[FFmpeg-devel] [PATCH 08/12] ffv1enc_vulkan: refactor shaders slightly to support sharing

Lynne dev at lynne.ee
Sun Jan 19 12:36:26 EET 2025


The shaders were written to support sharing, but needed slight
tweaking.
---
 libavcodec/ffv1enc_vulkan.c           | 75 +++++++++++++++------------
 libavcodec/vulkan/ffv1_common.comp    | 24 +++++++--
 libavcodec/vulkan/ffv1_enc_setup.comp | 18 +------
 libavcodec/vulkan/ffv1_reset.comp     |  3 +-
 libavcodec/vulkan/rangecoder.comp     | 27 +++++-----
 5 files changed, 82 insertions(+), 65 deletions(-)

diff --git a/libavcodec/ffv1enc_vulkan.c b/libavcodec/ffv1enc_vulkan.c
index 243f472568..53d648bcec 100644
--- a/libavcodec/ffv1enc_vulkan.c
+++ b/libavcodec/ffv1enc_vulkan.c
@@ -136,14 +136,15 @@ typedef struct FFv1VkResetParameters {
     uint32_t context_count;
     uint8_t codec_planes;
     uint8_t key_frame;
-    uint8_t padding[3];
+    uint8_t version;
+    uint8_t micro_version;
+    uint8_t padding[1];
 } FFv1VkResetParameters;
 
 typedef struct FFv1VkParameters {
     VkDeviceAddress slice_state;
     VkDeviceAddress scratch_data;
     VkDeviceAddress out_data;
-    uint64_t slice_size_max;
 
     int32_t sar[2];
     uint32_t chroma_shift[2];
@@ -151,6 +152,7 @@ typedef struct FFv1VkParameters {
     uint32_t plane_state_size;
     uint32_t context_count;
     uint32_t crcref;
+    uint32_t slice_size_max;
 
     uint8_t bits_per_raw_sample;
     uint8_t context_model;
@@ -175,7 +177,6 @@ static void add_push_data(FFVulkanShader *shd)
     GLSLC(1,    u8buf slice_state;                                            );
     GLSLC(1,    u8buf scratch_data;                                           );
     GLSLC(1,    u8buf out_data;                                               );
-    GLSLC(1,    uint64_t slice_size_max;                                      );
     GLSLC(0,                                                                  );
     GLSLC(1,    ivec2 sar;                                                    );
     GLSLC(1,    uvec2 chroma_shift;                                           );
@@ -183,6 +184,7 @@ static void add_push_data(FFVulkanShader *shd)
     GLSLC(1,    uint plane_state_size;                                        );
     GLSLC(1,    uint context_count;                                           );
     GLSLC(1,    uint32_t crcref;                                              );
+    GLSLC(1,    uint32_t slice_size_max;                                      );
     GLSLC(0,                                                                  );
     GLSLC(1,    uint8_t bits_per_raw_sample;                                  );
     GLSLC(1,    uint8_t context_model;                                        );
@@ -492,7 +494,6 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
         .slice_state = slice_data_buf->address + f->slice_count*256,
         .scratch_data = tmp_data_buf->address,
         .out_data = out_data_buf->address,
-        .slice_size_max = out_data_buf->size / f->slice_count,
         .bits_per_raw_sample = f->bits_per_raw_sample,
         .sar[0] = pict->sample_aspect_ratio.num,
         .sar[1] = pict->sample_aspect_ratio.den,
@@ -501,6 +502,7 @@ static int vulkan_encode_ffv1_submit_frame(AVCodecContext *avctx,
         .plane_state_size = plane_state_size,
         .context_count = context_count,
         .crcref = f->crcref,
+        .slice_size_max = out_data_buf->size / f->slice_count,
         .context_model = fv->ctx.context_model,
         .version = f->version,
         .micro_version = f->micro_version,
@@ -966,7 +968,6 @@ static void define_shared_code(AVCodecContext *avctx, FFVulkanShader *shd)
     GLSLF(0, #define TYPE int%i_t                                        ,smp_bits);
     GLSLF(0, #define VTYPE2 i%ivec2                                      ,smp_bits);
     GLSLF(0, #define VTYPE3 i%ivec3                                      ,smp_bits);
-    GLSLD(ff_source_common_comp);
     GLSLD(ff_source_rangecoder_comp);
 
     if (f->ac == AC_GOLOMB_RICE)
@@ -993,6 +994,10 @@ static int init_setup_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
                           1, 1, 1,
                           0));
 
+    /* Common codec header */
+    GLSLD(ff_source_common_comp);
+    add_push_data(shd);
+
     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);
@@ -1038,8 +1043,6 @@ static int init_setup_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
     };
     RET(ff_vk_shader_add_descriptor_set(&fv->s, shd, desc_set, 2, 0, 0));
 
-    add_push_data(shd);
-
     GLSLD(ff_source_ffv1_enc_setup_comp);
 
     RET(spv->compile_shader(&fv->s, spv, shd, &spv_data, &spv_len, "main",
@@ -1074,6 +1077,22 @@ static int init_reset_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
                           wg_dim, 1, 1,
                           0));
 
+    /* Common codec header */
+    GLSLD(ff_source_common_comp);
+
+    GLSLC(0, layout(push_constant, scalar) uniform pushConstants {             );
+    GLSLC(1,    u8buf slice_state;                                             );
+    GLSLC(1,    uint plane_state_size;                                         );
+    GLSLC(1,    uint context_count;                                            );
+    GLSLC(1,    uint8_t codec_planes;                                          );
+    GLSLC(1,    uint8_t key_frame;                                             );
+    GLSLC(1,    uint8_t version;                                               );
+    GLSLC(1,    uint8_t micro_version;                                         );
+    GLSLC(1,    uint8_t padding[1];                                            );
+    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);
@@ -1110,17 +1129,6 @@ static int init_reset_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
     };
     RET(ff_vk_shader_add_descriptor_set(&fv->s, shd, desc_set, 1, 0, 0));
 
-    GLSLC(0, layout(push_constant, scalar) uniform pushConstants {             );
-    GLSLC(1,    u8buf slice_state;                                             );
-    GLSLC(1,    uint plane_state_size;                                         );
-    GLSLC(1,    uint context_count;                                            );
-    GLSLC(1,    uint8_t codec_planes;                                          );
-    GLSLC(1,    uint8_t key_frame;                                             );
-    GLSLC(1,    uint8_t padding[3];                                            );
-    GLSLC(0, };                                                                );
-    ff_vk_shader_add_push_const(shd, 0, sizeof(FFv1VkResetParameters),
-                                VK_SHADER_STAGE_COMPUTE_BIT);
-
     GLSLD(ff_source_ffv1_reset_comp);
 
     RET(spv->compile_shader(&fv->s, spv, shd, &spv_data, &spv_len, "main",
@@ -1164,6 +1172,18 @@ static int init_rct_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
                           wg_count, wg_count, 1,
                           0));
 
+    define_shared_code(avctx, shd);
+
+    GLSLC(0, layout(push_constant, scalar) uniform pushConstants {             );
+    GLSLC(1,    int offset;                                                    );
+    GLSLC(1,    uint8_t bits;                                                  );
+    GLSLC(1,    uint8_t planar_rgb;                                            );
+    GLSLC(1,    uint8_t transparency;                                          );
+    GLSLC(1,    uint8_t padding[1];                                            );
+    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);
@@ -1187,8 +1207,6 @@ static int init_rct_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
     };
     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",
@@ -1220,16 +1238,6 @@ static int init_rct_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
     };
     RET(ff_vk_shader_add_descriptor_set(&fv->s, shd, desc_set, 3, 0, 0));
 
-    GLSLC(0, layout(push_constant, scalar) uniform pushConstants {             );
-    GLSLC(1,    int offset;                                                    );
-    GLSLC(1,    uint8_t bits;                                                  );
-    GLSLC(1,    uint8_t planar_rgb;                                            );
-    GLSLC(1,    uint8_t transparency;                                          );
-    GLSLC(1,    uint8_t padding[1];                                            );
-    GLSLC(0, };                                                                );
-    ff_vk_shader_add_push_const(shd, 0, sizeof(FFv1VkRCTParameters),
-                                VK_SHADER_STAGE_COMPUTE_BIT);
-
     GLSLD(ff_source_ffv1_enc_rct_comp);
 
     RET(spv->compile_shader(&fv->s, spv, shd, &spv_data, &spv_len, "main",
@@ -1268,6 +1276,11 @@ static int init_encode_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
                           1, 1, 1,
                           0));
 
+    /* Common codec header */
+    GLSLD(ff_source_common_comp);
+
+    add_push_data(shd);
+
     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);
@@ -1328,8 +1341,6 @@ static int init_encode_shader(AVCodecContext *avctx, FFVkSPIRVCompiler *spv)
     };
     RET(ff_vk_shader_add_descriptor_set(&fv->s, shd, desc_set, 3, 0, 0));
 
-    add_push_data(shd);
-
     /* Assemble the shader body */
     GLSLD(ff_source_ffv1_enc_common_comp);
 
diff --git a/libavcodec/vulkan/ffv1_common.comp b/libavcodec/vulkan/ffv1_common.comp
index 5b4a882367..604d03b2de 100644
--- a/libavcodec/vulkan/ffv1_common.comp
+++ b/libavcodec/vulkan/ffv1_common.comp
@@ -22,17 +22,18 @@
 
 struct SliceContext {
     RangeCoder c;
-
-#ifdef GOLOMB
     PutBitContext pb; /* 8*8 bytes */
-#endif
 
     ivec2 slice_dim;
     ivec2 slice_pos;
     ivec2 slice_rct_coef;
+    u8vec4 quant_table_idx;
+    uint context_count;
 
     uint hdr_len; // only used for golomb
-    int slice_coding_mode;
+
+    uint slice_coding_mode;
+    bool slice_reset_contexts;
 };
 
 /* -1, { -1, 0 } */
@@ -72,3 +73,18 @@ const uint32_t log2_run[41] = {
     16, 17, 18, 19, 20, 21, 22, 23,
     24,
 };
+
+uint slice_coord(uint width, uint sx, uint num_h_slices, uint chroma_shift)
+{
+    uint mpw = 1 << chroma_shift;
+    uint awidth = align(width, mpw);
+
+    if ((version < 4) || ((version == 4) && (micro_version < 3)))
+        return width * sx / num_h_slices;
+
+    sx = (2 * awidth * sx + num_h_slices * mpw) / (2 * num_h_slices * mpw) * mpw;
+    if (sx == awidth)
+        sx = width;
+
+    return sx;
+}
diff --git a/libavcodec/vulkan/ffv1_enc_setup.comp b/libavcodec/vulkan/ffv1_enc_setup.comp
index b861e25f74..23f09b2af6 100644
--- a/libavcodec/vulkan/ffv1_enc_setup.comp
+++ b/libavcodec/vulkan/ffv1_enc_setup.comp
@@ -20,21 +20,6 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
-uint slice_coord(uint width, uint sx, uint num_h_slices, uint chroma_shift)
-{
-    uint mpw = 1 << chroma_shift;
-    uint awidth = align(width, mpw);
-
-    if ((version < 4) || ((version == 4) && (micro_version < 3)))
-        return width * sx / num_h_slices;
-
-    sx = (2 * awidth * sx + num_h_slices * mpw) / (2 * num_h_slices * mpw) * mpw;
-    if (sx == awidth)
-        sx = width;
-
-    return sx;
-}
-
 void init_slice(out SliceContext sc, const uint slice_idx)
 {
     /* Set coordinates */
@@ -52,6 +37,7 @@ void init_slice(out SliceContext sc, const uint slice_idx)
     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;
 
     rac_init(sc.c,
              OFFBUF(u8buf, out_data, slice_idx * slice_size_max),
@@ -105,7 +91,7 @@ void write_slice_header(inout SliceContext sc, uint64_t state)
     put_symbol_unsigned(sc.c, state, sar.y);
 
     if (version >= 4) {
-        put_rac_full(sc.c, state, sc.slice_coding_mode == 1);
+        put_rac_full(sc.c, state, sc.slice_reset_contexts);
         put_symbol_unsigned(sc.c, state, sc.slice_coding_mode);
         if (sc.slice_coding_mode != 1 && colorspace == 1) {
             put_symbol_unsigned(sc.c, state, sc.slice_rct_coef.y);
diff --git a/libavcodec/vulkan/ffv1_reset.comp b/libavcodec/vulkan/ffv1_reset.comp
index c7c7962850..1b87ca754e 100644
--- a/libavcodec/vulkan/ffv1_reset.comp
+++ b/libavcodec/vulkan/ffv1_reset.comp
@@ -24,7 +24,8 @@ void main(void)
 {
     const uint slice_idx = gl_WorkGroupID.y*gl_NumWorkGroups.x + gl_WorkGroupID.x;
 
-    if (slice_ctx[slice_idx].slice_coding_mode == 0 && key_frame == 0)
+    if (key_frame == 0 &&
+        slice_ctx[slice_idx].slice_reset_contexts == false)
         return;
 
     uint64_t slice_state_off = uint64_t(slice_state) +
diff --git a/libavcodec/vulkan/rangecoder.comp b/libavcodec/vulkan/rangecoder.comp
index 848a056fb1..6e3b9c1238 100644
--- a/libavcodec/vulkan/rangecoder.comp
+++ b/libavcodec/vulkan/rangecoder.comp
@@ -21,8 +21,9 @@
  */
 
 struct RangeCoder {
-    u8buf bytestream_start;
-    u8buf bytestream;
+    uint64_t bytestream_start;
+    uint64_t bytestream;
+    uint64_t bytestream_end;
 
     uint low;
     uint16_t range;
@@ -34,28 +35,29 @@ struct RangeCoder {
 void renorm_encoder_full(inout RangeCoder c)
 {
     int bs_cnt = 0;
+    u8buf bytestream = u8buf(c.bytestream);
 
     if (c.outstanding_byte == 0xFF) {
         c.outstanding_byte = uint8_t(c.low >> 8);
     } else if (c.low <= 0xFF00) {
-        c.bytestream[bs_cnt++].v = c.outstanding_byte;
+        bytestream[bs_cnt++].v = c.outstanding_byte;
         uint16_t cnt = c.outstanding_count;
         for (; cnt > 0; cnt--)
-            c.bytestream[bs_cnt++].v = uint8_t(0xFF);
+            bytestream[bs_cnt++].v = uint8_t(0xFF);
         c.outstanding_count = uint16_t(0);
         c.outstanding_byte = uint8_t(c.low >> 8);
     } else if (c.low >= 0x10000) {
-        c.bytestream[bs_cnt++].v = c.outstanding_byte + uint8_t(1);
+        bytestream[bs_cnt++].v = c.outstanding_byte + uint8_t(1);
         uint16_t cnt = c.outstanding_count;
         for (; cnt > 0; cnt--)
-            c.bytestream[bs_cnt++].v = uint8_t(0x00);
+            bytestream[bs_cnt++].v = uint8_t(0x00);
         c.outstanding_count = uint16_t(0);
         c.outstanding_byte = uint8_t(bitfieldExtract(c.low, 8, 8));
     } else {
         c.outstanding_count++;
     }
 
-    c.bytestream = OFFBUF(u8buf, c.bytestream, bs_cnt);
+    c.bytestream += bs_cnt;
     c.range <<= 8;
     c.low = bitfieldInsert(0, c.low, 8, 8);
 }
@@ -74,10 +76,10 @@ void renorm_encoder(inout RangeCoder c)
         return;
     }
 
-    u8buf bs = c.bytestream;
+    u8buf bs = u8buf(c.bytestream);
     uint8_t outstanding_byte = c.outstanding_byte;
 
-    c.bytestream        = OFFBUF(u8buf, bs, oc);
+    c.bytestream        = uint64_t(bs) + oc;
     c.outstanding_count = uint16_t(0);
     c.outstanding_byte  = uint8_t(low >> 8);
 
@@ -179,10 +181,11 @@ uint32_t rac_terminate(inout RangeCoder c)
     return uint32_t(uint64_t(c.bytestream) - uint64_t(c.bytestream_start));
 }
 
-void rac_init(out RangeCoder r, u8buf data, uint64_t buf_size)
+void rac_init(out RangeCoder r, u8buf data, uint buf_size)
 {
-    r.bytestream_start = data;
-    r.bytestream = data;
+    r.bytestream_start = uint64_t(data);
+    r.bytestream = uint64_t(data);
+    r.bytestream_end = uint64_t(data) + buf_size;
     r.low = 0;
     r.range = uint16_t(0xFF00);
     r.outstanding_count = uint16_t(0);
-- 
2.47.1


More information about the ffmpeg-devel mailing list