birdulon/ChocolateBird
birdulon
/
ChocolateBird
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[BGM] slap highp everywhere to ward off destructive gpu driver "optimisations"

This commit is contained in:
Luke Hubmayer-Werner 2024-07-16 20:26:18 +09:30
parent 0b2a120b1d
commit 72dcc11945
1 changed files with 85 additions and 85 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

170
shaders/audio_renderer.gdshader
View File

@ -3,66 +3,66 @@
// Unfortunately, this loses type-checking on [0.0, 1.0] vs [0,255] etc. so a lot of this will involve comments declaring ranges.
shader_type canvas_item;
render_mode blend_premul_alpha;
uniform sampler2D instrument_samples;
uniform vec2 instrument_samples_size = vec2(2048.0, 128.0);
uniform int INT_OUTPUT_WIDTH = 4096;
uniform vec2 OUTPUT_FRAMEBUFFER_SIZE = vec2(4096.0, 4096.0);
uniform float reference_note = 71.0; // [0, 255], possibly [0, 127]
uniform float output_mixrate = 32000.0; // SNES SPC output is 32kHz
uniform vec2 midi_events_size = vec2(2048.0, 32.0);
uniform int tempo_scale_thousandths = 1000;
const int TEMPO_SCALE_MULTIPLIER = 1000;
uniform highp sampler2D instrument_samples;
uniform highp vec2 instrument_samples_size = vec2(2048.0, 128.0);
uniform highp int INT_OUTPUT_WIDTH = 4096;
uniform highp vec2 OUTPUT_FRAMEBUFFER_SIZE = vec2(4096.0, 4096.0);
uniform highp float reference_note = 71.0; // [0, 255], possibly [0, 127]
uniform highp float output_mixrate = 32000.0; // SNES SPC output is 32kHz
uniform highp vec2 midi_events_size = vec2(2048.0, 32.0);
uniform highp int tempo_scale_thousandths = 1000;
const highp int TEMPO_SCALE_MULTIPLIER = 1000;
// I feel like these magic numbers are a bit more intuitive in hex
const float x00FF = float(0x00FF); // 255.0
const float x0100 = float(0x0100); // 256.0
const float x7FFF = float(0x7FFF); // 32767.0
const float x8000 = float(0x8000); // 32768.0
const float xFF00 = float(0xFF00); // 65280.0
const float xFFFF = float(0xFFFF); // 65535.0
const float x10000 = float(0x10000); // 65536.0
const float x00FF0000 = float(0x00FF0000);
const float xFF000000 = float(0xFF000000);
const highp float x00FF = float(0x00FF); // 255.0
const highp float x0100 = float(0x0100); // 256.0
const highp float x7FFF = float(0x7FFF); // 32767.0
const highp float x8000 = float(0x8000); // 32768.0
const highp float xFF00 = float(0xFF00); // 65280.0
const highp float xFFFF = float(0xFFFF); // 65535.0
const highp float x10000 = float(0x10000); // 65536.0
const highp float x00FF0000 = float(0x00FF0000);
const highp float xFF000000 = float(0xFF000000);
const vec2 INT16_DOT_BE = vec2(xFF00, x00FF);
const vec2 INT16_DOT_LE = vec2(x00FF, xFF00);
const vec4 INT32_DOT_LE = vec4(x00FF, xFF00, x00FF0000, xFF000000);
const highp vec2 INT16_DOT_BE = vec2(xFF00, x00FF);
const highp vec2 INT16_DOT_LE = vec2(x00FF, xFF00);
const highp vec4 INT32_DOT_LE = vec4(x00FF, xFF00, x00FF0000, xFF000000);
float unpack_uint16(vec2 uint16) {
highp float unpack_uint16(highp vec2 uint16) {
// Convert packed 2byte integer, sampled as two [0.0, 1.0] range floats, to the original int value [0, 65535] in float32
return dot(uint16, INT16_DOT_LE);
}
float unpack_uint32_to_float(vec4 uint32) {
highp float unpack_uint32_to_float(highp vec4 uint32) {
// Convert packed 4byte integer, sampled as four [0.0, 1.0] range floats, to the original int value [0, 0xFFFFFFFF] in float32
// NOTE: THIS WILL LOSE PRECISION ON NUMBERS ABOVE 24BIT SIGNIFICANCE
// I CAN'T EVEN GUARANTEE THE 0xFF000000 CONSTANT WILL SURVIVE ROUNDING
return dot(uint32, INT32_DOT_LE);
}
int unpack_int32(vec4 int32) {
highp int unpack_int32(highp vec4 int32) {
// Convert packed 4byte integer, sampled as four [0.0, 1.0] range floats, to the original int value
// return int(unpack_uint16(int32.xy)) + (int(unpack_uint16(int32.zw)) << 16);
return int(unpack_uint16(int32.xy)) + (int(unpack_uint16(int32.zw)) * 0x10000);
}
float unpack_int16(vec2 int16) {
highp float unpack_int16(highp vec2 int16) {
// Convert packed 2byte integer, sampled as two [0.0, 1.0] range floats, to the original int value [-32768, 32767] in float32
float unsigned = dot(int16, INT16_DOT_LE);
highp float unsigned = dot(int16, INT16_DOT_LE);
return unsigned - (unsigned < x7FFF ? 0.0 : x10000);
}
float rescale_int16(float int16) {
highp float rescale_int16(highp float int16) {
// Rescale from [-32768, 32767] to [-1.0, 1.0)
return int16 / x8000;
}
vec2 pack_float_to_int16(float value) {
highp vec2 pack_float_to_int16(highp float value) {
// Convert a float in range [-1.0, 1.0) to a signed 2byte integer [-32768, 32767] packed into two [0.0, 1.0] floats
float scaled = value * x8000;
float unsigned = scaled + (scaled < 0.0 ? x10000 : 0.0);
float unsigned_div_256 = unsigned / x0100;
float MSB = trunc(unsigned_div_256) / x00FF;
float LSB = fract(unsigned_div_256) * x0100 / x00FF;
highp float scaled = value * x8000;
highp float unsigned = scaled + (scaled < 0.0 ? x10000 : 0.0);
highp float unsigned_div_256 = unsigned / x0100;
highp float MSB = trunc(unsigned_div_256) / x00FF;
highp float LSB = fract(unsigned_div_256) * x0100 / x00FF;
return vec2(LSB, MSB);
}
@ -100,50 +100,50 @@ vec2 pack_float_to_int16(float value) {
// With the 258 texel header, which uses 3 texels of margin, 255 would be subtracted from the above payload,
// leaving 261121 texels for the sample data.
const float HEADER_LENGTH_TEXELS = 5.0;
const int INSTRUMENT_SAMPLES_WIDTH = 2048;
float sinc(float x) {
const highp float HEADER_LENGTH_TEXELS = 5.0;
const highp int INSTRUMENT_SAMPLES_WIDTH = 2048;
highp float sinc(highp float x) {
x = abs(x) + 0.00000000000001; // Avoid division by zero
return min(sin(x)/x, 1.0);
}
float get_pitch_scale(float note) {
highp float get_pitch_scale(highp float note) {
return exp2((note - reference_note)/12.0);
}
vec2 get_inst_texel(vec2 xy) {
highp vec2 get_inst_texel(highp vec2 xy) {
return texture(instrument_samples, (xy+0.5)/instrument_samples_size).xw;
}
float get_inst_texel_int16(int smp) {
int x = smp % INSTRUMENT_SAMPLES_WIDTH;
int y = smp / INSTRUMENT_SAMPLES_WIDTH;
highp float get_inst_texel_int16(highp int smp) {
highp int x = smp % INSTRUMENT_SAMPLES_WIDTH;
highp int y = smp / INSTRUMENT_SAMPLES_WIDTH;
return unpack_int16(texture(instrument_samples, (vec2(float(x), float(y)) + 0.5)/instrument_samples_size).xw);
}
float get_instrument_sample(float instrument_index, float note, float t) {
float header_offset = instrument_index * HEADER_LENGTH_TEXELS;
int smp_start = unpack_int32(vec4(get_inst_texel(vec2(header_offset, 0.0)), get_inst_texel(vec2(header_offset + 1.0, 0.0)))); // The true start, after the prepended frames of silence
float smp_loop_begin = unpack_uint16(get_inst_texel(vec2(header_offset + 2.0, 0.0))); // padded past the true loop point for filter
float smp_loop_length = unpack_uint16(get_inst_texel(vec2(header_offset + 3.0, 0.0)));
float sample_mixrate = unpack_uint16(get_inst_texel(vec2(header_offset + 4.0, 0.0)));
highp float get_instrument_sample(highp float instrument_index, highp float note, highp float t) {
highp float header_offset = instrument_index * HEADER_LENGTH_TEXELS;
highp int smp_start = unpack_int32(vec4(get_inst_texel(vec2(header_offset, 0.0)), get_inst_texel(vec2(header_offset + 1.0, 0.0)))); // The true start, after the prepended frames of silence
highp float smp_loop_begin = unpack_uint16(get_inst_texel(vec2(header_offset + 2.0, 0.0))); // padded past the true loop point for filter
highp float smp_loop_length = unpack_uint16(get_inst_texel(vec2(header_offset + 3.0, 0.0)));
highp float sample_mixrate = unpack_uint16(get_inst_texel(vec2(header_offset + 4.0, 0.0)));
// Calculate the point we want to sample in linear space
float mixrate = sample_mixrate * get_pitch_scale(note);
float smp_t = t * mixrate;
highp float mixrate = sample_mixrate * get_pitch_scale(note);
highp float smp_t = t * mixrate;
// If we're past the end of the sample, we need to wrap it back to within the loop range
float overshoot = max(smp_t - smp_loop_begin, 0.0);
highp float overshoot = max(smp_t - smp_loop_begin, 0.0);
smp_t -= floor(overshoot/smp_loop_length) * smp_loop_length;
// if (smp_t > smp_loop_begin) {
// // return 0.0;
// smp_t = mod(smp_t - smp_loop_begin, smp_loop_length) + smp_loop_begin;
// }
int smp_window_start = smp_start + int(smp_t) - 6;
float smp_rel_filter_target = fract(smp_t) + 6.0;
float output = 0.0;
highp int smp_window_start = smp_start + int(smp_t) - 6;
highp float smp_rel_filter_target = fract(smp_t) + 6.0;
highp float output = 0.0;
for (int i = 0; i < 12; i++) {
int smp_filter = smp_window_start + i;
float s = get_inst_texel_int16(smp_filter);
highp int smp_filter = smp_window_start + i;
highp float s = get_inst_texel_int16(smp_filter);
// TODO: determine proper value for this. Might be based on instrument base mixrate.
output += s * sinc((smp_rel_filter_target - float(i)) * 3.1);
}
@ -153,60 +153,60 @@ float get_instrument_sample(float instrument_index, float note, float t) {
}
const int NUM_CHANNELS = 8;
const int MAX_CHANNEL_NOTE_EVENTS = 2048;
const highp int MAX_CHANNEL_NOTE_EVENTS = 2048;
const int NUM_CHANNEL_NOTE_PROBES = 11; // log2(MAX_CHANNEL_NOTE_EVENTS)
vec4 get_midi_texel(sampler2D tex, float x, float y) {
highp vec4 get_midi_texel(highp sampler2D tex, highp float x, highp float y) {
return texture(tex, vec2(x, y)/midi_events_size).xyzw;
}
int retime_smp(int smp) {
highp int retime_smp(highp int smp) {
// Overflow safety is important as our input values can go up to 2^24, and we multiply by around 2^10
int factor = smp / tempo_scale_thousandths;
int residue = smp % tempo_scale_thousandths;
int a = (residue * TEMPO_SCALE_MULTIPLIER) / tempo_scale_thousandths;
int b = factor * TEMPO_SCALE_MULTIPLIER;
highp int factor = smp / tempo_scale_thousandths;
highp int residue = smp % tempo_scale_thousandths;
highp int a = (residue * TEMPO_SCALE_MULTIPLIER) / tempo_scale_thousandths;
highp int b = factor * TEMPO_SCALE_MULTIPLIER;
return a + b;
}
vec4 render_song(sampler2D tex, int smp) {
highp vec4 render_song(highp sampler2D tex, highp int smp) {
// Each output texel rendered is a stereo S16LE frame representing 1/32000 of a second
// 2048 is an established safe texture dimension so may as well go 2048 wide
float t = float(smp)/output_mixrate;
vec2 downmixed_stereo = vec2(0.0);
highp float t = float(smp)/output_mixrate;
highp vec2 downmixed_stereo = vec2(0.0);
// Binary search the channels
for (int channel = 0; channel < NUM_CHANNELS; channel++) {
float row = float(channel * 4);
float event_idx = 0.0;
int smp_start;
highp float row = float(channel * 4);
highp float event_idx = 0.0;
highp int smp_start;
for (int i = 0; i < NUM_CHANNEL_NOTE_PROBES; i++) {
float step_size = exp2(float(NUM_CHANNEL_NOTE_PROBES - i - 1));
highp float step_size = exp2(float(NUM_CHANNEL_NOTE_PROBES - i - 1));
smp_start = retime_smp(int(unpack_int32(get_midi_texel(tex, event_idx + step_size, row))));
event_idx += (smp >= smp_start) ? step_size : 0.0;
}
smp_start = retime_smp(int(unpack_int32(get_midi_texel(tex, event_idx, row))));
int smp_end = retime_smp(int(unpack_int32(get_midi_texel(tex, event_idx, row+1.0))));
highp int smp_end = retime_smp(int(unpack_int32(get_midi_texel(tex, event_idx, row+1.0))));
vec4 note_event_supplement = get_midi_texel(tex, event_idx, row+2.0); // left as [0.0, 1.0]
float instrument_idx = trunc(note_event_supplement.x * 255.0);
float pitch_idx = note_event_supplement.y * 255.0;
float velocity = note_event_supplement.z;
float pan = note_event_supplement.w;
vec4 adsr = get_midi_texel(tex, event_idx, row+3.0); // left as [0.0, 1.0]
highp vec4 note_event_supplement = get_midi_texel(tex, event_idx, row+2.0); // left as [0.0, 1.0]
highp float instrument_idx = trunc(note_event_supplement.x * 255.0);
highp float pitch_idx = note_event_supplement.y * 255.0;
highp float velocity = note_event_supplement.z;
highp float pan = note_event_supplement.w;
highp vec4 adsr = get_midi_texel(tex, event_idx, row+3.0); // left as [0.0, 1.0]
// ====================At some point I'll look back into packing floats====================
// TBD = note_event_supplement.zw; - tremolo/vibrato/noise/pan_lfo/pitchbend/echo remain
// ====================At some point I'll look back into packing floats====================
float attack = 1.0 + adsr.x*255.0; //65535.0 + 1.0; // TODO: work out effective resolution for this
int smp_attack = int(attack) * 2; // Max value is 131072 samples = 4.096 seconds
highp float attack = 1.0 + adsr.x*255.0; //65535.0 + 1.0; // TODO: work out effective resolution for this
highp int smp_attack = int(attack) * 2; // Max value is 131072 samples = 4.096 seconds
// For now, just branch this
if (smp_start < smp) { // First sample may not start at zero!
int smp_overrun = smp - smp_end; // 256 samples of linear decay to 0 after note_off
highp int smp_overrun = smp - smp_end; // 256 samples of linear decay to 0 after note_off
smp_overrun = (smp_overrun < 0) ? 0 : smp_overrun;
if (smp_overrun < 256) {
float t_start = float(smp_start)/output_mixrate;
float attack_factor = min(float(smp - smp_start)/float(smp_attack), 1.0);
float release_factor = float(255-smp_overrun)/255.0; // 256 samples of linear decay to 0 after note_off
float samp = get_instrument_sample(instrument_idx, pitch_idx, t-t_start);
highp float t_start = float(smp_start)/output_mixrate;
highp float attack_factor = min(float(smp - smp_start)/float(smp_attack), 1.0);
highp float release_factor = float(255-smp_overrun)/255.0; // 256 samples of linear decay to 0 after note_off
highp float samp = get_instrument_sample(instrument_idx, pitch_idx, t-t_start);
samp *= velocity * attack_factor * release_factor;
// TODO: proper decay and sustain, revisit release
downmixed_stereo += samp * vec2(pan, 1.0-pan) * 0.5; // TODO: double it to maintain the mono level on each channel at center=0.5?
@ -219,9 +219,9 @@ vec4 render_song(sampler2D tex, int smp) {
void fragment() {
// GLES2
vec2 uv = vec2(UV.x, 1.0-UV.y);
highp vec2 uv = vec2(UV.x, 1.0-UV.y);
// uv = (trunc(uv*OUTPUT_FRAMEBUFFER_SIZE)+0.5)/OUTPUT_FRAMEBUFFER_SIZE;
// COLOR.xyzw = test_writeback(TEXTURE, uv);
ivec2 xy = ivec2(trunc(uv*OUTPUT_FRAMEBUFFER_SIZE));
highp ivec2 xy = ivec2(trunc(uv*OUTPUT_FRAMEBUFFER_SIZE));
COLOR.xyzw = render_song(TEXTURE, xy.x + (xy.y*INT_OUTPUT_WIDTH));
}