From: IndecisiveTurtle <geoste...@gmail.com>
Performance wise, encoding a 1080p 1-minute video is performed in about 2.5
minutes with the cpu encoder running on my Ryzen 5 4600H, while it takes about
30 seconds on my NVIDIA GTX 1650
Haar shader has a subgroup optimized variant that applies when configured
wavelet depth allows it
---
configure | 1 +
libavcodec/Makefile | 3 +
libavcodec/allcodecs.c | 1 +
libavcodec/vc2enc_vulkan.c | 959 +++++++++++++++++++
libavcodec/vulkan/vc2_dwt_haar.comp | 82 ++
libavcodec/vulkan/vc2_dwt_haar_subgroup.comp | 75 ++
libavcodec/vulkan/vc2_dwt_hor_legall.comp | 82 ++
libavcodec/vulkan/vc2_dwt_upload.comp | 96 ++
libavcodec/vulkan/vc2_dwt_ver_legall.comp | 78 ++
libavcodec/vulkan/vc2_encode.comp | 169 ++++
libavcodec/vulkan/vc2_slice_sizes.comp | 170 ++++
11 files changed, 1716 insertions(+)
create mode 100644 libavcodec/vc2enc_vulkan.c
create mode 100644 libavcodec/vulkan/vc2_dwt_haar.comp
create mode 100644 libavcodec/vulkan/vc2_dwt_haar_subgroup.comp
create mode 100644 libavcodec/vulkan/vc2_dwt_hor_legall.comp
create mode 100644 libavcodec/vulkan/vc2_dwt_upload.comp
create mode 100644 libavcodec/vulkan/vc2_dwt_ver_legall.comp
create mode 100644 libavcodec/vulkan/vc2_encode.comp
create mode 100644 libavcodec/vulkan/vc2_slice_sizes.comp
diff --git a/configure b/configure
index c94b8eac43..16ee163b05 100755
--- a/configure
+++ b/configure
@@ -3130,6 +3130,7 @@ utvideo_encoder_select="bswapdsp huffman llvidencdsp"
vble_decoder_select="llviddsp"
vbn_decoder_select="texturedsp"
vbn_encoder_select="texturedspenc"
+vc2_vulkan_encoder_select="vulkan spirv_compiler"
vmix_decoder_select="idctdsp"
vc1_decoder_select="blockdsp h264qpel intrax8 mpegvideodec qpeldsp vc1dsp"
vc1image_decoder_select="vc1_decoder"
diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index d4ebd86866..79505f8ef1 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -770,6 +770,9 @@ OBJS-$(CONFIG_VC1_MMAL_DECODER) += mmaldec.o
OBJS-$(CONFIG_VC1_QSV_DECODER) += qsvdec.o
OBJS-$(CONFIG_VC1_V4L2M2M_DECODER) += v4l2_m2m_dec.o
OBJS-$(CONFIG_VC2_ENCODER) += vc2enc.o vc2enc_dwt.o
vc2enc_common.o diractab.o
+OBJS-$(CONFIG_VC2_VULKAN_ENCODER) += vc2enc_vulkan.o vulkan/vc2_encode.o
vulkan/vc2_slice_sizes.o \
+ vulkan/vc2_dwt_hor_legall.o
vulkan/vc2_dwt_ver_legall.o \
+ vulkan/vc2_dwt_upload.o
vulkan/vc2_dwt_haar.o vulkan/vc2_dwt_haar_subgroup.o
OBJS-$(CONFIG_VCR1_DECODER) += vcr1.o
OBJS-$(CONFIG_VMDAUDIO_DECODER) += vmdaudio.o
OBJS-$(CONFIG_VMDVIDEO_DECODER) += vmdvideo.o
diff --git a/libavcodec/allcodecs.c b/libavcodec/allcodecs.c
index f10519617e..054b0d958b 100644
--- a/libavcodec/allcodecs.c
+++ b/libavcodec/allcodecs.c
@@ -365,6 +365,7 @@ extern const FFCodec ff_vc1image_decoder;
extern const FFCodec ff_vc1_mmal_decoder;
extern const FFCodec ff_vc1_qsv_decoder;
extern const FFCodec ff_vc1_v4l2m2m_decoder;
+extern const FFCodec ff_vc2_vulkan_encoder;
extern const FFCodec ff_vc2_encoder;
extern const FFCodec ff_vcr1_decoder;
extern const FFCodec ff_vmdvideo_decoder;
diff --git a/libavcodec/vc2enc_vulkan.c b/libavcodec/vc2enc_vulkan.c
new file mode 100644
index 0000000000..d90d65e36d
--- /dev/null
+++ b/libavcodec/vc2enc_vulkan.c
@@ -0,0 +1,959 @@
+/*
+ * Copyright (C) 2025 raphaelthegreat <geoste...@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/avassert.h"
+#include "libavutil/mem.h"
+#include "libavutil/pixdesc.h"
+#include "libavutil/opt.h"
+#include "libavutil/thread.h"
+#include "libavutil/version.h"
+#include "libavutil/vulkan_spirv.h"
+#include "libavutil/hwcontext_vulkan.h"
+#include "libavutil/vulkan_loader.h"
+#include "libavutil/vulkan.h"
+#include "codec_internal.h"
+#include "internal.h"
+#include "encode.h"
+#include "version.h"
+#include "vc2enc_common.h"
+#include "hwconfig.h"
+
+#define LEGALL_TILE_DIM 16
+#define LEGALL_WORKGROUP_X 64
+#define SLICE_WORKGROUP_X 128
+
+extern const char *ff_source_common_comp;
+extern const char *ff_source_vc2_encode_comp;
+extern const char *ff_source_vc2_dwt_hor_legall_comp;
+extern const char *ff_source_vc2_dwt_ver_legall_comp;
+extern const char *ff_source_vc2_slice_sizes_comp;
+extern const char *ff_source_vc2_dwt_upload_comp;
+extern const char *ff_source_vc2_dwt_haar_comp;
+extern const char *ff_source_vc2_dwt_haar_subgroup_comp;
+
+typedef struct VC2DwtPushData {
+ int s;
+ union {
+ int diff_offset;
+ int plane_idx;
+ };
+ int level;
+} VC2DwtPushData;
+
+typedef struct VC2EncAuxData {
+ int quant[MAX_DWT_LEVELS][4];
+ int ff_dirac_qscale_tab[116];
+} VC2EncAuxData;
+
+typedef struct VC2EncPushData {
+ VkDeviceAddress pb;
+ int num_x;
+ int num_y;
+ int wavelet_depth;
+ int size_scaler;
+ int prefix_bytes;
+} VC2EncPushData;
+
+typedef struct VC2EncSliceArgs {
+ int quant_idx;
+ int bytes;
+ int pb_start;
+ int pad;
+} VC2EncSliceArgs;
+
+typedef struct VC2EncSliceCalcPushData {
+ int num_x;
+ int num_y;
+ int wavelet_depth;
+ int size_scaler;
+ int prefix_bytes;
+ int bits_ceil;
+ int bits_floor;
+} VC2EncSliceCalcPushData;
+
+typedef struct VC2EncVulkanContext {
+ VC2EncContext base;
+ FFVkBuffer lut_buf;
+ FFVkBuffer slice_buf;
+ VC2EncSliceArgs *slice_args;
+
+ /* Vulkan state */
+ FFVulkanContext vkctx;
+ AVVulkanDeviceQueueFamily *qf;
+ FFVkExecPool e;
+
+ FFVulkanShader dwt_haar_shd;
+ FFVulkanShader dwt_upload_shd;
+ FFVulkanShader dwt_hor_shd, dwt_ver_shd;
+ FFVulkanShader slice_shd;
+ FFVulkanShader enc_shd;
+ AVBufferPool* dwt_buf_pool;
+ int haar_subgroup;
+
+ VkBuffer plane_buf;
+ VC2EncPushData enc_consts;
+ VC2DwtPushData dwt_consts;
+ VC2EncSliceCalcPushData calc_consts;
+
+ /* Intermediate frame pool */
+ AVBufferRef *intermediate_frames_ref[3];
+ AVFrame *intermediate_frame[AV_NUM_DATA_POINTERS];
+ VkImageView intermediate_views[AV_NUM_DATA_POINTERS];
+} VC2EncVulkanContext;
+
+static int init_vulkan_pipeline(VC2EncVulkanContext* s, FFVkSPIRVCompiler *spv,
+ FFVulkanShader* shd, int push_size,
+ int lg_x, int lg_y, int lg_z,
+ const char* pl_name, const char* pl_source,
+ int start_desc, int num_desc)
+{
+ int err = 0;
+ uint8_t *spv_data;
+ size_t spv_len;
+ void *spv_opaque = NULL;
+ FFVulkanContext *vkctx = &s->vkctx;
+ FFVulkanDescriptorSetBinding *desc;
+
+ ff_vk_shader_init(vkctx, shd, pl_name, VK_SHADER_STAGE_COMPUTE_BIT,
+ NULL, 0, lg_x, lg_y, lg_z, 0);
+
+ av_bprintf(&shd->src, "struct SliceArgs {int quant_idx;int bytes;int
pb_start;int pad;};\n");
+
+ desc = (FFVulkanDescriptorSetBinding []) {
+ {
+ .name = "src_planes",
+ .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+ .mem_layout = ff_vk_shader_rep_fmt(vkctx->frames->sw_format,
FF_VK_REP_UINT),
+ .dimensions = 2,
+ .elems = av_pix_fmt_count_planes(vkctx->frames->sw_format),
+ .stages = VK_SHADER_STAGE_COMPUTE_BIT,
+ },
+ {
+ .name = "coef_buf",
+ .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+ .mem_layout = "r32i",
+ .dimensions = 2,
+ .elems = 3,
+ .stages = VK_SHADER_STAGE_COMPUTE_BIT,
+ },
+ {
+ .name = "AuxData",
+ .type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
+ .stages = VK_SHADER_STAGE_COMPUTE_BIT,
+ .mem_layout = "scalar",
+ .buf_content = "int lut_quant[5][4]; int
ff_dirac_qscale_tab[116];",
+ },
+ {
+ .name = "SliceBuffer",
+ .type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+ .stages = VK_SHADER_STAGE_COMPUTE_BIT,
+ .mem_layout = "scalar",
+ .buf_content = "SliceArgs slice_args[];",
+ },
+ };
+ RET(ff_vk_shader_add_descriptor_set(vkctx, shd, desc + start_desc,
num_desc, 0, 0));
+
+ ff_vk_shader_add_push_const(shd, 0, push_size,
VK_SHADER_STAGE_COMPUTE_BIT);
+ av_bprintf(&shd->src, "#define PB_UNALIGNED\n");
+ av_bprintf(&shd->src, "#define PLANE_FMT %d\n", vkctx->frames->sw_format);
+ GLSLD(ff_source_common_comp);
+ GLSLD(pl_source);
+
+ /* Compile Haar shader */
+ RET(spv->compile_shader(vkctx, spv, shd, &spv_data, &spv_len, "main",
&spv_opaque));
+ RET(ff_vk_shader_link(vkctx, shd, spv_data, spv_len, "main"));
+ RET(ff_vk_shader_register_exec(vkctx, &s->e, shd));
+
+fail:
+ return err;
+}
+
+static int init_frame_pools(AVCodecContext *avctx)
+{
+ int i, err = 0;
+ VC2EncVulkanContext *sv = avctx->priv_data;
+ AVHWFramesContext *frames_ctx;
+ AVVulkanFramesContext *vk_frames;
+ enum AVPixelFormat sw_format = AV_PIX_FMT_GRAY32;
+
+ for (i = 0; i < 3; i++) {
+ sv->intermediate_frames_ref[i] =
av_hwframe_ctx_alloc(sv->vkctx.device_ref);
+ if (!sv->intermediate_frames_ref[i])
+ return AVERROR(ENOMEM);
+
+ frames_ctx = (AVHWFramesContext *)sv->intermediate_frames_ref[i]->data;
+ frames_ctx->format = AV_PIX_FMT_VULKAN;
+ frames_ctx->sw_format = sw_format;
+ frames_ctx->width = sv->base.plane[i].dwt_width;
+ frames_ctx->height = sv->base.plane[i].dwt_height;
+
+ vk_frames = frames_ctx->hwctx;
+ vk_frames->tiling = VK_IMAGE_TILING_OPTIMAL;
+ vk_frames->usage = VK_IMAGE_USAGE_STORAGE_BIT;
+ vk_frames->img_flags = VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
+
+ err = av_hwframe_ctx_init(sv->intermediate_frames_ref[i]);
+ if (err < 0) {
+ av_log(avctx, AV_LOG_ERROR, "Unable to initialize frame pool with
format %s: %s\n",
+ av_get_pix_fmt_name(sw_format), av_err2str(err));
+ av_buffer_unref(&sv->intermediate_frames_ref[i]);
+ return err;
+ }
+ }
+
+ return err;
+}
+
+static int vc2_init_vulkan(AVCodecContext *avctx)
+{
+ VC2EncVulkanContext *sv = avctx->priv_data;
+ VC2EncContext *s = &sv->base;
+ FFVulkanContext *vkctx = &sv->vkctx;
+ FFVkSPIRVCompiler *spv;
+ VC2EncAuxData *ad = NULL;
+ int err = 0;
+ unsigned int subgroup_size = vkctx->subgroup_props.maxSubgroupSize;
+
+ /* Initialize spirv compiler */
+ spv = ff_vk_spirv_init();
+ if (!spv) {
+ av_log(avctx, AV_LOG_ERROR, "Unable to initialize SPIR-V compiler!\n");
+ return -1;
+ }
+
+ ff_vk_exec_pool_init(vkctx, sv->qf, &sv->e, 1, 0, 0, 0, NULL);
+
+ /* Initialize Haar push data */
+ sv->dwt_consts.diff_offset = s->diff_offset;
+ sv->dwt_consts.s = s->wavelet_idx == VC2_TRANSFORM_HAAR_S ? 1 : 0;
+ sv->dwt_consts.level = 0;
+
+ /* Initializer slice calculation push data */
+ sv->calc_consts.num_x = s->num_x;
+ sv->calc_consts.num_y = s->num_y;
+ sv->calc_consts.wavelet_depth = s->wavelet_depth;
+ sv->calc_consts.prefix_bytes = s->prefix_bytes;
+
+ /* Initialize encoder push data */
+ sv->enc_consts.wavelet_depth = s->wavelet_depth;
+ sv->enc_consts.num_x = s->num_x;
+ sv->enc_consts.num_y = s->num_y;
+
+ /* Create buffer for encoder auxilary data. */
+ RET(ff_vk_create_buf(vkctx, &sv->lut_buf, sizeof(VC2EncAuxData), NULL,
NULL,
+ VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
+ VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT,
+ VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
+ VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT));
+ RET(ff_vk_map_buffer(vkctx, &sv->lut_buf, (void *)&ad, 0));
+ vc2_init_quant_matrix(s, ad->quant);
+ memcpy(ad->ff_dirac_qscale_tab, ff_dirac_qscale_tab,
sizeof(ff_dirac_qscale_tab));
+ RET(ff_vk_unmap_buffer(vkctx, &sv->lut_buf, 1));
+
+ /* Create buffer for encoder auxilary data. */
+ RET(ff_vk_create_buf(vkctx, &sv->slice_buf,
+ sizeof(VC2EncSliceArgs) * s->num_x * s->num_y,
+ NULL, NULL,
+ VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
+ VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT,
+ VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
+ VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT));
+ RET(ff_vk_map_buffer(vkctx, &sv->slice_buf, (void *)&sv->slice_args, 0));
+ memset(sv->slice_args, 0, sv->slice_buf.size);
+
+ /* Initialize intermediate frame pool. */
+ RET(init_frame_pools(avctx));
+
+ /* Initialize encoding pipelines */
+ init_vulkan_pipeline(sv, spv, &sv->dwt_upload_shd, sizeof(VC2DwtPushData),
+ 8, 8, 1, "dwt_upload_pl",
ff_source_vc2_dwt_upload_comp, 0, 2);
+ init_vulkan_pipeline(sv, spv, &sv->slice_shd, sizeof(VC2EncPushData),
+ SLICE_WORKGROUP_X, 1, 1, "slice_pl",
ff_source_vc2_slice_sizes_comp, 1, 3);
+ init_vulkan_pipeline(sv, spv, &sv->enc_shd, sizeof(VC2EncPushData),
+ SLICE_WORKGROUP_X, 1, 1, "enc_pl",
ff_source_vc2_encode_comp, 1, 3);
+ sv->haar_subgroup = 0;
+
+ if (s->wavelet_idx == VC2_TRANSFORM_HAAR || s->wavelet_idx ==
VC2_TRANSFORM_HAAR_S) {
+ if (subgroup_size == 32 && s->wavelet_depth < 3) {
+ init_vulkan_pipeline(sv, spv, &sv->dwt_haar_shd,
sizeof(VC2DwtPushData),
+ 64, 1, 1, "dwt_haar_pl",
ff_source_vc2_dwt_haar_subgroup_comp, 1, 1);
+ sv->haar_subgroup = 1;
+ } else if (subgroup_size == 64 && s->wavelet_depth < 4) {
+ init_vulkan_pipeline(sv, spv, &sv->dwt_haar_shd,
sizeof(VC2DwtPushData),
+ 64, 1, 1, "dwt_haar_pl",
ff_source_vc2_dwt_haar_subgroup_comp, 1, 1);
+ sv->haar_subgroup = 1;
+ } else {
+ init_vulkan_pipeline(sv, spv, &sv->dwt_haar_shd,
sizeof(VC2DwtPushData),
+ 32, 32, 1, "dwt_haar_pl",
ff_source_vc2_dwt_haar_comp, 1, 1);
+ }
+ } else if (s->wavelet_idx == VC2_TRANSFORM_5_3) {
+ init_vulkan_pipeline(sv, spv, &sv->dwt_hor_shd, sizeof(VC2DwtPushData),
+ LEGALL_WORKGROUP_X, 1, 1, "dwt_hor_pl",
ff_source_vc2_dwt_hor_legall_comp, 1, 1);
+ init_vulkan_pipeline(sv, spv, &sv->dwt_ver_shd, sizeof(VC2DwtPushData),
+ LEGALL_WORKGROUP_X, 1, 1, "dwt_ver_pl",
ff_source_vc2_dwt_ver_legall_comp, 1, 1);
+ }
+
+fail:
+ return err;
+}
+
+static void vulkan_bind_img_planes(FFVulkanContext *s, FFVkExecContext *e,
+ FFVulkanShader *shd, VkImageView *views,
+ int set, int binding)
+{
+ for (int i = 0; i < 3; i++)
+ ff_vk_shader_update_img(s, e, shd, set, binding, i,
+ views[i], VK_IMAGE_LAYOUT_GENERAL,
+ VK_NULL_HANDLE);
+}
+
+static void dwt_plane_haar(VC2EncVulkanContext *s, FFVkExecContext *exec,
+ VkImageMemoryBarrier2* img_bar, int nb_img_bar)
+{
+ int p, group_x, group_y;
+ FFVulkanContext *vkctx = &s->vkctx;
+ FFVulkanFunctions *vk = &vkctx->vkfn;
+ Plane* plane;
+
+ s->dwt_consts.level = s->base.wavelet_depth;
+ vulkan_bind_img_planes(vkctx, exec, &s->dwt_haar_shd,
s->intermediate_views, 0, 0);
+ ff_vk_exec_bind_shader(vkctx, exec, &s->dwt_haar_shd);
+
+ /* Haar pass */
+ for (p = 0; p < 3; p++) {
+ plane = &s->base.plane[p];
+ s->dwt_consts.plane_idx = p;
+ if (s->haar_subgroup) {
+ group_x = FFALIGN(plane->dwt_width, 8) >> 3;
+ group_y = FFALIGN(plane->dwt_height, 8) >> 3;
+ } else {
+ group_x = FFALIGN(plane->dwt_width, 32) >> 5;
+ group_y = FFALIGN(plane->dwt_height, 32) >> 5;
+ }
+
+ ff_vk_shader_update_push_const(vkctx, exec, &s->dwt_haar_shd,
VK_SHADER_STAGE_COMPUTE_BIT,
+ 0, sizeof(VC2DwtPushData),
&s->dwt_consts);
+ vk->CmdDispatch(exec->buf, group_x, group_y, 1);
+ }
+
+ /* Wait for haar dispatches to complete */
+ vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
+ .sType =
VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
+ .pImageMemoryBarriers = img_bar,
+ .imageMemoryBarrierCount =
nb_img_bar,
+ });
+}
+
+static void dwt_plane_legall(VC2EncVulkanContext *s, FFVkExecContext *exec,
+ VkImageMemoryBarrier2* img_bar, int nb_img_bar)
+{
+ FFVulkanContext *vkctx = &s->vkctx;
+ FFVulkanFunctions *vk = &vkctx->vkfn;
+ int legall_group_x = (s->base.plane[0].dwt_height + LEGALL_WORKGROUP_X -
1) >> 6;
+ int legall_group_y = (s->base.plane[0].dwt_width + LEGALL_WORKGROUP_X - 1)
>> 6;
+ int i;
+
+ /* Perform legall wavelet trasform */
+ for (i = 0; i < s->base.wavelet_depth; i++) {
+ s->dwt_consts.level = i;
+
+ /* Horizontal legall pass */
+ vulkan_bind_img_planes(vkctx, exec, &s->dwt_hor_shd,
s->intermediate_views, 0, 0);
+ ff_vk_exec_bind_shader(vkctx, exec, &s->dwt_hor_shd);
+ ff_vk_shader_update_push_const(vkctx, exec, &s->dwt_hor_shd,
VK_SHADER_STAGE_COMPUTE_BIT,
+ 0, sizeof(VC2DwtPushData),
&s->dwt_consts);
+ vk->CmdDispatch(exec->buf, legall_group_x, 1, 3);
+ vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
+ .sType =
VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
+ .pImageMemoryBarriers = img_bar,
+ .imageMemoryBarrierCount =
nb_img_bar,
+ });
+
+ /* Vertical legall pass */
+ vulkan_bind_img_planes(vkctx, exec, &s->dwt_ver_shd,
s->intermediate_views, 0, 0);
+ ff_vk_exec_bind_shader(vkctx, exec, &s->dwt_ver_shd);
+ ff_vk_shader_update_push_const(vkctx, exec, &s->dwt_ver_shd,
VK_SHADER_STAGE_COMPUTE_BIT,
+ 0, sizeof(VC2DwtPushData),
&s->dwt_consts);
+ vk->CmdDispatch(exec->buf, legall_group_y, 1, 3);
+ vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
+ .sType =
VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
+ .pImageMemoryBarriers = img_bar,
+ .imageMemoryBarrierCount =
nb_img_bar,
+ });
+ }
+}
+
+static int vulkan_dwt_plane(VC2EncVulkanContext *s, FFVkExecContext *exec,
AVFrame *frame)
+{
+ int i, err = 0, nb_img_bar = 0;
+ int wavelet_idx = s->base.wavelet_idx;
+ int group_x = s->base.plane[0].dwt_width >> 3;
+ int group_y = s->base.plane[0].dwt_height >> 3;
+ FFVulkanContext *vkctx = &s->vkctx;
+ FFVulkanFunctions *vk = &vkctx->vkfn;
+ VkImageView views[AV_NUM_DATA_POINTERS];
+ VkImageMemoryBarrier2 img_bar[AV_NUM_DATA_POINTERS];
+
+ /* Generate barriers and image views for frame images. */
+ RET(ff_vk_exec_add_dep_frame(vkctx, exec, frame,
+ VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
+ VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT));
+ RET(ff_vk_create_imageviews(vkctx, exec, views, frame, FF_VK_REP_UINT));
+ ff_vk_frame_barrier(vkctx, exec, 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);
+
+ /* Submit the image barriers. */
+ vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
+ .sType =
VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
+ .pImageMemoryBarriers = img_bar,
+ .imageMemoryBarrierCount =
nb_img_bar,
+ });
+
+ /* Create a temporaty frames */
+ nb_img_bar = 0;
+ for (i = 0; i < 3; i++) {
+ s->intermediate_frame[i] = av_frame_alloc();
+ if (!s->intermediate_frame[i])
+ return AVERROR(ENOMEM);
+
+ RET(av_hwframe_get_buffer(s->intermediate_frames_ref[i],
+ s->intermediate_frame[i], 0));
+ RET(ff_vk_exec_add_dep_frame(vkctx, exec, s->intermediate_frame[i],
+ VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
+ VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT));
+ RET(ff_vk_create_imageviews(vkctx, exec, &s->intermediate_views[i],
+ s->intermediate_frame[i], FF_VK_REP_INT));
+ ff_vk_frame_barrier(vkctx, exec, s->intermediate_frame[i], 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);
+ }
+
+ /* Submit the image barriers. */
+ vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
+ .sType =
VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
+ .pImageMemoryBarriers = img_bar,
+ .imageMemoryBarrierCount =
nb_img_bar,
+ });
+
+ /* Bind input images to the shader. */
+ ff_vk_shader_update_img_array(vkctx, exec, &s->dwt_upload_shd, frame,
views, 0, 0,
+ VK_IMAGE_LAYOUT_GENERAL, VK_NULL_HANDLE);
+ vulkan_bind_img_planes(vkctx, exec, &s->dwt_upload_shd,
s->intermediate_views, 0, 1);
+
+ /* Upload coefficients from planes to the buffer. */
+ s->dwt_consts.diff_offset = s->base.diff_offset;
+ ff_vk_exec_bind_shader(vkctx, exec, &s->dwt_upload_shd);
+ ff_vk_shader_update_push_const(vkctx, exec, &s->dwt_upload_shd,
VK_SHADER_STAGE_COMPUTE_BIT,
+ 0, sizeof(VC2DwtPushData), &s->dwt_consts);
+ vk->CmdDispatch(exec->buf, group_x, group_y, 1);
+ vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
+ .sType =
VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
+ .pImageMemoryBarriers = img_bar,
+ .imageMemoryBarrierCount =
nb_img_bar,
+ });
+
+ /* Perform wavelet trasform. */
+ if (wavelet_idx == VC2_TRANSFORM_HAAR || wavelet_idx ==
VC2_TRANSFORM_HAAR_S)
+ dwt_plane_haar(s, exec, img_bar, nb_img_bar);
+ else if (wavelet_idx == VC2_TRANSFORM_5_3)
+ dwt_plane_legall(s, exec, img_bar, nb_img_bar);
+
+fail:
+ return err;
+}
+
+static void vulkan_encode_slices(VC2EncVulkanContext *s, FFVkExecContext *exec)
+{
+ int num_slices = s->base.num_x * s->base.num_y;
+ int num_slice_groups = (num_slices + SLICE_WORKGROUP_X - 1) >> 7;
+ int i, skip = 0;
+ FFVulkanContext *vkctx = &s->vkctx;
+ FFVulkanFunctions *vk = &vkctx->vkfn;
+
+ /* Calculate slice sizes. */
+ vulkan_bind_img_planes(vkctx, exec, &s->slice_shd, s->intermediate_views,
0, 0);
+ ff_vk_shader_update_desc_buffer(vkctx, exec, &s->slice_shd,
+ 0, 1, 0, &s->lut_buf, 0,
+ sizeof(VC2EncAuxData),
+ VK_FORMAT_UNDEFINED);
+ ff_vk_shader_update_desc_buffer(vkctx, exec, &s->slice_shd,
+ 0, 2, 0, &s->slice_buf, 0,
+ s->slice_buf.size,
+ VK_FORMAT_UNDEFINED);
+ ff_vk_exec_bind_shader(vkctx, exec, &s->slice_shd);
+ ff_vk_shader_update_push_const(vkctx, exec, &s->slice_shd,
VK_SHADER_STAGE_COMPUTE_BIT,
+ 0, sizeof(VC2EncSliceCalcPushData),
&s->calc_consts);
+ vk->CmdDispatch(exec->buf, num_slice_groups, 1, 1);
+
+ flush_put_bits(&s->base.pb);
+ s->enc_consts.pb += put_bytes_output(&s->base.pb);
+
+ /* Wait for slice sizes to be written. */
+ vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
+ .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
+ .pBufferMemoryBarriers = &(VkBufferMemoryBarrier2) {
+ .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2,
+ .srcStageMask = VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
+ .dstStageMask = VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
+ .srcAccessMask = VK_ACCESS_2_SHADER_WRITE_BIT,
+ .dstAccessMask = VK_ACCESS_2_SHADER_READ_BIT,
+ .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+ .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+ .buffer = s->slice_buf.buf,
+ .size = sizeof(VC2EncSliceArgs) * num_slices,
+ .offset = 0,
+ },
+ .bufferMemoryBarrierCount = 1U,
+ });
+
+ /* Bind encoding shader. */
+ vulkan_bind_img_planes(vkctx, exec, &s->enc_shd, s->intermediate_views, 0,
0);
+ ff_vk_shader_update_desc_buffer(vkctx, exec, &s->enc_shd,
+ 0, 1, 0, &s->lut_buf, 0,
+ sizeof(VC2EncAuxData),
+ VK_FORMAT_UNDEFINED);
+ ff_vk_shader_update_desc_buffer(vkctx, exec, &s->enc_shd,
+ 0, 2, 0, &s->slice_buf, 0,
+ s->slice_buf.size,
+ VK_FORMAT_UNDEFINED);
+ ff_vk_exec_bind_shader(vkctx, exec, &s->enc_shd);
+ ff_vk_shader_update_push_const(vkctx, exec, &s->enc_shd,
VK_SHADER_STAGE_COMPUTE_BIT,
+ 0, sizeof(VC2EncPushData), &s->enc_consts);
+
+ vk->CmdDispatch(exec->buf, num_slice_groups, 1, 1);
+
+ ff_vk_exec_submit(vkctx, exec);
+ ff_vk_exec_wait(vkctx, exec);
+
+ for (int slice_y = 0; slice_y < s->base.num_y; slice_y++) {
+ for (int slice_x = 0; slice_x < s->base.num_x; slice_x++) {
+ VC2EncSliceArgs *args = &s->slice_args[s->base.num_x * slice_y +
slice_x];
+ skip += args->bytes;
+ }
+ }
+
+ /* Skip forward to write end header */
+ skip_put_bytes(&s->base.pb, skip);
+
+ /* Free allocated intermediate frames */
+ for (i = 0; i < 3; i++)
+ av_frame_free(&s->intermediate_frame[i]);
+}
+
+static int encode_frame(VC2EncVulkanContext *sv, AVPacket *avpkt, const
AVFrame *frame,
+ const char *aux_data, const int header_size, int field)
+{
+ int ret;
+ int64_t max_frame_bytes;
+ AVBufferRef *avpkt_buf = NULL;
+ FFVkBuffer* buf_vk = NULL;
+ VC2EncContext* s = &sv->base;
+ FFVulkanContext *vkctx = &sv->vkctx;
+ FFVkExecContext *exec = ff_vk_exec_get(vkctx, &sv->e);
+
+ ff_vk_exec_start(vkctx, exec);
+
+ /* Perform wavelet pass on the inpute frame. */
+ vulkan_dwt_plane(sv, exec, (AVFrame*)frame);
+
+ /* Allocate a buffer that can fit at all all 3 planes of data */
+ max_frame_bytes = header_size + s->avctx->width * s->avctx->height *
sizeof(dwtcoef);
+
+ /* Get a pooled device local host visible buffer for writing output data */
+ if (field < 2) {
+ ret = ff_vk_get_pooled_buffer(vkctx, &sv->dwt_buf_pool, &avpkt_buf,
+ VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
+
VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT, NULL,
+ max_frame_bytes << s->interlaced,
+ VK_MEMORY_PROPERTY_HOST_CACHED_BIT |
+ VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+ VK_MEMORY_PROPERTY_HOST_COHERENT_BIT);
+ avpkt->buf = avpkt_buf;
+ buf_vk = (FFVkBuffer *)avpkt_buf->data;
+ avpkt->data = buf_vk->mapped_mem;
+ avpkt->size = max_frame_bytes << s->interlaced;
+ sv->enc_consts.pb = buf_vk->address;
+ ff_vk_exec_add_dep_buf(vkctx, exec, &avpkt_buf, 1, 1);
+
+ if (ret < 0)
+ return ret;
+ init_put_bits(&s->pb, avpkt->data, avpkt->size);
+ }
+
+ /* Sequence header */
+ vc2_encode_parse_info(s, DIRAC_PCODE_SEQ_HEADER);
+ vc2_encode_seq_header(s);
+
+ /* Encoder version */
+ if (aux_data) {
+ vc2_encode_parse_info(s, DIRAC_PCODE_AUX);
+ ff_put_string(&s->pb, aux_data, 1);
+ }
+
+ /* Picture header */
+ vc2_encode_parse_info(s, DIRAC_PCODE_PICTURE_HQ);
+ vc2_encode_picture_start(s);
+
+ /* Encode slices */
+ vulkan_encode_slices(sv, exec);
+
+ /* End sequence */
+ vc2_encode_parse_info(s, DIRAC_PCODE_END_SEQ);
+
+ return 0;
+}
+
+static av_cold int vc2_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
+ const AVFrame *frame, int *got_packet)
+{
+ int ret = 0;
+ int slice_ceil, sig_size = 256;
+ VC2EncVulkanContext *sv = avctx->priv_data;
+ VC2EncContext *s = &sv->base;
+ const int bitexact = avctx->flags & AV_CODEC_FLAG_BITEXACT;
+ const char *aux_data = bitexact ? "Lavc" : LIBAVCODEC_IDENT;
+ const int aux_data_size = bitexact ? sizeof("Lavc") :
sizeof(LIBAVCODEC_IDENT);
+ const int header_size = 100 + aux_data_size;
+ int64_t r_bitrate = avctx->bit_rate >> (s->interlaced);
+
+ s->avctx = avctx;
+ s->size_scaler = 2;
+ s->prefix_bytes = 0;
+ s->last_parse_code = 0;
+ s->next_parse_offset = 0;
+
+ /* Rate control */
+ s->frame_max_bytes = (av_rescale(r_bitrate, s->avctx->time_base.num,
+ s->avctx->time_base.den) >> 3) -
header_size;
+ s->slice_max_bytes = slice_ceil = av_rescale(s->frame_max_bytes, 1,
s->num_x * s->num_y);
+
+ /* Find an appropriate size scaler */
+ while (sig_size > 255) {
+ int r_size = SSIZE_ROUND(s->slice_max_bytes);
+ if (r_size > slice_ceil) {
+ s->slice_max_bytes -= r_size - slice_ceil;
+ r_size = SSIZE_ROUND(s->slice_max_bytes);
+ }
+ sig_size = r_size/s->size_scaler; /* Signalled slize size */
+ s->size_scaler <<= 1;
+ }
+
+ s->slice_min_bytes = s->slice_max_bytes -
s->slice_max_bytes*(s->tolerance/100.0f);
+ if (s->slice_min_bytes < 0)
+ return AVERROR(EINVAL);
+
+ /* Update slice calc push data */
+ sv->calc_consts.size_scaler = s->size_scaler;
+ sv->calc_consts.bits_ceil = s->slice_max_bytes << 3;
+ sv->calc_consts.bits_floor = s->slice_min_bytes << 3;
+ sv->enc_consts.prefix_bytes = 0;
+ sv->enc_consts.size_scaler = s->size_scaler;
+
+ ret = encode_frame(sv, avpkt, frame, aux_data, header_size, s->interlaced);
+ if (ret)
+ return ret;
+ if (s->interlaced) {
+ ret = encode_frame(sv, avpkt, frame, aux_data, header_size, 2);
+ if (ret)
+ return ret;
+ }
+
+ flush_put_bits(&s->pb);
+ av_shrink_packet(avpkt, put_bytes_output(&s->pb));
+ avpkt->flags |= AV_PKT_FLAG_KEY;
+ *got_packet = 1;
+
+ return 0;
+}
+
+static av_cold int vc2_encode_end(AVCodecContext *avctx)
+{
+ VC2EncVulkanContext *sv = avctx->priv_data;
+ FFVulkanContext *vkctx = &sv->vkctx;
+ int i;
+
+ ff_vk_exec_pool_free(vkctx, &sv->e);
+
+ ff_vk_shader_free(vkctx, &sv->dwt_upload_shd);
+ ff_vk_shader_free(vkctx, &sv->dwt_haar_shd);
+ ff_vk_shader_free(vkctx, &sv->dwt_hor_shd);
+ ff_vk_shader_free(vkctx, &sv->dwt_ver_shd);
+ ff_vk_shader_free(vkctx, &sv->slice_shd);
+ ff_vk_shader_free(vkctx, &sv->enc_shd);
+
+ ff_vk_free_buf(vkctx, &sv->slice_buf);
+ ff_vk_free_buf(vkctx, &sv->lut_buf);
+
+ for (i = 0; i < 3; i++) {
+ ff_vc2enc_free_transforms(&sv->base.transform_args[i].t);
+ av_freep(&sv->base.plane[i].coef_buf);
+ av_buffer_unref(&sv->intermediate_frames_ref[i]);
+ }
+
+ av_buffer_pool_uninit(&sv->dwt_buf_pool);
+ ff_vk_uninit(vkctx);
+
+ return 0;
+}
+
+static av_cold int vc2_encode_init(AVCodecContext *avctx)
+{
+ static AVOnce init_static_once = AV_ONCE_INIT;
+ Plane *p;
+ SubBand *b;
+ int i, level, o, ret, depth;
+ const AVPixFmtDescriptor *fmt;
+ VC2EncVulkanContext *sv = avctx->priv_data;
+ VC2EncContext *s = &sv->base;
+ FFVulkanContext *vkctx = &sv->vkctx;
+
+ /* Init Vulkan */
+ ret = ff_vk_init(&sv->vkctx, avctx, NULL, avctx->hw_frames_ctx);
+ if (ret < 0)
+ return ret;
+
+ sv->qf = ff_vk_qf_find(vkctx, VK_QUEUE_COMPUTE_BIT, 0);
+ if (!sv->qf) {
+ av_log(avctx, AV_LOG_ERROR, "Device has no compute queues!\n");
+ return ret;
+ }
+
+ s->picture_number = 0;
+
+ /* Total allowed quantization range */
+ s->q_ceil = DIRAC_MAX_QUANT_INDEX;
+
+ s->ver.major = 2;
+ s->ver.minor = 0;
+ s->profile = 3;
+ s->level = 3;
+
+ s->base_vf = -1;
+ s->strict_compliance = 1;
+
+ s->q_avg = 0;
+ s->slice_max_bytes = 0;
+ s->slice_min_bytes = 0;
+
+ /* Mark unknown as progressive */
+ s->interlaced = !((avctx->field_order == AV_FIELD_UNKNOWN) ||
+ (avctx->field_order == AV_FIELD_PROGRESSIVE));
+ if (s->interlaced != 0) {
+ av_log(avctx, AV_LOG_ERROR, "Interlaced video is unsupported by this
encoder\n");
+ return AVERROR(ENOTSUP);
+ }
+
+ for (i = 0; i < base_video_fmts_len; i++) {
+ const VC2BaseVideoFormat *fmt = &base_video_fmts[i];
+ if (avctx->pix_fmt != fmt->pix_fmt || avctx->time_base.num !=
fmt->time_base.num ||
+ avctx->time_base.den != fmt->time_base.den || avctx->width !=
fmt->width ||
+ avctx->height != fmt->height || s->interlaced != fmt->interlaced)
+ continue;
+ s->base_vf = i;
+ s->level = base_video_fmts[i].level;
+ break;
+ }
+
+ if (s->interlaced)
+ av_log(avctx, AV_LOG_WARNING, "Interlacing enabled!\n");
+
+ if ((s->slice_width & (s->slice_width - 1)) ||
+ (s->slice_height & (s->slice_height - 1))) {
+ av_log(avctx, AV_LOG_ERROR, "Slice size is not a power of two!\n");
+ return AVERROR_UNKNOWN;
+ }
+
+ if ((s->slice_width > avctx->width) ||
+ (s->slice_height > avctx->height)) {
+ av_log(avctx, AV_LOG_ERROR, "Slice size is bigger than the image!\n");
+ return AVERROR_UNKNOWN;
+ }
+
+ if (s->base_vf <= 0) {
+ if (avctx->strict_std_compliance < FF_COMPLIANCE_STRICT) {
+ s->strict_compliance = s->base_vf = 0;
+ av_log(avctx, AV_LOG_WARNING, "Format does not strictly comply
with VC2 specs\n");
+ } else {
+ av_log(avctx, AV_LOG_WARNING, "Given format does not strictly
comply with "
+ "the specifications, decrease strictness to use it.\n");
+ return AVERROR_UNKNOWN;
+ }
+ } else {
+ av_log(avctx, AV_LOG_INFO, "Selected base video format = %i (%s)\n",
+ s->base_vf, base_video_fmts[s->base_vf].name);
+ }
+
+ /* Chroma subsampling */
+ ret = av_pix_fmt_get_chroma_sub_sample(vkctx->frames->sw_format,
&s->chroma_x_shift,
+
&s->chroma_y_shift);
+ if (ret)
+ return ret;
+
+ /* Bit depth and color range index */
+ fmt = av_pix_fmt_desc_get(vkctx->frames->sw_format);
+ depth = fmt->comp[0].depth;
+
+ /* 16-bit depth is unsupported by this encoder */
+ if (depth == 16) {
+ av_log(avctx, AV_LOG_ERROR, "16-bit pixel format depth is unsupported
by this encoder\n");
+ return AVERROR(ENOTSUP);
+ }
+
+ if (depth == 8 && avctx->color_range == AVCOL_RANGE_JPEG) {
+ s->bpp = 1;
+ s->bpp_idx = 1;
+ s->diff_offset = 128;
+ } else if (depth == 8 && (avctx->color_range == AVCOL_RANGE_MPEG ||
+ avctx->color_range == AVCOL_RANGE_UNSPECIFIED)) {
+ s->bpp = 1;
+ s->bpp_idx = 2;
+ s->diff_offset = 128;
+ } else if (depth == 10) {
+ s->bpp = 2;
+ s->bpp_idx = 3;
+ s->diff_offset = 512;
+ } else {
+ s->bpp = 2;
+ s->bpp_idx = 4;
+ s->diff_offset = 2048;
+ }
+
+ /* Planes initialization */
+ for (i = 0; i < 3; i++) {
+ int w, h;
+ p = &s->plane[i];
+ p->width = avctx->width >> (i ? s->chroma_x_shift : 0);
+ p->height = avctx->height >> (i ? s->chroma_y_shift : 0);
+ if (s->interlaced)
+ p->height >>= 1;
+ p->dwt_width = w = FFALIGN(p->width, (1 << s->wavelet_depth));
+ p->dwt_height = h = FFALIGN(p->height, (1 << s->wavelet_depth));
+ p->coef_stride = FFALIGN(p->dwt_width, 32);
+ for (level = s->wavelet_depth-1; level >= 0; level--) {
+ w = w >> 1;
+ h = h >> 1;
+ for (o = 0; o < 4; o++) {
+ b = &p->band[level][o];
+ b->width = w;
+ b->height = h;
+ b->stride = p->coef_stride;
+ b->shift = (o > 1)*b->height*b->stride + (o & 1)*b->width;
+ }
+ }
+
+ /* DWT init */
+ if (ff_vc2enc_init_transforms(&s->transform_args[i].t,
+ s->plane[i].coef_stride,
+ s->plane[i].dwt_height,
+ s->slice_width, s->slice_height))
+ return AVERROR(ENOMEM);
+ }
+
+ /* Slices */
+ s->num_x = s->plane[0].dwt_width/s->slice_width;
+ s->num_y = s->plane[0].dwt_height/s->slice_height;
+
+ s->slice_args = av_calloc(s->num_x*s->num_y, sizeof(SliceArgs));
+ if (!s->slice_args)
+ return AVERROR(ENOMEM);
+
+ for (i = 0; i < 116; i++) {
+ const uint64_t qf = ff_dirac_qscale_tab[i];
+ const uint32_t m = av_log2(qf);
+ const uint32_t t = (1ULL << (m + 32)) / qf;
+ const uint32_t r = (t*qf + qf) & UINT32_MAX;
+ if (!(qf & (qf - 1))) {
+ s->qmagic_lut[i][0] = 0xFFFFFFFF;
+ s->qmagic_lut[i][1] = 0xFFFFFFFF;
+ } else if (r <= 1 << m) {
+ s->qmagic_lut[i][0] = t + 1;
+ s->qmagic_lut[i][1] = 0;
+ } else {
+ s->qmagic_lut[i][0] = t;
+ s->qmagic_lut[i][1] = t;
+ }
+ }
+
+ ff_thread_once(&init_static_once, vc2_init_static_data);
+
+ vc2_init_vulkan(avctx);
+
+ return 0;
+}
+
+#define VC2ENC_FLAGS (AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
+static const AVOption vc2enc_options[] = {
+ {"tolerance", "Max undershoot in percent", offsetof(VC2EncContext,
tolerance), AV_OPT_TYPE_DOUBLE, {.dbl = 5.0f}, 0.0f, 45.0f, VC2ENC_FLAGS, .unit
= "tolerance"},
+ {"slice_width", "Slice width", offsetof(VC2EncContext, slice_width),
AV_OPT_TYPE_INT, {.i64 = 32}, 32, 1024, VC2ENC_FLAGS, .unit = "slice_width"},
+ {"slice_height", "Slice height", offsetof(VC2EncContext, slice_height),
AV_OPT_TYPE_INT, {.i64 = 8}, 8, 1024, VC2ENC_FLAGS, .unit = "slice_height"},
+ {"wavelet_depth", "Transform depth", offsetof(VC2EncContext,
wavelet_depth), AV_OPT_TYPE_INT, {.i64 = 4}, 1, 5, VC2ENC_FLAGS, .unit =
"wavelet_depth"},
+ {"wavelet_type", "Transform type", offsetof(VC2EncContext, wavelet_idx),
AV_OPT_TYPE_INT, {.i64 = VC2_TRANSFORM_HAAR_S}, 0, VC2_TRANSFORMS_NB,
VC2ENC_FLAGS, .unit = "wavelet_idx"},
+ {"5_3", "LeGall (5,3)", 0, AV_OPT_TYPE_CONST,
{.i64 = VC2_TRANSFORM_5_3}, INT_MIN, INT_MAX, VC2ENC_FLAGS, .unit =
"wavelet_idx"},
+ {"haar", "Haar (with shift)", 0, AV_OPT_TYPE_CONST,
{.i64 = VC2_TRANSFORM_HAAR_S}, INT_MIN, INT_MAX, VC2ENC_FLAGS, .unit =
"wavelet_idx"},
+ {"haar_noshift", "Haar (without shift)", 0, AV_OPT_TYPE_CONST,
{.i64 = VC2_TRANSFORM_HAAR}, INT_MIN, INT_MAX, VC2ENC_FLAGS, .unit =
"wavelet_idx"},
+ {"qm", "Custom quantization matrix", offsetof(VC2EncContext,
quant_matrix), AV_OPT_TYPE_INT, {.i64 = VC2_QM_DEF}, 0, VC2_QM_NB,
VC2ENC_FLAGS, .unit = "quant_matrix"},
+ {"default", "Default from the specifications", 0, AV_OPT_TYPE_CONST,
{.i64 = VC2_QM_DEF}, INT_MIN, INT_MAX, VC2ENC_FLAGS, .unit = "quant_matrix"},
+ {"color", "Prevents low bitrate discoloration", 0,
AV_OPT_TYPE_CONST, {.i64 = VC2_QM_COL}, INT_MIN, INT_MAX, VC2ENC_FLAGS, .unit =
"quant_matrix"},
+ {"flat", "Optimize for PSNR", 0, AV_OPT_TYPE_CONST, {.i64 =
VC2_QM_FLAT}, INT_MIN, INT_MAX, VC2ENC_FLAGS, .unit = "quant_matrix"},
+ {NULL}
+};
+
+static const AVClass vc2enc_class = {
+ .class_name = "vc2_vulkan_encoder",
+ .category = AV_CLASS_CATEGORY_ENCODER,
+ .option = vc2enc_options,
+ .item_name = av_default_item_name,
+ .version = LIBAVUTIL_VERSION_INT
+};
+
+static const FFCodecDefault vc2enc_defaults[] = {
+ { "b", "600000000" },
+ { NULL },
+};
+
+const AVCodecHWConfigInternal *const ff_vc2_hw_configs[] = {
+ HW_CONFIG_ENCODER_FRAMES(VULKAN, VULKAN),
+ HW_CONFIG_ENCODER_DEVICE(NONE, VULKAN),
+ NULL,
+};
+
+const FFCodec ff_vc2_vulkan_encoder = {
+ .p.name = "vc2_vulkan",
+ CODEC_LONG_NAME("SMPTE VC-2"),
+ .p.type = AVMEDIA_TYPE_VIDEO,
+ .p.id = AV_CODEC_ID_DIRAC,
+ .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_HARDWARE,
+ .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
+ .priv_data_size = sizeof(VC2EncVulkanContext),
+ .init = vc2_encode_init,
+ .close = vc2_encode_end,
+ FF_CODEC_ENCODE_CB(vc2_encode_frame),
+ .p.priv_class = &vc2enc_class,
+ .defaults = vc2enc_defaults,
+ CODEC_PIXFMTS(AV_PIX_FMT_VULKAN),
+ .hw_configs = ff_vc2_hw_configs,
+};
diff --git a/libavcodec/vulkan/vc2_dwt_haar.comp
b/libavcodec/vulkan/vc2_dwt_haar.comp
new file mode 100644
index 0000000000..4806cca729
--- /dev/null
+++ b/libavcodec/vulkan/vc2_dwt_haar.comp
@@ -0,0 +1,82 @@
+/*
+ * VC2 codec
+ *
+ * Copyright (c) 2025 raphaelthegreat <geoste...@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#extension GL_EXT_scalar_block_layout : require
+#extension GL_EXT_buffer_reference : require
+
+#define LOCAL_X 1024
+
+layout(push_constant, scalar) uniform ComputeInfo {
+ int s;
+ int plane_idx;
+ int wavelet_depth;
+};
+
+shared int local_coef[LOCAL_X];
+
+void main()
+{
+ ivec2 coord = ivec2(gl_GlobalInvocationID.xy);
+ ivec2 dwt_dim = imageSize(coef_buf[plane_idx]);
+ int value = imageLoad(coef_buf[plane_idx], coord).x;
+
+ /* Perform Haar wavelet on the 32x32 local workgroup with shared memory */
+ for (int i = 0; i < wavelet_depth; i++)
+ {
+ ivec2 mask = ivec2((1 << i) - 1);
+ if (any(notEqual(coord & mask, ivec2(0))))
+ break;
+
+ /* Offset between valid hor pixels for each level, +1, +2, +4 etc */
+ int dist = (1 << i);
+
+ local_coef[gl_LocalInvocationIndex] = value;
+ barrier();
+
+ /* Horizontal haar wavelet */
+ uint other_id = gl_LocalInvocationIndex ^ dist;
+ int other = local_coef[other_id];
+ int a = gl_LocalInvocationIndex < other_id ? value : other;
+ int b = gl_LocalInvocationIndex < other_id ? other : value;
+ int dst_b = (b - a) * (1 << s);
+ int dst_a = a * (1 << s) + ((dst_b + 1) >> 1);
+ value = gl_LocalInvocationIndex < other_id ? dst_a : dst_b;
+
+ /* Offset between valid ver pixels for each level, +1, +2, +4 etc */
+ dist <<= 5;
+
+ local_coef[gl_LocalInvocationIndex] = value;
+ barrier();
+
+ /* Vertical haar wavelet */
+ other_id = gl_LocalInvocationIndex ^ dist;
+ other = local_coef[other_id];
+ a = gl_LocalInvocationIndex < other_id ? value : other;
+ b = gl_LocalInvocationIndex < other_id ? other : value;
+ dst_b = b - a;
+ dst_a = a + ((dst_b + 1) >> 1);
+ value = gl_LocalInvocationIndex < other_id ? dst_a : dst_b;
+ }
+
+ /* Store value */
+ imageStore(coef_buf[plane_idx], coord, ivec4(value));
+}
diff --git a/libavcodec/vulkan/vc2_dwt_haar_subgroup.comp
b/libavcodec/vulkan/vc2_dwt_haar_subgroup.comp
new file mode 100644
index 0000000000..81b0964271
--- /dev/null
+++ b/libavcodec/vulkan/vc2_dwt_haar_subgroup.comp
@@ -0,0 +1,75 @@
+/*
+ * VC2 codec
+ *
+ * Copyright (c) 2025 raphaelthegreat <geoste...@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#extension GL_EXT_scalar_block_layout : require
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_shuffle : require
+
+#define TILE_DIM 8
+
+layout(push_constant, scalar) uniform ComputeInfo {
+ int s;
+ int plane_idx;
+ int wavelet_depth;
+};
+
+void main()
+{
+ ivec2 tile_coord = ivec2(gl_WorkGroupID.xy);
+ ivec2 local_coord = ivec2(gl_LocalInvocationIndex & 7,
gl_LocalInvocationIndex >> 3);
+ ivec2 coord = tile_coord * ivec2(TILE_DIM) + local_coord;
+
+ int value = imageLoad(coef_buf[plane_idx], coord).x;
+ for (int i = 0; i < wavelet_depth; i++)
+ {
+ ivec2 mask = ivec2((1 << i) - 1);
+ if (any(notEqual(local_coord & mask, ivec2(0))))
+ break;
+
+ /* Offset between valid hor pixels for each level, +1, +2, +4 etc */
+ int dist = (1 << i);
+
+ /* Horizontal haar wavelet */
+ uint other_sub_id = gl_SubgroupInvocationID ^ dist;
+ int other = subgroupShuffle(value, other_sub_id);
+ int a = gl_SubgroupInvocationID < other_sub_id ? value : other;
+ int b = gl_SubgroupInvocationID < other_sub_id ? other : value;
+ int dst_b = (b - a) * (1 << s);
+ int dst_a = a * (1 << s) + ((dst_b + 1) >> 1);
+ value = gl_SubgroupInvocationID < other_sub_id ? dst_a : dst_b;
+
+ /* Offset between valid ver pixels for each level, +1, +2, +4 etc */
+ dist <<= 3;
+
+ /* Vertical haar wavelet */
+ other_sub_id = gl_SubgroupInvocationID ^ dist;
+ other = subgroupShuffle(value, other_sub_id);
+ a = gl_SubgroupInvocationID < other_sub_id ? value : other;
+ b = gl_SubgroupInvocationID < other_sub_id ? other : value;
+ dst_b = b - a;
+ dst_a = a + ((dst_b + 1) >> 1);
+ value = gl_SubgroupInvocationID < other_sub_id ? dst_a : dst_b;
+ }
+
+ /* Store value */
+ imageStore(coef_buf[plane_idx], coord, ivec4(value));
+}
diff --git a/libavcodec/vulkan/vc2_dwt_hor_legall.comp
b/libavcodec/vulkan/vc2_dwt_hor_legall.comp
new file mode 100644
index 0000000000..bada2ee1fd
--- /dev/null
+++ b/libavcodec/vulkan/vc2_dwt_hor_legall.comp
@@ -0,0 +1,82 @@
+/*
+ * VC2 codec
+ *
+ * Copyright (c) 2025 raphaelthegreat <geoste...@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#extension GL_EXT_scalar_block_layout : require
+#extension GL_EXT_buffer_reference : require
+
+layout(push_constant, scalar) uniform ComputeInfo {
+ int s;
+ int diff_offset;
+ int level;
+};
+
+int image_load(int coord_x)
+{
+ int coord_y = int(gl_GlobalInvocationID.x);
+ return imageLoad(coef_buf[gl_GlobalInvocationID.z], ivec2(coord_x,
coord_y)).x;
+}
+
+void image_store(int coord_x, int value)
+{
+ int coord_y = int(gl_GlobalInvocationID.x);
+ imageStore(coef_buf[gl_GlobalInvocationID.z], ivec2(coord_x, coord_y),
ivec4(value));
+}
+
+void main()
+{
+ int coord_y = int(gl_GlobalInvocationID.x);
+ uint plane_idx = gl_GlobalInvocationID.z;
+ ivec2 work_area = imageSize(coef_buf[plane_idx]);
+ int dist = 1 << level;
+ if (coord_y >= work_area.y || (coord_y & (dist - 1)) != 0)
+ return;
+
+ // Shift in one bit that is used for additional precision
+ for (int x = 0; x < work_area.x; x += dist)
+ image_store(x, image_load(x) << 1);
+
+ // Lifting stage 2
+ for (int x = 0; x < work_area.x - 2 * dist; x += 2 * dist) {
+ int lhs = image_load(x);
+ int rhs = image_load(x + 2 * dist);
+ int value = image_load(x + dist);
+ value -= (lhs + rhs + 1) >> 1;
+ image_store(x + dist, value);
+ }
+ int lhs = image_load(work_area.x - 2 * dist);
+ int value = image_load(work_area.x - dist);
+ value -= (2 * lhs + 1) >> 1;
+ image_store(work_area.x - dist, value);
+
+ // Lifting stage 1
+ lhs = image_load(dist);
+ value = image_load(0);
+ value += (2 * lhs + 2) >> 2;
+ image_store(0, value);
+ for (int x = 2 * dist; x <= work_area.x - 2 * dist; x += 2 * dist) {
+ int lhs = image_load(x - dist);
+ int rhs = image_load(x + dist);
+ int value = image_load(x);
+ value += (lhs + rhs + 2) >> 2;
+ image_store(x, value);
+ }
+}
diff --git a/libavcodec/vulkan/vc2_dwt_upload.comp
b/libavcodec/vulkan/vc2_dwt_upload.comp
new file mode 100644
index 0000000000..c758fd867f
--- /dev/null
+++ b/libavcodec/vulkan/vc2_dwt_upload.comp
@@ -0,0 +1,96 @@
+/*
+ * VC2 codec
+ *
+ * Copyright (c) 2025 raphaelthegreat <geoste...@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#extension GL_EXT_scalar_block_layout : require
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
+#define AV_PIX_FMT_XV30 214
+#define AV_PIX_FMT_XV36 216
+#define AV_PIX_FMT_XV48 242
+#define AV_PIX_FMT_P212 222
+#define AV_PIX_FMT_P012 209
+#define AV_PIX_FMT_P210 198
+#define AV_PIX_FMT_P016 169
+#define AV_PIX_FMT_P010 158
+#define AV_PIX_FMT_NV16 101
+#define AV_PIX_FMT_NV12 23
+
+#define Y 0
+#define U 1
+#define V 2
+
+layout(push_constant, scalar) uniform ComputeInfo {
+ int s;
+ int diff_offset;
+ int level;
+};
+
+uvec4 load_plane(uint plane_idx)
+{
+ ivec2 coord = ivec2(gl_GlobalInvocationID.xy);
+ return imageLoad(src_planes[plane_idx], coord);
+}
+
+void store_plane(uint plane_idx, uint value)
+{
+ int result = int(value - diff_offset);
+ ivec2 coord = ivec2(gl_GlobalInvocationID.xy);
+ imageStore(coef_buf[plane_idx], coord, ivec4(result));
+}
+
+void main()
+{
+ uvec4 p0 = load_plane(0);
+#if PLANE_FMT == AV_PIX_FMT_XV30
+ store_plane(Y, (p0.x >> 10) & 0x3FF);
+ store_plane(U, p0.x & 0x3FF);
+ store_plane(V, (p0.x >> 20) & 0x3FF);
+#elif PLANE_FMT == AV_PIX_FMT_XV36
+ store_plane(Y, p0.y >> 4);
+ store_plane(U, p0.x >> 4);
+ store_plane(V, p0.z >> 4);
+#elif PLANE_FMT == AV_PIX_FMT_NV12
+ uvec4 p1 = load_plane(1);
+ store_plane(Y, p0.x | p0.y << 8);
+ store_plane(U, p1.x);
+ store_plane(V, p1.y);
+#elif PLANE_FMT == AV_PIX_FMT_NV16
+ uvec4 p1 = load_plane(1);
+ store_plane(Y, p0.x);
+ store_plane(U, p1.x);
+ store_plane(V, p1.y);
+#elif PLANE_FMT == AV_PIX_FMT_P010 || PLANE_FMT == AV_PIX_FMT_P210
+ uvec4 p1 = load_plane(1);
+ store_plane(Y, p0.x >> 6);
+ store_plane(U, p1.x >> 6);
+ store_plane(V, p1.y >> 6);
+#elif PLANE_FMT == AV_PIX_FMT_P012 || PLANE_FMT == AV_PIX_FMT_P212
+ uvec4 p1 = load_plane(1);
+ store_plane(Y, p0.x >> 4);
+ store_plane(U, p1.x >> 4);
+ store_plane(V, p1.y >> 4);
+#else
+ store_plane(Y, p0.x);
+ store_plane(U, load_plane(1).x);
+ store_plane(V, load_plane(2).x);
+#endif
+}
diff --git a/libavcodec/vulkan/vc2_dwt_ver_legall.comp
b/libavcodec/vulkan/vc2_dwt_ver_legall.comp
new file mode 100644
index 0000000000..ca391cc8d8
--- /dev/null
+++ b/libavcodec/vulkan/vc2_dwt_ver_legall.comp
@@ -0,0 +1,78 @@
+/*
+ * VC2 codec
+ *
+ * Copyright (c) 2025 raphaelthegreat <geoste...@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#extension GL_EXT_scalar_block_layout : require
+#extension GL_EXT_buffer_reference : require
+
+layout(push_constant, scalar) uniform ComputeInfo {
+ int s;
+ int diff_offset;
+ int level;
+};
+
+int image_load(int coord_y)
+{
+ int coord_x = int(gl_GlobalInvocationID.x);
+ return imageLoad(coef_buf[gl_GlobalInvocationID.z], ivec2(coord_x,
coord_y)).x;
+}
+
+void image_store(int coord_y, int value)
+{
+ int coord_x = int(gl_GlobalInvocationID.x);
+ imageStore(coef_buf[gl_GlobalInvocationID.z], ivec2(coord_x, coord_y),
ivec4(value));
+}
+
+void main()
+{
+ int coord_x = int(gl_GlobalInvocationID.x);
+ uint plane_idx = gl_GlobalInvocationID.z;
+ ivec2 work_area = imageSize(coef_buf[plane_idx]);
+ int dist = 1 << level;
+ if (coord_x >= work_area.x || (coord_x & (dist - 1)) != 0)
+ return;
+
+ // Lifting stage 2
+ for (int y = dist; y < work_area.y - 2 * dist; y += 2 * dist) {
+ int lhs = image_load(y - dist);
+ int rhs = image_load(y + dist);
+ int value = image_load(y);
+ value -= (lhs + rhs + 1) >> 1;
+ image_store(y, value);
+ }
+ int lhs = image_load(work_area.y - 2 * dist);
+ int value = image_load(work_area.y - dist);
+ value -= (2 * lhs + 1) >> 1;
+ image_store(work_area.y - dist, value);
+
+ // Lifting stage 1
+ lhs = image_load(dist);
+ value = image_load(0);
+ value += (2 * lhs + 2) >> 2;
+ image_store(0, value);
+ for (int y = 2 * dist; y <= work_area.y - 2 * dist; y += 2 * dist) {
+ int lhs = image_load(y + dist);
+ int rhs = image_load(y - dist);
+ int value = image_load(y);
+ value += (lhs + rhs + 2) >> 2;
+ image_store(y, value);
+ }
+}
diff --git a/libavcodec/vulkan/vc2_encode.comp
b/libavcodec/vulkan/vc2_encode.comp
new file mode 100644
index 0000000000..50b6214494
--- /dev/null
+++ b/libavcodec/vulkan/vc2_encode.comp
@@ -0,0 +1,169 @@
+/*
+ * VC2 codec
+ *
+ * Copyright (c) 2025 raphaelthegreat <geoste...@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+#extension GL_EXT_scalar_block_layout : require
+#extension GL_EXT_buffer_reference : require
+#extension GL_EXT_debug_printf : require
+
+#define MAX_DWT_LEVELS (5)
+
+layout(push_constant, scalar) uniform ComputeInfo {
+ u8buf bytestream;
+ ivec2 num_slices;
+ int wavelet_depth;
+ int size_scaler;
+ int prefix_bytes;
+};
+
+void put_vc2_ue_uint(inout PutBitContext pb, uint val)
+{
+ int pbits = 0, topbit = 1, maxval = 1, bits = 0;
+ if (val == 0)
+ {
+ put_bits(pb, 1, 1);
+ return;
+ }
+ val++;
+
+ while (val > maxval)
+ {
+ topbit <<= 1;
+ bits++;
+ maxval <<= 1;
+ maxval |= 1;
+ }
+
+ for (int i = 0; i < bits; i++)
+ {
+ topbit >>= 1;
+ pbits <<= 2;
+ if ((val & topbit) != 0)
+ pbits |= 1;
+ }
+
+ put_bits(pb, bits * 2 + 1, (pbits << 1) | 1);
+}
+
+int quants[MAX_DWT_LEVELS][4];
+
+int subband_coord(int index, int h, int lvl)
+{
+ int coord = index;
+ coord <<= 1;
+ coord |= h;
+ coord <<= (wavelet_depth-lvl-1);
+ return coord;
+}
+
+void main()
+{
+ int slice_index = int(gl_GlobalInvocationID.x);
+ int max_index = num_slices.x * num_slices.y;
+ if (slice_index >= max_index)
+ return;
+
+ /* Step 2. Quantize and encode */
+ int pb_start = slice_args[slice_index].pb_start;
+ int workgroup_x = int(gl_WorkGroupSize.x);
+ for (int i = 0, index = workgroup_x - 1; i < gl_WorkGroupID.x; i++) {
+ pb_start += slice_args[index].pb_start + slice_args[index].bytes;
+ index += workgroup_x;
+ }
+ ivec2 slice_coord = ivec2(slice_index % num_slices.x, slice_index /
num_slices.x);
+ int slice_bytes_max = slice_args[slice_index].bytes;
+ int quant_index = slice_args[slice_index].quant_idx;
+
+ PutBitContext pb;
+ init_put_bits(pb, OFFBUF(u8buf, bytestream, pb_start), slice_bytes_max);
+
+ for (int level = 0; level < wavelet_depth; level++)
+ for (int orientation = int(level > 0); orientation < 4; orientation++)
+ quants[level][orientation] = max(quant_index -
lut_quant[level][orientation], 0);
+
+ /* Write quant index for this slice */
+ put_bits(pb, 8, quant_index);
+
+ /* Luma + 2 Chroma planes */
+ for (int p = 0; p < 3; p++)
+ {
+ int pad_s, pad_c;
+ int bytes_start = int32_t(put_bytes_count(pb));
+
+ /* Save current location and write a zero value */
+ uint64_t write_ptr_start = pb.buf;
+ int bit_left_start = pb.bit_left;
+ put_bits(pb, 8, 0);
+
+ ivec2 dwt_dim = imageSize(coef_buf[p]);
+ for (int level = 0; level < wavelet_depth; level++)
+ {
+ ivec2 band_size = dwt_dim >> (wavelet_depth - level);
+ for (int o = int(level > 0); o < 4; o++)
+ {
+ /* Encode subband */
+ int left = band_size.x * (slice_coord.x) / num_slices.x;
+ int right = band_size.x * (slice_coord.x+1) / num_slices.x;
+ int top = band_size.y * (slice_coord.y) / num_slices.y;
+ int bottom = band_size.y * (slice_coord.y+1) / num_slices.y;
+
+ const int q_idx = quants[level][o];
+ const int qfactor = ff_dirac_qscale_tab[q_idx];
+
+ const int yh = o >> 1;
+ const int xh = o & 1;
+
+ for (int y = top; y < bottom; y++)
+ {
+ for (int x = left; x < right; x++)
+ {
+ int sx = subband_coord(x, xh, level);
+ int sy = subband_coord(y, yh, level);
+ int coef = imageLoad(coef_buf[p], ivec2(sx, sy)).x;
+ uint c_abs = uint(abs(coef));
+ c_abs = (c_abs << 2) / qfactor;
+ put_vc2_ue_uint(pb, c_abs);
+ if (c_abs != 0)
+ put_bits(pb, 1, int(coef < 0));
+ }
+ }
+ }
+ }
+ flush_put_bits(pb);
+ int bytes_len = int32_t(put_bytes_count(pb)) - bytes_start - 1;
+ if (p == 2)
+ {
+ int len_diff = slice_bytes_max - int32_t(put_bytes_count(pb));
+ pad_s = align((bytes_len + len_diff), size_scaler)/size_scaler;
+ pad_c = (pad_s*size_scaler) - bytes_len;
+ }
+ else
+ {
+ pad_s = align(bytes_len, size_scaler)/size_scaler;
+ pad_c = (pad_s*size_scaler) - bytes_len;
+ }
+ uint64_t start_ptr = write_ptr_start + ((BUF_BITS - bit_left_start) >>
3);
+ u8buf(start_ptr).v = uint8_t(pad_s);
+ /* vc2-reference uses that padding that decodes to '0' coeffs */
+ skip_put_bytes(pb, pad_c);
+ }
+}
diff --git a/libavcodec/vulkan/vc2_slice_sizes.comp
b/libavcodec/vulkan/vc2_slice_sizes.comp
new file mode 100644
index 0000000000..61070c1dc2
--- /dev/null
+++ b/libavcodec/vulkan/vc2_slice_sizes.comp
@@ -0,0 +1,170 @@
+/*
+ * VC2 codec
+ *
+ * Copyright (c) 2025 raphaelthegreat <geoste...@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+#extension GL_EXT_scalar_block_layout : require
+#extension GL_EXT_buffer_reference : require
+
+#define DIRAC_MAX_QUANT_INDEX 116
+#define MAX_DWT_LEVELS 5
+
+layout(push_constant, scalar) uniform ComputeInfo {
+ ivec2 num_slices;
+ int wavelet_depth;
+ int size_scaler;
+ int prefix_bytes;
+ int bits_ceil;
+ int bits_floor;
+};
+
+int count_vc2_ue_uint(uint val)
+{
+ return 2 * findMSB(val + 1) + 1;
+}
+
+int cache[DIRAC_MAX_QUANT_INDEX];
+int quants[MAX_DWT_LEVELS][4];
+shared int slice_sizes[gl_WorkGroupSize.x];
+
+int subband_coord(int index, int h, int lvl)
+{
+ int coord = index;
+ coord <<= 1;
+ coord |= h;
+ coord <<= (wavelet_depth-lvl-1);
+ return coord;
+}
+
+int count_hq_slice(int quant_index)
+{
+ int bits = 0;
+ if (cache[quant_index] != 0)
+ return cache[quant_index];
+
+ bits += 8*prefix_bytes;
+ bits += 8; /* quant_idx */
+
+ for (int level = 0; level < wavelet_depth; level++)
+ for (int orientation = int(level > 0); orientation < 4; orientation++)
+ quants[level][orientation] = max(quant_index -
lut_quant[level][orientation], 0);
+
+ int slice_index = int(gl_GlobalInvocationID.x);
+ ivec2 slice_coord = ivec2(slice_index % num_slices.x, slice_index /
num_slices.x);
+ for (int p = 0; p < 3; p++)
+ {
+ int bytes_start = bits >> 3;
+ bits += 8;
+
+ ivec2 dwt_dim = imageSize(coef_buf[p]);
+ for (int level = 0; level < wavelet_depth; level++)
+ {
+ ivec2 band_dim = dwt_dim >> (wavelet_depth - level);
+ for (int o = int(level > 0); o < 4; o++)
+ {
+ const int left = band_dim.x * slice_coord.x / num_slices.x;
+ const int right = band_dim.x * (slice_coord.x+1) /
num_slices.x;
+ const int top = band_dim.y * slice_coord.y / num_slices.y;
+ const int bottom = band_dim.y * (slice_coord.y+1) /
num_slices.y;
+
+ const int q_idx = quants[level][o];
+ const int qfactor = ff_dirac_qscale_tab[q_idx];
+
+ const int yh = o >> 1;
+ const int xh = o & 1;
+
+ for (int y = top; y < bottom; y++)
+ {
+ for (int x = left; x < right; x++)
+ {
+ int sx = subband_coord(x, xh, level);
+ int sy = subband_coord(y, yh, level);
+ int coef = imageLoad(coef_buf[p], ivec2(sx, sy)).x;
+ uint c_abs = uint(abs(coef));
+ c_abs = (c_abs << 2) / qfactor;
+ bits += count_vc2_ue_uint(c_abs);
+ bits += int(c_abs > 0);
+ }
+ }
+ }
+ }
+ bits += align(bits, 8) - bits;
+ int bytes_len = (bits >> 3) - bytes_start - 1;
+ int pad_s = align(bytes_len, size_scaler) / size_scaler;
+ int pad_c = (pad_s * size_scaler) - bytes_len;
+ bits += pad_c * 8;
+ }
+
+ cache[quant_index] = bits;
+ return bits;
+}
+
+int ssize_round(int b)
+{
+ return align(b, size_scaler) + 4 + prefix_bytes;
+}
+
+void main()
+{
+ int slice_index = int(gl_GlobalInvocationID.x);
+ int max_index = num_slices.x * num_slices.y;
+ if (slice_index >= max_index)
+ return;
+
+ for (int i = 0; i < DIRAC_MAX_QUANT_INDEX; i++)
+ cache[i] = 0;
+
+ const int q_ceil = DIRAC_MAX_QUANT_INDEX;
+ const int top = bits_ceil;
+ const int bottom = bits_floor;
+ int quant_buf[2] = int[2](-1, -1);
+ int quant = slice_args[slice_index].quant_idx;
+ int step = 1;
+ int bits_last = 0;
+ int bits = count_hq_slice(quant);
+ while ((bits > top) || (bits < bottom))
+ {
+ const int signed_step = bits > top ? +step : -step;
+ quant = clamp(quant + signed_step, 0, q_ceil-1);
+ bits = count_hq_slice(quant);
+ if (quant_buf[1] == quant)
+ {
+ quant = max(quant_buf[0], quant);
+ bits = quant == quant_buf[0] ? bits_last : bits;
+ break;
+ }
+ step = clamp(step / 2, 1, (q_ceil - 1) / 2);
+ quant_buf[1] = quant_buf[0];
+ quant_buf[0] = quant;
+ bits_last = bits;
+ }
+ int bytes = ssize_round(bits >> 3);
+ slice_args[slice_index].quant_idx = clamp(quant, 0, q_ceil-1);
+ slice_args[slice_index].bytes = bytes;
+ slice_sizes[gl_LocalInvocationIndex] = bytes;
+ barrier();
+
+ /* Prefix sum for all slices in current workgroup */
+ int total_bytes = 0;
+ for (int i = 0; i < gl_LocalInvocationIndex; i++)
+ total_bytes += slice_sizes[i];
+ slice_args[slice_index].pb_start = total_bytes;
+}
--
2.49.0
_______________________________________________
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".
