Speeds up everything on AMD by 3x.
This uses 32 local invocations to load state into cache, as well
as to do the RCT faster.
---
libavcodec/vulkan/ffv1_dec.comp | 71 ++++++++++++++++++++-------------
libavcodec/vulkan_ffv1.c | 7 +++-
2 files changed, 50 insertions(+), 28 deletions(-)
diff --git a/libavcodec/vulkan/ffv1_dec.comp b/libavcodec/vulkan/ffv1_dec.comp
index 9eba322b27..3c46ee1771 100644
--- a/libavcodec/vulkan/ffv1_dec.comp
+++ b/libavcodec/vulkan/ffv1_dec.comp
@@ -108,34 +108,37 @@ ivec2 get_pred(ivec2 sp, ivec2 off, int p, int sw,
uint8_t quant_table_idx)
#endif
#ifndef GOLOMB
-int get_isymbol(inout RangeCoder c, uint64_t state)
+#ifdef CACHED_SYMBOL_READER
+shared uint8_t state[CONTEXT_SIZE];
+#define READ(c, off) get_rac_direct(c, state[off])
+#else
+#define READ(c, off) get_rac(c, uint64_t(slice_state) + state_off + off)
+#endif
+
+int get_isymbol(inout RangeCoder c, uint state_off)
{
- if (expectEXT(get_rac(c, state), false))
+ if (expectEXT(READ(c, 0), false))
return 0;
- state += 1;
-
- int e;
- for (e = 0; e < 32; e++)
- if (!get_rac(c, state + min(e, 9)))
+ int e = 1;
+ for (; e < 33; e++)
+ if (!READ(c, min(e, 10)))
break;
- if (expectEXT(e == 0, false)) {
- return get_rac(c, state + 10) ? -1 : 1;
- } else if (expectEXT(e > 31, false)) {
+ if (expectEXT(e == 1, false)) {
+ return READ(c, 11) ? -1 : 1;
+ } else if (expectEXT(e == 33, false)) {
corrupt = true;
return 0;
}
- state += 21;
-
int a = 1;
- for (int i = e - 1; i >= 0; i--) {
+ for (int i = e + 20; i >= 22; i--) {
a <<= 1;
- a |= int(get_rac(c, state + min(i, 9))); // 22..31
+ a |= int(READ(c, min(i, 31)));
}
- return get_rac(c, state - 11 + min(e, 10)) ? -a : a;
+ return READ(c, min(e + 10, 21)) ? -a : a;
}
void decode_line_pcm(inout SliceContext sc, ivec2 sp, int w, int y, int p, int
bits)
@@ -157,7 +160,7 @@ void decode_line_pcm(inout SliceContext sc, ivec2 sp, int
w, int y, int p, int b
}
void decode_line(inout SliceContext sc, ivec2 sp, int w,
- int y, int p, int bits, uint64_t state,
+ int y, int p, int bits, uint state_off,
uint8_t quant_table_idx, const int run_index)
{
#ifndef RGB
@@ -171,19 +174,33 @@ void decode_line(inout SliceContext sc, ivec2 sp, int w,
ivec2 pr = get_pred(sp, ivec2(x, y), p, w,
quant_table_idx);
- int diff = get_isymbol(sc.c, state + CONTEXT_SIZE*abs(pr[0]));
- if (pr[0] < 0)
- diff = -diff;
+ uint context_off = state_off + CONTEXT_SIZE*abs(pr[0]);
+#ifdef CACHED_SYMBOL_READER
+ u8buf sb = u8buf(uint64_t(slice_state) + context_off +
gl_LocalInvocationID.x);
+ state[gl_LocalInvocationID.x] = sb.v;
+ barrier();
+ if (gl_LocalInvocationID.x == 0) {
- uint v = zero_extend(pr[1] + diff, bits);
- imageStore(dec[p], sp + LADDR(ivec2(x, y)), uvec4(v));
+#endif
+
+ int diff = get_isymbol(sc.c, context_off);
+ if (pr[0] < 0)
+ diff = -diff;
+
+ uint v = zero_extend(pr[1] + diff, bits);
+ imageStore(dec[p], sp + LADDR(ivec2(x, y)), uvec4(v));
+
+#ifdef CACHED_SYMBOL_READER
+ }
+ sb.v = state[gl_LocalInvocationID.x];
+#endif
}
}
#else /* GOLOMB */
void decode_line(inout SliceContext sc, ivec2 sp, int w,
- int y, int p, int bits, uint64_t state,
+ int y, int p, int bits, uint state_off,
uint8_t quant_table_idx, inout int run_index)
{
#ifndef RGB
@@ -202,7 +219,7 @@ void decode_line(inout SliceContext sc, ivec2 sp, int w,
ivec2 pr = get_pred(sp, ivec2(x, y), p, w,
quant_table_idx);
- VlcState sb = VlcState(state + VLC_STATE_SIZE*abs(pr[0]));
+ VlcState sb = VlcState(uint64_t(slice_state) + state_off +
VLC_STATE_SIZE*abs(pr[0]));
if (pr[0] == 0 && run_mode == 0)
run_mode = 1;
@@ -263,7 +280,7 @@ ivec4 transform_sample(ivec4 pix, ivec2 rct_coef)
void writeout_rgb(in SliceContext sc, ivec2 sp, int w, int y, bool apply_rct)
{
- for (int x = 0; x < w; x++) {
+ for (uint x = gl_LocalInvocationID.x; x < w; x += gl_WorkGroupSize.x) {
ivec2 lpos = sp + LADDR(ivec2(x, y));
ivec2 pos = sc.slice_pos + ivec2(x, y);
@@ -305,6 +322,8 @@ void decode_slice(inout SliceContext sc, const uint
slice_idx)
/* PCM coding */
#ifndef GOLOMB
if (sc.slice_coding_mode == 1) {
+ if (gl_LocalInvocationID.x > 0)
+ return;
#ifndef RGB
for (int p = 0; p < planes; p++) {
int h = sc.slice_dim.y;
@@ -328,9 +347,7 @@ void decode_slice(inout SliceContext sc, const uint
slice_idx)
#endif
{
u8vec4 quant_table_idx = sc.quant_table_idx.xyyz;
- u64vec4 slice_state_off = (uint64_t(slice_state) +
- slice_idx*plane_state_size*codec_planes) +
- plane_state_size*uvec4(0, 1, 1, 2);
+ u32vec4 slice_state_off = (slice_idx*codec_planes + uvec4(0, 1, 1,
2))*plane_state_size;
#ifndef RGB
for (int p = 0; p < planes; p++) {
diff --git a/libavcodec/vulkan_ffv1.c b/libavcodec/vulkan_ffv1.c
index c1875711bc..33c4e9114d 100644
--- a/libavcodec/vulkan_ffv1.c
+++ b/libavcodec/vulkan_ffv1.c
@@ -823,12 +823,14 @@ static int init_decode_shader(FFV1Context *f,
FFVulkanContext *s,
uint8_t *spv_data;
size_t spv_len;
void *spv_opaque = NULL;
+ int use_cached_reader = ac != AC_GOLOMB_RICE &&
+ s->driver_props.driverID == VK_DRIVER_ID_MESA_RADV;
RET(ff_vk_shader_init(s, shd, "ffv1_dec",
VK_SHADER_STAGE_COMPUTE_BIT,
(const char *[]) { "GL_EXT_buffer_reference",
"GL_EXT_buffer_reference2" }, 2,
- 1, 1, 1,
+ use_cached_reader ? 32 : 1, 1, 1,
0));
if (ac == AC_GOLOMB_RICE)
@@ -837,6 +839,9 @@ static int init_decode_shader(FFV1Context *f,
FFVulkanContext *s,
if (rgb)
av_bprintf(&shd->src, "#define RGB\n");
+ if (use_cached_reader)
+ av_bprintf(&shd->src, "#define CACHED_SYMBOL_READER 1\n");
+
/* Common codec header */
GLSLD(ff_source_common_comp);
--
2.47.2
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
