From d9483a9d39e6150615a1737d341e42d1a4125697 Mon Sep 17 00:00:00 2001 From: FallenSigh Date: Thu, 25 Jun 2026 20:10:31 +0800 Subject: [PATCH] Add timing instrumentation for profiling (prologue/loop/epilogue/crc32) --- sdk/software/examples/asm/user-sample.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-) diff --git a/sdk/software/examples/asm/user-sample.c b/sdk/software/examples/asm/user-sample.c index 84bdd42..b9d0bdc 100644 --- a/sdk/software/examples/asm/user-sample.c +++ b/sdk/software/examples/asm/user-sample.c @@ -160,7 +160,7 @@ int main(void) if (MATMUL_GROUP_NUM <= 0) return 0; volatile uint32_t *extram = (volatile uint32_t *)EXTRAM_BASE_UNCACHED; - uint32_t crc; + uint32_t crc, t_prologue, t_mainloop, t_epilogue, t_crc32; int g; // Ping-Pong Register Address Arrays (Physical for DMA) @@ -177,14 +177,17 @@ int main(void) REG_WR(MATMUL_STATUS, STATUS_DONE | STATUS_ERROR); /* ── Prologue: 第0组数据初始化 ────────────────────────── */ + t_prologue = get_ns(); // DMA Ch0: 从 ExtRAM 将 Group 0 的 A、B (128 bytes) 搬入 IP 的 Bank 0 dma_start_transfer(0, EXTRAM_PHYS_AB(0), phys_bank_AB[0], 128, 10); wait_dma_done_intr(0); // 启动 IP Bank 0 计算 REG_WR(MATMUL_CTRL, ctrl_start[0]); + t_prologue = get_ns() - t_prologue; /* ── Main Loop: 硬件完全流水线 (Overlapped Execution) ──── */ + t_mainloop = get_ns(); // 在 g 循环中: // IP 正在计算 prev_bank (g-1) // DMA_CH0 正在加载 cur_bank (g) @@ -213,21 +216,33 @@ int main(void) // 6. 立即启动当前组在 IP 中的计算 REG_WR(MATMUL_CTRL, ctrl_start[cur_bank]); } + t_mainloop = get_ns() - t_mainloop; /* ── Epilogue: 收尾最后一组 ───────────────────────────── */ + t_epilogue = get_ns(); wait_matmul_done_intr(); // 等待最后一次计算完成 wait_dma_done_intr(1); // 确保之前的 DMA CH1 空闲 // DMA CH1 将最后一组结果写回 dma_start_transfer(1, phys_bank_C[(MATMUL_GROUP_NUM - 1) % 2], EXTRAM_PHYS_C(MATMUL_GROUP_NUM - 1), 192, 20); wait_dma_done_intr(1); + t_epilogue = get_ns() - t_epilogue; /* ── CRC32 over result region ─────────────────────────────── */ + t_crc32 = get_ns(); crc = compute_crc32(extram); + t_crc32 = get_ns() - t_crc32; /* ── Final markers ────────────────────────────────────────── */ printf("MATMUL_CRC32=%08lx\n", (unsigned long)crc); printf("MATMUL_DONE\n"); + printf("TIME_PROLOGUE=%lu\n", (unsigned long)t_prologue); + printf("TIME_LOOP=%lu\n", (unsigned long)t_mainloop); + printf("TIME_LOOP_PER=%lu\n", (unsigned long)(t_mainloop / MATMUL_GROUP_NUM)); + printf("TIME_EPILOGUE=%lu\n", (unsigned long)t_epilogue); + printf("TIME_CRC32=%lu\n", (unsigned long)t_crc32); + printf("TIME_TOTAL=%lu\n", (unsigned long)(t_prologue + t_mainloop + t_epilogue + t_crc32)); + return 0; } \ No newline at end of file