PTX: move implementations of work-item and synchronisation functions

to lib, and add header files in generic.  Incorporates a patch by
Tom Stellard!

llvm-svn: 161313

libclc/generic/include/clc/synchronization/barrier.h

@@ -0,0 +1 @@
+_CLC_DECL void barrier(cl_mem_fence_flags flags);

libclc/generic/include/clc/workitem/get_global_id.h

@@ -0,0 +1 @@
+_CLC_DECL size_t get_global_id(uint dim);

libclc/generic/include/clc/workitem/get_global_size.h

@@ -0,0 +1 @@
+_CLC_DECL size_t get_global_size(uint dim);

libclc/generic/include/clc/workitem/get_group_id.h

@@ -0,0 +1 @@
+_CLC_DECL size_t get_group_id(uint dim);

libclc/generic/include/clc/workitem/get_local_id.h

@@ -0,0 +1 @@
+_CLC_DECL size_t get_local_id(uint dim);

libclc/generic/include/clc/workitem/get_local_size.h

@@ -0,0 +1 @@
+_CLC_DECL size_t get_local_size(uint dim);

libclc/generic/include/clc/workitem/get_num_groups.h

@@ -0,0 +1 @@
+_CLC_DECL size_t get_num_groups(uint dim);

libclc/generic/lib/SOURCES

@@ -12,3 +12,5 @@ integer/sub_sat.ll
 integer/sub_sat_impl.ll
 math/hypot.cl
 math/mad.cl
+workitem/get_global_id.cl
+workitem/get_global_size.cl

libclc/generic/lib/workitem/get_global_id.cl

@@ -0,0 +1,5 @@
+#include <clc/clc.h>
+
+_CLC_DEF size_t get_global_id(uint dim) {
+  return get_group_id(dim)*get_local_size(dim) + get_local_id(dim);
+}

libclc/generic/lib/workitem/get_global_size.cl

@@ -0,0 +1,5 @@
+#include <clc/clc.h>
+
+_CLC_DEF size_t get_global_size(uint dim) {
+  return get_num_groups(dim)*get_local_size(dim);
+}

libclc/ptx-nvidiacl/include/clc/workitem/get_global_id.h

@@ -1,8 +0,0 @@
-_CLC_INLINE size_t get_global_id(uint dim) {
-  switch (dim) {
-  case 0:  return __builtin_ptx_read_ctaid_x()*__builtin_ptx_read_ntid_x()+__builtin_ptx_read_tid_x();
-  case 1:  return __builtin_ptx_read_ctaid_y()*__builtin_ptx_read_ntid_y()+__builtin_ptx_read_tid_y();
-  case 2:  return __builtin_ptx_read_ctaid_z()*__builtin_ptx_read_ntid_z()+__builtin_ptx_read_tid_z();
-  default: return 0;
-  }
-}

libclc/ptx-nvidiacl/include/clc/workitem/get_global_size.h

@@ -1,8 +0,0 @@
-_CLC_INLINE size_t get_global_size(uint dim) {
-  switch (dim) {
-  case 0:  return __builtin_ptx_read_nctaid_x()*__builtin_ptx_read_ntid_x();
-  case 1:  return __builtin_ptx_read_nctaid_y()*__builtin_ptx_read_ntid_y();
-  case 2:  return __builtin_ptx_read_nctaid_z()*__builtin_ptx_read_ntid_z();
-  default: return 0;
-  }
-}

libclc/ptx-nvidiacl/lib/SOURCES

@@ -0,0 +1,4 @@
+workitem/get_group_id.cl
+workitem/get_local_id.cl
+workitem/get_local_size.cl
+workitem/get_num_groups.cl

libclc/ptx-nvidiacl/lib/synchronization/barrier.cl

@@ -1,4 +1,6 @@
-_CLC_INLINE void barrier(cl_mem_fence_flags flags) {
+#include <clc/clc.h>
+
+_CLC_DEF void barrier(cl_mem_fence_flags flags) {
   if (flags & CLK_LOCAL_MEM_FENCE) {
     __builtin_ptx_bar_sync(0);
   }

libclc/ptx-nvidiacl/lib/workitem/get_group_id.cl

@@ -1,4 +1,6 @@
-_CLC_INLINE size_t get_group_id(uint dim) {
+#include <clc/clc.h>
+
+_CLC_DEF size_t get_group_id(uint dim) {
   switch (dim) {
   case 0:  return __builtin_ptx_read_ctaid_x();
   case 1:  return __builtin_ptx_read_ctaid_y();

libclc/ptx-nvidiacl/lib/workitem/get_local_id.cl

@@ -1,4 +1,6 @@
-_CLC_INLINE size_t get_local_id(uint dim) {
+#include <clc/clc.h>
+
+_CLC_DEF size_t get_local_id(uint dim) {
   switch (dim) {
   case 0:  return __builtin_ptx_read_tid_x();
   case 1:  return __builtin_ptx_read_tid_y();

libclc/ptx-nvidiacl/lib/workitem/get_local_size.cl

@@ -1,4 +1,6 @@
-_CLC_INLINE size_t get_local_size(uint dim) {
+#include <clc/clc.h>
+
+_CLC_DEF size_t get_local_size(uint dim) {
   switch (dim) {
   case 0:  return __builtin_ptx_read_ntid_x();
   case 1:  return __builtin_ptx_read_ntid_y();

libclc/ptx-nvidiacl/lib/workitem/get_num_groups.cl

@@ -1,4 +1,6 @@
-_CLC_INLINE size_t get_num_groups(uint dim) {
+#include <clc/clc.h>
+
+_CLC_DEF size_t get_num_groups(uint dim) {
   switch (dim) {
   case 0:  return __builtin_ptx_read_nctaid_x();
   case 1:  return __builtin_ptx_read_nctaid_y();