Skip to content

V
vlc-1.1
Commit 62107e56 authored by Laurent Aimar's avatar Laurent Aimar
Browse files
Options

Improved performance when copying video surface in dxva2.

parent 106208c9
Show whitespace changes
Inline Side-by-side
Showing with 329 additions and 112 deletions
modules/codec/avcodec/dxva2.c
View file @ 62107e56
...@@ -240,6 +240,9 @@ typedef struct ...@@ -240,6 +240,9 @@ typedef struct
/* Option conversion */ /* Option conversion */
D3DFORMAT output; D3DFORMAT output;
uint8_t *surface_cache_base;
uint8_t *surface_cache;
size_t surface_cache_size;
/* */ /* */
struct dxva_context hw; struct dxva_context hw;
...@@ -282,6 +285,13 @@ static int DxResetVideoDecoder(vlc_va_dxva2_t *); ...@@ -282,6 +285,13 @@ static int DxResetVideoDecoder(vlc_va_dxva2_t *);
static void DxCreateVideoConversion(vlc_va_dxva2_t *); static void DxCreateVideoConversion(vlc_va_dxva2_t *);
static void DxDestroyVideoConversion(vlc_va_dxva2_t *); static void DxDestroyVideoConversion(vlc_va_dxva2_t *);
static void CopyFromNv12(picture_t *dst, const D3DLOCKED_RECT *src,
uint8_t *cache, size_t cache_size,
unsigned width, unsigned height);
static void CopyFromYv12(picture_t *dst, const D3DLOCKED_RECT *src,
uint8_t *cache, size_t cache_size,
unsigned width, unsigned height);
/* */ /* */
static int Setup(vlc_va_t *external, void **hw, vlc_fourcc_t *chroma, static int Setup(vlc_va_t *external, void **hw, vlc_fourcc_t *chroma,
int width, int height) int width, int height)
...@@ -330,86 +340,15 @@ ok: ...@@ -330,86 +340,15 @@ ok:
return VLC_SUCCESS; return VLC_SUCCESS;
} }
static void SplitUV(uint8_t *dstu, size_t dstu_pitch,
uint8_t *dstv, size_t dstv_pitch,
const uint8_t *src, size_t src_pitch,
unsigned w, unsigned h)
{
for (unsigned y = 0; y < h; y++) {
for (unsigned x = 0; x < w; x++) {
dstu[x] = src[2*x+0];
dstv[x] = src[2*x+1];
}
src += src_pitch;
dstu += dstu_pitch;
dstv += dstv_pitch;
}
}
/* FIXME CAN_COMPILE_SSSE3 seems undefined for me */
#if 0 && defined(CAN_COMPILE_SSSE3) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ > 0))
//#define HAVE_SPLITUV_SSSE3
static void SplitUVSSSE3(uint8_t *dstu, size_t dstu_pitch,
uint8_t *dstv, size_t dstv_pitch,
const uint8_t *src, size_t src_pitch,
unsigned w, unsigned h)
{
const uint8_t maskm[] = { 0, 2, 4, 6, 8, 10, 12, 14,
1, 3, 5, 7, 9, 11, 13, 15 };
for (unsigned y = 0; y < h; y++) {
unsigned x;
/* Prefetch the line */
for (x = 0; x < 2*w; x += 64) {
__asm__ volatile (
"prefetchnta 0(%[src])\n"
: : [src]"r"(((intptr_t)src & ~15)+ x));
}
for (x = 0; x < (w & ~31); x += 32) {
__asm__ volatile (
"movdqu (%[mask]), %%xmm0\n"
"movdqu 0(%[src]), %%xmm1\n"
"movdqu 16(%[src]), %%xmm2\n"
"movdqu 32(%[src]), %%xmm3\n"
"movdqu 48(%[src]), %%xmm4\n"
"pshufb %%xmm0, %%xmm1\n"
"pshufb %%xmm0, %%xmm2\n"
"pshufb %%xmm0, %%xmm3\n"
"pshufb %%xmm0, %%xmm4\n"
"movq %%xmm1, 0(%[dstu])\n"
"movq %%xmm2, 8(%[dstu])\n"
"movhpd %%xmm1, 0(%[dstv])\n"
"movhpd %%xmm2, 8(%[dstv])\n"
"movq %%xmm3, 16(%[dstu])\n"
"movq %%xmm4, 24(%[dstu])\n"
"movhpd %%xmm3, 16(%[dstv])\n"
"movhpd %%xmm4, 24(%[dstv])\n"
: : [dstu]"r"(&dstu[x]), [dstv]"r"(&dstv[x]), [src]"r"(&src[2*x]), [mask]"r"(maskm) : "memory");
}
/* Remaining */
for (; x < w; x++) {
dstu[x] = src[2*x+0];
dstv[x] = src[2*x+1];
}
src += src_pitch;
dstu += dstu_pitch;
dstv += dstv_pitch;
}
__asm__ volatile ( "emms" );
}
#endif
static int Extract(vlc_va_t *external, picture_t *picture, AVFrame *ff) static int Extract(vlc_va_t *external, picture_t *picture, AVFrame *ff)
{ {
vlc_va_dxva2_t *va = vlc_va_dxva2_Get(external); vlc_va_dxva2_t *va = vlc_va_dxva2_Get(external);
LPDIRECT3DSURFACE9 d3d = (LPDIRECT3DSURFACE9)(uintptr_t)ff->data[3]; LPDIRECT3DSURFACE9 d3d = (LPDIRECT3DSURFACE9)(uintptr_t)ff->data[3];
if (!va->surface_cache)
return VLC_EGENERIC;
/* */ /* */
assert(va->render == MAKEFOURCC('Y','V','1','2') ||
va->render == MAKEFOURCC('N','V','1','2'));
assert(va->output == MAKEFOURCC('Y','V','1','2')); assert(va->output == MAKEFOURCC('Y','V','1','2'));
/* */ /* */
...@@ -419,45 +358,15 @@ static int Extract(vlc_va_t *external, picture_t *picture, AVFrame *ff) ...@@ -419,45 +358,15 @@ static int Extract(vlc_va_t *external, picture_t *picture, AVFrame *ff)
return VLC_EGENERIC; return VLC_EGENERIC;
} }
/* Copy the Y plane */
plane_t src;
memset(&src, 0, sizeof(src));
src.p_pixels = lock.pBits;
src.i_pitch = lock.Pitch;
src.i_lines = va->surface_height;
src.i_visible_pitch = va->surface_width;
src.i_visible_lines = va->surface_height;
src.i_pixel_pitch = 1;
plane_CopyPixels(&picture->p[0], &src);
/* */
src.p_pixels += src.i_pitch * src.i_lines * src.i_pixel_pitch;
/* */
if (va->render == MAKEFOURCC('Y','V','1','2')) { if (va->render == MAKEFOURCC('Y','V','1','2')) {
src.i_pitch /= 2; CopyFromYv12(picture, &lock,
src.i_lines /= 2; va->surface_cache, va->surface_cache_size,
src.i_visible_pitch /= 2; va->surface_width, va->surface_height);
src.i_visible_lines /= 2;
for (unsigned n = 1; n < 3; n++) {
plane_CopyPixels(&picture->p[n], &src);
src.p_pixels += src.i_pitch * src.i_lines * src.i_pixel_pitch;
}
} else { } else {
plane_t *u = &picture->p[2]; assert(va->render == MAKEFOURCC('N','V','1','2'));
plane_t *v = &picture->p[1]; CopyFromNv12(picture, &lock,
va->surface_cache, va->surface_cache_size,
void (*split)(uint8_t *, size_t, uint8_t *, size_t, const uint8_t *, size_t, unsigned, unsigned); va->surface_width, va->surface_height);
#ifdef HAVE_SPLITUV_SSSE3
if (vlc_CPU() & CPU_CAPABILITY_SSSE3)
split = SplitUVSSSE3;
else
#endif
split = SplitUV;
split(u->p_pixels, u->i_pitch, v->p_pixels, v->i_pitch,
src.p_pixels, src.i_pitch,
va->surface_width/2, va->surface_height/2);
} }
/* */ /* */
...@@ -1013,11 +922,319 @@ static void DxCreateVideoConversion(vlc_va_dxva2_t *va) ...@@ -1013,11 +922,319 @@ static void DxCreateVideoConversion(vlc_va_dxva2_t *va)
va->output = va->render; va->output = va->render;
break; break;
} }
va->surface_cache_size = __MAX((va->surface_width + 0x0f) & ~ 0x0f, 4096);
va->surface_cache_base = malloc(16 + va->surface_cache_size);
va->surface_cache = &va->surface_cache_base[16 - ((intptr_t)va->surface_cache_base & 0x0f)];
} }
static void DxDestroyVideoConversion(vlc_va_dxva2_t *va) static void DxDestroyVideoConversion(vlc_va_dxva2_t *va)
{ {
VLC_UNUSED(va); free(va->surface_cache_base);
va->surface_cache_base = NULL;
va->surface_cache = NULL;
va->surface_cache_size = 0;
}
/* Copy 64 bytes from srcp to dsp loading data with the SSE>=2 instruction load and
* storing data with the SSE>=2 instruction store.
*/
#define COPY64(dstp, srcp, load, store) \
asm volatile ( \
load " 0(%[src]), %%xmm1\n" \
load " 16(%[src]), %%xmm2\n" \
load " 32(%[src]), %%xmm3\n" \
load " 48(%[src]), %%xmm4\n" \
store " %%xmm1, 0(%[dst])\n" \
store " %%xmm2, 16(%[dst])\n" \
store " %%xmm3, 32(%[dst])\n" \
store " %%xmm4, 48(%[dst])\n" \
: : [dst]"r"(dstp), [src]"r"(srcp) : "memory")
/* Execute the instruction op only if SSE2 is supported. */
#ifdef CAN_COMPILE_SSE2
# define ASM_SSE2(cpu, op) do { \
if (cpu & CPU_CAPABILITY_SSE2) \
asm volatile (op); \
} while (0)
#else
# define ASM_SSE2(cpu, op)
#endif
/* Optimized copy from "Uncacheable Speculative Write Combining" memory
* as used by some video surface.
* XXX It is really efficient only when SSE4.1 is available.
*/
static void CopyFromUswc(uint8_t *dst, size_t dst_pitch,
const uint8_t *src, size_t src_pitch,
unsigned unaligned,
unsigned width, unsigned height,
unsigned cpu)
{
assert(((intptr_t)dst & 0x0f) == 0 && (dst_pitch & 0x0f) == 0);
ASM_SSE2(cpu, "mfence");
for (unsigned y = 0; y < height; y++) {
unsigned x;
for (x = 0; x < unaligned; x++)
dst[x] = src[x];
#ifdef CAN_COMPILE_SSE4_1
if (cpu & CPU_CAPABILITY_SSE4_1) {
if (!unaligned) {
for (; x+63 < width; x += 64)
COPY64(&dst[x], &src[x], "movntdqa", "movdqa");
} else {
for (; x+63 < width; x += 64)
COPY64(&dst[x], &src[x], "movntdqa", "movdqu");
}
} else
#endif
#ifdef CAN_COMPILE_SSE2
if (cpu & CPU_CAPABILITY_SSE2) {
if (!unaligned) {
for (; x+63 < width; x += 64)
COPY64(&dst[x], &src[x], "movdqa", "movdqa");
} else {
for (; x+63 < width; x += 64)
COPY64(&dst[x], &src[x], "movdqa", "movdqu");
}
}
#endif
for (; x < width; x++)
dst[x] = src[x];
src += src_pitch;
dst += dst_pitch;
}
}
static void Copy2d(uint8_t *dst, size_t dst_pitch,
const uint8_t *src, size_t src_pitch,
unsigned width, unsigned height,
unsigned cpu)
{
assert(((intptr_t)src & 0x0f) == 0 && (src_pitch & 0x0f) == 0);
ASM_SSE2(cpu, "mfence");
for (unsigned y = 0; y < height; y++) {
unsigned x = 0;
bool unaligned = ((intptr_t)dst & 0x0f) != 0;
#ifdef CAN_COMPILE_SSE2
if (cpu & CPU_CAPABILITY_SSE2) {
if (!unaligned) {
for (; x+63 < width; x += 64)
COPY64(&dst[x], &src[x], "movdqa", "movntdq");
} else {
for (; x+63 < width; x += 64)
COPY64(&dst[x], &src[x], "movdqa", "movdqu");
}
}
#endif
for (; x < width; x++)
dst[x] = src[x];
src += src_pitch;
dst += dst_pitch;
}
}
static void SplitUV(uint8_t *dstu, size_t dstu_pitch,
uint8_t *dstv, size_t dstv_pitch,
const uint8_t *src, size_t src_pitch,
unsigned width, unsigned height, unsigned cpu)
{
const uint8_t shuffle[] = { 0, 2, 4, 6, 8, 10, 12, 14,
1, 3, 5, 7, 9, 11, 13, 15 };
const uint8_t mask[] = { 0xff, 0x00, 0xff, 0x00, 0xff, 0x00, 0xff, 0x00,
0xff, 0x00, 0xff, 0x00, 0xff, 0x00, 0xff, 0x00 };
assert(((intptr_t)src & 0x0f) == 0 && (src_pitch & 0x0f) == 0);
ASM_SSE2(cpu, "mfence");
for (unsigned y = 0; y < height; y++) {
unsigned x = 0;
#define LOAD64 \
"movdqa 0(%[src]), %%xmm0\n" \
"movdqa 16(%[src]), %%xmm1\n" \
"movdqa 32(%[src]), %%xmm2\n" \
"movdqa 48(%[src]), %%xmm3\n"
#define STORE2X32 \
"movq %%xmm0, 0(%[dst1])\n" \
"movq %%xmm1, 8(%[dst1])\n" \
"movhpd %%xmm0, 0(%[dst2])\n" \
"movhpd %%xmm1, 8(%[dst2])\n" \
"movq %%xmm2, 16(%[dst1])\n" \
"movq %%xmm3, 24(%[dst1])\n" \
"movhpd %%xmm2, 16(%[dst2])\n" \
"movhpd %%xmm3, 24(%[dst2])\n"
#ifdef CAN_COMPILE_SSSE3
if (cpu & CPU_CAPABILITY_SSSE3) {
for (x = 0; x < (width & ~31); x += 32) {
asm volatile (
"movdqu (%[shuffle]), %%xmm7\n"
LOAD64
"pshufb %%xmm7, %%xmm0\n"
"pshufb %%xmm7, %%xmm1\n"
"pshufb %%xmm7, %%xmm2\n"
"pshufb %%xmm7, %%xmm3\n"
STORE2X32
: : [dst1]"r"(&dstu[x]), [dst2]"r"(&dstv[x]), [src]"r"(&src[2*x]), [shuffle]"r"(shuffle) : "memory");
}
} else
#endif
#ifdef CAN_COMPILE_SSE2
if (cpu & CPU_CAPABILITY_SSE2) {
for (x = 0; x < (width & ~31); x += 32) {
asm volatile (
"movdqu (%[mask]), %%xmm7\n"
LOAD64
"movdqa %%xmm0, %%xmm4\n"
"movdqa %%xmm1, %%xmm5\n"
"movdqa %%xmm2, %%xmm6\n"
"psrlw $8, %%xmm0\n"
"psrlw $8, %%xmm1\n"
"pand %%xmm7, %%xmm4\n"
"pand %%xmm7, %%xmm5\n"
"pand %%xmm7, %%xmm6\n"
"packuswb %%xmm4, %%xmm0\n"
"packuswb %%xmm5, %%xmm1\n"
"pand %%xmm3, %%xmm7\n"
"psrlw $8, %%xmm2\n"
"psrlw $8, %%xmm3\n"
"packuswb %%xmm6, %%xmm2\n"
"packuswb %%xmm7, %%xmm3\n"
STORE2X32
: : [dst2]"r"(&dstu[x]), [dst1]"r"(&dstv[x]), [src]"r"(&src[2*x]), [mask]"r"(mask) : "memory");
}
}
#endif
#undef STORE2X32
#undef LOAD64
for (; x < width; x++) {
dstu[x] = src[2*x+0];
dstv[x] = src[2*x+1];
}
src += src_pitch;
dstu += dstu_pitch;
dstv += dstv_pitch;
}
}
static void CopyPlane(uint8_t *dst, size_t dst_pitch, const uint8_t *src, size_t src_pitch,
uint8_t *cache, size_t cache_size,
unsigned width, unsigned height,
unsigned cpu)
{
const unsigned w16 = (width+15) & ~15;
const unsigned hstep = cache_size / w16;
assert(hstep > 0);
for (unsigned y = 0; y < height; y += hstep) {
const unsigned unaligned = (intptr_t)src & 0x0f;
const unsigned hblock = __MIN(hstep, height - y);
/* Copy a bunch of line into our cache */
CopyFromUswc(cache, w16,
src, src_pitch,
unaligned,
width, hblock, cpu);
/* Copy from our cache to the destination */
Copy2d(dst, dst_pitch,
cache, w16,
width, hblock, cpu);
/* */
src += src_pitch * hblock;
dst += dst_pitch * hblock;
}
ASM_SSE2(cpu, "mfence");
}
static void SplitPlanes(uint8_t *dstu, size_t dstu_pitch,
uint8_t *dstv, size_t dstv_pitch,
const uint8_t *src, size_t src_pitch,
uint8_t *cache, size_t cache_size,
unsigned width, unsigned height,
unsigned cpu)
{
const unsigned w2_16 = (2*width+15) & ~15;
const unsigned hstep = cache_size / w2_16;
assert(hstep > 0);
for (unsigned y = 0; y < height; y += hstep) {
const unsigned unaligned = (intptr_t)src & 0x0f;
const unsigned hblock = __MIN(hstep, height - y);
/* Copy a bunch of line into our cache */
CopyFromUswc(cache, w2_16,
src, src_pitch,
unaligned,
2*width, hblock, cpu);
/* Copy from our cache to the destination */
SplitUV(dstu, dstu_pitch,
dstv, dstv_pitch,
cache, w2_16,
width, hblock, cpu);
/* */
src += src_pitch * hblock;
dstu += dstu_pitch * hblock;
dstv += dstv_pitch * hblock;
}
ASM_SSE2(cpu, "mfence");
}
static void CopyFromNv12(picture_t *dst, const D3DLOCKED_RECT *src,
uint8_t *cache, size_t cache_size,
unsigned width, unsigned height)
{
const unsigned cpu = vlc_CPU();
/* */
CopyPlane(dst->p[0].p_pixels, dst->p[0].i_pitch,
src->pBits, src->Pitch,
cache, cache_size,
width, height, cpu);
SplitPlanes(dst->p[2].p_pixels, dst->p[2].i_pitch,
dst->p[1].p_pixels, dst->p[1].i_pitch,
(const uint8_t*)src->pBits + src->Pitch * height, src->Pitch,
cache, cache_size,
width/2, height/2, cpu);
ASM_SSE2(cpu, "emms");
}
static void CopyFromYv12(picture_t *dst, const D3DLOCKED_RECT *src,
uint8_t *cache, size_t cache_size,
unsigned width, unsigned height)
{
const unsigned cpu = vlc_CPU();
/* */
for (unsigned n = 0, offset = 0; n < 3; n++) {
const unsigned d = n > 0 ? 2 : 1;
CopyPlane(dst->p[n].p_pixels, dst->p[n].i_pitch,
(const uint8_t*)src->pBits + offset, src->Pitch/d,
cache, cache_size,
width/d, height/d, cpu);
offset += (src->Pitch/d) * (height/d);
}
ASM_SSE2(cpu, "emms");
} }
#undef ASM_SSE2
#undef COPY64
#else #else
vlc_va_t *vlc_va_NewDxva2(vlc_object_t *log, int codec_id) vlc_va_t *vlc_va_NewDxva2(vlc_object_t *log, int codec_id)
{ {
......
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