Skip to content

F
ffmpeg.wasm-core
Commit 6532a1a8 authored by Michael Niedermayer's avatar Michael Niedermayer
Browse files
Options

sws/x86: split mmxext fast bilinear scaler out 

Signed-off-by: 's avatarMichael Niedermayer <michaelni@gmx.at>
parent e9f7c7ae
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 389 additions and 363 deletions
libswscale/swscale_internal.h
View file @ 6532a1a8
......@@ -864,6 +864,15 @@ void ff_hyscale_fast_c(SwsContext *c, int16_t *dst, int dstWidth,
void ff_hcscale_fast_c(SwsContext *c, int16_t *dst1, int16_t *dst2,
int dstWidth, const uint8_t *src1,
const uint8_t *src2, int srcW, int xInc);
int ff_init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
int16_t *filter, int32_t *filterPos,
int numSplits);
void ff_hyscale_fast_mmxext(SwsContext *c, int16_t *dst,
int dstWidth, const uint8_t *src,
int srcW, int xInc);
void ff_hcscale_fast_mmxext(SwsContext *c, int16_t *dst1, int16_t *dst2,
int dstWidth, const uint8_t *src1,
const uint8_t *src2, int srcW, int xInc);
static inline void fillPlane16(uint8_t *plane, int stride, int width, int height, int y,
int alpha, int bits, const int big_endian)
......
libswscale/utils.c
View file @ 6532a1a8
......@@ -681,172 +681,6 @@ fail:
return ret;
}
#if HAVE_MMXEXT_INLINE
static av_cold int init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
int16_t *filter, int32_t *filterPos,
int numSplits)
{
uint8_t *fragmentA;
x86_reg imm8OfPShufW1A;
x86_reg imm8OfPShufW2A;
x86_reg fragmentLengthA;
uint8_t *fragmentB;
x86_reg imm8OfPShufW1B;
x86_reg imm8OfPShufW2B;
x86_reg fragmentLengthB;
int fragmentPos;
int xpos, i;
// create an optimized horizontal scaling routine
/* This scaler is made of runtime-generated MMXEXT code using specially tuned
* pshufw instructions. For every four output pixels, if four input pixels
* are enough for the fast bilinear scaling, then a chunk of fragmentB is
* used. If five input pixels are needed, then a chunk of fragmentA is used.
*/
// code fragment
__asm__ volatile (
"jmp 9f \n\t"
// Begin
"0: \n\t"
"movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
"movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
"movd 1(%%"REG_c", %%"REG_S"), %%mm1 \n\t"
"punpcklbw %%mm7, %%mm1 \n\t"
"punpcklbw %%mm7, %%mm0 \n\t"
"pshufw $0xFF, %%mm1, %%mm1 \n\t"
"1: \n\t"
"pshufw $0xFF, %%mm0, %%mm0 \n\t"
"2: \n\t"
"psubw %%mm1, %%mm0 \n\t"
"movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
"pmullw %%mm3, %%mm0 \n\t"
"psllw $7, %%mm1 \n\t"
"paddw %%mm1, %%mm0 \n\t"
"movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
"add $8, %%"REG_a" \n\t"
// End
"9: \n\t"
// "int $3 \n\t"
"lea " LOCAL_MANGLE(0b) ", %0 \n\t"
"lea " LOCAL_MANGLE(1b) ", %1 \n\t"
"lea " LOCAL_MANGLE(2b) ", %2 \n\t"
"dec %1 \n\t"
"dec %2 \n\t"
"sub %0, %1 \n\t"
"sub %0, %2 \n\t"
"lea " LOCAL_MANGLE(9b) ", %3 \n\t"
"sub %0, %3 \n\t"
: "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
"=r" (fragmentLengthA)
);
__asm__ volatile (
"jmp 9f \n\t"
// Begin
"0: \n\t"
"movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
"movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
"punpcklbw %%mm7, %%mm0 \n\t"
"pshufw $0xFF, %%mm0, %%mm1 \n\t"
"1: \n\t"
"pshufw $0xFF, %%mm0, %%mm0 \n\t"
"2: \n\t"
"psubw %%mm1, %%mm0 \n\t"
"movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
"pmullw %%mm3, %%mm0 \n\t"
"psllw $7, %%mm1 \n\t"
"paddw %%mm1, %%mm0 \n\t"
"movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
"add $8, %%"REG_a" \n\t"
// End
"9: \n\t"
// "int $3 \n\t"
"lea " LOCAL_MANGLE(0b) ", %0 \n\t"
"lea " LOCAL_MANGLE(1b) ", %1 \n\t"
"lea " LOCAL_MANGLE(2b) ", %2 \n\t"
"dec %1 \n\t"
"dec %2 \n\t"
"sub %0, %1 \n\t"
"sub %0, %2 \n\t"
"lea " LOCAL_MANGLE(9b) ", %3 \n\t"
"sub %0, %3 \n\t"
: "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
"=r" (fragmentLengthB)
);
xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
fragmentPos = 0;
for (i = 0; i < dstW / numSplits; i++) {
int xx = xpos >> 16;
if ((i & 3) == 0) {
int a = 0;
int b = ((xpos + xInc) >> 16) - xx;
int c = ((xpos + xInc * 2) >> 16) - xx;
int d = ((xpos + xInc * 3) >> 16) - xx;
int inc = (d + 1 < 4);
uint8_t *fragment = inc ? fragmentB : fragmentA;
x86_reg imm8OfPShufW1 = inc ? imm8OfPShufW1B : imm8OfPShufW1A;
x86_reg imm8OfPShufW2 = inc ? imm8OfPShufW2B : imm8OfPShufW2A;
x86_reg fragmentLength = inc ? fragmentLengthB : fragmentLengthA;
int maxShift = 3 - (d + inc);
int shift = 0;
if (filterCode) {
filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
filterPos[i / 2] = xx;
memcpy(filterCode + fragmentPos, fragment, fragmentLength);
filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
((b + inc) << 2) |
((c + inc) << 4) |
((d + inc) << 6);
filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
(c << 4) |
(d << 6);
if (i + 4 - inc >= dstW)
shift = maxShift; // avoid overread
else if ((filterPos[i / 2] & 3) <= maxShift)
shift = filterPos[i / 2] & 3; // align
if (shift && i >= shift) {
filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
filterPos[i / 2] -= shift;
}
}
fragmentPos += fragmentLength;
if (filterCode)
filterCode[fragmentPos] = RET;
}
xpos += xInc;
}
if (filterCode)
filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
return fragmentPos + 1;
}
#endif /* HAVE_MMXEXT_INLINE */
static void fill_rgb2yuv_table(SwsContext *c, const int table[4], int dstRange)
{
int64_t W, V, Z, Cy, Cu, Cv;
......@@ -1400,9 +1234,9 @@ av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
#if HAVE_MMXEXT_INLINE
// can't downscale !!!
if (c->canMMXEXTBeUsed && (flags & SWS_FAST_BILINEAR)) {
c->lumMmxextFilterCodeSize = init_hscaler_mmxext(dstW, c->lumXInc, NULL,
c->lumMmxextFilterCodeSize = ff_init_hscaler_mmxext(dstW, c->lumXInc, NULL,
NULL, NULL, 8);
c->chrMmxextFilterCodeSize = init_hscaler_mmxext(c->chrDstW, c->chrXInc,
c->chrMmxextFilterCodeSize = ff_init_hscaler_mmxext(c->chrDstW, c->chrXInc,
NULL, NULL, NULL, 4);
#if USE_MMAP
......@@ -1443,9 +1277,9 @@ av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
FF_ALLOCZ_OR_GOTO(c, c->hLumFilterPos, (dstW / 2 / 8 + 8) * sizeof(int32_t), fail);
FF_ALLOCZ_OR_GOTO(c, c->hChrFilterPos, (c->chrDstW / 2 / 4 + 8) * sizeof(int32_t), fail);
init_hscaler_mmxext( dstW, c->lumXInc, c->lumMmxextFilterCode,
ff_init_hscaler_mmxext( dstW, c->lumXInc, c->lumMmxextFilterCode,
c->hLumFilter, (uint32_t*)c->hLumFilterPos, 8);
init_hscaler_mmxext(c->chrDstW, c->chrXInc, c->chrMmxextFilterCode,
ff_init_hscaler_mmxext(c->chrDstW, c->chrXInc, c->chrMmxextFilterCode,
c->hChrFilter, (uint32_t*)c->hChrFilterPos, 4);
#if USE_MMAP
......
libswscale/x86/Makefile
View file @ 6532a1a8
$(SUBDIR)x86/swscale_mmx.o: CFLAGS += $(NOREDZONE_FLAGS)
OBJS += x86/rgb2rgb.o \
OBJS += x86/hscale_fast_bilinear_simd.o \
x86/rgb2rgb.o \
x86/swscale.o \
x86/yuv2rgb.o \
......
libswscale/x86/hscale_fast_bilinear_simd.c 0 → 100644
View file @ 6532a1a8
This diff is collapsed.
libswscale/x86/swscale_template.c
View file @ 6532a1a8
......@@ -28,11 +28,6 @@
#undef MOVNTQ2
#undef PREFETCH
#if COMPILE_TEMPLATE_MMXEXT
#define PREFETCH "prefetchnta"
#else
#define PREFETCH " # nop"
#endif
#if COMPILE_TEMPLATE_MMXEXT
#define REAL_MOVNTQ(a,b) "movntq " #a ", " #b " \n\t"
......@@ -1470,191 +1465,6 @@ static void RENAME(yuv2yuyv422_1)(SwsContext *c, const int16_t *buf0,
);
}
}
#if COMPILE_TEMPLATE_MMXEXT
static void RENAME(hyscale_fast)(SwsContext *c, int16_t *dst,
int dstWidth, const uint8_t *src,
int srcW, int xInc)
{
int32_t *filterPos = c->hLumFilterPos;
int16_t *filter = c->hLumFilter;
void *mmxextFilterCode = c->lumMmxextFilterCode;
int i;
#if defined(PIC)
uint64_t ebxsave;
#endif
#if ARCH_X86_64
uint64_t retsave;
#endif
__asm__ volatile(
#if defined(PIC)
"mov %%"REG_b", %5 \n\t"
#if ARCH_X86_64
"mov -8(%%rsp), %%"REG_a" \n\t"
"mov %%"REG_a", %6 \n\t"
#endif
#else
#if ARCH_X86_64
"mov -8(%%rsp), %%"REG_a" \n\t"
"mov %%"REG_a", %5 \n\t"
#endif
#endif
"pxor %%mm7, %%mm7 \n\t"
"mov %0, %%"REG_c" \n\t"
"mov %1, %%"REG_D" \n\t"
"mov %2, %%"REG_d" \n\t"
"mov %3, %%"REG_b" \n\t"
"xor %%"REG_a", %%"REG_a" \n\t" // i
PREFETCH" (%%"REG_c") \n\t"
PREFETCH" 32(%%"REG_c") \n\t"
PREFETCH" 64(%%"REG_c") \n\t"
#if ARCH_X86_64
#define CALL_MMXEXT_FILTER_CODE \
"movl (%%"REG_b"), %%esi \n\t"\
"call *%4 \n\t"\
"movl (%%"REG_b", %%"REG_a"), %%esi \n\t"\
"add %%"REG_S", %%"REG_c" \n\t"\
"add %%"REG_a", %%"REG_D" \n\t"\
"xor %%"REG_a", %%"REG_a" \n\t"\
#else
#define CALL_MMXEXT_FILTER_CODE \
"movl (%%"REG_b"), %%esi \n\t"\
"call *%4 \n\t"\
"addl (%%"REG_b", %%"REG_a"), %%"REG_c" \n\t"\
"add %%"REG_a", %%"REG_D" \n\t"\
"xor %%"REG_a", %%"REG_a" \n\t"\
#endif /* ARCH_X86_64 */
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
#if defined(PIC)
"mov %5, %%"REG_b" \n\t"
#if ARCH_X86_64
"mov %6, %%"REG_a" \n\t"
"mov %%"REG_a", -8(%%rsp) \n\t"
#endif
#else
#if ARCH_X86_64
"mov %5, %%"REG_a" \n\t"
"mov %%"REG_a", -8(%%rsp) \n\t"
#endif
#endif
:: "m" (src), "m" (dst), "m" (filter), "m" (filterPos),
"m" (mmxextFilterCode)
#if defined(PIC)
,"m" (ebxsave)
#endif
#if ARCH_X86_64
,"m"(retsave)
#endif
: "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
#if !defined(PIC)
,"%"REG_b
#endif
);
for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--)
dst[i] = src[srcW-1]*128;
}
static void RENAME(hcscale_fast)(SwsContext *c, int16_t *dst1, int16_t *dst2,
int dstWidth, const uint8_t *src1,
const uint8_t *src2, int srcW, int xInc)
{
int32_t *filterPos = c->hChrFilterPos;
int16_t *filter = c->hChrFilter;
void *mmxextFilterCode = c->chrMmxextFilterCode;
int i;
#if defined(PIC)
DECLARE_ALIGNED(8, uint64_t, ebxsave);
#endif
#if ARCH_X86_64
DECLARE_ALIGNED(8, uint64_t, retsave);
#endif
__asm__ volatile(
#if defined(PIC)
"mov %%"REG_b", %7 \n\t"
#if ARCH_X86_64
"mov -8(%%rsp), %%"REG_a" \n\t"
"mov %%"REG_a", %8 \n\t"
#endif
#else
#if ARCH_X86_64
"mov -8(%%rsp), %%"REG_a" \n\t"
"mov %%"REG_a", %7 \n\t"
#endif
#endif
"pxor %%mm7, %%mm7 \n\t"
"mov %0, %%"REG_c" \n\t"
"mov %1, %%"REG_D" \n\t"
"mov %2, %%"REG_d" \n\t"
"mov %3, %%"REG_b" \n\t"
"xor %%"REG_a", %%"REG_a" \n\t" // i
PREFETCH" (%%"REG_c") \n\t"
PREFETCH" 32(%%"REG_c") \n\t"
PREFETCH" 64(%%"REG_c") \n\t"
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
"xor %%"REG_a", %%"REG_a" \n\t" // i
"mov %5, %%"REG_c" \n\t" // src
"mov %6, %%"REG_D" \n\t" // buf2
PREFETCH" (%%"REG_c") \n\t"
PREFETCH" 32(%%"REG_c") \n\t"
PREFETCH" 64(%%"REG_c") \n\t"
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
CALL_MMXEXT_FILTER_CODE
#if defined(PIC)
"mov %7, %%"REG_b" \n\t"
#if ARCH_X86_64
"mov %8, %%"REG_a" \n\t"
"mov %%"REG_a", -8(%%rsp) \n\t"
#endif
#else
#if ARCH_X86_64
"mov %7, %%"REG_a" \n\t"
"mov %%"REG_a", -8(%%rsp) \n\t"
#endif
#endif
:: "m" (src1), "m" (dst1), "m" (filter), "m" (filterPos),
"m" (mmxextFilterCode), "m" (src2), "m"(dst2)
#if defined(PIC)
,"m" (ebxsave)
#endif
#if ARCH_X86_64
,"m"(retsave)
#endif
: "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
#if !defined(PIC)
,"%"REG_b
#endif
);
for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--) {
dst1[i] = src1[srcW-1]*128;
dst2[i] = src2[srcW-1]*128;
}
}
#endif /* COMPILE_TEMPLATE_MMXEXT */
static av_cold void RENAME(sws_init_swscale)(SwsContext *c)
{
enum AVPixelFormat dstFormat = c->dstFormat;
......@@ -1723,8 +1533,8 @@ static av_cold void RENAME(sws_init_swscale)(SwsContext *c)
// Use the new MMX scaler if the MMXEXT one can't be used (it is faster than the x86 ASM one).
#if COMPILE_TEMPLATE_MMXEXT
if (c->flags & SWS_FAST_BILINEAR && c->canMMXEXTBeUsed) {
c->hyscale_fast = RENAME(hyscale_fast);
c->hcscale_fast = RENAME(hcscale_fast);
c->hyscale_fast = ff_hyscale_fast_mmxext;
c->hcscale_fast = ff_hcscale_fast_mmxext;
} else {
#endif /* COMPILE_TEMPLATE_MMXEXT */
c->hyscale_fast = NULL;
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment