* dnn:rvv: restore the fixed block size C0 == 8
#29304 made the net-wide block size follow the RVV vector width
(defaultC0 = max(8, vlanes())), so the blocked memory layout of every tensor
depended on the host's VLEN and on the LMUL of the universal intrinsics in core.
The same model gets a different layout on a 128-, 256- or 1024-bit machine.
Restore the fixed default of 8, which is the mainstream, well-tested setting on
every platform. C0/K0 become compile-time constants again on RVV as they already
were elsewhere, so the runtime block-size plumbing and the vlanes()-based scratch
sizing are dropped, and getConvFunc_ no longer advertises block sizes the kernels
cannot handle.
The blocked kernels currently require C0 == vlanes(), which this alone does not
satisfy; the next commit removes that requirement.
* dnn:rvv: pin the vector length to the channel block
The blocked kernels assumed one channel block is exactly one vector
(C0 == vlanes()). That is not a property of the layout, it is a property of the
hardware, so the kernels only worked where the two happened to coincide: with
C0 fixed at 8 they break wherever vlanes() != 8 -- under-filling the block when
the vector is narrower (silently leaving output channels unwritten) and
overrunning it when the vector is wider (CV_Assert).
The universal intrinsics cannot express this: v_float32 is always vlanes() wide
and there is no way to make it span exactly C0 elements. So the RVV paths drop to
native intrinsics and set the vector length explicitly:
const size_t _vl = __riscv_vsetvl_e32m2(C0); // == 8 on every VLEN >= 128
LMUL=2 is required, not incidental: vlmax(e32,m2) = VLEN/32*2 >= 8 even at
VLEN=128, the narrowest RVV configuration, so _vl == C0 on every RVV 1.0 core.
One vector then spans precisely one channel block regardless of VLEN, and the
C0-vs-vlanes asserts and VLEN-dependent branches are removed rather than extended.
The existing int8 kernels already choose LMUL=2 for the same reason
(conv2_int8_kernels.simd.hpp).
This is a native per-kernel choice and is independent of the LMUL the universal
intrinsics are built with; dnn no longer reads the vector width of
intrin_rvv_scalable.hpp at all.
conv, deconv, maxpool, avgpool and depthwise use native intrinsics. batch_norm
stays portable and instead replicates the C0-wide coefficient pattern across the
vector, which keeps full lane utilisation where a fixed vl would idle most of it.
The x86/ARM paths are restored to their pre-#29304 form and are unchanged.
Tested on SpaceMIT K3 (X100 VLEN=256, A100 VLEN=1024), opencv_test_dnn: no
regressions against upstream on either core.
* core:rvv: switch scalable universal intrinsics from LMUL=2 to LMUL=1
Move the RVV universal-intrinsic types (v_uint8..v_float64) from m2 (paired
registers, 16 usable) to m1 (individual registers, 32 usable) per #29454, to
relieve register pressure in the complex kernels that rely on universal
intrinsics. Simple kernels that regress can take native RVV branches; the complex
ones cannot, so the default should suit them.
Only intrin_rvv_scalable.hpp changes. The widening/narrowing cascades and
max_nlanes are adjusted, and helpers that spell an LMUL into the intrinsic name
shift down one step (m2->m1, m1->mf2): v_popcount, v_lut (vector and f64 paths),
the f64 dot-product reduction (#27003), v_matmul, and the byte-indexed v_lut
added by #29250, whose __riscv_vluxei8_v_u8m2 no longer matches the m1 base type.
VLEN=1024 needs no change: under m1, max_nlanes = CV_RVV_MAX_VLEN/SZ equals
vlanes() at 1024, and CV_RVV_MAX_VLEN already defaults to 1024. Verified on a
1024-bit core.
Tested on SpaceMIT K3, which exposes both a 256-bit (X100) and a 1024-bit (A100)
core. opencv_test_core: 5605/5612, the 7 failures (filestorage_base64, globbing)
non-SIMD and failing identically on the m2 baseline. Perf (x-factor = m2/m1,
>1 = m1 faster), median >= 300us: core 0.966 -> 0.996 and imgproc 0.950 -> 1.038
as VLEN goes 256 -> 1024, i.e. m2's width advantage disappears once both are wide
while m1's register headroom persists. m1 wins resize (1.50/1.41), Exp (1.28),
gaussianBlur (1.23), norm (1.22), scharr (1.20); m2 still wins the light
memory-streaming kernels addWeighted (0.76), compare (0.82) and dot (0.87), which
are the natural candidates for native RVV branches.
* Revert "dnn:rvv: pin the vector length to the channel block"
This reverts commit 40587c6dd9.
* dnn:rvv: disable CV_SIMD_SCALABLE in the new-engine conv kernel
Under the m1 switch the RVV conv kernel breaks at VLEN=128: it assumes
K0 == vlanes(), but m1 makes vlanes()=4 while the block stays C0=K0=8, so it
computes only half of each output block. Temporarily disable the RVV
(CV_SIMD_SCALABLE) branches so conv runs scalar on RVV, exactly as #29180 did for
the other new-engine layers. The vectorized path will be restored in a follow-up.
Only conv is affected: maxpool/avgpool/depthwise/deconv/batch_norm already handle
C0 == 2*vlanes (the SSE/NEON case) and stay on their universal-intrinsic paths.
* dnn:rvv: document the supported VLEN range and the v_setvlmax follow-up
With the fixed C0=8 the RVV blocked kernels cover VLEN=128 and 256 (C0 >= vlanes);
VLEN>=512 (C0 < vlanes) is not yet handled. Add TODO notes recording this and
pointing at the planned fix: a portable v_setvlmax<VT>(n) universal intrinsic that
caps the vector to the block (no-op on fixed-width ISAs, sets the vl/mask on
RVV/SVE), with an optimized multi-pixel variant as a later step, possibly an RVV
HAL for the new dnn engine. See the #29493 discussion.
* dnn:rvv: disable CV_SIMD_SCALABLE in max/avg pooling and depthwise
These three blocked kernels assert (C0 == nlanes || ...) at VLEN>=512, where the
fixed C0=8 is narrower than the vector register — breaking whole networks
(AlexNet/ResNet/YOLO) on wide RVV cores, though CI (VLEN=128) is unaffected.
Disable their RVV (CV_SIMD_SCALABLE) paths so they run scalar on RVV at every
VLEN, matching #29180's scope. Verified on a SpaceMIT K3: this removes the 92
VLEN=1024 assert failures, with no change at VLEN=128/256.
deconv and batch_norm are left enabled: they degrade to scalar at VLEN>=512
(deconv's C0 % vlanes guard, batch_norm's vector-loop bound) without asserting,
so they stay correct and keep their 128/256 vectorization. All of these
re-vectorize uniformly via the v_setvlmax<VT>(n) follow-up.
* experimental new arithmetics; work-in-progress
* continue working on new-gen arithmetic expressions
* * improved performance of the new add on small arrays
* added sub
* extended tests
* fixed potential bug when adding multi-channel array and a single-channel scalar
* improved const handling
* * added copyMask
* added mul/dev (without scale so far)
* * accelerated mul
* addedd scale to mul and div
* * done substantial refactoring; however a few more rounds of refactoring are ahead.
* added min, max, absdiff, addweighted.
* improved performance of the new arithmetic functions, but some of them are still slow, e.g. operations with mask have some bugs (that affect speed, not accuracy).
* * further (significantly) accelerated several functions, especially on small arrays: mul, binary ops with mask
* further polished the new arithmetic engine
* started integration of the new element-wise arithmetic engine into core
* big step forward. We now use the new engine inside cv::add, subtract, multiply, divide, absdiff, min and max.
* big progress:
* added bitwise operations
* fixed and accelerated compare
* ported regression tests to test new broadcasting behaviour of arithmetic functions
* lot's of improvements in compare, divide, addWeighted!
* lot's of small and big performance improvements in the new arithmetics
* * some more optimizations; parsing texpr-expressions is now faster as well
* port new_arithm to Linux/x86: dispatch guards, scalar-Mat compat fallback, dnn shape-contract fixes
Core:
- arithm.simd.hpp: CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY guards (the file is included
once per dispatched mode on x86), vx_load_expand instead of the 128-bit v_load_expand,
VTraits::vlanes() instead of ::nlanes
- arithm.cpp/precomp.hpp: compat fallback for scalar-like Mat operands (1x1, 1xcn/cnx1,
4x1 CV_64F - java/python tuples, operator-(Mat, Matx)): treated as a per-channel scalar
ONLY when the shapes are not broadcast-compatible, so every valid numpy-style broadcast
keeps its meaning and calls that would otherwise throw get the 4.x semantics
DNN (fallout of the stricter shape semantics, found by the new engine):
- dict.hpp: DictValue relied on fresh AutoBuffer having size()==fixed_size; allocate explicitly
- batch_norm: weights_/bias_ are 1-D [n] now; 0/1-D forward runs on exact-shape 1-D views
- net_impl2: extend the post-forward sanity check to non-temp outputs - a layer that
reallocates its preallocated output tensor now fails loudly instead of silently
detaching the result from the graph
- LSTM/LSTM2 batchwise (layout=1): getMemoryShapes now matches what forward() writes
(ONNX: Y=(batch,seq,dirs,hid), Yh/Yc=(batch,dirs,hid)); forward assembles seq-major
results in a local buffer and transposes INTO the preallocated outputs in place
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* u8/s8 multiply: exact integer SIMD path on non-FP16 builds (3-10x vs 5.x)
The unit-scale branch of vecBinaryKernel already supported a separate work-vector
type Wvec1 (used by the ARM f16 build and by u16/s16 everywhere), but on x86 the
u8/s8 same-type multiply still went through the f32 hub. Route it through
v_uint16/v_int16: products of 8-bit values fit exactly (255^2 < 2^16), the
saturating pack on store gives bit-exact results at half the vector traffic.
Also fix a latent kernel bug this exposed: the unit-scale branch stepped by
Wvec's lane count while loading/storing Wvec1 vectors. All previous Wvec1
instantiations had equal lane counts, but u8's v_uint16 has 2x the lanes of
v_float32 - the pairs overlapped (50% redundant work) and the tail backoff
could write VECSZ bytes past the row end. The branch now derives its step,
offsets and tail condition from Wvec1 itself.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* vecBinaryKernel: constexpr Op::useScalar instead of a runtime-only scale check
Every binary op functor now declares whether it consumes the scale scalar
(params[0]): true only for mul and the two div variants. Ops that ignore it
(add/sub/min/max/absdiff) take the fast 2-arg branch unconditionally - the
'scalar == 1' check used to fail for them (their params[0] is 0), sending them
through the preproc branch, and the condition now folds at compile time.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* restore cv::hal::mul8u as a wrapper over the element-wise engine
The symbol is still declared in core/hal/hal.hpp and called directly by external
code (the G-API fluid backend in opencv_contrib), but its implementation went
away with the old arithm kernels. Forward it to getMulFunc(CV_8U, CV_8U) - with
scale==1 it lands on the new exact integer SIMD path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* silence every new-arithm warning reported by CI (ARM64/Mac) and gcc 15
- arithm.cpp: bitwise_op_ocl and the actualScalarDepth/coerceTypes helpers are
consumed only by the OpenCL paths - guard them with HAVE_OPENCL; haveScalar in
cv::compare is read only inside CV_OCL_RUN - CV_UNUSED for OpenCL-less builds
- arithm_expr.hpp: declare getBitwiseFunc/getNotFunc/getAddWeightedFunc next to
the other per-op entry points (-Wmissing-prototypes in arithm.dispatch.cpp)
- arithm.simd.hpp: define CV_SIMD_16F to 0 when FP16 SIMD is absent (-Wundef);
{}-init the expandScalar staging buffers (-Wmaybe-uninitialized: they are
fully written before use, but the compiler cannot prove it with runtime
vector widths); rename the compare kernel's lambda parameter (-Wshadow)
- arithm_expr.cpp: rename the exec tile-lambda's hot-field locals that shadowed
TExpr members and outer locals (-Wshadow)
- test_new_arithm_extensive.cpp: rename the name-generator lambdas' parameter
shadowing the INSTANTIATE macro's own (-Wshadow), drop an unused variable
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* restore 4.x scalar semantics for the bindings' 4x1 CV_64F Scalar columns
The python/java bindings materialize numbers and tuples as a 2-D 4x1 CV_64F
Mat - or UMat, when the call carries UMat arguments. Three CI-reported python
failures came from those pseudo-scalars reaching the engine as arrays:
- absdiff(int_arr, 0): the (4,1) column is broadcast-COMPATIBLE with a 1-D
array, so numpy semantics silently won - an outer-product f64 result instead
of the int per-channel-scalar one;
- subtract(u8 4x8x4, (40,)): same, by the rows==4 coincidence;
- multiply(UMat, 2., dst=UMat): the scalar arrives as a UMAT, which the
scalar detection did not recognize at all.
isScalarArg now treats the exact bindings shape - 2-D 4x1 CV_64F single-channel
Mat/UMat against a <=4-channel array - as a scalar UNCONDITIONALLY (a 1-D [4]
array has dims==1 and still broadcasts). One exception, decided in arithm_op:
when the partner is itself a tiny scalar-shaped array, both are honest data and
ride the broadcast (compare(Mat 4x1, Mat 1x1) - issue #8999 - stays elementwise).
Small-array discipline, this all runs per engine call: the probes read Mat/UMat
fields directly (rows == 4 alone rejects almost everything, no _InputArray
getter dispatch), and a UMAT scalar's 32 bytes are copied into a caller-stack
buffer - no heap, no getMat mapping. Measured: no latency change on 4x4/16x16
element-wise calls.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cv::texpr: std::string_view -> const std::string& in the public API
string_view in an exported signature breaks some CUDA toolchain builds, and for
short expression strings the difference is immaterial (SSO, parsed once). The
parser internals keep string_view - the argument converts implicitly. Also drop
the now-unused <string_view> include from cvstd.hpp, so the header does not
reach every nvcc TU.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* compare boundary-rewrite: fixed 4-slot kind/bound arrays instead of AutoBuffers
A CONST operand is capped at 4 channels (addConst), so the per-channel
kind/bound staging needs no dynamic buffers - plain int[4]/double[4], with a
CV_Assert on the contract. This is also what gcc's -Wmaybe-uninitialized was
flagging (it could not see the AutoBuffer's inline storage get filled).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* element-wise engine: unary math kernels (sqrt/exp/log/sin/cos/tanh/erf/relu) + select
New dispatched pair math.simd.hpp / math.dispatch.cpp - the unary/ternary sibling
of arithm.simd.hpp:
- vecUnaryKernel: T -> T over f16/bf16/f32/f64 on top of the intrin_math
primitives (v_exp/v_log/v_sin/v_cos/v_sqrt/v_erf/v_max). f32/f64 compute
natively, f16/bf16 ride the f32 hub inside the kernel (vx_load_pair_as /
v_store_pair_as) - no materialized casts. Continuity collapse + the halide
right-edge backoff, suppressed in-place (it would re-apply Op to
already-written values). tanh = (e^2x-1)/(e^2x+1) with the input clamped to
+/-10 (f32) / +/-20 (f64) - unclamped saturation hits inf/inf = NaN. erf has
no f64 SIMD primitive: std::erf per lane.
- selectKernel(mask, x, y): 1-byte mask expanded to lane width and tested
against zero in the INTEGER domain (immune to DAZ/FTZ), branches of any
depth by element size, broadcast branches supported.
- emitUnary: math over a float input is T -> T now (f16 in -> f16 out, native
kernel when input and result depths match); integer inputs still compute in
the float domain and land in f32.
- emitTernary/select: literal branches are typed via typedConstFrom (an
OP_CAST of a depth-less flex const crashed); a non-1-byte mask is normalized
by an explicit "mask != 0" compare, never a value cast.
texpr already parsed the function names - they now execute. Tests: per-depth
accuracy of all 8 ops against the double std:: reference, integer input,
in-place, select over 4 depths / const branch / float mask.
Perf vs the classic kernels (1920x1080 f32, 16 threads, AVX2): exp 3.7x,
log 2.6x, sqrt 5.7x faster; polarToCart expressed as (r*cos(a), r*sin(a)) 4.0x.
Accuracy improves too (max rel err vs f64 reference): exp 8.1e-8 vs 2.1e-7,
log 8.0e-8 vs 1.5e-7.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* engine: OP_COPY_MASK folded into OP_SELECT; selectKernel moved to arithm.simd.hpp
copyMask(dst, mask, src) is select(mask, src, dst) - one masking primitive
instead of two. The compiler emits the masked-op tail as
addInsn(OP_SELECT, mask, r, out, out): the output slot rides as both arg2 and
the result, so unmasked elements are preserved by reading them back through
the b-branch. OP_COPY_MASK, copyMaskKernel and getCopyMaskFunc are gone.
selectKernel (moved from math.simd.hpp to arithm.simd.hpp) inherits every
copyMaskKernel optimization:
- the interleaved multichannel fast path (2..4 channels under a per-pixel
mask: expand the mask once per VECSZ rows, v_store_interleave across lanes);
- the per-row scalar path with the row-skip when the selected source row IS
dst (the "leave the output untouched" half of copyMask);
- plus the select-specific ones: branch broadcasts (stepx == 0) and the
right-edge tail backoff under dst-aliases-a-branch - safe because re-running
select over already-blended elements is idempotent; only dst == mask keeps
the backoff off (the store would rewrite mask bytes before the re-read).
Masked-add perf is on par with the old copyMask (1280x720, 1 thread: 8UC3
204 -> 198 us, 32FC3 1395 -> 1344, 8UC1/32FC1 within noise). Full core suite
24117 green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* engine: dedicated vectorized pow kernel (moved from arithm to math.simd.hpp)
pow was the last scalar-only binary op (scalarBinaryKernel + std::pow, f32/f64
only). The new powKernel keeps exact std::pow semantics and is T x T -> T over
all four float depths (f16/bf16 via the f32 hub):
- scalar exponent (the dominant call shape - texpr literals ride as 0-dim
broadcast consts) is dispatched PER ROW to the special cases:
y==2 -> x*x, y==3 -> x*x*x, y==0.5 -> v_sqrt, y==1 -> copy, y==0 -> fill 1;
- everything else - including a per-element exponent array - runs the general
vectorized exp(y * log(x)) path, valid for x > 0; a vector pair containing
any x <= 0 lane falls back to scalar std::pow for that pair (v_check_any),
which preserves every std::pow subtlety: signed results for integer y on
negative bases, NaN for fractional y, the x == 0 family;
- no right-edge tail backoff: pow is not idempotent, in-place calls finish
rows in the scalar tail.
Perf vs the classic cv::pow (1920x1080 f32, 16 threads, AVX2): p=2 1.4x
(classic special-cases it too), p=3 6.1x, p=0.5 6.3x, fractional p 4.5x with
slightly better accuracy (6.1e-7 vs 7.4e-7 max rel err). Tests: exponent
sweep 2/3/0.5/1/0/2.5/-1.5 vs the double std::pow reference on f32/f64,
negative bases (exact signed cubes, NaN for fractional), array exponent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* powKernel: halide right-edge tail backoff in every SIMD loop
Same shape as vecBinaryKernel: the final partial vector re-processes
[width - VECSZ*2, width) instead of finishing scalar, suppressed when dst
aliases an input (pow is not idempotent - the overlap region must be
recomputed from an untouched source, which the no-alias case guarantees).
Modest measured win (~2% on ROI rows for the special-cased exponents; the
general path tail was already cheap - modern libm powf is fast), no
regressions; mainly aligns the kernel with the house style, where every
SIMD loop ends vector-wide.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix MSVC 2019 C2975: function-local constexpr as a template argument inside a lambda
MSVC 2019 loses the constexpr-ness of function-local constants (LOCAL_OPS,
MAX_DIMS, ...) when they are used as template arguments inside a lambda body
(AutoBuffer<Slice, LOCAL_OPS> / std::array<int, MAX_DIMS> in the parallel
bodies of BroadcastOp::run and TExpr::exec). Hoist them to namespace scope -
no behavior change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ocl_arithm_op: route 16U multiply to the CPU engine on Apple OpenCL
The Apple OpenCL driver miscompiles the 16U multiply kernel: products near
the top of the u16 range come back wrapped instead of saturated (CPU vs GPU
NORM_INF up to 65535 in OCL_Arithm/Mul.Mat CV_16U cases). The same arithm.cl
kernel is correct on Intel NEO and NVIDIA drivers - verified not to reproduce
on Linux/Intel iGPU - so gate the decline to __APPLE__ only; the CPU engine
computes 16u multiply exactly (integer SIMD path).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* engine: neg/abs as compositions, clamp kernel, ** and ?: operators, abs(a-b) peephole
- OP_NEG and OP_ABS need no kernels: neg = sub(0, a), abs = absdiff(a, 0) -
including the engine absdiff auto-type rule (signed |a| lands in the
UNSIGNED type of the same width: |SHRT_MIN| fits u16 exactly instead of
saturating; NB the public cv::absdiff auto depth keeps the source type for
4.x compatibility - values agree, the depth rule is the engine own).
- peephole: abs(x - y) rewrites to absdiff(x, y) ALWAYS. On integers the
literal semantics differ (the subtract saturates first: u8 gives
max(x-y, 0)), but whoever writes abs(a - b) means absdiff - we deliberately
hand out the useful semantics instead of the saturation artifact. The just-
emitted OP_SUB is retired via the moveToOutput manoeuvre, so the program
shrinks to the single absdiff instruction. abs(x), abs(x - 0) and
absdiff(x, 0) all give one result.
- OP_CLAMP kernel (arithm.simd.hpp): v_min(v_max(x, lo), hi) over
u8/s8/u16/s16/u32/s32/f32 (+f64 with 64-bit SIMD), scalar f16/bf16/64-bit
ints; lo/hi may broadcast (the common clamp(img, a, b) shape) or be full
arrays; the tail backoff stays on under dst-aliases-x (clamp is idempotent).
emitTernary types literal bounds via typedConstFrom (same flex-const crash
select had) and keeps the auto result type pinned to x.
- parser: "a ** b" == pow(a, b), precedence above * /, RIGHT-associative
(a ** 2 ** 3 == a ** 8); "cond ? a : b" == select(cond, a, b), precedence
below everything, right-associative chains (f1 ? a : f2 ? b : c) work
without parentheses. parseTernary() is the expression entry point now.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cv::exp/log/sqrt on the engine via math_op; hal functions wrapped as engine kernels
math_op is the master function of the unary math family (the arithm_op
analogue): same-shape same-type output over f16/bf16/f32/f64 (classic
exp/log accepted f32/f64 only - the half floats are new), two tiers:
- small (<= 100000 elements) and continuous: call the kernel DIRECTLY over
the flattened data - no TExpr, no broadcastOp, no parallel_for setup;
- everything else: the usual single-instruction program via compile()/exec()
(parallelism for large arrays, real steps for ROIs).
getMathFunc routes OP_EXP/OP_LOG at f32/f64 through the full cv::hal stack -
an external vendor HAL (CALL_HAL), IPP, or the built-in table kernels,
whichever is installed - by wrapping hal::exp32f/exp64f/log32f/log64f as
engine kernels with the function pointer in TKernel::userdata, the same
mechanism castKernel uses for core BinaryFuncs. The engine adds tiling and
parallelism on top, so every tier gets the best available scalar-span
implementation. v_exp/v_log remain for f16/bf16 (the f32 hub) and the ops
hal has no entry points for.
v_log_default_32f: the degree-8 polynomial is evaluated by Estrin pairing
(4 dependent levels) instead of an 8-FMA Horner chain (~5% on the f16 hub
path). An exp64 Taylor-without-division rewrite was tried and benched SLOWER
than the Cephes Pade scheme (the evaluation is FMA-throughput-bound, and
vdivpd pipelines well enough) - reverted; a table-based reduction is the only
way further there.
cv::exp f32 (16 threads, AVX2+IPP build), old -> new: 640 elements
0.16 -> 0.15 us, 16k 2.79 -> 2.49, 640x480 47 -> 20, 1920x1080 452 -> 60 us
(the old CPU loops were single-threaded); f64 exp 1080p 1295 -> 161 us.
No size regresses; small arrays now run at installed-HAL speed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* getMathFunc: IPP tier + raw-HAL probing; the exp/log table kernels are deleted
The raw cv_hal_* entry points return int for a reason: without an installed
HAL they are stubs returning CV_HAL_ERROR_NOT_IMPLEMENTED. getMathFunc now
selects the exp/log implementation in three tiers:
1. HAVE_IPP && ipp::useIPP(): ippsExp/ippsLn through thin int adapters (IPP
is not routed through the cv_hal_ hooks, so it needs its own tier);
2. the raw cv_hal_exp32f/... hook, PROBED once with a 1-element call on the
safe input 1.0 (cached in magic statics): implemented -> wrapped as an
engine kernel with the function pointer in TKernel::userdata;
3. the engine own v_exp/v_log kernels.
Whichever wins, the engine adds tiling and parallelism on top.
The EXPTAB/LOGTAB table kernels and their tables (~790 lines in
mathfuncs_core.simd.hpp + mathfuncs.cpp) are DELETED: they benched within
~15% of v_exp/v_log, not worth a second implementation. The public
cv::hal::exp32f/exp64f/log32f/log64f keep their contract - CALL_HAL, then
IPP, then the built-in implementation - but the built-in is now the engine
vector kernel via ew::mathSpanEngine (one contiguous span, exported from
math.dispatch.cpp).
All unary math kernels (vec/scalar/hal wrappers, pow) also handle the
vertical-broadcast tile (s0y == 0, a row expanded into a matrix): the first
row is computed, the rest are memcpy of it - transcendentals cost far more
than a row copy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* texpr: hypot(x, y) binary op (alias: mag)
hypot = sqrt(x^2 + y^2), NAIVE like cv::magnitude (not the overflow-safe
std::hypot), computed in the float work type; kernels for the four float
depths only (T x T -> T; integer inputs ride the usual f32-compute + cast).
A 10-line EwHypot functor on top of vecBinaryKernel in arithm.simd.hpp -
broadcast branches, continuity collapse and the tail backoff come for free.
Registered in the parser as both "hypot" (the C/numpy name) and "mag" (the
cv::magnitude-flavored alias). A building block for the future cartToPolar.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* texpr: atan2(y, x) binary op - radians, standard C range
v_atan2 (arithm.simd.hpp, generic over the universal-intrinsic float vector):
the fastAtan2 minimax polynomial from mathfuncs_core v_atan_f32 reworked to
plain radians - the 180/pi factor dropped from the coefficients and the C
quadrant logic instead of the [0, 360) wrap, so the result matches std::atan2
over (-pi, pi]. Measured absolute accuracy ~1.6e-4 rad. (v_atan_f32 itself is
untouched - cv::phase/fastAtan2 keep their degree semantics.)
EwAtan2 rides vecBinaryKernel: f16/bf16/f32 through v_atan2 (the f32 hub),
f64 through exact scalar std::atan2. arg0 = y, arg1 = x, like std::atan2;
float depths only, same emitBinary policy as pow/hypot. Parser name "atan2".
Together with hypot this completes the cartToPolar building blocks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix the RISC-V RVV build and two ARM64 warnings
The new f64 kernel registrations (hypot, pow, the unary math family) gate on
CV_SIMD_64F || CV_SIMD_SCALABLE_64F, but the vx_setall_as(const double*,
v_float64&) helper family in arithm.simd.hpp was still CV_SIMD_64F-only -
scalable platforms (RVV) have v_float64 with CV_SIMD_64F == 0, so
vecBinaryKernel<double, ...> failed to instantiate there. Widen the helper
gate to match (verified with a riscv64 rv64gcv cross-build of opencv_core -
the engine f64 paths now vectorize on RVV instead of not compiling).
cv::exp/cv::log: the depth local is consumed by CV_OCL_RUN only - CV_UNUSED
for OpenCL-less builds (ARM64 -Wunused-variable).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* silence the remaining ARM64 gcc warnings
- compare boundary-rewrite: {}-init the fixed kind/bound arrays (filled for
every channel used below, but gcc cannot prove it across the cn <= 4 loop);
- cv::exp/log: [[maybe_unused]] on the depth local (consumed by CV_OCL_RUN
only), instead of the CV_UNUSED idiom;
- AutoBuffer::reserve: a targeted -Wmaybe-uninitialized suppression around
the live-element copy loop - only [0, sz) is read, all written before, but
gcc inlining a grow-from-inline-storage chain cannot see that. An
annotation for the analyzer, no behavior change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* saturating 32-bit add/sub kernels; cv::texpr python binding; two CI warnings
- v_add_sat/v_sub_sat for v_int32/v_uint32, local to arithm.simd.hpp for now
(the plan is to grow them into proper universal intrinsics later): NEON
single-instruction vqadd/vqsub, elsewhere the Hacker Delight bit tricks
over universal intrinsics (u32 add is 2 ops: or with the wrapped-compare
mask). EwAdd/EwSub overload vec() for the 32-bit lanes and getAddSubFunc
routes 32S/32U T->T through vecBinaryKernel instead of the former pure
scalar kernel. Semantics unchanged - the scalar int64 tail already
saturated; directed boundary tests added (both rails, 0 - INT_MIN, u32
cases, a full-range random block vs an exact int64 reference).
640x480 32S add: 0.48x of 5.x -> parity (memory-bound); 1080p: 6-8x.
- cv::texpr becomes CV_EXPORTS_W: python gets cv.texpr(expr, [inputs]) ->
tuple of ndarrays, so `res, = cv.texpr(...)` and `mag, ang = cv.texpr(...)`
unpacking both work. modules/python/test/test_expr.py covers arithmetic,
the fused abs(a-b), casts, broadcasting, ?: and ** operators, math
functions vs numpy, clamp, named temporaries, tuple outputs, the one-line
cartToPolar and the int32 saturation cases.
- warnings: {}-init the parser args array (gcc -Wmaybe-uninitialized on
Ubuntu 20/22); the compare short-row block gates sizeof(T) <= 4 as
constexpr so the f64 instantiation does not leave set-but-unused locals
(gcc 9 -Wunused-but-set-variable).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* arithm_op: direct-kernel fast path for small continuous arrays
Building and compiling the 1-instruction program plus the BroadcastOp setup
costs ~40-250ns per call - negligible on big images, dominant at 127x61-class
sizes where the classic 5.x functions were 1.3-2x faster. Mirror math_op two
tiers in arithm_op: two same-type same-shape continuous arrays, no mask, no
scalar, result depth == input depth, <= 100k elements -> call the T x T -> T
kernel directly over the flattened elements (checks ordered cheapest-first).
Applies to add/subtract/min/max/absdiff/multiply/addWeighted/and/or/xor;
compare and divide lower to more than a single kernel (boundary rewrites, int
guards) and keep the ordinary path. addWeighted falls through automatically
for the 32/64-bit int types whose lowering is wide-compute + cast
(getElemwiseFunc returns no direct kernel there).
127x61 vs 5.x, was -> now: add/subtract 8UC1 0.76x -> 1.3x, min/max u8
0.6x -> ~1x, addWeighted 1.0x -> 1.1-1.4x (32SC1 stays 4.5x); the one
remaining laggard is add/sub 32SC1 (0.74-0.82x) - the price of the new
SATURATING semantics (7-instruction AVX2 emulation vs the wrapping single
add of 5.x; single-instruction on NEON).
The 127x61 size is ADDED PERMANENTLY to the arithmetic/addWeighted/compare
perf grids: per-call overhead regressions in these base functions must be
caught by CI, not discovered by users.
dst creation goes through createSameSize (whole-shape transfer including
layout and future metadata, not piecemeal dims+sizes) here and in math_op.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cv::pow rebuilt on the engine; integer-exponent and 1/sqrt(x) kernel branches
Routing: p = 0/1/2 keep their early special cases (fill/copy/multiply); an
INTEGER array with an INTEGER power keeps the classic iPow multiply chain -
bit-exact compatibility, including its wrap-around quirks (iPow squares in
int, so e.g. pow(255,4) on u8 wraps negative and saturates to 0 - somebody
may rely on that). Everything else goes through the engine with the math_op
two-tier scheme: small continuous arrays call powKernel directly (the
exponent rides as a broadcast T scalar), the rest run the tiled parallel
program. Integer arrays with fractional powers compute in the float domain
and saturate back; the 32U/64-bit depths (classic iPow asserted on them) and
f16/bf16 (the classic float path misread them) now just work.
powKernel gets two new per-row exponent branches:
- p == -0.5: 1/v_sqrt(x) (the classic path used IPP ippsInvSqrt_A21, a
21-bit approximation; ours is exact - slightly slower on small arrays,
4.5x faster at 1080p via parallelism);
- any other INTEGER p (|p| <= 65536): LSB-first binary exponentiation, the
same multiply chain and order as iPow, fully vectorized - a few ulp
accurate vs ~2e-7 of exp(p*log x), and exact on non-positive bases (the
sign falls out of the multiplies, 0^negative divides to inf) - no scalar
patching.
cv::pow f32 vs 5.x: p=0.5 365 -> 61 us at 1080p (6x), p=3 7.7x, p=5 7x AND
faster at every size (the old scalar chain: 0.81 -> 0.47 us at 127x61),
p=2.5 5.3x. The s16^5 iPow path is untouched (118.7 == 118.5 us).
pow_exponents accuracy tests extended to 11 exponents x f32/f64 against the
double std::pow reference.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* perf: SANITY_CHECK_NOTHING for the tests whose grids got the 127x61 size
The 127x61 entry added to guard per-call overhead has no regression data in
opencv_extra, so the legacy SANITY_CHECK in addWeighted/compare failed on CI
(locally it passes silently without the test-data path). Accuracy of both
functions is covered by the accuracy suite; the perf tests should measure
time. PatchNaNs/finiteMask keep their SANITY_CHECK - their grids are
untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix a temp-buffer double release in the liveness pass ("d*d" crash)
When the same temp is passed as SEVERAL arguments of its last-use
instruction (e.g. the named-intermediate expression "d = {0} - {1}; d*d",
where the MUL consumes slot d twice), the buffer-reuse scan pushed the
temp's physical buffer onto the free list once per argument. That
overflows the ntemps-sized freeBufs array (caught by the AutoBuffer range
check in Debug: python test_expr.py::test_named_temporary) and, in larger
programs, would hand the same physical buffer to two live temps.
Release the buffer once by retiring lastUse[t] after the first hit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* executor: recalibrate opCost to vectorized cycles + clamp the stripe count
The per-element op costs fed into the parallel_for_ stripe hint were
scalar-era estimates (~20x above the vectorized reality: the atan2/exp
polynomials run at ~1.5 cycles/element, not 30). The hint therefore split
transcendental/divide work into hundreds of ~1 us jobs, which the
macOS/GCD backend dispatches poorly under sustained load, on top of the
P/E-core equal-share straggler effect. Measured on M4 Max (12P+4E),
sustained medians @1920x1080 f32: texpr atan2 320 -> 133 us, cv::exp
280 -> 141 us, cv::log 301 -> ~200 us, cv::pow(x,2.5) 388 -> 376 us;
the PR tables' math rows improved ~1.5-2x across the board.
- opCost is now in units of ~1/4 cycle/element of the SIMD kernels:
cheap ops 1 (unchanged), div/sqrt/hypot/convert_scale 10 -> 3,
transcendentals 30 -> 6.
- the stripe hint is clamped by min(4*nthreads, max(32, 3*nthreads)):
~4 stripes/thread is plenty of granularity for element-wise work, and
the ceiling is 32 pieces except on machines with many (heterogeneous)
cores, where anything coarser than ~3 pieces/worker turns the slow
cores into equal-share bottlenecks (measured: 32 stripes on 16 threads
is the worst point of the curve - 193 us vs 137 us at 48 for atan2).
getNumThreads() is clamped from below (WINRT/plugin backends may
report 0).
Not addressed here (needs cross-machine data, M2/M3 Ultra): streaming
memory-bound ops saturate the M4 Max fabric at ~8 fat stripes and E-core
participation only adds contention - a cost model cannot express that;
candidate follow-up is a bytes-aware clamp or a GCD-backend-level fix.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* multiply: v_mul_sat integer kernels (full product clamped to the type)
v_mul_sat(V, V) -> V for u8/s8/u16/s16/u32/s32 - the full-precision
product clamped to the lane type, which is exactly cv::multiply's integer
semantics at scale == 1. Local to arithm.simd.hpp for now, next to
v_add_sat/v_sub_sat, to be promoted into proper universal intrinsics
later. NEON: widening vmull + saturating narrow (vqmovn); other backends:
the portable v_mul_expand + saturating v_pack composition for 8/16-bit
lanes. 32-bit lanes have no universal widening multiply (no v_mul_expand
for s32), so the 32-bit integer fast path is NEON-only for now and the
other backends keep the previous f64 work-vector kernels (which measure
well on x86 with IPP-free AVX2).
EwMul::vec() now routes the integer lane types through v_mul_sat, and
getMulFunc_ passes the NATIVE lane vector as the scale==1 fast-path type:
whole-register loads/stores, the widening happens inside the multiply.
Replaces both the half-register widening loads (u8/s8/u16/s16) and the
scalar-equivalent f64 path for 32S/32U on NEON.
M4 Max, 640x480 (the sizes where the old kernels lost to carotene):
8U 22.0 -> 13.6 us, 8S 15.0 -> 11.1, 16S 21.2 -> 17.4, 32S 84.8 -> 30.5
(parity with the classic path everywhere, 32S was 0.38x). 1920x1080:
8S 1.30x -> 1.78x, 16S 2.57x -> 2.98x, 32S 2.20x -> 4.02x vs 5.x.
Correctness: exhaustive 8-bit (all 65536 pairs per sign), directed
saturation corners for 16/32-bit (46341^2, INT_MIN*-1, 65536*65536, all
sign combinations) against an exact int64 reference.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* NEON: make v_cvt_f64(v_int32) exact (was via f32, losing bits > 2^24)
The NEON implementations of v_cvt_f64/v_cvt_f64_high for v_int32 did
s32 -> f32 -> f64 (vcvt_f32_s32 + vcvt_f64_f32), silently rounding any
|x| > 2^24. Every vectorized f64 work path with int32 inputs on AArch64
was affected: the engine's addWeighted/divide 32S kernels, convertTo
32S -> 64F, etc. Found via addWeighted 32SC1 on values ~1e9: max error
was 32 vs the exact double reference (the classic carotene path is worse
still - it computes in f32 end-to-end with f32-truncated weights, max
error 96 on the same data).
The exact sequence sxtl + scvtf (vmovl_s32 + vcvtq_f64_s64) is the same
2 instructions, so there is no cost. addWeighted 32SC1 on the engine now
matches the exact-double reference bit-for-bit and stays at parity/1.4x
vs the classic path (640x480/1920x1080).
Pre-existing upstream bug (same code in 4.x) - worth a standalone
backport with directed large-value tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* math_op: drop the temporary direct-IPP tier for exp/log
Upstream moved the IPP math wrappers into the hal/ipp HAL module (cv_hal_exp32f/
log32f & co now resolve to ipp_hal_* which honor cv::ipp::useIPP via
CV_HAL_CHECK_USE_IPP). The engine's single probeHalUnary(cv_hal_*) probe already
picks that up uniformly, so the stopgap #ifdef HAVE_IPP ippExp/ippLog tier and its
ipp::useIPP() branch in getMathFunc are now redundant - removed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* arithm: restore cv::hal::and8u/or8u/xor8u/not8u as engine wrappers
These public CV_EXPORTS entry points (core/hal/hal.hpp) lost their definitions when
the bitwise ops moved to the element-wise engine, but they are still declared and called
by other modules (opencv_objdetect's aruco) and external code - the link broke with
undefined references to cv::hal::and8u/xor8u. Restore them as thin forwarders over the
engine's byte-wise bitwise kernels (getBitwiseFunc / getNotFunc), mirroring the existing
mul8u wrapper. CV_Assert guards the kernel lookup.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* blobdetector: guard empty contour when computing blob radius
The new AutoBuffer leaves its tail uninitialized for trivial types, which surfaced a
-Wmaybe-uninitialized in findBlobs where the median of per-point distances is read. Use a
std::vector, default the radius to 0, and compute the median only for a non-empty contour -
no unproven size invariant.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
FP8 support in core module #29369
Core part of https://github.com/opencv/opencv/issues/29313
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
Extracting IPP integaration as HAL for Canny #29434
Backport of : https://github.com/opencv/opencv/pull/29433
**Performance Numbers on Intel(R) Core(TM) i9-11900K:** https://docs.google.com/spreadsheets/d/1RMvUZP1tSQtdjuNQY9JasYmlx1L5iN9XWGhXtG5u1uI/edit?usp=sharing
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
Replace "static inline" by "inline" in headers #29453
This fixes#29436
I also replaced CV_INLINE which can be removed in a later PR.
I can make a similar PR for 4.x if you want.
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
Refactoring and optimizing cvtcolor in Imgproc #29389
**Key Changes:**
1. Unified three divergent SIMD swap implementations into a single shared `v_swap` helper.
2. Consolidated the repeated `CV_8U/CV_16U/CV_32F` depth-dispatch chains into a single `CvtColorLoopDepth` template.
3. Extracted the duplicated coefficient-selection loops in the YCrCb/YUV constructors into `selectYuvCoeffs`.
4. Collapsed the shared YUV420 store logic duplicated across both decode invokers into `storeYUV420block`.
5. Generalized the inline blue-channel coefficient swaps in `color_lab.cpp` into `swapBlueCoeffsCols`/`swapBlueCoeffsRows`.
6. Added a `v_dotprod`-based SIMD path (`v_RGB2Y`/`v_RGB2UV`) for the previously scalar `RGB8toYUV422Invoker`.
7. Migrated all inline `CV_IPP_CHECK` cvtColor paths out of the dispatch files and into the IPP HAL plugin (`hal/ipp/src/color_ipp.cpp`), replacing each call site with `CALL_HAL`.
**Performance Numbers on Intel(R) Core(TM) i9-11900K**: https://docs.google.com/spreadsheets/d/1pz0aHlTeG4Ao8RT7LipgdNSjhCJz92QfZG3GmNa63hU/edit?usp=sharing
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
imgproc: optimized LUT and equalizeHist with SIMD #29250
- optimized lut with SIMD
- support equalizeHist with v_lut
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
core(rvv): fix v_matmul/v_matmuladd scalable semantics and expand lane-group test coverage #29080
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
<!-- Note!!! If you are an automated agent, we have a special process for you: add 🤖🤖🤖 to the end of the PR title. -->
## Platform
SpacemiT X60 (K1), 8-core RISC-V RVV 1.0, VLEN=256, 16GB RAM, OS: Bianbu Linux (kernel 6.6.63), GCC 13.2.0, Build: OpenCV 4.14.0-pre, Release, HAL: YES (RVV HAL 0.0.1)
## Motivation
OpenCV's Universal Intrinsics `v_matmul` and `v_matmuladd` have a semantic bug in the RVV scalable backend (`modules/core/include/opencv2/core/hal/intrin_rvv_scalable.hpp`).
The current implementation uses `v_extract_n(v, 0/1/2/3)` with hardcoded indices, assuming the vector holds exactly 4 float lanes (128-bit fixed). On hardware with VLEN=256 (e.g. SpacemiT K1 / BPI-F3), `v_float32` with LMUL=2 holds 16 lanes. As a result, lanes 4–15 silently reuse the inputs from lanes 0–3, producing wrong results.
OpenCV itself acknowledges this in `modules/core/src/matmul.simd.hpp`:
// v_matmuladd for RVV is 128-bit only but not scalable,
// this will fail the test Core_Transform.accuracy
The RVV scalable `transform_32f` path has been disabled because of this bug. However, the existing `TheTest<R>::test_matmul()` only checked the first 4-lane group (the outer loop was effectively hardcoded to `int i = 0`), so the bug was never caught by CI even on wide-vector backends.
## Modification
**Test fix** (`modules/core/test/test_intrin_utils.hpp`): Expanded `test_matmul()` to iterate over all 4-lane groups:
// Before (only checked lane group i=0)
int i = 0;
for (int j = i; j < i + 4; ++j) { ... }
// After (checks all lane groups)
for (int i = 0; i < VTraits<R>::vlanes(); i += 4)
{
for (int j = i; j < i + 4; ++j) { ... }
}
**Kernel fix** (`modules/core/include/opencv2/core/hal/intrin_rvv_scalable.hpp`): Rewrote `v_matmul` and `v_matmuladd` to process all 4-lane groups correctly. Each group of 4 lanes now independently computes the full matrix multiply using its own `v[i], v[i+1], v[i+2], v[i+3]` inputs. The `transform_32f` RVV path in `matmul.simd.hpp` remains disabled as the autovectorized path shows better performance on current hardware.
## Experiment 1: Bug reproduced on SpacemiT K1 (VLEN=256)
./opencv_test_core --gtest_filter="*intrin*"
Result: `hal_intrin128.float32x4_BASELINE` FAILED with 24 failures, all from lane groups i=4, i=8, i=12 (lanes 4–15).
Representative failures from `v_matmul` (line 1526):
i=4 j=4: actual=158 expected=56
i=4 j=5: actual=166.39999 expected=59.200001
i=8 j=8: actual=314.39999 expected=68.800003
i=12 j=12: actual=512.40002 expected=81.599998
Representative failures from `v_matmuladd` (line 1540):
i=4 j=4: actual=147.5 expected=51.5
i=8 j=8: actual=284.70001 expected=60.700001
i=12 j=12: actual=453.89999 expected=69.900002
Lane group i=0 (j=0..3) passed correctly — confirming the bug only affects lanes beyond the first 4, exactly as expected from the hardcoded `v_extract_n(v, 0/1/2/3)` implementation.
## Experiment 2: Both tests pass after fixing the kernel
./opencv_test_core --gtest_filter='hal_intrin128.float32x4_BASELINE'
[ OK ] hal_intrin128.float32x4_BASELINE (1859 ms)
[ PASSED ] 1 test.
./opencv_test_core --gtest_filter='Core_Transform.accuracy'
[ OK ] Core_Transform.accuracy (819 ms)
[ PASSED ] 1 test.
## Experiment 3: RVV transform path remains disabled (performance regression)
After re-enabling the RVV scalable `transform_32f` path experimentally, benchmarks showed a significant regression vs the compiler-autovectorized scalar path (CV_32FC3):
Size RVV path Scalar path Ratio
640x480 7.83 ms 1.48 ms 5.3x slower
1280x720 23.96 ms 5.17 ms 4.6x slower
1920x1080 53.76 ms 10.12 ms 5.3x slower
The compiler-autovectorized path outperforms the hand-written RVV kernel for this workload, consistent with the original comment in `matmul.simd.hpp`. The `transform_32f` RVV path is therefore kept disabled in this PR. The kernel fix to `v_matmul`/`v_matmuladd` remains necessary for correctness on wide-vector hardware, and the expanded test ensures the bug cannot regress silently in future.
core: add YAML 1.2 support for FileStorage #28482Fixes: #26363
CI Update: https://github.com/opencv/ci-gha-workflow/pull/306
Summary:
- Bool true/false literals support
- Header-less YAMLs support for Python YAML module compatibility. Header-less files are parsed as YAMLs by default
- Added FORMAT_YAML_1_0 flag to FileStorage for fallback.
- New YAML1.2 header
Testing:
- Added Core_InputOutput.YAML_1_2_Compatibility test case in test_io.cpp.
- Added YAML interop test in Python
Better AutoBuffer API #28909
Mimic std::vector<> to help replacing std::vector<T> by cv::AutoBuffer<T> when possible
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
- Replace unsafe pointer arithmetic and direct buffer modification with std::string methods.
- Update documentation to clarify that the last digit is used as compression level and truncated from the actual filename.
- Add test cases for .gz and .gz0-9
doc: modernize Doxygen comments to support for v1.15.0 #28903
This PR addresses several documentation build failures encountered with modern Doxygen versions, particularly v1.15.0 (shipped with Ubuntu 26.04).
- flann module: Updated license headers from /**** to /*M****. This prevents Doxygen from misinterpreting the license text (specifically the unclosed backticks in ``AS IS'') as documentation blocks, which previously caused "Reached end of file" errors.
- core module: Fixed a typo in operations.hpp where a doubled backtick (``) caused parsing to fail.
- tutorials: Fixed a missing backtick in real_time_pose.markdown (around line 108) and corrected typos in the RobustMatcher class name.
- Links: Resolved explicit link request failures in calib3d.hpp by ensuring proper namespace resolution.
These fixes ensure that the documentation can be generated without errors on the latest toolchains.
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
- Emscripten 5.0.1+ changed version macros to uppercase and deprecated lowercase ones.
- Updated CMake to detect the version from both cases.
- Added macro aliases in intrin_wasm.hpp to avoid deprecated warnings and ensure compatibility.
Fixes#28396 : out-of-bounds read in SIMD type conversion #28397Fixes#28396Fixes#27080
The vx_load_expand function in WASM intrinsics was using
wasm_v128_load which always loads a full 128-bit register
(16 bytes), even when the function only needed 8 elements.
For example, when converting uint8 to float32:
- vx_load_expand needs 8 uint8 elements
- But wasm_v128_load reads 16 bytes from memory
- This causes an 8-byte out-of-bounds read
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
Block layout-based convolution in DNN #28585
merge together with https://github.com/opencv/opencv_extra/pull/1321
Some core parts of the new engine in DNN module have been revised substantially:
1. all tests seem to pass, except for `Test_Graph_Simplifier.ResizeSubgraph`, which has been disabled because it does not take the newly added `TransformLayoutLayer` into account. The test should be reworked perhaps.
1. convolution and related operations (maxpool/avgpool) now use so-called block layout (`DATA_LAYOUT_BLOCK`), where `NxCxHxW` tensors are represented as `NxC1xHxWxC0`, where `C1=(C + C0-1)/C0` and `C0` is a power-of-two (usually 4, 8, 16 or 32).
1. graph is now pre-processed and `TransformLayoutLayer` is inserted to convert data from NCHW or NHWC layout to the block layout or vice versa. The transformations are done in a lazy way only when they are really needed. For example, in the whole Resnet only 2 transformations are performed.
1. transformer-based models and other models that do not use convolutions will run as usual, without going to block layout.
1. there is yet another graph preprocessing stage added that embeds constant weights/scale and bias into convolution and batch norm layers.
1. 'batchnorm', 'activation' and 'adding a residual' are now fused with convolution, just like in the old engine. That brings some noticeable acceleration.
1. optimized convolution kernels have been added.
* depthwise convolution, as well as maxpool and avgpool support C0=4, 8, 16 etc. _as long as_ C0 is divisible by the number of fp32 lanes in a SIMD register of the target platform (e.g. on ARM with NEON there must be `C0 % 4 == 0`, on x64 with AVX2 `C0 % 8 == 0`).
* non-depthwise convolution only supports C0=8 for now. C0=8 seems to be a sweetspot for ARM with NEON, x64 with AVX2 or RISC-V with RVV (with 128- or 256-bit registers). For some platforms with dedicated matrix accelerators C0=16 or even C0=32 might be more efficient, but we could add the respective kernels later.
* only fp32 kernels have been added. fp16/bf16 kernels might be added a little later.
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
- [x] Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
Fix absdiff with int arguments #28267
I believe the fix to the undefined behavior described in #27080 is simply casting to unsigned before subtraction, because
1. casting int to unsigned is well-defined; negative values get represented modulo $2^{32}$
2. overflow in unsigned subtraction is well-defined and the results are modulo $2^{32}$
Since we are computing everything modulo $2^{32}$ and the result must always be a non-negative number below $2^{32}$, this computation should be well-defined and correct.
I have verified this on ARM Apple Clang and on x64 Linux gcc with `-O3` and both produce the correct values in the reproducer of #27080. The test I have added fails with the integer overflow on both platforms I have tested. Perhaps @fengyuentau could verify if this also fixes the issue on the platforms he has tested?
I have added a fix and removed the workarounds. I recommended this approach in #28229, but he decided to revert it.
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
core: add NEON support for cvFloor in fast_math.hpp #28243
- This PR adds NEON intrinsics-based implementation for the cvFloor function in fast_math.hpp for Windows-ARM64.
- Both float and double overloads now use NEON intrinsics for cvFloor Function.
- calchist and calchist1d function uses cvFloor function for its computations.
- After adding these changes both functions showed improvement in performance.
**Performance Benchmarks:**
<img width="956" height="273" alt="image" src="https://github.com/user-attachments/assets/a00c98cd-d245-4d11-a9fd-361a3bd89f59" />
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
velonica0
Varun Jaiswal
Vadim Pisarevsky
Abhishek Gola
Alexander Smorkalov
Prasad Ayush Kumar
Vincent Rabaud
Madan mohan Manokar
Kumataro
Peter Rekdal Khan-Sunde
Alexander Smorkalov
kevinylin88
Om Navin Rope
satyam yadav
Maksim Shabunin
Pierre Chatelier
Anshu
Murat Raimbekov
Adrian Kretz
Denis Barucic
pratham-mcw
Mark Harfouche