diff --git a/src/asahi/lib/agx_device.c b/src/asahi/lib/agx_device.c
index 0e1f356680c..3e75efe519a 100644
--- a/src/asahi/lib/agx_device.c
+++ b/src/asahi/lib/agx_device.c
@@ -75,6 +75,7 @@ static const struct debug_named_value agx_debug_options[] = {
    {"scratch",   AGX_DBG_SCRATCH,  "Debug scratch memory usage"},
    {"1queue",    AGX_DBG_1QUEUE,   "Force usage of a single queue for multiple contexts"},
    {"nosoft",    AGX_DBG_NOSOFT,   "Disable soft fault optimizations"},
+   {"nomerge",   AGX_DBG_NOMERGE,  "Disable control stream merging"},
    {"bodumpverbose", AGX_DBG_BODUMPVERBOSE,   "Include extra info with dumps"},
    DEBUG_NAMED_VALUE_END
 };
diff --git a/src/asahi/lib/agx_device.h b/src/asahi/lib/agx_device.h
index 17c900b67aa..52620660862 100644
--- a/src/asahi/lib/agx_device.h
+++ b/src/asahi/lib/agx_device.h
@@ -47,6 +47,7 @@ enum agx_dbg {
    AGX_DBG_NOSOFT = BITFIELD_BIT(19),
    AGX_DBG_FEEDBACK = BITFIELD_BIT(20),
    AGX_DBG_1QUEUE = BITFIELD_BIT(21),
+   AGX_DBG_NOMERGE = BITFIELD_BIT(22),
 };
 
 /* How many power-of-two levels in the BO cache do we want? 2^14 minimum chosen
diff --git a/src/asahi/vulkan/hk_cmd_buffer.c b/src/asahi/vulkan/hk_cmd_buffer.c
index e355502784a..92fab5452f3 100644
--- a/src/asahi/vulkan/hk_cmd_buffer.c
+++ b/src/asahi/vulkan/hk_cmd_buffer.c
@@ -259,6 +259,32 @@ hk_BeginCommandBuffer(VkCommandBuffer commandBuffer,
    return VK_SUCCESS;
 }
 
+/*
+ * Merge adjacent compute control streams. Except for reading timestamps, there
+ * is no reason to submit two CDM streams back-to-back in the same command
+ * buffer. However, it is challenging to avoid constructing such sequences due
+ * to the gymnastics required to reorder compute around graphics. Merging at
+ * EndCommandBuffer is cheap O(# of control streams) and lets us get away with
+ * the sloppiness.
+ */
+static void
+merge_control_streams(struct hk_cmd_buffer *cmd)
+{
+   struct hk_cs *last = NULL;
+
+   list_for_each_entry_safe(struct hk_cs, cs, &cmd->control_streams, node) {
+      if (cs->type == HK_CS_CDM && last && last->type == HK_CS_CDM &&
+          !last->timestamp.end.handle) {
+
+         hk_cs_merge_cdm(last, cs);
+         list_del(&cs->node);
+         hk_cs_destroy(cs);
+      } else {
+         last = cs;
+      }
+   }
+}
+
 VKAPI_ATTR VkResult VKAPI_CALL
 hk_EndCommandBuffer(VkCommandBuffer commandBuffer)
 {
@@ -271,6 +297,21 @@ hk_EndCommandBuffer(VkCommandBuffer commandBuffer)
    hk_cmd_buffer_end_compute(cmd);
    hk_cmd_buffer_end_compute_internal(cmd, &cmd->current_cs.post_gfx);
 
+   struct hk_device *dev = hk_cmd_buffer_device(cmd);
+   if (likely(!(dev->dev.debug & AGX_DBG_NOMERGE))) {
+      merge_control_streams(cmd);
+   }
+
+   /* We cannot terminate CDM control streams until after merging, since merging
+    * needs to append stream links late. Now that we've merged, insert all the
+    * missing stream terminates.
+    */
+   list_for_each_entry(struct hk_cs, cs, &cmd->control_streams, node) {
+      if (cs->type == HK_CS_CDM) {
+         cs->current = agx_cdm_terminate(cs->current);
+      }
+   }
+
    return vk_command_buffer_get_record_result(&cmd->vk);
 }
 
diff --git a/src/asahi/vulkan/hk_cmd_buffer.h b/src/asahi/vulkan/hk_cmd_buffer.h
index ef3bcc7fcd8..a1050819f8c 100644
--- a/src/asahi/vulkan/hk_cmd_buffer.h
+++ b/src/asahi/vulkan/hk_cmd_buffer.h
@@ -370,7 +370,7 @@ struct hk_cs {
 
    /* Statistics */
    struct {
-      uint32_t calls, cmds, flushes;
+      uint32_t calls, cmds, flushes, merged;
    } stats;
 
    /* Timestamp writes. Currently just compute end / fragment end. We could
@@ -406,6 +406,32 @@ struct hk_cs {
    uint32_t restart_index;
 };
 
+/*
+ * Helper to merge two compute control streams, concatenating the second control
+ * stream to the first one. Must sync with hk_cs.
+ */
+static inline void
+hk_cs_merge_cdm(struct hk_cs *a, const struct hk_cs *b)
+{
+   assert(a->type == HK_CS_CDM && b->type == HK_CS_CDM);
+   assert(a->cmd == b->cmd);
+   assert(!a->timestamp.end.handle);
+
+   agx_cdm_jump(a->current, b->addr);
+   a->current = b->current;
+   a->stream_linked = true;
+
+   a->scratch.cs.main |= b->scratch.cs.main;
+   a->scratch.cs.preamble |= b->scratch.cs.preamble;
+
+   a->timestamp = b->timestamp;
+
+   a->stats.calls += b->stats.calls;
+   a->stats.cmds += b->stats.cmds;
+   a->stats.flushes += b->stats.flushes;
+   a->stats.merged++;
+}
+
 static inline uint64_t
 hk_cs_current_addr(struct hk_cs *cs)
 {
@@ -660,8 +686,6 @@ hk_cmd_buffer_end_compute_internal(struct hk_cmd_buffer *cmd,
       if (cs->imm_writes.size) {
          hk_dispatch_imm_writes(cmd, cs);
       }
-
-      cs->current = agx_cdm_terminate(cs->current);
    }
 
    *ptr = NULL;
diff --git a/src/asahi/vulkan/hk_queue.c b/src/asahi/vulkan/hk_queue.c
index ab0a89b2c63..e1129294abe 100644
--- a/src/asahi/vulkan/hk_queue.c
+++ b/src/asahi/vulkan/hk_queue.c
@@ -838,8 +838,9 @@ queue_submit(struct hk_device *dev, struct hk_queue *queue,
          if (cs->type == HK_CS_CDM) {
             perf_debug(
                cmdbuf,
-               "%u: Submitting CDM with %u API calls, %u dispatches, %u flushes",
-               i, cs->stats.calls, cs->stats.cmds, cs->stats.flushes);
+               "%u: Submitting CDM with %u API calls, %u dispatches, %u flushes, %u merged",
+               i, cs->stats.calls, cs->stats.cmds, cs->stats.flushes,
+               cs->stats.merged);
 
             assert(cs->stats.cmds > 0 || cs->stats.flushes > 0 ||
                    cs->timestamp.end.handle);