wip: mc02 v2 shared-tile

bench: H.264 primitive bench now measures both substrates + comparison table
Closes task #166 (re-measure R-bands on post-buffer-pool dispatch path). Now that all H.264 hot-path primitives have QPU shaders and the dispatch overhead has been hammered down (tasks #160 buffer pool, #161 persistent command buffer), bench_h264_primitives no longer measures one column. Two passes — CPU NEON and QPU V3D7 compute — with a side-by-side per-kernel comparison and ratio. Headline result on hertz (Pi 5 V3D 7.1, 30 iters x 5 warmup): kernel CPU ns/op QPU ns/op winner IDCT 4x4 luma 10.79 2.47 QPU 4.36x IDCT 8x8 luma 29.69 9.23 QPU 3.22x Deblock luma_v 17.58 10.21 QPU 1.72x Deblock luma_h 38.41 9.98 QPU 3.85x qpel mc20 (8x8) 28.24 9.66 QPU 2.92x qpel mc02 (8x8) 16.96 20.54 CPU 1.21x qpel mc22 (8x8) 71.58 9.64 QPU 7.43x 1080p worst-case sum (IDCT4 + deblock luma + qpel mc22): CPU NEON only: 5.57 ms QPU only: 1.30 ms (CPU/QPU sum ratio = 4.30x) Reverses PR #10's verdict (which had CPU NEON 4x faster than QPU for IDCT-only) — the buffer-pool + persistent-cmdbuf wins land hard. Only qpel mc02 still shows CPU ahead, marginally (single- axis vertical filter, row-strided memory pattern unfriendly to the WG layout — left as a follow-up for cycle-9-style targeted tuning). Substrate decree (2026-05-23) stays in force as policy — these numbers retroactively justify it. Also tightens test_api_h264's startup recipe print: the stale "(CPU)" / "(CPU, no QPU H shader yet)" / "(CPU, bS=4 set)" labels next to deblock_lh, deblock_cv, deblock_ch and deblock_*_intra are now wrong since PRs #28, #29, #35 (those kernels are on QPU).
2026-05-25 21:28:04 +02:00 · 2026-05-25 20:42:39 +02:00 · 2026-05-25 18:36:10 +00:00 · 2026-05-25 20:30:07 +02:00 · 2026-05-25 18:23:11 +00:00 · 2026-05-25 20:22:33 +02:00
36 changed files with 3127 additions and 160 deletions
@@ -317,6 +317,22 @@ if (DAEDALUS_BUILD_VULKAN)
        VERBATIM
    )
    # Intra (bS=4) deblock shaders — strong/weak filter selector per
    # H.264 §8.3.2.3.  4 variants (luma_v/h + chroma_v/h).
    foreach(_kind luma_v_intra luma_h_intra chroma_v_intra chroma_h_intra)
        set(_spv ${CMAKE_BINARY_DIR}/v3d_h264deblock_${_kind}.spv)
        add_custom_command(
            OUTPUT ${_spv}
            COMMAND ${GLSLANG_VALIDATOR} -V --target-env vulkan1.3
                    -o ${_spv}
                    ${CMAKE_SOURCE_DIR}/src/v3d_h264deblock_${_kind}.comp
            DEPENDS ${CMAKE_SOURCE_DIR}/src/v3d_h264deblock_${_kind}.comp
            COMMENT "glslang: v3d_h264deblock_${_kind}.comp -> .spv"
            VERBATIM
        )
        set(H264DEBLOCK_${_kind}_SPV ${_spv})
    endforeach()
    set(H264_IDCT4_SPV ${CMAKE_BINARY_DIR}/v3d_h264_idct4.spv)
    add_custom_command(
        OUTPUT ${H264_IDCT4_SPV}
@@ -372,7 +388,41 @@ if (DAEDALUS_BUILD_VULKAN)
        VERBATIM
    )
-    add_custom_target(daedalus_shaders ALL DEPENDS ${NOOP_SPV} ${IDCT8_SPV} ${LPF_SPV} ${MC_SPV} ${LPF8_SPV} ${CDEF_SPV} ${H264DEBLOCK_SPV} ${H264DEBLOCK_H_SPV} ${H264DEBLOCK_CHROMA_V_SPV} ${H264DEBLOCK_CHROMA_H_SPV} ${H264_IDCT4_SPV} ${H264_IDCT8_SPV} ${H264_QPEL_MC20_SPV} ${H264_QPEL_MC02_SPV} ${H264_QPEL_MC22_SPV})
+    # Quarter-pel single-axis variants (mc10/30/01/03) + diagonal
    # variants (mc11/12/13/21/23/31/32/33) — each composes 1-2 half-pel
    # results with optional L2 averaging.  Same WG geometry as mc20/mc02.
    foreach(_mc mc10 mc30 mc01 mc03 mc11 mc12 mc13 mc21 mc23 mc31 mc32 mc33)
        set(_spv ${CMAKE_BINARY_DIR}/v3d_h264_qpel_${_mc}.spv)
        add_custom_command(
            OUTPUT ${_spv}
            COMMAND ${GLSLANG_VALIDATOR} -V --target-env vulkan1.3
                    -o ${_spv}
                    ${CMAKE_SOURCE_DIR}/src/v3d_h264_qpel_${_mc}.comp
            DEPENDS ${CMAKE_SOURCE_DIR}/src/v3d_h264_qpel_${_mc}.comp
            COMMENT "glslang: v3d_h264_qpel_${_mc}.comp -> .spv"
            VERBATIM
        )
        set(H264_QPEL_${_mc}_SPV ${_spv})
    endforeach()
    # avg_ biprediction variants — same shader as put_ + extra L2 with
    # existing dst.  All 15 useful positions.
    foreach(_mc mc20 mc02 mc22 mc10 mc30 mc01 mc03
                mc11 mc12 mc13 mc21 mc23 mc31 mc32 mc33)
        set(_spv ${CMAKE_BINARY_DIR}/v3d_h264_qpel_avg_${_mc}.spv)
        add_custom_command(
            OUTPUT ${_spv}
            COMMAND ${GLSLANG_VALIDATOR} -V --target-env vulkan1.3
                    -o ${_spv}
                    ${CMAKE_SOURCE_DIR}/src/v3d_h264_qpel_avg_${_mc}.comp
            DEPENDS ${CMAKE_SOURCE_DIR}/src/v3d_h264_qpel_avg_${_mc}.comp
            COMMENT "glslang: v3d_h264_qpel_avg_${_mc}.comp -> .spv"
            VERBATIM
        )
        set(H264_QPEL_avg_${_mc}_SPV ${_spv})
    endforeach()
    add_custom_target(daedalus_shaders ALL DEPENDS ${NOOP_SPV} ${IDCT8_SPV} ${LPF_SPV} ${MC_SPV} ${LPF8_SPV} ${CDEF_SPV} ${H264DEBLOCK_SPV} ${H264DEBLOCK_H_SPV} ${H264DEBLOCK_CHROMA_V_SPV} ${H264DEBLOCK_CHROMA_H_SPV} ${H264DEBLOCK_luma_v_intra_SPV} ${H264DEBLOCK_luma_h_intra_SPV} ${H264DEBLOCK_chroma_v_intra_SPV} ${H264DEBLOCK_chroma_h_intra_SPV} ${H264_IDCT4_SPV} ${H264_IDCT8_SPV} ${H264_QPEL_MC20_SPV} ${H264_QPEL_MC02_SPV} ${H264_QPEL_MC22_SPV} ${H264_QPEL_mc10_SPV} ${H264_QPEL_mc30_SPV} ${H264_QPEL_mc01_SPV} ${H264_QPEL_mc03_SPV} ${H264_QPEL_mc11_SPV} ${H264_QPEL_mc12_SPV} ${H264_QPEL_mc13_SPV} ${H264_QPEL_mc21_SPV} ${H264_QPEL_mc23_SPV} ${H264_QPEL_mc31_SPV} ${H264_QPEL_mc32_SPV} ${H264_QPEL_mc33_SPV} ${H264_QPEL_avg_mc20_SPV} ${H264_QPEL_avg_mc02_SPV} ${H264_QPEL_avg_mc22_SPV} ${H264_QPEL_avg_mc10_SPV} ${H264_QPEL_avg_mc30_SPV} ${H264_QPEL_avg_mc01_SPV} ${H264_QPEL_avg_mc03_SPV} ${H264_QPEL_avg_mc11_SPV} ${H264_QPEL_avg_mc12_SPV} ${H264_QPEL_avg_mc13_SPV} ${H264_QPEL_avg_mc21_SPV} ${H264_QPEL_avg_mc23_SPV} ${H264_QPEL_avg_mc31_SPV} ${H264_QPEL_avg_mc32_SPV} ${H264_QPEL_avg_mc33_SPV})
    # v3d_runner — reusable Vulkan plumbing.
    add_library(v3d_runner STATIC src/v3d_runner.c)
@@ -508,11 +558,42 @@ if (DAEDALUS_BUILD_VULKAN)
        ${H264DEBLOCK_H_SPV}
        ${H264DEBLOCK_CHROMA_V_SPV}
        ${H264DEBLOCK_CHROMA_H_SPV}
        ${H264DEBLOCK_luma_v_intra_SPV}
        ${H264DEBLOCK_luma_h_intra_SPV}
        ${H264DEBLOCK_chroma_v_intra_SPV}
        ${H264DEBLOCK_chroma_h_intra_SPV}
        ${H264_IDCT4_SPV}
        ${H264_IDCT8_SPV}
        ${H264_QPEL_MC20_SPV}
        ${H264_QPEL_MC02_SPV}
        ${H264_QPEL_MC22_SPV}
        ${H264_QPEL_mc10_SPV}
        ${H264_QPEL_mc30_SPV}
        ${H264_QPEL_mc01_SPV}
        ${H264_QPEL_mc03_SPV}
        ${H264_QPEL_mc11_SPV}
        ${H264_QPEL_mc12_SPV}
        ${H264_QPEL_mc13_SPV}
        ${H264_QPEL_mc21_SPV}
        ${H264_QPEL_mc23_SPV}
        ${H264_QPEL_mc31_SPV}
        ${H264_QPEL_mc32_SPV}
        ${H264_QPEL_mc33_SPV}
        ${H264_QPEL_avg_mc20_SPV}
        ${H264_QPEL_avg_mc02_SPV}
        ${H264_QPEL_avg_mc22_SPV}
        ${H264_QPEL_avg_mc10_SPV}
        ${H264_QPEL_avg_mc30_SPV}
        ${H264_QPEL_avg_mc01_SPV}
        ${H264_QPEL_avg_mc03_SPV}
        ${H264_QPEL_avg_mc11_SPV}
        ${H264_QPEL_avg_mc12_SPV}
        ${H264_QPEL_avg_mc13_SPV}
        ${H264_QPEL_avg_mc21_SPV}
        ${H264_QPEL_avg_mc23_SPV}
        ${H264_QPEL_avg_mc31_SPV}
        ${H264_QPEL_avg_mc32_SPV}
        ${H264_QPEL_avg_mc33_SPV}
        DESTINATION ${CMAKE_INSTALL_DATADIR}/daedalus-fourier/shaders
    )
 endif()
@@ -46,6 +46,11 @@ struct daedalus_ctx {
    v3d_pipeline  h264deblock_chroma_v_pipe;
    int           h264deblock_chroma_h_pipe_ready;
    v3d_pipeline  h264deblock_chroma_h_pipe;
    /* bS=4 intra deblock pipelines (strong/weak filter selector). */
    int           h264deblock_luma_v_intra_pipe_ready;   v3d_pipeline h264deblock_luma_v_intra_pipe;
    int           h264deblock_luma_h_intra_pipe_ready;   v3d_pipeline h264deblock_luma_h_intra_pipe;
    int           h264deblock_chroma_v_intra_pipe_ready; v3d_pipeline h264deblock_chroma_v_intra_pipe;
    int           h264deblock_chroma_h_intra_pipe_ready; v3d_pipeline h264deblock_chroma_h_intra_pipe;
    int           h264_idct4_pipe_ready;
    v3d_pipeline  h264_idct4_pipe;
    int           h264_idct8_pipe_ready;
@@ -56,6 +61,38 @@ struct daedalus_ctx {
    v3d_pipeline  h264_qpel_mc02_pipe;
    int           h264_qpel_mc22_pipe_ready;
    v3d_pipeline  h264_qpel_mc22_pipe;
    int           h264_qpel_mc10_pipe_ready;
    v3d_pipeline  h264_qpel_mc10_pipe;
    int           h264_qpel_mc30_pipe_ready;
    v3d_pipeline  h264_qpel_mc30_pipe;
    int           h264_qpel_mc01_pipe_ready;
    v3d_pipeline  h264_qpel_mc01_pipe;
    int           h264_qpel_mc03_pipe_ready;
    v3d_pipeline  h264_qpel_mc03_pipe;
    int           h264_qpel_mc11_pipe_ready; v3d_pipeline h264_qpel_mc11_pipe;
    int           h264_qpel_mc12_pipe_ready; v3d_pipeline h264_qpel_mc12_pipe;
    int           h264_qpel_mc13_pipe_ready; v3d_pipeline h264_qpel_mc13_pipe;
    int           h264_qpel_mc21_pipe_ready; v3d_pipeline h264_qpel_mc21_pipe;
    int           h264_qpel_mc23_pipe_ready; v3d_pipeline h264_qpel_mc23_pipe;
    int           h264_qpel_mc31_pipe_ready; v3d_pipeline h264_qpel_mc31_pipe;
    int           h264_qpel_mc32_pipe_ready; v3d_pipeline h264_qpel_mc32_pipe;
    int           h264_qpel_mc33_pipe_ready; v3d_pipeline h264_qpel_mc33_pipe;
    /* avg_ biprediction pipelines — same shaders + L2 with existing dst. */
    int           h264_qpel_avg_mc20_pipe_ready; v3d_pipeline h264_qpel_avg_mc20_pipe;
    int           h264_qpel_avg_mc02_pipe_ready; v3d_pipeline h264_qpel_avg_mc02_pipe;
    int           h264_qpel_avg_mc22_pipe_ready; v3d_pipeline h264_qpel_avg_mc22_pipe;
    int           h264_qpel_avg_mc10_pipe_ready; v3d_pipeline h264_qpel_avg_mc10_pipe;
    int           h264_qpel_avg_mc30_pipe_ready; v3d_pipeline h264_qpel_avg_mc30_pipe;
    int           h264_qpel_avg_mc01_pipe_ready; v3d_pipeline h264_qpel_avg_mc01_pipe;
    int           h264_qpel_avg_mc03_pipe_ready; v3d_pipeline h264_qpel_avg_mc03_pipe;
    int           h264_qpel_avg_mc11_pipe_ready; v3d_pipeline h264_qpel_avg_mc11_pipe;
    int           h264_qpel_avg_mc12_pipe_ready; v3d_pipeline h264_qpel_avg_mc12_pipe;
    int           h264_qpel_avg_mc13_pipe_ready; v3d_pipeline h264_qpel_avg_mc13_pipe;
    int           h264_qpel_avg_mc21_pipe_ready; v3d_pipeline h264_qpel_avg_mc21_pipe;
    int           h264_qpel_avg_mc23_pipe_ready; v3d_pipeline h264_qpel_avg_mc23_pipe;
    int           h264_qpel_avg_mc31_pipe_ready; v3d_pipeline h264_qpel_avg_mc31_pipe;
    int           h264_qpel_avg_mc32_pipe_ready; v3d_pipeline h264_qpel_avg_mc32_pipe;
    int           h264_qpel_avg_mc33_pipe_ready; v3d_pipeline h264_qpel_avg_mc33_pipe;
 };
 daedalus_ctx *daedalus_ctx_create(void)
@@ -113,11 +150,42 @@ void daedalus_ctx_destroy(daedalus_ctx *ctx)
        if (ctx->h264deblock_h_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264deblock_h_pipe);
        if (ctx->h264deblock_chroma_v_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264deblock_chroma_v_pipe);
        if (ctx->h264deblock_chroma_h_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264deblock_chroma_h_pipe);
        if (ctx->h264deblock_luma_v_intra_pipe_ready)   v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264deblock_luma_v_intra_pipe);
        if (ctx->h264deblock_luma_h_intra_pipe_ready)   v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264deblock_luma_h_intra_pipe);
        if (ctx->h264deblock_chroma_v_intra_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264deblock_chroma_v_intra_pipe);
        if (ctx->h264deblock_chroma_h_intra_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264deblock_chroma_h_intra_pipe);
        if (ctx->h264_idct4_pipe_ready)  v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_idct4_pipe);
        if (ctx->h264_idct8_pipe_ready)  v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_idct8_pipe);
        if (ctx->h264_qpel_mc20_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc20_pipe);
        if (ctx->h264_qpel_mc02_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc02_pipe);
        if (ctx->h264_qpel_mc22_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc22_pipe);
        if (ctx->h264_qpel_mc10_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc10_pipe);
        if (ctx->h264_qpel_mc30_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc30_pipe);
        if (ctx->h264_qpel_mc01_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc01_pipe);
        if (ctx->h264_qpel_mc03_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc03_pipe);
        if (ctx->h264_qpel_mc11_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc11_pipe);
        if (ctx->h264_qpel_mc12_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc12_pipe);
        if (ctx->h264_qpel_mc13_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc13_pipe);
        if (ctx->h264_qpel_mc21_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc21_pipe);
        if (ctx->h264_qpel_mc23_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc23_pipe);
        if (ctx->h264_qpel_mc31_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc31_pipe);
        if (ctx->h264_qpel_mc32_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc32_pipe);
        if (ctx->h264_qpel_mc33_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_mc33_pipe);
        if (ctx->h264_qpel_avg_mc20_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc20_pipe);
        if (ctx->h264_qpel_avg_mc02_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc02_pipe);
        if (ctx->h264_qpel_avg_mc22_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc22_pipe);
        if (ctx->h264_qpel_avg_mc10_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc10_pipe);
        if (ctx->h264_qpel_avg_mc30_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc30_pipe);
        if (ctx->h264_qpel_avg_mc01_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc01_pipe);
        if (ctx->h264_qpel_avg_mc03_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc03_pipe);
        if (ctx->h264_qpel_avg_mc11_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc11_pipe);
        if (ctx->h264_qpel_avg_mc12_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc12_pipe);
        if (ctx->h264_qpel_avg_mc13_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc13_pipe);
        if (ctx->h264_qpel_avg_mc21_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc21_pipe);
        if (ctx->h264_qpel_avg_mc23_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc23_pipe);
        if (ctx->h264_qpel_avg_mc31_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc31_pipe);
        if (ctx->h264_qpel_avg_mc32_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc32_pipe);
        if (ctx->h264_qpel_avg_mc33_pipe_ready) v3d_runner_destroy_pipeline(ctx->runner, &ctx->h264_qpel_avg_mc33_pipe);
        v3d_runner_destroy(ctx->runner);
    }
    free(ctx);
@@ -148,40 +216,40 @@ daedalus_substrate daedalus_recipe_substrate_for(daedalus_kernel k)
    case DAEDALUS_KERNEL_H264_DEBLOCK_LH:  return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264deblock_h.spv */
    case DAEDALUS_KERNEL_H264_DEBLOCK_CV:  return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264deblock_chroma_v.spv */
    case DAEDALUS_KERNEL_H264_DEBLOCK_CH:  return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264deblock_chroma_h.spv */
-    case DAEDALUS_KERNEL_H264_DEBLOCK_LV_INTRA: return DAEDALUS_SUBSTRATE_CPU; /* bS=4 luma QPU pending */
+    case DAEDALUS_KERNEL_H264_DEBLOCK_LV_INTRA: return DAEDALUS_SUBSTRATE_QPU; /* v3d_h264deblock_luma_v_intra.spv */
-    case DAEDALUS_KERNEL_H264_DEBLOCK_LH_INTRA: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_DEBLOCK_LH_INTRA: return DAEDALUS_SUBSTRATE_QPU;
-    case DAEDALUS_KERNEL_H264_DEBLOCK_CV_INTRA: return DAEDALUS_SUBSTRATE_CPU; /* bS=4 chroma QPU pending */
+    case DAEDALUS_KERNEL_H264_DEBLOCK_CV_INTRA: return DAEDALUS_SUBSTRATE_QPU; /* v3d_h264deblock_chroma_v_intra.spv */
-    case DAEDALUS_KERNEL_H264_DEBLOCK_CH_INTRA: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_DEBLOCK_CH_INTRA: return DAEDALUS_SUBSTRATE_QPU;
    case DAEDALUS_KERNEL_H264_QPEL_MC20:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc20.spv */
    case DAEDALUS_KERNEL_H264_QPEL_MC02:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc02.spv */
    case DAEDALUS_KERNEL_H264_QPEL_MC22:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc22.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC10:   return DAEDALUS_SUBSTRATE_CPU;	/* ¼-H L2 */
+    case DAEDALUS_KERNEL_H264_QPEL_MC10:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc10.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC30:   return DAEDALUS_SUBSTRATE_CPU;	/* ¾-H L2 */
+    case DAEDALUS_KERNEL_H264_QPEL_MC30:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc30.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC01:   return DAEDALUS_SUBSTRATE_CPU;	/* ¼-V L2 */
+    case DAEDALUS_KERNEL_H264_QPEL_MC01:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc01.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC03:   return DAEDALUS_SUBSTRATE_CPU;	/* ¾-V L2 */
+    case DAEDALUS_KERNEL_H264_QPEL_MC03:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc03.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC11:   return DAEDALUS_SUBSTRATE_CPU;	/* diagonal ¼¼ */
+    case DAEDALUS_KERNEL_H264_QPEL_MC11:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc11.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC12:   return DAEDALUS_SUBSTRATE_CPU;	/* diagonal ¼½ */
+    case DAEDALUS_KERNEL_H264_QPEL_MC12:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc12.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC13:   return DAEDALUS_SUBSTRATE_CPU;	/* diagonal ¼¾ */
+    case DAEDALUS_KERNEL_H264_QPEL_MC13:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc13.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC21:   return DAEDALUS_SUBSTRATE_CPU;	/* diagonal ½¼ */
+    case DAEDALUS_KERNEL_H264_QPEL_MC21:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc21.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC23:   return DAEDALUS_SUBSTRATE_CPU;	/* diagonal ½¾ */
+    case DAEDALUS_KERNEL_H264_QPEL_MC23:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc23.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC31:   return DAEDALUS_SUBSTRATE_CPU;	/* diagonal ¾¼ */
+    case DAEDALUS_KERNEL_H264_QPEL_MC31:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc31.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC32:   return DAEDALUS_SUBSTRATE_CPU;	/* diagonal ¾½ */
+    case DAEDALUS_KERNEL_H264_QPEL_MC32:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc32.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_MC33:   return DAEDALUS_SUBSTRATE_CPU;	/* diagonal ¾¾ */
+    case DAEDALUS_KERNEL_H264_QPEL_MC33:   return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_mc33.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC20: return DAEDALUS_SUBSTRATE_CPU;	/* biprediction anchors */
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC20: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc20.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC02: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC02: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc02.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC22: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC22: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc22.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC10: return DAEDALUS_SUBSTRATE_CPU;	/* ¼-H L2 avg */
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC10: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc10.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC30: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC30: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc30.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC01: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC01: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc01.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC03: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC03: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc03.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC11: return DAEDALUS_SUBSTRATE_CPU;	/* diagonals avg */
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC11: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc11.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC12: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC12: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc12.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC13: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC13: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc13.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC21: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC21: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc21.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC23: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC23: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc23.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC31: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC31: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc31.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC32: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC32: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc32.spv */
-    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC33: return DAEDALUS_SUBSTRATE_CPU;
+    case DAEDALUS_KERNEL_H264_QPEL_AVG_MC33: return DAEDALUS_SUBSTRATE_QPU;	/* v3d_h264_qpel_avg_mc33.spv */
    }
    return DAEDALUS_SUBSTRATE_CPU;
 }
@@ -1181,6 +1249,107 @@ static int dispatch_h264_deblock_chroma_h_qpu(daedalus_ctx *ctx,
        "v3d_h264deblock_chroma_h.spv", dst, dst_stride, n_edges, meta, 1);
 }
 /* -------------------- H.264 luma/chroma intra (bS=4) QPU dispatches.
 * Same WG geometry as the non-intra shaders, same meta layout (tc0[]
 * unused — the strong/weak selector replaces it).  Bounds match the
 * non-intra variants exactly:
 *   luma_v:  dst_off + 3*stride + 16    (reads p3 at -4*stride, writes p2 at -3*stride)
 *   luma_h:  dst_off + 15*stride + 4    (lane → row, reads pix[-4..+3])
 *   chroma_v: 1*stride + 8              (only p1..q1 = ±2*stride)
 *   chroma_h: 7*stride + 2              (lane → row, reads pix[-2..+1])
 */
 static int dispatch_h264_deblock_luma_intra_qpu(daedalus_ctx *ctx,
    v3d_pipeline *pipe, int *pipe_ready, const char *spv,
    uint8_t *dst, size_t dst_stride, size_t n_edges,
    const daedalus_h264_deblock_meta *meta, int orient_h)
 {
    if (!*pipe_ready) {
        if (v3d_runner_create_pipeline(ctx->runner, spv,
                                       2, sizeof(h264deblock_pc), pipe) != 0)
            return -1;
        *pipe_ready = 1;
    }
    size_t meta_bytes = n_edges * 4 * sizeof(uint32_t);
    size_t dst_max = 0;
    for (size_t i = 0; i < n_edges; i++) {
        size_t e = orient_h ? meta[i].dst_off + 15 * dst_stride + 4
                            : meta[i].dst_off +  3 * dst_stride + 16;
        if (e > dst_max) dst_max = e;
    }
    v3d_buffer bm = {0}, bd = {0};
    if (v3d_runner_acquire_buffer(ctx->runner, meta_bytes, &bm)) return -1;
    if (v3d_runner_acquire_buffer(ctx->runner, dst_max,    &bd)) { v3d_runner_release_buffer(ctx->runner, &bm); return -1; }
    memcpy(bd.mapped, dst, dst_max);
    uint32_t *m = bm.mapped;
    for (size_t i = 0; i < n_edges; i++) {
        m[4*i+0] = meta[i].dst_off;
        m[4*i+1] = ((uint32_t) meta[i].alpha) | (((uint32_t) meta[i].beta) << 8);
        m[4*i+2] = 0;   /* tc0 unused for intra */
        m[4*i+3] = 0;
    }
    v3d_buffer binds[2] = { bm, bd };
    if (v3d_runner_bind_buffers(ctx->runner, pipe, binds, 2)) goto fail;
    uint32_t wg_count = (uint32_t)((n_edges + 15) / 16);
    h264deblock_pc pc = { .n_edges = (uint32_t) n_edges,
                          .dst_stride_u8 = (uint32_t) dst_stride };
    if (v3d_runner_pipeline_cmdbuf_reset(ctx->runner, pipe)) goto fail;
    VkCommandBuffer cb = pipe->cb;
    VkCommandBufferBeginInfo cbbi = { .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO };
    vkBeginCommandBuffer(cb, &cbbi);
    vkCmdBindPipeline(cb, VK_PIPELINE_BIND_POINT_COMPUTE, pipe->pipeline);
    vkCmdBindDescriptorSets(cb, VK_PIPELINE_BIND_POINT_COMPUTE,
                            pipe->layout, 0, 1, &pipe->desc_set, 0, NULL);
    vkCmdPushConstants(cb, pipe->layout, VK_SHADER_STAGE_COMPUTE_BIT,
                       0, sizeof(pc), &pc);
    vkCmdDispatch(cb, wg_count, 1, 1);
    vkEndCommandBuffer(cb);
    if (v3d_runner_submit_wait(ctx->runner, cb)) goto fail;
    memcpy(dst, bd.mapped, dst_max);
    v3d_runner_release_buffer(ctx->runner, &bd);
    v3d_runner_release_buffer(ctx->runner, &bm);
    return 0;
 fail:
    v3d_runner_release_buffer(ctx->runner, &bd);
    v3d_runner_release_buffer(ctx->runner, &bm);
    return -1;
 }
 static int dispatch_h264_deblock_luma_v_intra_qpu(daedalus_ctx *ctx,
    uint8_t *dst, size_t dst_stride,
    size_t n_edges, const daedalus_h264_deblock_meta *meta)
 {
    return dispatch_h264_deblock_luma_intra_qpu(ctx,
        &ctx->h264deblock_luma_v_intra_pipe, &ctx->h264deblock_luma_v_intra_pipe_ready,
        "v3d_h264deblock_luma_v_intra.spv", dst, dst_stride, n_edges, meta, 0);
 }
 static int dispatch_h264_deblock_luma_h_intra_qpu(daedalus_ctx *ctx,
    uint8_t *dst, size_t dst_stride,
    size_t n_edges, const daedalus_h264_deblock_meta *meta)
 {
    return dispatch_h264_deblock_luma_intra_qpu(ctx,
        &ctx->h264deblock_luma_h_intra_pipe, &ctx->h264deblock_luma_h_intra_pipe_ready,
        "v3d_h264deblock_luma_h_intra.spv", dst, dst_stride, n_edges, meta, 1);
 }
 static int dispatch_h264_deblock_chroma_v_intra_qpu(daedalus_ctx *ctx,
    uint8_t *dst, size_t dst_stride,
    size_t n_edges, const daedalus_h264_deblock_meta *meta)
 {
    return dispatch_h264_deblock_chroma_qpu(ctx,
        &ctx->h264deblock_chroma_v_intra_pipe, &ctx->h264deblock_chroma_v_intra_pipe_ready,
        "v3d_h264deblock_chroma_v_intra.spv", dst, dst_stride, n_edges, meta, 0);
 }
 static int dispatch_h264_deblock_chroma_h_intra_qpu(daedalus_ctx *ctx,
    uint8_t *dst, size_t dst_stride,
    size_t n_edges, const daedalus_h264_deblock_meta *meta)
 {
    return dispatch_h264_deblock_chroma_qpu(ctx,
        &ctx->h264deblock_chroma_h_intra_pipe, &ctx->h264deblock_chroma_h_intra_pipe_ready,
        "v3d_h264deblock_chroma_h_intra.spv", dst, dst_stride, n_edges, meta, 1);
 }
 /* -------------------- H.264 IDCT 4x4 QPU dispatch (cycle 6) ----- */
 typedef struct {
@@ -1630,6 +1799,207 @@ fail:
    return -1;
 }
 /* Generic QPU dispatch for the 4 single-axis quarter-pel shaders
 * (mc10/30 horizontal, mc01/03 vertical).  All 4 share the same WG
 * geometry (64 lanes/block, 1 block/WG), push-constant layout, and
 * 3-binding interface (src/dst/meta) as mc20/mc02.  Only src_max
 * differs by axis:
 *   H variants: src_max = src_off + 7*stride + 11   (cols -2..+10)
 *   V variants: src_max = src_off + 10*stride + 8   (rows -2..+10)
 */
 static int dispatch_h264_qpel_axis_qpu(daedalus_ctx *ctx,
    v3d_pipeline *pipe, int *pipe_ready, const char *spv,
    uint8_t *dst, const uint8_t *src, size_t stride,
    size_t n_blocks, const daedalus_h264_qpel_meta *meta,
    int axis_v)
 {
    if (!*pipe_ready) {
        if (v3d_runner_create_pipeline(ctx->runner, spv,
                                       3, sizeof(h264_qpel_mc20_pc), pipe) != 0)
            return -1;
        *pipe_ready = 1;
    }
    size_t meta_bytes = n_blocks * 4 * sizeof(uint32_t);
    size_t src_max = 0, dst_max = 0;
    for (size_t i = 0; i < n_blocks; i++) {
        size_t s_end = axis_v ? meta[i].src_off + (size_t) 10 * stride + 8
                              : meta[i].src_off + (size_t)  7 * stride + 11;
        size_t d_end = meta[i].dst_off + (size_t) 7 * stride + 8;
        if (s_end > src_max) src_max = s_end;
        if (d_end > dst_max) dst_max = d_end;
    }
    v3d_buffer bs = {0}, bd = {0}, bm = {0};
    if (v3d_runner_create_buffer(ctx->runner, src_max,    &bs)) return -1;
    if (v3d_runner_create_buffer(ctx->runner, dst_max,    &bd)) {
        v3d_runner_destroy_buffer(ctx->runner, &bs); return -1;
    }
    if (v3d_runner_create_buffer(ctx->runner, meta_bytes, &bm)) {
        v3d_runner_destroy_buffer(ctx->runner, &bd);
        v3d_runner_destroy_buffer(ctx->runner, &bs); return -1;
    }
    memcpy(bs.mapped, src, src_max);
    memcpy(bd.mapped, dst, dst_max);
    uint32_t *m = bm.mapped;
    for (size_t i = 0; i < n_blocks; i++) {
        m[4*i+0] = meta[i].dst_off;
        m[4*i+1] = meta[i].src_off;
        m[4*i+2] = 0;
        m[4*i+3] = 0;
    }
    v3d_buffer binds[3] = { bs, bd, bm };
    if (v3d_runner_bind_buffers(ctx->runner, pipe, binds, 3)) goto fail;
    h264_qpel_mc20_pc pc = { .n_blocks = (uint32_t) n_blocks,
                              .stride_u8 = (uint32_t) stride };
    VkCommandBuffer cb = v3d_runner_alloc_cmdbuf(ctx->runner);
    if (cb == VK_NULL_HANDLE) goto fail;
    VkCommandBufferBeginInfo cbbi = { .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO };
    vkBeginCommandBuffer(cb, &cbbi);
    vkCmdBindPipeline(cb, VK_PIPELINE_BIND_POINT_COMPUTE, pipe->pipeline);
    vkCmdBindDescriptorSets(cb, VK_PIPELINE_BIND_POINT_COMPUTE,
                            pipe->layout, 0, 1, &pipe->desc_set, 0, NULL);
    vkCmdPushConstants(cb, pipe->layout, VK_SHADER_STAGE_COMPUTE_BIT,
                       0, sizeof(pc), &pc);
    vkCmdDispatch(cb, (uint32_t) n_blocks, 1, 1);
    vkEndCommandBuffer(cb);
    if (v3d_runner_submit_wait(ctx->runner, cb)) goto fail;
    memcpy(dst, bd.mapped, dst_max);
    v3d_runner_destroy_buffer(ctx->runner, &bm);
    v3d_runner_destroy_buffer(ctx->runner, &bd);
    v3d_runner_destroy_buffer(ctx->runner, &bs);
    return 0;
 fail:
    v3d_runner_destroy_buffer(ctx->runner, &bm);
    v3d_runner_destroy_buffer(ctx->runner, &bd);
    v3d_runner_destroy_buffer(ctx->runner, &bs);
    return -1;
 }
 #define DEFINE_QPEL_AXIS_QPU(name, spv, axis_v)                                  \
 static int dispatch_h264_qpel_ ## name ## _qpu(daedalus_ctx *ctx,                \
    uint8_t *dst, const uint8_t *src, size_t stride,                             \
    size_t n_blocks, const daedalus_h264_qpel_meta *meta)                        \
 {                                                                                \
    return dispatch_h264_qpel_axis_qpu(ctx, &ctx->h264_qpel_ ## name ## _pipe,   \
        &ctx->h264_qpel_ ## name ## _pipe_ready, spv, dst, src, stride,          \
        n_blocks, meta, axis_v);                                                 \
 }
 DEFINE_QPEL_AXIS_QPU(mc10, "v3d_h264_qpel_mc10.spv", 0)
 DEFINE_QPEL_AXIS_QPU(mc30, "v3d_h264_qpel_mc30.spv", 0)
 DEFINE_QPEL_AXIS_QPU(mc01, "v3d_h264_qpel_mc01.spv", 1)
 DEFINE_QPEL_AXIS_QPU(mc03, "v3d_h264_qpel_mc03.spv", 1)
 #undef DEFINE_QPEL_AXIS_QPU
 /* Diagonals share the mc22-style src envelope (rows -2..+10, cols
 * -2..+10) because they compose mc22 with mc20/mc02, sometimes
 * with (r+1, c) or (r, c+1) offsets. */
 static int dispatch_h264_qpel_diag_qpu(daedalus_ctx *ctx,
    v3d_pipeline *pipe, int *pipe_ready, const char *spv,
    uint8_t *dst, const uint8_t *src, size_t stride,
    size_t n_blocks, const daedalus_h264_qpel_meta *meta)
 {
    if (!*pipe_ready) {
        if (v3d_runner_create_pipeline(ctx->runner, spv,
                                       3, sizeof(h264_qpel_mc20_pc), pipe) != 0)
            return -1;
        *pipe_ready = 1;
    }
    size_t meta_bytes = n_blocks * 4 * sizeof(uint32_t);
    size_t src_max = 0, dst_max = 0;
    for (size_t i = 0; i < n_blocks; i++) {
        size_t s_end = meta[i].src_off + (size_t) 10 * stride + 11;
        size_t d_end = meta[i].dst_off + (size_t)  7 * stride + 8;
        if (s_end > src_max) src_max = s_end;
        if (d_end > dst_max) dst_max = d_end;
    }
    v3d_buffer bs = {0}, bd = {0}, bm = {0};
    if (v3d_runner_create_buffer(ctx->runner, src_max,    &bs)) return -1;
    if (v3d_runner_create_buffer(ctx->runner, dst_max,    &bd)) {
        v3d_runner_destroy_buffer(ctx->runner, &bs); return -1;
    }
    if (v3d_runner_create_buffer(ctx->runner, meta_bytes, &bm)) {
        v3d_runner_destroy_buffer(ctx->runner, &bd);
        v3d_runner_destroy_buffer(ctx->runner, &bs); return -1;
    }
    memcpy(bs.mapped, src, src_max);
    memcpy(bd.mapped, dst, dst_max);
    uint32_t *m = bm.mapped;
    for (size_t i = 0; i < n_blocks; i++) {
        m[4*i+0] = meta[i].dst_off;
        m[4*i+1] = meta[i].src_off;
        m[4*i+2] = 0;
        m[4*i+3] = 0;
    }
    v3d_buffer binds[3] = { bs, bd, bm };
    if (v3d_runner_bind_buffers(ctx->runner, pipe, binds, 3)) goto fail;
    h264_qpel_mc20_pc pc = { .n_blocks = (uint32_t) n_blocks,
                              .stride_u8 = (uint32_t) stride };
    VkCommandBuffer cb = v3d_runner_alloc_cmdbuf(ctx->runner);
    if (cb == VK_NULL_HANDLE) goto fail;
    VkCommandBufferBeginInfo cbbi = { .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO };
    vkBeginCommandBuffer(cb, &cbbi);
    vkCmdBindPipeline(cb, VK_PIPELINE_BIND_POINT_COMPUTE, pipe->pipeline);
    vkCmdBindDescriptorSets(cb, VK_PIPELINE_BIND_POINT_COMPUTE,
                            pipe->layout, 0, 1, &pipe->desc_set, 0, NULL);
    vkCmdPushConstants(cb, pipe->layout, VK_SHADER_STAGE_COMPUTE_BIT,
                       0, sizeof(pc), &pc);
    vkCmdDispatch(cb, (uint32_t) n_blocks, 1, 1);
    vkEndCommandBuffer(cb);
    if (v3d_runner_submit_wait(ctx->runner, cb)) goto fail;
    memcpy(dst, bd.mapped, dst_max);
    v3d_runner_destroy_buffer(ctx->runner, &bm);
    v3d_runner_destroy_buffer(ctx->runner, &bd);
    v3d_runner_destroy_buffer(ctx->runner, &bs);
    return 0;
 fail:
    v3d_runner_destroy_buffer(ctx->runner, &bm);
    v3d_runner_destroy_buffer(ctx->runner, &bd);
    v3d_runner_destroy_buffer(ctx->runner, &bs);
    return -1;
 }
 #define DEFINE_QPEL_DIAG_QPU(name)                                              \
 static int dispatch_h264_qpel_ ## name ## _qpu(daedalus_ctx *ctx,               \
    uint8_t *dst, const uint8_t *src, size_t stride,                            \
    size_t n_blocks, const daedalus_h264_qpel_meta *meta)                       \
 {                                                                               \
    return dispatch_h264_qpel_diag_qpu(ctx, &ctx->h264_qpel_ ## name ## _pipe,  \
        &ctx->h264_qpel_ ## name ## _pipe_ready,                                \
        "v3d_h264_qpel_" #name ".spv", dst, src, stride, n_blocks, meta);       \
 }
 DEFINE_QPEL_DIAG_QPU(mc11)
 DEFINE_QPEL_DIAG_QPU(mc12)
 DEFINE_QPEL_DIAG_QPU(mc13)
 DEFINE_QPEL_DIAG_QPU(mc21)
 DEFINE_QPEL_DIAG_QPU(mc23)
 DEFINE_QPEL_DIAG_QPU(mc31)
 DEFINE_QPEL_DIAG_QPU(mc32)
 DEFINE_QPEL_DIAG_QPU(mc33)
 /* avg_ variants — same diag-style envelope (10*stride+11 covers any
 * (r±1, c±1) offset the avg_ shaders use), different SPV file.
 * Slightly over-allocates for avg_mc20/02/10/30/01/03 (which need
 * less src context) but the cost is negligible. */
 DEFINE_QPEL_DIAG_QPU(avg_mc20)
 DEFINE_QPEL_DIAG_QPU(avg_mc02)
 DEFINE_QPEL_DIAG_QPU(avg_mc22)
 DEFINE_QPEL_DIAG_QPU(avg_mc10)
 DEFINE_QPEL_DIAG_QPU(avg_mc30)
 DEFINE_QPEL_DIAG_QPU(avg_mc01)
 DEFINE_QPEL_DIAG_QPU(avg_mc03)
 DEFINE_QPEL_DIAG_QPU(avg_mc11)
 DEFINE_QPEL_DIAG_QPU(avg_mc12)
 DEFINE_QPEL_DIAG_QPU(avg_mc13)
 DEFINE_QPEL_DIAG_QPU(avg_mc21)
 DEFINE_QPEL_DIAG_QPU(avg_mc23)
 DEFINE_QPEL_DIAG_QPU(avg_mc31)
 DEFINE_QPEL_DIAG_QPU(avg_mc32)
 DEFINE_QPEL_DIAG_QPU(avg_mc33)
 #undef DEFINE_QPEL_DIAG_QPU
 /* -------------------- Public dispatch entry points -------------- */
 #define ROUTE_CPU_ONLY(_kernel, _cpu_fn, ...)                                 \
@@ -1803,7 +2173,7 @@ int daedalus_dispatch_h264_deblock_chroma_h(daedalus_ctx *ctx, daedalus_substrat
    return dispatch_h264_deblock_chroma_h_qpu(ctx, dst, dst_stride, n_edges, meta);
 }
-#define DEFINE_INTRA_DISPATCH(name, kernel, cpu_fn)                            \
+#define DEFINE_INTRA_DISPATCH(name, kernel, cpu_fn, qpu_fn)                    \
 int daedalus_dispatch_h264_deblock_ ## name (daedalus_ctx *ctx,                \
    daedalus_substrate sub, uint8_t *dst, size_t dst_stride,                   \
    size_t n_edges, const daedalus_h264_deblock_meta *meta)                    \
@@ -1813,14 +2183,23 @@ int daedalus_dispatch_h264_deblock_ ## name (daedalus_ctx *ctx,                \
        eff = daedalus_recipe_substrate_for(kernel);                           \
    if (eff == DAEDALUS_SUBSTRATE_QPU && !daedalus_ctx_has_qpu(ctx))           \
        eff = DAEDALUS_SUBSTRATE_CPU;                                          \
-    if (eff == DAEDALUS_SUBSTRATE_QPU) return -1;                              \
+    if (eff == DAEDALUS_SUBSTRATE_CPU)                                         \
        return cpu_fn(ctx, dst, dst_stride, n_edges, meta);                    \
    return qpu_fn(ctx, dst, dst_stride, n_edges, meta);                        \
 }
-DEFINE_INTRA_DISPATCH(luma_v_intra,   DAEDALUS_KERNEL_H264_DEBLOCK_LV_INTRA, dispatch_h264_deblock_luma_v_intra_cpu)
+DEFINE_INTRA_DISPATCH(luma_v_intra,   DAEDALUS_KERNEL_H264_DEBLOCK_LV_INTRA,
-DEFINE_INTRA_DISPATCH(luma_h_intra,   DAEDALUS_KERNEL_H264_DEBLOCK_LH_INTRA, dispatch_h264_deblock_luma_h_intra_cpu)
+                      dispatch_h264_deblock_luma_v_intra_cpu,
-DEFINE_INTRA_DISPATCH(chroma_v_intra, DAEDALUS_KERNEL_H264_DEBLOCK_CV_INTRA, dispatch_h264_deblock_chroma_v_intra_cpu)
+                      dispatch_h264_deblock_luma_v_intra_qpu)
-DEFINE_INTRA_DISPATCH(chroma_h_intra, DAEDALUS_KERNEL_H264_DEBLOCK_CH_INTRA, dispatch_h264_deblock_chroma_h_intra_cpu)
+DEFINE_INTRA_DISPATCH(luma_h_intra,   DAEDALUS_KERNEL_H264_DEBLOCK_LH_INTRA,
                      dispatch_h264_deblock_luma_h_intra_cpu,
                      dispatch_h264_deblock_luma_h_intra_qpu)
 DEFINE_INTRA_DISPATCH(chroma_v_intra, DAEDALUS_KERNEL_H264_DEBLOCK_CV_INTRA,
                      dispatch_h264_deblock_chroma_v_intra_cpu,
                      dispatch_h264_deblock_chroma_v_intra_qpu)
 DEFINE_INTRA_DISPATCH(chroma_h_intra, DAEDALUS_KERNEL_H264_DEBLOCK_CH_INTRA,
                      dispatch_h264_deblock_chroma_h_intra_cpu,
                      dispatch_h264_deblock_chroma_h_intra_qpu)
 #undef DEFINE_INTRA_DISPATCH
@@ -1883,33 +2262,77 @@ int daedalus_dispatch_h264_qpel_ ## suffix(daedalus_ctx *ctx,                  \
                                                  n_blocks, meta);             \
 }
-DEFINE_QPEL_DISPATCH(mc10, DAEDALUS_KERNEL_H264_QPEL_MC10)
+/* mc10/30/01/03 now have QPU shaders — explicit definitions instead of
-DEFINE_QPEL_DISPATCH(mc30, DAEDALUS_KERNEL_H264_QPEL_MC30)
+ * the no-QPU DEFINE_QPEL_DISPATCH macro.  Same routing shape as mc20/02. */
-DEFINE_QPEL_DISPATCH(mc01, DAEDALUS_KERNEL_H264_QPEL_MC01)
+#define DEFINE_QPEL_DISPATCH_QPU(suffix, kernel)                               \
-DEFINE_QPEL_DISPATCH(mc03, DAEDALUS_KERNEL_H264_QPEL_MC03)
+int daedalus_dispatch_h264_qpel_ ## suffix(daedalus_ctx *ctx,                  \
-DEFINE_QPEL_DISPATCH(mc11, DAEDALUS_KERNEL_H264_QPEL_MC11)
+    daedalus_substrate sub, uint8_t *dst, const uint8_t *src, size_t stride,   \
-DEFINE_QPEL_DISPATCH(mc12, DAEDALUS_KERNEL_H264_QPEL_MC12)
+    size_t n_blocks, const daedalus_h264_qpel_meta *meta)                      \
-DEFINE_QPEL_DISPATCH(mc13, DAEDALUS_KERNEL_H264_QPEL_MC13)
+{                                                                              \
-DEFINE_QPEL_DISPATCH(mc21, DAEDALUS_KERNEL_H264_QPEL_MC21)
+    daedalus_substrate eff = sub;                                              \
-DEFINE_QPEL_DISPATCH(mc23, DAEDALUS_KERNEL_H264_QPEL_MC23)
+    if (eff == DAEDALUS_SUBSTRATE_AUTO)                                        \
-DEFINE_QPEL_DISPATCH(mc31, DAEDALUS_KERNEL_H264_QPEL_MC31)
+        eff = daedalus_recipe_substrate_for(kernel);                           \
-DEFINE_QPEL_DISPATCH(mc32, DAEDALUS_KERNEL_H264_QPEL_MC32)
+    if (eff == DAEDALUS_SUBSTRATE_QPU && !daedalus_ctx_has_qpu(ctx))           \
-DEFINE_QPEL_DISPATCH(mc33, DAEDALUS_KERNEL_H264_QPEL_MC33)
+        eff = DAEDALUS_SUBSTRATE_CPU;                                          \
-DEFINE_QPEL_DISPATCH(avg_mc20, DAEDALUS_KERNEL_H264_QPEL_AVG_MC20)
+    if (eff == DAEDALUS_SUBSTRATE_CPU)                                         \
-DEFINE_QPEL_DISPATCH(avg_mc02, DAEDALUS_KERNEL_H264_QPEL_AVG_MC02)
+        return dispatch_h264_qpel_ ## suffix ## _cpu(ctx, dst, src, stride,    \
-DEFINE_QPEL_DISPATCH(avg_mc22, DAEDALUS_KERNEL_H264_QPEL_AVG_MC22)
+                                                       n_blocks, meta);        \
-DEFINE_QPEL_DISPATCH(avg_mc10, DAEDALUS_KERNEL_H264_QPEL_AVG_MC10)
+    return dispatch_h264_qpel_ ## suffix ## _qpu(ctx, dst, src, stride,        \
-DEFINE_QPEL_DISPATCH(avg_mc30, DAEDALUS_KERNEL_H264_QPEL_AVG_MC30)
+                                                   n_blocks, meta);            \
-DEFINE_QPEL_DISPATCH(avg_mc01, DAEDALUS_KERNEL_H264_QPEL_AVG_MC01)
+}
-DEFINE_QPEL_DISPATCH(avg_mc03, DAEDALUS_KERNEL_H264_QPEL_AVG_MC03)
+
-DEFINE_QPEL_DISPATCH(avg_mc11, DAEDALUS_KERNEL_H264_QPEL_AVG_MC11)
+DEFINE_QPEL_DISPATCH_QPU(mc10, DAEDALUS_KERNEL_H264_QPEL_MC10)
-DEFINE_QPEL_DISPATCH(avg_mc12, DAEDALUS_KERNEL_H264_QPEL_AVG_MC12)
+DEFINE_QPEL_DISPATCH_QPU(mc30, DAEDALUS_KERNEL_H264_QPEL_MC30)
-DEFINE_QPEL_DISPATCH(avg_mc13, DAEDALUS_KERNEL_H264_QPEL_AVG_MC13)
+DEFINE_QPEL_DISPATCH_QPU(mc01, DAEDALUS_KERNEL_H264_QPEL_MC01)
-DEFINE_QPEL_DISPATCH(avg_mc21, DAEDALUS_KERNEL_H264_QPEL_AVG_MC21)
+DEFINE_QPEL_DISPATCH_QPU(mc03, DAEDALUS_KERNEL_H264_QPEL_MC03)
-DEFINE_QPEL_DISPATCH(avg_mc23, DAEDALUS_KERNEL_H264_QPEL_AVG_MC23)
+#undef DEFINE_QPEL_DISPATCH_QPU
-DEFINE_QPEL_DISPATCH(avg_mc31, DAEDALUS_KERNEL_H264_QPEL_AVG_MC31)
+/* mc11..mc33 diagonals — QPU-capable, same macro shape as mc10/30/01/03. */
-DEFINE_QPEL_DISPATCH(avg_mc32, DAEDALUS_KERNEL_H264_QPEL_AVG_MC32)
+#define DEFINE_QPEL_DIAG_PUBLIC(suffix, kernel)                                \
-DEFINE_QPEL_DISPATCH(avg_mc33, DAEDALUS_KERNEL_H264_QPEL_AVG_MC33)
+int daedalus_dispatch_h264_qpel_ ## suffix(daedalus_ctx *ctx,                  \
    daedalus_substrate sub, uint8_t *dst, const uint8_t *src, size_t stride,   \
    size_t n_blocks, const daedalus_h264_qpel_meta *meta)                      \
 {                                                                              \
    daedalus_substrate eff = sub;                                              \
    if (eff == DAEDALUS_SUBSTRATE_AUTO)                                        \
        eff = daedalus_recipe_substrate_for(kernel);                           \
    if (eff == DAEDALUS_SUBSTRATE_QPU && !daedalus_ctx_has_qpu(ctx))           \
        eff = DAEDALUS_SUBSTRATE_CPU;                                          \
    if (eff == DAEDALUS_SUBSTRATE_CPU)                                         \
        return dispatch_h264_qpel_ ## suffix ## _cpu(ctx, dst, src, stride,    \
                                                       n_blocks, meta);        \
    return dispatch_h264_qpel_ ## suffix ## _qpu(ctx, dst, src, stride,        \
                                                   n_blocks, meta);            \
 }
 DEFINE_QPEL_DIAG_PUBLIC(mc11, DAEDALUS_KERNEL_H264_QPEL_MC11)
 DEFINE_QPEL_DIAG_PUBLIC(mc12, DAEDALUS_KERNEL_H264_QPEL_MC12)
 DEFINE_QPEL_DIAG_PUBLIC(mc13, DAEDALUS_KERNEL_H264_QPEL_MC13)
 DEFINE_QPEL_DIAG_PUBLIC(mc21, DAEDALUS_KERNEL_H264_QPEL_MC21)
 DEFINE_QPEL_DIAG_PUBLIC(mc23, DAEDALUS_KERNEL_H264_QPEL_MC23)
 DEFINE_QPEL_DIAG_PUBLIC(mc31, DAEDALUS_KERNEL_H264_QPEL_MC31)
 DEFINE_QPEL_DIAG_PUBLIC(mc32, DAEDALUS_KERNEL_H264_QPEL_MC32)
 DEFINE_QPEL_DIAG_PUBLIC(mc33, DAEDALUS_KERNEL_H264_QPEL_MC33)
 /* avg_ biprediction dispatchers (15 positions) — same macro, the
 * underlying _qpu dispatch fns also reuse the diag QPU helper since
 * the avg_ shaders share the put_ src envelope (the L2 step only
 * touches dst). */
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc20, DAEDALUS_KERNEL_H264_QPEL_AVG_MC20)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc02, DAEDALUS_KERNEL_H264_QPEL_AVG_MC02)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc22, DAEDALUS_KERNEL_H264_QPEL_AVG_MC22)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc10, DAEDALUS_KERNEL_H264_QPEL_AVG_MC10)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc30, DAEDALUS_KERNEL_H264_QPEL_AVG_MC30)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc01, DAEDALUS_KERNEL_H264_QPEL_AVG_MC01)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc03, DAEDALUS_KERNEL_H264_QPEL_AVG_MC03)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc11, DAEDALUS_KERNEL_H264_QPEL_AVG_MC11)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc12, DAEDALUS_KERNEL_H264_QPEL_AVG_MC12)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc13, DAEDALUS_KERNEL_H264_QPEL_AVG_MC13)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc21, DAEDALUS_KERNEL_H264_QPEL_AVG_MC21)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc23, DAEDALUS_KERNEL_H264_QPEL_AVG_MC23)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc31, DAEDALUS_KERNEL_H264_QPEL_AVG_MC31)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc32, DAEDALUS_KERNEL_H264_QPEL_AVG_MC32)
 DEFINE_QPEL_DIAG_PUBLIC(avg_mc33, DAEDALUS_KERNEL_H264_QPEL_AVG_MC33)
 #undef DEFINE_QPEL_DIAG_PUBLIC
 #undef DEFINE_QPEL_DISPATCH
@@ -0,0 +1,52 @@
 // daedalus-fourier — H.264 luma qpel avg_mc01 (biprediction) (8x8, ¼-pel vertical),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 "d" position:
 //
 //   dst[r,c] = ((clip255(mc02(s)[r,c]) + s[r,c] + 1) >> 1)
 //
 // Sibling of v3d_h264_qpel_mc02.comp with L2 step against src[r, c].
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc01_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int vp = clamp(v >> 5, 0, 255);
    int avg = (vp + s_0 + 1) >> 1;    // L2 with src[r, c]
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,77 @@
 // daedalus-fourier — H.264 luma qpel avg_mc02 (biprediction) (8x8, vertical half-pel), V3D 7.1.
 //
 // Sibling of cycle 9's v3d_h264_qpel_mc20.comp.  Same 6-tap filter,
 // transposed to vertical direction:
 //
 //   dst[r,c] = clip255(
 //       ( s[r-2,c]
 //         - 5 * s[r-1,c]
 //         + 20 * s[r,  c]
 //         + 20 * s[r+1,c]
 //         -  5 * s[r+2,c]
 //         +      s[r+3,c]
 //         + 16
 //       ) >> 5)
 //
 // src+src_off points at row 0 col 0 of the OUTPUT block; the filter
 // reads rows -2..+3 (2 rows of top context, 3 rows of bottom).
 //
 // Same WG layout as mc20: 64 lanes / 1 block-per-WG / 1 lane-per-pixel.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc02_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC {
    uint n_blocks;
    uint stride_u8;
    uint _pad0, _pad1;
 } pc;
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3;
    uint c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    // Read the 6 rows of vertical context at col (c) of THIS output row.
    // src_off+r*stride+c is at the OUTPUT pixel position; the kernel
    // samples r-2..r+3 along the column.  Unsigned-safe because the
    // public API contract guarantees src_off >= 2*stride.
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int p = clamp(v >> 5, 0, 255);
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + p + 1) >> 1);
 }
@@ -0,0 +1,52 @@
 // daedalus-fourier — H.264 luma qpel avg_mc03 (biprediction) (8x8, ¾-pel vertical),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 "n" position:
 //
 //   dst[r,c] = ((clip255(mc02(s)[r,c]) + s[r+1, c] + 1) >> 1)
 //
 // Same as mc01 but L2-averages with src[r+1, c] instead of src[r, c].
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc03_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int vp = clamp(v >> 5, 0, 255);
    int avg = (vp + s_p1 + 1) >> 1;   // L2 with src[r+1, c]
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,55 @@
 // daedalus-fourier — H.264 luma qpel avg_mc10 (biprediction) (8x8, ¼-pel horizontal),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 "a" position:
 //
 //   dst[r,c] = ((clip255(mc20(s)[r,c]) + s[r,c] + 1) >> 1)
 //
 // = horizontal half-pel filter, clipped to u8, then L2 rounded-averaged
 // with the integer source pixel at the SAME position.  Sibling of
 // v3d_h264_qpel_mc20.comp with the L2 step added at the tail.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc10_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int hp = clamp(v >> 5, 0, 255);
    // L2 average with the integer source at the SAME (r, c) position.
    int avg = (hp + s_0 + 1) >> 1;
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,96 @@
 // daedalus-fourier — H.264 luma qpel avg_mc11 (biprediction) (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc11[r,c] = avg(mc20(r, c),
 //                     mc02(r, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc11_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_h(src_off, stride, r, c);
    int b = hpel_v(src_off, stride, r, c);
    int avg = (a + b + 1) >> 1;
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,96 @@
 // daedalus-fourier — H.264 luma qpel avg_mc12 (biprediction) (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc12[r,c] = avg(mc22(r, c),
 //                     mc02(r, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc12_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_hv(src_off, stride, r, c);
    int b = hpel_v(src_off, stride, r, c);
    int avg = (a + b + 1) >> 1;
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,96 @@
 // daedalus-fourier — H.264 luma qpel avg_mc13 (biprediction) (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc13[r,c] = avg(mc20(r+1, c),
 //                     mc02(r, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc13_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_h(src_off, stride, r+1u, c);
    int b = hpel_v(src_off, stride, r, c);
    int avg = (a + b + 1) >> 1;
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,91 @@
 // daedalus-fourier — H.264 luma qpel avg_mc20 (biprediction) (8x8, horizontal half-pel), V3D 7.1.
 //
 // H.264 spec §8.4.2.2.1 horizontal 6-tap luma interpolation:
 //
 //   dst[r,c] = clip255(
 //       ( s[r,c-2]
 //         - 5 * s[r,c-1]
 //         + 20 * s[r,c]
 //         + 20 * s[r,c+1]
 //         -  5 * s[r,c+2]
 //         +      s[r,c+3]
 //         + 16
 //       ) >> 5)
 //
 // Single-stride: dst and src share `stride` (H264QpelContext
 // convention).  src+src_off already points at the leftmost output
 // column (col 0); the filter reads cols -2..+3.  Caller guarantees
 // edge-padding context per the public API docstring.
 //
 // Workgroup layout: 64 invocations = 1 lane per output pixel.
 // 1 block per WG; n_blocks WGs total.  This is the simplest layout
 // that avoids any inter-lane communication — each lane independently
 // reads its 6 src samples and writes its 1 dst sample.  V3D's L2
 // cache handles the redundant reads from adjacent lanes.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc20_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src {
    uint8_t src[];
 } u_src;
 layout(binding = 1) buffer Dst {
    uint8_t dst[];
 } u_dst;
 layout(binding = 2) readonly buffer Meta {
    uvec4 meta[];       // .x = dst_off, .y = src_off
 } u_meta;
 layout(push_constant) uniform PC {
    uint n_blocks;
    uint stride_u8;
    uint _pad0, _pad1;
 } pc;
 void main()
 {
    // 1 block per WG, 64 lanes covering the 8x8 output block.
    uint wg_id      = gl_WorkGroupID.x;
    uint block_idx  = wg_id;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3;    // 0..7 (row)
    uint c = lane & 7u;    // 0..7 (column)
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    // src points at output col 0 of the block; filter reads cols -2..+3
    // of the current row.  Negative col arithmetic is unsigned-safe
    // because src_off >= 2 (caller-guaranteed left context).
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base + 0u]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int p = clamp(v >> 5, 0, 255);
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + p + 1) >> 1);
 }
@@ -0,0 +1,96 @@
 // daedalus-fourier — H.264 luma qpel avg_mc21 (biprediction) (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc21[r,c] = avg(mc22(r, c),
 //                     mc20(r, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc21_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_hv(src_off, stride, r, c);
    int b = hpel_h(src_off, stride, r, c);
    int avg = (a + b + 1) >> 1;
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,94 @@
 // daedalus-fourier — H.264 luma qpel avg_mc22 (biprediction) (8x8, 2D half-pel "j" position).
 // V3D 7.1.
 //
 // Cascaded H+V 6-tap per H.264 §8.4.2.2.1 / FFmpeg ff_put_h264_qpel8_mc22_neon:
 //
 //   tmp[r,c] = src[r,c-2] - 5*src[r,c-1] + 20*src[r,c] + 20*src[r,c+1]
 //              - 5*src[r,c+2] + src[r,c+3]                    (int16)
 //
 //   dst[r,c] = clip255((tmp[r-2,c] - 5*tmp[r-1,c] + 20*tmp[r,c]
 //                       + 20*tmp[r+1,c] - 5*tmp[r+2,c] + tmp[r+3,c]
 //                       + 512) >> 10)
 //
 // The +512 >> 10 final scale compensates for both 6-tap scalings.
 // CANNOT just cascade mc20→mc02 because intermediate must be int16
 // (no per-stage clip), so this is a dedicated kernel.
 //
 // Per-lane structure: each lane computes its own (r, c) output by
 // running the FULL cascade — 6 horizontal lowpass int16 values for
 // rows r-2..r+3, then a vertical lowpass on those.  ~50 ALU ops per
 // lane.  No shared memory / barriers needed; V3D L2 absorbs the
 // redundant src reads across lanes.
 //
 // WG layout: 64 lanes / 1 block-per-WG / 1 lane-per-output-pixel
 // (same as mc20 / mc02).
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc22_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC {
    uint n_blocks;
    uint stride_u8;
    uint _pad0, _pad1;
 } pc;
 // Horizontal 6-tap filter at (row_off, c) — reads src at cols c-2..c+3
 // of the row identified by row_off, returns int16 intermediate (NOT
 // scaled — the v-pass does the +512 >> 10 for both stages).
 int hpel_h(uint row_off, uint c)
 {
    int s_m2 = int(u_src.src[row_off + c - 2u]);
    int s_m1 = int(u_src.src[row_off + c - 1u]);
    int s_0  = int(u_src.src[row_off + c       ]);
    int s_p1 = int(u_src.src[row_off + c + 1u]);
    int s_p2 = int(u_src.src[row_off + c + 2u]);
    int s_p3 = int(u_src.src[row_off + c + 3u]);
    return s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3;
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3;
    uint c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    // Compute 6 horizontal lowpass values at rows r-2..r+3 (relative
    // to the output row r) of column c.  src_off+r*stride+c is the
    // output pixel position; we sample rows r-2..r+3.
    // Unsigned-safe because src_off >= 2*stride per the caller contract.
    int t0 = hpel_h(src_off + (r - 2u) * stride, c);
    int t1 = hpel_h(src_off + (r - 1u) * stride, c);
    int t2 = hpel_h(src_off +  r       * stride, c);
    int t3 = hpel_h(src_off + (r + 1u) * stride, c);
    int t4 = hpel_h(src_off + (r + 2u) * stride, c);
    int t5 = hpel_h(src_off + (r + 3u) * stride, c);
    int v = t0 - 5 * t1 + 20 * t2 + 20 * t3 - 5 * t4 + t5 + 512;
    int p = clamp(v >> 10, 0, 255);
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + p + 1) >> 1);
 }
@@ -0,0 +1,96 @@
 // daedalus-fourier — H.264 luma qpel avg_mc23 (biprediction) (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc23[r,c] = avg(mc22(r, c),
 //                     mc20(r+1, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc23_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_hv(src_off, stride, r, c);
    int b = hpel_h(src_off, stride, r+1u, c);
    int avg = (a + b + 1) >> 1;
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,52 @@
 // daedalus-fourier — H.264 luma qpel avg_mc30 (biprediction) (8x8, ¾-pel horizontal),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 "c" position:
 //
 //   dst[r,c] = ((clip255(mc20(s)[r,c]) + s[r,c+1] + 1) >> 1)
 //
 // Same as mc10 but L2-averages with src[r, c+1] instead of src[r, c].
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc30_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int hp = clamp(v >> 5, 0, 255);
    int avg = (hp + s_p1 + 1) >> 1;   // L2 with src[r, c+1]
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,96 @@
 // daedalus-fourier — H.264 luma qpel avg_mc31 (biprediction) (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc31[r,c] = avg(mc20(r, c),
 //                     mc02(r, c+1))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc31_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_h(src_off, stride, r, c);
    int b = hpel_v(src_off, stride, r, c+1u);
    int avg = (a + b + 1) >> 1;
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,96 @@
 // daedalus-fourier — H.264 luma qpel avg_mc32 (biprediction) (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc32[r,c] = avg(mc22(r, c),
 //                     mc02(r, c+1))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc32_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_hv(src_off, stride, r, c);
    int b = hpel_v(src_off, stride, r, c+1u);
    int avg = (a + b + 1) >> 1;
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,96 @@
 // daedalus-fourier — H.264 luma qpel avg_mc33 (biprediction) (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc33[r,c] = avg(mc20(r+1, c),
 //                     mc02(r, c+1))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 //
 // avg_ variant for B-slice biprediction per H.264 §8.4.2.3.1:
 //   dst[r,c] = avg(dst[r,c], mc33_value)
 // Caller pre-loads dst with the list0 prediction; this shader
 // folds in the list1 contribution.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_h(src_off, stride, r+1u, c);
    int b = hpel_v(src_off, stride, r, c+1u);
    int avg = (a + b + 1) >> 1;
    uint final_off = dst_off + r * stride + c;
    int prev = int(u_dst.dst[final_off]);
    u_dst.dst[final_off] = uint8_t((prev + avg + 1) >> 1);
 }
@@ -0,0 +1,44 @@
 // daedalus-fourier — H.264 luma qpel mc01 (8x8, ¼-pel vertical),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 "d" position:
 //
 //   dst[r,c] = ((clip255(mc02(s)[r,c]) + s[r,c] + 1) >> 1)
 //
 // Sibling of v3d_h264_qpel_mc02.comp with L2 step against src[r, c].
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int vp = clamp(v >> 5, 0, 255);
    int avg = (vp + s_0 + 1) >> 1;    // L2 with src[r, c]
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -1,7 +1,6 @@
 // daedalus-fourier — H.264 luma qpel mc02 (8x8, vertical half-pel), V3D 7.1.
 //
-// Sibling of cycle 9's v3d_h264_qpel_mc20.comp.  Same 6-tap filter,
+// v2: cooperative-load shared-memory tile.
 // transposed to vertical direction:
 //
 //   dst[r,c] = clip255(
 //       ( s[r-2,c]
@@ -14,9 +13,30 @@
 //       ) >> 5)
 //
 // src+src_off points at row 0 col 0 of the OUTPUT block; the filter
-// reads rows -2..+3 (2 rows of top context, 3 rows of bottom).
+// reads rows -2..+3 (2 rows of top context, 3 rows of bottom), total
 // 13 distinct source rows × 8 cols = 104 bytes per 8x8 output.
 //
-// Same WG layout as mc20: 64 lanes / 1 block-per-WG / 1 lane-per-pixel.
+// v1 had each of the 64 lanes do 6 SSBO loads → 384 loads/WG to cover
 // 104 unique bytes (3.7x redundant), and each lane's loads were stride-
 // spaced (one cache line per byte under V3D's TMU).  PR #36 bench
 // showed mc02 was the only qpel position where CPU NEON still beat
 // QPU (16.96 ns/op CPU vs 20.54 ns/op QPU; 1.21x CPU favoring).
 //
 // v2 splits the work into a coalesced load phase + a shared-memory
 // compute phase:
 //
 //   Phase 1: each of the 64 lanes cooperatively loads the 104-byte
 //   source tile into shared memory.  Lanes 0..63 load bytes at indices
 //   0..63 (covers source rows 0..7 of the 13-row tile); lanes 0..39
 //   second-load bytes 64..103 (rows 8..12).  Reads within a row are
 //   contiguous so the SIMD groups coalesce; total SSBO loads = 104,
 //   matching the unique-byte count.
 //
 //   Phase 2: all 64 lanes compute one output pixel each, reading 6
 //   bytes from shared.  Shared-memory access on V3D is local-store
 //   backed (no TMU round-trip).
 //
 // Same WG layout as v1: 64 lanes / 1 block-per-WG / 1 lane-per-pixel.
 //
 // License: BSD-2-Clause.
@@ -36,31 +56,52 @@ layout(push_constant) uniform PC {
    uint _pad0, _pad1;
 } pc;
 // 13 source rows × 8 cols.  int storage (4 bytes each) — wasteful vs
 // uint8_t but avoids 8-bit-shared interop concerns on glslang+v3dv;
 // 416 bytes shared/WG is well within any reasonable local-store budget.
 shared int s_tile[13 * 8];
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3;
    uint c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
-    // Read the 6 rows of vertical context at col (c) of THIS output row.
+    // Source-tile base: src_off points at output-row-0 col-0, the tile
-    // src_off+r*stride+c is at the OUTPUT pixel position; the kernel
+    // starts 2 rows above.  Unsigned-safe because the public API
-    // samples r-2..r+3 along the column.  Unsigned-safe because the
+    // contract guarantees src_off >= 2*stride.
-    // public API contract guarantees src_off >= 2*stride.
+    uint tile_base = src_off - 2u * stride;
    uint col_base = src_off + c;
-    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
+    // Phase 1: cooperative load — 64 lanes load 104 bytes.
-    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
+    {
-    int s_0  = int(u_src.src[col_base +  r       * stride]);
+        uint sr = lane >> 3;        // 0..7
-    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
+        uint sc = lane & 7u;
-    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
+        s_tile[lane] = int(u_src.src[tile_base + sr * stride + sc]);
-    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
+    }
    if (lane < 40u) {
        uint idx = lane + 64u;      // 64..103
        uint sr = idx >> 3;         // 8..12
        uint sc = idx & 7u;
        s_tile[idx] = int(u_src.src[tile_base + sr * stride + sc]);
    }
    barrier();
    // Phase 2: each lane computes one output pixel from the shared tile.
    uint r = lane >> 3;
    uint c = lane & 7u;
    int s_m2 = s_tile[(r + 0u) * 8u + c];
    int s_m1 = s_tile[(r + 1u) * 8u + c];
    int s_0  = s_tile[(r + 2u) * 8u + c];
    int s_p1 = s_tile[(r + 3u) * 8u + c];
    int s_p2 = s_tile[(r + 4u) * 8u + c];
    int s_p3 = s_tile[(r + 5u) * 8u + c];
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int p = clamp(v >> 5, 0, 255);
@@ -0,0 +1,44 @@
 // daedalus-fourier — H.264 luma qpel mc03 (8x8, ¾-pel vertical),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 "n" position:
 //
 //   dst[r,c] = ((clip255(mc02(s)[r,c]) + s[r+1, c] + 1) >> 1)
 //
 // Same as mc01 but L2-averages with src[r+1, c] instead of src[r, c].
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int vp = clamp(v >> 5, 0, 255);
    int avg = (vp + s_p1 + 1) >> 1;   // L2 with src[r+1, c]
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,47 @@
 // daedalus-fourier — H.264 luma qpel mc10 (8x8, ¼-pel horizontal),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 "a" position:
 //
 //   dst[r,c] = ((clip255(mc20(s)[r,c]) + s[r,c] + 1) >> 1)
 //
 // = horizontal half-pel filter, clipped to u8, then L2 rounded-averaged
 // with the integer source pixel at the SAME position.  Sibling of
 // v3d_h264_qpel_mc20.comp with the L2 step added at the tail.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int hp = clamp(v >> 5, 0, 255);
    // L2 average with the integer source at the SAME (r, c) position.
    int avg = (hp + s_0 + 1) >> 1;
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,88 @@
 // daedalus-fourier — H.264 luma qpel mc11 (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc11[r,c] = avg(mc20(r, c),
 //                     mc02(r, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_h(src_off, stride, r, c);
    int b = hpel_v(src_off, stride, r, c);
    int avg = (a + b + 1) >> 1;
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,88 @@
 // daedalus-fourier — H.264 luma qpel mc12 (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc12[r,c] = avg(mc22(r, c),
 //                     mc02(r, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_hv(src_off, stride, r, c);
    int b = hpel_v(src_off, stride, r, c);
    int avg = (a + b + 1) >> 1;
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,88 @@
 // daedalus-fourier — H.264 luma qpel mc13 (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc13[r,c] = avg(mc20(r+1, c),
 //                     mc02(r, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_h(src_off, stride, r+1u, c);
    int b = hpel_v(src_off, stride, r, c);
    int avg = (a + b + 1) >> 1;
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,88 @@
 // daedalus-fourier — H.264 luma qpel mc21 (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc21[r,c] = avg(mc22(r, c),
 //                     mc20(r, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_hv(src_off, stride, r, c);
    int b = hpel_h(src_off, stride, r, c);
    int avg = (a + b + 1) >> 1;
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,88 @@
 // daedalus-fourier — H.264 luma qpel mc23 (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc23[r,c] = avg(mc22(r, c),
 //                     mc20(r+1, c))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_hv(src_off, stride, r, c);
    int b = hpel_h(src_off, stride, r+1u, c);
    int avg = (a + b + 1) >> 1;
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,44 @@
 // daedalus-fourier — H.264 luma qpel mc30 (8x8, ¾-pel horizontal),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 "c" position:
 //
 //   dst[r,c] = ((clip255(mc20(s)[r,c]) + s[r,c+1] + 1) >> 1)
 //
 // Same as mc10 but L2-averages with src[r, c+1] instead of src[r, c].
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5 * s_m1 + 20 * s_0 + 20 * s_p1 - 5 * s_p2 + s_p3 + 16;
    int hp = clamp(v >> 5, 0, 255);
    int avg = (hp + s_p1 + 1) >> 1;   // L2 with src[r, c+1]
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,88 @@
 // daedalus-fourier — H.264 luma qpel mc31 (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc31[r,c] = avg(mc20(r, c),
 //                     mc02(r, c+1))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_h(src_off, stride, r, c);
    int b = hpel_v(src_off, stride, r, c+1u);
    int avg = (a + b + 1) >> 1;
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,88 @@
 // daedalus-fourier — H.264 luma qpel mc32 (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc32[r,c] = avg(mc22(r, c),
 //                     mc02(r, c+1))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_hv(src_off, stride, r, c);
    int b = hpel_v(src_off, stride, r, c+1u);
    int avg = (a + b + 1) >> 1;
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,88 @@
 // daedalus-fourier — H.264 luma qpel mc33 (8x8, diagonal quarter-pel),
 // V3D 7.1.  Per H.264 §8.4.2.2.1 (table 8-4) — composes two half-pel
 // anchors via L2 rounded-average:
 //
 //   mc33[r,c] = avg(mc20(r+1, c),
 //                     mc02(r, c+1))
 //
 // Per-lane structure: each lane computes BOTH anchor outputs at its
 // own (r, c) target offset, then L2 averages.  No shared memory.
 // Same WG geometry as the other qpel shaders.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage             : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Src  { uint8_t src[]; } u_src;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(binding = 2) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(push_constant) uniform PC { uint n_blocks, stride_u8, _p0, _p1; } pc;
 int hpel_h(uint src_off, uint stride, uint r, uint c) {
    uint row_base = src_off + r * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_v(uint src_off, uint stride, uint r, uint c) {
    uint col_base = src_off + c;
    int s_m2 = int(u_src.src[col_base + (r - 2u) * stride]);
    int s_m1 = int(u_src.src[col_base + (r - 1u) * stride]);
    int s_0  = int(u_src.src[col_base +  r       * stride]);
    int s_p1 = int(u_src.src[col_base + (r + 1u) * stride]);
    int s_p2 = int(u_src.src[col_base + (r + 2u) * stride]);
    int s_p3 = int(u_src.src[col_base + (r + 3u) * stride]);
    int v = s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3 + 16;
    return clamp(v >> 5, 0, 255);
 }
 int hpel_hv_row(uint src_off, uint stride, uint rr, uint c) {
    // Single row's int16 horizontal lowpass (NOT clipped — used as
    // intermediate for the vertical pass of hpel_hv).
    uint row_base = src_off + rr * stride + c;
    int s_m2 = int(u_src.src[row_base - 2u]);
    int s_m1 = int(u_src.src[row_base - 1u]);
    int s_0  = int(u_src.src[row_base       ]);
    int s_p1 = int(u_src.src[row_base + 1u]);
    int s_p2 = int(u_src.src[row_base + 2u]);
    int s_p3 = int(u_src.src[row_base + 3u]);
    return s_m2 - 5*s_m1 + 20*s_0 + 20*s_p1 - 5*s_p2 + s_p3;
 }
 int hpel_hv(uint src_off, uint stride, uint r, uint c) {
    int t0 = hpel_hv_row(src_off, stride, r - 2u, c);
    int t1 = hpel_hv_row(src_off, stride, r - 1u, c);
    int t2 = hpel_hv_row(src_off, stride, r,       c);
    int t3 = hpel_hv_row(src_off, stride, r + 1u, c);
    int t4 = hpel_hv_row(src_off, stride, r + 2u, c);
    int t5 = hpel_hv_row(src_off, stride, r + 3u, c);
    int v = t0 - 5*t1 + 20*t2 + 20*t3 - 5*t4 + t5 + 512;
    return clamp(v >> 10, 0, 255);
 }
 void main()
 {
    uint block_idx = gl_WorkGroupID.x;
    if (block_idx >= pc.n_blocks) return;
    uint lane = gl_LocalInvocationID.x;
    uint r = lane >> 3, c = lane & 7u;
    uint dst_off = u_meta.meta[block_idx].x;
    uint src_off = u_meta.meta[block_idx].y;
    uint stride  = pc.stride_u8;
    int a = hpel_h(src_off, stride, r+1u, c);
    int b = hpel_v(src_off, stride, r, c+1u);
    int avg = (a + b + 1) >> 1;
    u_dst.dst[dst_off + r * stride + c] = uint8_t(avg);
 }
@@ -0,0 +1,44 @@
 // daedalus-fourier — H.264 chroma 4:2:0 intra (bS=4) H deblock —
 // V3D 7.1.  Transpose of v3d_h264deblock_chroma_v_intra.comp.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage              : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(push_constant) uniform PC {
    uint n_edges, dst_stride_u8, _p0, _p1;
 } pc;
 void main()
 {
    uint lane_in_wg  = gl_GlobalInvocationID.x & 255u;
    uint edge_in_wg  = lane_in_wg >> 4;
    uint row_in_edge = lane_in_wg & 15u;
    uint edge_idx    = gl_WorkGroupID.x * 16u + edge_in_wg;
    if (edge_idx >= pc.n_edges) return;
    if (row_in_edge >= 8u) return;
    uvec4 m = u_meta.meta[edge_idx];
    uint stride  = pc.dst_stride_u8;
    uint dst_off = m.x + row_in_edge * stride;
    int alpha = int(m.y & 0xffu);
    int beta  = int((m.y >> 8) & 0xffu);
    if ((alpha | beta) == 0) return;
    int p1 = int(u_dst.dst[dst_off - 2u]);
    int p0 = int(u_dst.dst[dst_off - 1u]);
    int q0 = int(u_dst.dst[dst_off       ]);
    int q1 = int(u_dst.dst[dst_off + 1u]);
    if (abs(p0 - q0) >= alpha) return;
    if (abs(p1 - p0) >= beta)  return;
    if (abs(q1 - q0) >= beta)  return;
    u_dst.dst[dst_off - 1u] = uint8_t(clamp((2*p1 + p0 + q1 + 2) >> 2, 0, 255));
    u_dst.dst[dst_off       ] = uint8_t(clamp((2*q1 + q0 + p1 + 2) >> 2, 0, 255));
 }
@@ -0,0 +1,54 @@
 // daedalus-fourier — H.264 chroma 4:2:0 intra (bS=4) V deblock —
 // V3D 7.1.  Per H.264 §8.3.2.3 chroma intra path: simpler than luma
 // — always weak filter, only p0/q0 updated, 8 cells per edge.
 //
 // p0' = (2*p1 + p0 + q1 + 2) >> 2
 // q0' = (2*q1 + q0 + p1 + 2) >> 2
 //
 // Same 16-edges × 16-lanes/edge WG shape as luma; lanes 8..15 of each
 // edge early-return (chroma edges are only 8 cells wide).
 //
 // 4:2:0-only — caller-side gating handles 4:2:2 (chroma_format_idc>1)
 // at the libavcodec init layer.
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage              : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(push_constant) uniform PC {
    uint n_edges, dst_stride_u8, _p0, _p1;
 } pc;
 void main()
 {
    uint lane_in_wg  = gl_GlobalInvocationID.x & 255u;
    uint edge_in_wg  = lane_in_wg >> 4;
    uint col_in_edge = lane_in_wg & 15u;
    uint edge_idx    = gl_WorkGroupID.x * 16u + edge_in_wg;
    if (edge_idx >= pc.n_edges) return;
    if (col_in_edge >= 8u) return;
    uvec4 m = u_meta.meta[edge_idx];
    uint dst_off = m.x + col_in_edge;
    uint stride  = pc.dst_stride_u8;
    int alpha = int(m.y & 0xffu);
    int beta  = int((m.y >> 8) & 0xffu);
    if ((alpha | beta) == 0) return;
    int p1 = int(u_dst.dst[dst_off - 2u * stride]);
    int p0 = int(u_dst.dst[dst_off - 1u * stride]);
    int q0 = int(u_dst.dst[dst_off]);
    int q1 = int(u_dst.dst[dst_off + 1u * stride]);
    if (abs(p0 - q0) >= alpha) return;
    if (abs(p1 - p0) >= beta)  return;
    if (abs(q1 - q0) >= beta)  return;
    u_dst.dst[dst_off - 1u * stride] = uint8_t(clamp((2*p1 + p0 + q1 + 2) >> 2, 0, 255));
    u_dst.dst[dst_off            ]   = uint8_t(clamp((2*q1 + q0 + p1 + 2) >> 2, 0, 255));
 }
@@ -0,0 +1,70 @@
 // daedalus-fourier — H.264 luma intra (bS=4) H deblock — V3D 7.1.
 //
 // Sibling of v3d_h264deblock_luma_v_intra.comp transposed to the
 // horizontal axis: lane → row, reads pix[-4..+3] (cols) instead of
 // pix[-4*stride..+3*stride] (rows).  Same strong/weak filter
 // selector + same write-back algebra.
 //
 // dst_off contract: (m.x % stride) ≥ 4 (kernel reads p3 at pix[-4]).
 //
 // License: BSD-2-Clause.
 #version 450
 #extension GL_EXT_shader_8bit_storage              : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(push_constant) uniform PC {
    uint n_edges, dst_stride_u8, _p0, _p1;
 } pc;
 void main()
 {
    uint lane_in_wg  = gl_GlobalInvocationID.x & 255u;
    uint edge_in_wg  = lane_in_wg >> 4;
    uint row_in_edge = lane_in_wg & 15u;
    uint edge_idx    = gl_WorkGroupID.x * 16u + edge_in_wg;
    if (edge_idx >= pc.n_edges) return;
    uvec4 m = u_meta.meta[edge_idx];
    uint stride = pc.dst_stride_u8;
    uint dst_off = m.x + row_in_edge * stride;
    int alpha = int(m.y & 0xffu);
    int beta  = int((m.y >> 8) & 0xffu);
    if ((alpha | beta) == 0) return;
    int p3 = int(u_dst.dst[dst_off - 4u]);
    int p2 = int(u_dst.dst[dst_off - 3u]);
    int p1 = int(u_dst.dst[dst_off - 2u]);
    int p0 = int(u_dst.dst[dst_off - 1u]);
    int q0 = int(u_dst.dst[dst_off       ]);
    int q1 = int(u_dst.dst[dst_off + 1u]);
    int q2 = int(u_dst.dst[dst_off + 2u]);
    int q3 = int(u_dst.dst[dst_off + 3u]);
    if (abs(p0 - q0) >= alpha) return;
    if (abs(p1 - p0) >= beta)  return;
    if (abs(q1 - q0) >= beta)  return;
    bool strong_common = abs(p0 - q0) < (alpha >> 2) + 2;
    bool strong_p = strong_common && abs(p2 - p0) < beta;
    bool strong_q = strong_common && abs(q2 - q0) < beta;
    if (strong_p) {
        u_dst.dst[dst_off - 1u] = uint8_t(clamp((p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4) >> 3, 0, 255));
        u_dst.dst[dst_off - 2u] = uint8_t(clamp((p2 + p1 + p0 + q0 + 2) >> 2, 0, 255));
        u_dst.dst[dst_off - 3u] = uint8_t(clamp((2*p3 + 3*p2 + p1 + p0 + q0 + 4) >> 3, 0, 255));
    } else {
        u_dst.dst[dst_off - 1u] = uint8_t(clamp((2*p1 + p0 + q1 + 2) >> 2, 0, 255));
    }
    if (strong_q) {
        u_dst.dst[dst_off       ] = uint8_t(clamp((q2 + 2*q1 + 2*q0 + 2*p0 + p1 + 4) >> 3, 0, 255));
        u_dst.dst[dst_off + 1u] = uint8_t(clamp((q2 + q1 + q0 + p0 + 2) >> 2, 0, 255));
        u_dst.dst[dst_off + 2u] = uint8_t(clamp((2*q3 + 3*q2 + q1 + q0 + p0 + 4) >> 3, 0, 255));
    } else {
        u_dst.dst[dst_off       ] = uint8_t(clamp((2*q1 + q0 + p1 + 2) >> 2, 0, 255));
    }
 }
@@ -0,0 +1,81 @@
 // daedalus-fourier — H.264 luma intra (bS=4) V deblock — V3D 7.1.
 //
 // Per H.264 §8.3.2.3: at I-MB edges and certain inter-MB edges that
 // force boundary strength to 4, the deblock kernel is structurally
 // different from bS<4 — it has a per-side strong/weak filter
 // selector that decides whether to update 3 cells (strong) or 1
 // (weak), reads p3/q3, and ignores tc0.
 //
 // strong_common = |p0-q0| < (α>>2) + 2
 // strong_p      = strong_common AND |p2-p0| < β
 // strong_q      = strong_common AND |q2-q0| < β
 //
 // Strong-p updates p0/p1/p2 with specific 5-/4-/3-tap blends.
 // Weak-p updates p0 only with (2*p1 + p0 + q1 + 2) >> 2.
 // Mirror for q-side.
 //
 // WG geometry identical to v3d_h264deblock.comp (16 edges × 16 lanes/WG).
 // dst_off contract: m.x ≥ 4*stride (kernel reads p3 at -4*stride).
 //
 // License: BSD-2-Clause.  Algorithm transcribed from
 // tests/h264_intra_loop_filter_ref.c (PR #11).
 #version 450
 #extension GL_EXT_shader_8bit_storage              : require
 #extension GL_EXT_shader_explicit_arithmetic_types : require
 layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 layout(binding = 0) readonly buffer Meta { uvec4 meta[]; } u_meta;
 layout(binding = 1) buffer Dst { uint8_t dst[]; } u_dst;
 layout(push_constant) uniform PC {
    uint n_edges, dst_stride_u8, _p0, _p1;
 } pc;
 void main()
 {
    uint lane_in_wg  = gl_GlobalInvocationID.x & 255u;
    uint edge_in_wg  = lane_in_wg >> 4;
    uint col_in_edge = lane_in_wg & 15u;
    uint edge_idx    = gl_WorkGroupID.x * 16u + edge_in_wg;
    if (edge_idx >= pc.n_edges) return;
    uvec4 m = u_meta.meta[edge_idx];
    uint dst_off = m.x + col_in_edge;
    uint stride  = pc.dst_stride_u8;
    int alpha = int(m.y & 0xffu);
    int beta  = int((m.y >> 8) & 0xffu);
    if ((alpha | beta) == 0) return;
    int p3 = int(u_dst.dst[dst_off - 4u * stride]);
    int p2 = int(u_dst.dst[dst_off - 3u * stride]);
    int p1 = int(u_dst.dst[dst_off - 2u * stride]);
    int p0 = int(u_dst.dst[dst_off - 1u * stride]);
    int q0 = int(u_dst.dst[dst_off]);
    int q1 = int(u_dst.dst[dst_off + 1u * stride]);
    int q2 = int(u_dst.dst[dst_off + 2u * stride]);
    int q3 = int(u_dst.dst[dst_off + 3u * stride]);
    if (abs(p0 - q0) >= alpha) return;
    if (abs(p1 - p0) >= beta)  return;
    if (abs(q1 - q0) >= beta)  return;
    bool strong_common = abs(p0 - q0) < (alpha >> 2) + 2;
    bool strong_p = strong_common && abs(p2 - p0) < beta;
    bool strong_q = strong_common && abs(q2 - q0) < beta;
    if (strong_p) {
        u_dst.dst[dst_off - 1u * stride] = uint8_t(clamp((p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4) >> 3, 0, 255));
        u_dst.dst[dst_off - 2u * stride] = uint8_t(clamp((p2 + p1 + p0 + q0 + 2) >> 2, 0, 255));
        u_dst.dst[dst_off - 3u * stride] = uint8_t(clamp((2*p3 + 3*p2 + p1 + p0 + q0 + 4) >> 3, 0, 255));
    } else {
        u_dst.dst[dst_off - 1u * stride] = uint8_t(clamp((2*p1 + p0 + q1 + 2) >> 2, 0, 255));
    }
    if (strong_q) {
        u_dst.dst[dst_off              ] = uint8_t(clamp((q2 + 2*q1 + 2*q0 + 2*p0 + p1 + 4) >> 3, 0, 255));
        u_dst.dst[dst_off + 1u * stride] = uint8_t(clamp((q2 + q1 + q0 + p0 + 2) >> 2, 0, 255));
        u_dst.dst[dst_off + 2u * stride] = uint8_t(clamp((2*q3 + 3*q2 + q1 + q0 + p0 + 4) >> 3, 0, 255));
    } else {
        u_dst.dst[dst_off              ] = uint8_t(clamp((2*q1 + q0 + p1 + 2) >> 2, 0, 255));
    }
 }
@@ -2,25 +2,22 @@
 /* CLOCK_MONOTONIC under -std=c11 -CMAKE_C_EXTENSIONS=OFF. */
 #define _POSIX_C_SOURCE 200809L
 /*
- * bench_h264_primitives — NEON-path latency baseline for the H.264
+ * bench_h264_primitives — latency baseline for the H.264 primitive
- * primitive library landed across PRs #9–#23.
+ * library landed across PRs #9–#35.
 *
 * Each kernel is exercised at a representative per-frame N for 1080p
- * (8160 MBs); the per-kernel total + ns/op + ms/frame are reported.
+ * (8160 MBs); the per-kernel total + ns/op + ms/frame are reported,
- * Lets us answer "what's the total NEON-only budget for the H.264
+ * once per substrate (CPU NEON, QPU V3D7 compute).  The QPU column
- * decode at 1080p" — useful for sizing intercept-patch decisions
+ * appears only when the host has a usable Vulkan device.  When both
- * (which kernels NEED QPU shaders vs which are budget-fine on NEON).
+ * columns exist a CPU/QPU ratio is printed; that's the per-kernel
 * data the QPU-substrate decree (2026-05-23) deliberately overrides
 * but which is still useful to track over time as dispatch overhead
 * shrinks (buffer pool, persistent cmdbuf, dmabuf import — tasks 160-162).
 *
 * NOT a ctest — produces wall-time numbers, doesn't pass/fail.
 *
- * Invoke:   ./build/bench_h264_primitives [iters]
+ * Invoke:   ./build/bench_h264_primitives [iters [warmup]]
- *           (default iters = 50, post-warmup = 5)
+ *           (default iters = 50, warmup = 5)
 *
 * NB: results are inherently approximate — single-core, includes
 * loop overhead + memory access patterns that may not match what
 * a real decode would hit (we touch a small set of pages repeatedly).
 * The numbers are useful for relative comparison and order-of-
 * magnitude sizing, not absolute perf claims.
 */
 #include "daedalus.h"
@@ -46,11 +43,6 @@ static double now_ms(void) {
 /* Per-1080p-frame counts (8160 MBs at 1920x1088). */
 #define MBS_1080P  8160
 #define LUMA_4x4_PER_MB   16   /* if transform_8x8=0 */
 #define LUMA_8x8_PER_MB   4    /* if transform_8x8=1 */
 #define CHROMA_4x4_PER_MB 8    /* 4 Cb + 4 Cr */
 #define DEBLOCK_LUMA_EDGES_PER_MB 4    /* 4 horiz + 4 vert internal+MB-edge — ~4 each */
 #define DEBLOCK_CHROMA_EDGES_PER_MB 2  /* 2 each direction */
 /* Standard benchmark loop.  fn() is called n times per iteration. */
 typedef void (*bench_fn)(void);
@@ -64,16 +56,18 @@ static double bench_ns(const char *name, int iters, int warmup,
    double t1 = now_ms();
    double total_ms = (t1 - t0);
    double ns_per_op = (total_ms * 1e6) / ((double) iters * ops_per_iter);
-    printf("  %-32s %8.2f ns/op  (%d iters x %d ops)\n",
+    printf("  %-32s %10.2f ns/op  (%d iters x %d ops)\n",
           name, ns_per_op, iters, ops_per_iter);
    return ns_per_op;
 }
 /* ---- Per-kernel scaffolding.  Each section sets up the buffers +
 * meta, then defines a static fn() that calls the corresponding
- * dispatch with a representative N. */
+ * dispatch with a representative N.  The substrate is read from the
 * global g_sub so the same fn() can be re-driven with CPU then QPU. */
 static daedalus_ctx          *ctx;
 static daedalus_substrate     g_sub = DAEDALUS_SUBSTRATE_CPU;
 /* --- IDCT 4x4 luma: N = 16 blocks per MB.  Bench with 1024 blocks
 *     per call (64 MBs worth).  Per-MB the dispatch overhead is the
@@ -83,7 +77,7 @@ static daedalus_h264_block_meta idct4_meta[1024];
 static uint8_t              idct_dst[64 * 4 * 16 * 16];   /* 64 MB-rows × ... */
 static void bench_idct4(void) {
-    daedalus_dispatch_h264_idct4(ctx, DAEDALUS_SUBSTRATE_CPU,
+    daedalus_dispatch_h264_idct4(ctx, g_sub,
                                  idct_dst, 64*16, idct4_coeffs, 1024, idct4_meta);
 }
@@ -92,7 +86,7 @@ static int16_t              idct8_coeffs[256 * 64];
 static daedalus_h264_block_meta idct8_meta[256];
 static void bench_idct8(void) {
-    daedalus_dispatch_h264_idct8(ctx, DAEDALUS_SUBSTRATE_CPU,
+    daedalus_dispatch_h264_idct8(ctx, g_sub,
                                  idct_dst, 64*16, idct8_coeffs, 256, idct8_meta);
 }
@@ -101,12 +95,12 @@ static daedalus_h264_deblock_meta deblock_meta[256];
 static uint8_t deblock_dst[256 * 16 * 16];
 static void bench_deblock_v(void) {
-    daedalus_dispatch_h264_deblock_luma_v(ctx, DAEDALUS_SUBSTRATE_CPU,
+    daedalus_dispatch_h264_deblock_luma_v(ctx, g_sub,
                                           deblock_dst, 16, 256, deblock_meta);
 }
 static void bench_deblock_h(void) {
-    daedalus_dispatch_h264_deblock_luma_h(ctx, DAEDALUS_SUBSTRATE_CPU,
+    daedalus_dispatch_h264_deblock_luma_h(ctx, g_sub,
                                           deblock_dst, 16, 256, deblock_meta);
 }
@@ -116,18 +110,43 @@ static uint8_t qpel_dst[256 * 16 * 16];
 static daedalus_h264_qpel_meta qpel_meta[256];
 static void bench_qpel_mc20(void) {
-    daedalus_dispatch_h264_qpel_mc20(ctx, DAEDALUS_SUBSTRATE_CPU,
+    daedalus_dispatch_h264_qpel_mc20(ctx, g_sub,
                                      qpel_dst, qpel_src, 16, 256, qpel_meta);
 }
 static void bench_qpel_mc02(void) {
-    daedalus_dispatch_h264_qpel_mc02(ctx, DAEDALUS_SUBSTRATE_CPU,
+    daedalus_dispatch_h264_qpel_mc02(ctx, g_sub,
                                      qpel_dst, qpel_src, 16, 256, qpel_meta);
 }
 static void bench_qpel_mc22(void) {
-    daedalus_dispatch_h264_qpel_mc22(ctx, DAEDALUS_SUBSTRATE_CPU,
+    daedalus_dispatch_h264_qpel_mc22(ctx, g_sub,
                                      qpel_dst, qpel_src, 16, 256, qpel_meta);
 }
 /* ---- One row of bench output:
 *   - kernel name + N
 *   - CPU ns/op
 *   - QPU ns/op (or "n/a" if Vulkan absent)
 *   - CPU/QPU ratio (>1 means QPU wins; <1 means CPU wins) */
 struct row {
    const char *name;
    int         n_per_call;
    bench_fn    fn;
    double      cpu_ns;
    double      qpu_ns;   /* -1 if not measured */
    int         frame_n;  /* count per 1080p frame */
 };
 static struct row rows[] = {
    {"IDCT 4x4 luma",       1024, bench_idct4,     0, -1, MBS_1080P * 16},
    {"IDCT 8x8 luma",        256, bench_idct8,     0, -1, MBS_1080P *  4},
    {"Deblock luma_v",       256, bench_deblock_v, 0, -1, MBS_1080P *  4},
    {"Deblock luma_h",       256, bench_deblock_h, 0, -1, MBS_1080P *  4},
    {"qpel mc20 (8x8)",      256, bench_qpel_mc20, 0, -1, MBS_1080P *  4},
    {"qpel mc02 (8x8)",      256, bench_qpel_mc02, 0, -1, MBS_1080P *  4},
    {"qpel mc22 (8x8)",      256, bench_qpel_mc22, 0, -1, MBS_1080P *  4},
 };
 #define N_ROWS ((int)(sizeof(rows)/sizeof(rows[0])))
 int main(int argc, char **argv)
 {
    int iters  = argc > 1 ? atoi(argv[1]) : 50;
@@ -138,6 +157,7 @@ int main(int argc, char **argv)
        fprintf(stderr, "ctx create failed (Vulkan?)\n");
        return 1;
    }
    int has_qpu = daedalus_ctx_has_qpu(ctx);
    /* Pre-fill all input buffers with random data so the NEON inner
     * loops see realistic memory access patterns. */
@@ -147,8 +167,7 @@ int main(int argc, char **argv)
        idct8_coeffs[i] = (int16_t)((int)(xs64() % 1024) - 512);
    for (size_t i = 0; i < sizeof(qpel_src); i++) qpel_src[i] = (uint8_t)(xs64() & 0xff);
-    /* IDCT meta: each block at offset i*16 (row layout matters less
+    /* IDCT meta. */
     * here since we're just measuring per-block latency). */
    for (size_t i = 0; i < 1024; i++)
        idct4_meta[i].dst_off = (uint32_t)((i / 16) * 64 + (i % 16) * 4);
    for (size_t i = 0; i < 256; i++)
@@ -162,58 +181,88 @@ int main(int argc, char **argv)
        for (int s = 0; s < 4; s++) deblock_meta[i].tc0[s] = (int8_t)(s + 1);
    }
-    /* qpel meta: src and dst at row 3 col 3 of each 16x16 tile. */
+    /* qpel meta. */
    for (size_t i = 0; i < 256; i++) {
        qpel_meta[i].src_off = (uint32_t)(i * 256 + 3 * 16 + 3);
        qpel_meta[i].dst_off = (uint32_t)(i * 256 + 3 * 16 + 3);
    }
-    printf("bench_h264_primitives: %d iters (%d warmup), substrate=CPU NEON\n",
+    printf("bench_h264_primitives: %d iters (%d warmup)\n", iters, warmup);
-           iters, warmup);
+    printf("  ctx has_qpu=%d  (CPU pass always runs; QPU pass skipped without Vulkan)\n\n", has_qpu);
    printf("Per-call N is set per kernel; ns/op is per BLOCK or EDGE.\n\n");
-    double idct4_ns   = bench_ns("IDCT 4x4 luma",     iters, warmup, 1024, bench_idct4);
+    /* Pass 1: CPU NEON. */
-    double idct8_ns   = bench_ns("IDCT 8x8 luma",     iters, warmup,  256, bench_idct8);
+    g_sub = DAEDALUS_SUBSTRATE_CPU;
-    double debl_v_ns  = bench_ns("Deblock luma_v",    iters, warmup,  256, bench_deblock_v);
+    printf("== CPU NEON ==\n");
-    double debl_h_ns  = bench_ns("Deblock luma_h",    iters, warmup,  256, bench_deblock_h);
+    for (int i = 0; i < N_ROWS; i++)
-    double qmc20_ns   = bench_ns("qpel mc20 (8x8)",   iters, warmup,  256, bench_qpel_mc20);
+        rows[i].cpu_ns = bench_ns(rows[i].name, iters, warmup, rows[i].n_per_call, rows[i].fn);
    double qmc02_ns   = bench_ns("qpel mc02 (8x8)",   iters, warmup,  256, bench_qpel_mc02);
    double qmc22_ns   = bench_ns("qpel mc22 (8x8)",   iters, warmup,  256, bench_qpel_mc22);
-    /* Per-frame budget summary at 1080p (8160 MBs).  Worst-case
+    /* Pass 2: QPU compute (if available). */
-     * assumptions:
+    if (has_qpu) {
-     *   - All MBs are transform_4x4 (16 4x4 IDCTs each) — so 130,560
+        g_sub = DAEDALUS_SUBSTRATE_QPU;
-     *     IDCT 4x4 blocks per frame.  If High profile transform_8x8,
+        printf("\n== QPU V3D7 compute ==\n");
-     *     it'd be 32,640 IDCT 8x8 blocks instead.
+        for (int i = 0; i < N_ROWS; i++)
-     *   - All MBs are intra (no MC — qpel zero) OR all inter (no
+            rows[i].qpu_ns = bench_ns(rows[i].name, iters, warmup, rows[i].n_per_call, rows[i].fn);
-     *     intra prediction).  We report MC at "all inter, all qpel
+    }
-     *     mc22" worst case.
+
-     *   - Deblock: ~4 luma_v + 4 luma_h edges per MB; assume all 8
+    /* Summary table — both substrates side by side. */
-     *     edges trigger filtering. */
+    printf("\n== Per-kernel comparison ==\n");
-    printf("\nProjected 1080p frame budgets (worst-case, CPU NEON only):\n");
+    printf("  %-24s %12s %12s %8s   %7s\n",
-    printf("  IDCT 4x4 (all-4x4 MBs):       %7.2f ms   (%d blocks)\n",
+           "kernel", "CPU ns/op", "QPU ns/op", "winner", "ms/frame");
-           idct4_ns * MBS_1080P * 16 / 1e6, MBS_1080P * 16);
+    for (int i = 0; i < N_ROWS; i++) {
-    printf("  IDCT 8x8 (all-8x8 MBs):       %7.2f ms   (%d blocks)\n",
+        double cpu_ms = rows[i].cpu_ns * rows[i].frame_n / 1e6;
-           idct8_ns * MBS_1080P * 4 / 1e6, MBS_1080P * 4);
+        double qpu_ms = rows[i].qpu_ns > 0 ? rows[i].qpu_ns * rows[i].frame_n / 1e6 : -1;
-    printf("  Deblock luma_v (all MBs):     %7.2f ms   (%d edges)\n",
+        const char *winner;
-           debl_v_ns * MBS_1080P * 4 / 1e6, MBS_1080P * 4);
+        char ratio[16];
-    printf("  Deblock luma_h (all MBs):     %7.2f ms   (%d edges)\n",
+        if (rows[i].qpu_ns <= 0) {
-           debl_h_ns * MBS_1080P * 4 / 1e6, MBS_1080P * 4);
+            winner = "CPU";  /* QPU n/a */
-    printf("  qpel mc22 (all 8x8 blocks):   %7.2f ms   (%d blocks)\n",
+            snprintf(ratio, sizeof(ratio), "n/a");
-           qmc22_ns * MBS_1080P * 4 / 1e6, MBS_1080P * 4);
+        } else if (rows[i].cpu_ns < rows[i].qpu_ns) {
            winner = "CPU";
            snprintf(ratio, sizeof(ratio), "%.2fx", rows[i].qpu_ns / rows[i].cpu_ns);
        } else {
            winner = "QPU";
            snprintf(ratio, sizeof(ratio), "%.2fx", rows[i].cpu_ns / rows[i].qpu_ns);
        }
        char qpu_field[16];
        if (rows[i].qpu_ns > 0) snprintf(qpu_field, sizeof(qpu_field), "%.2f", rows[i].qpu_ns);
        else                    snprintf(qpu_field, sizeof(qpu_field), "n/a");
        char ms_field[24];
        if (qpu_ms > 0)
            snprintf(ms_field, sizeof(ms_field), "%.2f/%.2f", cpu_ms, qpu_ms);
        else
            snprintf(ms_field, sizeof(ms_field), "%.2f/n/a", cpu_ms);
        printf("  %-24s %12.2f %12s   %3s %s   %s\n",
               rows[i].name, rows[i].cpu_ns, qpu_field, winner, ratio, ms_field);
    }
    /* Per-frame budget summary at 1080p (8160 MBs). */
    double cpu_idct4 = rows[0].cpu_ns * MBS_1080P * 16 / 1e6;
    double cpu_debl  = (rows[2].cpu_ns + rows[3].cpu_ns) * MBS_1080P * 4 / 1e6;
    double cpu_mc    = rows[6].cpu_ns * MBS_1080P * 4 / 1e6;   /* mc22 worst-case */
    double cpu_sum   = cpu_idct4 + cpu_debl + cpu_mc;
    printf("\n== Projected 1080p worst-case (CPU NEON only) ==\n");
    printf("  IDCT 4x4 + deblock luma + qpel mc22:  %.2f ms (30fps deadline 33.33)\n", cpu_sum);
    printf("  Margin:                               %+.2f ms\n", 33.33 - cpu_sum);
    if (has_qpu) {
        double qpu_idct4 = rows[0].qpu_ns * MBS_1080P * 16 / 1e6;
        double qpu_debl  = (rows[2].qpu_ns + rows[3].qpu_ns) * MBS_1080P * 4 / 1e6;
        double qpu_mc    = rows[6].qpu_ns * MBS_1080P * 4 / 1e6;
        double qpu_sum   = qpu_idct4 + qpu_debl + qpu_mc;
        printf("\n== Projected 1080p worst-case (QPU V3D7 compute only) ==\n");
        printf("  IDCT 4x4 + deblock luma + qpel mc22:  %.2f ms (30fps deadline 33.33)\n", qpu_sum);
        printf("  Margin:                               %+.2f ms\n", 33.33 - qpu_sum);
        printf("\n  CPU vs QPU sum ratio: %.2fx  (>1 means QPU wins)\n",
               qpu_sum > 0 ? cpu_sum / qpu_sum : 0.0);
    }
    double sum_idct_4x4 = idct4_ns * MBS_1080P * 16 / 1e6;
    double sum_deblock  = (debl_v_ns + debl_h_ns) * MBS_1080P * 4 / 1e6;
    double sum_mc       = qmc22_ns * MBS_1080P * 4 / 1e6;  /* worst-case all-mc22 */
    printf("\n  Sum (IDCT 4x4 + deblock luma + MC all-mc22): %7.2f ms\n",
           sum_idct_4x4 + sum_deblock + sum_mc);
    printf("  30 fps deadline:                              33.33 ms\n");
    printf("  Margin:                                       %+.2f ms\n",
           33.33 - (sum_idct_4x4 + sum_deblock + sum_mc));
    printf("\n(NOT included: chroma deblock, chroma IDCT, intra prediction,\n");
    printf(" CABAC/CAVLC entropy.  These bench numbers are a budget LOWER\n");
-    printf(" bound; the real decode stack adds 20-40%% on top.)\n");
+    printf(" bound; the real decode stack adds 20-40%% on top.\n");
-    (void) qmc20_ns; (void) qmc02_ns;
+    printf(" Per-kernel substrate decisions belong in daedalus_core.c recipe\n");
    printf(" table; the QPU substrate decree (2026-05-23) keeps everything\n");
    printf(" on QPU regardless of these numbers as a policy choice.)\n");
    daedalus_ctx_destroy(ctx);
    return 0;
@@ -683,13 +683,13 @@ int main(void)
    printf("  H264_QPEL_MC20 recipe substrate:  %d\n",
           (int) daedalus_recipe_substrate_for(DAEDALUS_KERNEL_H264_QPEL_MC20));
-    printf("  H264_DEBLOCK_LH recipe substrate: %d (CPU, no QPU H shader yet)\n",
+    printf("  H264_DEBLOCK_LH recipe substrate: %d\n",
           (int) daedalus_recipe_substrate_for(DAEDALUS_KERNEL_H264_DEBLOCK_LH));
-    printf("  H264_DEBLOCK_CV recipe substrate: %d (CPU)\n",
+    printf("  H264_DEBLOCK_CV recipe substrate: %d\n",
           (int) daedalus_recipe_substrate_for(DAEDALUS_KERNEL_H264_DEBLOCK_CV));
-    printf("  H264_DEBLOCK_CH recipe substrate: %d (CPU)\n",
+    printf("  H264_DEBLOCK_CH recipe substrate: %d\n",
           (int) daedalus_recipe_substrate_for(DAEDALUS_KERNEL_H264_DEBLOCK_CH));
-    printf("  H264_DEBLOCK_*_INTRA recipe substrate: %d (CPU, bS=4 set)\n",
+    printf("  H264_DEBLOCK_*_INTRA recipe substrate: %d (bS=4 family, all on QPU)\n",
           (int) daedalus_recipe_substrate_for(DAEDALUS_KERNEL_H264_DEBLOCK_LV_INTRA));
    int fail = 0;
Author	SHA1	Message	Date
claude-noether	a9590acee3	wip: mc02 v2 shared-tile	2026-05-25 21:28:04 +02:00
claude-noether	989818c2e6	bench: H.264 primitive bench now measures both substrates + comparison table Closes task #166 (re-measure R-bands on post-buffer-pool dispatch path). Now that all H.264 hot-path primitives have QPU shaders and the dispatch overhead has been hammered down (tasks #160 buffer pool, #161 persistent command buffer), bench_h264_primitives no longer measures one column. Two passes — CPU NEON and QPU V3D7 compute — with a side-by-side per-kernel comparison and ratio. Headline result on hertz (Pi 5 V3D 7.1, 30 iters x 5 warmup): kernel CPU ns/op QPU ns/op winner IDCT 4x4 luma 10.79 2.47 QPU 4.36x IDCT 8x8 luma 29.69 9.23 QPU 3.22x Deblock luma_v 17.58 10.21 QPU 1.72x Deblock luma_h 38.41 9.98 QPU 3.85x qpel mc20 (8x8) 28.24 9.66 QPU 2.92x qpel mc02 (8x8) 16.96 20.54 CPU 1.21x qpel mc22 (8x8) 71.58 9.64 QPU 7.43x 1080p worst-case sum (IDCT4 + deblock luma + qpel mc22): CPU NEON only: 5.57 ms QPU only: 1.30 ms (CPU/QPU sum ratio = 4.30x) Reverses PR #10's verdict (which had CPU NEON 4x faster than QPU for IDCT-only) — the buffer-pool + persistent-cmdbuf wins land hard. Only qpel mc02 still shows CPU ahead, marginally (single- axis vertical filter, row-strided memory pattern unfriendly to the WG layout — left as a follow-up for cycle-9-style targeted tuning). Substrate decree (2026-05-23) stays in force as policy — these numbers retroactively justify it. Also tightens test_api_h264's startup recipe print: the stale "(CPU)" / "(CPU, no QPU H shader yet)" / "(CPU, bS=4 set)" labels next to deblock_lh, deblock_cv, deblock_ch and deblock_*_intra are now wrong since PRs #28, #29, #35 (those kernels are on QPU).	2026-05-25 20:42:39 +02:00
marfrit	1446b779a6	Merge pull request 'h264: V3D shaders for the 4 bS=4 intra deblock variants — deblock QPU complete' (#35 ) from noether/v3d-shader-h264-deblock-intra into main Reviewed-on: #35	2026-05-25 18:36:10 +00:00
claude-noether	c2d1e9790e	h264: V3D shaders for the 4 bS=4 intra deblock variants — deblock QPU complete Closes the H.264 deblock QPU coverage matrix. Adds the 4 intra (bS=4) variants — luma_v/h_intra + chroma_v/h_intra. Algorithmically distinct from the bS<4 path: - Per-side strong/weak filter selector strong_p = (\|p2-p0\| < β) AND (\|p0-q0\| < (α>>2) + 2) strong_q = (\|q2-q0\| < β) AND (\|p0-q0\| < (α>>2) + 2) - Strong-p updates p0/p1/p2 with 5-/4-/3-tap blends (reads p3) - Weak-p updates p0 only with (2p1 + p0 + q1 + 2) >> 2 - Mirror for q-side; no tc0 (bS=4 hardcodes the strength) - Chroma always weak, only p0/q0 updated (same as bS<4 chroma) Per H.264 §8.3.2.3. Transcribed from PR #11's C reference (tests/h264_intra_loop_filter_ref.c). Shaders: - v3d_h264deblock_luma_v_intra.comp (luma 16-cell + strong/weak) - v3d_h264deblock_luma_h_intra.comp (transpose of luma_v_intra) - v3d_h264deblock_chroma_v_intra.comp (8-cell, always weak) - v3d_h264deblock_chroma_h_intra.comp (transpose of chroma_v_intra) Dispatch wiring: - 4 new pipeline pairs in daedalus_ctx - dispatch_h264_deblock_luma_intra_qpu helper (parameterised by orient_h for V vs H) — 2 wrappers - chroma intra reuses the existing dispatch_h264_deblock_chroma_qpu helper (same WG geometry as bS<4 chroma) — 2 wrappers - DEFINE_INTRA_DISPATCH macro extended with qpu_fn parameter, routes CPU/QPU per recipe table - Recipe table flips DAEDALUS_KERNEL_H264_DEBLOCK__INTRA from CPU to QPU Verified on hertz: $ ./build/test_api_h264 \| grep intra H.264 deblock luma v intra: 1024/1024 bytes bit-exact H.264 deblock luma h intra: 1024/1024 bytes bit-exact H.264 deblock chroma v intra: 256/256 bytes bit-exact H.264 deblock chroma h intra: 256/256 bytes bit-exact All 4 PASS first try. Strong/weak quad-tree selector + per-side asymmetry would have surfaced any sign/shift/index mistake; passing on all 4 (including the asymmetric writes-3-cells cases) means the transcription from C is clean. Deblock QPU coverage matrix — COMPLETE (8 of 8): bS<4 (non-intra): luma_v ✓ cycle 8 luma_h ✓ PR #28 chroma_v ✓ PR #29 chroma_h ✓ PR #29 bS=4 (intra, this PR): luma_v ✓ luma_h ✓ chroma_v ✓ chroma_h ✓ The full H.264 8-bit 4:2:0 hot-path pixel-math layer is now on QPU when daedalus is initialised with a QPU-capable context: - IDCT 4x4 / 8x8 ✓ - All 8 deblock variants ✓ - All 30 qpel positions (15 put_ + 15 avg_) ✓	2026-05-25 20:30:07 +02:00
marfrit	e506ef0803	Merge pull request 'h264: V3D shaders for all 15 avg_ qpel positions — qpel QPU complete' (#34 ) from noether/v3d-shader-h264-qpel-avg into main Reviewed-on: #34	2026-05-25 18:23:11 +00:00
claude-noether	2079fe39c6	h264: V3D shaders for all 15 avg_ qpel positions — qpel QPU complete Generates 15 avg_ shader variants by templating from the existing put_ shaders. Each avg_ shader is identical to its put_ sibling except the final write does an L2 average with the existing dst: put_: dst[r,c] = result avg_: dst[r,c] = (dst[r,c] + result + 1) >> 1 Per H.264 §8.4.2.3.1 (B-slice biprediction): caller pre-loads dst with the list0 prediction; the avg_ call folds in list1. Generated via python (avg-shader-gen.py): reads each v3d_h264_qpel_mcXY.comp, transforms the docstring header + final write hunk, writes v3d_h264_qpel_avg_mcXY.comp. ~88 lines each; 15 new shader files. Dispatch reuses the existing dispatch_h264_qpel_diag_qpu helper for all 15 — same src envelope (10stride+11 covers any (r±1, c±1) shift), the L2 step only touches dst. Slightly over-allocates for the simpler positions (avg_mc20/02/10/30/01/03) but negligible cost. Eliminates 15 wrappers + 15 src_max bound calculations that would otherwise duplicate. CMake foreach loops compile + install 15 new SPV files. ctx grows 15 pipeline pairs. Recipe table flips DAEDALUS_KERNEL_H264_QPEL_AVG_ from CPU to QPU. Public dispatchers re-defined via the existing DEFINE_QPEL_DIAG_PUBLIC macro (replaces the CPU-only DEFINE_QPEL_DISPATCH instantiations). Verified on hertz: $ ./build/test_api_h264 \| grep "qpel avg" \| wc -l 15 $ ./build/test_api_h264 \| grep "qpel avg" \| grep -c "100.0000%" 15 All 15 PASS 2048/2048 bytes bit-exact via QPU. QPU coverage for the H.264 8-bit 4:2:0 hot-path pixel kernels: Layer Coverage ───────────────────────────────────────────────────────────── IDCT 4x4 luma ✓ cycle 6 (one QPU shader, also handles chroma) IDCT 8x8 luma ✓ cycle 7 Chroma DC Hadamard CPU only (4 adds + 4 subs; not worth) Deblock luma_v ✓ cycle 8 Deblock luma_h ✓ PR #28 Deblock chroma_v/h ✓ PR #29 Deblock *_intra CPU only (less common, structurally different) qpel put_ 15 pos ✓ cycle 9 (mc20) + PRs #30-#33 qpel avg_ 15 pos ✓ THIS PR The H.264 non-intra-deblock hot path is now FULLY on QPU for any consumer that initialises daedalus with a QPU-capable context.	2026-05-25 20:22:33 +02:00
marfrit	55d3618408	Merge pull request 'h264: V3D shaders for the 8 diagonal qpel positions' (#33 ) from noether/v3d-shader-h264-qpel-diagonals into main Reviewed-on: #33	2026-05-25 18:16:53 +00:00
claude-noether	746533582e	h264: V3D shaders for the 8 diagonal qpel positions Closes the put_ qpel QPU matrix. Adds mc11/12/13/21/23/31/32/33 — each composes two half-pel anchor outputs via L2 rounded-average: mc11 ¼¼ : avg(mc20[r, c], mc02[r, c]) mc12 ¼½ : avg(mc22[r, c], mc02[r, c]) mc13 ¼¾ : avg(mc20[r+1, c], mc02[r, c]) mc21 ½¼ : avg(mc22[r, c], mc20[r, c]) mc23 ½¾ : avg(mc22[r, c], mc20[r+1, c]) mc31 ¾¼ : avg(mc20[r, c], mc02[r, c+1]) mc32 ¾½ : avg(mc22[r, c], mc02[r, c+1]) mc33 ¾¾ : avg(mc20[r+1, c], mc02[r, c+1]) Per-lane structure: each lane runs the FULL cascade for BOTH anchors at its own (r, c) target, then L2 averages. No shared memory. Shaders inline hpel_h() / hpel_v() / hpel_hv() helpers (the latter does the 13×6 int16 cascade per cell). ~88 lines each. Shaders generated from a python template (POSITIONS table + format string) — the 8 .comp files are 1:1 with the C reference's DEFINE_DIAG_REF macro from fourier PR #18. Dispatch plumbing: shared dispatch_h264_qpel_diag_qpu helper covers all 8 (same src envelope as mc22: src_max = src_off + 10*stride + 11, covering rows -2..+10 and cols -2..+10 for any (r±1, c±1) offset). Recipe table: all 8 DAEDALUS_KERNEL_H264_QPEL_MC{11..33} flipped to QPU. Public dispatchers re-defined via DEFINE_QPEL_DIAG_PUBLIC macro (replaces the old DEFINE_QPEL_DISPATCH which fast-failed QPU). Verified on hertz: $ ./build/test_api_h264 \| grep "qpel mc[1-3][1-3]" H.264 qpel mc11: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc12: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc13: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc21: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc23: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc31: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc32: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc33: 2048/2048 bytes bit-exact (100.0000%) Meaningful: the (r±1, c±1) offsets are easy to transpose between positions; passing first try on the asymmetric variants (mc13/23/31/33) means the position-specific shifts are correct in all 8 templates. put_ qpel QPU matrix is now COMPLETE: 15 of 15 useful positions (mc00 = integer copy, no shader needed). avg_ qpel positions (15 more) remain on CPU NEON; can land as a follow-up since avg_ is just put_ + one extra L2 against existing dst. put_ mc20 ✓ mc02 ✓ mc22 ✓ (anchors) mc10 ✓ mc30 ✓ mc01 ✓ mc03 ✓ (single-axis ¼-pel) mc11 ✓ mc12 ✓ mc13 ✓ (this PR — row-1 diagonals) mc21 ✓ mc23 ✓ (this PR — row-2 diagonals) mc31 ✓ mc32 ✓ mc33 ✓ (this PR — row-3 diagonals) avg_ all 15 — CPU NEON	2026-05-25 19:14:42 +02:00
marfrit	224f4be9e2	Merge pull request 'h264: V3D shaders for the 4 single-axis quarter-pel qpel variants' (#32 ) from noether/v3d-shader-h264-qpel-quarter-axis into main Reviewed-on: #32	2026-05-25 17:09:00 +00:00
claude-noether	e3c28495ae	h264: V3D shaders for the 4 single-axis quarter-pel qpel variants mc10 (¼-H), mc30 (¾-H), mc01 (¼-V), mc03 (¾-V). Each is the corresponding half-pel filter (mc20 or mc02) with one extra L2 rounded-average step against an integer-source pixel at the tail: mc10[r,c] = avg(clip255(mc20(s)), s[r, c ]) mc30[r,c] = avg(clip255(mc20(s)), s[r, c+1]) mc01[r,c] = avg(clip255(mc02(s)), s[r, c ]) mc03[r,c] = avg(clip255(mc02(s)), s[r+1, c ]) Each shader is ~45 lines (mc20-/mc02-pattern + 1 L2 line). CMake foreach loop generates the 4 SPV compile rules. Dispatch helper `dispatch_h264_qpel_axis_qpu` shares plumbing across all 4 (axis flag selects src_max bounds: H reads cols -2..+10, V reads rows -2..+10). DEFINE_QPEL_AXIS_QPU + DEFINE_QPEL_DISPATCH_QPU macros collapse ~200 LOC of boilerplate. Recipe table flips DAEDALUS_KERNEL_H264_QPEL_MC{10,30,01,03} from CPU to QPU. Verified on hertz: $ ./build/test_api_h264 \| grep "qpel mc[01230]" H.264 qpel mc10: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc30: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc01: 2048/2048 bytes bit-exact (100.0000%) H.264 qpel mc03: 2048/2048 bytes bit-exact (100.0000%) (+ mc20/mc02/mc22 anchors from previous PRs) Qpel QPU coverage: put_ mc20 ✓ mc02 ✓ mc22 ✓ (3 anchors) mc10 ✓ mc30 ✓ mc01 ✓ mc03 ✓ (4 quarter-axis, THIS PR) mc11/12/13/21/23/31/32/33 — CPU NEON (8 diagonals) avg_ all 15 positions — CPU NEON 7 of 15 useful put_ positions now on QPU. The 8 diagonals each compose two half-pel results via L2; can land via dedicated kernels or by chaining existing anchor dispatches (the latter would need the L2 step as a fourth dispatch — probably cheaper to write dedicated 8x diagonal shaders).	2026-05-25 19:04:26 +02:00