imgproc: supports CV_SIMD_SCALABLE in pointSetBoundingRect #27479
+64
−131
Conversation
This file contains hidden or 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
|
Performance stats: SpacemiT K1 (RISC-V): i7-12700K: m2: |
fengyuentau
commented
Jun 25, 2025
| } | ||
| for(int j = 16; j < VTraits<v_uint8>::vlanes(); j*=2) | ||
| constexpr int max_nlanes = VTraits<v_int32>::max_nlanes; | ||
| int arr_minval[max_nlanes], arr_maxval[max_nlanes]; |
There was a problem hiding this comment.
We can do better than this if we have something like vtrn1q and vtrn2q in universal intrinsics, which does something like this:
v: [x1 y1 x2 y2]
vtrn1q(v, v): [x1 x1 x2 x2]
vtrn2q(v, v): [y1 y1 y2 y2]
There was a problem hiding this comment.
IMHO you can add extra branch with #ifdef CV_NEON. It gives you a chance to fix regressions on Mac M chips.
There was a problem hiding this comment.
Sure we can use CV_NEON for conditional compilation.
There was a problem hiding this comment.
Tested with vtrn and actually the performance did not go better on m2.
There was a problem hiding this comment.
Could you push it as separate branch? It may be useful with other ARM configs. I want to test it with Jetson.
There was a problem hiding this comment.
Let me give you a patch instead.
diff --git a/modules/imgproc/src/geometry.cpp b/modules/imgproc/src/geometry.cpp
index c98684f6d4..412460e0e2 100644
--- a/modules/imgproc/src/geometry.cpp
+++ b/modules/imgproc/src/geometry.cpp
@@ -739,6 +739,16 @@ static Rect pointSetBoundingRect( const Mat& points )
minval = v_min(ptXY2, minval);
maxval = v_max(ptXY2, maxval);
}
+ #if CV_NEON_AARCH64
+ int32x4_t x_minval = vtrn1q_s32(minval.val, minval.val);
+ int32x4_t y_minval = vtrn2q_s32(minval.val, minval.val);
+ int32x4_t x_maxval = vtrn1q_s32(maxval.val, maxval.val);
+ int32x4_t y_maxval = vtrn2q_s32(maxval.val, maxval.val);
+ xmin = vminvq_s32(x_minval);
+ ymin = vminvq_s32(y_minval);
+ xmax = vmaxvq_s32(x_maxval);
+ ymax = vmaxvq_s32(y_maxval);
+ #else
constexpr int max_nlanes = VTraits<v_int32>::max_nlanes;
int arr_minval[max_nlanes], arr_maxval[max_nlanes];
vx_store(arr_minval, minval);
@@ -754,6 +764,7 @@ static Rect pointSetBoundingRect( const Mat& points )
if (xmax < arr_maxval[2*j]) xmax = arr_maxval[2*j];
if (ymax < arr_maxval[2*j+1]) ymax = arr_maxval[2*j+1];
}
+ #endif
#endif
for( ; i < npoints; i++ )
{
@@ -794,6 +805,17 @@ static Rect pointSetBoundingRect( const Mat& points )
minval = v_min(ptXY2, minval);
maxval = v_max(ptXY2, maxval);
}
+ #if CV_NEON_AARCH64
+ v_int32 i_minval = v_floor(minval), i_maxval = v_floor(maxval);
+ int32x4_t x_minval = vtrn1q_s32(i_minval.val, i_minval.val);
+ int32x4_t y_minval = vtrn2q_s32(i_minval.val, i_minval.val);
+ int32x4_t x_maxval = vtrn1q_s32(i_maxval.val, i_maxval.val);
+ int32x4_t y_maxval = vtrn2q_s32(i_maxval.val, i_maxval.val);
+ xmin = vminvq_s32(x_minval);
+ ymin = vminvq_s32(y_minval);
+ xmax = vmaxvq_s32(x_maxval);
+ ymax = vmaxvq_s32(y_maxval);
+ #else
constexpr int max_nlanes = VTraits<v_int32>::max_nlanes;
int arr_minval[max_nlanes], arr_maxval[max_nlanes];
vx_store(arr_minval, v_floor(minval));
@@ -809,6 +831,7 @@ static Rect pointSetBoundingRect( const Mat& points )
if (xmax < arr_maxval[2*j]) xmax = arr_maxval[2*j];
if (ymax < arr_maxval[2*j+1]) ymax = arr_maxval[2*j+1];
}
+ #endif
#endif
for( ; i < npoints; i++ )
{
There was a problem hiding this comment.
I observed that with older version of gcc (8.3), the halide trick which deals with the tail of simd processing, can lead to some performance degradation. Newer compiler seems fine.
However, if the halide trick is dropped, the performance on i7-12700K also degrades.
asmorkalov
reviewed
Jul 8, 2025
vpisarev
reviewed
Jul 9, 2025
vpisarev
approved these changes
Jul 12, 2025
asmorkalov
approved these changes
Jul 14, 2025
|
Hm, there is strange fault on armv7: I'm debugging it right now. |
|
The test check the function behaviour with unaligned input. It was introduced by Maxim in c4ce942. armv7 is more sensitive to alignment and triggers sigbus: Newer generation just charges some performance penalties here. |
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.
Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
Patch to opencv_extra has the same branch name.