refactor: decode pattern once during initialization

font/tandy1k.data.h

@@ -6,9 +6,9 @@
 extern "C" {
 #endif
 
-extern uint32_t _binary_font_tandy1k_tandy1k_data_start __asm("_binary_font_tandy1k_tandy1k_data_start");
-extern uint32_t _binary_font_tandy1k_tandy1k_data_end __asm("_binary_font_tandy1k_tandy1k_data_end");
-extern uint32_t _binary_font_tandy1k_tandy1k_data_size __asm("_binary_font_tandy1k_tandy1k_data_size");
+extern uint32_t _binary_font_tandy1k_data_start __asm("_binary_font_tandy1k_data_start");
+extern uint32_t _binary_font_tandy1k_data_end __asm("_binary_font_tandy1k_data_end");
+extern uint32_t _binary_font_tandy1k_data_size __asm("_binary_font_tandy1k_data_size");
 
 #ifdef __cplusplus
 }

src/graphics.cpp

@@ -127,8 +127,8 @@ void transfer_textures()
 
   uint32_t offset = texture_memory_alloc.texture.start + 0;
   void * dst = reinterpret_cast<void *>(&ta_fifo_texture_memory[offset / 4]);
-  void * src = reinterpret_cast<void *>(&_binary_font_tandy1k_tandy1k_data_start);
-  int size = reinterpret_cast<int>(&_binary_font_tandy1k_tandy1k_data_size);
+  void * src = reinterpret_cast<void *>(&_binary_font_tandy1k_data_start);
+  int size = reinterpret_cast<int>(&_binary_font_tandy1k_data_size);
   transfer_ta_fifo_texture_memory_32byte(dst, src, size);
 }
 

src/interpreter.cpp

@@ -177,87 +177,94 @@ void rekey_note(int ch, xm_pattern_format_t * pf)
   }
 }
 
-int parse_pattern_line(xm_pattern_header_t * pattern_header, int note_offset, void (*func)(int, xm_pattern_format_t*))
+static inline void next_pattern()
 {
-  uint8_t * pattern = (uint8_t *)(((int)pattern_header) + s32(&pattern_header->pattern_header_length));
-
-  for (int i = 0; i < state.cache.number_of_channels; i++) {
-    int p = pattern[note_offset];
-    if (p & 0x80) {
-      note_offset += 1;
-      xm_pattern_format_t pf = {};
-      if (p & (1 << 0))
-        pf.note = pattern[note_offset++];
-      if (p & (1 << 1))
-        pf.instrument = pattern[note_offset++];
-      if (p & (1 << 2))
-        pf.volume_column_byte = pattern[note_offset++];
-      if (p & (1 << 3))
-        pf.effect_type = pattern[note_offset++];
-      if (p & (1 << 4))
-        pf.effect_parameter = pattern[note_offset++];
-      func(i, &pf);
-    } else {
-      xm_pattern_format_t * pf = (xm_pattern_format_t *)&pattern[note_offset];
-      func(i, pf);
-      note_offset += 5;
-    }
-  }
-  return note_offset;
-}
-
-static inline void next_pattern(int pattern_break)
-{
-  state.line_index = 0;
-  state.next_note_offset = 0;
+  if (state.pattern_break >= 0)
+    state.next_line_index = state.pattern_break;
+  else
+    state.next_line_index = 0;
   state.pattern_break = -1;
 
   state.pattern_order_table_index += 1;
-  printf("pattern_order_table_index: %d\n", state.pattern_order_table_index);
-  if (state.pattern_order_table_index >= state.cache.song_length)
+  if (state.pattern_order_table_index >= state.xm.song_length)
     state.pattern_order_table_index = 0;
+  printf("pattern_order_table_index: %d\n", state.pattern_order_table_index);
+
   state.pattern_index = state.xm.header->pattern_order_table[state.pattern_order_table_index];
+
+  printf("note_count: %d\n", state.xm.pattern_note_count[state.pattern_index]);
+}
+
+template <void (*F)(int, xm_pattern_format_t *)>
+void execute_line(int line_index)
+{
+  int line_pattern_index = line_index * state.xm.number_of_channels;
+  for (int ch = 0; ch < state.xm.number_of_channels; ch++) {
+    xm_pattern_format_t * pattern = state.xm.pattern[state.pattern_index];
+    F(ch, &pattern[line_pattern_index + ch]);
+  }
 }
 
 void interrupt()
 {
-  xm_pattern_header_t * pattern_header = state.xm.pattern_header[state.pattern_index];
-  int pattern_data_size = s16(&pattern_header->packed_pattern_data_size);
-
   bool keyoff_tick = (state.tick + 1) % (state.ticks_per_line * 2) == 0;
   bool note_tick = state.tick % (state.ticks_per_line * 2) == 0;
   bool effect_tick = (state.tick & 1) == 0;
   bool pattern_break_tick = (state.tick % (state.ticks_per_line * 2)) == (state.ticks_per_line * 2 - 1);
+
   if (keyoff_tick) {
     // execute keyoffs
-    parse_pattern_line(pattern_header, state.next_note_offset, rekey_note);
-    wait(); aica_sound.channel[0].KYONEX(1);
+    execute_line<rekey_note>(state.next_line_index);
   }
 
   if (state.pattern_break >= 0 && pattern_break_tick) {
     printf("pattern_break\n");
-    next_pattern(-1);
+    next_pattern();
   }
 
   if (note_tick) {
-    state.note_offset = state.next_note_offset;
-    //state.next_note_offset = parse_pattern_line(pattern_header, state.note_offset, play_debug_note);
-    state.next_note_offset = parse_pattern_line(pattern_header, state.note_offset, play_note);
-    state.line_index += 1;
-    wait(); aica_sound.channel[0].KYONEX(1);
-  }
-  if (effect_tick && !note_tick) {
-    // execute effects
-    parse_pattern_line(pattern_header, state.note_offset, play_note_effect);
-    wait(); aica_sound.channel[0].KYONEX(1);
-  }
+    // execute notes
+    state.line_index = state.next_line_index;
+    state.next_line_index += 1;
 
-  if (state.next_note_offset >= pattern_data_size && pattern_break_tick) {
-    printf("pattern_data_size\n");
-    next_pattern(-1);
+    execute_line<play_note>(state.line_index);
+  } else if (effect_tick) {
+    // execute effects
+    execute_line<play_note_effect>(state.line_index);
+  }
+  wait(); aica_sound.channel[0].KYONEX(1);
+
+  int note_index = state.next_line_index * state.xm.number_of_channels;
+  bool end_of_pattern = note_index >= state.xm.pattern_note_count[state.pattern_index];
+  if (end_of_pattern && pattern_break_tick) {
+    printf("end_of_pattern\n");
+    next_pattern();
   }
 
   state.tick += 1;
 }
 
+void init()
+{
+  // 195 = 1ms
+  // 2500 / bpm milliseconds
+
+  int default_bpm = s16(&state.xm.header->default_bpm);
+  int default_tempo = s16(&state.xm.header->default_tempo);
+  int tick_rate = 195.32 * 2500 / default_bpm;
+
+  printf("default_bpm %d\n", default_bpm);
+  printf("default_tempo %d\n", default_tempo);
+  printf("tick_rate %d\n", tick_rate);
+
+  state.tick_rate = tick_rate;
+  state.ticks_per_line = default_tempo;
+  state.tick = 0;
+  state.line_index = 0;
+  state.pattern_order_table_index = -1;
+  next_pattern();
+
+  printf("tick_rate %d\n", state.tick_rate);
+}
+
 }

src/interpreter.hpp

@@ -17,18 +17,16 @@ struct interpreter_state {
   int pattern_order_table_index;
   int pattern_break;
   int pattern_index;
-  int line_index; // within the current pattern (for debugging)
-  int note_offset; // within the current pattern
-  int next_note_offset;
+  int line_index;
+  int next_line_index; // within the current pattern
 
   struct xm_state xm;
 
-  struct xm_state_cache cache;
-
   struct channel_state channel[max_channels];
 };
 
 extern struct interpreter_state state;
 void interrupt();
+void init();
 
 }

src/main.cpp

@@ -11,6 +11,9 @@
 #include "interpreter.hpp"
 #include "sound.hpp"
 
+#include "xm/milkypack01.xm.h"
+#include "xm.h"
+
 void vbr100()
 {
   serial::string("vbr100\n");
@@ -171,11 +174,28 @@ void input_update()
   last_b = b;
 }
 
+void load_xm()
+{
+  using namespace interpreter;
+
+  int buf = (int)&_binary_xm_milkypack01_xm_start;
+
+  static uint8_t __attribute__((aligned(32))) sample_data[1024 * 1024 * 2];
+  const int sample_data_length = (sizeof (sample_data));
+
+  int sample_data_ix = xm_init(&state.xm,
+                               buf,
+                               sample_data,
+                               sample_data_length);
+  interpreter::init();
+  sound_init(sample_data, sample_data_ix, state.tick_rate);
+}
+
 void main()
 {
   serial::init(0);
 
-  sound_init();
+  load_xm();
   graphics_init();
   interrupt_init();
   //channel_sandbox_defaults();

src/malloc.c

@@ -0,0 +1,33 @@
+#include "assert.h"
+#include "malloc.h"
+
+struct arena {
+  uint8_t * mem;
+  uint32_t size;
+  uint32_t ix;
+};
+
+static uint8_t arena_mem[0x100000];
+
+static struct arena arena = {
+  .mem = arena_mem,
+  .size = (sizeof (arena_mem)),
+  .ix = 0,
+};
+
+void malloc_arena_reset()
+{
+  arena.ix = 0;
+}
+
+void * malloc_arena(uint32_t size)
+{
+  if (size == 0)
+    return nullptr;
+
+  assert((arena.ix & (~3)) == arena.ix);
+  void * ptr = &arena.mem[arena.ix];
+  size = (size + 3) & (~3);
+  arena.ix += size;
+  return ptr;
+}

src/malloc.h

@@ -0,0 +1,14 @@
+#pragma once
+
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void malloc_arena_reset();
+void * malloc_arena(uint32_t size);
+
+#ifdef __cplusplus
+}
+#endif

src/sound.cpp

@@ -17,10 +17,6 @@
 
 #include "assert.h"
 
-#include "xm/xm.h"
-#include "xm/milkypack01.xm.h"
-
-#include "xm.h"
 #include "interpreter.hpp"
 
 void g2_aica_dma(uint32_t g2_address, uint32_t system_address, int length)
@@ -73,20 +69,8 @@ void writeback(void const * const buf, uint32_t size)
 
 static uint8_t __attribute__((aligned(32))) zero[0x28c0] = {};
 
-void sound_init()
+void sound_init(uint8_t * sample_data, int sample_data_ix, int tick_rate)
 {
-  using namespace interpreter;
-
-  int buf = (int)&_binary_xm_milkypack01_xm_start;
-
-  static uint8_t __attribute__((aligned(32))) sample_data[1024 * 1024 * 2];
-  const int sample_data_length = (sizeof (sample_data));
-  int sample_data_ix = xm_init(&state.xm,
-                               &state.cache,
-                               buf,
-                               sample_data,
-                               sample_data_length);
-
   wait(); aica_sound.common.vreg_armrst = aica::vreg_armrst::ARMRST(1);
   wait(); aica_sound.common.dmea0_mrwinh = aica::dmea0_mrwinh::MRWINH(0b0111);
   system.ISTNRM = istnrm::end_of_dma_aica_dma;
@@ -99,14 +83,6 @@ void sound_init()
   g2_aica_dma((uint32_t)0x00703000, (int)zero, 0x15e0);
   g2_aica_dma_wait_complete();
 
-  printf("i[0] start %d size %d\n",
-         state.xm.sample_data_offset[0],
-         s32(&state.xm.sample_header[0]->sample_length));
-
-  printf("i[1] start %d size %d\n",
-         state.xm.sample_data_offset[1],
-         s32(&state.xm.sample_header[1]->sample_length));
-
   for (int i = 0; i < 16; i++) {
     serial::hexlify(&sample_data[i * 16], 16);
   }
@@ -175,27 +151,8 @@ void sound_init()
     | aica::mono_mem8mb_dac18b_ver_mvol::MVOL(0xc) // volume
     ;
 
-  // 195 = 1ms
-  // 2500 / bpm milliseconds
-  printf("default_bpm %d\n", s16(&state.xm.header->default_bpm));
-  printf("default_tempo %d\n", s16(&state.xm.header->default_tempo));
-
-  state.tick_rate = 195.32 * 2500 / s16(&state.xm.header->default_bpm);
-  state.ticks_per_line = s16(&state.xm.header->default_tempo);
-  state.tick = 0;
-  state.pattern_break = -1;
-  state.pattern_order_table_index = 0;
-  state.pattern_index = state.xm.header->pattern_order_table[state.pattern_order_table_index];
-  state.line_index = 0;
-  state.note_offset = 0;
-  state.next_note_offset = 0;
-
-  printf("tick_rate %d\n", state.tick_rate);
-
-  printf("pattern %d\n", state.pattern_index);
-
   sh7091.TMU.TSTR = 0; // stop all timers
-  sh7091.TMU.TCOR0 = state.tick_rate / 2;
+  sh7091.TMU.TCOR0 = tick_rate / 2;
   sh7091.TMU.TOCR = tmu::tocr::tcoe::tclk_is_external_clock_or_input_capture;
   sh7091.TMU.TCR0
     = tmu::tcr0::UNIE

src/sound.hpp

@@ -13,4 +13,4 @@ static inline void wait()
   };
 }
 
-void sound_init();
+void sound_init(uint8_t * sample_data, int sample_data_ix, int tick_rate);

src/xm.c

@@ -1,6 +1,7 @@
 #include "xm/xm.h"
 #include "printf/printf.h"
 #include "xm.h"
+#include "malloc.h"
 
 static int xm_unpack_sample(int buf,
                             int offset,
@@ -93,7 +94,6 @@ static int xm_samples_init(xm_state_t * xm,
                                           sample_header[i],
                                           sample_data,
                                           *sample_data_ix);
-      printf("%d\n", sample_data_length);
       assert(*sample_data_ix <= sample_data_length);
     }
     offset += sample_length;
@@ -101,8 +101,71 @@ static int xm_samples_init(xm_state_t * xm,
   return offset;
 }
 
+static inline xm_pattern_format_t parse_pattern_line(uint8_t * pattern, int * note_offset)
+{
+  int offset = *note_offset;
+
+  int p = pattern[offset];
+  if (p & 0x80) {
+    offset += 1;
+    xm_pattern_format_t pf = {};
+    if (p & (1 << 0))
+      pf.note = pattern[offset++];
+    if (p & (1 << 1))
+      pf.instrument = pattern[offset++];
+    if (p & (1 << 2))
+      pf.volume_column_byte = pattern[offset++];
+    if (p & (1 << 3))
+      pf.effect_type = pattern[offset++];
+    if (p & (1 << 4))
+      pf.effect_parameter = pattern[offset++];
+    *note_offset = offset;
+    return pf;
+  } else {
+    xm_pattern_format_t * pf = (xm_pattern_format_t *)&pattern[offset];
+    offset += 5;
+    *note_offset = offset;
+    return *pf;
+  }
+}
+
+static inline int count_pattern_notes(uint8_t * pattern, int pattern_data_size)
+{
+  int note_offset = 0;
+  int note_count = 0;
+
+  while (note_offset < pattern_data_size) {
+    parse_pattern_line(pattern, &note_offset);
+    note_count += 1;
+  }
+  assert(note_offset == pattern_data_size);
+
+  return note_count;
+}
+
+void xm_unpack_pattern(xm_state_t * xm,
+                       int pattern_index)
+{
+  xm_pattern_header_t * pattern_header = xm->pattern_header[pattern_index];
+  uint8_t * pattern = (uint8_t *)(((int)pattern_header) + s32(&pattern_header->pattern_header_length));
+
+  int pattern_data_size = s16(&pattern_header->packed_pattern_data_size);
+
+  int note_count = count_pattern_notes(pattern, pattern_data_size);
+
+  xm_pattern_format_t * pf = (xm_pattern_format_t *)malloc_arena((sizeof (xm_pattern_format_t)) * note_count);
+
+  xm->pattern[pattern_index] = pf;
+  xm->pattern_note_count[pattern_index] = note_count;
+
+  int note_offset = 0;
+  for (int i = 0; i < note_count; i++) {
+    pf[i] = parse_pattern_line(pattern, &note_offset);
+  }
+  assert(note_offset == pattern_data_size);
+}
+
 int xm_init(xm_state_t * xm,
-            xm_state_cache_t * cache,
             int buf,
             uint8_t * sample_data,
             int sample_data_length)
@@ -142,12 +205,19 @@ int xm_init(xm_state_t * xm,
   printf("end_of_instruments: %d\n", offset);
 
   int number_of_channels = s16(&xm->header->number_of_channels);
-  cache->number_of_channels = number_of_channels;
+  xm->number_of_channels = number_of_channels;
   printf("number_of_channels: %d\n", number_of_channels);
 
   int song_length = s16(&xm->header->song_length);
-  cache->song_length = song_length;
+  xm->song_length = song_length;
   printf("song_length: %d\n", song_length);
 
+  // reset arena
+  malloc_arena_reset();
+
+  for (int pattern_index = 0; pattern_index < number_of_patterns; pattern_index++) {
+    xm_unpack_pattern(xm, pattern_index);
+  }
+
   return sample_data_ix;
 }

src/xm.h

@@ -15,15 +15,14 @@ typedef struct xm_state {
   xm_instrument_header_t * instrument_header[xm_max_instruments];
   xm_sample_header_t * sample_header[xm_max_instruments]; // array
   int sample_data_offset[xm_max_instruments];
-} xm_state_t;
 
-typedef struct xm_state_cache {
   int number_of_channels;
   int song_length;
-} xm_state_cache_t;
+  xm_pattern_format_t * pattern[xm_max_patterns];
+  int pattern_note_count[xm_max_patterns];
+} xm_state_t;
 
 int xm_init(xm_state_t * xm,
-            xm_state_cache_t * cache,
             int buf,
             uint8_t * sample_data,
             int sample_data_length);

xm_player.mk

@@ -20,7 +20,8 @@ XM_PLAYER_OBJ = \
 	src/interpreter.o \
 	src/main.o \
 	src/sound.o \
-	src/xm.o
+	src/xm.o \
+	src/malloc.o
 
 xm_player.elf: LDSCRIPT = $(LIB)/main.lds
 xm_player.elf: $(START_OBJ) $(XM_PLAYER_OBJ) $(TEXTURE_OBJ) $(XM_OBJ) $(LIBGCC)