AIESW-27905: special handling for hint_bitmap 0 preemption point

specification/aie4/preemption_asm/generate_aie4_header.py

@@ -100,9 +100,46 @@ def generate_header(asm_dir, output_file):
 #include <map>
 #include <vector>
 #include <cstdint>
+#include <string>
 
 // C++17 inline variable - single instance across all translation units
-inline std::map<uint32_t, std::pair<std::vector<uint8_t>, std::vector<uint8_t>>> aie4_save_restore_map = {
+inline const std::map<uint32_t, std::pair<std::vector<std::string>, std::vector<std::string>>> aie4_save_restore_shimbd_map = {
+  {1, {{"DMAWRITE_data_125", "DMAWRITE_data_126"},
+       {"DMAWRITE_data_0", "DMAWRITE_data_1", "DMAWRITE_data_2", "DMAWRITE_data_3"}}},
+  {2, {{"DMAWRITE_data_250", "DMAWRITE_data_251", "DMAWRITE_data_252", "DMAWRITE_data_253"},
+       {"DMAWRITE_data_0", "DMAWRITE_data_1", "DMAWRITE_data_2", "DMAWRITE_data_3", "DMAWRITE_data_4", "DMAWRITE_data_5", "DMAWRITE_data_6", "DMAWRITE_data_7"}}},
+  {3, {{"DMAWRITE_data_375", "DMAWRITE_data_376", "DMAWRITE_data_377", "DMAWRITE_data_378", "DMAWRITE_data_379", "DMAWRITE_data_380"},
+       {"DMAWRITE_data_0", "DMAWRITE_data_1", "DMAWRITE_data_2", "DMAWRITE_data_3", "DMAWRITE_data_4", "DMAWRITE_data_5", "DMAWRITE_data_6", "DMAWRITE_data_7", "DMAWRITE_data_8", "DMAWRITE_data_9", "DMAWRITE_data_10", "DMAWRITE_data_11"}}},
+  {10, {{"DMAWRITE_data_125", "DMAWRITE_data_126"}, {"DMAWRITE_data_0", "DMAWRITE_data_1", "DMAWRITE_data_2", "DMAWRITE_data_3"}}},
+  {12, {{"DMAWRITE_data_125", "DMAWRITE_data_126"}, {"DMAWRITE_data_0", "DMAWRITE_data_1", "DMAWRITE_data_2", "DMAWRITE_data_3"}}},
+  {14, {{"DMAWRITE_data_125", "DMAWRITE_data_126"}, {"DMAWRITE_data_0", "DMAWRITE_data_1", "DMAWRITE_data_2", "DMAWRITE_data_3"}}}
+};
+
+inline const std::map<uint32_t, std::pair<std::vector<std::string>, std::vector<std::string>>>&
+get_aie4_save_restore_shimbd()
+{
+  return aie4_save_restore_shimbd_map;
+}
+
+inline const std::map<uint32_t, std::pair<std::vector<std::string>, std::vector<std::string>>> aie4_save_restore_membd_map = {
+  {1, {{"DMAWRITE_data_127", "DMAWRITE_data_128"},
+       {"DMAWRITE_data_4", "DMAWRITE_data_5", "DMAWRITE_data_6", "DMAWRITE_data_7"}}},
+  {2, {{"DMAWRITE_data_254", "DMAWRITE_data_255", "DMAWRITE_data_256", "DMAWRITE_data_257"},
+       {"DMAWRITE_data_8", "DMAWRITE_data_9", "DMAWRITE_data_10", "DMAWRITE_data_11", "DMAWRITE_data_12", "DMAWRITE_data_13", "DMAWRITE_data_14", "DMAWRITE_data_15"}}},
+  {3, {{"DMAWRITE_data_381", "DMAWRITE_data_382", "DMAWRITE_data_383", "DMAWRITE_data_384", "DMAWRITE_data_385", "DMAWRITE_data_386"},
+       {"DMAWRITE_data_12", "DMAWRITE_data_13", "DMAWRITE_data_14", "DMAWRITE_data_15", "DMAWRITE_data_16", "DMAWRITE_data_17", "DMAWRITE_data_18", "DMAWRITE_data_19", "DMAWRITE_data_20", "DMAWRITE_data_21", "DMAWRITE_data_22", "DMAWRITE_data_23"}}},
+  {10, {{"DMAWRITE_data_127", "DMAWRITE_data_128"}, {"DMAWRITE_data_4", "DMAWRITE_data_5", "DMAWRITE_data_6", "DMAWRITE_data_7"}}},
+  {12, {{"DMAWRITE_data_127", "DMAWRITE_data_128"}, {"DMAWRITE_data_4", "DMAWRITE_data_5", "DMAWRITE_data_6", "DMAWRITE_data_7"}}},
+  {14, {{"DMAWRITE_data_127", "DMAWRITE_data_128"}, {"DMAWRITE_data_4", "DMAWRITE_data_5", "DMAWRITE_data_6", "DMAWRITE_data_7"}}}
+};
+
+inline const std::map<uint32_t, std::pair<std::vector<std::string>, std::vector<std::string>>>&
+get_aie4_save_restore_membd()
+{
+  return aie4_save_restore_membd_map;
+}
+
+inline const std::map<uint32_t, std::pair<std::vector<uint8_t>, std::vector<uint8_t>>> aie4_save_restore_map = {
 """
 
     # Add entries
@@ -123,7 +160,7 @@ def generate_header(asm_dir, output_file):
     header += """
 };
 
-inline std::map<uint32_t, std::pair<std::vector<uint8_t>, std::vector<uint8_t>>>&
+inline const std::map<uint32_t, std::pair<std::vector<uint8_t>, std::vector<uint8_t>>>&
 get_aie4_save_restore()
 {
   return aie4_save_restore_map;

src/cpp/preprocessor/asm/asm_parser.cpp

@@ -1023,6 +1023,50 @@ asm_parser::inject_hintmap_save_restore(int col,
   for (const auto& hm : grp.hintmaps)
     m_hintmap_labels[col_prefix + hm] = grp.labels;
 
+  // When hint_bitmap is all-zero the scratchpad size is 0: no state needs to
+  // be saved or restored.  Inject minimal dummy jobs so the PREEMPT opcode
+  // still has valid save/restore label targets, but skip all BD configuration
+  // and stream-switch routing that the full templates would generate.
+  if (grp.size == 0) {
+    const std::string& slabel = grp.labels.first;
+    const std::string& rlabel = grp.labels.second;
+    log_info() << "hint_bitmap is 0 for col " << col
+               << ": injecting dummy save/restore jobs (labels @"
+               << slabel << " / @" << rlabel << ")" << std::endl;
+
+    const std::string save_asm =
+        slabel + ":\n"
+        "START_JOB 0\n"
+        "NOP\n"
+        "END_JOB\n"
+        "EOF\n"
+        "\n.endl " + slabel + "\n";
+
+    const std::string restore_asm =
+        rlabel + ":\n"
+        "START_JOB 0\n"
+        "LOAD_LAST_PDI\n"
+        "END_JOB\n"
+        "EOF\n"
+        "\n.endl " + rlabel + "\n";
+
+    m_current_col = col;
+    std::string dummy_save_file    = "0_0_" + save_file;
+    std::string dummy_restore_file = "0_0_" + restore_file;
+    log_info() << "Adding dummy save_file: " << dummy_save_file << " [size: " << save_asm.size()
+               << "], restore_file: " << dummy_restore_file << " [size: " << restore_asm.size() << "]" << std::endl;
+    set_data_state(false);
+    parse_lines(std::vector<char>(save_asm.begin(), save_asm.end()),
+                dummy_save_file);
+    pop_data_state();
+
+    set_data_state(false);
+    parse_lines(std::vector<char>(restore_asm.begin(), restore_asm.end()),
+                dummy_restore_file);
+    pop_data_state();
+    return;
+  }
+
   // Modify save_file like {scratchaddress}_{size}_aie4_save_3c.asm / {scratchaddress}_{size}_aie4_restore_3c.asm
   std::string save_file_mod = std::to_string(grp.scratchbase / CHUNK_SIZE) + "_" + std::to_string(grp.size / CHUNK_SIZE) + "_" + save_file;
   std::string restore_file_mod = std::to_string(grp.scratchbase / CHUNK_SIZE) + "_" + std::to_string(grp.size / CHUNK_SIZE) + "_" + restore_file;

test/aie4-ctrlcode/1col_preempt_0hint/CMakeLists.txt

@@ -0,0 +1,18 @@
+# SPDX-License-Identifier: MIT
+# Copyright (C) 2026 Advanced Micro Devices, Inc.
+
+# Assemble control code ASM with zero-hint_bitmap (hintmap_0 = 0x0) alongside a
+# non-zero hintmap (hintmap_1 = 0xf).  The zero-hintmap PREEMPT points must
+# produce minimal dummy save/restore jobs (NOP / LOAD_LAST_PDI) with no BD
+# configuration or stream-switch routing.
+add_test(NAME "aie4_1col_preempt_0hint_asm"
+  COMMAND aiebu-asm -t aie4_config -j "${CMAKE_CURRENT_SOURCE_DIR}/config.json" -o 1col-preempt-0hint.elf -f disabledump
+  WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+
+# Compare the md5sum of the generated ELF against the golden file
+add_test(NAME "aie4_1col_preempt_0hint_asm_md5sum"
+  COMMAND cmake -P "${AIEBU_SOURCE_DIR}/cmake/md5sum-compare.cmake" "${CMAKE_CURRENT_BINARY_DIR}/1col-preempt-0hint.elf" "${CMAKE_CURRENT_SOURCE_DIR}/gold.md5"
+  WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+
+set_tests_properties("aie4_1col_preempt_0hint_asm_md5sum" PROPERTIES DEPENDS "aie4_1col_preempt_0hint_asm")
+set_tests_properties("aie4_1col_preempt_0hint_asm" PROPERTIES LABELS memcheck)

test/aie4-ctrlcode/1col_preempt_0hint/config.json

@@ -0,0 +1,15 @@
+{
+    "xrt-kernels": [
+        {
+            "name": "DPU",
+            "arguments": [
+            ],
+            "instance": [
+                {
+                    "id": "dpu",
+                    "ctrl_code_file": "./test.asm"
+                }
+            ]
+        }
+    ]
+}

test/aie4-ctrlcode/1col_preempt_0hint/gold.md5

@@ -0,0 +1 @@
+691dbc2343144d54c5ad1a8876b4d196

test/aie4-ctrlcode/1col_preempt_0hint/pdi.asm

@@ -0,0 +1,6 @@
+pdi:
+START_JOB 0
+  NOP
+END_JOB
+EOF
+.endl pdi

test/aie4-ctrlcode/1col_preempt_0hint/test.asm

@@ -0,0 +1,32 @@
+.target	 aie4
+.partition	 1column
+.attach_to_group 0
+START_JOB 0
+	LOAD_PDI 0, @pdi
+END_JOB
+START_JOB 1
+	PREEMPT 0, @save, @restore , @hintmap_0
+END_JOB
+START_JOB 2
+	NOP
+END_JOB
+START_JOB 3
+	PREEMPT 1, @save, @restore , @hintmap_1
+END_JOB
+START_JOB 4
+	NOP
+END_JOB
+START_JOB 5
+	PREEMPT 2, @save, @restore , @hintmap_0
+END_JOB
+START_JOB 6
+	NOP
+END_JOB
+
+.include pdi.asm
+EOF
+.align 4
+hintmap_0:
+	.long 0x00000000
+hintmap_1:
+	.long 0x0000000f

test/aie4-ctrlcode/3col_preempt_0hint/CMakeLists.txt

@@ -0,0 +1,18 @@
+# SPDX-License-Identifier: MIT
+# Copyright (C) 2026 Advanced Micro Devices, Inc.
+
+# Assemble control code ASM with zero-hint_bitmap (hintmap_0 = 0x0) alongside a
+# non-zero hintmap (hintmap_1 = 0xf) for a 3-column partition.  The zero-hintmap
+# PREEMPT points must produce minimal dummy save/restore jobs (NOP / LOAD_LAST_PDI)
+# with no BD configuration or stream-switch routing.
+add_test(NAME "aie4_3col_preempt_0hint_asm"
+  COMMAND aiebu-asm -t aie4_config -j "${CMAKE_CURRENT_SOURCE_DIR}/config.json" -o 3col-preempt-0hint.elf -f disabledump
+  WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+
+# Compare the md5sum of the generated ELF against the golden file
+add_test(NAME "aie4_3col_preempt_0hint_asm_md5sum"
+  COMMAND cmake -P "${AIEBU_SOURCE_DIR}/cmake/md5sum-compare.cmake" "${CMAKE_CURRENT_BINARY_DIR}/3col-preempt-0hint.elf" "${CMAKE_CURRENT_SOURCE_DIR}/gold.md5"
+  WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+
+set_tests_properties("aie4_3col_preempt_0hint_asm_md5sum" PROPERTIES DEPENDS "aie4_3col_preempt_0hint_asm")
+set_tests_properties("aie4_3col_preempt_0hint_asm" PROPERTIES LABELS memcheck)

test/aie4-ctrlcode/3col_preempt_0hint/config.json

@@ -0,0 +1,15 @@
+{
+    "xrt-kernels": [
+        {
+            "name": "DPU",
+            "arguments": [
+            ],
+            "instance": [
+                {
+                    "id": "dpu",
+                    "ctrl_code_file": "./test.asm"
+                }
+            ]
+        }
+    ]
+}

test/aie4-ctrlcode/3col_preempt_0hint/gold.md5

@@ -0,0 +1 @@
+b27d37dceddc6f93b93ebd1bdcbc5f19

test/aie4-ctrlcode/3col_preempt_0hint/pdi.asm

@@ -0,0 +1,6 @@
+pdi:
+START_JOB 0
+  NOP
+END_JOB
+EOF
+.endl pdi

test/aie4-ctrlcode/3col_preempt_0hint/test.asm

@@ -0,0 +1,5 @@
+.include uc0.asm
+.eop
+.include uc2.asm
+.eop
+.include uc4.asm

test/aie4-ctrlcode/3col_preempt_0hint/uc0.asm

@@ -0,0 +1,35 @@
+.target	 aie4
+.aie_row_topology	 1-1-4-0
+.partition	 3column
+
+.attach_to_group 0
+
+START_JOB 0
+	LOAD_PDI 0, @pdi
+END_JOB
+START_JOB 1
+	PREEMPT 0, @save, @restore , @hintmap_0
+END_JOB
+START_JOB 2
+	NOP
+END_JOB
+START_JOB 3
+	PREEMPT 1, @save, @restore , @hintmap_1
+END_JOB
+START_JOB 4
+	NOP
+END_JOB
+START_JOB 5
+	PREEMPT 2, @save, @restore , @hintmap_0
+END_JOB
+START_JOB 6
+	NOP
+END_JOB
+
+.include pdi.asm
+EOF
+.align 4
+hintmap_0:
+	.long 0x00000000
+hintmap_1:
+	.long 0x0000000f

test/aie4-ctrlcode/3col_preempt_0hint/uc2.asm

@@ -0,0 +1,35 @@
+.target	 aie4
+.aie_row_topology	 1-1-4-0
+.partition	 3column
+
+.attach_to_group 2
+
+START_JOB 0
+	LOAD_PDI 0, @pdi
+END_JOB
+START_JOB 1
+	PREEMPT 0, @save, @restore , @hintmap_0
+END_JOB
+START_JOB 2
+	NOP
+END_JOB
+START_JOB 3
+	PREEMPT 1, @save, @restore , @hintmap_0
+END_JOB
+START_JOB 4
+	NOP
+END_JOB
+START_JOB 5
+	PREEMPT 2, @save, @restore , @hintmap_1
+END_JOB
+START_JOB 6
+	NOP
+END_JOB
+
+.include pdi.asm
+EOF
+.align 4
+hintmap_0:
+	.long 0x00000000
+hintmap_1:
+	.long 0x0000000f

test/aie4-ctrlcode/3col_preempt_0hint/uc4.asm

@@ -0,0 +1,35 @@
+.target	 aie4
+.aie_row_topology	 1-1-4-0
+.partition	 3column
+
+.attach_to_group 4
+
+START_JOB 0
+	LOAD_PDI 0, @pdi
+END_JOB
+START_JOB 1
+	PREEMPT 0, @save, @restore , @hintmap_0
+END_JOB
+START_JOB 2
+	NOP
+END_JOB
+START_JOB 3
+	PREEMPT 1, @save, @restore , @hintmap_1
+END_JOB
+START_JOB 4
+	NOP
+END_JOB
+START_JOB 5
+	PREEMPT 2, @save, @restore , @hintmap_0
+END_JOB
+START_JOB 6
+	NOP
+END_JOB
+
+.include pdi.asm
+EOF
+.align 4
+hintmap_0:
+	.long 0x00000000
+hintmap_1:
+	.long 0x0000000f

test/aie4-ctrlcode/CMakeLists.txt

@@ -5,6 +5,8 @@ if (${Boost_VERSION} VERSION_GREATER_EQUAL "1.71.0")
   add_subdirectory(basic)
   add_subdirectory(move_ddr_to_memtile_ctrlpkt)
   add_subdirectory(1col_preempt)
+  add_subdirectory(1col_preempt_0hint)
+  add_subdirectory(3col_preempt_0hint)
   add_subdirectory(multi_layer)
   add_subdirectory(compile_time/fused_hw_package_subgraph_0)
   add_subdirectory(multi_graph)

test/aie4-ctrlcode/compile_time/fused_hw_package_subgraph_0/gold.md5

@@ -1 +1 @@
-a91699474561b1d112d7885add3dd3d8
+75c70239ed04fc3d48399f0afffa0fe7