_BuiltInLocalInvocationId.z;
 }
 
-#endif // __SYCL_NVPTX__
+SYCL_EXTERNAL inline uint32_t __spirv_SubgroupSize() {
+  return __spirv_BuiltInSubgroupSize;
+}
+SYCL_EXTERNAL inline uint32_t __spirv_SubgroupMaxSize() {
+  return __spirv_BuiltInSubgroupMaxSize;
+}
+SYCL_EXTERNAL inline uint32_t __spirv_NumSubgroups() {
+  return __spirv_BuiltInNumSubgroups;
+}
+SYCL_EXTERNAL inline uint32_t __spirv_SubgroupId() {
+  return __spirv_BuiltInSubgroupId;
+}
+SYCL_EXTERNAL inline uint32_t __spirv_SubgroupLocalInvocationId() {
+  return __spirv_BuiltInSubgroupLocalInvocationId;
+}
 
-__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInSubgroupSize;
-__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInSubgroupMaxSize;
-__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInNumSubgroups;
-__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInNumEnqueuedSubgroups;
-__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInSubgroupId;
-__SPIRV_VAR_QUALIFIERS uint32_t __spirv_BuiltInSubgroupLocalInvocationId;
+#endif // __SYCL_NVPTX__
 
 #undef __SPIRV_VAR_QUALIFIERS
 

sycl/include/CL/sycl/ONEAPI/sub_group.hpp

@@ -109,7 +109,7 @@ struct sub_group {
 
   id_type get_local_id() const {
 #ifdef __SYCL_DEVICE_ONLY__
-    return __spirv_BuiltInSubgroupLocalInvocationId;
+    return __spirv_SubgroupLocalInvocationId();
 #else
     throw runtime_error("Sub-groups are not supported on host device.",
                         PI_INVALID_DEVICE);
@@ -127,7 +127,7 @@ struct sub_group {
 
   range_type get_local_range() const {
 #ifdef __SYCL_DEVICE_ONLY__
-    return __spirv_BuiltInSubgroupSize;
+    return __spirv_SubgroupSize();
 #else
     throw runtime_error("Sub-groups are not supported on host device.",
                         PI_INVALID_DEVICE);
@@ -136,7 +136,7 @@ struct sub_group {
 
   range_type get_max_local_range() const {
 #ifdef __SYCL_DEVICE_ONLY__
-    return __spirv_BuiltInSubgroupMaxSize;
+    return __spirv_SubgroupMaxSize();
 #else
     throw runtime_error("Sub-groups are not supported on host device.",
                         PI_INVALID_DEVICE);
@@ -145,7 +145,7 @@ struct sub_group {
 
   id_type get_group_id() const {
 #ifdef __SYCL_DEVICE_ONLY__
-    return __spirv_BuiltInSubgroupId;
+    return __spirv_SubgroupId();
 #else
     throw runtime_error("Sub-groups are not supported on host device.",
                         PI_INVALID_DEVICE);
@@ -163,7 +163,7 @@ struct sub_group {
 
   range_type get_group_range() const {
 #ifdef __SYCL_DEVICE_ONLY__
-    return __spirv_BuiltInNumSubgroups;
+    return __spirv_NumSubgroups();
 #else
     throw runtime_error("Sub-groups are not supported on host device.",
                         PI_INVALID_DEVICE);

sycl/test/sub_group/common.cpp

@@ -1,6 +1,3 @@
-// UNSUPPORTED: cuda
-// CUDA compilation and runtime do not yet support sub-groups.
-//
 // RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: env SYCL_DEVICE_TYPE=HOST %t.out
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
@@ -70,7 +67,7 @@ void check(queue &Queue, unsigned int G, unsigned int L) {
 }
 int main() {
   queue Queue;
-  if (!core_sg_supported(Queue.get_device())) {
+  if (Queue.get_device().is_host()) {
     std::cout << "Skipping test\n";
     return 0;
   }