From 115466353dbebac55dd91ec041ddff4e14a26d16 Mon Sep 17 00:00:00 2001 From: FallenSigh Date: Fri, 26 Jun 2026 00:34:44 +0800 Subject: [PATCH] Remove timing instrumentation from user-sample.c --- sdk/software/examples/asm/user-sample.c | 16 +--------------- 1 file changed, 1 insertion(+), 15 deletions(-) diff --git a/sdk/software/examples/asm/user-sample.c b/sdk/software/examples/asm/user-sample.c index ce5a566..1dfef2c 100644 --- a/sdk/software/examples/asm/user-sample.c +++ b/sdk/software/examples/asm/user-sample.c @@ -13,7 +13,6 @@ #include #include -#include #include /* ─── Configurable group count (default 10 for fast local testing) ─── */ @@ -23,8 +22,6 @@ /* ─── BSP globals ─────────────────────────────────────────────────── */ unsigned long UART_BASE = 0xbf000000; -unsigned long CONFREG_TIMER_BASE = 0xbf20f100; -unsigned long CONFREG_CLOCKS_PER_SEC = 50000000L; unsigned long CORE_CLOCKS_PER_SEC = 33000000L; /* ─── Memory layout ────────────────────────────────────────────────── */ @@ -111,7 +108,7 @@ int main(void) if (MATMUL_GROUP_NUM <= 0) return 0; volatile uint32_t *extram = (volatile uint32_t *)EXTRAM_BASE_UNCACHED; - uint32_t crc, t_prologue, t_mainloop, t_epilogue; + uint32_t crc; int g; // Ping-Pong Register Address Arrays (Physical for DMA) @@ -129,17 +126,14 @@ int main(void) REG_WR(MATMUL_CRC, 0); // 复位硬件 CRC32 为 0xFFFFFFFF /* ── 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) @@ -168,28 +162,20 @@ 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; /* ── Read hardware CRC32 result ──────────────────────────── */ crc = REG_RD(MATMUL_CRC); /* ── Final markers ────────────────────────────────────────── */ printf("MATMUL_CRC32=%08lx\n", (unsigned long)crc); - 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_TOTAL=%lu\n", (unsigned long)(t_prologue + t_mainloop + t_epilogue)); printf("MATMUL_DONE\n"); return 0;