int main(int argc, char **argv) {
// show_qpu_frament(address, instruction count);
for (int i=8; i<17; i++) {
int size = 1<<i;
char filename[256];
sprintf(filename, "shader_%d%s.s", size<1024?size:size/1024, size<1024?"":"k");
FILE *fp = freopen (filename, "w", stdout);
puts(
"/*\n"
"Disassembly of Andrew Holme's QPU FFT for Raspberry Pi.\n"
"From: https://github.com/raspberrypi/firmware/tree/master/opt/vc/src/hello_pi/hello_fft\n"
"*/\n"
);
copyright();
printf("shader_%d%s:\n", (1<<i)<=512?(1<<i):(1<<i)/1024, (1<<i)<=512?"":"k");
show_qpu_fragment(gpu_fft_shader_code(i), gpu_fft_shader_size(i)/4);
printf("\n");
fclose(fp);
}
}
int gpu_fft_prepare(
int mb, // mailbox file_desc
int log2_N, // log2(FFT_length) = 8...22
int direction, // GPU_FFT_FWD: fft(); GPU_FFT_REV: ifft()
int jobs, // number of transforms in batch
struct GPU_FFT **fft) {
unsigned info_bytes, twid_bytes, fft_bytes, data_bytes,
fft_bytes_4k, buff_bytes, num_buff,
code_bytes, unif_bytes, mail_bytes;
unsigned size, *uptr, vc_tw, vc_buff, vc_data;
int i, j, q, shared, unique, passes, ret;
struct GPU_FFT_PTR ptr;
struct GPU_FFT *info;
if (gpu_fft_twiddle_size(log2_N, &shared, &unique, &passes)) return -2;
info_bytes = 4096;
fft_bytes = sizeof(COMPLEX)<<log2_N;
code_bytes = gpu_fft_shader_size(log2_N);
twid_bytes = sizeof(COMPLEX)*16*(shared+GPU_FFT_QPUS*unique);
unif_bytes = sizeof(int)*GPU_FFT_QPUS*(4+2*jobs*passes);
mail_bytes = sizeof(int)*GPU_FFT_QPUS*2;
// (MM) Use as few temporary buffers as possible
if (fft_bytes >= GPU_FFT_MIN_BUFF_SIZE)
// FFT size larger than cache => use exactly one buffer
num_buff = 1;
else
{ // use up to GPU_FFT_TEMP_BUFF_SIZE bytes buffer, but no more than needed
num_buff = GPU_FFT_MIN_BUFF_SIZE / fft_bytes;
i = (passes-1) * jobs;
if (i > GPU_FFT_MAX_BUFF_COUNT)
i = GPU_FFT_MAX_BUFF_COUNT;
if (num_buff > i)
num_buff = i;
}
// Missing 4k alignment for 256 points
fft_bytes_4k = ((fft_bytes-1)|4095)+1;
buff_bytes = fft_bytes_4k * num_buff;
data_bytes = fft_bytes * jobs;
size = info_bytes + // header
buff_bytes + // additional buffers
data_bytes + // source data
code_bytes + // shader, aligned
twid_bytes + // twiddles
unif_bytes + // uniforms
mail_bytes; // mailbox message
//fprintf(stderr, \
"fft_bytes = %x\n" \
"code_bytes = %x\n" \
"twid_bytes = %x\n" \
"unif_bytes = %x\n" \
"mail_bytes = %x\n" \
"buff_bytes = %x\n" \
"data_bytes = %x\n" \
"num_buff = %x\n" \
"size = %x\n", \
fft_bytes, code_bytes, twid_bytes, unif_bytes, mail_bytes, buff_bytes, data_bytes, num_buff, size);
ret = gpu_fft_alloc(mb, size, &ptr);
if (ret) return ret;
// Header
info = (struct GPU_FFT *) ptr.arm.vptr;
gpu_fft_ptr_inc(&ptr, info_bytes);
// For transpose
info->x = 1<<log2_N;
info->y = jobs;
// (MM) Use dedicated temporary buffers so all transforms can operate in place,
// except if we only have 1 buffer.
info->out = ptr.arm.cptr;
vc_buff = gpu_fft_ptr_inc(&ptr, buff_bytes);
info->step = fft_bytes / sizeof(COMPLEX);
info->in = ptr.arm.cptr;
if (num_buff > 1 || !(passes & 1))
info->out = info->in;
vc_data = gpu_fft_ptr_inc(&ptr, data_bytes);
//fprintf(stderr, \
"vc_buff = %x\n" \
"vc_data = %x\n" \
"in = %x\n" \
"out = %x\n", \
vc_buff, vc_data, info->in, info->out);
// Shader code
memcpy(ptr.arm.vptr, gpu_fft_shader_code(log2_N), code_bytes);
info->base.vc_code = gpu_fft_ptr_inc(&ptr, code_bytes);
// Twiddles
gpu_fft_twiddle_data(log2_N, direction, ptr.arm.fptr);
vc_tw = gpu_fft_ptr_inc(&ptr, twid_bytes);
uptr = ptr.arm.uptr;
// Uniforms
for (q=0; q<GPU_FFT_QPUS; q++) {
int current_buff = num_buff-1;
*uptr++ = vc_tw;
*uptr++ = vc_tw + sizeof(COMPLEX)*16*(shared + q*unique);
*uptr++ = q;
for (i=0; i < jobs; ++i) {
unsigned data = vc_data + i * fft_bytes;
*uptr++ = data;
if (num_buff == 1)
{ // use ping pong buffers
unsigned buff = vc_buff + current_buff * fft_bytes_4k;
for (j = 1; j < passes; j++)
{ uptr[0] = uptr[1] = buff;
// swap buffers
buff = data;
data = uptr[0];
uptr += 2;
}
if (passes & 1)
++current_buff;
*uptr++ = buff;
} else
{ // use dedicated buffers
for (j = 1; j < passes; j++)
{ // round robin
current_buff = (current_buff+1) % num_buff;
uptr[0] = uptr[1] = vc_buff + current_buff * fft_bytes_4k;
uptr += 2;
}
*uptr++ = data;
}
}
*uptr++ = 0;
//if (q == 0) for (i = -(4+2*jobs*passes); i < 0; ++i) fprintf(stderr, "%x\n", uptr[i]);
info->base.vc_unifs[q] = gpu_fft_ptr_inc(&ptr, sizeof(int)*(4+2*jobs*passes));
}
if ((jobs<<log2_N) <= GPU_FFT_BUSY_WAIT_LIMIT) {
// Direct register poking with busy wait
info->base.vc_msg = 0;
}
else {
// Mailbox message
for (q=0; q<GPU_FFT_QPUS; q++) {
*uptr++ = info->base.vc_unifs[q];
*uptr++ = info->base.vc_code;
}
info->base.vc_msg = ptr.vc;
}
*fft = info;
return 0;
}
int gpu_fft_prepare(
int mb, // mailbox file_desc
int log2_N, // log2(FFT_length) = 8...22
int direction, // GPU_FFT_FWD: fft(); GPU_FFT_REV: ifft()
int jobs, // number of transforms in batch
struct GPU_FFT **fft) {
uint32_t *uptr;
unsigned info_bytes, twid_bytes, data_bytes, code_bytes, unif_bytes, mail_bytes;
unsigned size, vc_tw, vc_data;
int i, q, shared, unique, passes, ret;
struct GPU_FFT_BASE *base;
struct GPU_FFT_PTR ptr;
struct GPU_FFT *info;
if (gpu_fft_twiddle_size(log2_N, &shared, &unique, &passes)) return -2;
info_bytes = 4096;
data_bytes = (1+((sizeof(COMPLEX)<<log2_N)|4095));
code_bytes = gpu_fft_shader_size(log2_N);
twid_bytes = sizeof(COMPLEX)*16*(shared+GPU_FFT_QPUS*unique);
unif_bytes = sizeof(int32_t)*GPU_FFT_QPUS*(5+jobs*2);
mail_bytes = sizeof(int32_t)*GPU_FFT_QPUS*2;
size = info_bytes + // header
data_bytes*jobs*2 + // ping-pong data, aligned
code_bytes + // shader, aligned
twid_bytes + // twiddles
unif_bytes + // uniforms
mail_bytes; // mailbox message
ret = gpu_fft_alloc(mb, size, &ptr);
if (ret) return ret;
// Header
info = (struct GPU_FFT *) ptr.arm.vptr;
base = (struct GPU_FFT_BASE *) info;
gpu_fft_ptr_inc(&ptr, info_bytes);
// For transpose
info->x = 1<<log2_N;
info->y = jobs;
// Ping-pong buffers leave results in or out of place
info->in = info->out = ptr.arm.cptr;
info->step = data_bytes / sizeof(COMPLEX);
if (passes&1) info->out += info->step * jobs; // odd => out of place
vc_data = gpu_fft_ptr_inc(&ptr, data_bytes*jobs*2);
// Shader code
memcpy(ptr.arm.vptr, gpu_fft_shader_code(log2_N), code_bytes);
base->vc_code = gpu_fft_ptr_inc(&ptr, code_bytes);
// Twiddles
gpu_fft_twiddle_data(log2_N, direction, ptr.arm.fptr);
vc_tw = gpu_fft_ptr_inc(&ptr, twid_bytes);
uptr = ptr.arm.uptr;
// Uniforms
for (q=0; q<GPU_FFT_QPUS; q++) {
*uptr++ = vc_tw;
*uptr++ = vc_tw + sizeof(COMPLEX)*16*(shared + q*unique);
*uptr++ = q;
for (i=0; i<jobs; i++) {
*uptr++ = vc_data + data_bytes*i;
*uptr++ = vc_data + data_bytes*i + data_bytes*jobs;
}
*uptr++ = 0;
*uptr++ = (q==0); // For mailbox: IRQ enable, master only
base->vc_unifs[q] = gpu_fft_ptr_inc(&ptr, sizeof(int32_t)*(5+jobs*2));
}
if ((jobs<<log2_N) <= GPU_FFT_BUSY_WAIT_LIMIT) {
// Direct register poking with busy wait
base->vc_msg = 0;
}
else {
// Mailbox message
for (q=0; q<GPU_FFT_QPUS; q++) {
*uptr++ = base->vc_unifs[q];
*uptr++ = base->vc_code;
}
base->vc_msg = ptr.vc;
}
*fft = info;
return 0;
}
