[FFmpeg-devel] [PATCH 11/16] ffv1enc_vulkan: implement RCT search for level >= 4
Lynne
dev at lynne.ee
Wed May 14 22:02:40 EEST 2025
---
libavcodec/ffv1enc_vulkan.c | 204 ++++++++++++++++++++++++-
libavcodec/vulkan/Makefile | 2 +-
libavcodec/vulkan/ffv1_enc_setup.comp | 6 +-
libavcodec/vulkan/ffv1_rct_search.comp | 139 +++++++++++++++++
4 files changed, 346 insertions(+), 5 deletions(-)
create mode 100644 libavcodec/vulkan/ffv1_rct_search.comp
diff --git a/libavcodec/ffv1enc_vulkan.c b/libavcodec/ffv1enc_vulkan.c
index 5de16d5b02..d9e12f5fae 100644
--- a/libavcodec/ffv1enc_vulkan.c
+++ b/libavcodec/ffv1enc_vulkan.c
@@ -74,6 +74,7 @@ typedef struct VulkanEncodeFFv1Context {
size_t max_heap_size;
FFVulkanShader setup;
+ FFVulkanShader rct_search;
FFVulkanShader reset;
FFVulkanShader enc;
@@ -101,6 +102,7 @@ typedef struct VulkanEncodeFFv1Context {
int num_h_slices;
int num_v_slices;
int force_pcm;
+ int optimize_rct;
int is_rgb;
int ppi;
@@ -112,6 +114,7 @@ extern const char *ff_source_rangecoder_comp;
extern const char *ff_source_ffv1_vlc_comp;
extern const char *ff_source_ffv1_common_comp;
extern const char *ff_source_ffv1_reset_comp;
+extern const char *ff_source_ffv1_rct_search_comp;
extern const char *ff_source_ffv1_enc_setup_comp;
extern const char *ff_source_ffv1_enc_comp;
@@ -147,7 +150,8 @@ typedef struct FFv1VkParameters {
uint8_t ec;
uint8_t ppi;
uint8_t chunks;
- uint8_t padding[4];
+ uint8_t rct_search;
+ uint8_t padding[3];
} FFv1VkParameters;
static void add_push_data(FFVulkanShader *shd)
@@ -184,12 +188,76 @@ static void add_push_data(FFVulkanShader *shd)
GLSLC(1, uint8_t ec; );
GLSLC(1, uint8_t ppi; );
GLSLC(1, uint8_t chunks; );
- GLSLC(1, uint8_t padding[4]; );
+ GLSLC(1, uint8_t rct_search; );
+ GLSLC(1, uint8_t padding[3]; );
GLSLC(0, }; );
ff_vk_shader_add_push_const(shd, 0, sizeof(FFv1VkParameters),
VK_SHADER_STAGE_COMPUTE_BIT);
}
+typedef struct FFv1VkRCTSearchParameters {
+ int fmt_lut[4];
+ int rct_offset;
+ uint8_t planar_rgb;
+ uint8_t transparency;
+ uint8_t key_frame;
+ uint8_t force_pcm;
+ uint8_t version;
+ uint8_t micro_version;
+ uint8_t padding[2];
+} FFv1VkRCTSearchParameters;
+
+static int run_rct_search(AVCodecContext *avctx, FFVkExecContext *exec,
+ AVFrame *enc_in, VkImageView *enc_in_views,
+ FFVkBuffer *slice_data_buf, uint32_t slice_data_size)
+{
+ VulkanEncodeFFv1Context *fv = avctx->priv_data;
+ FFV1Context *f = &fv->ctx;
+ FFVulkanFunctions *vk = &fv->s.vkfn;
+ AVHWFramesContext *src_hwfc = (AVHWFramesContext *)enc_in->hw_frames_ctx->data;
+ FFv1VkRCTSearchParameters pd;
+
+ /* Update descriptors */
+ ff_vk_shader_update_desc_buffer(&fv->s, exec, &fv->rct_search,
+ 0, 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_search,
+ enc_in, enc_in_views,
+ 0, 1,
+ VK_IMAGE_LAYOUT_GENERAL,
+ VK_NULL_HANDLE);
+
+ ff_vk_exec_bind_shader(&fv->s, exec, &fv->rct_search);
+
+ pd = (FFv1VkRCTSearchParameters) {
+ .rct_offset = 1 << 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,
+ .key_frame = f->key_frame,
+ .force_pcm = fv->force_pcm,
+ .version = f->version,
+ .micro_version = f->micro_version,
+ };
+
+ 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);
+
+ ff_vk_shader_update_push_const(&fv->s, exec, &fv->rct_search,
+ VK_SHADER_STAGE_COMPUTE_BIT,
+ 0, sizeof(pd), &pd);
+
+ vk->CmdDispatch(exec->buf, fv->ctx.num_h_slices, fv->ctx.num_v_slices, 1);
+
+ return 0;
+}
+
static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
FFVkExecContext *exec,
const AVFrame *pict)
@@ -366,6 +434,25 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
};
}
+ if (fv->optimize_rct) {
+ RET(run_rct_search(avctx, exec,
+ src, src_views,
+ slice_data_buf, slice_data_size));
+
+ buf_bar[nb_buf_bar++] = (VkBufferMemoryBarrier2) {
+ .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2,
+ .srcStageMask = slice_data_buf->stage,
+ .dstStageMask = VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
+ .srcAccessMask = slice_data_buf->access,
+ .dstAccessMask = VK_ACCESS_2_SHADER_STORAGE_WRITE_BIT,
+ .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+ .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+ .buffer = slice_data_buf->buf,
+ .size = slice_data_size*f->slice_count,
+ .offset = 0,
+ };
+ }
+
vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
.pImageMemoryBarriers = img_bar,
@@ -412,6 +499,7 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
.ec = f->ec,
.ppi = fv->ppi,
.chunks = fv->chunks,
+ .rct_search = fv->optimize_rct,
};
/* For some reason the C FFv1 encoder/decoder treats these differently */
@@ -920,6 +1008,103 @@ static void define_shared_code(AVCodecContext *avctx, FFVulkanShader *shd)
GLSLD(ff_source_ffv1_common_comp);
}
+static int init_rct_search_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
+{
+ int err;
+ VulkanEncodeFFv1Context *fv = avctx->priv_data;
+ FFVulkanShader *shd = &fv->rct_search;
+ FFVulkanDescriptorSetBinding *desc_set;
+
+ uint8_t *spv_data;
+ size_t spv_len;
+ void *spv_opaque = NULL;
+
+ RET(ff_vk_shader_init(&fv->s, shd, "ffv1_rct_search",
+ VK_SHADER_STAGE_COMPUTE_BIT,
+ (const char *[]) { "GL_EXT_buffer_reference",
+ "GL_EXT_buffer_reference2",
+ "GL_EXT_null_initializer" }, 3,
+ 32, 32, 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 rct_offset; );
+ GLSLC(1, uint8_t planar_rgb; );
+ GLSLC(1, uint8_t transparency; );
+ GLSLC(1, uint8_t key_frame; );
+ GLSLC(1, uint8_t force_pcm; );
+ GLSLC(1, uint8_t version; );
+ GLSLC(1, uint8_t micro_version; );
+ GLSLC(1, uint8_t padding[3]; );
+ GLSLC(0, }; );
+ ff_vk_shader_add_push_const(shd, 0, sizeof(FFv1VkResetParameters),
+ 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);
+
+ /* Never used */
+ 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, 1));
+
+ define_shared_code(avctx, shd);
+
+ desc_set = (FFVulkanDescriptorSetBinding []) {
+ {
+ .name = "slice_data_buf",
+ .type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+ .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,
+ },
+ };
+ RET(ff_vk_shader_add_descriptor_set(&fv->s, shd, desc_set, 2, 0, 0));
+
+ GLSLD(ff_source_ffv1_rct_search_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_setup_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
{
int err;
@@ -1417,6 +1602,17 @@ static av_cold int vulkan_encode_ffv1_init(AVCodecContext *avctx)
if (!fv->is_rgb && f->bits_per_raw_sample > 8)
fv->rep_fmt = FF_VK_REP_INT;
+ /* Init rct search shader */
+ fv->optimize_rct = fv->is_rgb && f->version >= 4 &&
+ !fv->force_pcm && fv->optimize_rct;
+ if (fv->optimize_rct) {
+ err = init_rct_search_shader(avctx, spv);
+ if (err < 0) {
+ spv->uninit(&spv);
+ return err;
+ }
+ }
+
/* Init setup shader */
err = init_setup_shader(avctx, spv);
if (err < 0) {
@@ -1528,6 +1724,7 @@ static av_cold int vulkan_encode_ffv1_close(AVCodecContext *avctx)
ff_vk_shader_free(&fv->s, &fv->enc);
ff_vk_shader_free(&fv->s, &fv->reset);
ff_vk_shader_free(&fv->s, &fv->setup);
+ ff_vk_shader_free(&fv->s, &fv->rct_search);
if (fv->exec_ctx_info) {
for (int i = 0; i < fv->async_depth; i++) {
@@ -1591,6 +1788,9 @@ static const AVOption vulkan_encode_ffv1_options[] = {
{ "force_pcm", "Code all slices with no prediction", OFFSET(force_pcm), AV_OPT_TYPE_BOOL,
{ .i64 = 0 }, 0, 1, VE },
+ { "rct_search", "Run a search for RCT parameters (level 4 only)", OFFSET(optimize_rct), AV_OPT_TYPE_BOOL,
+ { .i64 = 1 }, 0, 1, VE },
+
{ "async_depth", "Internal parallelization depth", OFFSET(async_depth), AV_OPT_TYPE_INT,
{ .i64 = 1 }, 1, INT_MAX, VE },
diff --git a/libavcodec/vulkan/Makefile b/libavcodec/vulkan/Makefile
index 4bbcb38c6a..729cb4f15c 100644
--- a/libavcodec/vulkan/Makefile
+++ b/libavcodec/vulkan/Makefile
@@ -7,7 +7,7 @@ OBJS-$(CONFIG_FFV1_VULKAN_ENCODER) += vulkan/common.o \
vulkan/rangecoder.o vulkan/ffv1_vlc.o \
vulkan/ffv1_common.o vulkan/ffv1_reset.o \
vulkan/ffv1_enc_rct.o vulkan/ffv1_enc_setup.o \
- vulkan/ffv1_enc.o
+ vulkan/ffv1_rct_search.o vulkan/ffv1_enc.o
OBJS-$(CONFIG_FFV1_VULKAN_HWACCEL) += vulkan/common.o \
vulkan/rangecoder.o vulkan/ffv1_vlc.o \
diff --git a/libavcodec/vulkan/ffv1_enc_setup.comp b/libavcodec/vulkan/ffv1_enc_setup.comp
index 6f21e47523..5f8e6704b0 100644
--- a/libavcodec/vulkan/ffv1_enc_setup.comp
+++ b/libavcodec/vulkan/ffv1_enc_setup.comp
@@ -22,7 +22,7 @@
uint8_t state[CONTEXT_SIZE];
-void init_slice(out SliceContext sc, const uint slice_idx)
+void init_slice(inout SliceContext sc, const uint slice_idx)
{
/* Set coordinates */
uvec2 img_size = imageSize(src[0]);
@@ -37,11 +37,13 @@ void init_slice(out SliceContext sc, const uint slice_idx)
sc.slice_pos = ivec2(sxs, sys);
sc.slice_dim = ivec2(sxe - sxs, sye - sys);
- sc.slice_rct_coef = ivec2(1, 1);
sc.slice_coding_mode = int(force_pcm == 1);
sc.slice_reset_contexts = sc.slice_coding_mode == 1;
sc.quant_table_idx = u8vec3(context_model);
+ if ((rct_search == 0) || (sc.slice_coding_mode == 1))
+ sc.slice_rct_coef = ivec2(1, 1);
+
rac_init(sc.c,
OFFBUF(u8buf, out_data, slice_idx * slice_size_max),
slice_size_max);
diff --git a/libavcodec/vulkan/ffv1_rct_search.comp b/libavcodec/vulkan/ffv1_rct_search.comp
new file mode 100644
index 0000000000..055bde46c4
--- /dev/null
+++ b/libavcodec/vulkan/ffv1_rct_search.comp
@@ -0,0 +1,139 @@
+/*
+ * FFv1 codec
+ *
+ * Copyright (c) 2024 Lynne <dev at lynne.ee>
+ *
+ * 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
+ */
+
+ivec3 load_components(ivec2 pos)
+{
+ ivec3 pix = ivec3(imageLoad(src[0], pos));
+ if (planar_rgb != 0) {
+ for (int i = 1; i < 3; i++)
+ pix[i] = int(imageLoad(src[i], pos)[0]);
+ }
+
+ return ivec3(pix[fmt_lut[0]], pix[fmt_lut[1]], pix[fmt_lut[2]]);
+}
+
+#define NUM_CHECKS 15
+const ivec2 rct_y_coeff[NUM_CHECKS] = {
+ ivec2(0, 0), // 4G
+
+ ivec2(0, 1), // 3G + B
+ ivec2(1, 0), // R + 3G
+ ivec2(1, 1), // R + 2G + B
+
+ ivec2(0, 2), // 2G + 2B
+ ivec2(2, 0), // 2R + 2G
+ ivec2(2, 2), // 2R + 2B
+
+ ivec2(0, 3), // 1G + 3B
+ ivec2(3, 0), // 3R + 1G
+
+ ivec2(0, 4), // 4B
+ ivec2(4, 0), // 4R
+
+ ivec2(1, 2), // R + G + 2B
+ ivec2(2, 1), // 2R + G + B
+
+ ivec2(3, 1), // 3R + B
+ ivec2(1, 3), // R + 3B
+};
+
+shared ivec3 pix_buf[gl_WorkGroupSize.x + 1][gl_WorkGroupSize.y + 1] = { };
+
+ivec3 transform_sample(ivec3 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;
+ return pix;
+}
+
+uint get_dist(ivec3 cur)
+{
+ ivec3 LL = pix_buf[gl_LocalInvocationID.x + 0][gl_LocalInvocationID.y + 1];
+ ivec3 TL = pix_buf[gl_LocalInvocationID.x + 0][gl_LocalInvocationID.y + 0];
+ ivec3 TT = pix_buf[gl_LocalInvocationID.x + 1][gl_LocalInvocationID.y + 0];
+
+ ivec3 pred = ivec3(predict(LL.r, ivec2(TL.r, TT.r)),
+ predict(LL.g, ivec2(TL.g, TT.g)),
+ predict(LL.b, ivec2(TL.b, TT.b)));
+
+ uvec3 c = abs(pred - cur);
+ return mid_pred(c.r, c.g, c.b);
+}
+
+shared uint score_cols[gl_WorkGroupSize.y] = { };
+shared uint score_mode[16] = { };
+
+void process(ivec2 pos)
+{
+ ivec3 pix = load_components(pos);
+
+ for (int i = 0; i < NUM_CHECKS; i++) {
+ ivec3 tx_pix = transform_sample(pix, rct_y_coeff[i]);
+ pix_buf[gl_LocalInvocationID.x + 1][gl_LocalInvocationID.y + 1] = tx_pix;
+ memoryBarrierShared();
+
+ uint dist = get_dist(tx_pix);
+ atomicAdd(score_mode[i], dist);
+ }
+}
+
+void coeff_search(inout SliceContext sc)
+{
+ uvec2 img_size = imageSize(src[0]);
+ uint sxs = slice_coord(img_size.x, gl_WorkGroupID.x + 0,
+ gl_NumWorkGroups.x, 0);
+ uint sxe = slice_coord(img_size.x, gl_WorkGroupID.x + 1,
+ gl_NumWorkGroups.x, 0);
+ uint sys = slice_coord(img_size.y, gl_WorkGroupID.y + 0,
+ gl_NumWorkGroups.y, 0);
+ uint sye = slice_coord(img_size.y, gl_WorkGroupID.y + 1,
+ gl_NumWorkGroups.y, 0);
+
+ for (uint y = sys + gl_LocalInvocationID.y; y < sye; y += gl_WorkGroupSize.y) {
+ for (uint x = sxs + gl_LocalInvocationID.x; x < sxe; x += gl_WorkGroupSize.x) {
+ process(ivec2(x, y));
+ }
+ }
+
+ if (gl_LocalInvocationID.x == 0 && gl_LocalInvocationID.y == 0) {
+ uint min_score = 0xFFFFFFFF;
+ uint min_idx = 3;
+ for (int i = 0; i < NUM_CHECKS; i++) {
+ if (score_mode[i] < min_score) {
+ min_score = score_mode[i];
+ min_idx = i;
+ }
+ }
+ sc.slice_rct_coef = rct_y_coeff[min_idx];
+ }
+}
+
+void main(void)
+{
+ if (force_pcm == 1)
+ return;
+ const uint slice_idx = gl_WorkGroupID.y*gl_NumWorkGroups.x + gl_WorkGroupID.x;
+ coeff_search(slice_ctx[slice_idx]);
+}
--
2.49.0.395.g12beb8f557c
More information about the ffmpeg-devel
mailing list