diff --git a/src/asahi/genxml/cmdbuf.xml b/src/asahi/genxml/cmdbuf.xml
index 7e07d54e132..387420a1034 100644
--- a/src/asahi/genxml/cmdbuf.xml
+++ b/src/asahi/genxml/cmdbuf.xml
@@ -1010,14 +1010,17 @@
     <value name="Indirect local" value="2"/>
   </enum>
 
-  <struct name="CDM Launch" size="8">
+  <struct name="CDM Launch Word 0" size="4">
     <field name="Uniform register count" size="3" start="1" type="uint" modifier="groups(64)"/>
     <field name="Texture state register count" size="5" start="4" type="uint" modifier="groups(8)"/>
     <field name="Sampler state register count" size="3" start="9" type="Sampler states"/>
     <field name="Preshader register count" size="4" start="12" type="uint" modifier="groups(16)"/>
     <field name="Mode" size="2" start="27" type="CDM Mode"/>
     <field name="Block Type" size="3" start="29" type="CDM Block Type" default="Launch"/>
-    <field name="Pipeline" size="26" start="1:6" type="address" modifier="shr(6)"/>
+  </struct>
+
+  <struct name="CDM Launch Word 1" size="4">
+    <field name="Pipeline" size="26" start="6" type="address" modifier="shr(6)"/>
   </struct>
 
   <struct name="CDM Unk G14X" size="8">
diff --git a/src/asahi/lib/decode.c b/src/asahi/lib/decode.c
index 1689f2e959c..35055450d67 100644
--- a/src/asahi/lib/decode.c
+++ b/src/asahi/lib/decode.c
@@ -665,7 +665,8 @@ agxdecode_cdm(struct agxdecode_ctx *ctx, const uint8_t *map, uint64_t *link,
 
    switch (block_type) {
    case AGX_CDM_BLOCK_TYPE_LAUNCH: {
-      size_t length = AGX_CDM_LAUNCH_LENGTH;
+      size_t length =
+         AGX_CDM_LAUNCH_WORD_0_LENGTH + AGX_CDM_LAUNCH_WORD_1_LENGTH;
 
 #define CDM_PRINT(STRUCT_NAME, human)                                          \
    do {                                                                        \
@@ -674,17 +675,20 @@ agxdecode_cdm(struct agxdecode_ctx *ctx, const uint8_t *map, uint64_t *link,
       length += AGX_CDM_##STRUCT_NAME##_LENGTH;                                \
    } while (0);
 
-      agx_unpack(agxdecode_dump_stream, map, CDM_LAUNCH, hdr);
-      agxdecode_stateful(ctx, hdr.pipeline, "Pipeline", agxdecode_usc, verbose,
-                         params, &hdr.sampler_state_register_count);
-      DUMP_UNPACKED(CDM_LAUNCH, hdr, "Compute\n");
-      map += AGX_CDM_LAUNCH_LENGTH;
+      agx_unpack(agxdecode_dump_stream, map + 0, CDM_LAUNCH_WORD_0, hdr0);
+      agx_unpack(agxdecode_dump_stream, map + 4, CDM_LAUNCH_WORD_1, hdr1);
+
+      agxdecode_stateful(ctx, hdr1.pipeline, "Pipeline", agxdecode_usc, verbose,
+                         params, &hdr0.sampler_state_register_count);
+      DUMP_UNPACKED(CDM_LAUNCH_WORD_0, hdr0, "Compute\n");
+      DUMP_UNPACKED(CDM_LAUNCH_WORD_1, hdr1, "Compute\n");
+      map += 8;
 
       /* Added in G14X */
       if (params->gpu_generation >= 14 && params->num_clusters_total > 1)
          CDM_PRINT(UNK_G14X, "Unknown G14X");
 
-      switch (hdr.mode) {
+      switch (hdr0.mode) {
       case AGX_CDM_MODE_DIRECT:
          CDM_PRINT(GLOBAL_SIZE, "Global size");
          CDM_PRINT(LOCAL_SIZE, "Local size");
@@ -697,7 +701,7 @@ agxdecode_cdm(struct agxdecode_ctx *ctx, const uint8_t *map, uint64_t *link,
          CDM_PRINT(INDIRECT, "Indirect buffer");
          break;
       default:
-         fprintf(agxdecode_dump_stream, "Unknown CDM mode: %u\n", hdr.mode);
+         fprintf(agxdecode_dump_stream, "Unknown CDM mode: %u\n", hdr0.mode);
          break;
       }
 
diff --git a/src/gallium/drivers/asahi/agx_state.c b/src/gallium/drivers/asahi/agx_state.c
index 494cc2a7532..d41301a1410 100644
--- a/src/gallium/drivers/asahi/agx_state.c
+++ b/src/gallium/drivers/asahi/agx_state.c
@@ -4037,9 +4037,10 @@ agx_launch_gs_prerast(struct agx_batch *batch,
 
    agx_ensure_cmdbuf_has_space(
       batch, &batch->cdm,
-      8 * (AGX_CDM_LAUNCH_LENGTH + AGX_CDM_UNK_G14X_LENGTH +
-           AGX_CDM_INDIRECT_LENGTH + AGX_CDM_GLOBAL_SIZE_LENGTH +
-           AGX_CDM_LOCAL_SIZE_LENGTH + AGX_CDM_BARRIER_LENGTH));
+      8 * (AGX_CDM_LAUNCH_WORD_0_LENGTH + AGX_CDM_LAUNCH_WORD_1_LENGTH +
+           AGX_CDM_UNK_G14X_LENGTH + AGX_CDM_INDIRECT_LENGTH +
+           AGX_CDM_GLOBAL_SIZE_LENGTH + AGX_CDM_LOCAL_SIZE_LENGTH +
+           AGX_CDM_BARRIER_LENGTH));
 
    assert(!info->primitive_restart && "should have been lowered");
 
@@ -5216,7 +5217,7 @@ agx_launch(struct agx_batch *batch, const struct pipe_grid_info *info,
    /* TODO: Ensure space if we allow multiple kernels in a batch */
    uint8_t *out = batch->cdm.current;
 
-   agx_push(out, CDM_LAUNCH, cfg) {
+   agx_push(out, CDM_LAUNCH_WORD_0, cfg) {
       if (info->indirect)
          cfg.mode = AGX_CDM_MODE_INDIRECT_GLOBAL;
       else
@@ -5227,6 +5228,9 @@ agx_launch(struct agx_batch *batch, const struct pipe_grid_info *info,
       cfg.texture_state_register_count = agx_nr_tex_descriptors(batch, cs);
       cfg.sampler_state_register_count =
          translate_sampler_state_count(ctx, cs, stage);
+   }
+
+   agx_push(out, CDM_LAUNCH_WORD_1, cfg) {
       cfg.pipeline =
          agx_build_pipeline(batch, cs, linked, PIPE_SHADER_COMPUTE,
                             info->variable_shared_mem, subgroups_per_core);
@@ -5367,9 +5371,10 @@ agx_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
     * in practice, we can use CDM stream links.
     */
    size_t dispatch_upper_bound =
-      AGX_CDM_LAUNCH_LENGTH + AGX_CDM_UNK_G14X_LENGTH +
-      AGX_CDM_INDIRECT_LENGTH + AGX_CDM_GLOBAL_SIZE_LENGTH +
-      AGX_CDM_LOCAL_SIZE_LENGTH + AGX_CDM_BARRIER_LENGTH;
+      AGX_CDM_LAUNCH_WORD_0_LENGTH + AGX_CDM_LAUNCH_WORD_1_LENGTH +
+      AGX_CDM_UNK_G14X_LENGTH + AGX_CDM_INDIRECT_LENGTH +
+      AGX_CDM_GLOBAL_SIZE_LENGTH + AGX_CDM_LOCAL_SIZE_LENGTH +
+      AGX_CDM_BARRIER_LENGTH;
 
    if (batch->cdm.current + dispatch_upper_bound >= batch->cdm.end)
       agx_flush_batch_for_reason(ctx, batch, "CDM overfull");