Formatting

src/rewrite_quantization.cpp

@@ -64,23 +64,26 @@ void apply_quantizelinear(module& m, instruction_ref ins)
         max_quant = qt.max();
         min_quant = qt.min();
     });
-    if (enabled(MIGRAPHX_BROADCAST_Q{}))
+    if(enabled(MIGRAPHX_BROADCAST_Q{}))
     {
         auto s       = add_zero_point->get_shape();
         auto min_arg = m.add_literal(literal{shape{s.type()}, {min_quant}});
         auto max_arg = m.add_literal(literal{shape{s.type()}, {max_quant}});
         // auto min_mbcast =
-        //     m.insert_instruction(ins, make_op("multibroadcast", {{"out_lens", s.lens()}}), min_arg);
+        //     m.insert_instruction(ins, make_op("multibroadcast", {{"out_lens", s.lens()}}),
+        //     min_arg);
         // auto max_mbcast =
-        //     m.insert_instruction(ins, make_op("multibroadcast", {{"out_lens", s.lens()}}), max_arg);
-        
+        //     m.insert_instruction(ins, make_op("multibroadcast", {{"out_lens", s.lens()}}),
+        //     max_arg);
+
         // auto saturate =
         //     m.insert_instruction(ins, make_op("clip"), add_zero_point, min_mbcast, max_mbcast);
-        auto saturate = insert_common_op(m, ins, make_op("clip"), {add_zero_point, min_arg, max_arg});
+        auto saturate =
+            insert_common_op(m, ins, make_op("clip"), {add_zero_point, min_arg, max_arg});
         m.replace_instruction(
             ins, make_op("convert", {{"target_type", ins->get_shape().type()}}), saturate);
     }
-    else 
+    else
     {
         auto s = add_zero_point->get_shape();
         std::vector<int> min_data(s.elements(), min_quant);
@@ -88,7 +91,8 @@ void apply_quantizelinear(module& m, instruction_ref ins)
         auto min_arg = m.add_literal(literal(s, min_data));
         auto max_arg = m.add_literal(literal(s, max_data));
 
-        auto saturate = m.insert_instruction(ins, make_op("clip"), add_zero_point, min_arg, max_arg);
+        auto saturate =
+            m.insert_instruction(ins, make_op("clip"), add_zero_point, min_arg, max_arg);
         m.replace_instruction(
             ins, make_op("convert", {{"target_type", ins->get_shape().type()}}), saturate);
     }

tools/gemm_perf.py

 import subprocess, csv, re, datetime
 
+
 class CSVFile:
+
     def __init__(self, path="output.csv"):
         self.path = path
 
@@ -18,20 +20,23 @@ def get_device_name():
     matches = re.findall("gfx\d*[a-z]*", str(out.stdout))
     return matches[0]
 
-def run_perf(model, batch_size, int8=False, use_ck=False, use_large_k=False, disable_fusion=False):
+
+def run_perf(model,
+             batch_size,
+             int8=False,
+             use_ck=False,
+             use_large_k=False,
+             disable_fusion=False):
     env_vars = ""
     if use_ck:
-        env_vars += "MIGRAPHX_ENABLE_CK=1 " 
+        env_vars += "MIGRAPHX_ENABLE_CK=1 "
         if use_large_k:
-            env_vars += "MIGRAPHX_USE_LARGE_K=1 " 
+            env_vars += "MIGRAPHX_USE_LARGE_K=1 "
         if disable_fusion:
             env_vars += "MIGRAPHX_DISABLE_CK_FUSION=1 "
     int8_str = "--int8" if int8 else ""
     cmd = f"{env_vars} ../build/bin/driver perf {model} --fill1 input_ids --input-dim @input_ids {batch_size} 384 --batch {batch_size} --fp16 {int8_str}  --exhaustive-tune"
-    out = subprocess.run(cmd,
-                         capture_output=True,
-                         check=True,
-                         shell=True)
+    out = subprocess.run(cmd, capture_output=True, check=True, shell=True)
 
     summary = re.findall("Summary.*", str(out.stdout))[0].replace("\\n", "\n")
     total_time = re.findall("Total time: \d+\.\d*", summary)[0]
@@ -42,13 +47,14 @@ def run_perf(model, batch_size, int8=False, use_ck=False, use_large_k=False, dis
         ck_gemm_time = re.findall("\d+\.\d*", ck_gemm_time[0])[0]
     else:
         ck_gemm_time = "0.0"
-    
-    rb_gemm_time = re.findall("gpu::quant_gemm: \d+\.\d*|gpu::gemm: \d+\.\d*", summary)
+
+    rb_gemm_time = re.findall("gpu::quant_gemm: \d+\.\d*|gpu::gemm: \d+\.\d*",
+                              summary)
     if rb_gemm_time:
         rb_gemm_time = re.findall("\d+\.\d*", rb_gemm_time[0])[0]
     else:
         rb_gemm_time = "0.0"
-    
+
     gemm_pack_time = re.findall("gpu::int8_gemm_pack_a: \d+\.\d*", summary)
     if gemm_pack_time:
         gemm_pack_time = re.findall("\d+\.\d*", gemm_pack_time[0])[0]
@@ -64,19 +70,19 @@ def run_perf(model, batch_size, int8=False, use_ck=False, use_large_k=False, dis
     with open("perf_summaries.txt", "a+") as f:
         f.write(cmd + "\n")
         f.write(summary + "\n\n")
-    
+
     return [total_time] + gemm_times
 
+
 def run_ck_perf(model, batch_size, int8=False, use_large_k=False):
-    # CK with fusions 
+    # CK with fusions
     total_time = run_perf(model, batch_size, int8, True, use_large_k, False)[0]
-    # CK without fusions 
+    # CK without fusions
     gemm_times = run_perf(model, batch_size, int8, True, use_large_k, True)
 
     return [total_time] + gemm_times[1:]
 
 
-
 if __name__ == "__main__":
     device_id = get_device_name()
     model = "/code/bert_base_cased_1_fp16_gpu.onnx"
@@ -84,7 +90,10 @@ if __name__ == "__main__":
     cf.write_row([str(datetime.datetime.now())])
     cf.write_row([device_id])
     cf.write_row([model])
-    headers = ["", "Total Time (ms)", "CK GEMM Time (ms)", "RB GEMM Time (ms)", "GEMM Pack Time (ms)", "Total GEMM Time (ms)"]
+    headers = [
+        "", "Total Time (ms)", "CK GEMM Time (ms)", "RB GEMM Time (ms)",
+        "GEMM Pack Time (ms)", "Total GEMM Time (ms)"
+    ]
 
     batch_size = "1"
     # int8:
@@ -95,7 +104,8 @@ if __name__ == "__main__":
     # CK Only
     cf.write_row(["CK"] + run_ck_perf(model, batch_size, quantize, True))
     # CK + rocBLAS (k>2048)
-    cf.write_row(["CK + rocBLAS(k>2048)"] + run_ck_perf(model, batch_size, quantize, False))
+    cf.write_row(["CK + rocBLAS(k>2048)"] +
+                 run_ck_perf(model, batch_size, quantize, False))
     # rocBLAS Only
     cf.write_row(["rocBLAS"] + run_perf(model, batch_size, quantize))
     cf.write_row()
@@ -108,7 +118,8 @@ if __name__ == "__main__":
     # CK Only
     cf.write_row(["CK"] + run_ck_perf(model, batch_size, quantize, True))
     # CK + rocBLAS (k>2048)
-    cf.write_row(["CK + rocBLAS(k>2048)"] + run_ck_perf(model, batch_size, quantize, False))
+    cf.write_row(["CK + rocBLAS(k>2048)"] +
+                 run_ck_perf(model, batch_size, quantize, False))
     # rocBLAS Only
     cf.write_row(["rocBLAS"] + run_perf(model, batch_size, quantize))
     cf.write_row()
@@ -122,7 +133,8 @@ if __name__ == "__main__":
     # CK Only
     cf.write_row(["CK"] + run_ck_perf(model, batch_size, quantize, True))
     # CK + rocBLAS (k>2048)
-    cf.write_row(["CK + rocBLAS(k>2048)"] + run_ck_perf(model, batch_size, quantize, False))
+    cf.write_row(["CK + rocBLAS(k>2048)"] +
+                 run_ck_perf(model, batch_size, quantize, False))
     # rocBLAS Only
     cf.write_row(["rocBLAS"] + run_perf(model, batch_size, quantize))
     cf.write_row()
@@ -135,7 +147,8 @@ if __name__ == "__main__":
     # CK Only
     cf.write_row(["CK"] + run_ck_perf(model, batch_size, quantize, True))
     # CK + rocBLAS (k>2048)
-    cf.write_row(["CK + rocBLAS(k>2048)"] + run_ck_perf(model, batch_size, quantize, False))
+    cf.write_row(["CK + rocBLAS(k>2048)"] +
+                 run_ck_perf(model, batch_size, quantize, False))
     # rocBLAS Only
     cf.write_row(["rocBLAS"] + run_perf(model, batch_size, quantize))
-    cf.write_row()
\ No newline at end of file
+    cf.write_row()