aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorFrank Barchard <fbarchard@google.com>2022-03-17 15:50:29 -0700
committerlibyuv LUCI CQ <libyuv-scoped@luci-project-accounts.iam.gserviceaccount.com>2022-03-18 07:22:36 +0000
commit95b14b24462d67aede96e30243694732f9471e63 (patch)
tree0cbcd60504d41cf6d588ca59d59fe6ee31d7901d
parent3aebf69d668177e7ee6dbbe0025e5c3dbb525ff2 (diff)
downloadlibyuv-95b14b24462d67aede96e30243694732f9471e63.tar.gz
RAWToJ400 faster version for ARM
- Unrolled to 16 pixels - Take constants via structure, allowing different colorspace and channel order - Use ADDHN to add 16.5 and take upper 8 bits of 16 bit values, narrowing to 8 bits - clang-format applied, affecting mips code On Cortex A510 Was RAWToJ400_Opt (1623 ms) Now RAWToJ400_Opt (862 ms) C RAWToJ400_Opt (1627 ms) Bug: b/220171611 Change-Id: I06a9baf9650ebe2802fb6ff6dfbd524e2c06ada0 Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/3534023 Reviewed-by: Wan-Teh Chang <wtc@google.com> Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Diffstat
-rw-r--r--README.chromium4
-rw-r--r--include/libyuv/convert_argb.h4
-rw-r--r--include/libyuv/loongson_intrinsics.h62
-rw-r--r--include/libyuv/macros_msa.h54
-rw-r--r--include/libyuv/row.h12
-rw-r--r--include/libyuv/version.h2
-rw-r--r--source/convert.cc80
-rw-r--r--source/convert_from_argb.cc32
-rw-r--r--source/planar_functions.cc2
-rw-r--r--source/row_any.cc20
-rw-r--r--source/row_lasx.cc187
-rw-r--r--source/row_neon.cc349
-rw-r--r--source/row_neon64.cc355
-rw-r--r--unit_test/convert_test.cc22
14 files changed, 533 insertions, 652 deletions
diff --git a/README.chromium b/README.chromium
index 940a0142..5bcfae87 100644
--- a/README.chromium
+++ b/README.chromium
@@ -1,8 +1,8 @@
Name: libyuv
URL: http://code.google.com/p/libyuv/
-Version: 1813
+Version: 1814
License: BSD
License File: LICENSE
Description:
-libyuv is an open source project that includes YUV conversion and scaling functionality. \ No newline at end of file
+libyuv is an open source project that includes YUV conversion and scaling functionality.
diff --git a/include/libyuv/convert_argb.h b/include/libyuv/convert_argb.h
index 9b0529cd..f66d20ce 100644
--- a/include/libyuv/convert_argb.h
+++ b/include/libyuv/convert_argb.h
@@ -14,7 +14,7 @@
#include "libyuv/basic_types.h"
#include "libyuv/rotate.h" // For enum RotationMode.
-#include "libyuv/scale.h" // For enum FilterMode.
+#include "libyuv/scale.h" // For enum FilterMode.
#ifdef __cplusplus
namespace libyuv {
@@ -2117,7 +2117,7 @@ int P210ToARGBMatrixFilter(const uint16_t* src_y,
const struct YuvConstants* yuvconstants,
int width,
int height,
- enum FilterMode filter) ;
+ enum FilterMode filter);
// Convert P010 to AR30 with matrix and UV filter mode.
LIBYUV_API
diff --git a/include/libyuv/loongson_intrinsics.h b/include/libyuv/loongson_intrinsics.h
index 79b5d0e4..1d613def 100644
--- a/include/libyuv/loongson_intrinsics.h
+++ b/include/libyuv/loongson_intrinsics.h
@@ -91,7 +91,8 @@
* out : 23,40,41,26, 23,40,41,26
* =============================================================================
*/
-static inline __m128i __lsx_vdp2add_h_b(__m128i in_c, __m128i in_h,
+static inline __m128i __lsx_vdp2add_h_b(__m128i in_c,
+ __m128i in_h,
__m128i in_l) {
__m128i out;
@@ -117,7 +118,8 @@ static inline __m128i __lsx_vdp2add_h_b(__m128i in_c, __m128i in_h,
* out : 23,40,41,26, 23,40,41,26
* =============================================================================
*/
-static inline __m128i __lsx_vdp2add_h_bu(__m128i in_c, __m128i in_h,
+static inline __m128i __lsx_vdp2add_h_bu(__m128i in_c,
+ __m128i in_h,
__m128i in_l) {
__m128i out;
@@ -143,7 +145,8 @@ static inline __m128i __lsx_vdp2add_h_bu(__m128i in_c, __m128i in_h,
* out : -4,-24,-60,-112, 6,26,62,114
* =============================================================================
*/
-static inline __m128i __lsx_vdp2add_h_bu_b(__m128i in_c, __m128i in_h,
+static inline __m128i __lsx_vdp2add_h_bu_b(__m128i in_c,
+ __m128i in_h,
__m128i in_l) {
__m128i out;
@@ -169,7 +172,8 @@ static inline __m128i __lsx_vdp2add_h_bu_b(__m128i in_c, __m128i in_h,
* out : 23,40,41,26
* =============================================================================
*/
-static inline __m128i __lsx_vdp2add_w_h(__m128i in_c, __m128i in_h,
+static inline __m128i __lsx_vdp2add_w_h(__m128i in_c,
+ __m128i in_h,
__m128i in_l) {
__m128i out;
@@ -414,8 +418,8 @@ static inline __m128i __lsx_vclip255_w(__m128i _in) {
_out0, _out1, _out2, _out3, _out4, _out5, _out6, \
_out7) \
{ \
- __m128i zero = { 0 }; \
- __m128i shuf8 = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 }; \
+ __m128i zero = {0}; \
+ __m128i shuf8 = {0x0F0E0D0C0B0A0908, 0x1716151413121110}; \
__m128i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7; \
\
_t0 = __lsx_vilvl_b(_in2, _in0); \
@@ -828,7 +832,8 @@ static inline __m256i __lasx_xvdp2_w_hu_h(__m256i in_h, __m256i in_l) {
* Example : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
* =============================================================================
*/
-static inline __m256i __lasx_xvdp2add_h_b(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvdp2add_h_b(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i out;
@@ -851,7 +856,8 @@ static inline __m256i __lasx_xvdp2add_h_b(__m256i in_c, __m256i in_h,
* Example : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
* =============================================================================
*/
-static inline __m256i __lasx_xvdp2add_h_bu(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvdp2add_h_bu(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i out;
@@ -874,7 +880,8 @@ static inline __m256i __lasx_xvdp2add_h_bu(__m256i in_c, __m256i in_h,
* Example : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
* =============================================================================
*/
-static inline __m256i __lasx_xvdp2add_h_bu_b(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvdp2add_h_bu_b(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i out;
@@ -901,7 +908,8 @@ static inline __m256i __lasx_xvdp2add_h_bu_b(__m256i in_c, __m256i in_h,
* out : 23,40,41,26, 23,40,41,26
* =============================================================================
*/
-static inline __m256i __lasx_xvdp2add_w_h(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvdp2add_w_h(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i out;
@@ -924,7 +932,8 @@ static inline __m256i __lasx_xvdp2add_w_h(__m256i in_c, __m256i in_h,
* Example : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
* =============================================================================
*/
-static inline __m256i __lasx_xvdp2add_w_hu(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvdp2add_w_hu(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i out;
@@ -947,7 +956,8 @@ static inline __m256i __lasx_xvdp2add_w_hu(__m256i in_c, __m256i in_h,
* Example : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
* =============================================================================
*/
-static inline __m256i __lasx_xvdp2add_w_hu_h(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvdp2add_w_hu_h(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i out;
@@ -971,7 +981,8 @@ static inline __m256i __lasx_xvdp2add_w_hu_h(__m256i in_c, __m256i in_h,
* Example : See out = __lasx_xvdp2sub_w_h(in_c, in_h, in_l)
* =============================================================================
*/
-static inline __m256i __lasx_xvdp2sub_h_bu(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvdp2sub_h_bu(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i out;
@@ -1000,7 +1011,8 @@ static inline __m256i __lasx_xvdp2sub_h_bu(__m256i in_c, __m256i in_h,
* out : -7,-3,0,0, 0,-1,0,-1
* =============================================================================
*/
-static inline __m256i __lasx_xvdp2sub_w_h(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvdp2sub_w_h(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i out;
@@ -1201,7 +1213,8 @@ static inline __m256i __lasx_xvaddw_w_w_h(__m256i in_h, __m256i in_l) {
* out : 201, 602,1203,2004, -995, -1794,-2793,-3992
* =============================================================================
*/
-static inline __m256i __lasx_xvmaddwl_w_h(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvmaddwl_w_h(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i tmp0, tmp1, out;
@@ -1225,7 +1238,8 @@ static inline __m256i __lasx_xvmaddwl_w_h(__m256i in_c, __m256i in_h,
* Example : See out = __lasx_xvmaddwl_w_h(in_c, in_h, in_l)
* =============================================================================
*/
-static inline __m256i __lasx_xvmaddwh_w_h(__m256i in_c, __m256i in_h,
+static inline __m256i __lasx_xvmaddwh_w_h(__m256i in_c,
+ __m256i in_h,
__m256i in_l) {
__m256i tmp0, tmp1, out;
@@ -1303,7 +1317,7 @@ static inline __m256i __lasx_xvmulwh_w_h(__m256i in_h, __m256i in_l) {
*/
static inline __m256i __lasx_xvsaddw_hu_hu_bu(__m256i in_h, __m256i in_l) {
__m256i tmp1, out;
- __m256i zero = { 0 };
+ __m256i zero = {0};
tmp1 = __lasx_xvilvl_b(zero, in_l);
out = __lasx_xvsadd_hu(in_h, tmp1);
@@ -1921,12 +1935,14 @@ static inline __m256i __lasx_xvsplati_h_h(__m256i in, int idx) {
* VP:1,2,3,4,
* =============================================================================
*/
-#define VECT_PRINT(RTYPE, element_num, in0, enter) \
- { \
- RTYPE _tmp0 = (RTYPE)in0; \
- int _i = 0; \
- if (enter) printf("\nVP:"); \
- for (_i = 0; _i < element_num; _i++) printf("%d,", _tmp0[_i]); \
+#define VECT_PRINT(RTYPE, element_num, in0, enter) \
+ { \
+ RTYPE _tmp0 = (RTYPE)in0; \
+ int _i = 0; \
+ if (enter) \
+ printf("\nVP:"); \
+ for (_i = 0; _i < element_num; _i++) \
+ printf("%d,", _tmp0[_i]); \
}
#endif /* LOONGSON_INTRINSICS_H */
diff --git a/include/libyuv/macros_msa.h b/include/libyuv/macros_msa.h
index 7a1dc4af..b9a44fcc 100644
--- a/include/libyuv/macros_msa.h
+++ b/include/libyuv/macros_msa.h
@@ -81,37 +81,35 @@
})
#endif // !(__mips == 64)
#else // !(__mips_isa_rev >= 6)
-#define LW(psrc) \
- ({ \
- uint8_t *psrc_lw_m = (uint8_t *) (psrc); \
- uint32_t val_lw_m; \
- \
- __asm__ volatile ( \
- "lwr %[val_lw_m], 0(%[psrc_lw_m]) \n\t" \
- "lwl %[val_lw_m], 3(%[psrc_lw_m]) \n\t" \
- \
- : [val_lw_m] "=&r"(val_lw_m) \
- : [psrc_lw_m] "r"(psrc_lw_m) \
- ); \
- \
- val_lw_m; \
+#define LW(psrc) \
+ ({ \
+ uint8_t* psrc_lw_m = (uint8_t*)(psrc); \
+ uint32_t val_lw_m; \
+ \
+ __asm__ volatile( \
+ "lwr %[val_lw_m], 0(%[psrc_lw_m]) \n\t" \
+ "lwl %[val_lw_m], 3(%[psrc_lw_m]) \n\t" \
+ \
+ : [val_lw_m] "=&r"(val_lw_m) \
+ : [psrc_lw_m] "r"(psrc_lw_m)); \
+ \
+ val_lw_m; \
})
#if (__mips == 64)
-#define LD(psrc) \
- ({ \
- uint8_t *psrc_ld_m = (uint8_t *) (psrc); \
- uint64_t val_ld_m = 0; \
- \
- __asm__ volatile ( \
- "ldr %[val_ld_m], 0(%[psrc_ld_m]) \n\t" \
- "ldl %[val_ld_m], 7(%[psrc_ld_m]) \n\t" \
- \
- : [val_ld_m] "=&r" (val_ld_m) \
- : [psrc_ld_m] "r" (psrc_ld_m) \
- ); \
- \
- val_ld_m; \
+#define LD(psrc) \
+ ({ \
+ uint8_t* psrc_ld_m = (uint8_t*)(psrc); \
+ uint64_t val_ld_m = 0; \
+ \
+ __asm__ volatile( \
+ "ldr %[val_ld_m], 0(%[psrc_ld_m]) \n\t" \
+ "ldl %[val_ld_m], 7(%[psrc_ld_m]) \n\t" \
+ \
+ : [val_ld_m] "=&r"(val_ld_m) \
+ : [psrc_ld_m] "r"(psrc_ld_m)); \
+ \
+ val_ld_m; \
})
#else // !(__mips == 64)
#define LD(psrc) \
diff --git a/include/libyuv/row.h b/include/libyuv/row.h
index d611b8b3..2be84338 100644
--- a/include/libyuv/row.h
+++ b/include/libyuv/row.h
@@ -1380,7 +1380,9 @@ void BGRAToYRow_LSX(const uint8_t* src_bgra, uint8_t* dst_y, int width);
void ABGRToYRow_LSX(const uint8_t* src_abgr, uint8_t* dst_y, int width);
void RGBAToYRow_LSX(const uint8_t* src_rgba, uint8_t* dst_y, int width);
void ARGB1555ToYRow_LSX(const uint8_t* src_argb1555, uint8_t* dst_y, int width);
-void ARGB1555ToYRow_LASX(const uint8_t* src_argb1555, uint8_t* dst_y, int width);
+void ARGB1555ToYRow_LASX(const uint8_t* src_argb1555,
+ uint8_t* dst_y,
+ int width);
void RGB565ToYRow_LSX(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
void RGB565ToYRow_LASX(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
void RGB24ToYRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
@@ -2748,7 +2750,9 @@ void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
int width);
void RGB24ToARGBRow_MSA(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
void RGB24ToARGBRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
-void RGB24ToARGBRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
+void RGB24ToARGBRow_LASX(const uint8_t* src_rgb24,
+ uint8_t* dst_argb,
+ int width);
void RAWToARGBRow_NEON(const uint8_t* src_raw, uint8_t* dst_argb, int width);
void RAWToRGBARow_NEON(const uint8_t* src_raw, uint8_t* dst_rgba, int width);
void RAWToARGBRow_MSA(const uint8_t* src_raw, uint8_t* dst_argb, int width);
@@ -2851,8 +2855,8 @@ void RGB24ToARGBRow_Any_LSX(const uint8_t* src_ptr,
uint8_t* dst_ptr,
int width);
void RGB24ToARGBRow_Any_LASX(const uint8_t* src_ptr,
- uint8_t* dst_ptr,
- int width);
+ uint8_t* dst_ptr,
+ int width);
void RAWToARGBRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
void RAWToRGBARow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
void RAWToARGBRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
diff --git a/include/libyuv/version.h b/include/libyuv/version.h
index e9ad82d9..d0a46371 100644
--- a/include/libyuv/version.h
+++ b/include/libyuv/version.h
@@ -11,6 +11,6 @@
#ifndef INCLUDE_LIBYUV_VERSION_H_
#define INCLUDE_LIBYUV_VERSION_H_
-#define LIBYUV_VERSION 1813
+#define LIBYUV_VERSION 1814
#endif // INCLUDE_LIBYUV_VERSION_H_
diff --git a/source/convert.cc b/source/convert.cc
index 8f02636d..45590a7b 100644
--- a/source/convert.cc
+++ b/source/convert.cc
@@ -1422,7 +1422,7 @@ int ARGBToI420(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYRow = ARGBToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYRow = ARGBToYRow_NEON;
}
}
@@ -1658,7 +1658,7 @@ int ABGRToI420(const uint8_t* src_abgr,
#if defined(HAS_ABGRTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ABGRToYRow = ABGRToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ABGRToYRow = ABGRToYRow_NEON;
}
}
@@ -1754,7 +1754,7 @@ int RGBAToI420(const uint8_t* src_rgba,
#if defined(HAS_RGBATOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
RGBAToYRow = RGBAToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
RGBAToYRow = RGBAToYRow_NEON;
}
}
@@ -1855,11 +1855,9 @@ int RGB24ToI420(const uint8_t* src_rgb24,
if (TestCpuFlag(kCpuHasNEON)) {
RGB24ToUVRow = RGB24ToUVRow_Any_NEON;
RGB24ToYRow = RGB24ToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
RGB24ToYRow = RGB24ToYRow_NEON;
- if (IS_ALIGNED(width, 16)) {
- RGB24ToUVRow = RGB24ToUVRow_NEON;
- }
+ RGB24ToUVRow = RGB24ToUVRow_NEON;
}
}
#endif
@@ -2031,11 +2029,9 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
if (TestCpuFlag(kCpuHasNEON)) {
RGB24ToUVJRow = RGB24ToUVJRow_Any_NEON;
RGB24ToYJRow = RGB24ToYJRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
RGB24ToYJRow = RGB24ToYJRow_NEON;
- if (IS_ALIGNED(width, 16)) {
- RGB24ToUVJRow = RGB24ToUVJRow_NEON;
- }
+ RGB24ToUVJRow = RGB24ToUVJRow_NEON;
}
}
#endif
@@ -2095,18 +2091,18 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
#endif
#endif // HAS_RGB24TOYJROW
- {
+{
#if !defined(HAS_RGB24TOYJROW)
- // Allocate 2 rows of ARGB.
- const int kRowSize = (width * 4 + 31) & ~31;
- align_buffer_64(row, kRowSize * 2);
+ // Allocate 2 rows of ARGB.
+ const int kRowSize = (width * 4 + 31) & ~31;
+ align_buffer_64(row, kRowSize * 2);
#endif
- for (y = 0; y < height - 1; y += 2) {
+ for (y = 0; y < height - 1; y += 2) {
#if defined(HAS_RGB24TOYJROW)
- RGB24ToUVJRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
- RGB24ToYJRow(src_rgb24, dst_y, width);
- RGB24ToYJRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
+ RGB24ToUVJRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
+ RGB24ToYJRow(src_rgb24, dst_y, width);
+ RGB24ToYJRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
#else
RGB24ToARGBRow(src_rgb24, row, width);
RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + kRowSize, width);
@@ -2114,26 +2110,26 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
ARGBToYJRow(row, dst_y, width);
ARGBToYJRow(row + kRowSize, dst_y + dst_stride_y, width);
#endif
- src_rgb24 += src_stride_rgb24 * 2;
- dst_y += dst_stride_y * 2;
- dst_u += dst_stride_u;
- dst_v += dst_stride_v;
- }
- if (height & 1) {
+ src_rgb24 += src_stride_rgb24 * 2;
+ dst_y += dst_stride_y * 2;
+ dst_u += dst_stride_u;
+ dst_v += dst_stride_v;
+ }
+ if (height & 1) {
#if defined(HAS_RGB24TOYJROW)
- RGB24ToUVJRow(src_rgb24, 0, dst_u, dst_v, width);
- RGB24ToYJRow(src_rgb24, dst_y, width);
+ RGB24ToUVJRow(src_rgb24, 0, dst_u, dst_v, width);
+ RGB24ToYJRow(src_rgb24, dst_y, width);
#else
RGB24ToARGBRow(src_rgb24, row, width);
ARGBToUVJRow(row, 0, dst_u, dst_v, width);
ARGBToYJRow(row, dst_y, width);
#endif
- }
+ }
#if !defined(HAS_RGB24TOYJROW)
- free_aligned_buffer_64(row);
+ free_aligned_buffer_64(row);
#endif
- }
- return 0;
+}
+return 0;
}
#undef HAS_RGB24TOYJROW
@@ -2187,11 +2183,9 @@ int RAWToI420(const uint8_t* src_raw,
if (TestCpuFlag(kCpuHasNEON)) {
RAWToUVRow = RAWToUVRow_Any_NEON;
RAWToYRow = RAWToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
RAWToYRow = RAWToYRow_NEON;
- if (IS_ALIGNED(width, 16)) {
- RAWToUVRow = RAWToUVRow_NEON;
- }
+ RAWToUVRow = RAWToUVRow_NEON;
}
}
#endif
@@ -2363,11 +2357,9 @@ int RAWToJ420(const uint8_t* src_raw,
if (TestCpuFlag(kCpuHasNEON)) {
RAWToUVJRow = RAWToUVJRow_Any_NEON;
RAWToYJRow = RAWToYJRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
RAWToYJRow = RAWToYJRow_NEON;
- if (IS_ALIGNED(width, 16)) {
- RAWToUVJRow = RAWToUVJRow_NEON;
- }
+ RAWToUVJRow = RAWToUVJRow_NEON;
}
}
#endif
@@ -2521,8 +2513,8 @@ int RGB565ToI420(const uint8_t* src_rgb565,
}
}
// MSA version does direct RGB565 to YUV.
-#elif (defined(HAS_RGB565TOYROW_MSA) || defined(HAS_RGB565TOYROW_LSX) \
- || defined(HAS_RGB565TOYROW_LASX))
+#elif (defined(HAS_RGB565TOYROW_MSA) || defined(HAS_RGB565TOYROW_LSX) || \
+ defined(HAS_RGB565TOYROW_LASX))
#if defined(HAS_RGB565TOYROW_MSA) && defined(HAS_RGB565TOUVROW_MSA)
if (TestCpuFlag(kCpuHasMSA)) {
RGB565ToUVRow = RGB565ToUVRow_Any_MSA;
@@ -2701,8 +2693,8 @@ int ARGB1555ToI420(const uint8_t* src_argb1555,
}
}
// MSA version does direct ARGB1555 to YUV.
-#elif (defined(HAS_ARGB1555TOYROW_MSA) || defined(HAS_ARGB1555TOYROW_LSX) \
- || defined(HAS_ARGB1555TOYROW_LASX))
+#elif (defined(HAS_ARGB1555TOYROW_MSA) || defined(HAS_ARGB1555TOYROW_LSX) || \
+ defined(HAS_ARGB1555TOYROW_LASX))
#if defined(HAS_ARGB1555TOYROW_MSA) && defined(HAS_ARGB1555TOUVROW_MSA)
if (TestCpuFlag(kCpuHasMSA)) {
ARGB1555ToUVRow = ARGB1555ToUVRow_Any_MSA;
@@ -3067,7 +3059,7 @@ int RGB24ToJ400(const uint8_t* src_rgb24,
#if defined(HAS_RGB24TOYJROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
RGB24ToYJRow = RGB24ToYJRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
RGB24ToYJRow = RGB24ToYJRow_NEON;
}
}
diff --git a/source/convert_from_argb.cc b/source/convert_from_argb.cc
index 89856e25..e50c2af3 100644
--- a/source/convert_from_argb.cc
+++ b/source/convert_from_argb.cc
@@ -103,7 +103,7 @@ int ARGBToI444(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYRow = ARGBToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYRow = ARGBToYRow_NEON;
}
}
@@ -205,7 +205,7 @@ int ARGBToI422(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYRow = ARGBToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYRow = ARGBToYRow_NEON;
}
}
@@ -283,7 +283,7 @@ int ARGBToNV12(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYRow = ARGBToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYRow = ARGBToYRow_NEON;
}
}
@@ -477,7 +477,7 @@ int ARGBToNV21(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYRow = ARGBToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYRow = ARGBToYRow_NEON;
}
}
@@ -638,7 +638,7 @@ int ABGRToNV12(const uint8_t* src_abgr,
#if defined(HAS_ABGRTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ABGRToYRow = ABGRToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ABGRToYRow = ABGRToYRow_NEON;
}
}
@@ -790,7 +790,7 @@ int ABGRToNV21(const uint8_t* src_abgr,
#if defined(HAS_ABGRTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ABGRToYRow = ABGRToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ABGRToYRow = ABGRToYRow_NEON;
}
}
@@ -947,7 +947,7 @@ int ARGBToYUY2(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYRow = ARGBToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYRow = ARGBToYRow_NEON;
}
}
@@ -1110,7 +1110,7 @@ int ARGBToUYVY(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYRow = ARGBToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYRow = ARGBToYRow_NEON;
}
}
@@ -1249,7 +1249,7 @@ int ARGBToI400(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYRow = ARGBToYRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYRow = ARGBToYRow_NEON;
}
}
@@ -1906,7 +1906,7 @@ int ARGBToJ420(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYJROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYJRow = ARGBToYJRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYJRow = ARGBToYJRow_NEON;
}
}
@@ -2029,7 +2029,7 @@ int ARGBToJ422(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYJROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYJRow = ARGBToYJRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYJRow = ARGBToYJRow_NEON;
}
}
@@ -2248,7 +2248,7 @@ int ARGBToJ400(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYJROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYJRow = ARGBToYJRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYJRow = ARGBToYJRow_NEON;
}
}
@@ -2314,7 +2314,7 @@ int RGBAToJ400(const uint8_t* src_rgba,
#if defined(HAS_RGBATOYJROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
RGBAToYJRow = RGBAToYJRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
RGBAToYJRow = RGBAToYJRow_NEON;
}
}
@@ -2387,11 +2387,9 @@ int RAWToJNV21(const uint8_t* src_raw,
if (TestCpuFlag(kCpuHasNEON)) {
RAWToUVJRow = RAWToUVJRow_Any_NEON;
RAWToYJRow = RAWToYJRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
RAWToYJRow = RAWToYJRow_NEON;
- if (IS_ALIGNED(width, 16)) {
- RAWToUVJRow = RAWToUVJRow_NEON;
- }
+ RAWToUVJRow = RAWToUVJRow_NEON;
}
}
#endif
diff --git a/source/planar_functions.cc b/source/planar_functions.cc
index 0c2ae2b1..43dc1892 100644
--- a/source/planar_functions.cc
+++ b/source/planar_functions.cc
@@ -4186,7 +4186,7 @@ static int ARGBSobelize(const uint8_t* src_argb,
#if defined(HAS_ARGBTOYJROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToYJRow = ARGBToYJRow_Any_NEON;
- if (IS_ALIGNED(width, 8)) {
+ if (IS_ALIGNED(width, 16)) {
ARGBToYJRow = ARGBToYJRow_NEON;
}
}
diff --git a/source/row_any.cc b/source/row_any.cc
index e79d23de..2d30e0a5 100644
--- a/source/row_any.cc
+++ b/source/row_any.cc
@@ -987,7 +987,7 @@ ANY11(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_SSSE3, 0, 4, 1, 15)
ANY11(RGBAToYJRow_Any_SSSE3, RGBAToYJRow_SSSE3, 0, 4, 1, 15)
#endif
#ifdef HAS_ARGBTOYROW_NEON
-ANY11(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 0, 4, 1, 7)
+ANY11(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 0, 4, 1, 15)
#endif
#ifdef HAS_ARGBTOYROW_MSA
ANY11(ARGBToYRow_Any_MSA, ARGBToYRow_MSA, 0, 4, 1, 15)
@@ -996,10 +996,10 @@ ANY11(ARGBToYRow_Any_MSA, ARGBToYRow_MSA, 0, 4, 1, 15)
ANY11(ARGBToYRow_Any_LASX, ARGBToYRow_LASX, 0, 4, 1, 31)
#endif
#ifdef HAS_ARGBTOYJROW_NEON
-ANY11(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 0, 4, 1, 7)
+ANY11(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 0, 4, 1, 15)
#endif
#ifdef HAS_RGBATOYJROW_NEON
-ANY11(RGBAToYJRow_Any_NEON, RGBAToYJRow_NEON, 0, 4, 1, 7)
+ANY11(RGBAToYJRow_Any_NEON, RGBAToYJRow_NEON, 0, 4, 1, 15)
#endif
#ifdef HAS_ARGBTOYJROW_MSA
ANY11(ARGBToYJRow_Any_MSA, ARGBToYJRow_MSA, 0, 4, 1, 15)
@@ -1011,7 +1011,7 @@ ANY11(ARGBToYJRow_Any_LSX, ARGBToYJRow_LSX, 0, 4, 1, 15)
ANY11(ARGBToYJRow_Any_LASX, ARGBToYJRow_LASX, 0, 4, 1, 31)
#endif
#ifdef HAS_BGRATOYROW_NEON
-ANY11(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 0, 4, 1, 7)
+ANY11(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 0, 4, 1, 15)
#endif
#ifdef HAS_BGRATOYROW_MSA
ANY11(BGRAToYRow_Any_MSA, BGRAToYRow_MSA, 0, 4, 1, 15)
@@ -1020,7 +1020,7 @@ ANY11(BGRAToYRow_Any_MSA, BGRAToYRow_MSA, 0, 4, 1, 15)
ANY11(BGRAToYRow_Any_LSX, BGRAToYRow_LSX, 0, 4, 1, 15)
#endif
#ifdef HAS_ABGRTOYROW_NEON
-ANY11(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 0, 4, 1, 7)
+ANY11(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 0, 4, 1, 15)
#endif
#ifdef HAS_ABGRTOYROW_MSA
ANY11(ABGRToYRow_Any_MSA, ABGRToYRow_MSA, 0, 4, 1, 7)
@@ -1029,7 +1029,7 @@ ANY11(ABGRToYRow_Any_MSA, ABGRToYRow_MSA, 0, 4, 1, 7)
ANY11(ABGRToYRow_Any_LSX, ABGRToYRow_LSX, 0, 4, 1, 15)
#endif
#ifdef HAS_RGBATOYROW_NEON
-ANY11(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 0, 4, 1, 7)
+ANY11(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 0, 4, 1, 15)
#endif
#ifdef HAS_RGBATOYROW_MSA
ANY11(RGBAToYRow_Any_MSA, RGBAToYRow_MSA, 0, 4, 1, 15)
@@ -1038,7 +1038,7 @@ ANY11(RGBAToYRow_Any_MSA, RGBAToYRow_MSA, 0, 4, 1, 15)
ANY11(RGBAToYRow_Any_LSX, RGBAToYRow_LSX, 0, 4, 1, 15)
#endif
#ifdef HAS_RGB24TOYROW_NEON
-ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 7)
+ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 15)
#endif
#ifdef HAS_RGB24TOYJROW_AVX2
ANY11(RGB24ToYJRow_Any_AVX2, RGB24ToYJRow_AVX2, 0, 3, 1, 31)
@@ -1047,7 +1047,7 @@ ANY11(RGB24ToYJRow_Any_AVX2, RGB24ToYJRow_AVX2, 0, 3, 1, 31)
ANY11(RGB24ToYJRow_Any_SSSE3, RGB24ToYJRow_SSSE3, 0, 3, 1, 15)
#endif
#ifdef HAS_RGB24TOYJROW_NEON
-ANY11(RGB24ToYJRow_Any_NEON, RGB24ToYJRow_NEON, 0, 3, 1, 7)
+ANY11(RGB24ToYJRow_Any_NEON, RGB24ToYJRow_NEON, 0, 3, 1, 15)
#endif
#ifdef HAS_RGB24TOYROW_MSA
ANY11(RGB24ToYRow_Any_MSA, RGB24ToYRow_MSA, 0, 3, 1, 15)
@@ -1059,7 +1059,7 @@ ANY11(RGB24ToYRow_Any_LSX, RGB24ToYRow_LSX, 0, 3, 1, 15)
ANY11(RGB24ToYRow_Any_LASX, RGB24ToYRow_LASX, 0, 3, 1, 31)
#endif
#ifdef HAS_RAWTOYROW_NEON
-ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 7)
+ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 15)
#endif
#ifdef HAS_RAWTOYJROW_AVX2
ANY11(RAWToYJRow_Any_AVX2, RAWToYJRow_AVX2, 0, 3, 1, 31)
@@ -1068,7 +1068,7 @@ ANY11(RAWToYJRow_Any_AVX2, RAWToYJRow_AVX2, 0, 3, 1, 31)
ANY11(RAWToYJRow_Any_SSSE3, RAWToYJRow_SSSE3, 0, 3, 1, 15)
#endif
#ifdef HAS_RAWTOYJROW_NEON
-ANY11(RAWToYJRow_Any_NEON, RAWToYJRow_NEON, 0, 3, 1, 7)
+ANY11(RAWToYJRow_Any_NEON, RAWToYJRow_NEON, 0, 3, 1, 15)
#endif
#ifdef HAS_RAWTOYROW_MSA
ANY11(RAWToYRow_Any_MSA, RAWToYRow_MSA, 0, 3, 1, 15)
diff --git a/source/row_lasx.cc b/source/row_lasx.cc
index 12891ff5..7dd18f40 100644
--- a/source/row_lasx.cc
+++ b/source/row_lasx.cc
@@ -23,14 +23,14 @@ extern "C" {
#define ALPHA_VAL (-1)
// Fill YUV -> RGB conversion constants into vectors
-#define YUVTORGB_SETUP(yuvconst, ub, vr, ug, vg, yg, yb) \
- { \
- ub = __lasx_xvreplgr2vr_h(yuvconst->kUVToB[0]); \
- vr = __lasx_xvreplgr2vr_h(yuvconst->kUVToR[1]); \
- ug = __lasx_xvreplgr2vr_h(yuvconst->kUVToG[0]); \
- vg = __lasx_xvreplgr2vr_h(yuvconst->kUVToG[1]); \
- yg = __lasx_xvreplgr2vr_h(yuvconst->kYToRgb[0]); \
- yb = __lasx_xvreplgr2vr_w(yuvconst->kYBiasToRgb[0]); \
+#define YUVTORGB_SETUP(yuvconst, ub, vr, ug, vg, yg, yb) \
+ { \
+ ub = __lasx_xvreplgr2vr_h(yuvconst->kUVToB[0]); \
+ vr = __lasx_xvreplgr2vr_h(yuvconst->kUVToR[1]); \
+ ug = __lasx_xvreplgr2vr_h(yuvconst->kUVToG[0]); \
+ vg = __lasx_xvreplgr2vr_h(yuvconst->kUVToG[1]); \
+ yg = __lasx_xvreplgr2vr_h(yuvconst->kYToRgb[0]); \
+ yb = __lasx_xvreplgr2vr_w(yuvconst->kYBiasToRgb[0]); \
}
// Load 32 YUV422 pixel data
@@ -191,27 +191,26 @@ extern "C" {
pdst_argb += 64; \
}
-#define RGBTOUV(_tmpb, _tmpg, _tmpr, _nexb, _nexg, _nexr, _reg0, _reg1) \
- { \
- __m256i _tmp0, _tmp1, _tmp2, _tmp3; \
- _tmp0 = __lasx_xvaddwev_h_bu(_tmpb, _nexb); \
- _tmp1 = __lasx_xvaddwod_h_bu(_tmpb, _nexb); \
- _tmp2 = __lasx_xvaddwev_h_bu(_tmpg, _nexg); \
- _tmp3 = __lasx_xvaddwod_h_bu(_tmpg, _nexg); \
- _reg0 = __lasx_xvaddwev_h_bu(_tmpr, _nexr); \
- _reg1 = __lasx_xvaddwod_h_bu(_tmpr, _nexr); \
- _tmpb = __lasx_xvavgr_hu(_tmp0, _tmp1); \
- _tmpg = __lasx_xvavgr_hu(_tmp2, _tmp3); \
- _tmpr = __lasx_xvavgr_hu(_reg0, _reg1); \
- _reg0 = __lasx_xvmadd_h(const_8080, const_112, _tmpb); \
- _reg1 = __lasx_xvmadd_h(const_8080, const_112, _tmpr); \
- _reg0 = __lasx_xvmsub_h(_reg0, const_74, _tmpg); \
- _reg1 = __lasx_xvmsub_h(_reg1, const_94, _tmpg); \
- _reg0 = __lasx_xvmsub_h(_reg0, const_38, _tmpr); \
- _reg1 = __lasx_xvmsub_h(_reg1, const_18, _tmpb); \
+#define RGBTOUV(_tmpb, _tmpg, _tmpr, _nexb, _nexg, _nexr, _reg0, _reg1) \
+ { \
+ __m256i _tmp0, _tmp1, _tmp2, _tmp3; \
+ _tmp0 = __lasx_xvaddwev_h_bu(_tmpb, _nexb); \
+ _tmp1 = __lasx_xvaddwod_h_bu(_tmpb, _nexb); \
+ _tmp2 = __lasx_xvaddwev_h_bu(_tmpg, _nexg); \
+ _tmp3 = __lasx_xvaddwod_h_bu(_tmpg, _nexg); \
+ _reg0 = __lasx_xvaddwev_h_bu(_tmpr, _nexr); \
+ _reg1 = __lasx_xvaddwod_h_bu(_tmpr, _nexr); \
+ _tmpb = __lasx_xvavgr_hu(_tmp0, _tmp1); \
+ _tmpg = __lasx_xvavgr_hu(_tmp2, _tmp3); \
+ _tmpr = __lasx_xvavgr_hu(_reg0, _reg1); \
+ _reg0 = __lasx_xvmadd_h(const_8080, const_112, _tmpb); \
+ _reg1 = __lasx_xvmadd_h(const_8080, const_112, _tmpr); \
+ _reg0 = __lasx_xvmsub_h(_reg0, const_74, _tmpg); \
+ _reg1 = __lasx_xvmsub_h(_reg1, const_94, _tmpg); \
+ _reg0 = __lasx_xvmsub_h(_reg0, const_38, _tmpr); \
+ _reg1 = __lasx_xvmsub_h(_reg1, const_18, _tmpb); \
}
-
void MirrorRow_LASX(const uint8_t* src, uint8_t* dst, int width) {
int x;
int len = width / 64;
@@ -596,8 +595,8 @@ void I422ToARGB1555Row_LASX(const uint8_t* src_y,
__m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
__m256i vec_ubvr, vec_ugvg;
__m256i const_0x80 = __lasx_xvldi(0x80);
- __m256i alpha = {0x8000800080008000, 0x8000800080008000,
- 0x8000800080008000, 0x8000800080008000};
+ __m256i alpha = {0x8000800080008000, 0x8000800080008000, 0x8000800080008000,
+ 0x8000800080008000};
YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
@@ -1507,14 +1506,14 @@ void RGB24ToARGBRow_LASX(const uint8_t* src_rgb24,
__m256i dst0, dst1, dst2, dst3;
__m256i reg0, reg1, reg2, reg3;
__m256i alpha = __lasx_xvldi(0xFF);
- __m256i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514,
- 0x131211100F0E0D0C, 0x1B1A191817161514};
- __m256i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100,
- 0x1F1E1D1C1B1A1918, 0x0706050403020100};
- __m256i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C,
- 0x0B0A090807060504, 0x131211100F0E0D0C};
- __m256i shuf3 = {0x1005040310020100, 0x100B0A0910080706,
- 0x1005040310020100, 0x100B0A0910080706};
+ __m256i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514, 0x131211100F0E0D0C,
+ 0x1B1A191817161514};
+ __m256i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100, 0x1F1E1D1C1B1A1918,
+ 0x0706050403020100};
+ __m256i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C, 0x0B0A090807060504,
+ 0x131211100F0E0D0C};
+ __m256i shuf3 = {0x1005040310020100, 0x100B0A0910080706, 0x1005040310020100,
+ 0x100B0A0910080706};
for (x = 0; x < len; x++) {
reg0 = __lasx_xvld(src_rgb24, 0);
@@ -1523,7 +1522,8 @@ void RGB24ToARGBRow_LASX(const uint8_t* src_rgb24,
src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
- DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0, tmp1);
+ DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0,
+ tmp1);
tmp2 = __lasx_xvshuf_b(src1, src2, shuf2);
DUP4_ARG3(__lasx_xvshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
tmp1, shuf3, alpha, tmp2, shuf3, reg0, reg1, reg2, reg3);
@@ -1545,14 +1545,14 @@ void RAWToARGBRow_LASX(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
__m256i tmp0, tmp1, tmp2, reg0, reg1, reg2, reg3;
__m256i dst0, dst1, dst2, dst3;
__m256i alpha = __lasx_xvldi(0xFF);
- __m256i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514,
- 0x131211100F0E0D0C, 0x1B1A191817161514};
- __m256i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100,
- 0x1F1E1D1C1B1A1918, 0x0706050403020100};
- __m256i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C,
- 0x0B0A090807060504, 0x131211100F0E0D0C};
- __m256i shuf3 = {0x1003040510000102, 0x10090A0B10060708,
- 0x1003040510000102, 0x10090A0B10060708};
+ __m256i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514, 0x131211100F0E0D0C,
+ 0x1B1A191817161514};
+ __m256i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100, 0x1F1E1D1C1B1A1918,
+ 0x0706050403020100};
+ __m256i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C, 0x0B0A090807060504,
+ 0x131211100F0E0D0C};
+ __m256i shuf3 = {0x1003040510000102, 0x10090A0B10060708, 0x1003040510000102,
+ 0x10090A0B10060708};
for (x = 0; x < len; x++) {
reg0 = __lasx_xvld(src_raw, 0);
@@ -1561,7 +1561,8 @@ void RAWToARGBRow_LASX(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
- DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0, tmp1);
+ DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0,
+ tmp1);
tmp2 = __lasx_xvshuf_b(src1, src2, shuf2);
DUP4_ARG3(__lasx_xvshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
tmp1, shuf3, alpha, tmp2, shuf3, reg0, reg1, reg2, reg3);
@@ -1577,8 +1578,8 @@ void RAWToARGBRow_LASX(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
}
void ARGB1555ToYRow_LASX(const uint8_t* src_argb1555,
- uint8_t* dst_y,
- int width) {
+ uint8_t* dst_y,
+ int width) {
int x;
int len = width / 32;
__m256i src0, src1;
@@ -1646,8 +1647,8 @@ void ARGB1555ToUVRow_LASX(const uint8_t* src_argb1555,
0x8080808080808080, 0x8080808080808080};
for (x = 0; x < len; x++) {
- DUP4_ARG2(__lasx_xvld, src_argb1555, 0, src_argb1555, 32, next_argb1555,
- 0, next_argb1555, 32, src0, src1, src2, src3);
+ DUP4_ARG2(__lasx_xvld, src_argb1555, 0, src_argb1555, 32, next_argb1555, 0,
+ next_argb1555, 32, src0, src1, src2, src3);
DUP2_ARG2(__lasx_xvpickev_b, src1, src0, src3, src2, tmp0, tmp2);
DUP2_ARG2(__lasx_xvpickod_b, src1, src0, src3, src2, tmp1, tmp3);
tmpb = __lasx_xvandi_b(tmp0, 0x1F);
@@ -1821,14 +1822,14 @@ void RGB24ToYRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
0x4219421942194219, 0x4219421942194219};
__m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
0x1080108010801080, 0x1080108010801080};
- __m256i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C,
- 0x0B09080605030200, 0x17151412110F0E0C};
- __m256i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604,
- 0x0301001E1D1B1A18, 0x0F0D0C0A09070604};
- __m256i shuff2 = {0x000A000700040001, 0x001600130010000D,
- 0x000A000700040001, 0x001600130010000D};
- __m256i shuff3 = {0x0002001F001C0019, 0x000E000B00080005,
- 0x0002001F001C0019, 0x000E000B00080005};
+ __m256i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C, 0x0B09080605030200,
+ 0x17151412110F0E0C};
+ __m256i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604, 0x0301001E1D1B1A18,
+ 0x0F0D0C0A09070604};
+ __m256i shuff2 = {0x000A000700040001, 0x001600130010000D, 0x000A000700040001,
+ 0x001600130010000D};
+ __m256i shuff3 = {0x0002001F001C0019, 0x000E000B00080005, 0x0002001F001C0019,
+ 0x000E000B00080005};
for (x = 0; x < len; x++) {
reg0 = __lasx_xvld(src_rgb24, 0);
@@ -1887,8 +1888,8 @@ void RGB24ToUVRow_LASX(const uint8_t* src_rgb24,
DUP4_ARG2(__lasx_xvld, src_rgb24, 0, src_rgb24, 32, src_rgb24, 64,
next_rgb24, 0, reg0, reg1, reg2, tmp0);
DUP2_ARG2(__lasx_xvld, next_rgb24, 32, next_rgb24, 64, tmp1, tmp2);
- DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x30, reg2, reg0, 0x21, reg2,
- reg1, 0x30, tmp1, tmp0, 0x30, src0, src1, src2, nex0);
+ DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x30, reg2, reg0, 0x21, reg2, reg1,
+ 0x30, tmp1, tmp0, 0x30, src0, src1, src2, nex0);
DUP2_ARG3(__lasx_xvpermi_q, tmp2, tmp0, 0x21, tmp2, tmp1, 0x30, nex1, nex2);
DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
nexb);
@@ -1926,14 +1927,14 @@ void RAWToYRow_LASX(const uint8_t* src_raw, uint8_t* dst_y, int width) {
0x1942194219421942, 0x1942194219421942};
__m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
0x1080108010801080, 0x1080108010801080};
- __m256i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C,
- 0x0B09080605030200, 0x17151412110F0E0C};
- __m256i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604,
- 0x0301001E1D1B1A18, 0x0F0D0C0A09070604};
- __m256i shuff2 = {0x000A000700040001, 0x001600130010000D,
- 0x000A000700040001, 0x001600130010000D};
- __m256i shuff3 = {0x0002001F001C0019, 0x000E000B00080005,
- 0x0002001F001C0019, 0x000E000B00080005};
+ __m256i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C, 0x0B09080605030200,
+ 0x17151412110F0E0C};
+ __m256i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604, 0x0301001E1D1B1A18,
+ 0x0F0D0C0A09070604};
+ __m256i shuff2 = {0x000A000700040001, 0x001600130010000D, 0x000A000700040001,
+ 0x001600130010000D};
+ __m256i shuff3 = {0x0002001F001C0019, 0x000E000B00080005, 0x0002001F001C0019,
+ 0x000E000B00080005};
for (x = 0; x < len; x++) {
reg0 = __lasx_xvld(src_raw, 0);
@@ -1989,24 +1990,24 @@ void RAWToUVRow_LASX(const uint8_t* src_raw,
0x0706050403020100, 0x1F1C191613100908};
for (x = 0; x < len; x++) {
- DUP4_ARG2(__lasx_xvld, src_raw, 0, src_raw, 32, src_raw, 64,
- next_raw, 0, reg0, reg1, reg2, tmp0);
+ DUP4_ARG2(__lasx_xvld, src_raw, 0, src_raw, 32, src_raw, 64, next_raw, 0,
+ reg0, reg1, reg2, tmp0);
DUP2_ARG2(__lasx_xvld, next_raw, 32, next_raw, 64, tmp1, tmp2);
- DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x30, reg2, reg0, 0x21, reg2,
- reg1, 0x30, tmp1, tmp0, 0x30, src0, src1, src2, nex0);
+ DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x30, reg2, reg0, 0x21, reg2, reg1,
+ 0x30, tmp1, tmp0, 0x30, src0, src1, src2, nex0);
DUP2_ARG3(__lasx_xvpermi_q, tmp2, tmp0, 0x21, tmp2, tmp1, 0x30, nex1, nex2);
- DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b,
- tmpb, nexb);
- DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g,
- tmpg, nexg);
- DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r,
- tmpr, nexr);
- DUP2_ARG3(__lasx_xvshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b,
- tmpb, nexb);
- DUP2_ARG3(__lasx_xvshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g,
- tmpg, nexg);
- DUP2_ARG3(__lasx_xvshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r,
- tmpr, nexr);
+ DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
+ nexb);
+ DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g, tmpg,
+ nexg);
+ DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r, tmpr,
+ nexr);
+ DUP2_ARG3(__lasx_xvshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b, tmpb,
+ nexb);
+ DUP2_ARG3(__lasx_xvshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g, tmpg,
+ nexg);
+ DUP2_ARG3(__lasx_xvshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r, tmpr,
+ nexr);
RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, reg0, reg1);
dst0 = __lasx_xvpickod_b(reg1, reg0);
__lasx_xvstelm_d(dst0, dst_u, 0, 0);
@@ -2071,8 +2072,8 @@ void NV12ToRGB565Row_LASX(const uint8_t* src_y,
vec_vu = __lasx_xvld(src_uv, 0);
vec_vu = __lasx_xvsub_b(vec_vu, const_0x80);
vec_vu = __lasx_vext2xv_h_b(vec_vu);
- YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb,
- out_r, out_g, out_b);
+ YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_r, out_g,
+ out_b);
out_b = __lasx_xvsrli_h(out_b, 3);
out_g = __lasx_xvsrli_h(out_g, 2);
out_r = __lasx_xvsrli_h(out_r, 3);
@@ -2109,8 +2110,8 @@ void NV21ToARGBRow_LASX(const uint8_t* src_y,
vec_uv = __lasx_xvld(src_uv, 0);
vec_uv = __lasx_xvsub_b(vec_uv, const_0x80);
vec_uv = __lasx_vext2xv_h_b(vec_uv);
- YUVTORGB(vec_y, vec_uv, vec_ubvr, vec_ugvg, vec_yg, vec_yb, out_b,
- out_g, out_r);
+ YUVTORGB(vec_y, vec_uv, vec_ubvr, vec_ugvg, vec_yg, vec_yb, out_b, out_g,
+ out_r);
STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
src_y += 16;
src_uv += 16;
@@ -2127,8 +2128,8 @@ void ARGBToYJRow_LASX(const uint8_t* src_argb, uint8_t* dst_y, int width) {
__m256i const_150 = __lasx_xvldi(0x96);
__m256i const_br = {0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D,
0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D};
- __m256i shuff = {0x0000000400000000, 0x0000000500000001,
- 0x0000000600000002, 0x0000000700000003};
+ __m256i shuff = {0x0000000400000000, 0x0000000500000001, 0x0000000600000002,
+ 0x0000000700000003};
for (x = 0; x < len; x++) {
DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
@@ -2169,8 +2170,8 @@ void ARGBToUVJRow_LASX(const uint8_t* src_argb,
__m256i const_10 = __lasx_xvldi(0x40A);
__m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
0x8080808080808080, 0x8080808080808080};
- __m256i shuff = {0x1614060412100200, 0x1E1C0E0C1A180A08,
- 0x1715070513110301, 0x1F1D0F0D1B190B09};
+ __m256i shuff = {0x1614060412100200, 0x1E1C0E0C1A180A08, 0x1715070513110301,
+ 0x1F1D0F0D1B190B09};
for (x = 0; x < len; x++) {
DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
diff --git a/source/row_neon.cc b/source/row_neon.cc
index e1063236..aa3ac70a 100644
--- a/source/row_neon.cc
+++ b/source/row_neon.cc
@@ -1645,29 +1645,6 @@ void ARGBToARGB4444Row_NEON(const uint8_t* src_argb,
: "cc", "memory", "q0", "q1", "q2", "q3");
}
-void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
- asm volatile(
- "vmov.u8 d24, #25 \n" // B * 0.1016 coefficient
- "vmov.u8 d25, #129 \n" // G * 0.5078 coefficient
- "vmov.u8 d26, #66 \n" // R * 0.2578 coefficient
- "vmov.u8 d27, #16 \n" // Add 16 constant
- "1: \n"
- "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q2, d0, d24 \n" // B
- "vmlal.u8 q2, d1, d25 \n" // G
- "vmlal.u8 q2, d2, d26 \n" // R
- "vqrshrn.u16 d0, q2, #8 \n" // 16 bit to 8 bit Y
- "vqadd.u8 d0, d27 \n"
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
- "bgt 1b \n"
- : "+r"(src_argb), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "q0", "q1", "q2", "q12", "q13");
-}
-
void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
uint8_t* dst_a,
int width) {
@@ -1686,48 +1663,6 @@ void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
);
}
-void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
- asm volatile(
- "vmov.u8 d24, #29 \n" // B * 0.1140 coefficient
- "vmov.u8 d25, #150 \n" // G * 0.5870 coefficient
- "vmov.u8 d26, #77 \n" // R * 0.2990 coefficient
- "1: \n"
- "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q2, d0, d24 \n" // B
- "vmlal.u8 q2, d1, d25 \n" // G
- "vmlal.u8 q2, d2, d26 \n" // R
- "vqrshrn.u16 d0, q2, #8 \n" // 16 bit to 8 bit Y
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
- "bgt 1b \n"
- : "+r"(src_argb), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "q0", "q1", "q2", "q12", "q13");
-}
-
-void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
- asm volatile(
- "vmov.u8 d24, #29 \n" // B * 0.1140 coefficient
- "vmov.u8 d25, #150 \n" // G * 0.5870 coefficient
- "vmov.u8 d26, #77 \n" // R * 0.2990 coefficient
- "1: \n"
- "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 RGBA pixels.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q2, d1, d24 \n" // B
- "vmlal.u8 q2, d2, d25 \n" // G
- "vmlal.u8 q2, d3, d26 \n" // R
- "vqrshrn.u16 d0, q2, #8 \n" // 16 bit to 8 bit Y
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
- "bgt 1b \n"
- : "+r"(src_rgba), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "q0", "q1", "q2", "q12", "q13");
-}
-
// 8x1 pixels.
void ARGBToUV444Row_NEON(const uint8_t* src_argb,
uint8_t* dst_u,
@@ -1747,15 +1682,13 @@ void ARGBToUV444Row_NEON(const uint8_t* src_argb,
"vmull.u8 q2, d0, d24 \n" // B
"vmlsl.u8 q2, d1, d25 \n" // G
"vmlsl.u8 q2, d2, d26 \n" // R
- "vadd.u16 q2, q2, q15 \n" // +128 -> unsigned
"vmull.u8 q3, d2, d24 \n" // R
"vmlsl.u8 q3, d1, d28 \n" // G
"vmlsl.u8 q3, d0, d27 \n" // B
- "vadd.u16 q3, q3, q15 \n" // +128 -> unsigned
- "vqshrn.u16 d0, q2, #8 \n" // 16 bit to 8 bit U
- "vqshrn.u16 d1, q3, #8 \n" // 16 bit to 8 bit V
+ "vaddhn.u16 d0, q2, q15 \n" // +128 -> unsigned
+ "vaddhn.u16 d1, q3, q15 \n" // +128 -> unsigned
"vst1.8 {d0}, [%1]! \n" // store 8 pixels U.
"vst1.8 {d1}, [%2]! \n" // store 8 pixels V.
@@ -1775,13 +1708,11 @@ void ARGBToUV444Row_NEON(const uint8_t* src_argb,
"vmul.s16 q8, " #QB ", q10 \n" /* B */ \
"vmls.s16 q8, " #QG ", q11 \n" /* G */ \
"vmls.s16 q8, " #QR ", q12 \n" /* R */ \
- "vadd.u16 q8, q8, q15 \n" /* +128 -> unsigned */ \
"vmul.s16 q9, " #QR ", q10 \n" /* R */ \
"vmls.s16 q9, " #QG ", q14 \n" /* G */ \
"vmls.s16 q9, " #QB ", q13 \n" /* B */ \
- "vadd.u16 q9, q9, q15 \n" /* +128 -> unsigned */ \
- "vqshrn.u16 d0, q8, #8 \n" /* 16 bit to 8 bit U */ \
- "vqshrn.u16 d1, q9, #8 \n" /* 16 bit to 8 bit V */
+ "vaddhn.u16 d0, q8, q15 \n" /* +128 -> unsigned */ \
+ "vaddhn.u16 d1, q9, q15 \n" /* +128 -> unsigned */
// clang-format on
// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
@@ -2559,161 +2490,169 @@ void AB64ToARGBRow_NEON(const uint16_t* src_ab64,
: "cc", "memory", "q0", "q1", "q2", "q3", "q4");
}
-void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+struct RgbConstants {
+ uint8_t kRGBToY[4];
+ uint16_t kAddY;
+ uint16_t pad;
+};
+
+// RGB to JPeg coefficients
+// B * 0.1140 coefficient = 29
+// G * 0.5870 coefficient = 150
+// R * 0.2990 coefficient = 77
+// Add 0.5 = 0x80
+struct RgbConstants kRgb24JPEGConstants = {{29, 150, 77, 0}, 128};
+
+struct RgbConstants kRawJPEGConstants = {{77, 150, 29, 0}, 128};
+
+// RGB to BT.601 coefficients
+// B * 0.1016 coefficient = 25
+// G * 0.5078 coefficient = 129
+// R * 0.2578 coefficient = 66
+// Add 16.5 = 0x1080
+
+struct RgbConstants kRgb24I601Constants = {{25, 129, 66, 0}, 0x1080};
+
+struct RgbConstants kRawI601Constants = {{66, 129, 25, 0}, 0x1080};
+
+// ARGB expects first 3 values to contain RGB and 4th value is ignored.
+void ARGBToYMatrixRow_NEON(const uint8_t* src_argb,
+ uint8_t* dst_y,
+ int width,
+ const struct RgbConstants* rgbconstants) {
asm volatile(
- "vmov.u8 d6, #25 \n" // B * 0.1016 coefficient
- "vmov.u8 d5, #129 \n" // G * 0.5078 coefficient
- "vmov.u8 d4, #66 \n" // R * 0.2578 coefficient
- "vmov.u8 d7, #16 \n" // Add 16 constant
+ "vld1.8 {d0}, [%3] \n" // load rgbconstants
+ "vdup.u8 d20, d0[0] \n"
+ "vdup.u8 d21, d0[1] \n"
+ "vdup.u8 d22, d0[2] \n"
+ "vdup.u16 q12, d0[2] \n"
"1: \n"
- "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of BGRA.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q8, d1, d4 \n" // R
- "vmlal.u8 q8, d2, d5 \n" // G
- "vmlal.u8 q8, d3, d6 \n" // B
- "vqrshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit Y
- "vqadd.u8 d0, d7 \n"
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
+ "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 16 pixels of ARGB
+ "vld4.8 {d1, d3, d5, d7}, [%0]! \n"
+ "subs %2, %2, #16 \n" // 16 processed per loop.
+ "vmull.u8 q8, d0, d20 \n" // B
+ "vmull.u8 q9, d1, d20 \n"
+ "vmlal.u8 q8, d2, d21 \n" // G
+ "vmlal.u8 q9, d3, d21 \n"
+ "vmlal.u8 q8, d4, d22 \n" // R
+ "vmlal.u8 q9, d5, d22 \n"
+ "vaddhn.u16 d0, q8, q12 \n" // 16 bit to 8 bit Y
+ "vaddhn.u16 d1, q9, q12 \n"
+ "vst1.8 {d0, d1}, [%1]! \n" // store 16 pixels Y.
"bgt 1b \n"
- : "+r"(src_bgra), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+ : "+r"(src_argb), // %0
+ "+r"(dst_y), // %1
+ "+r"(width) // %2
+ : "r"(rgbconstants) // %3
+ : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "d20", "d21", "d22",
+ "q12");
+}
+
+void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+ ARGBToYMatrixRow_NEON(src_argb, dst_y, width, &kRgb24I601Constants);
+}
+
+void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_yj, int width) {
+ ARGBToYMatrixRow_NEON(src_argb, dst_yj, width, &kRgb24JPEGConstants);
}
void ABGRToYRow_NEON(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+ ARGBToYMatrixRow_NEON(src_abgr, dst_y, width, &kRawI601Constants);
+}
+
+// RGBA expects first value to be A and ignored, then 3 values to contain RGB.
+// Same code as ARGB, except the LD4
+void RGBAToYMatrixRow_NEON(const uint8_t* src_rgba,
+ uint8_t* dst_y,
+ int width,
+ const struct RgbConstants* rgbconstants) {
asm volatile(
- "vmov.u8 d6, #25 \n" // B * 0.1016 coefficient
- "vmov.u8 d5, #129 \n" // G * 0.5078 coefficient
- "vmov.u8 d4, #66 \n" // R * 0.2578 coefficient
- "vmov.u8 d7, #16 \n" // Add 16 constant
+ "vld1.8 {d0}, [%3] \n" // load rgbconstants
+ "vdup.u8 d20, d0[0] \n"
+ "vdup.u8 d21, d0[1] \n"
+ "vdup.u8 d22, d0[2] \n"
+ "vdup.u16 q12, d0[2] \n"
"1: \n"
- "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of ABGR.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q8, d0, d4 \n" // R
- "vmlal.u8 q8, d1, d5 \n" // G
- "vmlal.u8 q8, d2, d6 \n" // B
- "vqrshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit Y
- "vqadd.u8 d0, d7 \n"
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
- "bgt 1b \n"
- : "+r"(src_abgr), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+ "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 16 pixels of RGBA
+ "vld4.8 {d1, d3, d5, d7}, [%0]! \n"
+ "subs %2, %2, #16 \n" // 16 processed per loop.
+ "vmull.u8 q8, d2, d20 \n" // B
+ "vmull.u8 q9, d3, d20 \n"
+ "vmlal.u8 q8, d4, d21 \n" // G
+ "vmlal.u8 q9, d5, d21 \n"
+ "vmlal.u8 q8, d6, d22 \n" // R
+ "vmlal.u8 q9, d7, d22 \n"
+ "vaddhn.u16 d0, q8, q12 \n" // 16 bit to 8 bit Y
+ "vaddhn.u16 d1, q9, q12 \n"
+ "vst1.8 {d0, d1}, [%1]! \n" // store 16 pixels Y.
+ "bgt 1b \n"
+ : "+r"(src_rgba), // %0
+ "+r"(dst_y), // %1
+ "+r"(width) // %2
+ : "r"(rgbconstants) // %3
+ : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "d20", "d21", "d22",
+ "q12");
}
void RGBAToYRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
- asm volatile(
- "vmov.u8 d4, #25 \n" // B * 0.1016 coefficient
- "vmov.u8 d5, #129 \n" // G * 0.5078 coefficient
- "vmov.u8 d6, #66 \n" // R * 0.2578 coefficient
- "vmov.u8 d7, #16 \n" // Add 16 constant
- "1: \n"
- "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of RGBA.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q8, d1, d4 \n" // B
- "vmlal.u8 q8, d2, d5 \n" // G
- "vmlal.u8 q8, d3, d6 \n" // R
- "vqrshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit Y
- "vqadd.u8 d0, d7 \n"
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
- "bgt 1b \n"
- : "+r"(src_rgba), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+ RGBAToYMatrixRow_NEON(src_rgba, dst_y, width, &kRgb24I601Constants);
}
-void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
- asm volatile(
- "vmov.u8 d4, #25 \n" // B * 0.1016 coefficient
- "vmov.u8 d5, #129 \n" // G * 0.5078 coefficient
- "vmov.u8 d6, #66 \n" // R * 0.2578 coefficient
- "vmov.u8 d7, #16 \n" // Add 16 constant
- "1: \n"
- "vld3.8 {d0, d1, d2}, [%0]! \n" // load 8 pixels of RGB24.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q8, d0, d4 \n" // B
- "vmlal.u8 q8, d1, d5 \n" // G
- "vmlal.u8 q8, d2, d6 \n" // R
- "vqrshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit Y
- "vqadd.u8 d0, d7 \n"
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
- "bgt 1b \n"
- : "+r"(src_rgb24), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_yj, int width) {
+ RGBAToYMatrixRow_NEON(src_rgba, dst_yj, width, &kRgb24JPEGConstants);
}
-void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+ RGBAToYMatrixRow_NEON(src_bgra, dst_y, width, &kRawI601Constants);
+}
+
+void RGBToYMatrixRow_NEON(const uint8_t* src_rgb,
+ uint8_t* dst_y,
+ int width,
+ const struct RgbConstants* rgbconstants) {
asm volatile(
- "vmov.u8 d6, #25 \n" // B * 0.1016 coefficient
- "vmov.u8 d5, #129 \n" // G * 0.5078 coefficient
- "vmov.u8 d4, #66 \n" // R * 0.2578 coefficient
- "vmov.u8 d7, #16 \n" // Add 16 constant
+ "vld1.8 {d0}, [%3] \n" // load rgbconstants
+ "vdup.u8 d20, d0[0] \n"
+ "vdup.u8 d21, d0[1] \n"
+ "vdup.u8 d22, d0[2] \n"
+ "vdup.u16 q12, d0[2] \n"
"1: \n"
- "vld3.8 {d0, d1, d2}, [%0]! \n" // load 8 pixels of RAW.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q8, d0, d4 \n" // B
- "vmlal.u8 q8, d1, d5 \n" // G
- "vmlal.u8 q8, d2, d6 \n" // R
- "vqrshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit Y
- "vqadd.u8 d0, d7 \n"
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
- "bgt 1b \n"
- : "+r"(src_raw), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+ "vld3.8 {d2, d4, d6}, [%0]! \n" // load 16 pixels of
+ // RGB24.
+ "vld3.8 {d3, d5, d7}, [%0]! \n"
+ "subs %2, %2, #16 \n" // 16 processed per loop.
+ "vmull.u8 q8, d2, d20 \n" // B
+ "vmull.u8 q9, d3, d20 \n"
+ "vmlal.u8 q8, d4, d21 \n" // G
+ "vmlal.u8 q9, d5, d21 \n"
+ "vmlal.u8 q8, d6, d22 \n" // R
+ "vmlal.u8 q9, d7, d22 \n"
+ "vaddhn.u16 d0, q8, q12 \n" // 16 bit to 8 bit Y
+ "vaddhn.u16 d1, q9, q12 \n"
+ "vst1.8 {d0, d1}, [%1]! \n" // store 16 pixels Y.
+ "bgt 1b \n"
+ : "+r"(src_rgb), // %0
+ "+r"(dst_y), // %1
+ "+r"(width) // %2
+ : "r"(rgbconstants) // %3
+ : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "d20", "d21", "d22",
+ "q12");
}
void RGB24ToYJRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
- asm volatile(
- "vmov.u8 d4, #29 \n" // B * 0.1140 coefficient
- "vmov.u8 d5, #150 \n" // G * 0.5870 coefficient
- "vmov.u8 d6, #77 \n" // R * 0.2990 coefficient
- "1: \n"
- "vld3.8 {d0, d1, d2}, [%0]! \n" // load 8 pixels of RGB24.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q4, d0, d4 \n" // B
- "vmlal.u8 q4, d1, d5 \n" // G
- "vmlal.u8 q4, d2, d6 \n" // R
- "vqrshrn.u16 d0, q4, #8 \n" // 16 bit to 8 bit Y
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
- "bgt 1b \n"
- : "+r"(src_rgb24), // %0
- "+r"(dst_yj), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "q4");
+ RGBToYMatrixRow_NEON(src_rgb24, dst_yj, width, &kRgb24JPEGConstants);
}
void RAWToYJRow_NEON(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
- asm volatile(
- "vmov.u8 d6, #29 \n" // B * 0.1140 coefficient
- "vmov.u8 d5, #150 \n" // G * 0.5870 coefficient
- "vmov.u8 d4, #77 \n" // R * 0.2990 coefficient
- "1: \n"
- "vld3.8 {d0, d1, d2}, [%0]! \n" // load 8 pixels of RAW.
- "subs %2, %2, #8 \n" // 8 processed per loop.
- "vmull.u8 q4, d0, d4 \n" // R
- "vmlal.u8 q4, d1, d5 \n" // G
- "vmlal.u8 q4, d2, d6 \n" // B
- "vqrshrn.u16 d0, q4, #8 \n" // 16 bit to 8 bit Y
- "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y.
- "bgt 1b \n"
- : "+r"(src_raw), // %0
- "+r"(dst_yj), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "q4");
+ RGBToYMatrixRow_NEON(src_raw, dst_yj, width, &kRawJPEGConstants);
+}
+
+void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
+ RGBToYMatrixRow_NEON(src_rgb24, dst_y, width, &kRgb24I601Constants);
+}
+
+void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+ RGBToYMatrixRow_NEON(src_raw, dst_y, width, &kRawI601Constants);
}
// Bilinear filter 16x2 -> 16x1
diff --git a/source/row_neon64.cc b/source/row_neon64.cc
index 7139ead7..ad355f2a 100644
--- a/source/row_neon64.cc
+++ b/source/row_neon64.cc
@@ -2021,30 +2021,6 @@ void AB64ToARGBRow_NEON(const uint16_t* src_ab64,
: "cc", "memory", "v0", "v1", "v2", "v3", "v4");
}
-void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
- asm volatile(
- "movi v4.8b, #25 \n" // B * 0.1016 coefficient
- "movi v5.8b, #129 \n" // G * 0.5078 coefficient
- "movi v6.8b, #66 \n" // R * 0.2578 coefficient
- "movi v7.8b, #16 \n" // Add 16 constant
- "1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v3.8h, v0.8b, v4.8b \n" // B
- "prfm pldl1keep, [%0, 448] \n"
- "umlal v3.8h, v1.8b, v5.8b \n" // G
- "umlal v3.8h, v2.8b, v6.8b \n" // R
- "uqrshrn v0.8b, v3.8h, #8 \n" // 16 bit to 8 bit Y
- "uqadd v0.8b, v0.8b, v7.8b \n"
- "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y.
- "b.gt 1b \n"
- : "+r"(src_argb), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
-}
-
void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
uint8_t* dst_a,
int width) {
@@ -2063,50 +2039,6 @@ void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
);
}
-void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
- asm volatile(
- "movi v4.8b, #29 \n" // B * 0.1140 coefficient
- "movi v5.8b, #150 \n" // G * 0.5870 coefficient
- "movi v6.8b, #77 \n" // R * 0.2990 coefficient
- "1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v3.8h, v0.8b, v4.8b \n" // B
- "prfm pldl1keep, [%0, 448] \n"
- "umlal v3.8h, v1.8b, v5.8b \n" // G
- "umlal v3.8h, v2.8b, v6.8b \n" // R
- "uqrshrn v0.8b, v3.8h, #8 \n" // 16 bit to 8 bit Y
- "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y.
- "b.gt 1b \n"
- : "+r"(src_argb), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6");
-}
-
-void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
- asm volatile(
- "movi v4.8b, #29 \n" // B * 0.1140 coefficient
- "movi v5.8b, #150 \n" // G * 0.5870 coefficient
- "movi v6.8b, #77 \n" // R * 0.2990 coefficient
- "1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 RGBA
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v0.8h, v1.8b, v4.8b \n" // B
- "prfm pldl1keep, [%0, 448] \n"
- "umlal v0.8h, v2.8b, v5.8b \n" // G
- "umlal v0.8h, v3.8b, v6.8b \n" // R
- "uqrshrn v3.8b, v0.8h, #8 \n" // 16 bit to 8 bit Y
- "st1 {v3.8b}, [%1], #8 \n" // store 8 pixels Y.
- "b.gt 1b \n"
- : "+r"(src_rgba), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6");
-}
-
// 8x1 pixels.
void ARGBToUV444Row_NEON(const uint8_t* src_argb,
uint8_t* dst_u,
@@ -2124,18 +2056,16 @@ void ARGBToUV444Row_NEON(const uint8_t* src_argb,
"ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB
"subs %w3, %w3, #8 \n" // 8 processed per loop.
"umull v4.8h, v0.8b, v24.8b \n" // B
- "prfm pldl1keep, [%0, 448] \n"
"umlsl v4.8h, v1.8b, v25.8b \n" // G
"umlsl v4.8h, v2.8b, v26.8b \n" // R
- "add v4.8h, v4.8h, v29.8h \n" // +128 -> unsigned
+ "prfm pldl1keep, [%0, 448] \n"
"umull v3.8h, v2.8b, v24.8b \n" // R
"umlsl v3.8h, v1.8b, v28.8b \n" // G
"umlsl v3.8h, v0.8b, v27.8b \n" // B
- "add v3.8h, v3.8h, v29.8h \n" // +128 -> unsigned
- "uqshrn v0.8b, v4.8h, #8 \n" // 16 bit to 8 bit U
- "uqshrn v1.8b, v3.8h, #8 \n" // 16 bit to 8 bit V
+ "addhn v0.8b, v4.8h, v29.8h \n" // +128 -> unsigned
+ "addhn v1.8b, v3.8h, v29.8h \n" // +128 -> unsigned
"st1 {v0.8b}, [%1], #8 \n" // store 8 pixels U.
"st1 {v1.8b}, [%2], #8 \n" // store 8 pixels V.
@@ -2166,10 +2096,8 @@ void ARGBToUV444Row_NEON(const uint8_t* src_argb,
"mls v4.8h, " #QG ",v24.8h \n" /* G */ \
"mls v3.8h, " #QR ",v22.8h \n" /* R */ \
"mls v4.8h, " #QB ",v23.8h \n" /* B */ \
- "add v3.8h, v3.8h, v25.8h \n" /* +128 -> unsigned */ \
- "add v4.8h, v4.8h, v25.8h \n" /* +128 -> unsigned */ \
- "uqshrn v0.8b, v3.8h, #8 \n" /* 16 bit to 8 bit U */ \
- "uqshrn v1.8b, v4.8h, #8 \n" /* 16 bit to 8 bit V */
+ "addhn v0.8b, v3.8h, v25.8h \n" /* +128 -> unsigned */ \
+ "addhn v1.8b, v4.8h, v25.8h \n" /* +128 -> unsigned */
// clang-format on
// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
@@ -2807,168 +2735,169 @@ void ARGB4444ToYRow_NEON(const uint8_t* src_argb4444,
: "cc", "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27");
}
-void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+struct RgbConstants {
+ uint8_t kRGBToY[4];
+ uint16_t kAddY;
+ uint16_t pad;
+};
+
+// RGB to JPeg coefficients
+// B * 0.1140 coefficient = 29
+// G * 0.5870 coefficient = 150
+// R * 0.2990 coefficient = 77
+// Add 0.5 = 0x80
+struct RgbConstants kRgb24JPEGConstants = {{29, 150, 77, 0}, 128};
+
+struct RgbConstants kRawJPEGConstants = {{77, 150, 29, 0}, 128};
+
+// RGB to BT.601 coefficients
+// B * 0.1016 coefficient = 25
+// G * 0.5078 coefficient = 129
+// R * 0.2578 coefficient = 66
+// Add 16.5 = 0x1080
+
+struct RgbConstants kRgb24I601Constants = {{25, 129, 66, 0}, 0x1080};
+
+struct RgbConstants kRawI601Constants = {{66, 129, 25, 0}, 0x1080};
+
+// ARGB expects first 3 values to contain RGB and 4th value is ignored.
+void ARGBToYMatrixRow_NEON(const uint8_t* src_argb,
+ uint8_t* dst_y,
+ int width,
+ const struct RgbConstants* rgbconstants) {
asm volatile(
- "movi v4.8b, #66 \n" // R * 0.2578 coefficient
- "movi v5.8b, #129 \n" // G * 0.5078 coefficient
- "movi v6.8b, #25 \n" // B * 0.1016 coefficient
- "movi v7.8b, #16 \n" // Add 16 constant
+ "ldr d0, [%3] \n" // load rgbconstants
+ "dup v6.16b, v0.b[0] \n"
+ "dup v7.16b, v0.b[1] \n"
+ "dup v16.16b, v0.b[2] \n"
+ "dup v17.8h, v0.h[2] \n"
"1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 pixels.
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v16.8h, v1.8b, v4.8b \n" // R
+ "ld4 {v2.16b,v3.16b,v4.16b,v5.16b}, [%0], #64 \n" // load 16
+ // pixels.
+ "subs %w2, %w2, #16 \n" // 16 processed per loop.
+ "umull v0.8h, v2.8b, v6.8b \n" // B
+ "umull2 v1.8h, v2.16b, v6.16b \n"
"prfm pldl1keep, [%0, 448] \n"
- "umlal v16.8h, v2.8b, v5.8b \n" // G
- "umlal v16.8h, v3.8b, v6.8b \n" // B
- "uqrshrn v0.8b, v16.8h, #8 \n" // 16 bit to 8 bit Y
- "uqadd v0.8b, v0.8b, v7.8b \n"
- "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y.
+ "umlal v0.8h, v3.8b, v7.8b \n" // G
+ "umlal2 v1.8h, v3.16b, v7.16b \n"
+ "umlal v0.8h, v4.8b, v16.8b \n" // R
+ "umlal2 v1.8h, v4.16b, v16.16b \n"
+ "addhn v0.8b, v0.8h, v17.8h \n" // 16 bit to 8 bit Y
+ "addhn v1.8b, v1.8h, v17.8h \n"
+ "st1 {v0.8b, v1.8b}, [%1], #16 \n" // store 16 pixels Y.
"b.gt 1b \n"
- : "+r"(src_bgra), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
+ : "+r"(src_argb), // %0
+ "+r"(dst_y), // %1
+ "+r"(width) // %2
+ : "r"(rgbconstants) // %3
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+ "v17");
+}
+
+void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+ ARGBToYMatrixRow_NEON(src_argb, dst_y, width, &kRgb24I601Constants);
+}
+
+void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_yj, int width) {
+ ARGBToYMatrixRow_NEON(src_argb, dst_yj, width, &kRgb24JPEGConstants);
}
void ABGRToYRow_NEON(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+ ARGBToYMatrixRow_NEON(src_abgr, dst_y, width, &kRawI601Constants);
+}
+
+// RGBA expects first value to be A and ignored, then 3 values to contain RGB.
+// Same code as ARGB, except the LD4
+void RGBAToYMatrixRow_NEON(const uint8_t* src_rgba,
+ uint8_t* dst_y,
+ int width,
+ const struct RgbConstants* rgbconstants) {
asm volatile(
- "movi v6.8b, #25 \n" // B * 0.1016 coefficient
- "movi v5.8b, #129 \n" // G * 0.5078 coefficient
- "movi v4.8b, #66 \n" // R * 0.2578 coefficient
- "movi v7.8b, #16 \n" // Add 16 constant
+ "ldr d0, [%3] \n" // load rgbconstants
+ "dup v6.16b, v0.b[0] \n"
+ "dup v7.16b, v0.b[1] \n"
+ "dup v16.16b, v0.b[2] \n"
+ "dup v17.8h, v0.h[2] \n"
"1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 pixels.
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v16.8h, v0.8b, v4.8b \n" // R
+ "ld4 {v1.16b,v2.16b,v3.16b,v4.16b}, [%0], #64 \n" // load 16
+ // pixels.
+ "subs %w2, %w2, #16 \n" // 16 processed per loop.
+ "umull v0.8h, v2.8b, v6.8b \n" // B
+ "umull2 v1.8h, v2.16b, v6.16b \n"
"prfm pldl1keep, [%0, 448] \n"
- "umlal v16.8h, v1.8b, v5.8b \n" // G
- "umlal v16.8h, v2.8b, v6.8b \n" // B
- "uqrshrn v0.8b, v16.8h, #8 \n" // 16 bit to 8 bit Y
- "uqadd v0.8b, v0.8b, v7.8b \n"
- "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y.
- "b.gt 1b \n"
- : "+r"(src_abgr), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
+ "umlal v0.8h, v3.8b, v7.8b \n" // G
+ "umlal2 v1.8h, v3.16b, v7.16b \n"
+ "umlal v0.8h, v4.8b, v16.8b \n" // R
+ "umlal2 v1.8h, v4.16b, v16.16b \n"
+ "addhn v0.8b, v0.8h, v17.8h \n" // 16 bit to 8 bit Y
+ "addhn v1.8b, v1.8h, v17.8h \n"
+ "st1 {v0.8b, v1.8b}, [%1], #16 \n" // store 16 pixels Y.
+ "b.gt 1b \n"
+ : "+r"(src_rgba), // %0
+ "+r"(dst_y), // %1
+ "+r"(width) // %2
+ : "r"(rgbconstants) // %3
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+ "v17");
}
void RGBAToYRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
- asm volatile(
- "movi v4.8b, #25 \n" // B * 0.1016 coefficient
- "movi v5.8b, #129 \n" // G * 0.5078 coefficient
- "movi v6.8b, #66 \n" // R * 0.2578 coefficient
- "movi v7.8b, #16 \n" // Add 16 constant
- "1: \n"
- "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 pixels.
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v16.8h, v1.8b, v4.8b \n" // B
- "prfm pldl1keep, [%0, 448] \n"
- "umlal v16.8h, v2.8b, v5.8b \n" // G
- "umlal v16.8h, v3.8b, v6.8b \n" // R
- "uqrshrn v0.8b, v16.8h, #8 \n" // 16 bit to 8 bit Y
- "uqadd v0.8b, v0.8b, v7.8b \n"
- "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y.
- "b.gt 1b \n"
- : "+r"(src_rgba), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
+ RGBAToYMatrixRow_NEON(src_rgba, dst_y, width, &kRgb24I601Constants);
}
-void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
- asm volatile(
- "movi v4.8b, #25 \n" // B * 0.1016 coefficient
- "movi v5.8b, #129 \n" // G * 0.5078 coefficient
- "movi v6.8b, #66 \n" // R * 0.2578 coefficient
- "movi v7.8b, #16 \n" // Add 16 constant
- "1: \n"
- "ld3 {v0.8b,v1.8b,v2.8b}, [%0], #24 \n" // load 8 pixels.
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v16.8h, v0.8b, v4.8b \n" // B
- "prfm pldl1keep, [%0, 448] \n"
- "umlal v16.8h, v1.8b, v5.8b \n" // G
- "umlal v16.8h, v2.8b, v6.8b \n" // R
- "uqrshrn v0.8b, v16.8h, #8 \n" // 16 bit to 8 bit Y
- "uqadd v0.8b, v0.8b, v7.8b \n"
- "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y.
- "b.gt 1b \n"
- : "+r"(src_rgb24), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
+void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_yj, int width) {
+ RGBAToYMatrixRow_NEON(src_rgba, dst_yj, width, &kRgb24JPEGConstants);
}
-void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width) {
- asm volatile(
- "movi v6.8b, #25 \n" // B * 0.1016 coefficient
- "movi v5.8b, #129 \n" // G * 0.5078 coefficient
- "movi v4.8b, #66 \n" // R * 0.2578 coefficient
- "movi v7.8b, #16 \n" // Add 16 constant
+void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+ RGBAToYMatrixRow_NEON(src_bgra, dst_y, width, &kRawI601Constants);
+}
+
+void RGBToYMatrixRow_NEON(const uint8_t* src_rgb,
+ uint8_t* dst_y,
+ int width,
+ const struct RgbConstants* rgbconstants) {
+ asm volatile(
+ "ldr d0, [%3] \n" // load rgbconstants
+ "dup v5.16b, v0.b[0] \n"
+ "dup v6.16b, v0.b[1] \n"
+ "dup v7.16b, v0.b[2] \n"
+ "dup v16.8h, v0.h[2] \n"
"1: \n"
- "ld3 {v0.8b,v1.8b,v2.8b}, [%0], #24 \n" // load 8 pixels.
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v16.8h, v0.8b, v4.8b \n" // B
+ "ld3 {v2.16b,v3.16b,v4.16b}, [%0], #48 \n" // load 16 pixels.
+ "subs %w2, %w2, #16 \n" // 16 processed per loop.
+ "umull v0.8h, v2.8b, v5.8b \n" // B
+ "umull2 v1.8h, v2.16b, v5.16b \n"
"prfm pldl1keep, [%0, 448] \n"
- "umlal v16.8h, v1.8b, v5.8b \n" // G
- "umlal v16.8h, v2.8b, v6.8b \n" // R
- "uqrshrn v0.8b, v16.8h, #8 \n" // 16 bit to 8 bit Y
- "uqadd v0.8b, v0.8b, v7.8b \n"
- "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y.
- "b.gt 1b \n"
- : "+r"(src_raw), // %0
- "+r"(dst_y), // %1
- "+r"(width) // %2
- :
+ "umlal v0.8h, v3.8b, v6.8b \n" // G
+ "umlal2 v1.8h, v3.16b, v6.16b \n"
+ "umlal v0.8h, v4.8b, v7.8b \n" // R
+ "umlal2 v1.8h, v4.16b, v7.16b \n"
+ "addhn v0.8b, v0.8h, v16.8h \n" // 16 bit to 8 bit Y
+ "addhn v1.8b, v1.8h, v16.8h \n"
+ "st1 {v0.8b, v1.8b}, [%1], #16 \n" // store 16 pixels Y.
+ "b.gt 1b \n"
+ : "+r"(src_rgb), // %0
+ "+r"(dst_y), // %1
+ "+r"(width) // %2
+ : "r"(rgbconstants) // %3
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
}
void RGB24ToYJRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
- asm volatile(
- "movi v4.8b, #29 \n" // B * 0.1140 coefficient
- "movi v5.8b, #150 \n" // G * 0.5870 coefficient
- "movi v6.8b, #77 \n" // R * 0.2990 coefficient
- "1: \n"
- "ld3 {v0.8b,v1.8b,v2.8b}, [%0], #24 \n" // load 8 pixels.
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v0.8h, v0.8b, v4.8b \n" // B
- "prfm pldl1keep, [%0, 448] \n"
- "umlal v0.8h, v1.8b, v5.8b \n" // G
- "umlal v0.8h, v2.8b, v6.8b \n" // R
- "uqrshrn v0.8b, v0.8h, #8 \n" // 16 bit to 8 bit Y
- "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y.
- "b.gt 1b \n"
- : "+r"(src_rgb24), // %0
- "+r"(dst_yj), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6");
+ RGBToYMatrixRow_NEON(src_rgb24, dst_yj, width, &kRgb24JPEGConstants);
}
void RAWToYJRow_NEON(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
- asm volatile(
- "movi v6.8b, #29 \n" // B * 0.1140 coefficient
- "movi v5.8b, #150 \n" // G * 0.5870 coefficient
- "movi v4.8b, #77 \n" // R * 0.2990 coefficient
- "1: \n"
- "ld3 {v0.8b,v1.8b,v2.8b}, [%0], #24 \n" // load 8 pixels.
- "subs %w2, %w2, #8 \n" // 8 processed per loop.
- "umull v0.8h, v0.8b, v4.8b \n" // B
- "prfm pldl1keep, [%0, 448] \n"
- "umlal v0.8h, v1.8b, v5.8b \n" // G
- "umlal v0.8h, v2.8b, v6.8b \n" // R
- "uqrshrn v0.8b, v0.8h, #8 \n" // 16 bit to 8 bit Y
- "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y.
- "b.gt 1b \n"
- : "+r"(src_raw), // %0
- "+r"(dst_yj), // %1
- "+r"(width) // %2
- :
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6");
+ RGBToYMatrixRow_NEON(src_raw, dst_yj, width, &kRawJPEGConstants);
+}
+
+void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
+ RGBToYMatrixRow_NEON(src_rgb24, dst_y, width, &kRgb24I601Constants);
+}
+
+void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+ RGBToYMatrixRow_NEON(src_raw, dst_y, width, &kRawI601Constants);
}
// Bilinear filter 16x2 -> 16x1
diff --git a/unit_test/convert_test.cc b/unit_test/convert_test.cc
index de90f660..185c5aa4 100644
--- a/unit_test/convert_test.cc
+++ b/unit_test/convert_test.cc
@@ -1314,7 +1314,7 @@ TESTATOBIPLANAR(ARGB, 1, 4, NV12, 2, 2)
TESTATOBIPLANAR(ARGB, 1, 4, NV21, 2, 2)
TESTATOBIPLANAR(ABGR, 1, 4, NV12, 2, 2)
TESTATOBIPLANAR(ABGR, 1, 4, NV21, 2, 2)
-TESTATOBIPLANAR(RAW, 1, 3, JNV21, 2, 2)
+TESTATOBIPLANAR(RAW, 1, 3, JNV21, 2, 2)
TESTATOBIPLANAR(YUY2, 2, 4, NV12, 2, 2)
TESTATOBIPLANAR(UYVY, 2, 4, NV12, 2, 2)
TESTATOBIPLANAR(AYUV, 1, 4, NV12, 2, 2)
@@ -3813,14 +3813,18 @@ TESTQPLANAR16TOB(I210Alpha, 2, 1, ARGBFilter, 4, 4, 1, 10)
#define P216ToAR30(a, b, c, d, e, f, g, h) \
P216ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
-#define P010ToARGBFilter(a, b, c, d, e, f, g, h) \
- P010ToARGBMatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, kFilterBilinear)
-#define P210ToARGBFilter(a, b, c, d, e, f, g, h) \
- P210ToARGBMatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, kFilterBilinear)
-#define P010ToAR30Filter(a, b, c, d, e, f, g, h) \
- P010ToAR30MatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, kFilterBilinear)
-#define P210ToAR30Filter(a, b, c, d, e, f, g, h) \
- P210ToAR30MatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, kFilterBilinear)
+#define P010ToARGBFilter(a, b, c, d, e, f, g, h) \
+ P010ToARGBMatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+ kFilterBilinear)
+#define P210ToARGBFilter(a, b, c, d, e, f, g, h) \
+ P210ToARGBMatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+ kFilterBilinear)
+#define P010ToAR30Filter(a, b, c, d, e, f, g, h) \
+ P010ToAR30MatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+ kFilterBilinear)
+#define P210ToAR30Filter(a, b, c, d, e, f, g, h) \
+ P210ToAR30MatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+ kFilterBilinear)
#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
TESTBIPLANAR16TOB(P010, 2, 2, ARGB, 4, 4, 1, 10)
generated by cgit v1.2.3 (git 2.25.1) at 2024-07-08 08:25:46 +0000