freedreno/crashdec: Add support for CP_BV_MEMPOOL

The BV mempool is even further cut down compared to the "small mem pool"
layout which seems to be used by a610. It also shrinks the block size to
4 chunks instead of 8. This layout happens to be shared by a702, so
abstract out the layout into a "mempool size" enum.

While we're here, fix a bug with how the mempool offset for chunks is
printed. This accounts for the test diff.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36590>

src/freedreno/decode/crashdec-mempool.c

@@ -71,7 +71,7 @@ dump_mem_pool_chunk(const uint32_t *chunk)
 }
 
 void
-dump_cp_mem_pool(uint32_t *mempool)
+dump_cp_mem_pool(uint32_t *mempool, bool is_bv)
 {
    /* The mem pool is a shared pool of memory used for storing in-flight
     * register writes. There are 6 different queues, one for each
@@ -82,11 +82,12 @@ dump_cp_mem_pool(uint32_t *mempool)
     *
     * The queues are conceptually divided into 128-bit "chunks", and the
     * read and write pointers are in units of chunks.  These chunks are
-    * organized internally into 8-chunk "blocks", and memory is allocated
-    * dynamically in terms of blocks. Each queue is represented as a
-    * singly-linked list of blocks, as well as 3-bit start/end chunk
-    * pointers that point within the first/last block.  The next pointers
-    * are located in a separate array, rather than inline.
+    * organized internally into 8-chunk (or on certain cut-down parts,
+    * 4-chunk) "blocks", and memory is allocated dynamically in terms of
+    * blocks. Each queue is represented as a singly-linked list of blocks, as
+    * well as 3-bit start/end chunk pointers that point within the first/last
+    * block.  The next pointers are located in a separate array, rather than
+    * inline.
     */
 
    /* TODO: The firmware CP_MEM_POOL save/restore routines do something
@@ -111,17 +112,60 @@ dump_cp_mem_pool(uint32_t *mempool)
     */
    bool small_mem_pool = false;
 
-   /* The array of next pointers for each block. */
-   const uint32_t *next_pointers =
-      small_mem_pool ? &mempool[0x800] : &mempool[0x1000];
+   enum {
+      MEMPOOL_FULL,
+      MEMPOOL_HALF,
+      MEMPOOL_QUARTER,
+   } mempool_size = MEMPOOL_FULL;
 
+   if (small_mem_pool)
+      mempool_size = MEMPOOL_HALF;
+   else if (is_bv)
+      mempool_size = MEMPOOL_QUARTER;
+
+   uint32_t next_pointers_offset;
+   uint32_t data1_offset;
+   uint32_t data2_offset;
    /* Maximum number of blocks in the pool, also the size of the pointers
     * array.
     */
-   const int num_blocks = small_mem_pool ? 0x30 : 0x80;
+   int num_blocks;
+   /* log2 of number of chunks per block. */
+   int block_size_log2;
 
-   /* Number of queues */
-   const unsigned num_queues = is_a6xx() ? 6 : 7;
+   switch (mempool_size) {
+   case MEMPOOL_FULL:
+      next_pointers_offset = 0x1000;
+      data1_offset = 0x1800;
+      data2_offset = 0x2000;
+      num_blocks = 0x80;
+      block_size_log2 = 3;
+      break;
+   case MEMPOOL_HALF:
+      next_pointers_offset = 0x800;
+      data1_offset = 0xc00;
+      data2_offset = 0x1000;
+      num_blocks = 0x30;
+      block_size_log2 = 3;
+      break;
+   case MEMPOOL_QUARTER:
+      next_pointers_offset = 0x400;
+      data1_offset = 0x600;
+      data2_offset = 0x800;
+      num_blocks = 0x40;
+      block_size_log2 = 2;
+      break;
+   default:
+      UNREACHABLE("unknown mempool size");
+   }
+   const uint32_t *next_pointers = &mempool[next_pointers_offset];
+
+   const int block_size = 1 << block_size_log2;
+
+   /* Number of queues or clusters. On a7xx the clusters are shuffled a bit to
+    * increase the count by one, but BV seems to not use the last cluster.
+    */
+   const unsigned num_queues = (is_a6xx() || is_bv) ? 6 : 7;
 
    /* Unfortunately the per-queue state is a little more complicated than
     * a simple pair of begin/end pointers. Instead of a single beginning
@@ -182,11 +226,8 @@ dump_cp_mem_pool(uint32_t *mempool)
       uint32_t unk0;
       uint32_t padding0[7]; /* Mirrors of unk0 */
 
-      struct {
-         uint32_t chunk : 3;
-         uint32_t first_block : 32 - 3;
-      } writer[6];
-      uint32_t padding1[2]; /* Mirror of writer[5] */
+      uint32_t writer_first_block_chunk[7];
+      uint32_t padding1[1]; /* Mirror of writer_first_block_chunk[6] */
 
       uint32_t unk1;
       uint32_t padding2[7]; /* Mirrors of unk1 */
@@ -197,11 +238,8 @@ dump_cp_mem_pool(uint32_t *mempool)
       uint32_t unk2[7];
       uint32_t padding4[1];
 
-      struct {
-         uint32_t chunk : 3;
-         uint32_t first_block : 32 - 3;
-      } reader[7];
-      uint32_t padding5[1]; /* Mirror of reader[5] */
+      uint32_t reader_first_block_chunk[7];
+      uint32_t padding5[1]; /* Mirror of reader_first_block_chunk[6] */
 
       uint32_t unk3;
       uint32_t padding6[7]; /* Mirrors of unk3 */
@@ -216,15 +254,13 @@ dump_cp_mem_pool(uint32_t *mempool)
       uint32_t padding9[7]; /* Mirrors of unk4 */
    } data1;
 
-   const uint32_t *data1_ptr =
-      small_mem_pool ? &mempool[0xc00] : &mempool[0x1800];
+   const uint32_t *data1_ptr = &mempool[data1_offset];
    memcpy(&data1, data1_ptr, sizeof(data1));
 
    /* Based on the kernel, the first dword is the mem pool size (in
     * blocks?) and mirrors CP_MEM_POOL_DBG_SIZE.
     */
-   const uint32_t *data2_ptr =
-      small_mem_pool ? &mempool[0x1000] : &mempool[0x2000];
+   const uint32_t *data2_ptr = &mempool[data2_offset];
    const int data2_size = 0x60;
 
    /* This seems to be the size of each queue in chunks. */
@@ -249,28 +285,34 @@ dump_cp_mem_pool(uint32_t *mempool)
       printf("\tCLUSTER_%s:\n\n",
              is_a6xx() ? cluster_names_a6xx[queue] : cluster_names_a7xx[queue]);
 
+      uint32_t writer_first_block =
+         data1.writer_first_block_chunk[queue] >> block_size_log2;
+      uint32_t writer_chunk =
+         data1.writer_first_block_chunk[queue] & (block_size - 1);
+      uint32_t reader_first_block =
+         data1.reader_first_block_chunk[queue] >> block_size_log2;
+      uint32_t reader_chunk =
+         data1.reader_first_block_chunk[queue] & (block_size - 1);
       if (verbose) {
-         printf("\t\twriter_first_block: 0x%x\n",
-                data1.writer[queue].first_block);
+         printf("\t\twriter_first_block: 0x%x\n", writer_first_block);
          printf("\t\twriter_second_block: 0x%x\n",
                 data1.writer_second_block[queue]);
-         printf("\t\twriter_chunk: %d\n", data1.writer[queue].chunk);
-         printf("\t\treader_first_block: 0x%x\n",
-                data1.reader[queue].first_block);
+         printf("\t\twriter_chunk: %d\n", writer_chunk);
+         printf("\t\treader_first_block: 0x%x\n", reader_first_block);
          printf("\t\treader_second_block: 0x%x\n",
                 data1.reader_second_block[queue]);
-         printf("\t\treader_chunk: %d\n", data1.reader[queue].chunk);
+         printf("\t\treader_chunk: %d\n", reader_chunk);
          printf("\t\tblock_count: %d\n", data1.block_count[queue]);
          printf("\t\tunk2: 0x%x\n", data1.unk2[queue]);
          printf("\t\tqueue_size: %d\n\n", queue_sizes[queue]);
       }
 
-      uint32_t cur_chunk = data1.reader[queue].chunk;
-      uint32_t cur_block = cur_chunk > 3 ? data1.reader[queue].first_block
-                                         : data1.reader_second_block[queue];
-      uint32_t last_chunk = data1.writer[queue].chunk;
-      uint32_t last_block = last_chunk > 3 ? data1.writer[queue].first_block
-                                           : data1.writer_second_block[queue];
+      uint32_t cur_chunk = reader_chunk;
+      uint32_t cur_block = cur_chunk >= block_size / 2 ?
+         reader_first_block : data1.reader_second_block[queue];
+      uint32_t last_chunk = writer_chunk;
+      uint32_t last_block = last_chunk >= block_size / 2 ?
+         writer_first_block : data1.writer_second_block[queue];
 
       if (verbose)
          printf("\tblock %x\n", cur_block);
@@ -281,16 +323,16 @@ dump_cp_mem_pool(uint32_t *mempool)
       unsigned calculated_queue_size = 0;
       while (cur_block != last_block || cur_chunk != last_chunk) {
          calculated_queue_size++;
-         uint32_t *chunk_ptr = &mempool[cur_block * 0x20 + cur_chunk * 4];
+         uint32_t *chunk_ptr = &mempool[(cur_block * block_size + cur_chunk) * 4];
 
          dump_mem_pool_chunk(chunk_ptr);
 
          printf("\t%05x: %08x %08x %08x %08x\n",
-                4 * (cur_block * 0x20 + cur_chunk + 4), chunk_ptr[0],
+                16 * (cur_block * block_size + cur_chunk), chunk_ptr[0],
                 chunk_ptr[1], chunk_ptr[2], chunk_ptr[3]);
 
          cur_chunk++;
-         if (cur_chunk == 8) {
+         if (cur_chunk == block_size) {
             cur_block = next_pointers[cur_block];
             if (verbose)
                printf("\tblock %x\n", cur_block);

src/freedreno/decode/crashdec.c

@@ -900,7 +900,9 @@ decode_indexed_registers(void)
             dump_cp_ucode_dbg(buf);
 
          if (!strcmp(name, "CP_MEM_POOL_DBG"))
-            dump_cp_mem_pool(buf);
+            dump_cp_mem_pool(buf, false);
+         if (!strcmp(name, "CP_BV_MEM_POOL_DBG"))
+            dump_cp_mem_pool(buf, true);
 
          if (dump)
             dump_hex_ascii(buf, 4 * sizedwords, 1);

src/freedreno/decode/crashdec.h

@@ -85,7 +85,7 @@ is_gmu_legacy(void)
 }
 
 void dump_register(struct regacc *r);
-void dump_cp_mem_pool(uint32_t *mempool);
+void dump_cp_mem_pool(uint32_t *mempool, bool is_bv);
 void handle_prefetch(uint32_t *dwords, uint32_t sizedwords);
 
 struct a6xx_hfi_state {

src/freedreno/tests/reference/crash_prefetch.log

@@ -14655,346 +14655,346 @@ indexed-registers:
 	000170: 00000026 00000026 00000026 00000026	|&...&...&...&...|
 	CLUSTER_FE:
 
-	01fa0: 00000000 00000000 00000000 00000000
+	01fc0: 00000000 00000000 00000000 00000000
 		write PC_CNTL (09b00) context 1
 			PC_CNTL: { 0 }
-	01fa4: 00000002 00000000 b0000000 00000209
+	01fd0: 00000002 00000000 b0000000 00000209
 
 	CLUSTER_SP_VS:
 
-	02310: 00000000 00000000 00000000 00000000
+	02300: 00000000 00000000 00000000 00000000
 
 	CLUSTER_PC_VS:
 
-	01f18: 00000000 00000000 00000000 00000000
+	01f20: 00000000 00000000 00000000 00000000
 		write PC_CNTL (09b00) context 1
 			PC_CNTL: { 0 }
-	01f1c: 00000002 00000000 b0000000 00000209
+	01f30: 00000002 00000000 b0000000 00000209
 
 	CLUSTER_GRAS:
 
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x22 }
-	01c1c: 00000002 00000088 60100000 0000000d
+	01c30: 00000002 00000088 60100000 0000000d
 		write CP_EVENT_START (0d600) context 1
 			CP_EVENT_START: { STATE_ID = 0x23 }
-	01c20: 00000002 0000008c 60000000 0000020d
+	01c40: 00000002 0000008c 60000000 0000020d
 		write CP_EVENT_END (0d601) context 1
 			CP_EVENT_END: { STATE_ID = 0x23 }
-	01c24: 00000002 0000008c 60100000 0000020d
+	01c50: 00000002 0000008c 60100000 0000020d
 		write CP_EVENT_START (0d600) context 0
 			CP_EVENT_START: { STATE_ID = 0x24 }
-	01c28: 00000002 00000090 60000000 0000000d
+	01c60: 00000002 00000090 60000000 0000000d
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x24 }
-	01c2c: 00000002 00000090 60100000 0000000d
+	01c70: 00000002 00000090 60100000 0000000d
 		write CP_EVENT_START (0d600) context 1
 			CP_EVENT_START: { STATE_ID = 0x25 }
-	02b90: 00000002 00000094 60000000 0000020d
+	02b80: 00000002 00000094 60000000 0000020d
 		write CP_EVENT_END (0d601) context 1
 			CP_EVENT_END: { STATE_ID = 0x25 }
-	02b94: 00000002 00000094 60100000 0000020d
+	02b90: 00000002 00000094 60100000 0000020d
 		write CP_EVENT_START (0d600) context 0
 			CP_EVENT_START: { STATE_ID = 0x26 }
-	02b98: 00000002 00000098 60000000 0000000d
+	02ba0: 00000002 00000098 60000000 0000000d
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x26 }
-	02b9c: 00000002 00000098 60100000 0000000d
-	02ba0: 00000000 00000000 00000000 00000000
-	02ba4: 00000000 00000000 00000000 00000000
+	02bb0: 00000002 00000098 60100000 0000000d
+	02bc0: 00000000 00000000 00000000 00000000
+	02bd0: 00000000 00000000 00000000 00000000
 
 	CLUSTER_SP_PS:
 
 		write CP_EVENT_END (0d601) context 1
 			CP_EVENT_END: { STATE_ID = 0x21 }
-	01a9c: 00000002 00000084 60100000 0000020d
+	01ab0: 00000002 00000084 60100000 0000020d
 		write SP_PS_MRT[0].REG (0a996) context 0
 			SP_PS_MRT[0].REG: { COLOR_FORMAT = FMT6_16_16_16_16_UNORM }
 		write SP_PS_MRT[0x1].REG (0a997) context 0
 			SP_PS_MRT[0x1].REG: { COLOR_FORMAT = FMT6_5_6_5_UNORM }
-	01aa0: 00000183 00000038 9972a658 0000000a
+	01ac0: 00000183 00000038 9972a658 0000000a
 		write SP_PS_MRT[0x2].REG (0a998) context 0
 			SP_PS_MRT[0x2].REG: { COLOR_FORMAT = FMT6_16_16_UNORM }
 		write SP_PS_MRT[0x3].REG (0a999) context 0
 			SP_PS_MRT[0x3].REG: { COLOR_FORMAT = FMT6_16_16_16_16_SNORM }
-	01aa4: 0000010f 00000184 9992a660 0000000a
+	01ad0: 0000010f 00000184 9992a660 0000000a
 		write SP_SRGB_CNTL (0a98a) context 0
 			SP_SRGB_CNTL: { 0 }
-	01aa8: 00000002 00000000 98a00000 0000000a
+	01ae0: 00000002 00000000 98a00000 0000000a
 		write CP_EVENT_START (0d600) context 0
 			CP_EVENT_START: { STATE_ID = 0x22 }
-	01aac: 00000002 00000088 60000000 0000000d
+	01af0: 00000002 00000088 60000000 0000000d
 		write SP_EVENT_INITIATOR (0bb02) context 0
 			SP_EVENT_INITIATOR: { STATE_ID = 0x22 | EVENT = PC_CCU_FLUSH_COLOR_TS }
 		write SP_EVENT_INITIATOR (0bb02) context 0
 			SP_EVENT_INITIATOR: { STATE_ID = 0x22 | EVENT = CONTEXT_DONE }
-	02790: 00880077 00880014 b022ec08 0000000b
+	02780: 00880077 00880014 b022ec08 0000000b
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x22 }
-	02794: 00000002 00000088 60100000 0000000d
+	02790: 00000002 00000088 60100000 0000000d
 		write CP_EVENT_START (0d600) context 1
 			CP_EVENT_START: { STATE_ID = 0x23 }
-	02798: 00000002 0000008c 60000000 0000020d
+	027a0: 00000002 0000008c 60000000 0000020d
 		write SP_EVENT_INITIATOR (0bb02) context 1
 			SP_EVENT_INITIATOR: { STATE_ID = 0x23 | EVENT = PC_CCU_FLUSH_DEPTH_TS }
 		write SP_EVENT_INITIATOR (0bb02) context 1
 			SP_EVENT_INITIATOR: { STATE_ID = 0x23 | EVENT = CONTEXT_DONE }
-	0279c: 008c0073 008c0014 b022ec08 0000030b
+	027b0: 008c0073 008c0014 b022ec08 0000030b
 		write CP_EVENT_END (0d601) context 1
 			CP_EVENT_END: { STATE_ID = 0x23 }
-	027a0: 00000002 0000008c 60100000 0000020d
+	027c0: 00000002 0000008c 60100000 0000020d
 		write CP_EVENT_START (0d600) context 0
 			CP_EVENT_START: { STATE_ID = 0x24 }
-	027a4: 00000002 00000090 60000000 0000000d
+	027d0: 00000002 00000090 60000000 0000000d
 		write SP_EVENT_INITIATOR (0bb02) context 0
 			SP_EVENT_INITIATOR: { STATE_ID = 0x24 | EVENT = PC_CCU_INVALIDATE_COLOR }
 		write SP_EVENT_INITIATOR (0bb02) context 0
 			SP_EVENT_INITIATOR: { STATE_ID = 0x24 | EVENT = CONTEXT_DONE }
-	027a8: 00900067 00900014 b022ec08 0000000b
+	027e0: 00900067 00900014 b022ec08 0000000b
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x24 }
-	027ac: 00000002 00000090 60100000 0000000d
+	027f0: 00000002 00000090 60100000 0000000d
 		write CP_EVENT_START (0d600) context 1
 			CP_EVENT_START: { STATE_ID = 0x25 }
-	02d10: 00000002 00000094 60000000 0000020d
+	02d00: 00000002 00000094 60000000 0000020d
 		write SP_EVENT_INITIATOR (0bb02) context 1
 			SP_EVENT_INITIATOR: { STATE_ID = 0x25 | EVENT = PC_CCU_INVALIDATE_DEPTH }
 		write SP_EVENT_INITIATOR (0bb02) context 1
 			SP_EVENT_INITIATOR: { STATE_ID = 0x25 | EVENT = CONTEXT_DONE }
-	02d14: 00940063 00940014 b022ec08 0000030b
+	02d10: 00940063 00940014 b022ec08 0000030b
 		write CP_EVENT_END (0d601) context 1
 			CP_EVENT_END: { STATE_ID = 0x25 }
-	02d18: 00000002 00000094 60100000 0000020d
+	02d20: 00000002 00000094 60100000 0000020d
 		write CP_EVENT_START (0d600) context 0
 			CP_EVENT_START: { STATE_ID = 0x26 }
-	02d1c: 00000002 00000098 60000000 0000000d
+	02d30: 00000002 00000098 60000000 0000000d
 		write SP_EVENT_INITIATOR (0bb02) context 0
 			SP_EVENT_INITIATOR: { STATE_ID = 0x26 | EVENT = CACHE_INVALIDATE }
 		write SP_EVENT_INITIATOR (0bb02) context 0
 			SP_EVENT_INITIATOR: { STATE_ID = 0x26 | EVENT = CONTEXT_DONE }
-	02d20: 009800c7 00980014 b022ec08 0000000b
+	02d40: 009800c7 00980014 b022ec08 0000000b
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x26 }
-	02d24: 00000002 00000098 60100000 0000000d
-	02d28: 00000000 00000000 00000000 00000000
-	02d2c: 00000000 00000000 00000000 00000000
+	02d50: 00000002 00000098 60100000 0000000d
+	02d60: 00000000 00000000 00000000 00000000
+	02d70: 00000000 00000000 00000000 00000000
 
 	CLUSTER_PS:
 
 		write CP_2D_EVENT_END (0d701) context 1
 			CP_2D_EVENT_END: { STATE_ID = 0x9 }
-	01a24: 00000002 00000024 70100000 0000020d
+	01a50: 00000002 00000024 70100000 0000020d
 		write EVENT_CMD (e7) pipe
 			EVENT_CMD: { EVENT_TYPE = EVENT | TS_WRITE }
 		write EVENT_TS_ADDR (e8) pipe
 			EVENT_TS_ADDR: 0x1000
-	01a28: 00000017 00004000 0e80039c 000000c0
+	01a60: 00000017 00004000 0e80039c 000000c0
 		write EVENT_TS_ADDR+0x1 (e9) pipe
 			EVENT_TS_ADDR+0x1: 0x1
 		write EVENT_TS_CTRL (ea) pipe
 			EVENT_TS_CTRL: { EVENT = CCU_FLUSH_COLOR }
-	01a2c: 00000007 00001400 0ea003a4 000000c0
+	01a70: 00000007 00001400 0ea003a4 000000c0
 		write EVENT_TS_DATA (eb) pipe
 			EVENT_TS_DATA: 0
-	01990: 00000002 00000000 0eb00000 00000080
+	01980: 00000002 00000000 0eb00000 00000080
 		write CP_EVENT_START (0d600) context 0
 			CP_EVENT_START: { STATE_ID = 0x20 }
-	01994: 00000002 00000080 60000000 0000000d
+	01990: 00000002 00000080 60000000 0000000d
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x20 }
-	01998: 00000002 00000080 60100000 0000000d
+	019a0: 00000002 00000080 60100000 0000000d
 		write EVENT_CMD (e7) pipe
 			EVENT_CMD: { EVENT_TYPE = EVENT }
-	0199c: 00000002 00000004 0e700000 00000080
+	019b0: 00000002 00000004 0e700000 00000080
 		write CP_EVENT_START (0d600) context 1
 			CP_EVENT_START: { STATE_ID = 0x21 }
-	019a0: 00000002 00000084 60000000 0000020d
+	019c0: 00000002 00000084 60000000 0000020d
 		write CP_EVENT_END (0d601) context 1
 			CP_EVENT_END: { STATE_ID = 0x21 }
-	019a4: 00000002 00000084 60100000 0000020d
+	019d0: 00000002 00000084 60100000 0000020d
 		write RB_DEPTH_BUFFER_INFO (08872) context 0
 			RB_DEPTH_BUFFER_INFO: { DEPTH_FORMAT = DEPTH6_NONE }
 		write RB_DEPTH_BUFFER_PITCH (08873) context 0
 			RB_DEPTH_BUFFER_PITCH: 0
-	019a8: 00000003 00000000 873221c8 00000008
+	019e0: 00000003 00000000 873221c8 00000008
 		write RB_DEPTH_BUFFER_ARRAY_PITCH (08874) context 0
 			RB_DEPTH_BUFFER_ARRAY_PITCH: 0
 		write RB_DEPTH_BUFFER_BASE (08875) context 0
 			RB_DEPTH_BUFFER_BASE: 0
-	019ac: 00000003 00000000 875221d0 00000008
+	019f0: 00000003 00000000 875221d0 00000008
 		write RB_DEPTH_BUFFER_BASE_HI (08876) context 0
 			RB_DEPTH_BUFFER_BASE_HI: 0
 		write RB_DEPTH_GMEM_BASE (08877) context 0
 			RB_DEPTH_GMEM_BASE: 0
-	02110: 00000003 00000000 877221d8 00000008
+	02100: 00000003 00000000 877221d8 00000008
 		write RB_STENCIL_BUFFER_INFO (08881) context 0
 			RB_STENCIL_BUFFER_INFO: { 0 }
 		write RB_MRT[0].BUF_INFO (08822) context 0
 			RB_MRT[0].BUF_INFO: { COLOR_FORMAT = FMT6_16_16_16_16_UNORM | COLOR_TILE_MODE = TILE6_3 | COLOR_SWAP = WZYX }
-	02114: 00000003 00000d80 82222204 00000008
+	02110: 00000003 00000d80 82222204 00000008
 		write RB_MRT[0].PITCH (08823) context 0
 			RB_MRT[0].PITCH: 512
 		write RB_MRT[0].ARRAY_PITCH (08824) context 0
 			RB_MRT[0].ARRAY_PITCH: 40960
-	02118: 00000023 00000a00 8242208c 00000008
+	02120: 00000023 00000a00 8242208c 00000008
 		write RB_MRT[0].BASE (08825) context 0
 			RB_MRT[0].BASE: 0x9c000
 		write RB_MRT[0].BASE_HI (08826) context 0
 			RB_MRT[0].BASE_HI: 0x1
-	0211c: 00270003 00000004 82622094 00000008
+	02130: 00270003 00000004 82622094 00000008
 		write RB_MRT[0].BASE_GMEM (08827) context 0
 			RB_MRT[0].BASE_GMEM: 0x2e000
 		write RB_COLOR_FLAG_BUFFER[0].ADDR (08903) context 0
 			RB_COLOR_FLAG_BUFFER[0].ADDR: 0x9c000
-	02120: 000b8003 00270000 9032209c 00000008
+	02140: 000b8003 00270000 9032209c 00000008
 		write RB_COLOR_FLAG_BUFFER[0].ADDR_HI (08904) context 0
 			RB_COLOR_FLAG_BUFFER[0].ADDR_HI: 0x1
 		write RB_COLOR_FLAG_BUFFER[0].PITCH (08905) context 0
 			RB_COLOR_FLAG_BUFFER[0].PITCH: { PITCH = 0 | ARRAY_PITCH = 0 }
-	02124: 00000007 00000000 90522410 00000008
+	02150: 00000007 00000000 90522410 00000008
 		write RB_MRT[0x1].BUF_INFO (0882a) context 0
 			RB_MRT[0x1].BUF_INFO: { COLOR_FORMAT = FMT6_5_6_5_UNORM | COLOR_TILE_MODE = TILE6_3 | COLOR_SWAP = WZYX }
 		write RB_MRT[0x1].PITCH (0882b) context 0
 			RB_MRT[0x1].PITCH: 256
-	02128: 00000c3b 00000010 82b220a8 00000008
+	02160: 00000c3b 00000010 82b220a8 00000008
 		write RB_MRT[0x1].ARRAY_PITCH (0882c) context 0
 			RB_MRT[0x1].ARRAY_PITCH: 20480
 		write RB_MRT[0x1].BASE (0882d) context 0
 			RB_MRT[0x1].BASE: 0xaf000
-	0212c: 00000503 002bc000 82d220b0 00000008
+	02170: 00000503 002bc000 82d220b0 00000008
 		write RB_MRT[0x1].BASE_HI (0882e) context 0
 			RB_MRT[0x1].BASE_HI: 0x1
 		write RB_MRT[0x1].BASE_GMEM (0882f) context 0
 			RB_MRT[0x1].BASE_GMEM: 0x52000
-	01e10: 00000007 00148000 82f220b8 00000008
+	01e00: 00000007 00148000 82f220b8 00000008
 		write RB_COLOR_FLAG_BUFFER[0x1].ADDR (08906) context 0
 			RB_COLOR_FLAG_BUFFER[0x1].ADDR: 0xae000
 		write RB_COLOR_FLAG_BUFFER[0x1].ADDR_HI (08907) context 0
 			RB_COLOR_FLAG_BUFFER[0x1].ADDR_HI: 0x1
-	01e14: 002b8003 00000004 90722418 00000008
+	01e10: 002b8003 00000004 90722418 00000008
 		write RB_COLOR_FLAG_BUFFER[0x1].PITCH (08908) context 0
 			RB_COLOR_FLAG_BUFFER[0x1].PITCH: { PITCH = 64 | ARRAY_PITCH = 1024 }
 		write RB_MRT[0x2].BUF_INFO (08832) context 0
 			RB_MRT[0x2].BUF_INFO: { COLOR_FORMAT = FMT6_16_16_UNORM | COLOR_TILE_MODE = TILE6_3 | COLOR_SWAP = WZYX }
-	01e18: 00010007 00000d0c 83222420 00000008
+	01e20: 00010007 00000d0c 83222420 00000008
 		write RB_MRT[0x2].PITCH (08833) context 0
 			RB_MRT[0x2].PITCH: 256
 		write RB_MRT[0x2].ARRAY_PITCH (08834) context 0
 			RB_MRT[0x2].ARRAY_PITCH: 20480
-	01e1c: 00000013 00000500 834220cc 00000008
+	01e30: 00000013 00000500 834220cc 00000008
 		write RB_MRT[0x2].BASE (08835) context 0
 			RB_MRT[0x2].BASE: 0xb6000
 		write RB_MRT[0x2].BASE_HI (08836) context 0
 			RB_MRT[0x2].BASE_HI: 0x1
-	01e20: 002d8003 00000004 836220d4 00000008
+	01e40: 002d8003 00000004 836220d4 00000008
 		write RB_MRT[0x2].BASE_GMEM (08837) context 0
 			RB_MRT[0x2].BASE_GMEM: 0
 		write RB_COLOR_FLAG_BUFFER[0x2].ADDR (08909) context 0
 			RB_COLOR_FLAG_BUFFER[0x2].ADDR: 0xb6000
-	01e24: 00000003 002d8000 909220dc 00000008
+	01e50: 00000003 002d8000 909220dc 00000008
 		write RB_COLOR_FLAG_BUFFER[0x2].ADDR_HI (0890a) context 0
 			RB_COLOR_FLAG_BUFFER[0x2].ADDR_HI: 0x1
 		write RB_COLOR_FLAG_BUFFER[0x2].PITCH (0890b) context 0
 			RB_COLOR_FLAG_BUFFER[0x2].PITCH: { PITCH = 0 | ARRAY_PITCH = 0 }
-	01e28: 00000007 00000000 90b22428 00000008
+	01e60: 00000007 00000000 90b22428 00000008
 		write RB_MRT[0x3].BUF_INFO (0883a) context 0
 			RB_MRT[0x3].BUF_INFO: { COLOR_FORMAT = FMT6_16_16_16_16_SNORM | COLOR_TILE_MODE = TILE6_3 | COLOR_SWAP = WZYX }
 		write RB_MRT[0x3].PITCH (0883b) context 0
 			RB_MRT[0x3].PITCH: 512
-	01e2c: 00000d87 00000020 83b220e8 00000008
+	01e70: 00000d87 00000020 83b220e8 00000008
 		write RB_MRT[0x3].ARRAY_PITCH (0883c) context 0
 			RB_MRT[0x3].ARRAY_PITCH: 40960
 		write RB_MRT[0x3].BASE (0883d) context 0
 			RB_MRT[0x3].BASE: 0xbf000
-	02410: 00000a03 002fc000 83d220f0 00000008
+	02400: 00000a03 002fc000 83d220f0 00000008
 		write RB_MRT[0x3].BASE_HI (0883e) context 0
 			RB_MRT[0x3].BASE_HI: 0x1
 		write RB_MRT[0x3].BASE_GMEM (0883f) context 0
 			RB_MRT[0x3].BASE_GMEM: 0x5c000
-	02414: 00000007 00170000 83f220f8 00000008
+	02410: 00000007 00170000 83f220f8 00000008
 		write RB_COLOR_FLAG_BUFFER[0x3].ADDR (0890c) context 0
 			RB_COLOR_FLAG_BUFFER[0x3].ADDR: 0xbf000
 		write RB_COLOR_FLAG_BUFFER[0x3].ADDR_HI (0890d) context 0
 			RB_COLOR_FLAG_BUFFER[0x3].ADDR_HI: 0x1
-	02418: 002fc003 00000004 90d22430 00000008
+	02420: 002fc003 00000004 90d22430 00000008
 		write RB_COLOR_FLAG_BUFFER[0x3].PITCH (0890e) context 0
 			RB_COLOR_FLAG_BUFFER[0x3].PITCH: { PITCH = 0 | ARRAY_PITCH = 0 }
 		write RB_SRGB_CNTL (0880f) context 0
 			RB_SRGB_CNTL: { 0 }
-	0241c: 00000003 00000000 80f22438 00000008
+	02430: 00000003 00000000 80f22438 00000008
 		write RB_RENDER_CNTL (08801) context 0
 			RB_RENDER_CNTL: { CCUSINGLECACHELINESIZE = 0x2 | RASTER_MODE = TYPE_TILED | RASTER_DIRECTION = LR_TB | FLAG_MRTS = 0x2 }
 		write EVENT_CMD (e7) pipe
 			EVENT_CMD: { EVENT_TYPE = EVENT | TS_WRITE }
-	02420: 00080043 00000014 0e722004 00000080
+	02440: 00080043 00000014 0e722004 00000080
 		write EVENT_TS_ADDR (e8) pipe
 			EVENT_TS_ADDR: 0x1000
 		write EVENT_TS_ADDR+0x1 (e9) pipe
 			EVENT_TS_ADDR+0x1: 0x1
-	02424: 00004003 00000004 0e9003a0 000000c0
+	02450: 00004003 00000004 0e9003a0 000000c0
 		write EVENT_TS_CTRL (ea) pipe
 			EVENT_TS_CTRL: { EVENT = CCU_FLUSH_COLOR }
 		write EVENT_TS_DATA (eb) pipe
 			EVENT_TS_DATA: 0
-	02428: 00001403 00000000 0eb003a8 000000c0
+	02460: 00001403 00000000 0eb003a8 000000c0
 		write CP_EVENT_START (0d600) context 0
 			CP_EVENT_START: { STATE_ID = 0x22 }
-	0242c: 00000002 00000088 60000000 0000000d
+	02470: 00000002 00000088 60000000 0000000d
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x22 }
-	01b90: 00000002 00000088 60100000 0000000d
+	01b80: 00000002 00000088 60100000 0000000d
 		write EVENT_CMD (e7) pipe
 			EVENT_CMD: { EVENT_TYPE = EVENT | TS_WRITE }
 		write EVENT_TS_ADDR (e8) pipe
 			EVENT_TS_ADDR: 0x1000
-	01b94: 00000017 00004000 0e80039c 000000c0
+	01b90: 00000017 00004000 0e80039c 000000c0
 		write EVENT_TS_ADDR+0x1 (e9) pipe
 			EVENT_TS_ADDR+0x1: 0x1
 		write EVENT_TS_CTRL (ea) pipe
 			EVENT_TS_CTRL: { EVENT = CCU_FLUSH_DEPTH }
-	01b98: 00000007 00001000 0ea003a4 000000c0
+	01ba0: 00000007 00001000 0ea003a4 000000c0
 		write EVENT_TS_DATA (eb) pipe
 			EVENT_TS_DATA: 0
-	01b9c: 00000002 00000000 0eb00000 00000080
+	01bb0: 00000002 00000000 0eb00000 00000080
 		write CP_EVENT_START (0d600) context 1
 			CP_EVENT_START: { STATE_ID = 0x23 }
-	01ba0: 00000002 0000008c 60000000 0000020d
+	01bc0: 00000002 0000008c 60000000 0000020d
 		write CP_EVENT_END (0d601) context 1
 			CP_EVENT_END: { STATE_ID = 0x23 }
-	01ba4: 00000002 0000008c 60100000 0000020d
+	01bd0: 00000002 0000008c 60100000 0000020d
 		write EVENT_CMD (e7) pipe
 			EVENT_CMD: { EVENT_TYPE = EVENT }
-	01ba8: 00000002 00000004 0e700000 00000080
+	01be0: 00000002 00000004 0e700000 00000080
 		write CP_EVENT_START (0d600) context 0
 			CP_EVENT_START: { STATE_ID = 0x24 }
-	01bac: 00000002 00000090 60000000 0000000d
+	01bf0: 00000002 00000090 60000000 0000000d
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x24 }
-	02190: 00000002 00000090 60100000 0000000d
+	02180: 00000002 00000090 60100000 0000000d
 		write EVENT_CMD (e7) pipe
 			EVENT_CMD: { EVENT_TYPE = EVENT }
-	02194: 00000002 00000004 0e700000 00000080
+	02190: 00000002 00000004 0e700000 00000080
 		write CP_EVENT_START (0d600) context 1
 			CP_EVENT_START: { STATE_ID = 0x25 }
-	02198: 00000002 00000094 60000000 0000020d
+	021a0: 00000002 00000094 60000000 0000020d
 		write CP_EVENT_END (0d601) context 1
 			CP_EVENT_END: { STATE_ID = 0x25 }
-	0219c: 00000002 00000094 60100000 0000020d
+	021b0: 00000002 00000094 60100000 0000020d
 		write EVENT_CMD (e7) pipe
 			EVENT_CMD: { EVENT_TYPE = EVENT }
-	021a0: 00000002 00000004 0e700000 00000080
+	021c0: 00000002 00000004 0e700000 00000080
 		write CP_EVENT_START (0d600) context 0
 			CP_EVENT_START: { STATE_ID = 0x26 }
-	021a4: 00000002 00000098 60000000 0000000d
+	021d0: 00000002 00000098 60000000 0000000d
 		write CP_EVENT_END (0d601) context 0
 			CP_EVENT_END: { STATE_ID = 0x26 }
-	021a8: 00000002 00000098 60100000 0000000d
-	021ac: 00000000 00000000 00000000 00000000
+	021e0: 00000002 00000098 60100000 0000000d
+	021f0: 00000000 00000000 00000000 00000000
 		write RBBM_WAIT_FOR_GPU_IDLE_CMD (0001c) context 0
 			RBBM_WAIT_FOR_GPU_IDLE_CMD: { WAIT_GPU_IDLE }
 		write WFI_PEND_DECR (81) pipe
-	02710: 00000007 00000004 08100070 00000080
-	02714: 00000000 00000000 00000000 00000000
+	02700: 00000007 00000004 08100070 00000080
+	02710: 00000000 00000000 00000000 00000000
 
 shader-blocks:
   - type: A6XX_TP0_TMO_DATA