[Clang] Fix cross-lane scan when given divergent lanes #127703
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
jhuber6
requested review from
arsenm,
Artem-B,
jdoerfert,
JonChesterfield,
jplehr,
shiltian and
yxsamliu
llvmbot
added
clang
|
@llvm/pr-subscribers-backend-x86 @llvm/pr-subscribers-clang Author: Joseph Huber (jhuber6) ChangesSummary: Needs to be backported for correct behavior and because it fixes a Full diff: https://github.com/llvm/llvm-project/pull/127703.diff 4 Files Affected:
diff --git a/clang/lib/Headers/gpuintrin.h b/clang/lib/Headers/gpuintrin.h
index 11c87e85cd497..75271edbc84a2 100644
--- a/clang/lib/Headers/gpuintrin.h
+++ b/clang/lib/Headers/gpuintrin.h
@@ -154,7 +154,7 @@ __gpu_shuffle_idx_f64(uint64_t __lane_mask, uint32_t __idx, double __x,
#define __DO_LANE_SUM(__type, __suffix) \
_DEFAULT_FN_ATTRS static __inline__ __type __gpu_lane_sum_##__suffix( \
uint64_t __lane_mask, __type __x) { \
- for (uint32_t __step = __gpu_num_lanes() / 2; __step > 0; __step /= 2) { \
+ for (uint32_t __step = 1; __step < __gpu_num_lanes(); __step *= 2) { \
uint32_t __index = __step + __gpu_lane_id(); \
__x += __gpu_shuffle_idx_##__suffix(__lane_mask, __index, __x, \
__gpu_num_lanes()); \
@@ -171,14 +171,29 @@ __DO_LANE_SUM(double, f64); // double __gpu_lane_sum_f64(m, x)
#define __DO_LANE_SCAN(__type, __bitmask_type, __suffix) \
_DEFAULT_FN_ATTRS static __inline__ uint32_t __gpu_lane_scan_##__suffix( \
uint64_t __lane_mask, uint32_t __x) { \
- for (uint32_t __step = 1; __step < __gpu_num_lanes(); __step *= 2) { \
- uint32_t __index = __gpu_lane_id() - __step; \
- __bitmask_type bitmask = __gpu_lane_id() >= __step; \
- __x += __builtin_bit_cast( \
- __type, -bitmask & __builtin_bit_cast(__bitmask_type, \
- __gpu_shuffle_idx_##__suffix( \
- __lane_mask, __index, __x, \
- __gpu_num_lanes()))); \
+ uint64_t __first = __lane_mask >> __builtin_ctzll(__lane_mask); \
+ bool __divergent = __gpu_read_first_lane_##__suffix( \
+ __lane_mask, __first & (__first + 1)); \
+ if (__divergent) { \
+ __type __accum = 0; \
+ for (uint64_t __mask = __lane_mask; __mask; __mask &= __mask - 1) { \
+ __type __index = __builtin_ctzll(__mask); \
+ __type __tmp = __gpu_shuffle_idx_##__suffix(__lane_mask, __index, __x, \
+ __gpu_num_lanes()); \
+ __x = __gpu_lane_id() == __index ? __accum + __tmp : __x; \
+ __accum += __tmp; \
+ } \
+ } else { \
+ for (uint32_t __step = 1; __step < __gpu_num_lanes(); __step *= 2) { \
+ uint32_t __index = __gpu_lane_id() - __step; \
+ __bitmask_type bitmask = __gpu_lane_id() >= __step; \
+ __x += __builtin_bit_cast( \
+ __type, \
+ -bitmask & __builtin_bit_cast(__bitmask_type, \
+ __gpu_shuffle_idx_##__suffix( \
+ __lane_mask, __index, __x, \
+ __gpu_num_lanes()))); \
+ } \
} \
return __x; \
}
diff --git a/clang/lib/Headers/nvptxintrin.h b/clang/lib/Headers/nvptxintrin.h
index f857a87b5f4c7..29d0adcabc82f 100644
--- a/clang/lib/Headers/nvptxintrin.h
+++ b/clang/lib/Headers/nvptxintrin.h
@@ -155,8 +155,11 @@ _DEFAULT_FN_ATTRS static __inline__ void __gpu_sync_lane(uint64_t __lane_mask) {
_DEFAULT_FN_ATTRS static __inline__ uint32_t
__gpu_shuffle_idx_u32(uint64_t __lane_mask, uint32_t __idx, uint32_t __x,
uint32_t __width) {
+ // Mask out inactive lanes to match AMDGPU behavior.
uint32_t __mask = (uint32_t)__lane_mask;
- return __nvvm_shfl_sync_idx_i32(__mask, __x, __idx,
+ bool __bitmask = (1ull << __idx) & __lane_mask;
+ return -__bitmask &
+ __nvvm_shfl_sync_idx_i32(__mask, __x, __idx,
((__gpu_num_lanes() - __width) << 8u) | 0x1f);
}
diff --git a/libc/test/integration/src/__support/GPU/CMakeLists.txt b/libc/test/integration/src/__support/GPU/CMakeLists.txt
index e066830f6cc0d..a750f2302b511 100644
--- a/libc/test/integration/src/__support/GPU/CMakeLists.txt
+++ b/libc/test/integration/src/__support/GPU/CMakeLists.txt
@@ -6,6 +6,8 @@ add_integration_test(
SUITE libc-support-gpu-tests
SRCS
scan_reduce.cpp
+ DEPENDS
+ libc.src.stdio.printf
LOADER_ARGS
--threads 64
)
diff --git a/libc/test/integration/src/__support/GPU/scan_reduce.cpp b/libc/test/integration/src/__support/GPU/scan_reduce.cpp
index bc621c3300cbe..7739137f5b200 100644
--- a/libc/test/integration/src/__support/GPU/scan_reduce.cpp
+++ b/libc/test/integration/src/__support/GPU/scan_reduce.cpp
@@ -8,6 +8,7 @@
#include "src/__support/CPP/bit.h"
#include "src/__support/GPU/utils.h"
+#include "src/stdio/printf.h"
#include "test/IntegrationTest/test.h"
using namespace LIBC_NAMESPACE;
@@ -53,10 +54,59 @@ static void test_scan() {
EXPECT_EQ(z, gpu::get_lane_id() % 2 ? gpu::get_lane_id() / 2 + 1 : 0);
}
+static uint32_t random(uint64_t *rand_next) {
+ uint64_t x = *rand_next;
+ x ^= x >> 12;
+ x ^= x << 25;
+ x ^= x >> 27;
+ *rand_next = x;
+ return static_cast<uint32_t>((x * 0x2545F4914F6CDD1Dul) >> 32);
+}
+
+// Scan operations can break down under thread divergence, make sure that the
+// function works under some random divergence. We do this by trivially
+// implementing a scan with shared scratch memory and then comparing the
+// results.
+static void test_scan_divergent() {
+ static uint32_t input[64] = {0};
+ static uint32_t result[64] = {0};
+ uint64_t state = gpu::processor_clock() * __gpu_lane_id();
+
+ for (int i = 0; i < 64; ++i) {
+ uint64_t lanemask = gpu::get_lane_mask();
+ if (random(&state) & (1ull << gpu::get_lane_id())) {
+ uint64_t divergent = gpu::get_lane_mask();
+ uint32_t value = random(&state) % 256;
+ input[gpu::get_lane_id()] = value;
+
+ if (gpu::is_first_lane(divergent)) {
+ uint32_t accumulator = 0;
+ for (uint32_t lane = 0; lane < gpu::get_lane_size(); ++lane) {
+ uint32_t tmp = input[lane];
+ result[lane] = tmp + accumulator;
+ accumulator += tmp;
+ }
+ }
+ gpu::sync_lane(divergent);
+
+ uint32_t scan = gpu::scan(divergent, value);
+ EXPECT_EQ(scan, result[gpu::get_lane_id()]);
+ }
+ if (gpu::is_first_lane(lanemask))
+ __builtin_memset(input, 0, sizeof(input));
+ gpu::sync_lane(lanemask);
+ }
+}
+
TEST_MAIN(int argc, char **argv, char **envp) {
+ if (gpu::get_thread_id() >= gpu::get_lane_size())
+ return 0;
+
test_reduce();
test_scan();
+ test_scan_divergent();
+
return 0;
}
|
|
@llvm/pr-subscribers-libc Author: Joseph Huber (jhuber6) ChangesSummary: Needs to be backported for correct behavior and because it fixes a Full diff: https://github.com/llvm/llvm-project/pull/127703.diff 4 Files Affected:
diff --git a/clang/lib/Headers/gpuintrin.h b/clang/lib/Headers/gpuintrin.h
index 11c87e85cd497..75271edbc84a2 100644
--- a/clang/lib/Headers/gpuintrin.h
+++ b/clang/lib/Headers/gpuintrin.h
@@ -154,7 +154,7 @@ __gpu_shuffle_idx_f64(uint64_t __lane_mask, uint32_t __idx, double __x,
#define __DO_LANE_SUM(__type, __suffix) \
_DEFAULT_FN_ATTRS static __inline__ __type __gpu_lane_sum_##__suffix( \
uint64_t __lane_mask, __type __x) { \
- for (uint32_t __step = __gpu_num_lanes() / 2; __step > 0; __step /= 2) { \
+ for (uint32_t __step = 1; __step < __gpu_num_lanes(); __step *= 2) { \
uint32_t __index = __step + __gpu_lane_id(); \
__x += __gpu_shuffle_idx_##__suffix(__lane_mask, __index, __x, \
__gpu_num_lanes()); \
@@ -171,14 +171,29 @@ __DO_LANE_SUM(double, f64); // double __gpu_lane_sum_f64(m, x)
#define __DO_LANE_SCAN(__type, __bitmask_type, __suffix) \
_DEFAULT_FN_ATTRS static __inline__ uint32_t __gpu_lane_scan_##__suffix( \
uint64_t __lane_mask, uint32_t __x) { \
- for (uint32_t __step = 1; __step < __gpu_num_lanes(); __step *= 2) { \
- uint32_t __index = __gpu_lane_id() - __step; \
- __bitmask_type bitmask = __gpu_lane_id() >= __step; \
- __x += __builtin_bit_cast( \
- __type, -bitmask & __builtin_bit_cast(__bitmask_type, \
- __gpu_shuffle_idx_##__suffix( \
- __lane_mask, __index, __x, \
- __gpu_num_lanes()))); \
+ uint64_t __first = __lane_mask >> __builtin_ctzll(__lane_mask); \
+ bool __divergent = __gpu_read_first_lane_##__suffix( \
+ __lane_mask, __first & (__first + 1)); \
+ if (__divergent) { \
+ __type __accum = 0; \
+ for (uint64_t __mask = __lane_mask; __mask; __mask &= __mask - 1) { \
+ __type __index = __builtin_ctzll(__mask); \
+ __type __tmp = __gpu_shuffle_idx_##__suffix(__lane_mask, __index, __x, \
+ __gpu_num_lanes()); \
+ __x = __gpu_lane_id() == __index ? __accum + __tmp : __x; \
+ __accum += __tmp; \
+ } \
+ } else { \
+ for (uint32_t __step = 1; __step < __gpu_num_lanes(); __step *= 2) { \
+ uint32_t __index = __gpu_lane_id() - __step; \
+ __bitmask_type bitmask = __gpu_lane_id() >= __step; \
+ __x += __builtin_bit_cast( \
+ __type, \
+ -bitmask & __builtin_bit_cast(__bitmask_type, \
+ __gpu_shuffle_idx_##__suffix( \
+ __lane_mask, __index, __x, \
+ __gpu_num_lanes()))); \
+ } \
} \
return __x; \
}
diff --git a/clang/lib/Headers/nvptxintrin.h b/clang/lib/Headers/nvptxintrin.h
index f857a87b5f4c7..29d0adcabc82f 100644
--- a/clang/lib/Headers/nvptxintrin.h
+++ b/clang/lib/Headers/nvptxintrin.h
@@ -155,8 +155,11 @@ _DEFAULT_FN_ATTRS static __inline__ void __gpu_sync_lane(uint64_t __lane_mask) {
_DEFAULT_FN_ATTRS static __inline__ uint32_t
__gpu_shuffle_idx_u32(uint64_t __lane_mask, uint32_t __idx, uint32_t __x,
uint32_t __width) {
+ // Mask out inactive lanes to match AMDGPU behavior.
uint32_t __mask = (uint32_t)__lane_mask;
- return __nvvm_shfl_sync_idx_i32(__mask, __x, __idx,
+ bool __bitmask = (1ull << __idx) & __lane_mask;
+ return -__bitmask &
+ __nvvm_shfl_sync_idx_i32(__mask, __x, __idx,
((__gpu_num_lanes() - __width) << 8u) | 0x1f);
}
diff --git a/libc/test/integration/src/__support/GPU/CMakeLists.txt b/libc/test/integration/src/__support/GPU/CMakeLists.txt
index e066830f6cc0d..a750f2302b511 100644
--- a/libc/test/integration/src/__support/GPU/CMakeLists.txt
+++ b/libc/test/integration/src/__support/GPU/CMakeLists.txt
@@ -6,6 +6,8 @@ add_integration_test(
SUITE libc-support-gpu-tests
SRCS
scan_reduce.cpp
+ DEPENDS
+ libc.src.stdio.printf
LOADER_ARGS
--threads 64
)
diff --git a/libc/test/integration/src/__support/GPU/scan_reduce.cpp b/libc/test/integration/src/__support/GPU/scan_reduce.cpp
index bc621c3300cbe..7739137f5b200 100644
--- a/libc/test/integration/src/__support/GPU/scan_reduce.cpp
+++ b/libc/test/integration/src/__support/GPU/scan_reduce.cpp
@@ -8,6 +8,7 @@
#include "src/__support/CPP/bit.h"
#include "src/__support/GPU/utils.h"
+#include "src/stdio/printf.h"
#include "test/IntegrationTest/test.h"
using namespace LIBC_NAMESPACE;
@@ -53,10 +54,59 @@ static void test_scan() {
EXPECT_EQ(z, gpu::get_lane_id() % 2 ? gpu::get_lane_id() / 2 + 1 : 0);
}
+static uint32_t random(uint64_t *rand_next) {
+ uint64_t x = *rand_next;
+ x ^= x >> 12;
+ x ^= x << 25;
+ x ^= x >> 27;
+ *rand_next = x;
+ return static_cast<uint32_t>((x * 0x2545F4914F6CDD1Dul) >> 32);
+}
+
+// Scan operations can break down under thread divergence, make sure that the
+// function works under some random divergence. We do this by trivially
+// implementing a scan with shared scratch memory and then comparing the
+// results.
+static void test_scan_divergent() {
+ static uint32_t input[64] = {0};
+ static uint32_t result[64] = {0};
+ uint64_t state = gpu::processor_clock() * __gpu_lane_id();
+
+ for (int i = 0; i < 64; ++i) {
+ uint64_t lanemask = gpu::get_lane_mask();
+ if (random(&state) & (1ull << gpu::get_lane_id())) {
+ uint64_t divergent = gpu::get_lane_mask();
+ uint32_t value = random(&state) % 256;
+ input[gpu::get_lane_id()] = value;
+
+ if (gpu::is_first_lane(divergent)) {
+ uint32_t accumulator = 0;
+ for (uint32_t lane = 0; lane < gpu::get_lane_size(); ++lane) {
+ uint32_t tmp = input[lane];
+ result[lane] = tmp + accumulator;
+ accumulator += tmp;
+ }
+ }
+ gpu::sync_lane(divergent);
+
+ uint32_t scan = gpu::scan(divergent, value);
+ EXPECT_EQ(scan, result[gpu::get_lane_id()]);
+ }
+ if (gpu::is_first_lane(lanemask))
+ __builtin_memset(input, 0, sizeof(input));
+ gpu::sync_lane(lanemask);
+ }
+}
+
TEST_MAIN(int argc, char **argv, char **envp) {
+ if (gpu::get_thread_id() >= gpu::get_lane_size())
+ return 0;
+
test_reduce();
test_scan();
+ test_scan_divergent();
+
return 0;
}
|
jhuber6
force-pushed
the
ScanFix
branch
2 times, most recently
from
bf81a92 to
ba202ea
Compare
Summary: The scan operation implemented here only works if there are contiguous ones in the executation mask that can be used to propagate the result. There are two solutions to this, one is to enter 'whole-wave-mode' and forcibly turn them back on, or to do this serially. This implementation does the latter because it's more portable, but checks to see if the parallel fast-path is applicable. Needs to be backported for correct behavior and because it fixes a failing libc test.
shiltian
reviewed
Feb 19, 2025
| @@ -188,6 +186,32 @@ __DO_LANE_SCAN(float, uint32_t, f32); // float __gpu_lane_scan_f32(m, x) | |||
| __DO_LANE_SCAN(double, uint64_t, f64); // double __gpu_lane_scan_f64(m, x) | |||
| #undef __DO_LANE_SCAN | |||
|
|
|||
| // Gets the sum of all lanes inside the warp or wavefront. | |||
There was a problem hiding this comment.
nit: why do you need to sink this macro?
shiltian
approved these changes
Feb 19, 2025
|
/cherry-pick 6cc7ca0 |
|
/pull-request #128085 |
swift-ci
pushed a commit
to swiftlang/llvm-project
that referenced
this pull request
Feb 21, 2025
Summary: The scan operation implemented here only works if there are contiguous ones in the executation mask that can be used to propagate the result. There are two solutions to this, one is to enter 'whole-wave-mode' and forcibly turn them back on, or to do this serially. This implementation does the latter because it's more portable, but checks to see if the parallel fast-path is applicable. Needs to be backported for correct behavior and because it fixes a failing libc test. (cherry picked from commit 6cc7ca0)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
The scan operation implemented here only works if there are contiguous
ones in the executation mask that can be used to propagate the result.
There are two solutions to this, one is to enter 'whole-wave-mode' and
forcibly turn them back on, or to do this serially. This implementation
does the latter because it's more portable, but checks to see if the
parallel fast-path is applicable.
Needs to be backported for correct behavior and because it fixes a
failing libc test.