virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-08-04 22:03:09 +02:00
Code Issues Releases Activity

ffv1enc_vulkan: switch to 2-line cache, unify prediction code

Browse Source
This commit is contained in:
Lynne
2025-05-01 19:45:08 +02:00
parent 52595025c5
commit bd41838b60
5 changed files with 288 additions and 429 deletions
Download Patch File Download Diff File Expand all files Collapse all files

379
libavcodec/ffv1enc_vulkan.c
View File

@ -37,6 +37,9 @@
#define LG_ALIGN_W 32 #define LG_ALIGN_W 32
#define LG_ALIGN_H 32 #define LG_ALIGN_H 32
/* Unlike the decoder, we need 4 lines (but really only 3) */
#define RGB_LINECACHE 4
typedef struct VulkanEncodeFFv1FrameData { typedef struct VulkanEncodeFFv1FrameData {
/* Output data */ /* Output data */
AVBufferRef *out_data_ref; AVBufferRef *out_data_ref;
@ -72,7 +75,6 @@ typedef struct VulkanEncodeFFv1Context {
FFVulkanShader setup; FFVulkanShader setup;
FFVulkanShader reset; FFVulkanShader reset;
FFVulkanShader rct;
FFVulkanShader enc; FFVulkanShader enc;
/* Constant read-only buffers */ /* Constant read-only buffers */
@ -111,7 +113,6 @@ extern const char *ff_source_rangecoder_comp;
extern const char *ff_source_ffv1_vlc_comp; extern const char *ff_source_ffv1_vlc_comp;
extern const char *ff_source_ffv1_common_comp; extern const char *ff_source_ffv1_common_comp;
extern const char *ff_source_ffv1_reset_comp; extern const char *ff_source_ffv1_reset_comp;
extern const char *ff_source_ffv1_enc_rct_comp;
extern const char *ff_source_ffv1_enc_setup_comp; extern const char *ff_source_ffv1_enc_setup_comp;
extern const char *ff_source_ffv1_enc_comp; extern const char *ff_source_ffv1_enc_comp;
@ -120,6 +121,7 @@ typedef struct FFv1VkParameters {
VkDeviceAddress scratch_data; VkDeviceAddress scratch_data;
VkDeviceAddress out_data; VkDeviceAddress out_data;
int32_t fmt_lut[4];
int32_t sar[2]; int32_t sar[2];
uint32_t chroma_shift[2]; uint32_t chroma_shift[2];
@ -127,7 +129,9 @@ typedef struct FFv1VkParameters {
uint32_t context_count; uint32_t context_count;
uint32_t crcref; uint32_t crcref;
uint32_t slice_size_max; uint32_t slice_size_max;
int rct_offset;
uint8_t extend_lookup[8];
uint8_t bits_per_raw_sample; uint8_t bits_per_raw_sample;
uint8_t context_model; uint8_t context_model;
uint8_t version; uint8_t version;
@ -137,13 +141,14 @@ typedef struct FFv1VkParameters {
uint8_t components; uint8_t components;
uint8_t planes; uint8_t planes;
uint8_t codec_planes; uint8_t codec_planes;
uint8_t planar_rgb;
uint8_t transparency; uint8_t transparency;
uint8_t colorspace; uint8_t colorspace;
uint8_t pic_mode; uint8_t pic_mode;
uint8_t ec; uint8_t ec;
uint8_t ppi; uint8_t ppi;
uint8_t chunks; uint8_t chunks;
uint8_t padding[1]; uint8_t padding[4];
} FFv1VkParameters; } FFv1VkParameters;
static void add_push_data(FFVulkanShader *shd) static void add_push_data(FFVulkanShader *shd)
@ -153,6 +158,7 @@ static void add_push_data(FFVulkanShader *shd)
GLSLC(1, u8buf scratch_data; ); GLSLC(1, u8buf scratch_data; );
GLSLC(1, u8buf out_data; ); GLSLC(1, u8buf out_data; );
GLSLC(0, ); GLSLC(0, );
GLSLC(1, ivec4 fmt_lut; );
GLSLC(1, ivec2 sar; ); GLSLC(1, ivec2 sar; );
GLSLC(1, uvec2 chroma_shift; ); GLSLC(1, uvec2 chroma_shift; );
GLSLC(0, ); GLSLC(0, );
@ -160,7 +166,9 @@ static void add_push_data(FFVulkanShader *shd)
GLSLC(1, uint context_count; ); GLSLC(1, uint context_count; );
GLSLC(1, uint32_t crcref; ); GLSLC(1, uint32_t crcref; );
GLSLC(1, uint32_t slice_size_max; ); GLSLC(1, uint32_t slice_size_max; );
GLSLC(1, int rct_offset; );
GLSLC(0, ); GLSLC(0, );
GLSLC(1, uint8_t extend_lookup[8]; );
GLSLC(1, uint8_t bits_per_raw_sample; ); GLSLC(1, uint8_t bits_per_raw_sample; );
GLSLC(1, uint8_t context_model; ); GLSLC(1, uint8_t context_model; );
GLSLC(1, uint8_t version; ); GLSLC(1, uint8_t version; );
@ -170,122 +178,19 @@ static void add_push_data(FFVulkanShader *shd)
GLSLC(1, uint8_t components; ); GLSLC(1, uint8_t components; );
GLSLC(1, uint8_t planes; ); GLSLC(1, uint8_t planes; );
GLSLC(1, uint8_t codec_planes; ); GLSLC(1, uint8_t codec_planes; );
GLSLC(1, uint8_t planar_rgb; );
GLSLC(1, uint8_t transparency; ); GLSLC(1, uint8_t transparency; );
GLSLC(1, uint8_t colorspace; ); GLSLC(1, uint8_t colorspace; );
GLSLC(1, uint8_t pic_mode; ); GLSLC(1, uint8_t pic_mode; );
GLSLC(1, uint8_t ec; ); GLSLC(1, uint8_t ec; );
GLSLC(1, uint8_t ppi; ); GLSLC(1, uint8_t ppi; );
GLSLC(1, uint8_t chunks; ); GLSLC(1, uint8_t chunks; );
GLSLC(1, uint8_t padding[1]; ); GLSLC(1, uint8_t padding[4]; );
GLSLC(0, }; ); GLSLC(0, }; );
ff_vk_shader_add_push_const(shd, 0, sizeof(FFv1VkParameters), ff_vk_shader_add_push_const(shd, 0, sizeof(FFv1VkParameters),
VK_SHADER_STAGE_COMPUTE_BIT); VK_SHADER_STAGE_COMPUTE_BIT);
} }
static int run_rct(AVCodecContext *avctx, FFVkExecContext *exec,
AVFrame *enc_in, VkImageView *enc_in_views,
AVFrame **intermediate_frame, VkImageView *intermediate_views,
VkImageMemoryBarrier2 *img_bar, int *nb_img_bar,
VkBufferMemoryBarrier2 *buf_bar, int *nb_buf_bar,
FFVkBuffer *slice_data_buf, uint32_t slice_data_size)
{
int err;
VulkanEncodeFFv1Context *fv = avctx->priv_data;
FFV1Context *f = &fv->ctx;
FFVulkanFunctions *vk = &fv->s.vkfn;
AVHWFramesContext *src_hwfc = (AVHWFramesContext *)enc_in->hw_frames_ctx->data;
FFv1VkRCTParameters pd;
/* Create a temporaty frame */
*intermediate_frame = av_frame_alloc();
if (!(*intermediate_frame))
return AVERROR(ENOMEM);
RET(av_hwframe_get_buffer(fv->intermediate_frames_ref,
*intermediate_frame, 0));
RET(ff_vk_exec_add_dep_frame(&fv->s, exec, *intermediate_frame,
VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT));
RET(ff_vk_create_imageviews(&fv->s, exec, intermediate_views,
*intermediate_frame,
fv->rep_fmt));
/* Update descriptors */
ff_vk_shader_update_desc_buffer(&fv->s, exec, &fv->rct,
1, 0, 0,
slice_data_buf,
0, slice_data_size*f->slice_count,
VK_FORMAT_UNDEFINED);
ff_vk_shader_update_img_array(&fv->s, exec, &fv->rct,
enc_in, enc_in_views,
1, 1,
VK_IMAGE_LAYOUT_GENERAL,
VK_NULL_HANDLE);
ff_vk_shader_update_img_array(&fv->s, exec, &fv->rct,
*intermediate_frame, intermediate_views,
1, 2,
VK_IMAGE_LAYOUT_GENERAL,
VK_NULL_HANDLE);
ff_vk_frame_barrier(&fv->s, exec, *intermediate_frame, img_bar, nb_img_bar,
VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_SHADER_WRITE_BIT,
VK_IMAGE_LAYOUT_GENERAL,
VK_QUEUE_FAMILY_IGNORED);
/* Prep the input/output images */
vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
.pImageMemoryBarriers = img_bar,
.imageMemoryBarrierCount = *nb_img_bar,
.pBufferMemoryBarriers = buf_bar,
.bufferMemoryBarrierCount = *nb_buf_bar,
});
*nb_img_bar = 0;
if (*nb_buf_bar) {
slice_data_buf->stage = buf_bar[0].dstStageMask;
slice_data_buf->access = buf_bar[0].dstAccessMask;
*nb_buf_bar = 0;
}
/* Run the shader */
ff_vk_exec_bind_shader(&fv->s, exec, &fv->rct);
pd = (FFv1VkRCTParameters) {
.offset = 1 << f->bits_per_raw_sample,
.bits = f->bits_per_raw_sample,
.planar_rgb = ff_vk_mt_is_np_rgb(src_hwfc->sw_format) &&
(ff_vk_count_images((AVVkFrame *)enc_in->data[0]) > 1),
.transparency = f->transparency,
};
/* For some reason the C FFv1 encoder/decoder treats these differently */
if (src_hwfc->sw_format == AV_PIX_FMT_GBRP10 ||
src_hwfc->sw_format == AV_PIX_FMT_GBRP12 ||
src_hwfc->sw_format == AV_PIX_FMT_GBRP14)
memcpy(pd.fmt_lut, (int [4]) { 2, 1, 0, 3 }, 4*sizeof(int));
else
ff_vk_set_perm(src_hwfc->sw_format, pd.fmt_lut, 1);
ff_vk_shader_update_push_const(&fv->s, exec, &fv->rct,
VK_SHADER_STAGE_COMPUTE_BIT,
0, sizeof(pd), &pd);
vk->CmdDispatch(exec->buf, fv->ctx.num_h_slices, fv->ctx.num_v_slices, 1);
/* Add a post-dispatch barrier before encoding */
ff_vk_frame_barrier(&fv->s, exec, *intermediate_frame, img_bar, nb_img_bar,
VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_SHADER_READ_BIT,
VK_IMAGE_LAYOUT_GENERAL,
VK_QUEUE_FAMILY_IGNORED);
fail:
return err;
}
static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx, static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
FFVkExecContext *exec, FFVkExecContext *exec,
const AVFrame *pict) const AVFrame *pict)
@ -298,8 +203,6 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
VulkanEncodeFFv1FrameData *fd = exec->opaque; VulkanEncodeFFv1FrameData *fd = exec->opaque;
FFv1VkParameters pd; FFv1VkParameters pd;
AVFrame *intermediate_frame = NULL;
/* Slice data */ /* Slice data */
AVBufferRef *slice_data_ref; AVBufferRef *slice_data_ref;
FFVkBuffer *slice_data_buf; FFVkBuffer *slice_data_buf;
@ -318,11 +221,11 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
uint32_t context_count = f->context_count[f->context_model]; uint32_t context_count = f->context_count[f->context_model];
const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(avctx->sw_pix_fmt); const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
VkImageView in_views[AV_NUM_DATA_POINTERS]; AVFrame *src = (AVFrame *)pict;
VkImageView intermediate_views[AV_NUM_DATA_POINTERS]; VkImageView src_views[AV_NUM_DATA_POINTERS];
AVFrame *enc_in = (AVFrame *)pict; AVFrame *tmp = NULL;
VkImageView *enc_in_views = in_views; VkImageView tmp_views[AV_NUM_DATA_POINTERS];
VkImageMemoryBarrier2 img_bar[37]; VkImageMemoryBarrier2 img_bar[37];
int nb_img_bar = 0; int nb_img_bar = 0;
@ -402,27 +305,44 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
ff_vk_exec_add_dep_buf(&fv->s, exec, &fd->out_data_ref, 1, 1); ff_vk_exec_add_dep_buf(&fv->s, exec, &fd->out_data_ref, 1, 1);
/* Prepare input frame */ /* Prepare input frame */
RET(ff_vk_exec_add_dep_frame(&fv->s, exec, enc_in, RET(ff_vk_exec_add_dep_frame(&fv->s, exec, src,
VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT)); VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT));
RET(ff_vk_create_imageviews(&fv->s, exec, enc_in_views, enc_in, RET(ff_vk_create_imageviews(&fv->s, exec, src_views, src,
fv->rep_fmt)); fv->rep_fmt));
ff_vk_frame_barrier(&fv->s, exec, enc_in, img_bar, &nb_img_bar, ff_vk_frame_barrier(&fv->s, exec, src, img_bar, &nb_img_bar,
VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_SHADER_READ_BIT, VK_ACCESS_SHADER_READ_BIT,
VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_GENERAL,
VK_QUEUE_FAMILY_IGNORED); VK_QUEUE_FAMILY_IGNORED);
/* Setup shader needs the original input */ if (fv->is_rgb) {
/* Create a temporaty frame */
tmp = av_frame_alloc();
if (!(tmp))
return AVERROR(ENOMEM);
RET(av_hwframe_get_buffer(fv->intermediate_frames_ref,
tmp, 0));
RET(ff_vk_exec_add_dep_frame(&fv->s, exec, tmp,
VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT));
RET(ff_vk_create_imageviews(&fv->s, exec, tmp_views,
tmp,
fv->rep_fmt));
}
/* Setup shader */
ff_vk_shader_update_desc_buffer(&fv->s, exec, &fv->setup, ff_vk_shader_update_desc_buffer(&fv->s, exec, &fv->setup,
1, 0, 0, 1, 0, 0,
slice_data_buf, slice_data_buf,
0, slice_data_size*f->slice_count, 0, slice_data_size*f->slice_count,
VK_FORMAT_UNDEFINED); VK_FORMAT_UNDEFINED);
ff_vk_shader_update_img_array(&fv->s, exec, &fv->setup, ff_vk_shader_update_img_array(&fv->s, exec, &fv->setup,
enc_in, enc_in_views, src, src_views,
1, 1, 1, 1,
VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_GENERAL,
VK_NULL_HANDLE); VK_NULL_HANDLE);
@ -471,6 +391,7 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
.plane_state_size = plane_state_size, .plane_state_size = plane_state_size,
.context_count = context_count, .context_count = context_count,
.crcref = f->crcref, .crcref = f->crcref,
.rct_offset = 1 << f->bits_per_raw_sample,
.slice_size_max = out_data_buf->size / f->slice_count, .slice_size_max = out_data_buf->size / f->slice_count,
.context_model = fv->ctx.context_model, .context_model = fv->ctx.context_model,
.version = f->version, .version = f->version,
@ -480,6 +401,8 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
.components = fmt_desc->nb_components, .components = fmt_desc->nb_components,
.planes = av_pix_fmt_count_planes(avctx->sw_pix_fmt), .planes = av_pix_fmt_count_planes(avctx->sw_pix_fmt),
.codec_planes = f->plane_count, .codec_planes = f->plane_count,
.planar_rgb = ff_vk_mt_is_np_rgb(avctx->sw_pix_fmt) &&
(ff_vk_count_images((AVVkFrame *)src->data[0]) > 1),
.transparency = f->transparency, .transparency = f->transparency,
.colorspace = f->colorspace, .colorspace = f->colorspace,
.pic_mode = !(pict->flags & AV_FRAME_FLAG_INTERLACED) ? 3 : .pic_mode = !(pict->flags & AV_FRAME_FLAG_INTERLACED) ? 3 :
@ -488,11 +411,35 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
.ppi = fv->ppi, .ppi = fv->ppi,
.chunks = fv->chunks, .chunks = fv->chunks,
}; };
/* For some reason the C FFv1 encoder/decoder treats these differently */
if (avctx->sw_pix_fmt == AV_PIX_FMT_GBRP10 ||
avctx->sw_pix_fmt == AV_PIX_FMT_GBRP12 ||
avctx->sw_pix_fmt == AV_PIX_FMT_GBRP14)
memcpy(pd.fmt_lut, (int [4]) { 2, 1, 0, 3 }, 4*sizeof(int));
else
ff_vk_set_perm(avctx->sw_pix_fmt, pd.fmt_lut, 1);
for (int i = 0; i < f->quant_table_count; i++)
pd.extend_lookup[i] = (f->quant_tables[i][3][127] != 0) ||
(f->quant_tables[i][4][127] != 0);
ff_vk_shader_update_push_const(&fv->s, exec, &fv->setup, ff_vk_shader_update_push_const(&fv->s, exec, &fv->setup,
VK_SHADER_STAGE_COMPUTE_BIT, VK_SHADER_STAGE_COMPUTE_BIT,
0, sizeof(pd), &pd); 0, sizeof(pd), &pd);
vk->CmdDispatch(exec->buf, fv->ctx.num_h_slices, fv->ctx.num_v_slices, 1); vk->CmdDispatch(exec->buf, fv->ctx.num_h_slices, fv->ctx.num_v_slices, 1);
/* Clean up temporary image */
if (fv->is_rgb) {
AVVkFrame *vkf = (AVVkFrame *)tmp->data[0];
vk->CmdClearColorImage(exec->buf, vkf->img[0], VK_IMAGE_LAYOUT_GENERAL,
&((VkClearColorValue) { 0 }),
1, &((VkImageSubresourceRange) {
.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.levelCount = 1,
.layerCount = 1,
}));
}
/* Setup shader modified the slice data buffer */ /* Setup shader modified the slice data buffer */
buf_bar[nb_buf_bar++] = (VkBufferMemoryBarrier2) { buf_bar[nb_buf_bar++] = (VkBufferMemoryBarrier2) {
.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2, .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2,
@ -546,19 +493,6 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
f->plane_count); f->plane_count);
} }
/* Run RCT shader */
if (fv->is_rgb) {
RET(run_rct(avctx, exec,
enc_in, enc_in_views,
&intermediate_frame, intermediate_views,
img_bar, &nb_img_bar, buf_bar, &nb_buf_bar,
slice_data_buf, slice_data_size));
/* Use the new frame */
enc_in = intermediate_frame;
enc_in_views = intermediate_views;
}
/* If the reset shader ran, insert a barrier now. */ /* If the reset shader ran, insert a barrier now. */
if (f->key_frame || f->version > 3) { if (f->key_frame || f->version > 3) {
/* Reset shader modified the slice data buffer */ /* Reset shader modified the slice data buffer */
@ -577,6 +511,15 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
}; };
} }
if (fv->is_rgb) {
ff_vk_frame_barrier(&fv->s, exec, tmp, img_bar, &nb_img_bar,
VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
VK_IMAGE_LAYOUT_GENERAL,
VK_QUEUE_FAMILY_IGNORED);
}
/* Final barrier before encoding */ /* Final barrier before encoding */
vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) { vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO, .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
@ -599,7 +542,7 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
0, slice_data_size*f->slice_count, 0, slice_data_size*f->slice_count,
VK_FORMAT_UNDEFINED); VK_FORMAT_UNDEFINED);
ff_vk_shader_update_img_array(&fv->s, exec, &fv->enc, ff_vk_shader_update_img_array(&fv->s, exec, &fv->enc,
enc_in, enc_in_views, src, src_views,
1, 1, 1, 1,
VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_GENERAL,
VK_NULL_HANDLE); VK_NULL_HANDLE);
@ -608,6 +551,12 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
results_data_buf, results_data_buf,
0, results_data_buf->size, 0, results_data_buf->size,
VK_FORMAT_UNDEFINED); VK_FORMAT_UNDEFINED);
if (fv->is_rgb)
ff_vk_shader_update_img_array(&fv->s, exec, &fv->enc,
tmp, tmp_views,
1, 3,
VK_IMAGE_LAYOUT_GENERAL,
VK_NULL_HANDLE);
ff_vk_exec_bind_shader(&fv->s, exec, &fv->enc); ff_vk_exec_bind_shader(&fv->s, exec, &fv->enc);
ff_vk_shader_update_push_const(&fv->s, exec, &fv->enc, ff_vk_shader_update_push_const(&fv->s, exec, &fv->enc,
@ -624,11 +573,11 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
/* This, if needed, was referenced by the execution context /* This, if needed, was referenced by the execution context
* as it was declared as a dependency. */ * as it was declared as a dependency. */
av_frame_free(&intermediate_frame); av_frame_free(&tmp);
return 0; return 0;
fail: fail:
av_frame_free(&intermediate_frame); av_frame_free(&tmp);
ff_vk_exec_discard_deps(&fv->s, exec); ff_vk_exec_discard_deps(&fv->s, exec);
return err; return err;
@ -846,6 +795,7 @@ static int init_indirect(AVCodecContext *avctx, enum AVPixelFormat sw_format)
{ {
int err; int err;
VulkanEncodeFFv1Context *fv = avctx->priv_data; VulkanEncodeFFv1Context *fv = avctx->priv_data;
FFV1Context *f = &fv->ctx;
AVHWFramesContext *frames_ctx; AVHWFramesContext *frames_ctx;
AVVulkanFramesContext *vk_frames; AVVulkanFramesContext *vk_frames;
@ -856,12 +806,13 @@ static int init_indirect(AVCodecContext *avctx, enum AVPixelFormat sw_format)
frames_ctx = (AVHWFramesContext *)fv->intermediate_frames_ref->data; frames_ctx = (AVHWFramesContext *)fv->intermediate_frames_ref->data;
frames_ctx->format = AV_PIX_FMT_VULKAN; frames_ctx->format = AV_PIX_FMT_VULKAN;
frames_ctx->sw_format = sw_format; frames_ctx->sw_format = sw_format;
frames_ctx->width = FFALIGN(fv->s.frames->width, 32); frames_ctx->width = fv->s.frames->width;
frames_ctx->height = FFALIGN(fv->s.frames->height, 32); frames_ctx->height = f->num_v_slices*RGB_LINECACHE;
vk_frames = frames_ctx->hwctx; vk_frames = frames_ctx->hwctx;
vk_frames->tiling = VK_IMAGE_TILING_OPTIMAL; vk_frames->tiling = VK_IMAGE_TILING_OPTIMAL;
vk_frames->usage = VK_IMAGE_USAGE_STORAGE_BIT; vk_frames->usage = VK_IMAGE_USAGE_STORAGE_BIT |
VK_IMAGE_USAGE_TRANSFER_DST_BIT;
vk_frames->img_flags = VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT; vk_frames->img_flags = VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
err = av_hwframe_ctx_init(fv->intermediate_frames_ref); err = av_hwframe_ctx_init(fv->intermediate_frames_ref);
@ -929,6 +880,7 @@ static void define_shared_code(AVCodecContext *avctx, FFVulkanShader *shd)
FFV1Context *f = &fv->ctx; FFV1Context *f = &fv->ctx;
int smp_bits = fv->ctx.use32bit ? 32 : 16; int smp_bits = fv->ctx.use32bit ? 32 : 16;
av_bprintf(&shd->src, "#define RGB_LINECACHE %i\n" ,RGB_LINECACHE);
av_bprintf(&shd->src, "#define CONTEXT_SIZE %i\n" ,CONTEXT_SIZE); av_bprintf(&shd->src, "#define CONTEXT_SIZE %i\n" ,CONTEXT_SIZE);
av_bprintf(&shd->src, "#define MAX_QUANT_TABLE_MASK 0x%x\n" ,MAX_QUANT_TABLE_MASK); av_bprintf(&shd->src, "#define MAX_QUANT_TABLE_MASK 0x%x\n" ,MAX_QUANT_TABLE_MASK);
@ -1120,122 +1072,6 @@ fail:
return err; return err;
} }
static int init_rct_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
{
int err;
VulkanEncodeFFv1Context *fv = avctx->priv_data;
FFVulkanShader *shd = &fv->rct;
FFVulkanDescriptorSetBinding *desc_set;
uint8_t *spv_data;
size_t spv_len;
void *spv_opaque = NULL;
int wg_count = sqrt(fv->s.props.properties.limits.maxComputeWorkGroupInvocations);
enum AVPixelFormat intermediate_fmt = get_supported_rgb_buffer_fmt(avctx);
if (intermediate_fmt == AV_PIX_FMT_NONE) {
av_log(avctx, AV_LOG_ERROR, "Unable to find a supported compatible "
"pixel format for RCT buffer!\n");
return AVERROR(ENOTSUP);
}
RET(init_indirect(avctx, intermediate_fmt));
RET(ff_vk_shader_init(&fv->s, shd, "ffv1_rct",
VK_SHADER_STAGE_COMPUTE_BIT,
(const char *[]) { "GL_EXT_buffer_reference",
"GL_EXT_buffer_reference2" }, 2,
wg_count, wg_count, 1,
0));
/* Common codec header */
GLSLD(ff_source_common_comp);
GLSLC(0, layout(push_constant, scalar) uniform pushConstants { );
GLSLC(1, ivec4 fmt_lut; );
GLSLC(1, int offset; );
GLSLC(1, uint8_t bits; );
GLSLC(1, uint8_t planar_rgb; );
GLSLC(1, uint8_t color_planes; );
GLSLC(1, uint8_t transparency; );
GLSLC(1, uint8_t version; );
GLSLC(1, uint8_t micro_version; );
GLSLC(1, uint8_t padding[2]; );
GLSLC(0, }; );
ff_vk_shader_add_push_const(shd, 0, sizeof(FFv1VkRCTParameters),
VK_SHADER_STAGE_COMPUTE_BIT);
av_bprintf(&shd->src, "#define MAX_QUANT_TABLES %i\n", MAX_QUANT_TABLES);
av_bprintf(&shd->src, "#define MAX_CONTEXT_INPUTS %i\n", MAX_CONTEXT_INPUTS);
av_bprintf(&shd->src, "#define MAX_QUANT_TABLE_SIZE %i\n", MAX_QUANT_TABLE_SIZE);
desc_set = (FFVulkanDescriptorSetBinding []) {
{
.name = "rangecoder_static_buf",
.type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.mem_layout = "scalar",
.buf_content = "uint8_t zero_one_state[512];",
},
{
.name = "quant_buf",
.type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.mem_layout = "scalar",
.buf_content = "int16_t quant_table[MAX_QUANT_TABLES]"
"[MAX_CONTEXT_INPUTS][MAX_QUANT_TABLE_SIZE];",
},
};
RET(ff_vk_shader_add_descriptor_set(&fv->s, shd, desc_set, 2, 1, 0));
define_shared_code(avctx, shd);
desc_set = (FFVulkanDescriptorSetBinding []) {
{
.name = "slice_data_buf",
.type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
.mem_quali = "readonly",
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.buf_content = "SliceContext slice_ctx[1024];",
},
{
.name = "src",
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.dimensions = 2,
.mem_layout = ff_vk_shader_rep_fmt(fv->s.frames->sw_format,
fv->rep_fmt),
.elems = av_pix_fmt_count_planes(fv->s.frames->sw_format),
.mem_quali = "readonly",
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
},
{
.name = "dst",
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.dimensions = 2,
.mem_layout = ff_vk_shader_rep_fmt(intermediate_fmt,
fv->rep_fmt),
.elems = av_pix_fmt_count_planes(intermediate_fmt),
.mem_quali = "writeonly",
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
},
};
RET(ff_vk_shader_add_descriptor_set(&fv->s, shd, desc_set, 3, 0, 0));
GLSLD(ff_source_ffv1_enc_rct_comp);
RET(spv->compile_shader(&fv->s, spv, shd, &spv_data, &spv_len, "main",
&spv_opaque));
RET(ff_vk_shader_link(&fv->s, shd, spv_data, spv_len, "main"));
RET(ff_vk_shader_register_exec(&fv->s, &fv->exec_pool, shd));
fail:
if (spv_opaque)
spv->free_shader(spv, &spv_opaque);
return err;
}
static int init_encode_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv) static int init_encode_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
{ {
int err; int err;
@ -1243,10 +1079,6 @@ static int init_encode_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
FFVulkanShader *shd = &fv->enc; FFVulkanShader *shd = &fv->enc;
FFVulkanDescriptorSetBinding *desc_set; FFVulkanDescriptorSetBinding *desc_set;
AVHWFramesContext *frames_ctx = fv->intermediate_frames_ref ?
(AVHWFramesContext *)fv->intermediate_frames_ref->data :
fv->s.frames;
uint8_t *spv_data; uint8_t *spv_data;
size_t spv_len; size_t spv_len;
void *spv_opaque = NULL; void *spv_opaque = NULL;
@ -1307,9 +1139,9 @@ static int init_encode_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
.name = "src", .name = "src",
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.dimensions = 2, .dimensions = 2,
.mem_layout = ff_vk_shader_rep_fmt(frames_ctx->sw_format, .mem_layout = ff_vk_shader_rep_fmt(fv->s.frames->sw_format,
fv->rep_fmt), fv->rep_fmt),
.elems = av_pix_fmt_count_planes(frames_ctx->sw_format), .elems = av_pix_fmt_count_planes(fv->s.frames->sw_format),
.mem_quali = "readonly", .mem_quali = "readonly",
.stages = VK_SHADER_STAGE_COMPUTE_BIT, .stages = VK_SHADER_STAGE_COMPUTE_BIT,
}, },
@ -1321,7 +1153,19 @@ static int init_encode_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
.buf_content = "uint64_t slice_results[2048];", .buf_content = "uint64_t slice_results[2048];",
}, },
}; };
RET(ff_vk_shader_add_descriptor_set(&fv->s, shd, desc_set, 3, 0, 0)); if (fv->is_rgb) {
AVHWFramesContext *intermediate_frames_ctx;
intermediate_frames_ctx = (AVHWFramesContext *)fv->intermediate_frames_ref->data;
desc_set[3] = (FFVulkanDescriptorSetBinding) {
.name = "tmp",
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.dimensions = 2,
.mem_layout = ff_vk_shader_rep_fmt(intermediate_frames_ctx->sw_format,
FF_VK_REP_NATIVE),
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
};
}
RET(ff_vk_shader_add_descriptor_set(&fv->s, shd, desc_set, 3 + fv->is_rgb, 0, 0));
GLSLD(ff_source_ffv1_enc_comp); GLSLD(ff_source_ffv1_enc_comp);
@ -1566,13 +1410,15 @@ static av_cold int vulkan_encode_ffv1_init(AVCodecContext *avctx)
return err; return err;
} }
/* Init RCT shader */
if (fv->is_rgb) { if (fv->is_rgb) {
err = init_rct_shader(avctx, spv); enum AVPixelFormat intermediate_fmt = get_supported_rgb_buffer_fmt(avctx);
if (err < 0) { if (intermediate_fmt == AV_PIX_FMT_NONE) {
spv->uninit(&spv); av_log(avctx, AV_LOG_ERROR, "Unable to find a supported compatible "
return err; "pixel format for RCT buffer!\n");
return AVERROR(ENOTSUP);
} }
RET(init_indirect(avctx, intermediate_fmt));
} }
/* Encode shader */ /* Encode shader */
@ -1659,7 +1505,6 @@ static av_cold int vulkan_encode_ffv1_close(AVCodecContext *avctx)
ff_vk_exec_pool_free(&fv->s, &fv->transfer_exec_pool); ff_vk_exec_pool_free(&fv->s, &fv->transfer_exec_pool);
ff_vk_shader_free(&fv->s, &fv->enc); ff_vk_shader_free(&fv->s, &fv->enc);
ff_vk_shader_free(&fv->s, &fv->rct);
ff_vk_shader_free(&fv->s, &fv->reset); ff_vk_shader_free(&fv->s, &fv->reset);
ff_vk_shader_free(&fv->s, &fv->setup); ff_vk_shader_free(&fv->s, &fv->setup);

87
libavcodec/vulkan/ffv1_common.comp
View File

@ -92,3 +92,90 @@ uint slice_coord(uint width, uint sx, uint num_h_slices, uint chroma_shift)
return sx; return sx;
} }
#ifdef RGB
#define RGB_LBUF (RGB_LINECACHE - 1)
#define LADDR(p) (ivec2((p).x, ((p).y & RGB_LBUF)))
ivec2 get_pred(readonly uimage2D pred, ivec2 sp, ivec2 off,
int comp, int sw, uint8_t quant_table_idx, bool extend_lookup)
{
const ivec2 yoff_border1 = expectEXT(off.x == 0, false) ? ivec2(1, -1) : ivec2(0, 0);
/* Thanks to the same coincidence as below, we can skip checking if off == 0, 1 */
VTYPE3 top = VTYPE3(TYPE(imageLoad(pred, sp + LADDR(off + ivec2(-1, -1) + yoff_border1))[comp]),
TYPE(imageLoad(pred, sp + LADDR(off + ivec2(0, -1)))[comp]),
TYPE(imageLoad(pred, sp + LADDR(off + ivec2(min(1, sw - off.x - 1), -1)))[comp]));
/* Normally, we'd need to check if off != ivec2(0, 0) here, since otherwise, we must
* return zero. However, ivec2(-1, 0) + ivec2(1, -1) == ivec2(0, -1), e.g. previous
* row, 0 offset, same slice, which is zero since we zero out the buffer for RGB */
TYPE cur = TYPE(imageLoad(pred, sp + LADDR(off + ivec2(-1, 0) + yoff_border1))[comp]);
int base = quant_table[quant_table_idx][0][(cur - top[0]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][1][(top[0] - top[1]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][2][(top[1] - top[2]) & MAX_QUANT_TABLE_MASK];
if (expectEXT(extend_lookup, false)) {
TYPE cur2 = TYPE(0);
if (expectEXT(off.x > 0, true)) {
const ivec2 yoff_border2 = expectEXT(off.x == 1, false) ? ivec2(-1, -1) : ivec2(-2, 0);
cur2 = TYPE(imageLoad(pred, sp + LADDR(off + yoff_border2))[comp]);
}
base += quant_table[quant_table_idx][3][(cur2 - cur) & MAX_QUANT_TABLE_MASK];
/* top-2 became current upon swap */
TYPE top2 = TYPE(imageLoad(pred, sp + LADDR(off))[comp]);
base += quant_table[quant_table_idx][4][(top2 - top[1]) & MAX_QUANT_TABLE_MASK];
}
/* context, prediction */
return ivec2(base, predict(cur, VTYPE2(top)));
}
#else /* RGB */
#define LADDR(p) (p)
ivec2 get_pred(readonly uimage2D pred, ivec2 sp, ivec2 off,
int comp, int sw, uint8_t quant_table_idx, bool extend_lookup)
{
const ivec2 yoff_border1 = off.x == 0 ? ivec2(1, -1) : ivec2(0, 0);
sp += off;
VTYPE3 top = VTYPE3(TYPE(0),
TYPE(0),
TYPE(0));
if (off.y > 0 && off != ivec2(0, 1))
top[0] = TYPE(imageLoad(pred, sp + ivec2(-1, -1) + yoff_border1)[comp]);
if (off.y > 0) {
top[1] = TYPE(imageLoad(pred, sp + ivec2(0, -1))[comp]);
top[2] = TYPE(imageLoad(pred, sp + ivec2(min(1, sw - off.x - 1), -1))[comp]);
}
TYPE cur = TYPE(0);
if (off != ivec2(0, 0))
cur = TYPE(imageLoad(pred, sp + ivec2(-1, 0) + yoff_border1)[comp]);
int base = quant_table[quant_table_idx][0][(cur - top[0]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][1][(top[0] - top[1]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][2][(top[1] - top[2]) & MAX_QUANT_TABLE_MASK];
if (expectEXT(extend_lookup, false)) {
TYPE cur2 = TYPE(0);
if (off.x > 0 && off != ivec2(1, 0)) {
const ivec2 yoff_border2 = off.x == 1 ? ivec2(1, -1) : ivec2(0, 0);
cur2 = TYPE(imageLoad(pred, sp + ivec2(-2, 0) + yoff_border2)[comp]);
}
base += quant_table[quant_table_idx][3][(cur2 - cur) & MAX_QUANT_TABLE_MASK];
TYPE top2 = TYPE(0);
if (off.y > 1)
top2 = TYPE(imageLoad(pred, sp + ivec2(0, -2))[comp]);
base += quant_table[quant_table_idx][4][(top2 - top[1]) & MAX_QUANT_TABLE_MASK];
}
/* context, prediction */
return ivec2(base, predict(cur, VTYPE2(top)));
}
#endif

91
libavcodec/vulkan/ffv1_dec.comp
View File

@ -20,93 +20,6 @@
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/ */
#ifndef RGB
#define LADDR(p) (p)
#else
#define RGB_LINECACHE 2
#define RGB_LBUF (RGB_LINECACHE - 1)
#define LADDR(p) (ivec2((p).x, ((p).y & RGB_LBUF)))
#endif
#ifdef RGB
ivec2 get_pred(readonly uimage2D pred, ivec2 sp, ivec2 off, int comp, int sw, uint8_t quant_table_idx)
{
const ivec2 yoff_border1 = expectEXT(off.x == 0, false) ? ivec2(1, -1) : ivec2(0, 0);
/* Thanks to the same coincidence as below, we can skip checking if off == 0, 1 */
VTYPE3 top = VTYPE3(TYPE(imageLoad(pred, sp + LADDR(off + ivec2(-1, -1) + yoff_border1))[comp]),
TYPE(imageLoad(pred, sp + LADDR(off + ivec2(0, -1)))[comp]),
TYPE(imageLoad(pred, sp + LADDR(off + ivec2(min(1, sw - off.x - 1), -1)))[comp]));
/* Normally, we'd need to check if off != ivec2(0, 0) here, since otherwise, we must
* return zero. However, ivec2(-1, 0) + ivec2(1, -1) == ivec2(0, -1), e.g. previous
* row, 0 offset, same slice, which is zero since we zero out the buffer for RGB */
TYPE cur = TYPE(imageLoad(pred, sp + LADDR(off + ivec2(-1, 0) + yoff_border1))[comp]);
int base = quant_table[quant_table_idx][0][(cur - top[0]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][1][(top[0] - top[1]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][2][(top[1] - top[2]) & MAX_QUANT_TABLE_MASK];
if (expectEXT(extend_lookup[quant_table_idx] > 0, false)) {
TYPE cur2 = TYPE(0);
if (expectEXT(off.x > 0, true)) {
const ivec2 yoff_border2 = expectEXT(off.x == 1, false) ? ivec2(-1, -1) : ivec2(-2, 0);
cur2 = TYPE(imageLoad(pred, sp + LADDR(off + yoff_border2))[comp]);
}
base += quant_table[quant_table_idx][3][(cur2 - cur) & MAX_QUANT_TABLE_MASK];
/* top-2 became current upon swap */
TYPE top2 = TYPE(imageLoad(pred, sp + LADDR(off))[comp]);
base += quant_table[quant_table_idx][4][(top2 - top[1]) & MAX_QUANT_TABLE_MASK];
}
/* context, prediction */
return ivec2(base, predict(cur, VTYPE2(top)));
}
#else
ivec2 get_pred(readonly uimage2D pred, ivec2 sp, ivec2 off, int comp, int sw, uint8_t quant_table_idx)
{
const ivec2 yoff_border1 = off.x == 0 ? ivec2(1, -1) : ivec2(0, 0);
sp += off;
VTYPE3 top = VTYPE3(TYPE(0),
TYPE(0),
TYPE(0));
if (off.y > 0 && off != ivec2(0, 1))
top[0] = TYPE(imageLoad(pred, sp + ivec2(-1, -1) + yoff_border1)[comp]);
if (off.y > 0) {
top[1] = TYPE(imageLoad(pred, sp + ivec2(0, -1))[comp]);
top[2] = TYPE(imageLoad(pred, sp + ivec2(min(1, sw - off.x - 1), -1))[comp]);
}
TYPE cur = TYPE(0);
if (off != ivec2(0, 0))
cur = TYPE(imageLoad(pred, sp + ivec2(-1, 0) + yoff_border1)[comp]);
int base = quant_table[quant_table_idx][0][(cur - top[0]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][1][(top[0] - top[1]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][2][(top[1] - top[2]) & MAX_QUANT_TABLE_MASK];
if ((quant_table[quant_table_idx][3][127] != 0) ||
(quant_table[quant_table_idx][4][127] != 0)) {
TYPE cur2 = TYPE(0);
if (off.x > 0 && off != ivec2(1, 0)) {
const ivec2 yoff_border2 = off.x == 1 ? ivec2(1, -1) : ivec2(0, 0);
cur2 = TYPE(imageLoad(pred, sp + ivec2(-2, 0) + yoff_border2)[comp]);
}
base += quant_table[quant_table_idx][3][(cur2 - cur) & MAX_QUANT_TABLE_MASK];
TYPE top2 = TYPE(0);
if (off.y > 1)
top2 = TYPE(imageLoad(pred, sp + ivec2(0, -2))[comp]);
base += quant_table[quant_table_idx][4][(top2 - top[1]) & MAX_QUANT_TABLE_MASK];
}
/* context, prediction */
return ivec2(base, predict(cur, VTYPE2(top)));
}
#endif
#ifndef GOLOMB #ifndef GOLOMB
#ifdef CACHED_SYMBOL_READER #ifdef CACHED_SYMBOL_READER
shared uint8_t state[CONTEXT_SIZE]; shared uint8_t state[CONTEXT_SIZE];
@ -172,7 +85,7 @@ void decode_line(inout SliceContext sc, ivec2 sp, int w,
for (int x = 0; x < w; x++) { for (int x = 0; x < w; x++) {
ivec2 pr = get_pred(dec[p], sp, ivec2(x, y), 0, w, ivec2 pr = get_pred(dec[p], sp, ivec2(x, y), 0, w,
quant_table_idx); quant_table_idx, extend_lookup[quant_table_idx] > 0);
uint context_off = state_off + CONTEXT_SIZE*abs(pr[0]); uint context_off = state_off + CONTEXT_SIZE*abs(pr[0]);
#ifdef CACHED_SYMBOL_READER #ifdef CACHED_SYMBOL_READER
@ -217,7 +130,7 @@ void decode_line(inout SliceContext sc, ivec2 sp, int w,
ivec2 pos = sp + ivec2(x, y); ivec2 pos = sp + ivec2(x, y);
int diff; int diff;
ivec2 pr = get_pred(dec[p], sp, ivec2(x, y), 0, w, ivec2 pr = get_pred(dec[p], sp, ivec2(x, y), 0, w,
quant_table_idx); quant_table_idx, extend_lookup[quant_table_idx] > 0);
VlcState sb = VlcState(uint64_t(slice_state) + state_off + VLC_STATE_SIZE*abs(pr[0])); VlcState sb = VlcState(uint64_t(slice_state) + state_off + VLC_STATE_SIZE*abs(pr[0]));

155
libavcodec/vulkan/ffv1_enc.comp
View File

@ -20,48 +20,6 @@
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/ */
ivec2 get_pred(readonly uimage2D pred, ivec2 sp, ivec2 off, int comp, int sw, uint8_t quant_table_idx)
{
const ivec2 yoff_border1 = off.x == 0 ? ivec2(1, -1) : ivec2(0, 0);
sp += off;
VTYPE3 top = VTYPE3(TYPE(0),
TYPE(0),
TYPE(0));
if (off.y > 0 && off != ivec2(0, 1))
top[0] = TYPE(imageLoad(pred, sp + ivec2(-1, -1) + yoff_border1)[comp]);
if (off.y > 0) {
top[1] = TYPE(imageLoad(pred, sp + ivec2(0, -1))[comp]);
top[2] = TYPE(imageLoad(pred, sp + ivec2(min(1, sw - off.x - 1), -1))[comp]);
}
TYPE cur = TYPE(0);
if (off != ivec2(0, 0))
cur = TYPE(imageLoad(pred, sp + ivec2(-1, 0) + yoff_border1)[comp]);
int base = quant_table[quant_table_idx][0][(cur - top[0]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][1][(top[0] - top[1]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][2][(top[1] - top[2]) & MAX_QUANT_TABLE_MASK];
if ((quant_table[quant_table_idx][3][127] != 0) ||
(quant_table[quant_table_idx][4][127] != 0)) {
TYPE cur2 = TYPE(0);
if (off.x > 0 && off != ivec2(1, 0)) {
const ivec2 yoff_border2 = off.x == 1 ? ivec2(1, -1) : ivec2(0, 0);
cur2 = TYPE(imageLoad(pred, sp + ivec2(-2, 0) + yoff_border2)[comp]);
}
base += quant_table[quant_table_idx][3][(cur2 - cur) & MAX_QUANT_TABLE_MASK];
TYPE top2 = TYPE(0);
if (off.y > 1)
top2 = TYPE(imageLoad(pred, sp + ivec2(0, -2))[comp]);
base += quant_table[quant_table_idx][4][(top2 - top[1]) & MAX_QUANT_TABLE_MASK];
}
/* context, prediction */
return ivec2(base, predict(cur, VTYPE2(top)));
}
#ifndef GOLOMB #ifndef GOLOMB
/* Note - only handles signed values */ /* Note - only handles signed values */
void put_symbol(inout RangeCoder c, uint64_t state, int v) void put_symbol(inout RangeCoder c, uint64_t state, int v)
@ -86,38 +44,42 @@ void put_symbol(inout RangeCoder c, uint64_t state, int v)
put_rac(c, state - 11 + min(e, 10), v < 0); put_rac(c, state - 11 + min(e, 10), v < 0);
} }
void encode_line_pcm(inout SliceContext sc, int y, int p, int comp, void encode_line_pcm(inout SliceContext sc, readonly uimage2D img,
int bits) ivec2 sp, int y, int p, int comp, int bits)
{ {
ivec2 sp = sc.slice_pos;
int w = sc.slice_dim.x; int w = sc.slice_dim.x;
#ifndef RGB
if (p > 0 && p < 3) { if (p > 0 && p < 3) {
w >>= chroma_shift.x; w >>= chroma_shift.x;
sp >>= chroma_shift; sp >>= chroma_shift;
} }
#endif
for (int x = 0; x < w; x++) { for (int x = 0; x < w; x++) {
uint v = imageLoad(src[p], (sp + ivec2(x, y)))[comp]; uint v = imageLoad(img, sp + LADDR(ivec2(x, y)))[comp];
for (int i = (bits - 1); i >= 0; i--) for (int i = (bits - 1); i >= 0; i--)
put_rac_equi(sc.c, bool(bitfieldExtract(v, i, 1))); put_rac_equi(sc.c, bool(bitfieldExtract(v, i, 1)));
} }
} }
void encode_line(inout SliceContext sc, uint64_t state, void encode_line(inout SliceContext sc, readonly uimage2D img, uint64_t state,
int y, int p, int comp, int bits, ivec2 sp, int y, int p, int comp, int bits,
uint8_t quant_table_idx, const int run_index) uint8_t quant_table_idx, const int run_index)
{ {
ivec2 sp = sc.slice_pos;
int w = sc.slice_dim.x; int w = sc.slice_dim.x;
#ifndef RGB
if (p > 0 && p < 3) { if (p > 0 && p < 3) {
w >>= chroma_shift.x; w >>= chroma_shift.x;
sp >>= chroma_shift; sp >>= chroma_shift;
} }
#endif
for (int x = 0; x < w; x++) { for (int x = 0; x < w; x++) {
ivec2 d = get_pred(src[p], sp, ivec2(x, y), comp, w, quant_table_idx); ivec2 d = get_pred(img, sp, ivec2(x, y), comp, w,
d[1] = int(imageLoad(src[p], sp + ivec2(x, y))[comp]) - d[1]; quant_table_idx, extend_lookup[quant_table_idx] > 0);
d[1] = int(imageLoad(img, sp + LADDR(ivec2(x, y)))[comp]) - d[1];
if (d[0] < 0) if (d[0] < 0)
d = -d; d = -d;
@ -130,24 +92,26 @@ void encode_line(inout SliceContext sc, uint64_t state,
#else /* GOLOMB */ #else /* GOLOMB */
void encode_line(inout SliceContext sc, uint64_t state, void encode_line(inout SliceContext sc, readonly uimage2D img, uint64_t state,
int y, int p, int comp, int bits, ivec2 sp, int y, int p, int comp, int bits,
uint8_t quant_table_idx, inout int run_index) uint8_t quant_table_idx, inout int run_index)
{ {
ivec2 sp = sc.slice_pos;
int w = sc.slice_dim.x; int w = sc.slice_dim.x;
#ifndef RGB
if (p > 0 && p < 3) { if (p > 0 && p < 3) {
w >>= chroma_shift.x; w >>= chroma_shift.x;
sp >>= chroma_shift; sp >>= chroma_shift;
} }
#endif
int run_count = 0; int run_count = 0;
bool run_mode = false; bool run_mode = false;
for (int x = 0; x < w; x++) { for (int x = 0; x < w; x++) {
ivec2 d = get_pred(src[p], sp, ivec2(x, y), comp, w, quant_table_idx); ivec2 d = get_pred(img, sp, ivec2(x, y), comp, w,
d[1] = int(imageLoad(src[p], sp + ivec2(x, y))[comp]) - d[1]; quant_table_idx, extend_lookup[quant_table_idx] > 0);
d[1] = int(imageLoad(img, sp + LADDR(ivec2(x, y)))[comp]) - d[1];
if (d[0] < 0) if (d[0] < 0)
d = -d; d = -d;
@ -198,14 +162,56 @@ void encode_line(inout SliceContext sc, uint64_t state,
} }
#endif #endif
#ifdef RGB
ivec4 load_components(ivec2 pos)
{
ivec4 pix = ivec4(imageLoad(src[0], pos));
if (planar_rgb != 0) {
for (int i = 1; i < (3 + transparency); i++)
pix[i] = int(imageLoad(src[i], pos)[0]);
}
return ivec4(pix[fmt_lut[0]], pix[fmt_lut[1]],
pix[fmt_lut[2]], pix[fmt_lut[3]]);
}
void transform_sample(inout ivec4 pix, ivec2 rct_coef)
{
pix.b -= pix.g;
pix.r -= pix.g;
pix.g += (pix.r*rct_coef.x + pix.b*rct_coef.y) >> 2;
pix.b += rct_offset;
pix.r += rct_offset;
}
void preload_rgb(in SliceContext sc, ivec2 sp, int w, int y, bool apply_rct)
{
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);
ivec4 pix = load_components(pos);
if (expectEXT(apply_rct, true))
transform_sample(pix, sc.slice_rct_coef);
imageStore(tmp, lpos, pix);
}
}
#endif
void encode_slice(inout SliceContext sc, const uint slice_idx) void encode_slice(inout SliceContext sc, const uint slice_idx)
{ {
ivec2 sp = sc.slice_pos;
#ifndef RGB #ifndef RGB
int bits = bits_per_raw_sample; int bits = bits_per_raw_sample;
#else #else
int bits = 9; int bits = 9;
if (bits != 8 || sc.slice_coding_mode != 0) if (bits != 8 || sc.slice_coding_mode != 0)
bits = bits_per_raw_sample + int(sc.slice_coding_mode != 1); bits = bits_per_raw_sample + int(sc.slice_coding_mode != 1);
sp.y = int(gl_WorkGroupID.y)*RGB_LINECACHE;
#endif #endif
#ifndef GOLOMB #ifndef GOLOMB
@ -222,15 +228,17 @@ void encode_slice(inout SliceContext sc, const uint slice_idx)
int comp = c - p; int comp = c - p;
for (int y = 0; y < h; y++) for (int y = 0; y < h; y++)
encode_line_pcm(sc, y, p, comp, bits); encode_line_pcm(sc, src[p], sp, y, p, comp, bits);
} }
#else #else
for (int y = 0; y < sc.slice_dim.y; y++) { for (int y = 0; y < sc.slice_dim.y; y++) {
encode_line_pcm(sc, y, 0, 1, bits); preload_rgb(sc, sp, sc.slice_dim.x, y, false);
encode_line_pcm(sc, y, 0, 2, bits);
encode_line_pcm(sc, y, 0, 0, bits); encode_line_pcm(sc, tmp, sp, y, 0, 1, bits);
encode_line_pcm(sc, tmp, sp, y, 0, 2, bits);
encode_line_pcm(sc, tmp, sp, y, 0, 0, bits);
if (transparency == 1) if (transparency == 1)
encode_line_pcm(sc, y, 0, 3, bits); encode_line_pcm(sc, tmp, sp, y, 0, 3, bits);
} }
#endif #endif
} else } else
@ -252,7 +260,8 @@ void encode_slice(inout SliceContext sc, const uint slice_idx)
int comp = c - p; int comp = c - p;
for (int y = 0; y < h; y++) for (int y = 0; y < h; y++)
encode_line(sc, slice_state_off, y, p, comp, bits, quant_table_idx[c], run_index); encode_line(sc, src[p], slice_state_off, sp, y, p,
comp, bits, quant_table_idx[c], run_index);
/* For the second chroma plane, reuse the first plane's state */ /* For the second chroma plane, reuse the first plane's state */
if (c != 1) if (c != 1)
@ -261,15 +270,17 @@ void encode_slice(inout SliceContext sc, const uint slice_idx)
#else #else
int run_index = 0; int run_index = 0;
for (int y = 0; y < sc.slice_dim.y; y++) { for (int y = 0; y < sc.slice_dim.y; y++) {
encode_line(sc, slice_state_off + plane_state_size*0, preload_rgb(sc, sp, sc.slice_dim.x, y, true);
y, 0, 1, bits, quant_table_idx[0], run_index);
encode_line(sc, slice_state_off + plane_state_size*1, encode_line(sc, tmp, slice_state_off + plane_state_size*0,
y, 0, 2, bits, quant_table_idx[1], run_index); sp, y, 0, 1, bits, quant_table_idx[0], run_index);
encode_line(sc, slice_state_off + plane_state_size*1, encode_line(sc, tmp, slice_state_off + plane_state_size*1,
y, 0, 0, bits, quant_table_idx[2], run_index); sp, y, 0, 2, bits, quant_table_idx[1], run_index);
encode_line(sc, tmp, slice_state_off + plane_state_size*1,
sp, y, 0, 0, bits, quant_table_idx[2], run_index);
if (transparency == 1) if (transparency == 1)
encode_line(sc, slice_state_off + plane_state_size*2, encode_line(sc, tmp, slice_state_off + plane_state_size*2,
y, 0, 3, bits, quant_table_idx[3], run_index); sp, y, 0, 3, bits, quant_table_idx[3], run_index);
} }
#endif #endif
} }

5
libavcodec/vulkan_ffv1.c
View File

@ -26,6 +26,8 @@
#include "libavutil/vulkan_spirv.h" #include "libavutil/vulkan_spirv.h"
#include "libavutil/mem.h" #include "libavutil/mem.h"
#define RGB_LINECACHE 2
extern const char *ff_source_common_comp; extern const char *ff_source_common_comp;
extern const char *ff_source_rangecoder_comp; extern const char *ff_source_rangecoder_comp;
extern const char *ff_source_ffv1_vlc_comp; extern const char *ff_source_ffv1_vlc_comp;
@ -610,6 +612,7 @@ static void define_shared_code(FFVulkanShader *shd, int use32bit)
GLSLC(0, #define DECODE ); GLSLC(0, #define DECODE );
av_bprintf(&shd->src, "#define RGB_LINECACHE %i\n" ,RGB_LINECACHE);
av_bprintf(&shd->src, "#define CONTEXT_SIZE %i\n" ,CONTEXT_SIZE); av_bprintf(&shd->src, "#define CONTEXT_SIZE %i\n" ,CONTEXT_SIZE);
av_bprintf(&shd->src, "#define MAX_QUANT_TABLE_MASK 0x%x\n" ,MAX_QUANT_TABLE_MASK); av_bprintf(&shd->src, "#define MAX_QUANT_TABLE_MASK 0x%x\n" ,MAX_QUANT_TABLE_MASK);
@ -936,7 +939,7 @@ static int init_indirect(AVCodecContext *avctx, FFVulkanContext *s,
frames_ctx->format = AV_PIX_FMT_VULKAN; frames_ctx->format = AV_PIX_FMT_VULKAN;
frames_ctx->sw_format = sw_format; frames_ctx->sw_format = sw_format;
frames_ctx->width = s->frames->width; frames_ctx->width = s->frames->width;
frames_ctx->height = f->num_v_slices*2; frames_ctx->height = f->num_v_slices*RGB_LINECACHE;
vk_frames = frames_ctx->hwctx; vk_frames = frames_ctx->hwctx;
vk_frames->tiling = VK_IMAGE_TILING_OPTIMAL; vk_frames->tiling = VK_IMAGE_TILING_OPTIMAL;