potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-21 19:33:03 +04:00
Code Packages Releases Wiki Activity

feat(rvv-hal): optimize FAST pre-screen with deferred u8mf2 widening

Browse Source 
Defer the u8→i16 widening (vzext) in the pre-screen phase until after
the quick-reject check passes. This avoids 5 unnecessary widen
operations per pixel-strip when the pre-screen rejects (which happens
~70-80% of the time at the default threshold=20).

The approach uses u8mf2 (same VL as i16m1) for the pre-screen
comparison with vssubu/vsaddu for bounds and vmsgtu for unsigned
comparison. When the pre-screen passes, the already-loaded u8mf2
variables are widened to i16m1 for the score computation, reusing
the same data without redundant memory loads.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Shengyu Wu
2026-05-31 20:49:03 +08:00
parent 20742dc581
commit aee03ffbb2
1 changed files with 24 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hal/riscv-rvv/src/features2d/fast.cpp
+24 -21
View File
@@ -56,32 +56,35 @@ inline int fast_16(const uchar* src_data, size_t src_step,
size_t vl;
for (; j < width - 3; j += vl, ptr += vl)
{
vl = __riscv_vsetvl_e16m1(width - 3 - j);
/* u8mf2 pre-screen: same VL as i16m1, defer vzext until needed */
vl = __riscv_vsetvl_e8mf2(width - 3 - j);
vuint8mf2_t vcen_8 = __riscv_vle8_v_u8mf2(ptr, vl);
vuint8mf2_t vlo_8 = __riscv_vssubu_vx_u8mf2(vcen_8, (uint8_t)threshold, vl);
vuint8mf2_t vhi_8 = __riscv_vsaddu_vx_u8mf2(vcen_8, (uint8_t)threshold, vl);
vuint8mf2_t vk0_8 = __riscv_vle8_v_u8mf2(ptr + pixel[0], vl);
vuint8mf2_t vk4_8 = __riscv_vle8_v_u8mf2(ptr + pixel[4], vl);
vuint8mf2_t vk8_8 = __riscv_vle8_v_u8mf2(ptr + pixel[8], vl);
vuint8mf2_t vk12_8 = __riscv_vle8_v_u8mf2(ptr + pixel[12], vl);
/* Load center pixel and widen to int16 */
vint16m1_t vcen = __riscv_vreinterpret_v_u16m1_i16m1(
__riscv_vzext_vf2(__riscv_vle8_v_u8mf2(ptr, vl), vl));
vint16m1_t vlo = __riscv_vsub_vx_i16m1(vcen, threshold, vl);
vint16m1_t vhi = __riscv_vadd_vx_i16m1(vcen, threshold, vl);
/* 4-direction quick reject */
vint16m1_t vk0 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(__riscv_vle8_v_u8mf2(ptr + pixel[0], vl), vl));
vint16m1_t vk4 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(__riscv_vle8_v_u8mf2(ptr + pixel[4], vl), vl));
vint16m1_t vk8 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(__riscv_vle8_v_u8mf2(ptr + pixel[8], vl), vl));
vint16m1_t vk12 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(__riscv_vle8_v_u8mf2(ptr + pixel[12], vl), vl));
vbool16_t bright = __riscv_vmand_mm_b16(__riscv_vmsgt_vv_i16m1_b16(vk0, vhi, vl), __riscv_vmsgt_vv_i16m1_b16(vk4, vhi, vl), vl);
vbool16_t dark = __riscv_vmand_mm_b16(__riscv_vmsgt_vv_i16m1_b16(vlo, vk0, vl), __riscv_vmsgt_vv_i16m1_b16(vlo, vk4, vl), vl);
bright = __riscv_vmor_mm_b16(bright, __riscv_vmand_mm_b16(__riscv_vmsgt_vv_i16m1_b16(vk4, vhi, vl), __riscv_vmsgt_vv_i16m1_b16(vk8, vhi, vl), vl), vl);
dark = __riscv_vmor_mm_b16(dark, __riscv_vmand_mm_b16(__riscv_vmsgt_vv_i16m1_b16(vlo, vk4, vl), __riscv_vmsgt_vv_i16m1_b16(vlo, vk8, vl), vl), vl);
bright = __riscv_vmor_mm_b16(bright, __riscv_vmand_mm_b16(__riscv_vmsgt_vv_i16m1_b16(vk8, vhi, vl), __riscv_vmsgt_vv_i16m1_b16(vk12, vhi, vl), vl), vl);
dark = __riscv_vmor_mm_b16(dark, __riscv_vmand_mm_b16(__riscv_vmsgt_vv_i16m1_b16(vlo, vk8, vl), __riscv_vmsgt_vv_i16m1_b16(vlo, vk12, vl), vl), vl);
bright = __riscv_vmor_mm_b16(bright, __riscv_vmand_mm_b16(__riscv_vmsgt_vv_i16m1_b16(vk12, vhi, vl), __riscv_vmsgt_vv_i16m1_b16(vk0, vhi, vl), vl), vl);
dark = __riscv_vmor_mm_b16(dark, __riscv_vmand_mm_b16(__riscv_vmsgt_vv_i16m1_b16(vlo, vk12, vl), __riscv_vmsgt_vv_i16m1_b16(vlo, vk0, vl), vl), vl);
vbool16_t bright = __riscv_vmand_mm_b16(__riscv_vmsgtu_vv_u8mf2_b16(vk0_8, vhi_8, vl), __riscv_vmsgtu_vv_u8mf2_b16(vk4_8, vhi_8, vl), vl);
vbool16_t dark = __riscv_vmand_mm_b16(__riscv_vmsgtu_vv_u8mf2_b16(vlo_8, vk0_8, vl), __riscv_vmsgtu_vv_u8mf2_b16(vlo_8, vk4_8, vl), vl);
bright = __riscv_vmor_mm_b16(bright, __riscv_vmand_mm_b16(__riscv_vmsgtu_vv_u8mf2_b16(vk4_8, vhi_8, vl), __riscv_vmsgtu_vv_u8mf2_b16(vk8_8, vhi_8, vl), vl), vl);
dark = __riscv_vmor_mm_b16(dark, __riscv_vmand_mm_b16(__riscv_vmsgtu_vv_u8mf2_b16(vlo_8, vk4_8, vl), __riscv_vmsgtu_vv_u8mf2_b16(vlo_8, vk8_8, vl), vl), vl);
bright = __riscv_vmor_mm_b16(bright, __riscv_vmand_mm_b16(__riscv_vmsgtu_vv_u8mf2_b16(vk8_8, vhi_8, vl), __riscv_vmsgtu_vv_u8mf2_b16(vk12_8, vhi_8, vl), vl), vl);
dark = __riscv_vmor_mm_b16(dark, __riscv_vmand_mm_b16(__riscv_vmsgtu_vv_u8mf2_b16(vlo_8, vk8_8, vl), __riscv_vmsgtu_vv_u8mf2_b16(vlo_8, vk12_8, vl), vl), vl);
bright = __riscv_vmor_mm_b16(bright, __riscv_vmand_mm_b16(__riscv_vmsgtu_vv_u8mf2_b16(vk12_8, vhi_8, vl), __riscv_vmsgtu_vv_u8mf2_b16(vk0_8, vhi_8, vl), vl), vl);
dark = __riscv_vmor_mm_b16(dark, __riscv_vmand_mm_b16(__riscv_vmsgtu_vv_u8mf2_b16(vlo_8, vk12_8, vl), __riscv_vmsgtu_vv_u8mf2_b16(vlo_8, vk0_8, vl), vl), vl);
if (__riscv_vfirst_m_b16(__riscv_vmor_mm_b16(bright, dark, vl), vl) < 0)
continue;
/* Widen pre-loaded u8mf2 to i16m1 for score computation */
vint16m1_t vcen = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(vcen_8, vl));
vint16m1_t vk0 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(vk0_8, vl));
vint16m1_t vk4 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(vk4_8, vl));
vint16m1_t vk8 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(vk8_8, vl));
vint16m1_t vk12 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(vk12_8, vl));
/* Load remaining 12 neighbors */
vint16m1_t vk1 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(__riscv_vle8_v_u8mf2(ptr + pixel[1], vl), vl));
vint16m1_t vk2 = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2(__riscv_vle8_v_u8mf2(ptr + pixel[2], vl), vl));