This is an automated email from the git hooks/post-receive script. Git pushed a commit to branch master in repository ffmpeg. commit 58bd5ad6309c9a0aa27b9aad8bad6dc1415bcb36 Author: Lynne <dev@lynne.ee> AuthorDate: Wed Dec 24 04:10:39 2025 +0100 Commit: Lynne <dev@lynne.ee> CommitDate: Wed Dec 31 15:00:47 2025 +0100 vulkan/prores_raw_idct: use the same prores_idct method for loading coeffs This saves 2 barriers. Also implement workbank avoidance. --- libavcodec/vulkan/prores_raw_idct.comp | 26 ++++++++++++++------------ 1 file changed, 14 insertions(+), 12 deletions(-) diff --git a/libavcodec/vulkan/prores_raw_idct.comp b/libavcodec/vulkan/prores_raw_idct.comp index ffd71d1d73..c9850d17d7 100644 --- a/libavcodec/vulkan/prores_raw_idct.comp +++ b/libavcodec/vulkan/prores_raw_idct.comp @@ -63,30 +63,32 @@ void main(void) uint8_t qmat_buf[64] = qmat; [[unroll]] - for (uint i = gl_LocalInvocationID.x; i < 64; i += gl_WorkGroupSize.x) { - int v = int(imageLoad(dst, offs + 2*ivec2(BLOCK_ID*8, 0) + scan[i])[0]); + for (uint y = 0; y < 8; y++) { + uint block_off = y*8 + ROW_ID; + int v = int(imageLoad(dst, offs + 2*ivec2(BLOCK_ID*8, 0) + scan[block_off])[0]); float vf = float(sign_extend(v, 16)) / 32768.0; - vf *= qmat_buf[i] * qscale; - blocks[BLOCK_ID][COMP_ID*64 + i] = (vf / (64*4.56)) * - idct_scale[i]; + vf *= qmat_buf[block_off] * qscale; + blocks[BLOCK_ID][COMP_ID*72 + y*9 + ROW_ID] = (vf / (64*4.56)) * + idct_scale[block_off]; } + /* Column-wise iDCT */ + idct8(BLOCK_ID, COMP_ID*72 + ROW_ID, 9); barrier(); - idct8(BLOCK_ID, COMP_ID*64 + ROW_ID*8, 1); - blocks[BLOCK_ID][COMP_ID*64 + ROW_ID] += 0.5; + blocks[BLOCK_ID][COMP_ID*72 + ROW_ID * 9] += 0.5f; + /* Row-wise iDCT */ + idct8(BLOCK_ID, COMP_ID*72 + ROW_ID * 9, 1); barrier(); - idct8(BLOCK_ID, COMP_ID*64 + ROW_ID, 8); - barrier(); [[unroll]] - for (uint i = gl_LocalInvocationID.x; i < 64; i += gl_WorkGroupSize.x) { - int v = int(round(blocks[BLOCK_ID][COMP_ID*64 + i]*4095.0)); + for (uint y = 0; y < 8; y++) { + int v = int(round(blocks[BLOCK_ID][COMP_ID*72 + y*9 + ROW_ID]*4095.0)); v = clamp(v, 0, 4095); v <<= 4; imageStore(dst, - offs + 2*ivec2(BLOCK_ID*8 + (i & 7), i >> 3), + offs + 2*ivec2(BLOCK_ID*8 + ROW_ID, y), ivec4(v)); } }