Add RP2040.memcpyDMA for DMA-managed memory copies

RP2040::memcpyDMA implements a DMA-controlled memory copy call identical in
function to standard memcpy, but using an onboard DMA engine.  For large
memory transfers this can be significantly faster than using the CPU-based
memcpy.  Only 4-byte aligned source, destination, and counts are allowed.
If any inputs are not 4-byte aligned, then standard memcpy will occur so
it will behave correctly for any inputs.

Author: earlephilhower

cores/rp2040/RP2040Support.h

@@ -27,6 +27,7 @@
 #include <hardware/structs/rosc.h>
 #include <hardware/structs/systick.h>
 #include <pico/multicore.h>
+#include <hardware/dma.h>
 #include <pico/rand.h>
 #include <pico/util/queue.h>
 #include <pico/bootrom.h>
@@ -311,6 +312,36 @@ class RP2040 {
         return id;
     }
 
+#pragma GCC push_options
+#pragma GCC optimize ("Os")
+    void *memcpyDMA(void *dest, const void *src, size_t n) {
+        // Allocate a DMA channel on 1st call, reuse it every call after
+        if (memcpyDMAChannel < 1) {
+            memcpyDMAChannel = dma_claim_unused_channel(true);
+            dma_channel_config c = dma_channel_get_default_config(memcpyDMAChannel);
+            channel_config_set_transfer_data_size(&c, DMA_SIZE_32);
+            channel_config_set_read_increment(&c, true);
+            channel_config_set_write_increment(&c, true);
+            channel_config_set_irq_quiet(&c, true);
+            dma_channel_set_config(memcpyDMAChannel, &c, false);
+        }
+        // If there's any misalignment or too small, use regular memcpy which can handle it
+        if ((n < 64) || (((uint32_t)dest) | ((uint32_t)src) | n) & 3) {
+            return memcpy(dest, src, n);
+        }
+
+        int words = n / 4;
+        dma_channel_set_read_addr(memcpyDMAChannel, src, false);
+        dma_channel_set_write_addr(memcpyDMAChannel, dest, false);
+        dma_channel_set_trans_count(memcpyDMAChannel, words, false);
+        dma_channel_start(memcpyDMAChannel);
+        while (dma_channel_is_busy(memcpyDMAChannel)) {
+            /* busy wait dma */
+        }
+        return dest;
+    }
+#pragma GCC pop_options
+
     // Multicore comms FIFO
     _MFIFO fifo;
 
@@ -336,4 +367,5 @@ class RP2040 {
     PIO _pio;
     int _sm;
     PIOProgram *_ccountPgm;
+    int memcpyDMAChannel = -1;
 };

docs/rp2040.rst

@@ -99,3 +99,13 @@ bootloader.
 void rp2040.rebootToBootloader()
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 Will reboot the RP2040 into USB UF2 upload mode.
+
+DMA-based MEMCPY
+----------------
+The onboard DMA engines can copy 4-byte aligned quantities faster than the CPU.
+
+void \*rp2040.memcpyDMA(void \*dest, const void \*src, size_t count);
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Uses a DMA engine to transfer data from src to dest in 4-byte chunks without CPU
+intervention.  If any arguments are not 4-byte aligned, or if the count is not a
+multiple of 4, then it will fall back to CPU-managed memcpy.

libraries/rp2040/examples/DMAMemcpy/DMAMemcpy.ino

@@ -0,0 +1,51 @@
+// Simple speed and functionality test for DMA memcpy
+// Released to the public domain by Earle F. Philhower, III, 2024
+
+
+uint32_t src[1024];
+uint32_t dest[1024];
+
+void setup() {
+  // Set up a simple test pattern in src, verify after each pass
+  srand(0xc0de);
+  for (int i = 0; i < 1024; i++) {
+    src[i] = ((~i) & 1024) | ((i) << 22);
+  }
+}
+
+void verify(const char *name, uint32_t *src) {
+  for (int i = 0; i < 1024; i++) {
+    uint32_t expected = ((~i) & 1024) | ((i) << 22);
+    if (expected != src[i]) {
+      Serial.printf("ERROR, mismatch @ %d on %s memcpy\n", i, name);
+      while (true) {
+        // Idle forever, this is fatal!
+      }
+    }
+  }
+}
+
+void loop() {
+  uint64_t start, stop;
+
+  start = rp2040.getCycleCount();
+  for (int i = 0; i < 1000; i++) {
+    memcpy(dest, src, 4 * 1024);
+    memcpy(src, dest, 4 * 1024);
+  }
+  stop = rp2040.getCycleCount();
+  verify("CPU", src);
+  verify("CPU", dest);
+  Serial.printf("CPU: %lld clock cycles for 4K\n", (stop - start) / 1000);
+
+  start = rp2040.getCycleCount();
+  for (int i = 0; i < 1000; i++) {
+    rp2040.memcpyDMA(dest, src, 4 * 1024);
+    rp2040.memcpyDMA(src, dest, 4 * 1024);
+  }
+  stop = rp2040.getCycleCount();
+  verify("DMA", src);
+  verify("DMA", dest);
+  Serial.printf("DMA: %lld clock cycles for 4K\n\n\n", (stop - start) / 1000);
+  delay(1000);
+}