Add perf test

4 files changed, 137 insertions, 7 deletions

diff --git a/test/e2e/emulation_speed/bootstrap.atk16 b/test/e2e/emulation_speed/bootstrap.atk16
new file mode 100644
index 0000000..9b024e6
--- /dev/null
+++ b/test/e2e/emulation_speed/bootstrap.atk16
@@ -0,0 +1,78 @@
+@let SP              RH
+
+; vector table fields
+@let vector_table    0x10
+@let vt_ISR0         0x10 ; ISR0
+@let vt_ISR1         0x11 ; ISR1
+@let vt_ISR2         0x12 ; ISR2
+@let vt_ISR3         0x13 ; ISR3
+@let vt_stack_addr   0x14 ; Stack address
+@let vt_term_pp_addr 0x15 ; Terminal peripheral address
+@let vt_kb_pp_addr   0x16 ; Keyboard peripheral address
+@let vt_gr_mode_addr 0x17 ; Graphics mode setting address
+@let vt_sprite_mem   0x18 ; Sprite memory address
+@let vt_text_mem     0x19 ; Text memory buffer address
+
+; memory segments
+@let stack_segment   0x8000
+@let mmio_segment    0xE000
+@let terminal_addr   0xE000
+@let keyboard_addr   0xE001
+@let gr_mode_addr    0xE002
+@let sprite_mem      0xE800
+@let text_mem        0xF800 ; to 0xFFFF
+; note: sprite memory & text memory can use the same space
+; since they are never used at the same time
+
+; graphics mode settings
+@let gr_disabled_mode 0b00
+@let gr_text_mode     0b01
+@let gr_sprite_mode   0b10
+
+@address vector_table
+  keyboard_isr   ; 0x10
+  hlt_isr        ; 0x11
+  hlt_isr        ; 0x12
+  hlt_isr        ; 0x13
+  stack_segment  ; 0x14
+  terminal_addr  ; 0x15
+  keyboard_addr  ; 0x16
+  gr_mode_addr   ; 0x17
+  sprite_mem     ; 0x18
+  text_mem       ; 0x19
+
+@label hlt_isr
+  ldi 0x55 RE
+  hlt
+
+@label keyboard_isr
+  spu RA
+  spu RB
+  ldi vt_kb_pp_addr RA    ; RA := Keyboard peripheral address pointer
+  ldr RA RA               ; RA := Keyboard peripheral address deref
+  ldr RA RA               ; RA := <16-bit keyboard character code> from MMIO register
+  ldi vt_term_pp_addr RB  ; RB := Terminal peripheral address pointer
+  ldr RB RB               ; RB := Terminal peripheral address deref
+  str RA RB               ; terminal <- character code (ASCII?)
+  mov RA RE               ; RE := character code (for debugging)
+  spo RB
+  spo RA
+  rti
+@label program_segment
+
+; define short prelude that sets up vital instructions
+; NOTE: must be shorter than the vector table offset
+@address 0x0
+; set up stack pointer to point to beginning of stack segment
+  ldi vt_stack_addr RA
+  ldr RA SP
+  jpi program_segment
+
+@address program_segment
+; set graphics mode to disabled
+  ldi vt_gr_mode_addr RA
+  ldr RA RA
+  ldi gr_disabled_mode RB
+  str RB RA
+; jump to main (user defined)
+  jpi main
diff --git a/test/e2e/emulation_speed/long_loop.atk16 b/test/e2e/emulation_speed/long_loop.atk16
new file mode 100644
index 0000000..2303311
--- /dev/null
+++ b/test/e2e/emulation_speed/long_loop.atk16
@@ -0,0 +1,16 @@
+; bootstrap code (must be called to setup mandatory data)
+; will jump to label main
+@include bootstrap
+
+@label iters
+  10000
+
+@label main
+  ldi iters RA
+  ldr RA RA
+@label loop
+  dec RA
+  bri zero done
+  jpi loop
+@label done
+  hlt
diff --git a/test/e2e/emulation_speed/test_emulation_speed.py b/test/e2e/emulation_speed/test_emulation_speed.py
new file mode 100644
index 0000000..f2bcd22
--- /dev/null
+++ b/test/e2e/emulation_speed/test_emulation_speed.py
@@ -0,0 +1,43 @@
+import time
+from atk16_asm import assemble
+from atk16_emu import Machine
+from test.utils import pad_bytearray
+
+def test_emulation_speed():
+  filename = "test/e2e/emulation_speed/long_loop.atk16"
+  with open(filename, "r") as f:
+    source = f.read()
+
+  obj = assemble(source, filename)
+  rom_image = pad_bytearray(obj.program)
+
+  machine = Machine()
+  machine.load_rom_image(rom_image)
+  machine.reset()
+  start = time.perf_counter_ns()
+  machine.run_until_halted()
+  end = time.perf_counter_ns()
+  machine.print_state_summary()
+
+  elapsed = end - start
+  print("Elapsed time:", elapsed, "ns")
+  steps_taken = machine.steps_taken
+  mean_ns_per_instr = elapsed / steps_taken
+  print("Instructions / second:", steps_taken / (elapsed / 1e9))
+  print("ns / instruction:", mean_ns_per_instr)
+
+  # VGA frequency is 25.175 MHz
+  # Divided by 16, that's 1.5734375 MHz (main clock)
+  # Instructions take 3–5 cycles, the average instruction takes around 4 cycles
+  # 1.5734375 MHz / 4 = 393359.375 Hz
+  # That is the goal instruction frequency
+  # TODO limit emulation speed to 393359.375 Hz ~= 2542 ns per instruction
+  #      this is in order to have similar speed as the actual machine
+
+  expected = 2542
+  allowed_range = 0.05 # 0..1
+  expected_min = expected * (1 - allowed_range)
+  expected_max = expected * (1 + allowed_range)
+  percentage = 100 * mean_ns_per_instr / expected
+  assert expected_min < mean_ns_per_instr, f"Emulator seems to run too fast ({percentage:.2f}% of expected ns per instruction)"
+  assert mean_ns_per_instr < expected_max, f"Emulator seems to run too slow ({percentage:.2f}% of expected ns per instruction)"
diff --git a/test/e2e/fibo/test_fibo.py b/test/e2e/fibo/test_fibo.py
index 65af07a..664e701 100644
--- a/test/e2e/fibo/test_fibo.py
+++ b/test/e2e/fibo/test_fibo.py
@@ -22,13 +22,6 @@ def test_fibo():
   elapsed = end - start
   print("Elapsed time:", elapsed, "ns")
   steps_taken = machine.steps_taken
-
-  # VGA frequency is 25.175 MHz
-  # Divided by 16, that's 1.5734375 MHz (main clock)
-  # Instructions take 3–5 cycles, the average instruction takes around 4 cycles
-  # 1.5734375 MHz / 4 = 393359.375 Hz
-  # That is the goal instruction frequency
-  # TODO limit emulation speed to 393359.375 Hz
   print("Instructions / second:", steps_taken / (elapsed / 1e9))
 
   assert machine.ra.value == 41443