[clang] [Headers][NFC] Steps to allow sharing code between gpu intrin.h headers (PR #131134)
https://github.com/JonChesterfield updated
https://github.com/llvm/llvm-project/pull/131134
>From 7347ebc6a0aadd1b9676e329bdf7705dbfae7875 Mon Sep 17 00:00:00 2001
From: Jon Chesterfield
Date: Thu, 13 Mar 2025 12:49:42 +
Subject: [PATCH] [libc][nfc] Steps to allow sharing code between gpu intrin.h
headers
---
clang/lib/Headers/amdgpuintrin.h | 22 ++
clang/lib/Headers/gpuintrin.h| 26 +++---
clang/lib/Headers/nvptxintrin.h | 27 ---
3 files changed, 25 insertions(+), 50 deletions(-)
diff --git a/clang/lib/Headers/amdgpuintrin.h b/clang/lib/Headers/amdgpuintrin.h
index 839a05175cf3e..56748f6c3e818 100644
--- a/clang/lib/Headers/amdgpuintrin.h
+++ b/clang/lib/Headers/amdgpuintrin.h
@@ -13,11 +13,8 @@
#error "This file is intended for AMDGPU targets or offloading to AMDGPU"
#endif
-#include
-
-#if !defined(__cplusplus)
-_Pragma("push_macro(\"bool\")");
-#define bool _Bool
+#ifndef __GPUINTRIN_H
+#error "Never use directly; include instead"
#endif
_Pragma("omp begin declare target device_type(nohost)");
@@ -146,17 +143,6 @@ __gpu_shuffle_idx_u32(uint64_t __lane_mask, uint32_t
__idx, uint32_t __x,
return __builtin_amdgcn_ds_bpermute(__lane << 2, __x);
}
-// Shuffles the the lanes inside the wavefront according to the given index.
-_DEFAULT_FN_ATTRS static __inline__ uint64_t
-__gpu_shuffle_idx_u64(uint64_t __lane_mask, uint32_t __idx, uint64_t __x,
- uint32_t __width) {
- uint32_t __hi = (uint32_t)(__x >> 32ull);
- uint32_t __lo = (uint32_t)(__x & 0x);
- return ((uint64_t)__gpu_shuffle_idx_u32(__lane_mask, __idx, __hi, __width)
- << 32ull) |
- ((uint64_t)__gpu_shuffle_idx_u32(__lane_mask, __idx, __lo, __width));
-}
-
// Returns a bitmask marking all lanes that have the same value of __x.
_DEFAULT_FN_ATTRS static __inline__ uint64_t
__gpu_match_any_u32(uint64_t __lane_mask, uint32_t __x) {
@@ -238,8 +224,4 @@ _DEFAULT_FN_ATTRS static __inline__ void
__gpu_thread_suspend(void) {
_Pragma("omp end declare variant");
_Pragma("omp end declare target");
-#if !defined(__cplusplus)
-_Pragma("pop_macro(\"bool\")");
-#endif
-
#endif // __AMDGPUINTRIN_H
diff --git a/clang/lib/Headers/gpuintrin.h b/clang/lib/Headers/gpuintrin.h
index 4181628d18048..ac79d685337c5 100644
--- a/clang/lib/Headers/gpuintrin.h
+++ b/clang/lib/Headers/gpuintrin.h
@@ -25,6 +25,13 @@
#endif
#endif
+#include
+
+#if !defined(__cplusplus)
+_Pragma("push_macro(\"bool\")");
+#define bool _Bool
+#endif
+
#if defined(__NVPTX__)
#include
#elif defined(__AMDGPU__)
@@ -33,13 +40,6 @@
#error "This header is only meant to be used on GPU architectures."
#endif
-#include
-
-#if !defined(__cplusplus)
-_Pragma("push_macro(\"bool\")");
-#define bool _Bool
-#endif
-
_Pragma("omp begin declare target device_type(nohost)");
_Pragma("omp begin declare variant match(device = {kind(gpu)})");
@@ -141,6 +141,18 @@ __gpu_read_first_lane_f64(uint64_t __lane_mask, double
__x) {
__builtin_bit_cast(uint64_t, __x)));
}
+// Shuffles the the lanes according to the given index.
+_DEFAULT_FN_ATTRS static __inline__ uint64_t
+__gpu_shuffle_idx_u64(uint64_t __lane_mask, uint32_t __idx, uint64_t __x,
+ uint32_t __width) {
+ uint32_t __hi = (uint32_t)(__x >> 32ull);
+ uint32_t __lo = (uint32_t)(__x & 0x);
+ uint32_t __mask = (uint32_t)__lane_mask;
+ return ((uint64_t)__gpu_shuffle_idx_u32(__mask, __idx, __hi, __width)
+ << 32ull) |
+ ((uint64_t)__gpu_shuffle_idx_u32(__mask, __idx, __lo, __width));
+}
+
// Shuffles the the lanes according to the given index.
_DEFAULT_FN_ATTRS static __inline__ float
__gpu_shuffle_idx_f32(uint64_t __lane_mask, uint32_t __idx, float __x,
diff --git a/clang/lib/Headers/nvptxintrin.h b/clang/lib/Headers/nvptxintrin.h
index d00a5f6de3950..10ad7a682d4cd 100644
--- a/clang/lib/Headers/nvptxintrin.h
+++ b/clang/lib/Headers/nvptxintrin.h
@@ -13,15 +13,12 @@
#error "This file is intended for NVPTX targets or offloading to NVPTX"
#endif
-#ifndef __CUDA_ARCH__
-#define __CUDA_ARCH__ 0
+#ifndef __GPUINTRIN_H
+#error "Never use directly; include instead"
#endif
-#include
-
-#if !defined(__cplusplus)
-_Pragma("push_macro(\"bool\")");
-#define bool _Bool
+#ifndef __CUDA_ARCH__
+#define __CUDA_ARCH__ 0
#endif
_Pragma("omp begin declare target device_type(nohost)");
@@ -153,18 +150,6 @@ __gpu_shuffle_idx_u32(uint64_t __lane_mask, uint32_t
__idx, uint32_t __x,
((__gpu_num_lanes() - __width) << 8u) |
0x1f);
}
-// Shuffles the the lanes inside the warp according to the given index.
-_DEFAULT_FN_ATTRS static __inline__ uint64_t
-__gpu_shuffle_idx_u64(uint64_t __lane_mask, uint32_t __idx, uint64_t __x,
- uint32_t __width) {
- uint32_t __hi = (uint32_t)(__x >> 32ull);
- uint32_t __lo = (uint32_t)(__x & 0x);
- uint32_t __mask =
[clang] [Headers][NFC] Steps to allow sharing code between gpu intrin.h headers (PR #131134)
https://github.com/JonChesterfield closed https://github.com/llvm/llvm-project/pull/131134 ___ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Headers][NFC] Steps to allow sharing code between gpu intrin.h headers (PR #131134)
https://github.com/jhuber6 edited https://github.com/llvm/llvm-project/pull/131134 ___ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Headers][NFC] Steps to allow sharing code between gpu intrin.h headers (PR #131134)
https://github.com/jhuber6 approved this pull request. LG https://github.com/llvm/llvm-project/pull/131134 ___ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Headers][NFC] Steps to allow sharing code between gpu intrin.h headers (PR #131134)
https://github.com/JonChesterfield edited https://github.com/llvm/llvm-project/pull/131134 ___ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Headers][NFC] Steps to allow sharing code between gpu intrin.h headers (PR #131134)
https://github.com/JonChesterfield edited https://github.com/llvm/llvm-project/pull/131134 ___ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Headers][NFC] Steps to allow sharing code between gpu intrin.h headers (PR #131134)
https://github.com/JonChesterfield updated
https://github.com/llvm/llvm-project/pull/131134
>From 4c04f6979409642eb6bc9dc3c48b5e3636210ef0 Mon Sep 17 00:00:00 2001
From: Jon Chesterfield
Date: Thu, 13 Mar 2025 12:49:42 +
Subject: [PATCH] [libc][nfc] Steps to allow sharing code between gpu intrin.h
headers
---
clang/lib/Headers/amdgpuintrin.h | 24 ++--
clang/lib/Headers/gpuintrin.h| 32 +++-
clang/lib/Headers/nvptxintrin.h | 29 -
3 files changed, 29 insertions(+), 56 deletions(-)
diff --git a/clang/lib/Headers/amdgpuintrin.h b/clang/lib/Headers/amdgpuintrin.h
index 839a05175cf3e..6a2b46b0c511e 100644
--- a/clang/lib/Headers/amdgpuintrin.h
+++ b/clang/lib/Headers/amdgpuintrin.h
@@ -13,14 +13,10 @@
#error "This file is intended for AMDGPU targets or offloading to AMDGPU"
#endif
-#include
-
-#if !defined(__cplusplus)
-_Pragma("push_macro(\"bool\")");
-#define bool _Bool
+#ifndef __GPUINTRIN_H
+#error "Never use directly; include instead"
#endif
-_Pragma("omp begin declare target device_type(nohost)");
_Pragma("omp begin declare variant match(device = {arch(amdgcn)})");
// Type aliases to the address spaces used by the AMDGPU backend.
@@ -146,17 +142,6 @@ __gpu_shuffle_idx_u32(uint64_t __lane_mask, uint32_t
__idx, uint32_t __x,
return __builtin_amdgcn_ds_bpermute(__lane << 2, __x);
}
-// Shuffles the the lanes inside the wavefront according to the given index.
-_DEFAULT_FN_ATTRS static __inline__ uint64_t
-__gpu_shuffle_idx_u64(uint64_t __lane_mask, uint32_t __idx, uint64_t __x,
- uint32_t __width) {
- uint32_t __hi = (uint32_t)(__x >> 32ull);
- uint32_t __lo = (uint32_t)(__x & 0x);
- return ((uint64_t)__gpu_shuffle_idx_u32(__lane_mask, __idx, __hi, __width)
- << 32ull) |
- ((uint64_t)__gpu_shuffle_idx_u32(__lane_mask, __idx, __lo, __width));
-}
-
// Returns a bitmask marking all lanes that have the same value of __x.
_DEFAULT_FN_ATTRS static __inline__ uint64_t
__gpu_match_any_u32(uint64_t __lane_mask, uint32_t __x) {
@@ -236,10 +221,5 @@ _DEFAULT_FN_ATTRS static __inline__ void
__gpu_thread_suspend(void) {
}
_Pragma("omp end declare variant");
-_Pragma("omp end declare target");
-
-#if !defined(__cplusplus)
-_Pragma("pop_macro(\"bool\")");
-#endif
#endif // __AMDGPUINTRIN_H
diff --git a/clang/lib/Headers/gpuintrin.h b/clang/lib/Headers/gpuintrin.h
index 4181628d18048..d369e60b2fcac 100644
--- a/clang/lib/Headers/gpuintrin.h
+++ b/clang/lib/Headers/gpuintrin.h
@@ -25,14 +25,6 @@
#endif
#endif
-#if defined(__NVPTX__)
-#include
-#elif defined(__AMDGPU__)
-#include
-#elif !defined(_OPENMP)
-#error "This header is only meant to be used on GPU architectures."
-#endif
-
#include
#if !defined(__cplusplus)
@@ -41,6 +33,15 @@ _Pragma("push_macro(\"bool\")");
#endif
_Pragma("omp begin declare target device_type(nohost)");
+
+#if defined(__NVPTX__)
+#include
+#elif defined(__AMDGPU__)
+#include
+#elif !defined(_OPENMP)
+#error "This header is only meant to be used on GPU architectures."
+#endif
+
_Pragma("omp begin declare variant match(device = {kind(gpu)})");
#define __GPU_X_DIM 0
@@ -141,6 +142,18 @@ __gpu_read_first_lane_f64(uint64_t __lane_mask, double
__x) {
__builtin_bit_cast(uint64_t, __x)));
}
+// Shuffles the the lanes according to the given index.
+_DEFAULT_FN_ATTRS static __inline__ uint64_t
+__gpu_shuffle_idx_u64(uint64_t __lane_mask, uint32_t __idx, uint64_t __x,
+ uint32_t __width) {
+ uint32_t __hi = (uint32_t)(__x >> 32ull);
+ uint32_t __lo = (uint32_t)(__x & 0x);
+ uint32_t __mask = (uint32_t)__lane_mask;
+ return ((uint64_t)__gpu_shuffle_idx_u32(__mask, __idx, __hi, __width)
+ << 32ull) |
+ ((uint64_t)__gpu_shuffle_idx_u32(__mask, __idx, __lo, __width));
+}
+
// Shuffles the the lanes according to the given index.
_DEFAULT_FN_ATTRS static __inline__ float
__gpu_shuffle_idx_f32(uint64_t __lane_mask, uint32_t __idx, float __x,
@@ -223,7 +236,8 @@ __DO_LANE_SUM(double, f64); // double
__gpu_lane_sum_f64(m, x)
#undef __DO_LANE_SUM
_Pragma("omp end declare variant");
-_Pragma("omp end declare target");
+
+_Pragma("omp end declare target"); // nohost
#if !defined(__cplusplus)
_Pragma("pop_macro(\"bool\")");
diff --git a/clang/lib/Headers/nvptxintrin.h b/clang/lib/Headers/nvptxintrin.h
index d00a5f6de3950..8e6b0de72aae1 100644
--- a/clang/lib/Headers/nvptxintrin.h
+++ b/clang/lib/Headers/nvptxintrin.h
@@ -13,18 +13,14 @@
#error "This file is intended for NVPTX targets or offloading to NVPTX"
#endif
-#ifndef __CUDA_ARCH__
-#define __CUDA_ARCH__ 0
+#ifndef __GPUINTRIN_H
+#error "Never use directly; include instead"
#endif
-#include
-
-#if !defined(__cplusplus)
-_Pragma("push_macro(\"bool\")");
-#define bool _Bool
+#ifndef __CUDA_ARCH__
+#define __CUDA_ARCH__ 0
#endif
-_Pragma("omp begin declare tar
