#5480: Add test dispatch buffer overflow assert, and update test cmd …

…line to avoid overflow

 - Generalize (use soc desc) prefetcher buffer overflow assert, and add
   one for dispatch buffer too.  My 8KB page size paged-write unit test
   was unknowingly overflowing L1, this assert catches it.

 - Adjust test cmd line to reduce pages per dispatch buffer block in
   half (48 -> 24) to keep as 768KB dispatch buffer size. This fixes
   newly added assert and hangs, mismatches seen in various versions of
   this test.

tests/scripts/run_cpp_fd2_tests.sh

@@ -20,18 +20,18 @@ fi
 #############################################
 echo "Running test_prefetcher tests now...";
 
-TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 0 -i 3  # TrueSmoke Test
-TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 1 -i 3  # Smoke Test
-TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 2 -i 3  # Random Test
-TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 3 -i 3  # PCIE Test
-TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 4 -i 3  # Paged DRAM Read Test
-TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 5 -i 3  # Paged DRAM Write + Read Test
-TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 6 -i 3  # Host Test
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 0 -i 3  # TrueSmoke Test
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 1 -i 3  # Smoke Test
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 2 -i 3  # Random Test
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 3 -i 3  # PCIE Test
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 4 -i 3  # Paged DRAM Read Test
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 5 -i 3  # Paged DRAM Write + Read Test
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 6 -i 3  # Host Test
 
 # Testcase: Paged Write Cmd to DRAM. 256 pages, 224b size.
-TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 4 -i 1 -dpgs 224 -dpgr 256
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 4 -i 1 -dpgs 224 -dpgr 256
 # Testcase: Paged Write Cmd to DRAM. 120 pages, 64b size.
-TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 4 -i 1 -dpgs 64 -dpgr 120
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_prefetcher -t 4 -i 1 -dpgs 64 -dpgr 120
 
 #############################################
 # TEST_DISPATCHER TESTS                     #
@@ -56,8 +56,8 @@ echo "Running test_dispatcher tests now...";
 # Testcase: Arbitrary non-even numbers. This caught some test issues with overflowing start_page one test implementation.
 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_dispatcher -i 1 -w 0 -t 2 -min 16 -max 16 -lps 5 -pbs 275 -np 13
 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_dispatcher -i 1 -w 0 -t 3 -min 16 -max 16 -lps 5 -pbs 275 -np 13
-# 11.7 GB/s whb0 - DRAM.   Have to reduce number of pages to not exceed 1MB L1 for GS.
-./build/test/tt_metal/perf_microbenchmark/dispatch/test_dispatcher -w 1000 -t 2 -min 8192 -max 8192 -lps 13 -pbs 2 -np 100 -i 1000
+# 11.885 GB/s whb0 - DRAM.   Have to reduce number of pages to not exceed 1MB L1 for GS. Also, number of pages per block.
+./build/test/tt_metal/perf_microbenchmark/dispatch/test_dispatcher -w 10 -t 2 -min 8192 -max 8192 -lps 13 -pbs 2 -np 100 -i 1 -pi 5000 -bs 24
 
 # Packed Write
 ./build/test/tt_metal/perf_microbenchmark/dispatch/test_dispatcher -i 3 -w 5 -t 4 -min 256 -max 256

tests/tt_metal/tt_metal/perf_microbenchmark/dispatch/run_paged_tests.sh

@@ -141,6 +141,5 @@ TT_METAL_SLOW_DISPATCH_MODE=1 ${TT_METAL_HOME}/build/test/tt_metal/perf_microben
 # 9.600 GB/s whb0
 TT_METAL_SLOW_DISPATCH_MODE=1 ${TT_METAL_HOME}/build/test/tt_metal/perf_microbenchmark/dispatch/test_dispatcher -w 10 -t 2 -wx 0 -wy 1 -min 4096 -max 4096 -lps 12 -pbs 2 -np 128 -c -i 1 -pi 10000 |& tee ${DIR}/perf_write_128_page_4096b_size_dispatch_buffer_4096b_pages_10000_iter_dram_pbs2.log
 
-# 11.7 GB/s whb0
-TT_METAL_SLOW_DISPATCH_MODE=1 ${TT_METAL_HOME}/build/test/tt_metal/perf_microbenchmark/dispatch/test_dispatcher -w 1000 -t 2 -wx 0 -wy 1 -min 8192 -max 8192 -lps 13 -pbs 2 -np 128 -c -i 1000 |& tee ${DIR}/perf_write_128_page_8192b_size_dispatch_buffer_8192b_pages_1000_iter_dram_pbs2.log
-# FIXME Hangs TT_METAL_SLOW_DISPATCH_MODE=1 ${TT_METAL_HOME}/build/test/tt_metal/perf_microbenchmark/dispatch/test_dispatcher -w 10 -t 2 -wx 0 -wy 1 -min 8192 -max 8192 -lps 13 -pbs 2 -np 128 -c -i 1 -pi 10000 |& tee ${DIR}/perf_write_128_page_8192b_size_dispatch_buffer_8192b_pages_10000_iter_dram_pbs2.log
+# 11.872 GB/s whb0 - reduced number of pages per block in half otherwise uses 1536 KB L1 (exceeds for GS, WH)
+TT_METAL_SLOW_DISPATCH_MODE=1 ${TT_METAL_HOME}/build/test/tt_metal/perf_microbenchmark/dispatch/test_dispatcher -w 10 -t 2 -wx 0 -wy 1 -min 8192 -max 8192 -lps 13 -pbs 2 -np 128 -c -i 1 -pi 5000 -bs 24 |& tee ${DIR}/perf_write_128_page_8192b_size_dispatch_buffer_8192b_pages_10000_iter_dram_pbs2.log

tests/tt_metal/tt_metal/perf_microbenchmark/dispatch/test_dispatcher.cpp

@@ -111,6 +111,8 @@ void init(int argc, char **argv) {
     dispatch_buffer_block_size_pages_g = test_args::get_command_option_uint32(input_args, "-bs", DEFAULT_DISPATCH_BUFFER_BLOCK_SIZE_PAGES);
     dispatch_buffer_size_blocks_g = test_args::get_command_option_uint32(input_args, "-b", DEFAULT_DISPATCH_BUFFER_SIZE_BLOCKS);
     dispatch_buffer_size_g = dispatch_buffer_page_size_g * dispatch_buffer_block_size_pages_g * dispatch_buffer_size_blocks_g;
+    log_info(tt::LogTest, "Computed dispatch_buffer_size_g: {} from page_size: {} block_size_pages: {} blocks: {}",
+        dispatch_buffer_size_g, dispatch_buffer_page_size_g, dispatch_buffer_block_size_pages_g, dispatch_buffer_size_blocks_g);
 
     prefetcher_page_batch_size_g = test_args::get_command_option_uint32(input_args, "-ppbs", DEFAULT_PREFETCHER_PAGE_BATCH_SIZE);
 
@@ -119,6 +121,8 @@ void init(int argc, char **argv) {
     // divide the batch size evenlly, one page for terminate
     pbs_pages = pbs_pages / prefetcher_page_batch_size_g * prefetcher_page_batch_size_g + terminate_cmd_pages;
     prefetcher_buffer_size_g = pbs_pages * dispatch_buffer_page_size_g;
+    log_info(tt::LogTest, "Computed prefetcher_buffer_size_g: {} from page_size: {} prefetch_buffer_pages: {}",
+        prefetcher_buffer_size_g, dispatch_buffer_page_size_g, pbs_pages);
 
     max_xfer_size_bytes_g = test_args::get_command_option_uint32(input_args, "-max", max_xfer_size_bytes_g);
     min_xfer_size_bytes_g = test_args::get_command_option_uint32(input_args, "-min", min_xfer_size_bytes_g);
@@ -355,8 +359,15 @@ int main(int argc, char **argv) {
         // Want different buffers on each core, instead use big buffer and self-manage it
         uint32_t l1_buf_base = align(DISPATCH_L1_UNRESERVED_BASE, dispatch_buffer_page_size_g);
         TT_ASSERT((l1_buf_base & (dispatch_buffer_page_size_g - 1)) == 0);
-        if (prefetcher_buffer_size_g + l1_buf_base > 1024 * 1024) {
-            log_fatal(LogTest, "Error, prefetcher buffer size too large\n");
+
+        // Make sure user doesn't exceed available L1 space with cmd line arguments.
+        auto &soc_desc = tt::Cluster::instance().get_soc_desc(device->id());
+        if (prefetcher_buffer_size_g + l1_buf_base > soc_desc.worker_l1_size) {
+            log_fatal(LogTest, "Prefetcher buffer size too large. {} exceeds l1_worker_size: {}", dispatch_buffer_size_g, soc_desc.worker_l1_size);
+            exit(-1);
+        }
+        if (dispatch_buffer_size_g + l1_buf_base > soc_desc.worker_l1_size) {
+            log_fatal(LogTest, "Dispatcher buffer size too large. {} exceeds l1_worker_size: {}", dispatch_buffer_size_g, soc_desc.worker_l1_size);
             exit(-1);
         }