Use unplaced GEMM, python code refactor, and add throughput metric

Author: kurtis-b-1

example/gemm/CMakeLists.txt

@@ -16,7 +16,7 @@ function(add_aie_gemm_xclbin m k n NUM_COLUMNS B_COL_MAJ C_COL_MAJ TRACE_SIZE AR
     add_aie_design(${XCLBIN_NAME}
                    XCLBIN_ONLY
                    PYTHON gemm.py
-                   PYTHON_FLAGS --dev ${DEVICE} -M ${DUMMY_M} -K ${DUMMY_K} -N ${DUMMY_N} -m ${m} -k ${k} -n ${n} --n-aie-cols ${NUM_COLUMNS} --b-col-maj ${B_COL_MAJ} --c-col-maj ${C_COL_MAJ} --dtype_in ${INPUT_DATA_TYPE_PYTHON} --dtype_out ${OUTPUT_DATA_TYPE_PYTHON} --trace_size ${TRACE_SIZE} ${EMULATE_STR} --prio-accuracy --output-file-path ${CMAKE_BINARY_DIR}/aie/${XCLBIN_NAME}.mlir
+                   PYTHON_FLAGS --dev ${DEVICE} -M ${DUMMY_M} -K ${DUMMY_K} -N ${DUMMY_N} -m ${m} -k ${k} -n ${n} --n-aie-cols ${NUM_COLUMNS} --b-col-maj ${B_COL_MAJ} --c-col-maj ${C_COL_MAJ} --dtype_in ${INPUT_DATA_TYPE_PYTHON} --dtype_out ${OUTPUT_DATA_TYPE_PYTHON} --trace_size ${TRACE_SIZE} ${EMULATE_STR} ${PRIO_ACC_STR} --output-file-path ${CMAKE_BINARY_DIR}/aie/${XCLBIN_NAME}.mlir
                    AIE_CORE_KERNELS ${ARCHIVE_NAME}
                    EXTRA_AIECC_FLAGS --dynamic-objFifos
                    OUTPUT_XCLBIN GEMM_XCLBIN_${XCLBIN_NAME})
@@ -43,7 +43,7 @@ function(add_aie_gemm_design M K N m k n NUM_COLUMNS B_COL_MAJ C_COL_MAJ TRACE_S
     add_aie_design(${EXAMPLE}
                    INSTS_ONLY
                    PYTHON gemm.py
-                   PYTHON_FLAGS --dev ${DEVICE} -M ${M} -K ${K} -N ${N} -m ${m} -k ${k} -n ${n} --n-aie-cols ${NUM_COLUMNS} --b-col-maj ${B_COL_MAJ} --c-col-maj ${C_COL_MAJ} --dtype_in ${INPUT_DATA_TYPE_PYTHON} --dtype_out ${OUTPUT_DATA_TYPE_PYTHON} --trace_size ${TRACE_SIZE} ${EMULATE_STR} --prio-accuracy --output-file-path ${CMAKE_BINARY_DIR}/aie/${EXAMPLE}.mlir
+                   PYTHON_FLAGS --dev ${DEVICE} -M ${M} -K ${K} -N ${N} -m ${m} -k ${k} -n ${n} --n-aie-cols ${NUM_COLUMNS} --b-col-maj ${B_COL_MAJ} --c-col-maj ${C_COL_MAJ} --dtype_in ${INPUT_DATA_TYPE_PYTHON} --dtype_out ${OUTPUT_DATA_TYPE_PYTHON} --trace_size ${TRACE_SIZE} ${EMULATE_STR} ${PRIO_ACC_STR} --output-file-path ${CMAKE_BINARY_DIR}/aie/${EXAMPLE}.mlir
                    EXTRA_AIECC_FLAGS --dynamic-objFifos
                    OUTPUT_INSTS GEMM_INSTS)
 
@@ -69,7 +69,8 @@ function(add_aie_gemm_design M K N m k n NUM_COLUMNS B_COL_MAJ C_COL_MAJ TRACE_S
                         "PASS!"
                     METRICS
                         "Latency" [=[Latency \(us\): (?P<metric>\d+)]=]
-                        "Bandwidth" [=[Effective Bandwidth: (?P<metric>[\d\.e\+-]+) GB/s]=])
+                        "Bandwidth" [=[Effective Bandwidth: (?P<metric>[\d\.e\+-]+) GB/s]=]
+                        "GFLOP/s" [=[Throughput:\s*(?P<metric>[\d\.e\+-]+) GFLOP/s]=])
 endfunction()
 
 set(M_LIST "2048")
@@ -86,15 +87,22 @@ set(ARCHIVE_NAME "gemm_${m}x${k}x${n}_archive.a")
 set(AIE_BUILD_DIR ${CMAKE_BINARY_DIR}/aie)
 set(ARCHIVE_PATH "${AIE_BUILD_DIR}/${ARCHIVE_NAME}")
 
+set(PRIO_ACCURACY True)
 set(EMULATE_BFLOAT16_MMUL_WITH_BFP16 False)
 set (B_COL_MAJ 0)
 set (C_COL_MAJ 0)
 
-set(MM_KERNEL_DEFINES "DIM_M=${m}" "DIM_K=${k}" "DIM_N=${n}" "bf16_f32_ONLY" "ROUND_CONV_EVEN")
+set(MM_KERNEL_DEFINES "DIM_M=${m}" "DIM_K=${k}" "DIM_N=${n}" "ROUND_CONV_EVEN")
 if (EMULATE_BFLOAT16_MMUL_WITH_BFP16)
     set(EMULATE_STR --emulate-bf16-mmul-with-bfp16)
     list(APPEND MM_KERNEL_DEFINES "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16")
 endif()
+if (PRIO_ACCURACY)
+    set(PRIO_ACC_STR --prio-accuracy)
+    list(APPEND MM_KERNEL_DEFINES "bf16_f32_ONLY")
+else()
+    list(APPEND MM_KERNEL_DEFINES "bf16_bf16_ONLY")
+endif()
 if (B_COL_MAJ)
     list(APPEND MM_KERNEL_DEFINES "B_COL_MAJ")
 endif()

example/gemm/gemm.cpp

@@ -160,6 +160,11 @@ int main(int argc, const char *argv[])
     double bandwidth_GBps = (total_bytes / (1024 * 1024 * 1024)) / (npu_time * 1e-6);
     std::cout << "Effective Bandwidth: " << bandwidth_GBps << " GB/s" << std::endl;
 
+    // Need to cast to long long since M/K/N=2048 would give 17*10^9 ops for example
+    unsigned long long n_ops = static_cast<unsigned long long>(M) * K * N * 2;
+    float throughput = n_ops / npu_time / 1e3; // GOP/s
+    std::cout << "Throughput: " << throughput << " GFLOP/s" << std::endl;
+
     std::bfloat16_t *bufOut1 = bo_out.map<std::bfloat16_t *>();
 
     // Compare with golden reference