Add warmups and loops params to GEMV benchmarks

gemv/Makefile

+GPU_ARCH ?= gfx936
 CXX ?= hipcc
 CXX_FLAGS ?= -std=c++17 -O3
-GPU_ARCH ?= gfx936
+HIPCC_FLAGS = --offload-arch=$(GPU_ARCH)
+NVCC_FLAGS = -arch=$(GPU_ARCH) -x cu
 
 TARGET := gemv_bench
 SRC := main.cpp
@@ -11,15 +13,14 @@ DEP := gemv_bf16.h gemv_utils.h hip_compat.h
 all: $(TARGET)
 
 # 根据 CXX 变量判断编译器类型
-ifneq (,$(findstring nvcc,$(CXX)))
-
-# NVCC 编译
+ifneq (,$(findstring hipcc,$(CXX)))
+# HIPCC 编译
 $(TARGET): $(SRC) $(DEP)
-	$(CXX) $(CXX_FLAGS) -arch=$(GPU_ARCH) -x cu $< -o $@
+	$(CXX) $(CXX_FLAGS) $(HIPCC_FLAGS) $< -o $@
 else
-# HIPCC 编译
+# NVCC 编译
 $(TARGET): $(SRC) $(DEP)
-	$(CXX) $(CXX_FLAGS) --offload-arch=$(GPU_ARCH) $< -o $@
+	$(CXX) $(CXX_FLAGS) $(NVCC_FLAGS) $< -o $@
 endif
 
 clean:

gemv/gemv_utils.h

@@ -106,7 +106,8 @@ struct KernelCase {
   KernelLauncher func;
 };
 
-inline void run_benchmark(const std::vector<KernelCase> &cases, int M, int K,
+inline void run_benchmark(int warmups, int loops,
+                          const std::vector<KernelCase> &cases, int M, int K,
                           float alpha, const hip_bfloat16 *A, int lda,
                           const hip_bfloat16 *x, float beta, hip_bfloat16 *y,
                           bool do_verify) {
@@ -144,6 +145,7 @@ inline void run_benchmark(const std::vector<KernelCase> &cases, int M, int K,
   printf("lda=%d\n", lda);
   printf("sizeof(A)=%lu MB\n", M * lda * sizeof(hip_bfloat16) / 1024 / 1024);
   printf("L2 cache=%d MB\n", get_l2_cache_size());
+  printf("Warmups=%d, Loops=%d\n", warmups, loops);
   printf("%s\n", std::string(w_table, '-').c_str());
   printf("%-38s %10s %10s %10s %8s\n", "Kernel Name", "Time (us)", "GFLOPS",
          "BW (GB/s)", "Result");
@@ -172,13 +174,13 @@ inline void run_benchmark(const std::vector<KernelCase> &cases, int M, int K,
     }
 
     // 2. Warmup
-    for (int i = 0; i < 100; ++i) {
+    for (int i = 0; i < warmups; ++i) {
       k.func(M, K, alpha, A, lda, x, beta, y);
     }
     checkHipErrors(hipDeviceSynchronize());
 
     // 3. Timing
-    int num_runs = 100;
+    int num_runs = loops;
     checkHipErrors(hipEventRecord(start));
     for (int i = 0; i < num_runs; ++i) {
       k.func(M, K, alpha, A, lda, x, beta, y);

gemv/main.cpp

 #include "gemv_bf16.h"
 
 int main(int argc, char **argv) {
+  int warmups = 100;
+  int loops = 2000;
   bool do_verify = false;
   float alpha = 1.0f;
   float beta = 0.0f;
@@ -10,6 +12,14 @@ int main(int argc, char **argv) {
   int lda = K;
   int block_size = 256;
 
+  if (char *value = getCmdOption(argv, argv + argc, "--warmups")) {
+    warmups = std::stoi(value);
+  }
+
+  if (char *value = getCmdOption(argv, argv + argc, "--loops")) {
+    loops = std::stoi(value);
+  }
+
   if (char *value = getCmdOption(argv, argv + argc, "--verify")) {
     do_verify = std::stoi(value) == 1;
   }
@@ -244,7 +254,8 @@ int main(int argc, char **argv) {
        }});
 
   // 运行所有测试
-  run_benchmark(kernels, M, K, alpha, d_A, lda, d_x, beta, d_y, do_verify);
+  run_benchmark(warmups, loops, kernels, M, K, alpha, d_A, lda, d_x, beta, d_y,
+                do_verify);
 
   // 清理
   checkHipErrors(hipFree(d_A));

gemv/run-all.sh

@@ -4,23 +4,24 @@ set -e
 export HIP_VISIBLE_DEVICES=1
 BIND_CMD="numactl -N 0 -m 0"
 
-CXX=hipcc make
+make clean
+CXX=hipcc make GPU_ARCH=gfx936
 # CXX=nvcc make GPU_ARCH=sm_80
 if [[ "$*" == *"--pmc"* ]]; then
-    PROF_CMD="hipprof --trace-off --pmc --pmc-type 3"
-    ${PROF_CMD} -o log/pmc-k1 ${BIND_CMD} ./gemv_bench --verify 1 -M 11264 -K 4096
-    ${PROF_CMD} -o log/pmc-k2 ${BIND_CMD} ./gemv_bench --verify 1 -M 4096 -K 11264
-    ${PROF_CMD} -o log/pmc-k3 ${BIND_CMD} ./gemv_bench --verify 1 -M 12288 -K 4096
-    ${PROF_CMD} -o log/pmc-k4 ${BIND_CMD} ./gemv_bench --verify 1 -M 4096 -K 4096
+    PROF_CMD="hipprof --trace-off --pmc"
+    ${PROF_CMD} -o log/pmc-w1 ${BIND_CMD} ./gemv_bench --warmups 10 --loops 20 --verify 1 -M 11264 -K 4096
+    ${PROF_CMD} -o log/pmc-w2 ${BIND_CMD} ./gemv_bench --warmups 10 --loops 20 --verify 1 -M 4096 -K 11264
+    ${PROF_CMD} -o log/pmc-w3 ${BIND_CMD} ./gemv_bench --warmups 10 --loops 20 --verify 1 -M 12288 -K 4096
+    ${PROF_CMD} -o log/pmc-w4 ${BIND_CMD} ./gemv_bench --warmups 10 --loops 20 --verify 1 -M 4096 -K 4096
 elif [[ "$*" == *"--trace"* ]]; then
     PROF_CMD="hipprof --hip-trace"
-    ${PROF_CMD} -o log/trace-k1 ${BIND_CMD} ./gemv_bench --verify 1 -M 11264 -K 4096
-    ${PROF_CMD} -o log/trace-k2 ${BIND_CMD} ./gemv_bench --verify 1 -M 4096 -K 11264
-    ${PROF_CMD} -o log/trace-k3 ${BIND_CMD} ./gemv_bench --verify 1 -M 12288 -K 4096
-    ${PROF_CMD} -o log/trace-k4 ${BIND_CMD} ./gemv_bench --verify 1 -M 4096 -K 4096
+    ${PROF_CMD} -o log/trace-w1 ${BIND_CMD} ./gemv_bench --warmups 100 --loops 1000 --verify 1 -M 11264 -K 4096
+    ${PROF_CMD} -o log/trace-w2 ${BIND_CMD} ./gemv_bench --warmups 100 --loops 1000 --verify 1 -M 4096 -K 11264
+    ${PROF_CMD} -o log/trace-w3 ${BIND_CMD} ./gemv_bench --warmups 100 --loops 1000 --verify 1 -M 12288 -K 4096
+    ${PROF_CMD} -o log/trace-w4 ${BIND_CMD} ./gemv_bench --warmups 100 --loops 1000 --verify 1 -M 4096 -K 4096
 else
-    ${BIND_CMD} ./gemv_bench --verify 1 -M 11264 -K 4096
-    ${BIND_CMD} ./gemv_bench --verify 1 -M 4096 -K 11264
-    ${BIND_CMD} ./gemv_bench --verify 1 -M 12288 -K 4096
-    ${BIND_CMD} ./gemv_bench --verify 1 -M 4096 -K 4096
+    ${BIND_CMD} ./gemv_bench --warmups 100 --loops 2000 --verify 1 -M 11264 -K 4096
+    ${BIND_CMD} ./gemv_bench --warmups 100 --loops 2000 --verify 1 -M 4096 -K 11264
+    ${BIND_CMD} ./gemv_bench --warmups 100 --loops 2000 --verify 1 -M 12288 -K 4096
+    ${BIND_CMD} ./gemv_bench --warmups 100 --loops 2000 --verify 1 -M 4096 -K 4096
 fi