ii-sound/generate_player.py

575 lines
18 KiB
Python
Raw Normal View History

import enum
import itertools
import numpy
from typing import Iterable, List, Tuple
import opcodes_6502
SLOW_PATH_TRAMPOLINE = [
opcodes_6502.Literal("eof_slow_path:", indent=0),
opcodes_6502.STA_C030,
opcodes_6502.Opcode(4, 3, "LDX WDATA"),
opcodes_6502.Opcode(4, 3, "STX @jmp+1"),
opcodes_6502.Literal("@jmp:", indent=0),
opcodes_6502.Opcode(6, 3, "JMP ($2000)") # TODO: 5 cycles on 6502
2022-06-04 13:06:52 +00:00
]
SLOW_PATH_EOF = [
opcodes_6502.Literal(
"; We've read exactly 2KB from the socket buffer. Before continuing "
"we need to ACK this read,"),
opcodes_6502.Literal(
"; and make sure there's at least another 2KB in the buffer."),
opcodes_6502.Literal(";"),
opcodes_6502.Literal(
"; Save the W5100 address pointer so we can continue reading the "
"socket buffer once we are done."),
opcodes_6502.Literal(
"; We know the low-order byte is 0 because Socket RX memory is "
"page-aligned and so is 2K frame."),
opcodes_6502.Literal(
"; IMPORTANT - from now on until we restore this below, we can't "
"trash the Y register!"),
opcodes_6502.Opcode(4, 4, "LDY WADRH"),
opcodes_6502.Literal("; Update new Received Read pointer."),
opcodes_6502.Literal(";"),
opcodes_6502.Literal(
"; We know we have received exactly 2KB, so we don't need to read the "
"current value from the"),
opcodes_6502.Literal(
"; hardware. We can track it ourselves instead, which saves a "
"few cycles."),
opcodes_6502.Opcode(2, 2, "LDA #>S0RXRD"),
opcodes_6502.Opcode(4, 3, "STA WADRH"),
opcodes_6502.Opcode(2, 2, "LDA #<S0RXRD"),
opcodes_6502.Opcode(4, 3, "STA WADRL"),
opcodes_6502.Opcode(4, 3,
"LDA RXRD ; TODO: in principle we could update RXRD outside of "
"the EOF path"),
opcodes_6502.Opcode(2, 1, "CLC"),
opcodes_6502.Opcode(2, 2, "ADC #$08"),
opcodes_6502.Opcode(4, 3,
"STA WDATA ; Store new high byte of received read pointer"),
opcodes_6502.Opcode(4, 3, "STA RXRD ; Save for next time"),
opcodes_6502.Literal("; Send the Receive command"),
opcodes_6502.Opcode(2, 2, "LDA #<S0CR"),
opcodes_6502.Opcode(4, 3, "STA WADRL"),
opcodes_6502.Opcode(2, 2, "LDA #SCRECV"),
opcodes_6502.Opcode(4, 3, "STA WDATA"),
opcodes_6502.Literal(
"; Make sure we have at least 2KB more in the socket buffer so we can "
"start another frame."
),
opcodes_6502.Opcode(2, 2, "LDA #$07"),
opcodes_6502.Opcode(2, 2, "LDX #<S0RXRSR ; Socket 0 Received Size "
"register"),
opcodes_6502.Literal(
"; we might loop an unknown number of times here waiting for data but "
"the default should be to"),
opcodes_6502.Literal("; fall straight through"),
opcodes_6502.Literal("@0:", indent=0),
opcodes_6502.Opcode(4, 3, "STX WADRL"),
opcodes_6502.Opcode(4, 3, "CMP WDATA ; High byte of received size"),
opcodes_6502.Opcode(2, 2,
"BCS @0 ; 2 cycles in common case when there is already sufficient "
"data waiting."),
opcodes_6502.Literal(
"; We're good to go for another frame. Restore W5100 address pointer "
"where we last found it, to"),
opcodes_6502.Literal(
"; begin iterating through the next 2KB of the socket buffer."),
opcodes_6502.Literal(";"),
opcodes_6502.Literal(
"; It turns out that the W5100 automatically wraps the address pointer "
"at the end of the 8K"),
opcodes_6502.Literal(
"; RX/TX buffers. Since we're using an 8K socket, that means we don't "
"have to do any work to"),
opcodes_6502.Literal("; manage the read pointer!"),
opcodes_6502.Opcode(4, 3, "STY WADRH"),
opcodes_6502.Opcode(2, 2, "LDA #$00"),
opcodes_6502.Opcode(4, 3, "STA WADRL"),
opcodes_6502.Opcode(6, 3, "JMP (WDATA)"),
]
# Fast path really only requires 1 byte but we have to burn an extra one to
# sync up to 2KB boundary
FAST_PATH_EOF = [opcodes_6502.Opcode(4, 3, "LDA WDATA")] + SLOW_PATH_EOF
def fast_path_trampoline(label: str) -> List[opcodes_6502.Opcode]:
return [
opcodes_6502.Literal("eof_fast_path_%s:", indent=0),
opcodes_6502.STA_C030,
opcodes_6502.Opcode(3, 3, "JMP _eof_fast_path_%s" % label)
]
# def _make_end_of_frame_voltages(cycles) -> numpy.ndarray:
# """Voltage sequence for end-of-frame TCP processing."""
# c = []
# voltage_high = True
# for i, skip_cycles in enumerate(cycles):
# c.extend([1.0 if voltage_high else -1.0] * (skip_cycles - 1))
# if i != len(cycles) - 1:
# voltage_high = not voltage_high
# c.append(1.0 if voltage_high else -1.0)
# return numpy.array(c, dtype=numpy.float32)
#
# # These are duty cycles
# eof_cycles = [
# # (16,6),
# # (14,6),
# # (12,8), # -0.15
# # (14, 10), # -0.10
# # (12,10), # -0.05
# # (4, 40, 4, 40, 4, 40, 4, 6),
# # (4, 38, 6, 38, 6, 38, 6, 6),
# # (4, 36, 8, 36, 8, 36, 8, 6),
# # (4, 34, 10, 34, 10, 34, 10, 6),
# # (4, 32, 12, 32, 12, 32, 12, 6),
# # (4, 30, 14, 30, 14, 30, 14, 6),
# (4, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 6), # 0.0
# (4, 11, 10, 11, 10, 11, 10, 11, 10, 11, 10, 11, 10, 11, 6), # 0.046
# (4, 24, 20, 24, 20, 24, 20, 6), # 0.09
# (4, 10, 8, 10, 8, 10, 8, 10, 8, 10, 8, 10, 8, 10, 8, 10, 6), # 0.11
# (4, 13, 10, 13, 10, 13, 10, 13, 10, 13, 10, 13, 10, 13, 6), # 0.13
# (4, 28, 20, 28, 20, 28, 20, 6), # 0.166
# (4, 26, 18, 26, 18, 26, 18, 6), # 0.18
# (4, 24, 16, 24, 16, 24, 16, 6), # 0.2
#
# # (10, 8, 10, 10, 10, 8), # 0.05
# # (12, 10, 12, 8, 10, 10), # 0.1
# # (4, 10, 8, 10, 8, 10, 8, 10, 8, 10, 8, 10, 8, 10, 8, 10, 8, 10, 6), # 0.15
# # (10, 6, 12, 6), # 0.20
# # (10, 4), # 0.25
# # (14, 4, 10, 6), # 0.30
# # (12, 4), # 0.35
# # (14, 4), # 0.40
# ]
#
import itertools
def _make_end_of_frame_voltages2(cycles) -> numpy.ndarray:
"""Voltage sequence for end-of-frame TCP processing."""
max_len = 140
voltage_high = False
c = [1.0, 1.0, 1.0, -1.0] # STA $C030
for i, skip_cycles in enumerate(itertools.cycle(cycles)):
c.extend([1.0 if voltage_high else -1.0] * (skip_cycles - 1))
voltage_high = not voltage_high
c.append(1.0 if voltage_high else -1.0)
if len(c) >= max_len:
break
c.extend([1.0 if voltage_high else -1.0] * 6) # JMP (WDATA)
return numpy.array(c, dtype=numpy.float32)
def _duty_cycles():
# The player sequence for periods of silence (i.e. 0-valued waveform)
# is a sequence of 10 cycle ticks, so we need to support maintaining
# this during EOF in order to avoid introducing noise during such periods.
res = {0.0: [(20, 10, 10)]}
for i in range(4, 42, 1):
if i == 5:
continue
for j in range(i, 42, 1):
# if j == 5:
# continue
if j < i:
continue
# When first duty cycle is small enough to fit in the stage 1
# trampoline, we can't fit the second duty cycle in the stage 2
# trampoline because we'd need too may variants.
#
# e.g.
#
# eof_trampoline_4:
# STA $C030 ; 4 cycles
# STA $C030 ; 4 cycles
# JMP eof_trampoline_4_stage2 ; 3 cycles
#
# eof_trampoline_4_stage2:
# LDA WDATA ; 4
# STA @0+1 ; 4
# @0: JMP (xxyy) ; 6
#
# eof_trampoline_4_b_stage3:
# ; second duty cycle must land here
if i in {4, 6, 8} and j < 21:
continue
# XXX compute value for duty cycle 2
# Limit to min 22Khz carrier
if (i + j) > 45:
continue
duty = j / (i + j) * 2 - 1
res.setdefault(duty, []).append((i + j, i, j))
cycles = []
for c in sorted(list(res.keys())):
pair = sorted(res[c], reverse=False)[0][1:]
cycles.append(pair)
print(c, pair)
print(len(cycles))
return sorted(cycles, key=lambda p: p[0] + p[1])
eof_cycles = _duty_cycles()
# 1181 error
# Fast EOF
# STA $C030
# JMP $xxxx
# --> 6 bytes, 7 cycles, 7 cycle gap
# Slow EOF
# STA $C030
# LDA WDATA
# STA @0+1
# @0:
# JMP ($20xx)
# 12 bytes, 18 cycles, 18 cycle gap
# 2-stage jump?
# STA $C030
# NOP
# NOP
# STA $C030
# JMP
# 12 cycles
# LDA #$07
# STA $C0x4
# JMP (WDATA) --> indirect to page 7
#
# (7, ... duty cycles)
# aaaa:
# STA $C030
# LDA WDATA
# STA @0+1
# @0:
# JMP $(20xx)
# (7, 18) and up
# (9, ... duty cycles)
# bbbb:
# NOP
# STA $C030
# ....
def audio_opcodes() -> Iterable[opcodes_6502.Opcode]:
# Interleave 3 x STA_C030 with 0,2,4 intervening NOP cycles
# We don't need to worry about 6 or more cycle paddings because
# these can always be achieved by chaining JMP (WDATA) to itself
# for i in range(0, 6, 2):
# for j in range(0, 6, 2):
# ops = []
# # don't need to worry about 0 or 2 since they are subsequences
# ops.extend(opcodes_6502.nops(4))
# ops.append(opcodes_6502.STA_C030)
# if i:
# ops.extend(opcodes_6502.nops(i))
# ops.append(opcodes_6502.STA_C030)
# if j:
# ops.extend(opcodes_6502.nops(j))
# ops.append(opcodes_6502.STA_C030)
# ops.append(opcodes_6502.JMP_WDATA)
# yield tuple(ops)
# Add a NOP sled so we can more efficiently chain together longer
# runs of NOPs without wasting bytes in the TCP frame by chaining
# together JMP (WDATA)
yield tuple(
[nop for nop in opcodes_6502.nops(4)] + [
opcodes_6502.STA_C030, opcodes_6502.JMP_WDATA])
yield tuple(
[nop for nop in opcodes_6502.nops(4)] + [
opcodes_6502.Opcode(3, 2, "STA zpdummy"),
opcodes_6502.STA_C030, opcodes_6502.JMP_WDATA])
# yield tuple(
# [nop for nop in opcodes_6502.nops(20)] + [opcodes_6502.JMP_WDATA])
#
# yield tuple(
# [nop for nop in opcodes_6502.nops(18)] + [
# opcodes_6502.Opcode(3, 2, "STA zpdummy"), opcodes_6502.JMP_WDATA])
# Fill rest of page 3 with trampoline:
# trampoline_7:
# STA $C030
# JMP $aaaa ; 8 bytes each
# trampoline_9:
# STA $C030
# JMP $bbbb
# trampoline_10:
# STA $C030
# JMP $bbbb
# ...
# trampoline_33:
# STA $C030
# JMP $ffff
# 8 bytes each
# Can also do
# trampoline_4:
# STA $C030
# STA $C030
# JMP $1234 ; 12 bytes
# trampoline_6:
# STA $C030
# NOP
# STA $C030
# JMP $2345 ; 13 bytes
# trampoline_8:
# STA $C030
# NOP
# NOP
# STA $C030
# JMP $2345 ; 14 bytes
def eof_trampoline_stage1(cycles):
ops = [
opcodes_6502.Literal(
"eof_trampoline_%d:" % cycles, indent=0
),
opcodes_6502.STA_C030,
]
if cycles == 4:
return ops + [
opcodes_6502.STA_C030,
opcodes_6502.Opcode(3, 3, "JMP eof_trampoline_%d_stage2" % cycles)
]
if cycles == 5:
return None
if cycles == 6:
return ops + [
opcodes_6502.Opcode(2, 1, "NOP"),
opcodes_6502.STA_C030,
opcodes_6502.Opcode(3, 3, "JMP eof_trampoline_%d_stage2" % cycles)
]
if cycles == 8:
return ops + [
opcodes_6502.Opcode(2, 1, "NOP"),
opcodes_6502.Opcode(2, 1, "NOP"),
opcodes_6502.STA_C030,
opcodes_6502.Opcode(3, 3, "JMP eof_trampoline_%d_stage2" % cycles)
]
return ops + [
opcodes_6502.Opcode(3, 3, "JMP eof_trampoline_%d_stage2" % cycles)
]
def eof_trampoline_stage2(cycles, trampoline_page):
label = [
opcodes_6502.Literal(
"eof_trampoline_%d_stage2:" % cycles, indent=0
)
]
ops = [
opcodes_6502.Opcode(4, 3, "LDA WDATA"),
opcodes_6502.Opcode(4, 3, "STA @0+1"),
opcodes_6502.Literal("@0:", indent=0),
opcodes_6502.Opcode(
6, 3, "JMP (eof_trampoline_stage3_page%d)" % trampoline_page)
]
if cycles < 7 or cycles == 8:
return label + ops
# For cycles == 7 or > 8 we need to interleave a STA $C030 into stage 2
# because we couldn't fit it in stage 1
interleave_ops = [
opcodes_6502.padding(cycles - 7),
opcodes_6502.STA_C030,
opcodes_6502.padding(100)
]
return label + list(opcodes_6502.interleave_opcodes(interleave_ops, ops))
def assign_eof_trampoline_stage3_pages(duty_cycles):
second_cycles = {}
for a, b in sorted(duty_cycles):
second_cycles.setdefault(a, []).append(b)
# bin-pack the (a, b) duty cycles into pages so we can set up indirect
# jump tables to dispatch the third stage trampoline. A greedy algorithm
# works fine here
pages = []
page = []
longest_first_cycles = sorted(
list(second_cycles.items()), key=lambda c: len(c[1]), reverse=True)
left = len(longest_first_cycles)
while left:
for i, cycles in enumerate(longest_first_cycles):
if cycles is None:
continue
cycle1, cycles2 = cycles
if len(page) < (128 - len(cycles2)):
page.extend((cycle1, cycle2) for cycle2 in cycles2)
longest_first_cycles[i] = None
left -= 1
pages.append(page)
page = []
page_offsets = {}
for page_idx, page in enumerate(pages):
offset = 0
for a, b in page:
offset += 2
page_offsets[(a, b)] = (page_idx, offset)
return page_offsets
def generate_player(
player_ops: Iterable[opcodes_6502.Opcode],
opcode_filename: str,
player_stage1_filename: str,
player_stage2_filename: str
):
num_bytes = 0
seen_op_suffix_toggles = set()
offset = 0
unique_entrypoints = {}
toggles = {}
with open(player_stage1_filename, "w+") as f:
for i, ops in enumerate(player_ops):
player_op = []
for j, op in enumerate(ops):
op_suffix_toggles = opcodes_6502.toggles(ops[j:])
if op_suffix_toggles not in seen_op_suffix_toggles:
# new subsequence
seen_op_suffix_toggles.add(op_suffix_toggles)
player_op.append(
opcodes_6502.Literal(
"tick_%02x: ; voltages %s" % (
offset, op_suffix_toggles), indent=0))
unique_entrypoints[offset] = op_suffix_toggles
player_op.append(op)
offset += op.bytes
assert unique_entrypoints
player_op_len = opcodes_6502.total_bytes(player_op)
# Make sure we reserve 9 bytes for END_OF_FRAME and EXIT
assert (num_bytes + player_op_len) <= (256 - 9)
for op in player_op:
f.write("%s\n" % str(op))
num_bytes += player_op_len
f.write("\n")
duty_cycles = _duty_cycles()
duty_cycle_first = sorted(list(set(dc[0] for dc in duty_cycles)))
for eof_stage1_cycles in duty_cycle_first:
eof_stage1_ops = eof_trampoline_stage1(eof_stage1_cycles)
if not eof_stage1_ops:
continue
for op in eof_stage1_ops:
f.write("%s\n" % str(op))
f.write("\n")
num_bytes += opcodes_6502.total_bytes(eof_stage1_ops)
f.write("; %d entrypoints, %d bytes\n" % (
len(unique_entrypoints), num_bytes))
# XXX if we're spilling the STA $C030 onto stage 2 then we can accommodate
# lower values for b
# it's only the ones where we have the STA $C030 on stage 1 that we have
# a lower bound of b >= 14 cycles
with open(player_stage2_filename, "w+") as f:
eof_stage3_page_offsets = assign_eof_trampoline_stage3_pages(
duty_cycles)
# We bin pack each (a, b) duty cycle onto the same jump table page
pages_by_first_duty_cycle = {}
for ab, po in eof_stage3_page_offsets.items():
pages_by_first_duty_cycle[ab[0]] = po[0]
for eof_stage1_cycles in duty_cycle_first:
page = pages_by_first_duty_cycle[eof_stage1_cycles]
eof_stage2_ops = eof_trampoline_stage2(eof_stage1_cycles, page)
if not eof_stage2_ops:
continue
for op in eof_stage2_ops:
f.write("%s\n" % str(op))
f.write("\n")
for a, b in duty_cycles:
duty_cycle_ops = itertools.chain(
opcodes_6502.padding(b -
)
with open(opcode_filename, "w") as f:
f.write("import enum\nimport numpy\n\n\n")
f.write("class Opcode(enum.Enum):\n")
for o in unique_entrypoints.keys():
f.write(" TICK_%02x = 0x%02x\n" % (o, o))
f.write(" EXIT = 0x%02x\n" % num_bytes)
# f.write(" END_OF_FRAME = 0x%02x\n" % (num_bytes + 3))
for i, _ in enumerate(eof_cycles):
f.write(" END_OF_FRAME_%d = 0x%02x\n" % (i, num_bytes + 4 + i))
f.write("\n\nVOLTAGE_SCHEDULE = {\n")
for o, v in unique_entrypoints.items():
f.write(
" Opcode.TICK_%02x: numpy.array(%s, dtype=numpy.float32),"
"\n" % (o, v))
for i, skip_cycles in enumerate(eof_cycles):
f.write(" Opcode.END_OF_FRAME_%d: numpy.array([%s], "
"dtype=numpy.float32), # %s\n" % (i, ", ".join(
str(f) for f in _make_end_of_frame_voltages2(
skip_cycles)), skip_cycles))
f.write("}\n")
#
# f.write("\n\nTOGGLES = {\n")
# for o, v in toggles.items():
# f.write(
# " Opcode.TICK_%02x: %d,\n" % (o, v)
# )
# f.write("}\n")
#
f.write("\n\nEOF_OPCODES = (\n")
for i in range(len(eof_cycles)):
f.write(" Opcode.END_OF_FRAME_%d,\n" % i)
f.write(")\n")
def main():
player_ops = audio_opcodes()
generate_player(
player_ops,
2020-10-15 13:27:37 +00:00
opcode_filename="opcodes_generated.py",
player_stage1_filename="player/player_generated.s",
player_stage2_filename="player/player_stage2_generated.s"
)
if __name__ == "__main__":
main()