add matmul scripts

base_test/matmul_test/READE.md

@@ -0,0 +1,68 @@
+Matmul 自动化测试
+# 1. 脚本说明
+matmul 放置位置：
+```shell
+# mudnn_bench 默认存放在 /usr/local/musa/ 下
+mudnn_bench
+├── bench_test_matmul.sh
+├── bin
+│   ├── mudnn_bench -> mudnn_bench-x.x.x
+│   └── mudnn_bench-x.x.x
+├── matmul_test
+```
+mudnn_bench 示例：  
+**部分旧版本mudnn_bench和mudnn版本不支持混合精度测试，需要和开发者做确认.**
+```shell
+
+# 示例 1：单卡，大矩阵，f32
+MUSA_VISIBLE_DEVICES=4 ./bin/mudnn_bench -m --mm_m 6144 --mm_n 3584 --mm_k 6144 --warmup 30 --tm i --tmv 1000 -p -t f32
+
+# 示例 2：多卡，标准尺寸，bf16
+MUSA_VISIBLE_DEVICES=0,1 ./bin/mudnn_bench -m --mm_m 4096 --mm_n 4096 --mm_k 4096 --warmup 30 --tm i --tmv 1000 -p -t bf16
+
+# 示例 3：单卡，特殊组合，int8
+MUSA_VISIBLE_DEVICES=2 ./bin/mudnn_bench -m --mm_m 8192 --mm_n 8192 --mm_k 768 --warmup 30 --tm i --tmv 1000 -p -t int8
+
+# 示例 4：使用混合精度格式
+MUSA_VISIBLE_DEVICES=3 ./bin/mudnn_bench -m --mm_m 2048 --mm_n 2048 --mm_k 2048 --warmup 30 --tm i --tmv 1000 -p -t bf16:q4:bf16:bf16
+```
+
+# 2. 测试
+可在测试脚本中自行批量配置测试MNK，warmup，iter等。
+## 2.1 fp64, tf32 测试
+注意：fp64和tf32 数据类型调用非 mudnn 接口
+```shell
+# 1. 编译
+cd ./fp64_tf32_src
+
+bash build_gemm_tf32.sh
+bash build_gemm_fp64.sh
+
+## 2. 测试
+bash test_gemm_fp64_tf32.sh
+```
+
+## 2.2 f32_f16_bf16_q8_fp8 测试
+mudnn_bench 测试矩阵value默认说明：
+- 浮点：-0.5~0.5  
+- fp8: 整型-10~10转浮点  
+- qint4：-7～7 
+- 整型：-127~127  
+> 部分版本 mudnn_bench 工具支持全 0 测试(参数 `-z` 实现)，需要和开发者确认
+```shell
+bash test_gemm_f32_f16_bf16_q8_fp8.sh
+```
+
+## 2.3 混合精度测试
+```shell
+# A,B: fp16, C,D: f32: "f16:f16:f32:f32"
+# A,B: bf16, C,D: f32: "bf16:bf16:f32:f32"
+# A,B: tf32, C,D: f32: "f32"
+# A,B: int8, C,D: int32: "int8"
+# W8A8: "q8:q8:f32:f32"
+# W4A16: "bf16:q4:bf16:bf16"
+# A,B: fp8, C,D: fp16: "float8_e4m3:float8_e4m3:f16:f16"
+
+bash test_gemm_mixed.sh
+```
+

base_test/matmul_test/exetrct_log_tools/summarize_f32_f16_bf16_q8_fp8_log.py

@@ -0,0 +1,122 @@
+import re
+import os
+import sys
+from typing import List, Dict, Optional
+
+def extract_matmul_data(log_path: str) -> List[Dict[str, str]]:
+    patterns = {
+        "datatype": re.compile(r"DataType (\w+)"),
+        "mat_params": re.compile(r"m (\d+), n (\d+), k (\d+)"),
+        "elapsed_time": re.compile(r"AverageElapsedTime\(ms\) : (\d+\.\d+)"),
+        "throughput_gops": re.compile(r"Throughput (\d+\.\d+) GOPS")
+    }
+
+    extracted = []
+    current_block = {}
+
+    try:
+        with open(log_path, 'r', encoding='utf-8') as f:
+            for line in f:
+                line = line.strip()
+
+                dt_match = patterns["datatype"].search(line)
+                if dt_match:
+                    current_block["datatype"] = dt_match.group(1)
+
+                mp_match = patterns["mat_params"].search(line)
+                if mp_match:
+                    current_block["m"] = mp_match.group(1)
+                    current_block["n"] = mp_match.group(2)
+                    current_block["k"] = mp_match.group(3)
+
+                et_match = patterns["elapsed_time"].search(line)
+                if et_match:
+                    current_block["elapsed_time"] = et_match.group(1)
+
+                tp_match = patterns["throughput_gops"].search(line)
+                if tp_match:
+                    tops = round(float(tp_match.group(1)) / 1000, 4)
+                    current_block["throughput_tops"] = str(tops)
+
+                if line == "==============================" and current_block:
+                    required = ["datatype", "m", "n", "k", "elapsed_time", "throughput_tops"]
+                    if all(key in current_block for key in required):
+                        dim = f"{current_block['m']}-{current_block['n']}-{current_block['k']}"
+                        extracted.append({
+                            "datatype": current_block["datatype"],
+                            "shape": dim,
+                            "Throughput(TOPS)": current_block["throughput_tops"],
+                            "AverageElapsedTime(ms)": current_block["elapsed_time"]
+                        })
+                    current_block = {}
+
+        required = ["datatype", "m", "n", "k", "elapsed_time", "throughput_tops"]
+        if current_block and all(key in current_block for key in required):
+            dim = f"{current_block['m']}×{current_block['n']}×{current_block['k']}"
+            extracted.append({
+                "datatype": current_block["datatype"],
+                "shape": dim,
+                "Throughput(TOPS)": current_block["throughput_tops"],
+                "AverageElapsedTime(ms)": current_block["elapsed_time"]
+            })
+
+    except Exception as e:
+        print(f"❌ 读取日志失败：{str(e)}")
+        return []
+
+    return extracted
+
+def generate_csv(data: List[Dict[str, str]], output_path: str) -> bool:
+    if not data:
+        print("⚠️  未提取到有效数据，跳过CSV生成")
+        return False
+
+    headers = ["datatype", "shape", "Throughput(TOPS)", "AverageElapsedTime(ms)"]
+
+    try:
+        with open(output_path, 'w', encoding='utf-8') as f:
+            f.write(", ".join(headers) + "\n")
+            for item in data:
+                row = [item[h] for h in headers]
+                f.write(", ".join(row) + "\n")
+        print(f"✅ CSV生成成功：{output_path}")
+        return True
+    except Exception as e:
+        print(f"❌ 生成CSV失败：{str(e)}")
+        return False
+
+def main(input_log: str, output_csv: Optional[str] = None):
+    if not os.path.isfile(input_log):
+        print(f"❌ 输入日志文件不存在：{input_log}")
+        return
+
+    if not output_csv:
+        log_dir = os.path.dirname(input_log)
+        log_name = os.path.splitext(os.path.basename(input_log))[0]
+        output_csv = os.path.join(log_dir, f"{log_name}_summary.csv")
+
+    print(f"📊 开始提取日志数据：{input_log}")
+    matmul_data = extract_matmul_data(input_log)
+
+    if not matmul_data:
+        print("❌ 未提取到任何有效测试数据")
+        return
+
+    print(f"✅ 成功提取 {len(matmul_data)} 条测试记录")
+
+    generate_csv(matmul_data, output_csv)
+    print("🎯 所有操作完成！")
+
+if __name__ == "__main__":
+    # 修正sys.argv判断（sys.argv[0]是脚本名，需至少传入1个输入文件路径）
+    if len(sys.argv) < 2:
+        print("用法：")
+        print("  python summarize_fp64_tf32_log.py <输入日志文件路径>")
+        print("示例：")
+        print("  python summarize_fp64_tf32_log.py bench.log")
+        sys.exit(1)
+
+    input_path = sys.argv[1]
+    output_path = sys.argv[1].replace('.log', '.csv')  # 日志文件同名CSV输出
+    main(input_path, output_path)
+

base_test/matmul_test/exetrct_log_tools/summarize_fp64_tf32_log.py

@@ -0,0 +1,120 @@
+import re
+import sys
+import os
+from typing import List, Dict, Optional
+
+def extract_matmul_data(log_path: str) -> List[Dict[str, str]]:
+    patterns = {
+        "datatype": re.compile(r"MatMul (\w+) Test \(MUSA\)"),
+        "mat_params": re.compile(r"m = (\d+), n = (\d+), k = (\d+)"),
+        "duration_us": re.compile(r"Duration:(\s*[\d\.]+) us"),
+        "tflops": re.compile(r"computation-\w+=(\s*[\d\.]+)")
+    }
+
+    extracted = []
+    current_block = {}
+
+    try:
+        with open(log_path, 'r', encoding='utf-8') as f:
+            for line in f:
+                line = line.strip()
+
+                dt_match = patterns["datatype"].search(line)
+                if dt_match:
+                    current_block["datatype"] = dt_match.group(1)
+
+                mp_match = patterns["mat_params"].search(line)
+                if mp_match:
+                    current_block["m"] = mp_match.group(1)
+                    current_block["n"] = mp_match.group(2)
+                    current_block["k"] = mp_match.group(3)
+
+                dur_match = patterns["duration_us"].search(line)
+                if dur_match:
+                    us_val = float(dur_match.group(1).strip())
+                    ms_val = round(us_val / 1000, 6)
+                    current_block["duration_ms"] = str(ms_val)
+
+                tf_match = patterns["tflops"].search(line)
+                if tf_match:
+                    tf_val = tf_match.group(1).strip()
+                    current_block["tflops"] = str(round(float(tf_val), 6))
+
+                if line == "========================================" and current_block:
+                    required = ["datatype", "m", "n", "k", "duration_ms", "tflops"]
+                    if all(key in current_block for key in required):
+                        shape = f"{current_block['m']}-{current_block['n']}-{current_block['k']}"
+                        extracted.append({
+                            "DataType": current_block["datatype"],
+                            "shape": shape,
+                            "Compute_ability(TFLOPS)": current_block["tflops"],
+                            "AverageElapsedTime(ms)": current_block["duration_ms"]
+                        })
+                    current_block = {}
+
+        required = ["datatype", "m", "n", "k", "duration_ms", "tflops"]
+        if current_block and all(key in current_block for key in required):
+            shape = f"{current_block['m']}-{current_block['n']}-{current_block['k']}"
+            extracted.append({
+                "DataType": current_block["datatype"],
+                "shape": shape,
+                "Compute_ability(TFLOPS)": current_block["tflops"],
+                "AverageElapsedTime(ms)": current_block["duration_ms"]
+            })
+
+    except Exception as e:
+        print(f"❌ 读取日志失败：{str(e)}")
+        return []
+
+    return extracted
+
+def generate_csv(data: List[Dict[str, str]], output_path: str) -> bool:
+    if not data:
+        print("⚠️  未提取到有效数据，跳过CSV生成")
+        return False
+
+    headers = ["DataType", "shape", "Compute_ability(TFLOPS)", "AverageElapsedTime(ms)"]
+    try:
+        with open(output_path, 'w', encoding='utf-8') as f:
+            f.write(", ".join(headers) + "\n")
+            for item in data:
+                row = [item[h] for h in headers]
+                f.write(",".join(row) + "\n")
+        print(f"✅ CSV生成成功：{output_path}")
+        return True
+    except Exception as e:
+        print(f"❌ 生成CSV失败：{str(e)}")
+        return False
+
+def main(input_log: str, output_csv: Optional[str] = None):
+    if not os.path.isfile(input_log):
+        print(f"❌ 输入日志文件不存在：{input_log}")
+        return
+
+    if not output_csv:
+        log_dir = os.path.dirname(input_log)
+        log_name = os.path.splitext(os.path.basename(input_log))[0]
+        output_csv = os.path.join(log_dir, f"{log_name}_summary.csv")
+
+    print(f"📊 开始提取日志数据：{input_log}")
+    matmul_data = extract_matmul_data(input_log)
+
+    if not matmul_data:
+        print("❌ 未提取到任何有效测试数据")
+        return
+
+    print(f"✅ 成功提取 {len(matmul_data)} 条测试记录")
+    generate_csv(matmul_data, output_csv)
+    print("🎯 所有操作完成！")
+
+if __name__ == "__main__":
+    if len(sys.argv) < 2:
+        print("用法：")
+        print("  python summarize_fp64_tf32_log.py <输入日志文件路径>")
+        print("示例：")
+        print("  python summarize_fp64_tf32_log.py bench.log")
+        sys.exit(1)
+
+    input_path = sys.argv[1]
+    output_path = sys.argv[1].replace('.log', '.csv')
+    main(input_path, output_path)

base_test/matmul_test/exetrct_log_tools/summary_mixed_data.py

@@ -0,0 +1,64 @@
+import re
+import sys
+import csv
+import os
+
+if len(sys.argv) < 2:
+    print("Usage: python summary_fix_data.py <log_file>")
+    sys.exit(1)
+
+log_file = sys.argv[1]
+print(f"📊 正在读取并解析日志：{log_file}")
+
+if not os.path.exists(log_file):
+    print("❌ 日志文件不存在")
+    sys.exit(1)
+
+# 收集结果
+records = []
+
+# 正则模式
+re_start = re.compile(r"测试:\s*M=(\d+),\s*N=(\d+),\s*K=(\d+),\s*Type=([\w:]+)")
+re_result = re.compile(r"AverageElapsedTime\(ms\)\s*:\s*([\d\.]+)\s*,\s*Throughput\s*([\d\.]+)\s*GOPS")
+
+cur_M = cur_N = cur_K = cur_type = None
+
+with open(log_file, "r", encoding="utf-8") as f:
+    for line in f:
+        line = line.strip()
+
+        # 匹配开始参数
+        m1 = re_start.search(line)
+        if m1:
+            cur_M, cur_N, cur_K, cur_type = m1.groups()
+            continue
+
+        # 匹配结果
+        m2 = re_result.search(line)
+        if m2 and cur_M is not None:
+            elapsed, gops = m2.groups()
+            records.append({
+                "M": cur_M,
+                "N": cur_N,
+                "K": cur_K,
+                "Type": cur_type,
+                "AvgTime(ms)": elapsed,
+                "GOPS": gops
+            })
+            # 清空当前块（防止串行）
+            cur_M = cur_N = cur_K = cur_type = None
+
+# 输出 CSV
+if not records:
+    print("⚠️ 未提取到任何有效数据")
+    sys.exit(0)
+
+csv_path = log_file.replace(".log", ".csv")
+with open(csv_path, "w", newline="", encoding="utf-8") as csvfile:
+    writer = csv.DictWriter(csvfile, fieldnames=records[0].keys())
+    writer.writeheader()
+    writer.writerows(records)
+
+print(f"✅ 解析完成，共 {len(records)} 条数据")
+print(f"📄 CSV 已生成：{csv_path}")
+

base_test/matmul_test/fp64_tf32_src/build_gemm_fp64.sh

@@ -0,0 +1 @@
+mcc gemm_fp64.mu -lmusart -lmublas -o gemm_fp64 --offload-arch=mp_31

base_test/matmul_test/fp64_tf32_src/build_gemm_tf32.sh

@@ -0,0 +1 @@
+g++ gemm_tf32.cpp -std=c++17 -I/usr/local/musa/include -L /usr/local/musa/lib/ -fopenmp -lmudnn -lmusart -o gemm_tf32 -O2