Merge remote-tracking branch 'github/simd'

This commit is contained in:
Jörg Thalheim 2015-03-10 15:23:22 +01:00
commit 608a750738
3 changed files with 135 additions and 93 deletions

View File

@ -126,6 +126,7 @@ int fix_fft_org(fixed fr[], fixed fi[], int m, int inverse)
wr >>= 1; wr >>= 1;
wi >>= 1; wi >>= 1;
} }
for(i=m; i<n; i+=istep) for(i=m; i<n; i+=istep)
{ {
@ -148,6 +149,7 @@ int fix_fft_org(fixed fr[], fixed fi[], int m, int inverse)
fi[i] = qi + ti; fi[i] = qi + ti;
} }
} }
--k; --k;
l = istep; l = istep;
} }

22
fft.c
View File

@ -37,6 +37,14 @@
#include <stdio.h> #include <stdio.h>
#define FFT_COMBINED_STORE(_fr, _fi, _i, _simd_r, _simd_i) \
asm ("{" "\n" \
" fft_simd_store %1, %0, %2" "\n" \
" nop" "\n" \
" fft_simd_store %3, %0, %4" "\n" \
"}" \
:: "r" (_i), "r" (_fr), "r" (_simd_r), "r" (_fi), "r" (_simd_i));
/* /*
* fix_fft() - perform fast Fourier transform. * fix_fft() - perform fast Fourier transform.
* *
@ -112,8 +120,21 @@ int fix_fft(fixed fr[], fixed fi[], int m, int inverse)
istep = l << 1; //step width of current butterfly istep = l << 1; //step width of current butterfly
FFT_REG reg; FFT_REG reg;
FFT_REG_SIMD simd_r, simd_i;
fixed *reg_s = ((fixed*) &reg); fixed *reg_s = ((fixed*) &reg);
if(l == 1)
{
for(i=0; i<n; i+=8)
{
simd_r = FFT_simd_load(fr, i);
simd_i = FFT_simd_load(fi, i);
FFT_SIMD_FIRST(simd_r, simd_i, (xtbool) shift);
FFT_COMBINED_STORE(fr, fi, i, simd_r, simd_i);
}
}
else
{
for(m=0; m<n; m+=istep) for(m=0; m<n; m+=istep)
{ {
for(i=m; i<m+l; ++i) for(i=m; i<m+l; ++i)
@ -133,6 +154,7 @@ int fix_fft(fixed fr[], fixed fi[], int m, int inverse)
fi[j] = reg_s[0]; fi[j] = reg_s[0];
} }
} }
}
--k; --k;
l = istep; l = istep;
} }

View File

@ -132,6 +132,62 @@ table SIN_WAVE 16 1024 {
regfile FFT_REG 64 2 fftv regfile FFT_REG 64 2 fftv
regfile FFT_REG_SIMD 128 2 fftsv
function [31:0] FFT_VAR_SHIFT([31:0] data, [3:0] sh)
{
assign FFT_VAR_SHIFT = TIEmux(sh,
data[31:0],
{data[30:0], 1'b0},
{data[29:0], 2'b0},
{data[28:0], 3'b0},
{data[27:0], 4'b0},
{data[26:0], 5'b0},
{data[25:0], 6'b0},
{data[24:0], 7'b0},
{data[23:0], 8'b0},
{data[22:0], 9'b0},
{data[21:0], 10'b0},
{data[20:0], 11'b0},
{data[19:0], 12'b0},
{data[18:0], 13'b0},
{data[17:0], 14'b0},
{data[16:0], 15'b0}
);
}
operation FFT_SIMD_LOAD {in AR *base, in AR offset, out FFT_REG_SIMD data} {out VAddr, in MemDataIn128}
{
assign VAddr = TIEadd(base, {offset[30:0], 1'b0}, 1'b0);
wire [15:0] o1 = MemDataIn128[15:0];
wire [15:0] o2 = MemDataIn128[31:16];
wire [15:0] o3 = MemDataIn128[47:32];
wire [15:0] o4 = MemDataIn128[63:48];
wire [15:0] o5 = MemDataIn128[79:64];
wire [15:0] o6 = MemDataIn128[95:80];
wire [15:0] o7 = MemDataIn128[111:96];
wire [15:0] o8 = MemDataIn128[127:112];
assign data = {o1, o2, o3, o4, o5, o6, o7, o8 };
}
operation FFT_SIMD_STORE {in AR *base, in AR offset, in FFT_REG_SIMD data} {out VAddr, out MemDataOut128}
{
assign VAddr = TIEadd(base, {offset[30:0], 1'b0}, 1'b0);
wire [15:0] o1 = data[15:0];
wire [15:0] o2 = data[31:16];
wire [15:0] o3 = data[47:32];
wire [15:0] o4 = data[63:48];
wire [15:0] o5 = data[79:64];
wire [15:0] o6 = data[95:80];
wire [15:0] o7 = data[111:96];
wire [15:0] o8 = data[127:112];
assign MemDataOut128 = {o1, o2, o3, o4, o5, o6, o7, o8 };
}
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); assign VAddr = TIEadd(base, offset[31:1], 1'b0);
@ -243,59 +299,21 @@ operation FFT_CALC {inout FFT_REG data, in AR i, in BR shift, in BR inverse} {}
assign data = FFT_BUTTERFLY(data, wr, wi, shift); assign data = FFT_BUTTERFLY(data, wr, wi, shift);
} }
//////////////////////////////////////////////////////////////////////////// // 4 butterflies at once
// operation FFT_SIMD_FIRST {inout FFT_REG_SIMD fr, inout FFT_REG_SIMD fi, in BR shift} {}
// Generated by XPRES v4.0.4 {
// Sat Mar 07 17:29:10 2015
// wire [15:0] wr = TIEmux(shift, 16'h7fff, 16'h3fff);
// Register Files wire [15:0] wi = 16'b0;
// AR (a): 16 entries, 32 bits per entry, ports 3r / 2w
// vec (v): 16 entries, 160 bits per entry, ports 3r / 1w wire [63:0] res1 = FFT_butterfly({fr[127:96], fi[127:96]}, wr, wi, shift);
// valign (u): 4 entries, 128 bits per entry, ports 2r / 1w wire [63:0] res2 = FFT_butterfly({fr[95:64], fi[95:64]}, wr, wi, shift);
// sel (s): 8 entries, 32 bits per entry, ports 2r / 1w wire [63:0] res3 = FFT_butterfly({fr[63:32], fi[63:32]}, wr, wi, shift);
// FFT_reg (fftv): 2 entries, 64 bits per entry, ports 1r / 1w wire [63:0] res4 = FFT_butterfly({fr[31:0], fi[31:0]}, wr, wi, shift);
//
// FLIX Formats assign fr = { res1[63:32], res2[63:32], res3[63:32], res4[63:32] };
// x24: size 24 bits, 1 slot assign fi = { res1[31:0], res2[31:0], res3[31:0], res4[31:0] };
// slot Inst: size 24 bits }
// opcodes { }
// x64: size 64 bits, 3 slots
// slot vsLDST: size 24 bits
// opcodes { }
// slot vsMAC: size 18 bits
// opcodes { }
// slot vsALU: size 18 bits
// opcodes { }
// flix64_0: size 64 bits, 2 slots
// slot flix64_0_slot0: size 24 bits
// opcodes { L16SI NOP ld.FFT_reg mv.FFT_reg st.FFT_reg }
// slot flix64_0_slot1: size 10 bits
// opcodes { MOV.N NOP }
//
// This TIE requires the following configuration settings:
//
// Required Endian: Little
// Required Instruction Width: 64 bits
// Minimum Data-Memory Width: 128 bits
// Required Load/Store Units: 1
// Requires Byte Enables: Yes
// Requires Booleans: No
// Pipeline Length: 5 stages
//
// This TIE was generated on a processor configuration with the
// following ISA instruction options enabled:
//
// CLAMPS
// MUL16
// NSA/NSAU
// MIN/MAX and MINU/MAXU
// Sign Extend to 32 Bits
// Enable Density Instructions
// Enable Boolean Registers
// Zero Overhead Loop Instructions
// Vectra LX DSP Coprocessor Instruction Family
//
////////////////////////////////////////////////////////////////////////////
//-------------------------------------------------------------------------- //--------------------------------------------------------------------------
@ -314,10 +332,9 @@ immediate_range LD.FFT_REG_immed2 -32 24 8
// flix64_0, format width 64 bits, 2 slots // flix64_0, format width 64 bits, 2 slots
format flix64_0 64 { flix64_0_slot0, flix64_0_slot1, flix64_0_slot2 } format flix64_0 64 { flix64_0_slot0, flix64_0_slot1, flix64_0_slot2 }
//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, FFT_SIMD_LOAD, FFT_SIMD_STORE }
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_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 } slot_opcodes flix64_0_slot2 { S32I, ADDI.N, L32I, L16SI, ADDX2, NOP, FFT_SHIFT_CHECK, J, MOVI, SSL, MOV.N, S16I, FFT_SIMD_LOAD, FFT_SIMD_STORE }
//-------------------------------------------------------------------------- //--------------------------------------------------------------------------
@ -328,3 +345,4 @@ slot_opcodes flix64_0_slot2 { S32I, ADDI.N, L32I, L16SI, ADDX2, NOP, FFT_SHIFT_C
//-------------------------------------------------------------------------- //--------------------------------------------------------------------------
ctype FFT_REG 64 64 FFT_REG default ctype FFT_REG 64 64 FFT_REG default
ctype FFT_REG_SIMD 128 128 FFT_REG_SIMD default