Überflix-Optimierung

This commit is contained in:
Jörg Thalheim 2015-03-10 15:01:22 +01:00
parent 17df4c1c65
commit 893f42fca7
2 changed files with 25 additions and 32 deletions

8
fft.c
View File

@ -66,7 +66,7 @@ int fix_fft(fixed fr[], fixed fi[], int m, int inverse)
/* decimation in time - re-order data */
for(m=0;;) {
FFT_bit_reverse(m, mr, mm);
FFT_BIT_REVERSE(m, mr, mm);
if(m >= nn) break;
if(mr <= m) continue;
@ -90,7 +90,7 @@ int fix_fft(fixed fr[], fixed fi[], int m, int inverse)
shift = 0;
for(i=0; i<n/8; i+=8)
{
if(FFT_shift_check(fr, i) | FFT_shift_check(fi, i))
if(FFT_SHIFT_CHECK(fr, i) | FFT_SHIFT_CHECK(fi, i))
{
shift = 1;
++scale;
@ -111,7 +111,7 @@ int fix_fft(fixed fr[], fixed fi[], int m, int inverse)
on each data point exactly once, during this pass. */
istep = l << 1; //step width of current butterfly
FFT_reg reg;
FFT_REG reg;
fixed *reg_s = ((fixed*) &reg);
for(m=0; m<n; m+=istep)
@ -125,7 +125,7 @@ int fix_fft(fixed fr[], fixed fi[], int m, int inverse)
reg_s[1] = fi[i];
reg_s[0] = fi[j];
FFT_calc(reg, i << k, (xtbool) shift, inverse);
FFT_CALC(reg, i << k, (xtbool) shift, inverse);
fr[i] = reg_s[3];
fr[j] = reg_s[2];

View File

@ -1,4 +1,4 @@
table sin_wave 16 1024 {
table SIN_WAVE 16 1024 {
0, 201, 402, 603, 804, 1005, 1206, 1406,
1607, 1808, 2009, 2209, 2410, 2610, 2811, 3011,
3211, 3411, 3611, 3811, 4011, 4210, 4409, 4608,
@ -130,9 +130,9 @@ table sin_wave 16 1024 {
63929, 64130, 64330, 64531, 64732, 64933, 65134, 65335
}
regfile fft_reg 64 2 fftv
regfile FFT_REG 64 2 fftv
operation fft_shift_check {in AR *base, in AR offset, out AR needs_shift} {out VAddr, in MemDataIn128}
operation FFT_SHIFT_CHECK {in AR *base, in AR offset, out AR needs_shift} {out VAddr, in MemDataIn128}
{
assign VAddr = TIEadd(base, offset[31:1], 1'b0);
@ -158,7 +158,7 @@ operation fft_shift_check {in AR *base, in AR offset, out AR needs_shift} {out V
}
operation fft_bit_reverse {inout AR m, out AR mr, in AR mm} {}
operation FFT_BIT_REVERSE {inout AR m, out AR mr, in AR mm} {}
{
wire [15:0] x = TIEadd(m, 1'b0, 1'b1);
assign mr = {
@ -181,22 +181,22 @@ operation fft_bit_reverse {inout AR m, out AR mr, in AR mm} {}
assign m = x;
}
function [31:0] fft_twiddle ([31:0] j, [0:0] shift, [0:0] inverse)
function [31:0] FFT_TWIDDLE ([31:0] j, [0:0] shift, [0:0] inverse)
{
// 256 = N_WAVE / 4
wire [9:0] idx = TIEadd(j, 256, 1'b0);
wire [15:0] sin = sin_wave[j];
wire [15:0] wr1 = sin_wave[idx];
wire [15:0] sin = SIN_WAVE[j];
wire [15:0] wr1 = SIN_WAVE[idx];
wire [15:0] wi1 = TIEadd(~sin, 16'b0, 1'b1);
wire [15:0] wi2 = TIEmux(inverse, wi1, sin);
assign fft_twiddle = {
assign FFT_TWIDDLE = {
TIEmux(shift, wr1, {wr1[15], wr1[15:1]}),
TIEmux(shift, wi2, {wi2[15], wi2[15:1]})
};
}
function [63:0] fft_butterfly ([63:0] data, [15:0] wr, [15:0] wi, [0:0] shift) {
function [63:0] FFT_BUTTERFLY ([63:0] data, [15:0] wr, [15:0] wi, [0:0] shift) {
// operands real parts
wire [15:0] r1 = data[63:48];
@ -230,17 +230,17 @@ function [63:0] fft_butterfly ([63:0] data, [15:0] wr, [15:0] wi, [0:0] shift) {
wire [15:0] resi1 = TIEadd(eveni, oddi, 1'b0);
wire [15:0] resi2 = TIEadd(eveni, ~oddi, 1'b1);
assign fft_butterfly = { resr1, resr2, resi1, resi2 };
assign FFT_BUTTERFLY = { resr1, resr2, resi1, resi2 };
}
operation fft_calc {inout fft_reg data, in AR i, in BR shift, in BR inverse} {}
operation FFT_CALC {inout FFT_REG data, in AR i, in BR shift, in BR inverse} {}
{
wire [31:0] twiddle = fft_twiddle(i, shift, inverse);
wire [31:0] twiddle = FFT_TWIDDLE(i, shift, inverse);
wire [15:0] wr = twiddle[31:16];
wire [15:0] wi = twiddle[15:0];
assign data = fft_butterfly(data, wr, wi, shift);
assign data = FFT_BUTTERFLY(data, wr, wi, shift);
}
////////////////////////////////////////////////////////////////////////////
@ -304,29 +304,22 @@ operation fft_calc {inout fft_reg data, in AR i, in BR shift, in BR inverse} {}
//
//--------------------------------------------------------------------------
immediate_range st.FFT_reg_immed2 -32 24 8
immediate_range ld.FFT_reg_immed2 -32 24 8
immediate_range ST.FFT_REG_immed2 -32 24 8
immediate_range LD.FFT_REG_immed2 -32 24 8
//--------------------------------------------------------------------------
//
// FLIX Formats and Slots
//
// flix64_0, format width 64 bits, 2 slots
//
//--------------------------------------------------------------------------
format flix64_0 64 { flix64_0_slot0, flix64_0_slot1, flix64_0_slot2 }
slot_opcodes Inst {
FFT_bit_reverse, FFT_calc, ld.FFT_reg,
mv.FFT_reg, st.FFT_reg }
slot_opcodes flix64_0_slot0 {
ADD, BEQZ, BGEU, BLTU, BNEZ, L16SI, MOV.N, NOP, S16I, SEXT, SUB, FFT_shift_check }
slot_opcodes flix64_0_slot1 {
ADD, ADDX2, MOV.N, MOVNEZ, NEG, NOP }
slot_opcodes flix64_0_slot2 {
ADD, L16SI, NOP, S16I, SUB, FFT_shift_check }
//Full slots:
slot_opcodes flix64_0_slot0 { MOVI, J, ADDX2, L16SI, S16I, FFT_BIT_REVERSE, S32I, L32I, FFT_SHIFT_CHECK, OR, NOP, ADD, FFT_CALC }
slot_opcodes flix64_0_slot1 { SSL, SLL, MOVI, ADDX2, NOP, ADDI.N, ANDBC, ADD, MOV.N, J }
slot_opcodes flix64_0_slot2 { S32I, ADDI.N, L32I, L16SI, ADDX2, NOP, FFT_SHIFT_CHECK, J, MOVI, SSL, MOV.N, S16I }
//--------------------------------------------------------------------------
//--------------------------------------------------------------------------
//
@ -334,4 +327,4 @@ slot_opcodes flix64_0_slot2 {
//
//--------------------------------------------------------------------------
ctype fft_reg 64 64 fft_reg default
ctype FFT_REG 64 64 FFT_REG default