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] Enable ScaleRowDown34_RVV & ScaleRowDown34_{0,1}_Box_RVV

Browse Source 
Run on SiFive internal FPGA:

Test case                       RVV function			Speedup
I444ScaleDownBy3by4_None	ScaleRowDown34_RVV	        5.8
I444ScaleDownBy3by4_Linear	ScaleRowDown34_0/1_Box_RVV	6.5
I444ScaleDownBy3by4_Bilinear	ScaleRowDown34_0/1_Box_RVV	6.3

Bug: libyuv:956
Change-Id: I8ef221ab14d631e14f1ba1aaa25d2b30d4e710db
Signed-off-by: Darren Hsieh <darren.hsieh@sifive.com>
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/4607777
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
This commit is contained in:
Darren Hsieh 2023-06-07 01:17:43 -07:00 committed by libyuv LUCI CQ
parent 2a5d7e2fbc
commit 552571e8b2
3 changed files with 160 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
include/libyuv/scale_row.h
View File

@ -180,6 +180,7 @@ extern "C" {
#define HAS_SCALEARGBROWDOWN2_RVV #define HAS_SCALEARGBROWDOWN2_RVV
#define HAS_SCALEARGBROWDOWNEVEN_RVV #define HAS_SCALEARGBROWDOWNEVEN_RVV
#define HAS_SCALEROWDOWN2_RVV #define HAS_SCALEROWDOWN2_RVV
#define HAS_SCALEROWDOWN34_RVV
#define HAS_SCALEROWDOWN4_RVV #define HAS_SCALEROWDOWN4_RVV
#define HAS_SCALEUVROWDOWN2_RVV #define HAS_SCALEUVROWDOWN2_RVV
#define HAS_SCALEUVROWDOWN2LINEAR_RVV #define HAS_SCALEUVROWDOWN2LINEAR_RVV
@ -1822,6 +1823,18 @@ void ScaleRowDown4Box_RVV(const uint8_t* src_ptr,
ptrdiff_t src_stride, ptrdiff_t src_stride,
uint8_t* dst_ptr, uint8_t* dst_ptr,
int dst_width); int dst_width);
void ScaleRowDown34_RVV(const uint8_t* src_ptr,
ptrdiff_t src_stride,
uint8_t* dst_ptr,
int dst_width);
void ScaleRowDown34_0_Box_RVV(const uint8_t* src_ptr,
ptrdiff_t src_stride,
uint8_t* dst_ptr,
int dst_width);
void ScaleRowDown34_1_Box_RVV(const uint8_t* src_ptr,
ptrdiff_t src_stride,
uint8_t* dst_ptr,
int dst_width);
#ifdef __cplusplus #ifdef __cplusplus
} // extern "C" } // extern "C"

11
source/scale.cc
View File

@ -485,6 +485,17 @@ static void ScalePlaneDown34(int src_width,
} }
} }
#endif #endif
#if defined(HAS_SCALEROWDOWN34_RVV)
if (TestCpuFlag(kCpuHasRVV)) {
if (!filtering) {
ScaleRowDown34_0 = ScaleRowDown34_RVV;
ScaleRowDown34_1 = ScaleRowDown34_RVV;
} else {
ScaleRowDown34_0 = ScaleRowDown34_0_Box_RVV;
ScaleRowDown34_1 = ScaleRowDown34_1_Box_RVV;
}
}
#endif
for (y = 0; y < dst_height - 2; y += 3) { for (y = 0; y < dst_height - 2; y += 3) {
ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width); ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);

136
source/scale_rvv.cc
View File

@ -330,6 +330,142 @@ void ScaleRowDown4Box_RVV(const uint8_t* src_ptr,
} while (w > 0); } while (w > 0);
} }
void ScaleRowDown34_RVV(const uint8_t* src_ptr,
ptrdiff_t src_stride,
uint8_t* dst_ptr,
int dst_width) {
size_t w = (size_t)dst_width / 3u;
do {
size_t vl = __riscv_vsetvl_e8m2(w);
vuint8m2_t v_s0, v_s1, v_s2, v_s3;
__riscv_vlseg4e8_v_u8m2(&v_s0, &v_s1, &v_s2, &v_s3, src_ptr, vl);
__riscv_vsseg3e8_v_u8m2(dst_ptr, v_s0, v_s1, v_s3, vl);
w -= vl;
src_ptr += 4 * vl;
dst_ptr += 3 * vl;
} while (w > 0);
}
void ScaleRowDown34_0_Box_RVV(const uint8_t* src_ptr,
ptrdiff_t src_stride,
uint8_t* dst_ptr,
int dst_width) {
size_t w = (size_t)dst_width / 3u;
const uint8_t* s = src_ptr;
const uint8_t* t = src_ptr + src_stride;
// NOTE: To match behavior on other platforms, vxrm (fixed-point rounding mode
// register) is set to round-to-nearest-up mode(0).
asm volatile("csrwi vxrm, 0");
do {
vuint8m2_t v_s0, v_s1, v_s2, v_s3;
vuint16m4_t v_t0_u16, v_t1_u16, v_t2_u16, v_t3_u16;
vuint8m2_t v_u0, v_u1, v_u2, v_u3;
vuint16m4_t v_u1_u16;
vuint8m2_t v_a0, v_a1, v_a2;
size_t vl = __riscv_vsetvl_e8m2(w);
__riscv_vlseg4e8_v_u8m2(&v_s0, &v_s1, &v_s2, &v_s3, s, vl);
if (src_stride == 0) {
v_t0_u16 = __riscv_vwaddu_vx_u16m4(v_s0, 2, vl);
v_t1_u16 = __riscv_vwaddu_vx_u16m4(v_s1, 2, vl);
v_t2_u16 = __riscv_vwaddu_vx_u16m4(v_s2, 2, vl);
v_t3_u16 = __riscv_vwaddu_vx_u16m4(v_s3, 2, vl);
} else {
vuint8m2_t v_t0, v_t1, v_t2, v_t3;
__riscv_vlseg4e8_v_u8m2(&v_t0, &v_t1, &v_t2, &v_t3, t, vl);
v_t0_u16 = __riscv_vwaddu_vx_u16m4(v_t0, 0, vl);
v_t1_u16 = __riscv_vwaddu_vx_u16m4(v_t1, 0, vl);
v_t2_u16 = __riscv_vwaddu_vx_u16m4(v_t2, 0, vl);
v_t3_u16 = __riscv_vwaddu_vx_u16m4(v_t3, 0, vl);
t += 4 * vl;
}
v_t0_u16 = __riscv_vwmaccu_vx_u16m4(v_t0_u16, 3, v_s0, vl);
v_t1_u16 = __riscv_vwmaccu_vx_u16m4(v_t1_u16, 3, v_s1, vl);
v_t2_u16 = __riscv_vwmaccu_vx_u16m4(v_t2_u16, 3, v_s2, vl);
v_t3_u16 = __riscv_vwmaccu_vx_u16m4(v_t3_u16, 3, v_s3, vl);
// Use round-to-nearest-up mode for vnclip & averaging add
v_u0 = __riscv_vnclipu_wx_u8m2(v_t0_u16, 2, vl);
v_u1 = __riscv_vnclipu_wx_u8m2(v_t1_u16, 2, vl);
v_u2 = __riscv_vnclipu_wx_u8m2(v_t2_u16, 2, vl);
v_u3 = __riscv_vnclipu_wx_u8m2(v_t3_u16, 2, vl);
// a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
v_u1_u16 = __riscv_vwaddu_vx_u16m4(v_u1, 0, vl);
v_u1_u16 = __riscv_vwmaccu_vx_u16m4(v_u1_u16, 3, v_u0, vl);
v_a0 = __riscv_vnclipu_wx_u8m2(v_u1_u16, 2, vl);
// a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
v_a1 = __riscv_vaaddu_vv_u8m2(v_u1, v_u2, vl);
// a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
v_u1_u16 = __riscv_vwaddu_vx_u16m4(v_u2, 0, vl);
v_u1_u16 = __riscv_vwmaccu_vx_u16m4(v_u1_u16, 3, v_u3, vl);
v_a2 = __riscv_vnclipu_wx_u8m2(v_u1_u16, 2, vl);
__riscv_vsseg3e8_v_u8m2(dst_ptr, v_a0, v_a1, v_a2, vl);
w -= vl;
s += 4 * vl;
dst_ptr += 3 * vl;
} while (w > 0);
}
void ScaleRowDown34_1_Box_RVV(const uint8_t* src_ptr,
ptrdiff_t src_stride,
uint8_t* dst_ptr,
int dst_width) {
size_t w = (size_t)dst_width / 3u;
const uint8_t* s = src_ptr;
const uint8_t* t = src_ptr + src_stride;
// NOTE: To match behavior on other platforms, vxrm (fixed-point rounding mode
// register) is set to round-to-nearest-up mode(0).
asm volatile("csrwi vxrm, 0");
do {
vuint8m2_t v_s0, v_s1, v_s2, v_s3;
vuint8m2_t v_ave0, v_ave1, v_ave2, v_ave3;
vuint16m4_t v_u1_u16;
vuint8m2_t v_a0, v_a1, v_a2;
size_t vl = __riscv_vsetvl_e8m2(w);
__riscv_vlseg4e8_v_u8m2(&v_s0, &v_s1, &v_s2, &v_s3, s, vl);
// Use round-to-nearest-up mode for vnclip & averaging add
if (src_stride == 0) {
v_ave0 = __riscv_vaaddu_vv_u8m2(v_s0, v_s0, vl);
v_ave1 = __riscv_vaaddu_vv_u8m2(v_s1, v_s1, vl);
v_ave2 = __riscv_vaaddu_vv_u8m2(v_s2, v_s2, vl);
v_ave3 = __riscv_vaaddu_vv_u8m2(v_s3, v_s3, vl);
} else {
vuint8m2_t v_t0, v_t1, v_t2, v_t3;
__riscv_vlseg4e8_v_u8m2(&v_t0, &v_t1, &v_t2, &v_t3, t, vl);
v_ave0 = __riscv_vaaddu_vv_u8m2(v_s0, v_t0, vl);
v_ave1 = __riscv_vaaddu_vv_u8m2(v_s1, v_t1, vl);
v_ave2 = __riscv_vaaddu_vv_u8m2(v_s2, v_t2, vl);
v_ave3 = __riscv_vaaddu_vv_u8m2(v_s3, v_t3, vl);
t += 4 * vl;
}
// a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
v_u1_u16 = __riscv_vwaddu_vx_u16m4(v_ave1, 0, vl);
v_u1_u16 = __riscv_vwmaccu_vx_u16m4(v_u1_u16, 3, v_ave0, vl);
v_a0 = __riscv_vnclipu_wx_u8m2(v_u1_u16, 2, vl);
// a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
v_a1 = __riscv_vaaddu_vv_u8m2(v_ave1, v_ave2, vl);
// a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
v_u1_u16 = __riscv_vwaddu_vx_u16m4(v_ave2, 0, vl);
v_u1_u16 = __riscv_vwmaccu_vx_u16m4(v_u1_u16, 3, v_ave3, vl);
v_a2 = __riscv_vnclipu_wx_u8m2(v_u1_u16, 2, vl);
__riscv_vsseg3e8_v_u8m2(dst_ptr, v_a0, v_a1, v_a2, vl);
w -= vl;
s += 4 * vl;
dst_ptr += 3 * vl;
} while (w > 0);
}
void ScaleUVRowDown2_RVV(const uint8_t* src_uv, void ScaleUVRowDown2_RVV(const uint8_t* src_uv,
ptrdiff_t src_stride, ptrdiff_t src_stride,
uint8_t* dst_uv, uint8_t* dst_uv,