Merge branch 'feature/psram'

malcolm-sparkfun · malcolm-sparkfun · commit 87bc47d7404c · 2025-03-04T10:07:41.000-07:00
diff --git a/ports/rp2/CMakeLists.txt b/ports/rp2/CMakeLists.txt
@@ -173,6 +173,12 @@ set(MICROPY_SOURCE_PORT
     ${CMAKE_BINARY_DIR}/pins_${MICROPY_BOARD}.c
 )
 
+if(PICO_RP2350)
+    list(APPEND MICROPY_SOURCE_PORT
+        rp2_psram.c
+    )
+endif()
+
 set(MICROPY_SOURCE_QSTR
     ${MICROPY_SOURCE_PY}
     ${MICROPY_DIR}/shared/readline/readline.c
diff --git a/ports/rp2/main.c b/ports/rp2/main.c
@@ -26,6 +26,8 @@
 
 #include <stdio.h>
 
+#include "rp2_psram.h"
+#include "rp2_flash.h"
 #include "py/compile.h"
 #include "py/cstack.h"
 #include "py/runtime.h"
@@ -93,6 +95,13 @@ int main(int argc, char **argv) {
     // Hook for setting up anything that needs to be super early in the boot-up process.
     MICROPY_BOARD_STARTUP();
 
+    // Set the flash divisor to an appropriate value
+    rp2_flash_set_timing();
+
+    #if defined(MICROPY_HW_PSRAM_CS_PIN) && MICROPY_HW_ENABLE_PSRAM
+    size_t psram_size = psram_init(MICROPY_HW_PSRAM_CS_PIN);
+    #endif
+
     #if MICROPY_HW_ENABLE_UART_REPL
     bi_decl(bi_program_feature("UART REPL"))
     setup_default_uart();
@@ -120,7 +129,21 @@ int main(int argc, char **argv) {
 
     // Initialise stack extents and GC heap.
     mp_cstack_init_with_top(&__StackTop, &__StackTop - &__StackBottom);
+
+    #if defined(MICROPY_HW_PSRAM_CS_PIN) && MICROPY_HW_ENABLE_PSRAM
+    if (psram_size) {
+        #if MICROPY_GC_SPLIT_HEAP
+        gc_init(&__GcHeapStart, &__GcHeapEnd);
+        gc_add((void *)PSRAM_LOCATION, (void *)(PSRAM_LOCATION + psram_size));
+        #else
+        gc_init((void *)PSRAM_LOCATION, (void *)(PSRAM_LOCATION + psram_size));
+        #endif
+    } else {
+        gc_init(&__GcHeapStart, &__GcHeapEnd);
+    }
+    #else
     gc_init(&__GcHeapStart, &__GcHeapEnd);
+    #endif
 
     #if MICROPY_PY_LWIP
     // lwIP doesn't allow to reinitialise itself by subsequent calls to this function
diff --git a/ports/rp2/modmachine.c b/ports/rp2/modmachine.c
@@ -31,6 +31,8 @@
 #include "mp_usbd.h"
 #include "modmachine.h"
 #include "uart.h"
+#include "rp2_psram.h"
+#include "rp2_flash.h"
 #include "clocks_extra.h"
 #include "hardware/pll.h"
 #include "hardware/structs/rosc.h"
@@ -94,6 +96,11 @@ static mp_obj_t mp_machine_get_freq(void) {
 
 static void mp_machine_set_freq(size_t n_args, const mp_obj_t *args) {
     mp_int_t freq = mp_obj_get_int(args[0]);
+
+    // If necessary, increase the flash divider before increasing the clock speed
+    const int old_freq = clock_get_hz(clk_sys);
+    rp2_flash_set_timing_for_freq(MAX(freq, old_freq));
+
     if (!set_sys_clock_khz(freq / 1000, false)) {
         mp_raise_ValueError(MP_ERROR_TEXT("cannot change frequency"));
     }
@@ -111,10 +118,19 @@ static void mp_machine_set_freq(size_t n_args, const mp_obj_t *args) {
             }
         }
     }
+
+    // If clock speed was reduced, maybe we can reduce the flash divider
+    if (freq < old_freq) {
+        rp2_flash_set_timing_for_freq(freq);
+    }
+
     #if MICROPY_HW_ENABLE_UART_REPL
     setup_default_uart();
     mp_uart_init();
     #endif
+    #if defined(MICROPY_HW_PSRAM_CS_PIN) && MICROPY_HW_ENABLE_PSRAM
+    psram_init(MICROPY_HW_PSRAM_CS_PIN);
+    #endif
 }
 
 static void mp_machine_idle(void) {
diff --git a/ports/rp2/mpconfigport.h b/ports/rp2/mpconfigport.h
@@ -71,8 +71,20 @@
 #define MICROPY_CONFIG_ROM_LEVEL                (MICROPY_CONFIG_ROM_LEVEL_EXTRA_FEATURES)
 #endif
 
+#ifndef MICROPY_HW_ENABLE_PSRAM
+#define MICROPY_HW_ENABLE_PSRAM (0)
+#endif
+
 // Memory allocation policies
+#if MICROPY_HW_ENABLE_PSRAM
+#define MICROPY_GC_STACK_ENTRY_TYPE             uint32_t
+#define MICROPY_ALLOC_GC_STACK_SIZE             (1024) // Avoid slowdown when GC stack overflow causes a full sweep of PSRAM-backed heap
+#else
 #define MICROPY_GC_STACK_ENTRY_TYPE             uint16_t
+#endif
+#ifndef MICROPY_GC_SPLIT_HEAP
+#define MICROPY_GC_SPLIT_HEAP                   MICROPY_HW_ENABLE_PSRAM // whether PSRAM is added to or replaces the heap
+#endif
 #define MICROPY_ALLOC_PATH_MAX                  (128)
 #define MICROPY_QSTR_BYTES_IN_HASH              (1)
 
diff --git a/ports/rp2/rp2_flash.c b/ports/rp2/rp2_flash.c
@@ -26,12 +26,19 @@
 
 #include <string.h>
 
+#include "rp2_psram.h"
 #include "py/mphal.h"
 #include "py/runtime.h"
 #include "extmod/vfs.h"
 #include "modrp2.h"
 #include "hardware/flash.h"
 #include "pico/binary_info.h"
+#ifdef PICO_RP2350
+#include "hardware/structs/ioqspi.h"
+#include "hardware/structs/qmi.h"
+#else
+#include "hardware/structs/ssi.h"
+#endif
 
 #define BLOCK_SIZE_BYTES (FLASH_SECTOR_SIZE)
 
@@ -80,16 +87,75 @@ static bool use_multicore_lockout(void) {
     ;
 }
 
+// Function to set the flash divisor to the correct divisor, assumes interrupts disabled
+// and core1 locked out if relevant.
+static void __no_inline_not_in_flash_func(rp2_flash_set_timing_internal)(int clock_hz) {
+
+    // Use the minimum divisor assuming a 133MHz flash.
+    const int max_flash_freq = 133000000;
+    int divisor = (clock_hz + max_flash_freq - 1) / max_flash_freq;
+
+    #if PICO_RP2350
+    // Make sure flash is deselected - QMI doesn't appear to have a busy flag(!)
+    while ((ioqspi_hw->io[1].status & IO_QSPI_GPIO_QSPI_SS_STATUS_OUTTOPAD_BITS) != IO_QSPI_GPIO_QSPI_SS_STATUS_OUTTOPAD_BITS) {
+        ;
+    }
+
+    // RX delay equal to the divisor means sampling at the same time as the next falling edge of SCK after the
+    // falling edge that generated the data.  This is pretty tight at 133MHz but seems to work with the Winbond flash chips.
+    const int rxdelay = divisor;
+    qmi_hw->m[0].timing = (1 << QMI_M0_TIMING_COOLDOWN_LSB) |
+        rxdelay << QMI_M1_TIMING_RXDELAY_LSB |
+        divisor << QMI_M1_TIMING_CLKDIV_LSB;
+
+    // Force a read through XIP to ensure the timing is applied
+    volatile uint32_t *ptr = (volatile uint32_t *)0x14000000;
+    (void)*ptr;
+    #else
+    // RP2040 SSI hardware only supports even divisors
+    if (divisor & 1) {
+        divisor += 1;
+    }
+
+    // Wait for SSI not busy
+    while (ssi_hw->sr & SSI_SR_BUSY_BITS) {
+        ;
+    }
+
+    // Disable, set the new divisor, and re-enable
+    hw_clear_bits(&ssi_hw->ssienr, SSI_SSIENR_SSI_EN_BITS);
+    ssi_hw->baudr = divisor;
+    hw_set_bits(&ssi_hw->ssienr, SSI_SSIENR_SSI_EN_BITS);
+    #endif
+}
+
 // Flash erase and write must run with interrupts disabled and the other core suspended,
 // because the XIP bit gets disabled.
 static uint32_t begin_critical_flash_section(void) {
     if (use_multicore_lockout()) {
         multicore_lockout_start_blocking();
     }
-    return save_and_disable_interrupts();
+    uint32_t state = save_and_disable_interrupts();
+
+    #if defined(MICROPY_HW_PSRAM_CS_PIN) && MICROPY_HW_ENABLE_PSRAM
+    // We're about to invalidate the XIP cache, clean it first to commit any dirty writes to PSRAM
+    // Use the upper 16k of the maintenance space (0x1bffc000 through 0x1bffffff) to workaround
+    // incorrect behaviour of the XIP clean operation, where it also alters the tag of the associated
+    // cache line: https://forums.raspberrypi.com/viewtopic.php?t=378249#p2263677
+    volatile uint8_t *maintenance_ptr = (volatile uint8_t *)(XIP_SRAM_BASE + (XIP_MAINTENANCE_BASE - XIP_BASE));
+    for (int i = 1; i < 16 * 1024; i += 8) {
+        maintenance_ptr[i] = 0;
+    }
+    #endif
+
+    return state;
 }
 
 static void end_critical_flash_section(uint32_t state) {
+    rp2_flash_set_timing_internal(clock_get_hz(clk_sys));
+    #if defined(MICROPY_HW_PSRAM_CS_PIN) && MICROPY_HW_ENABLE_PSRAM
+    psram_init(MICROPY_HW_PSRAM_CS_PIN);
+    #endif
     restore_interrupts(state);
     if (use_multicore_lockout()) {
         multicore_lockout_end_blocking();
@@ -155,11 +221,16 @@ static mp_obj_t rp2_flash_readblocks(size_t n_args, const mp_obj_t *args) {
 }
 static MP_DEFINE_CONST_FUN_OBJ_VAR_BETWEEN(rp2_flash_readblocks_obj, 3, 4, rp2_flash_readblocks);
 
+static inline size_t min_size(size_t a, size_t b) {
+    return a < b ? a : b;
+}
+
 static mp_obj_t rp2_flash_writeblocks(size_t n_args, const mp_obj_t *args) {
     rp2_flash_obj_t *self = MP_OBJ_TO_PTR(args[0]);
     uint32_t offset = mp_obj_get_int(args[1]) * BLOCK_SIZE_BYTES;
     mp_buffer_info_t bufinfo;
     mp_get_buffer_raise(args[2], &bufinfo, MP_BUFFER_READ);
+
     if (n_args == 3) {
         mp_uint_t atomic_state = begin_critical_flash_section();
         flash_range_erase(self->flash_base + offset, bufinfo.len);
@@ -169,10 +240,31 @@ static mp_obj_t rp2_flash_writeblocks(size_t n_args, const mp_obj_t *args) {
     } else {
         offset += mp_obj_get_int(args[3]);
     }
-    mp_uint_t atomic_state = begin_critical_flash_section();
-    flash_range_program(self->flash_base + offset, bufinfo.buf, bufinfo.len);
-    end_critical_flash_section(atomic_state);
-    mp_event_handle_nowait();
+
+    if ((uintptr_t)bufinfo.buf >= SRAM_BASE) {
+        mp_uint_t atomic_state = begin_critical_flash_section();
+        flash_range_program(self->flash_base + offset, bufinfo.buf, bufinfo.len);
+        end_critical_flash_section(atomic_state);
+        mp_event_handle_nowait();
+    } else {
+        size_t bytes_left = bufinfo.len;
+        size_t bytes_offset = 0;
+        static uint8_t copy_buffer[BLOCK_SIZE_BYTES] = {0};
+
+        while (bytes_left) {
+            memcpy(copy_buffer, bufinfo.buf + bytes_offset, min_size(bytes_left, BLOCK_SIZE_BYTES));
+            mp_uint_t atomic_state = begin_critical_flash_section();
+            flash_range_program(self->flash_base + offset + bytes_offset, copy_buffer, min_size(bytes_left, BLOCK_SIZE_BYTES));
+            end_critical_flash_section(atomic_state);
+            bytes_offset += BLOCK_SIZE_BYTES;
+            if (bytes_left <= BLOCK_SIZE_BYTES) {
+                break;
+            }
+            bytes_left -= BLOCK_SIZE_BYTES;
+            mp_event_handle_nowait();
+        }
+    }
+
     // TODO check return value
     return mp_const_none;
 }
@@ -220,3 +312,23 @@ MP_DEFINE_CONST_OBJ_TYPE(
     make_new, rp2_flash_make_new,
     locals_dict, &rp2_flash_locals_dict
     );
+
+// Modify the flash timing.  Ensure flash access is suspended while
+// the timings are altered.
+void rp2_flash_set_timing_for_freq(int clock_hz) {
+    if (multicore_lockout_victim_is_initialized(1 - get_core_num())) {
+        multicore_lockout_start_blocking();
+    }
+    uint32_t state = save_and_disable_interrupts();
+
+    rp2_flash_set_timing_internal(clock_hz);
+
+    restore_interrupts(state);
+    if (multicore_lockout_victim_is_initialized(1 - get_core_num())) {
+        multicore_lockout_end_blocking();
+    }
+}
+
+void rp2_flash_set_timing() {
+    rp2_flash_set_timing_for_freq(clock_get_hz(clk_sys));
+}
diff --git a/ports/rp2/rp2_flash.h b/ports/rp2/rp2_flash.h
@@ -0,0 +1,7 @@
+#ifndef MICROPY_INCLUDED_RP2_MACHINE_FLASH_H
+#define MICROPY_INCLUDED_RP2_MACHINE_FLASH_H
+
+extern void rp2_flash_set_timing_for_freq(int clock_hz);
+extern void rp2_flash_set_timing();
+
+#endif
diff --git a/ports/rp2/rp2_psram.c b/ports/rp2/rp2_psram.c
diff --git a/ports/rp2/rp2_psram.h b/ports/rp2/rp2_psram.h
