coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2025-12-06 08:46:47 +08:00
Code Issues Packages Projects Releases Wiki Activity

[AArch64] Add SME implementation of I444ToARGBRow

Browse Source 
This is based on an unrolled version of the existing SVE2 code. The
implementation in this case is a pure streaming-SVE (SSVE)
implementation based on the existing SVE2 implementation, we do not use
the ZA tile.

Change-Id: I83d8e58aafd814125b3446fb1c9ec4a5fb56fe3e
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5913882
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Justin Green <greenjustin@google.com>
This commit is contained in:
George Steed 2024-08-22 13:56:00 +01:00 committed by Frank Barchard
parent 22c5c18778
commit 237f39cb8c
3 changed files with 109 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
include/libyuv/row.h
View File

@ -573,6 +573,7 @@ extern "C" {
#if !defined(LIBYUV_DISABLE_SME) && defined(CLANG_HAS_SME) && \
defined(__aarch64__)
#define HAS_I422TOARGBROW_SME
#define HAS_I444TOARGBROW_SME
#endif
// The following are available on AArch64 platforms:
@ -1072,6 +1073,12 @@ void I444ToARGBRow_SVE2(const uint8_t* src_y,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width);
void I444ToARGBRow_SME(const uint8_t* src_y,
const uint8_t* src_u,
const uint8_t* src_v,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width);
void I444ToRGB24Row_NEON(const uint8_t* src_y,
const uint8_t* src_u,
const uint8_t* src_v,

15
source/convert_argb.cc
View File

@ -631,6 +631,11 @@ int I444ToARGBMatrix(const uint8_t* src_y,
I444ToARGBRow = I444ToARGBRow_SVE2;
}
#endif
#if defined(HAS_I444TOARGBROW_SME)
if (TestCpuFlag(kCpuHasSME)) {
I444ToARGBRow = I444ToARGBRow_SME;
}
#endif
#if defined(HAS_I444TOARGBROW_MSA)
if (TestCpuFlag(kCpuHasMSA)) {
I444ToARGBRow = I444ToARGBRow_Any_MSA;
@ -6482,6 +6487,11 @@ static int I420ToARGBMatrixBilinear(const uint8_t* src_y,
I444ToARGBRow = I444ToARGBRow_SVE2;
}
#endif
#if defined(HAS_I444TOARGBROW_SME)
if (TestCpuFlag(kCpuHasSME)) {
I444ToARGBRow = I444ToARGBRow_SME;
}
#endif
#if defined(HAS_I444TOARGBROW_MSA)
if (TestCpuFlag(kCpuHasMSA)) {
I444ToARGBRow = I444ToARGBRow_Any_MSA;
@ -6634,6 +6644,11 @@ static int I422ToARGBMatrixLinear(const uint8_t* src_y,
I444ToARGBRow = I444ToARGBRow_SVE2;
}
#endif
#if defined(HAS_I444TOARGBROW_SME)
if (TestCpuFlag(kCpuHasSME)) {
I444ToARGBRow = I444ToARGBRow_SME;
}
#endif
#if defined(HAS_I444TOARGBROW_MSA)
if (TestCpuFlag(kCpuHasMSA)) {
I444ToARGBRow = I444ToARGBRow_Any_MSA;

87
source/row_sme.cc
View File

@ -44,6 +44,21 @@ extern "C" {
"trn2 z1.b, z0.b, z0.b \n" \
"trn1 z0.b, z0.b, z0.b \n"
// Read twice as much data from YUV, putting the even elements from the Y data
// in z0.h and odd elements in z1.h.
#define READYUV444_SVE_2X \
"ld1b {z0.b}, p1/z, [%[src_y]] \n" \
"ld1b {z2.b}, p1/z, [%[src_u]] \n" \
"ld1b {z3.b}, p1/z, [%[src_v]] \n" \
"incb %[src_y] \n" \
"incb %[src_u] \n" \
"incb %[src_v] \n" \
"prfm pldl1keep, [%[src_y], 448] \n" \
"prfm pldl1keep, [%[src_u], 128] \n" \
"prfm pldl1keep, [%[src_v], 128] \n" \
"trn2 z1.b, z0.b, z0.b \n" \
"trn1 z0.b, z0.b, z0.b \n"
// The U/V component multiplies do not need to be duplicated in I422, we just
// need to combine them with Y0/Y1 correctly.
#define I422TORGB_SVE_2X \
@ -67,6 +82,30 @@ extern "C" {
"uqsub z18.h, z18.h, z27.h \n" /* R0 */ \
"uqsub z22.h, z22.h, z27.h \n" /* R1 */
#define I444TORGB_SVE_2X \
"umulh z0.h, z24.h, z0.h \n" /* Y0 */ \
"umulh z1.h, z24.h, z1.h \n" /* Y1 */ \
"umullb z6.h, z30.b, z2.b \n" \
"umullt z7.h, z30.b, z2.b \n" \
"umullb z4.h, z28.b, z2.b \n" /* DB */ \
"umullt z2.h, z28.b, z2.b \n" /* DB */ \
"umlalb z6.h, z31.b, z3.b \n" /* DG */ \
"umlalt z7.h, z31.b, z3.b \n" /* DG */ \
"umullb z5.h, z29.b, z3.b \n" /* DR */ \
"umullt z3.h, z29.b, z3.b \n" /* DR */ \
"add z17.h, z0.h, z26.h \n" /* G */ \
"add z21.h, z1.h, z26.h \n" /* G */ \
"add z16.h, z0.h, z4.h \n" /* B */ \
"add z20.h, z1.h, z2.h \n" /* B */ \
"add z18.h, z0.h, z5.h \n" /* R */ \
"add z22.h, z1.h, z3.h \n" /* R */ \
"uqsub z17.h, z17.h, z6.h \n" /* G */ \
"uqsub z21.h, z21.h, z7.h \n" /* G */ \
"uqsub z16.h, z16.h, z25.h \n" /* B */ \
"uqsub z20.h, z20.h, z25.h \n" /* B */ \
"uqsub z18.h, z18.h, z27.h \n" /* R */ \
"uqsub z22.h, z22.h, z27.h \n" /* R */
#define RGBTOARGB8_SVE_2X \
/* Inputs: B: z16.h, G: z17.h, R: z18.h, A: z19.b */ \
"uqshrnb z16.b, z16.h, #6 \n" /* B0 */ \
@ -81,6 +120,54 @@ extern "C" {
"z20", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", \
"z31", "p0", "p1", "p2", "p3"
__arm_locally_streaming void I444ToARGBRow_SME(
const uint8_t* src_y,
const uint8_t* src_u,
const uint8_t* src_v,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
// Streaming-SVE only, no use of ZA tile.
uint64_t vl;
asm volatile(
"cntb %[vl] \n"
"ptrue p0.b \n" //
YUVTORGB_SVE_SETUP
"dup z19.b, #255 \n" // A
"subs %w[width], %w[width], %w[vl] \n"
"b.lt 2f \n"
// Run bulk of computation with an all-true predicate to avoid predicate
// generation overhead.
"ptrue p1.b \n"
"1: \n" //
READYUV444_SVE_2X I444TORGB_SVE_2X RGBTOARGB8_SVE_2X
"subs %w[width], %w[width], %w[vl] \n"
"st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"
"incb %[dst_argb], all, mul #4 \n"
"b.ge 1b \n"
"2: \n"
"adds %w[width], %w[width], %w[vl] \n"
"b.eq 99f \n"
// Calculate a predicate for the final iteration to deal with the tail.
"whilelt p1.b, wzr, %w[width] \n" //
READYUV444_SVE_2X I444TORGB_SVE_2X RGBTOARGB8_SVE_2X
"st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"
"99: \n"
: [src_y] "+r"(src_y), // %[src_y]
[src_u] "+r"(src_u), // %[src_u]
[src_v] "+r"(src_v), // %[src_v]
[dst_argb] "+r"(dst_argb), // %[dst_argb]
[width] "+r"(width), // %[width]
[vl] "=&r"(vl) // %[vl]
: [kUVCoeff] "r"(&yuvconstants->kUVCoeff), // %[kUVCoeff]
[kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias) // %[kRGBCoeffBias]
: "cc", "memory", YUVTORGB_SVE_REGS);
}
__arm_locally_streaming void I422ToARGBRow_SME(
const uint8_t* src_y,
const uint8_t* src_u,