Update GEMV benchmarks

evo2/bm/Makefile

@@ -3,7 +3,7 @@ CXXFLAGS ?= -std=c++17 -O3
 OFFLOAD_ARCH ?= gfx936
 
 TARGET := gemv_bench
-SRC := gemv_bf16.hip
+SRC := gemv_bf16.cpp
 DEP := gemv_utils.h
 
 .PHONY: all clean

evo2/bm/gemv_utils.h

@@ -37,7 +37,8 @@ inline char *getCmdOption(char **begin, char **end, const std::string &option) {
 // --------------------------------------------------------------------------------
 
 inline void gemv_cpu(int M, int K, float alpha, const hip_bfloat16 *h_A,
-                     int lda, const hip_bfloat16 *h_x, hip_bfloat16 *h_y) {
+                     int lda, const hip_bfloat16 *h_x, float beta,
+                     hip_bfloat16 *h_y) {
   for (int m = 0; m < M; ++m) {
     float sum = 0.0f;
     for (int k = 0; k < K; ++k) {
@@ -45,7 +46,7 @@ inline void gemv_cpu(int M, int K, float alpha, const hip_bfloat16 *h_A,
       float val_x = static_cast<float>(h_x[k]);
       sum += val_a * val_x;
     }
-    h_y[m] = hip_bfloat16(alpha * sum);
+    h_y[m] = hip_bfloat16(alpha * sum + beta * h_y[m]);
   }
 
   return;
@@ -85,9 +86,9 @@ inline bool verify_result(int M, const hip_bfloat16 *h_y_gpu,
 // --------------------------------------------------------------------------------
 
 // 定义统一的 Kernel Launcher 签名
-using KernelLauncher =
-    std::function<void(int M, int K, float alpha, const hip_bfloat16 *A,
-                       int lda, const hip_bfloat16 *x, hip_bfloat16 *y)>;
+using KernelLauncher = std::function<void(
+    int M, int K, float alpha, const hip_bfloat16 *A, int lda,
+    const hip_bfloat16 *x, float beta, hip_bfloat16 *y)>;
 
 struct KernelCase {
   std::string name;
@@ -96,7 +97,7 @@ struct KernelCase {
 
 inline void run_benchmark(const std::vector<KernelCase> &cases, int M, int K,
                           float alpha, const hip_bfloat16 *A, int lda,
-                          const hip_bfloat16 *x, hip_bfloat16 *y,
+                          const hip_bfloat16 *x, float beta, hip_bfloat16 *y,
                           bool do_verify) {
 
   std::cout << "GEMV Benchmarks" << std::endl;
@@ -120,7 +121,7 @@ inline void run_benchmark(const std::vector<KernelCase> &cases, int M, int K,
                              hipMemcpyDeviceToHost));
 
     // 计算 CPU Reference
-    gemv_cpu(M, K, alpha, h_A.data(), lda, h_x.data(), h_y_ref.data());
+    gemv_cpu(M, K, alpha, h_A.data(), lda, h_x.data(), beta, h_y_ref.data());
   }
 
   // 列宽
@@ -143,7 +144,7 @@ inline void run_benchmark(const std::vector<KernelCase> &cases, int M, int K,
       checkHipErrors(hipMemset(y, 0, M * sizeof(hip_bfloat16)));
 
       // 运行一次
-      k.func(M, K, alpha, A, lda, x, y);
+      k.func(M, K, alpha, A, lda, x, beta, y);
       checkHipErrors(hipDeviceSynchronize());
 
       // 拷回结果
@@ -159,7 +160,7 @@ inline void run_benchmark(const std::vector<KernelCase> &cases, int M, int K,
 
     // 2. Warmup
     for (int i = 0; i < 100; ++i) {
-      k.func(M, K, alpha, A, lda, x, y);
+      k.func(M, K, alpha, A, lda, x, beta, y);
     }
     checkHipErrors(hipDeviceSynchronize());
 
@@ -167,7 +168,7 @@ inline void run_benchmark(const std::vector<KernelCase> &cases, int M, int K,
     int num_runs = 1000;
     checkHipErrors(hipEventRecord(start));
     for (int i = 0; i < num_runs; ++i) {
-      k.func(M, K, alpha, A, lda, x, y);
+      k.func(M, K, alpha, A, lda, x, beta, y);
     }
     checkHipErrors(hipEventRecord(stop));
     checkHipErrors(hipEventSynchronize(stop));
@@ -184,8 +185,8 @@ inline void run_benchmark(const std::vector<KernelCase> &cases, int M, int K,
     double bytes_moved = (double)(M * K + K + M) * sizeof(hip_bfloat16);
     double bw = bytes_moved / (avg_ms * 1e-3) / 1e9;
 
-    printf("%-38s %10.1f %10.2f %10.2f %8s\n", k.name.c_str(), avg_ms * 1e3, gflops,
-           bw, result_status.c_str());
+    printf("%-38s %10.1f %10.2f %10.2f %8s\n", k.name.c_str(), avg_ms * 1e3,
+           gflops, bw, result_status.c_str());
   }
 
   std::cout << std::string(w_table, '-') << std::endl;

evo2/bm/run-all.sh

 #!/bin/bash
+# BW150
+export HIP_VISIBLE_DEVICES=1
+BIND_CMD="numactl -N 0 -m 0"
 
 make
 
-# BW150
-export HIP_VISIBLE_DEVICES=4
-
-hipprof numactl -N 0 -m 0 ./gemv_bench --verify 1 -M 11264 -K 4096
-hipprof numactl -N 0 -m 0 ./gemv_bench --verify 1 -M 4096 -K 11264
-hipprof numactl -N 0 -m 0 ./gemv_bench --verify 1 -M 12288 -K 4096
-hipprof numactl -N 0 -m 0 ./gemv_bench --verify 1 -M 4096 -K 4096
\ No newline at end of file
+if [[ "$*" == *"--trace"* ]]; then
+    PROF_CMD="hipprof --trace-off --pmc"
+    ${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
+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
+fi