coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2025-12-07 01:06:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

[RVV] Fix TestARGBInterpolate test fail

Browse Source 
Root cause:
Because InterpolateRow_RVV doesn't setup rounding mode to round-to-nearest-up when y1_fraction == 128.
The rounding mode register is set to round-down in ARGBAttenuateRow_RVV.
It cause InterpolateRow_RVV(y1_fraction == 128) runs on round-down mode.
Running on round-down mode make output result differs from round-to-nearest-up mode.

Solved by: ensure to use correct rounding mode in InterpolateRow_RVV.

Also, removing unnecessary rounding mode setup in ARGBAttenuateRow_RVV.

Bug: libyuv:956
Change-Id: Ib5265d42bad76b036e42b8f91ee42a9afe1f768d
Signed-off-by: Bruce Lai <bruce.lai@sifive.com>
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/4624492
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
This commit is contained in:
Bruce Lai 2023-06-18 04:54:00 -07:00 committed by libyuv LUCI CQ
parent 4472b5b849
commit 873d0db989
1 changed files with 5 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
source/row_rvv.cc
View File

@ -555,13 +555,16 @@ void InterpolateRow_RVV(uint8_t* dst_ptr,
} while (dst_w > 0); } while (dst_w > 0);
return; return;
} }
// To match behavior on other platforms, vxrm (fixed-point rounding mode
// register) is set to round-to-nearest-up(0).
asm volatile("csrwi vxrm, 0");
// Blend 50 / 50. // Blend 50 / 50.
if (y1_fraction == 128) { if (y1_fraction == 128) {
do { do {
size_t vl = __riscv_vsetvl_e8m8(dst_w); size_t vl = __riscv_vsetvl_e8m8(dst_w);
vuint8m8_t row0 = __riscv_vle8_v_u8m8(src_ptr, vl); vuint8m8_t row0 = __riscv_vle8_v_u8m8(src_ptr, vl);
vuint8m8_t row1 = __riscv_vle8_v_u8m8(src_ptr1, vl); vuint8m8_t row1 = __riscv_vle8_v_u8m8(src_ptr1, vl);
// Averaging add // Use round-to-nearest-up mode for averaging add
vuint8m8_t row_out = __riscv_vaaddu_vv_u8m8(row0, row1, vl); vuint8m8_t row_out = __riscv_vaaddu_vv_u8m8(row0, row1, vl);
__riscv_vse8_v_u8m8(dst_ptr, row_out, vl); __riscv_vse8_v_u8m8(dst_ptr, row_out, vl);
dst_w -= vl; dst_w -= vl;
@ -572,15 +575,13 @@ void InterpolateRow_RVV(uint8_t* dst_ptr,
return; return;
} }
// General purpose row blend. // General purpose row blend.
// To match behavior on other platforms, vxrm (fixed-point rounding mode
// register) is set to round-to-nearest-up(0).
asm volatile("csrwi vxrm, 0");
do { do {
size_t vl = __riscv_vsetvl_e8m4(dst_w); size_t vl = __riscv_vsetvl_e8m4(dst_w);
vuint8m4_t row0 = __riscv_vle8_v_u8m4(src_ptr, vl); vuint8m4_t row0 = __riscv_vle8_v_u8m4(src_ptr, vl);
vuint16m8_t acc = __riscv_vwmulu_vx_u16m8(row0, y0_fraction, vl); vuint16m8_t acc = __riscv_vwmulu_vx_u16m8(row0, y0_fraction, vl);
vuint8m4_t row1 = __riscv_vle8_v_u8m4(src_ptr1, vl); vuint8m4_t row1 = __riscv_vle8_v_u8m4(src_ptr1, vl);
acc = __riscv_vwmaccu_vx_u16m8(acc, y1_fraction, row1, vl); acc = __riscv_vwmaccu_vx_u16m8(acc, y1_fraction, row1, vl);
// Use round-to-nearest-up mode for vnclip
__riscv_vse8_v_u8m4(dst_ptr, __riscv_vnclipu_wx_u8m4(acc, 8, vl), vl); __riscv_vse8_v_u8m4(dst_ptr, __riscv_vnclipu_wx_u8m4(acc, 8, vl), vl);
dst_w -= vl; dst_w -= vl;
src_ptr += vl; src_ptr += vl;
@ -929,9 +930,6 @@ void ARGBAttenuateRow_RVV(const uint8_t* src_argb,
uint8_t* dst_argb, uint8_t* dst_argb,
int width) { int width) {
size_t w = (size_t)width; size_t w = (size_t)width;
// To match behavior on other platforms, vxrm (fixed-point rounding mode
// register) is set to round-down(2).
asm volatile("csrwi vxrm, 2");
do { do {
vuint8m2_t v_b, v_g, v_r, v_a; vuint8m2_t v_b, v_g, v_r, v_a;
vuint16m4_t v_ba_16, v_ga_16, v_ra_16; vuint16m4_t v_ba_16, v_ga_16, v_ra_16;