Add transpose test

6 files changed, 280 insertions, 129 deletions

diff --git a/fw/.gitignore b/fw/.gitignore
index 02752e1..2d9d6e6 100644
--- a/fw/.gitignore
+++ b/fw/.gitignore
@@ -7,3 +7,4 @@
 *.bin
 *.hex
 *.pp
+transpose.elf
diff --git a/fw/Makefile b/fw/Makefile
index 6ff1686..8296708 100644
--- a/fw/Makefile
+++ b/fw/Makefile
@@ -34,6 +34,9 @@ LDFLAGS += -L$(CMSIS_PATH)/Lib/GCC -larm_cortexM0l_math
 
 all: main.elf main.pdf
 
+cmsis_exports.c: $(CMSIS_DEV_PATH)/Include/stm32f030x6.h $(CMSIS_PATH)/Include/core_cm0.h
+	python3 gen_cmsis_exports.py $^ > $@
+
 %.o: %.c
 	$(CC) -c $(CFLAGS) -o $@ $^
 	$(CC) -E $(CFLAGS) -o $(@:.o=.pp) $^
@@ -45,7 +48,7 @@ all: main.elf main.pdf
 %.dot: %.elf
 	r2 -a arm -qc 'aa;agC' $< 2>/dev/null >$@
 
-main.elf: main.o startup_stm32f030x6.o system_stm32f0xx.o $(HAL_PATH)/Src/stm32f0xx_ll_utils.o base.o semihosting.o
+main.elf: main.o startup_stm32f030x6.o system_stm32f0xx.o $(HAL_PATH)/Src/stm32f0xx_ll_utils.o base.o semihosting.o cmsis_exports.o transpose.o
 	$(CC) $(CFLAGS) $(LDFLAGS) -o $@ $^ $(LIBS)
 	$(OBJCOPY) -O ihex $@ $(@:.elf=.hex)
 	$(OBJCOPY) -O binary $@ $(@:.elf=.bin)
@@ -55,6 +58,13 @@ main.elf: main.o startup_stm32f030x6.o system_stm32f0xx.o $(HAL_PATH)/Src/stm32f
 program: main.elf openocd.cfg
 	openocd -f openocd.cfg -c "program $< verify reset exit"
 
+transpose.elf: transpose.c transpose_main.c
+	gcc -g -Wall -Wpedantic -o $@ $^
+
+.PHONY: transpose_test
+transpose_test: transpose.elf
+	./transpose.elf
+
 clean:
 	rm -f **.o
 	rm -f main.elf main.hex main.bin main.map main.lst
diff --git a/fw/main.c b/fw/main.c
index 092028b..aa01b6c 100644
--- a/fw/main.c
+++ b/fw/main.c
@@ -13,6 +13,8 @@
 #include <string.h>
 #include <unistd.h>
 
+#include "transpose.h"
+
 /* 
  * Part number: STM32F030F4C6
  */
@@ -43,28 +45,6 @@ void strobe_leds(void) {
 
 static volatile unsigned int sys_time = 0;
 
-enum Segment { SegA, SegB, SegC, SegD, SegE, SegF, SegG, SegDP, nsegments };
-enum {
-    nrows = 4,
-    ncols = 8,
-    nbits = 10,
-};
-enum {
-    frame_size_words = nrows*ncols*nsegments/32,
-};
-struct framebuf {
-    /* Multiplexing order: first Digits, then Time/bits, last Segments */
-    union {
-        uint32_t data[nbits*frame_size_words];
-        struct {
-            struct {
-                uint32_t data[frame_size_words];
-            } frame[nbits];
-        };
-    };
-    uint8_t brightness; /* 0 or 1; controls global brighntess control */
-};
-
 volatile struct framebuf fb[2] = {0};
 volatile struct framebuf *read_fb=fb+0, *write_fb=fb+1;
 volatile int led_state = 0;
@@ -78,48 +58,6 @@ volatile uint8_t this_addr = 0x05; /* FIXME */
 #define SR_ILED_HIGH 0x0080
 #define SR_ILED_LOW  0x0040
 
-void transpose_data(volatile uint8_t *rx_buf, volatile struct framebuf *out_fb) {
-    memset((uint8_t *)out_fb, 0, sizeof(*out_fb));
-    struct data_format {
-        union {
-            uint8_t high[8];
-            struct { uint8_t ah, bh, ch, dh, eh, fh, gh, dph; };
-        };
-        union {
-            uint16_t low;
-            struct { uint8_t dpl:2, gl:2, fl:2, el:2, dl:2, cl:2, bl:2, al:2; };
-        };
-    };
-    struct data_format *rxp = (struct data_format *)rx_buf;
-    for (int bit=0; bit<8; bit++) { /* bits */
-        uint32_t bit_mask = 1U<<bit;
-        volatile uint32_t *frame_data = out_fb->frame[bit+2].data;
-        uint8_t *start_inp = rxp->high;
-        for (volatile uint32_t *outp=frame_data; outp<frame_data+8; outp++) { /* segments */
-            uint32_t acc = 0;
-            uint8_t *inp = start_inp++;
-            for (int digit=0; digit<32; digit++) {
-                acc |= (*inp & bit_mask) >> bit << digit;
-                inp += sizeof(struct data_format);
-            }
-            *outp = acc;
-        }
-    }
-    for (int bit=0; bit<2; bit++) { /* bits */
-        volatile uint32_t *frame_data = out_fb->frame[bit].data;
-        uint16_t *inp = &rxp->low;
-        for (int seg=0; seg<8; seg++) { /* segments */
-            uint32_t mask = 1 << bit << seg;
-            uint32_t acc = 0;
-            for (int digit=0; digit<32; digit++) {
-                acc |= (*inp & mask) >> bit >> seg << digit;
-                inp += sizeof(struct data_format)/sizeof(uint16_t);
-            }
-            frame_data[seg] = acc;
-        }
-    }
-}
-
 void shift_aux(int global_current, int leds, int active_segment) {
     spi_send(
               (global_current ? SR_ILED_HIGH : SR_ILED_LOW)
@@ -128,6 +66,14 @@ void shift_aux(int global_current, int leds, int active_segment) {
     strobe_aux();
 }
 
+inline unsigned int stk_start() {
+    return SysTick->VAL;
+}
+
+inline unsigned int stk_end(unsigned int start) {
+    return (start - SysTick->VAL) & 0xffffff;
+}
+
 static volatile int frame_duration;
 /* returns new bit time in cycles */
 int shift_data() {
@@ -176,15 +122,16 @@ void cfg_timer3() {
     TIM3->EGR  |= TIM_EGR_UG;
     TIM3->CR1   = TIM_CR1_ARPE;
     TIM3->CR1  |= TIM_CR1_CEN;
+    NVIC_EnableIRQ(TIM3_IRQn);
+    NVIC_SetPriority(TIM3_IRQn, 2);
 }
 
-TIM_TypeDef *tim3 = TIM3;
 
 void TIM3_IRQHandler() {
     //TIM3->CR1 &= ~TIM_CR1_CEN_Msk; FIXME
 
     /* This takes about 10us */
-    int period = 0; /* FIXME DEBUG shift_data(); */
+    int period = shift_data();
     TIM3->CCR1 = period;
     TIM3->CNT = 0xffff; /* To not enable OC1 right away */
 
@@ -199,8 +146,11 @@ enum Command {
     CMD_GET_DESC,
 };
 
+int gaddr;
 void USART1_IRQHandler() {
-    int addr = USART1->RDR;
+    gaddr = USART1->RDR;
+    USART1->RQR |= USART_RQR_RXFRQ;
+    int addr = gaddr; /* FIXME DEBUG */
     int cmd = addr>>5;
     addr &= 0x1F;
     /* Are we addressed? */
@@ -212,32 +162,78 @@ void USART1_IRQHandler() {
         switch (cmd) {
         case CMD_SET_FB:
             /* Are we ready to process new frame data? */
-            if (fb_op != FB_WRITE)
+            if (fb_op != FB_WRITE) {
+                //asm("bkpt"); /* FIXME DEBUG */
                 goto errout; /* Error: Not yet ready to receive new packet */
+            }
             /* Disable this RX interrupt for duration of DMA transfer */
             USART1->CR1 &= ~USART_CR1_RXNEIE_Msk;
             /* Enable DMA transfer to write buffer */
+            DMA1->IFCR |= DMA_IFCR_CGIF3;
             DMA1_Channel3->CCR |= DMA_CCR_EN;
-            /* Kick off formatting code in main loop outside interrupt context */
-            fb_op = FB_FORMAT;
+            USART1->CR3 |= USART_CR3_DMAR;
             break;
         }
     }
 errout:
     /* FIXME */
-    return;
+return;
 }
 
 void DMA1_Channel2_3_IRQHandler() {
     /* DMA Transfer complete, re-enable receive interrupt */
     USART1->CR1 |= USART_CR1_RXNEIE;
+    /* ...and disable this DMA channel */
+    USART1->CR3 &= ~USART_CR3_DMAR_Msk;
+    DMA1_Channel3->CCR &= ~DMA_CCR_EN_Msk;
+    /* Kick off formatting code in main loop outside interrupt context */
+    fb_op = FB_FORMAT;
+    DMA1->IFCR |= DMA_IFCR_CGIF3;
+}
+
+void uart_config(void) {
+    USART1->CR1 = /* 8-bit -> M1, M0 clear */
+        /* RTOIE clear */
+          (8 << USART_CR1_DEAT_Pos) /* 8 sample cycles/1 bit DE assertion time */
+        | (8 << USART_CR1_DEDT_Pos) /* 8 sample cycles/1 bit DE assertion time */
+        //| USART_CR1_OVER8 FIXME debug?
+        /* CMIF clear */
+        | USART_CR1_MME
+        /* WAKE clear */
+        /* PCE, PS clear */
+        | USART_CR1_RXNEIE
+        /* other interrupts clear */
+        | USART_CR1_TE
+        | USART_CR1_RE;
+    //USART1->CR2 = USART_CR2_RTOEN; /* Timeout enable */
+    USART1->CR3 = USART_CR3_DEM; /* RS485 DE enable (output on RTS) */
+    int usartdiv = 25;
+    USART1->BRR = usartdiv;
+    USART1->CR1 |= USART_CR1_UE;
+
+    /* Configure DMA for USART frame data reception */
+    DMA1_Channel3->CPAR = (unsigned int)&USART1->RDR;
+    DMA1_Channel3->CMAR = (unsigned int)rx_buf;
+    DMA1_Channel3->CNDTR = sizeof(rx_buf);
+    DMA1_Channel3->CCR = (0<<DMA_CCR_PL_Pos);
+    DMA1_Channel3->CCR |=
+          (0<<DMA_CCR_MSIZE_Pos)
+        | (0<<DMA_CCR_PSIZE_Pos)
+        | DMA_CCR_MINC
+        | DMA_CCR_TCIE
+        | DMA_CCR_CIRC;
+
+    NVIC_EnableIRQ(USART1_IRQn);
+    NVIC_SetPriority(USART1_IRQn, 3);
+    NVIC_EnableIRQ(DMA1_Channel2_3_IRQn);
+    NVIC_SetPriority(DMA1_Channel2_3_IRQn, 3);
 }
 
 int main(void) {
     RCC->CR |= RCC_CR_HSEON;
     while (!(RCC->CR&RCC_CR_HSERDY));
     RCC->CFGR &= ~RCC_CFGR_PLLMUL_Msk & ~RCC_CFGR_SW_Msk & ~RCC_CFGR_PPRE_Msk & ~RCC_CFGR_HPRE_Msk;
-    RCC->CFGR |= (1<<RCC_CFGR_PLLMUL_Pos) | RCC_CFGR_PLLSRC_HSE_PREDIV; /* PLL x4 -> 50.0MHz */
+    RCC->CFGR |= (2<<RCC_CFGR_PLLMUL_Pos) | RCC_CFGR_PLLSRC_HSE_PREDIV; /* PLL x4 -> 50.0MHz */
     RCC->CFGR2 &= ~RCC_CFGR2_PREDIV_Msk;
     RCC->CFGR2 |= RCC_CFGR2_PREDIV_DIV2; /* prediv :2 -> 12.5MHz */
     RCC->CR |= RCC_CR_PLLON;
@@ -245,25 +241,7 @@ int main(void) {
     RCC->CFGR |= (2<<RCC_CFGR_SW_Pos);
     SystemCoreClockUpdate();
 
-    //SysTick_Config(SystemCoreClock/1000); /* 1ms interval */
-    //NVIC_DisableIRQ(SysTick_IRQn);
-    SysTick->VAL = 0U;
-    SysTick->LOAD = 0x00FFFFFF;
-    SysTick->CTRL = SysTick_CTRL_CLKSOURCE_Msk | SysTick_CTRL_ENABLE_Msk;
-    while (42) {
-        static unsigned int cvr __attribute__((used));
-        cvr = SysTick->VAL;
-        //if (fb_op == FB_FORMAT) {
-            transpose_data(rx_buf, write_fb);
-            write_fb->brightness = rx_buf[offsetof(struct framebuf, brightness)];
-        //    fb_op = FB_UPDATE;
-        //    while (fb_op == FB_UPDATE)
-        //        ;
-        //}
-        cvr = cvr -  SysTick->VAL;
-        asm volatile ("bkpt");
-    }
-    //LL_Init1msTick(SystemCoreClock);
+    LL_Init1msTick(SystemCoreClock);
 
     RCC->AHBENR  |= RCC_AHBENR_GPIOAEN | RCC_AHBENR_DMAEN;
     RCC->APB2ENR |= RCC_APB2ENR_SPI1EN | RCC_APB2ENR_USART1EN;
@@ -309,38 +287,6 @@ int main(void) {
     SPI1->CR2 &= ~SPI_CR2_DS_Msk;
     SPI1->CR2 |= LL_SPI_DATAWIDTH_16BIT;
 
-    /* Configure USART1 */
-    USART1->CR1 = /* 8-bit -> M1, M0 clear */
-        /* RTOIE clear */
-          (8 << USART_CR1_DEAT_Pos) /* 8 sample cycles/1 bit DE assertion time */
-        | (8 << USART_CR1_DEDT_Pos) /* 8 sample cycles/1 bit DE assertion time */
-        | USART_CR1_OVER8
-        /* CMIF clear */
-        | USART_CR1_MME
-        /* WAKE clear */
-        /* PCE, PS clear */
-        | USART_CR1_RXNEIE
-        /* other interrupts clear */
-        | USART_CR1_TE
-        | USART_CR1_RE;
-    //USART1->CR2 = USART_CR2_RTOEN; /* Timeout enable */
-    USART1->CR3 = USART_CR3_DEM; /* RS485 DE enable (output on RTS) */
-    int usartdiv = 25;
-    USART1->BRR = (usartdiv&0xFFF0) | ((usartdiv>>1) & 0x7);
-    USART1->CR1 |= USART_CR1_UE;
-
-    /* Configure DMA for USART frame data reception */
-    USART1->CR3 |= USART_CR3_DMAR;
-    DMA1_Channel3->CPAR = (unsigned int)&USART1->RDR;
-    DMA1_Channel3->CMAR = (unsigned int)rx_buf;
-    DMA1_Channel3->CNDTR = sizeof(rx_buf);
-    DMA1_Channel3->CCR = (0<<DMA_CCR_PL_Pos);
-    DMA1_Channel3->CCR |=
-          (0<<DMA_CCR_MSIZE_Pos)
-        | (0<<DMA_CCR_PSIZE_Pos)
-        | DMA_CCR_MINC
-        | DMA_CCR_TCIE;
-
     /* Baud rate PCLK/2 -> 25MHz */
     SPI1->CR1 =
           SPI_CR1_BIDIMODE
@@ -354,19 +300,30 @@ int main(void) {
         | SPI_CR1_CPHA;
     /* FIXME maybe try w/o BIDI */
 
+    /* Clear frame buffer */
     read_fb->brightness = 1;
     for (int i=0; i<sizeof(read_fb->data)/sizeof(uint32_t); i++) {
-        read_fb->data[i] = 0xffffffff;
+        read_fb->data[i] = 0xffffffff; /*FIXME debug replace with 0x00000000 */
     }
-    cfg_timer3();
 
-    NVIC_EnableIRQ(TIM3_IRQn);
-    NVIC_SetPriority(TIM3_IRQn, 2);
- 
+    cfg_timer3();
     SysTick_Config(SystemCoreClock/1000); /* 1ms interval */
+    uart_config();
 
+    uint8_t i=0;
+    int last_sys_time=0;
     while (42) {
         led_state = (sys_time>>8)&7;
+        if ((sys_time ^ last_sys_time) & (1<<4)) {
+            USART1->TDR = i++;
+        }
+        last_sys_time = sys_time;
+        if (fb_op == FB_FORMAT) {
+            //transpose_data(rx_buf, read_fb);
+            fb_op = FB_UPDATE;
+            while (fb_op == FB_UPDATE)
+                ;
+        }
     }
 }
 
diff --git a/fw/transpose.c b/fw/transpose.c
new file mode 100644
index 0000000..1c5f982
--- /dev/null
+++ b/fw/transpose.c
@@ -0,0 +1,61 @@
+
+#include <unistd.h>
+#include <string.h>
+
+#include "transpose.h"
+
+void transpose_data(volatile uint8_t *rx_buf, volatile struct framebuf *out_fb) {
+    memset((uint8_t *)out_fb, 0, sizeof(*out_fb));
+    struct data_format *rxp = (struct data_format *)rx_buf;
+    for (int bit=0; bit<8; bit++) { /* bits */
+        uint32_t bit_mask = 1U<<bit;
+        volatile uint32_t *frame_data = out_fb->frame[bit+2].data;
+        uint8_t *start_inp = rxp->high;
+        for (volatile uint32_t *outp=frame_data; outp<frame_data+8; outp++) { /* segments */
+            uint32_t acc = 0;
+            uint8_t *inp = start_inp++;
+            for (int digit=0; digit<32; digit++) {
+                acc |= (*inp & bit_mask) >> bit << digit;
+                inp += sizeof(struct data_format);
+            }
+            *outp = acc;
+        }
+    }
+    for (int bit=0; bit<2; bit++) { /* bits */
+        volatile uint32_t *frame_data = out_fb->frame[bit].data;
+        uint16_t *inp = &rxp->low;
+        for (int seg=0; seg<8; seg++) { /* segments */
+            uint32_t mask = 1 << bit << (seg*2);
+            uint32_t acc = 0;
+            for (int digit=0; digit<32; digit++) {
+                acc |= (*inp & mask) >> bit >> seg << digit;
+                inp += sizeof(struct data_format)/sizeof(uint16_t);
+            }
+            frame_data[seg] = acc;
+        }
+    }
+}
+
+
+void untranspose_data(struct framebuf *fb, uint8_t *txbuf) {
+    memset(txbuf, 0, sizeof(*fb));
+
+    struct data_format *tx = (struct data_format *) txbuf;
+
+    for (size_t i=0; i<32; i++) { /* digit */
+        for (size_t j=0; j<8; j++) { /* segment */
+            for (size_t k=0; k<8; k++) { /* bit */
+                tx[i].high[j] |= (fb->frame[k+2].data[j] & (1<<i)) ? (1<<k) : 0;
+            }
+        }
+    }
+
+    for (size_t i=0; i<32; i++) { /* digit */
+        for (size_t j=0; j<8; j++) { /* segment */
+            for (size_t k=0; k<2; k++) { /* bit */
+                tx[i].low |= (fb->frame[k].data[j] & (1<<i)) ? (1<<k<<(j*2)) : 0;
+            }
+        }
+    }
+}
+
diff --git a/fw/transpose.h b/fw/transpose.h
new file mode 100644
index 0000000..1927a72
--- /dev/null
+++ b/fw/transpose.h
@@ -0,0 +1,45 @@
+#ifndef __TRANSPOSE_H__
+#define __TRANSPOSE_H__
+
+#include <stdint.h>
+
+enum Segment { SegA, SegB, SegC, SegD, SegE, SegF, SegG, SegDP, nsegments };
+enum {
+    nrows = 4,
+    ncols = 8,
+    nbits = 10,
+};
+
+enum {
+    frame_size_words = nrows*ncols*nsegments/32,
+};
+
+struct framebuf {
+    /* Multiplexing order: first Digits, then Time/bits, last Segments */
+    union {
+        uint32_t data[nbits*frame_size_words];
+        struct {
+            struct {
+                uint32_t data[frame_size_words];
+            } frame[nbits];
+        };
+    };
+    uint8_t brightness; /* 0 or 1; controls global brighntess control */
+};
+
+struct data_format {
+    union {
+        uint8_t high[8];
+        struct { uint8_t ah, bh, ch, dh, eh, fh, gh, dph; };
+    };
+    union {
+        uint16_t low;
+        struct { uint8_t dpl:2, gl:2, fl:2, el:2, dl:2, cl:2, bl:2, al:2; };
+    };
+} __attribute__((packed));
+
+
+void transpose_data(volatile uint8_t *rx_buf, volatile struct framebuf *out_fb);
+void untranspose_data(struct framebuf *fb, uint8_t *txbuf);
+
+#endif//__TRANSPOSE_H__
diff --git a/fw/transpose_main.c b/fw/transpose_main.c
new file mode 100644
index 0000000..793872d
--- /dev/null
+++ b/fw/transpose_main.c
@@ -0,0 +1,77 @@
+
+#include <stdio.h>
+#include <stdint.h>
+#include <malloc.h>
+#include <string.h>
+#include <assert.h>
+
+#include "transpose.h"
+
+void rxbuf_dump(struct data_format *rx) {
+    fprintf(stdout, "\033[38;5;244m          high          |       low\n");
+    fprintf(stdout, " a  b  c  d  e  f  g  . | a b c d e f g .\033[0m\n");
+
+    for (int i=0; i<32; i++) {
+        for (int j=0; j<8; j++)
+            fprintf(stdout, "%02x ", rx[i].high[j]);
+        fprintf(stdout, "\033[38;5;244m|\033[0m");
+        for (int j=0; j<8; j++)
+            fprintf(stdout, " %1x", (rx[i].low>>(2*j))&0x3);
+        fprintf(stdout, "\n");
+    }
+}
+
+void fb_dump(struct framebuf *fb) {
+    fprintf(stdout, "\033[38;5;244mbit | ");
+    for (int i=0; i<8; i++)
+        fprintf(stdout, "       %d ", i);
+    fprintf(stdout, "\033[0m\n");
+
+    for (int i=0; i<10; i++) {
+        fprintf(stdout, " %2d |", i);
+        for (int j=0; j<8; j++)
+           fprintf(stdout, " %08x", fb->frame[i].data[j]);
+        fprintf(stdout, "\n");
+    }
+}
+
+int main(int argc, char **argv) {
+    FILE *randf = fopen("/dev/urandom", "r");
+    if (!randf)
+        return 2;
+
+    struct framebuf *fb = malloc(sizeof(struct framebuf));
+    if (!fb)
+        return 2;
+    assert(sizeof(*fb) == 32*sizeof(struct data_format)+4);
+
+    uint8_t *rxbuf1 = malloc(sizeof(*fb));
+    if (!rxbuf1)
+        return 2;
+    if (fread(rxbuf1, 1, sizeof(*fb), randf) != sizeof(*fb))
+        return 2;
+
+    uint8_t *rxbuf2 = malloc(sizeof(*fb));
+    if (!rxbuf2)
+        return 2;
+    memcpy(rxbuf2, rxbuf1, sizeof(*fb));
+
+    memset((void *)fb, 0, sizeof(*fb));
+
+    transpose_data(rxbuf1, fb);
+    untranspose_data(fb, rxbuf1);
+
+    assert(!memcmp(rxbuf1, rxbuf2, sizeof(fb)));
+    
+    fprintf(stdout, "\n\033[93mDUMP of ORIGINAL\033[0m\n");
+    rxbuf_dump((struct data_format *)rxbuf2);
+
+    fprintf(stdout, "\n\033[93mDUMP of FRAME BUFFER\033[0m\n");
+    fb_dump(fb);
+
+    fprintf(stdout, "\n\033[93mDUMP of RESULT\033[0m\n");
+    rxbuf_dump((struct data_format *)rxbuf1);
+
+    fprintf(stderr, "PASS");
+    return 0;
+}