Improve time_critical test

More DMA, and check the cycles after - now works on RP2040 and RP2350

test/pico_xip_sram_test/CMakeLists.txt

@@ -1,19 +1,13 @@
 # XIP SRAM only binary
 add_executable(pico_xip_sram_test pico_xip_sram_test.c)
-target_link_libraries(pico_xip_sram_test PRIVATE pico_stdlib pico_multicore hardware_dma)
+target_link_libraries(pico_xip_sram_test PRIVATE pico_stdlib)
 pico_set_binary_type(pico_xip_sram_test xip_sram)
 pico_minimize_runtime(pico_xip_sram_test INCLUDE PRINTF PRINTF_MINIMAL DEFAULT_ALARM_POOL PANIC)
 target_compile_definitions(pico_xip_sram_test PRIVATE PICO_HEAP_SIZE=0x200)
 pico_add_extra_outputs(pico_xip_sram_test)
 
 # Use XIP SRAM for time critical functions
-add_executable(pico_critical_xip_sram_test pico_xip_sram_test.c)
+add_executable(pico_critical_xip_sram_test pico_critical_xip_sram_test.c)
 target_link_libraries(pico_critical_xip_sram_test PRIVATE pico_stdlib pico_multicore hardware_dma)
 pico_set_binary_type(pico_critical_xip_sram_test copy_to_ram_using_xip_ram)
 pico_add_extra_outputs(pico_critical_xip_sram_test)
-
-# Same as above, but without XIP SRAM, to show the difference
-add_executable(pico_critical_no_xip_sram_test pico_xip_sram_test.c)
-target_link_libraries(pico_critical_no_xip_sram_test PRIVATE pico_stdlib pico_multicore hardware_dma)
-pico_set_binary_type(pico_critical_no_xip_sram_test copy_to_ram)
-pico_add_extra_outputs(pico_critical_no_xip_sram_test)

test/pico_xip_sram_test/pico_critical_xip_sram_test.c

@@ -0,0 +1,106 @@
+#include <stdio.h>
+#include "pico/stdlib.h"
+#include "pico/multicore.h"
+#include "hardware/dma.h"
+#include "hardware/structs/systick.h"
+#include "hardware/structs/busctrl.h"
+
+
+int __time_critical_func(test_func_xip)(void) {
+    systick_hw->rvr = 0x00ffffff;
+    systick_hw->cvr = 0;
+
+    volatile uint32_t i = 0;
+    i += 4;
+    i += i;
+
+    return systick_hw->rvr - systick_hw->cvr;
+}
+
+int __not_in_flash_func(test_func_sram)(void) {
+    systick_hw->rvr = 0x00ffffff;
+    systick_hw->cvr = 0;
+
+    volatile uint32_t i = 0;
+    i += 4;
+    i += i;
+
+    return systick_hw->rvr - systick_hw->cvr;
+}
+
+
+void core1_entry() {
+    // Just read memory repeatedly
+    pico_default_asm_volatile(
+        "1:\n"
+        "ldr r0, =%0\n"
+        "ldmia r0!, {r1-r4}\n"
+        "ldmia r0!, {r1-r4}\n"
+        "ldmia r0!, {r1-r4}\n"
+        "ldmia r0!, {r1-r4}\n"
+        "ldmia r0!, {r1-r4}\n"
+        "ldmia r0!, {r1-r4}\n"
+        "ldmia r0!, {r1-r4}\n"
+        "ldmia r0!, {r1-r4}\n"
+        "b 1b\n"
+        : : "i" (SRAM_BASE) : "r0", "r1", "r2", "r3", "r4"
+    );
+}
+
+
+void trigger_dma(void) {
+    int dat[8];
+    for (int i = 0; i < count_of(dat); i++) {
+        int chan = dma_claim_unused_channel(true);
+        dma_channel_config c = dma_channel_get_default_config(chan);
+        channel_config_set_transfer_data_size(&c, DMA_SIZE_32);
+        channel_config_set_read_increment(&c, true);
+        channel_config_set_write_increment(&c, true);
+        uint32_t from = SRAM_BASE;
+        uint32_t to = SRAM_BASE + ((SRAM4_BASE - SRAM_BASE) / 2);
+        uint32_t size = ((SRAM4_BASE - SRAM_BASE) / 2) / 4;
+        dma_channel_configure(chan, &c, (uint32_t*)to, (uint32_t*)from, size, true);
+        dat[i] = chan;
+    }
+    for (int i = 0; i < count_of(dat); i++) {
+        dma_channel_unclaim(dat[i]);
+    }
+}
+
+
+int main(void) {
+    stdio_init_all();
+    printf("pico_xip_sram_test begins\n");
+
+    multicore_launch_core1(core1_entry);
+
+    systick_hw->csr = 0x4 | 0x1; // clock source and enable
+
+    // Give core1 and DMA high priority
+    hw_set_bits(&busctrl_hw->priority, BUSCTRL_BUS_PRIORITY_PROC1_BITS | BUSCTRL_BUS_PRIORITY_DMA_R_BITS | BUSCTRL_BUS_PRIORITY_DMA_W_BITS);
+    hw_clear_bits(&busctrl_hw->priority, BUSCTRL_BUS_PRIORITY_PROC0_BITS);
+
+    int test_func_xip_cycles = 0;
+    int test_func_sram_cycles = 0;
+    for (int i = 0; i < 5; i++) {
+        printf("running... %d\n", i);
+        trigger_dma();
+        int tmp = test_func_xip();
+        test_func_xip_cycles += tmp;
+        printf("test_func_xip: %d\n", tmp);
+        tmp = test_func_sram();
+        test_func_sram_cycles += tmp;
+        printf("test_func_sram: %d\n", tmp);
+        sleep_ms(500);
+    }
+
+    if (test_func_xip_cycles >= test_func_sram_cycles) {
+        printf("ERROR: test_func_xip_cycles (%d) >= test_func_sram_cycles (%d)\n", test_func_xip_cycles, test_func_sram_cycles);
+        return 1;
+    } else {
+        printf("SUCCESS: test_func_xip_cycles (%d) < test_func_sram_cycles (%d)\n", test_func_xip_cycles, test_func_sram_cycles);
+    }
+
+    printf("pico_xip_sram_test ends\n");
+    return 0;
+}

test/pico_xip_sram_test/pico_xip_sram_test.c

@@ -1,72 +1,13 @@
 #include <stdio.h>
 #include "pico/stdlib.h"
-#include "pico/multicore.h"
-#include "hardware/dma.h"
-#include "hardware/structs/systick.h"
-#include "hardware/structs/busctrl.h"
-
-
-int __time_critical_func(test_func)(void) {
-    systick_hw->rvr = 0x00ffffff;
-    systick_hw->cvr = 0;
-
-    volatile uint32_t i = 0;
-    i += 4;
-    i += i;
-
-    return systick_hw->rvr - systick_hw->cvr;
-}
-
-
-void core1_entry() {
-    // Just read memory repeatedly
-    pico_default_asm_volatile(
-        "1:\n"
-        "ldr r0, =%0\n"
-        "ldmia r0!, {r1-r4}\n"
-        "ldmia r0!, {r1-r4}\n"
-        "ldmia r0!, {r1-r4}\n"
-        "ldmia r0!, {r1-r4}\n"
-        "ldmia r0!, {r1-r4}\n"
-        "ldmia r0!, {r1-r4}\n"
-        "ldmia r0!, {r1-r4}\n"
-        "ldmia r0!, {r1-r4}\n"
-        "b 1b\n"
-        : : "i" (SRAM_BASE) : "r0", "r1", "r2", "r3", "r4"
-    );
-}
-
-
-void trigger_dma(void) {
-    int chan = dma_claim_unused_channel(true);
-    dma_channel_config c = dma_channel_get_default_config(chan);
-    channel_config_set_transfer_data_size(&c, DMA_SIZE_32);
-    channel_config_set_read_increment(&c, true);
-    channel_config_set_write_increment(&c, true);
-    uint32_t from = SRAM_BASE;
-    uint32_t to = SRAM_BASE + ((SRAM_STRIPED_END - SRAM_BASE) / 2);
-    uint32_t size = ((SRAM_STRIPED_END - SRAM_BASE) / 2) / 4;
-    dma_channel_configure(chan, &c, (uint32_t*)to, (uint32_t*)from, size, true);
-    dma_channel_unclaim(chan);
-}
 
 
 int main(void) {
     stdio_init_all();
     printf("pico_xip_sram_test begins\n");
 
-    multicore_launch_core1(core1_entry);
-
-    systick_hw->csr = 0x4 | 0x1; // clock source and enable
-
-    // Give core1 and DMA high priority
-    hw_set_bits(&busctrl_hw->priority, BUSCTRL_BUS_PRIORITY_PROC1_BITS | BUSCTRL_BUS_PRIORITY_DMA_R_BITS | BUSCTRL_BUS_PRIORITY_DMA_W_BITS);
-    hw_clear_bits(&busctrl_hw->priority, BUSCTRL_BUS_PRIORITY_PROC0_BITS);
-
     for (int i = 0; i < 5; i++) {
         printf("running... %d\n", i);
-        trigger_dma();
-        printf("test_func: %d\n", test_func());
         sleep_ms(500);
     }