1. CPU GPU
Ultimate CGRA w/ high-speed compiler
CGRA for Energy-efficient Cryptography
Beyond-Neuromorphic Systems
Non-Deterministic Computing
1
ナレータ VOICEVOX:もち子(cv 明日葉よもぎ)
はらぺこエンジニアに贈るCGRAの世界2022
（13.HW/SW協調設計編）

2. 20220202
2
Templates for IMAX programming
exe(OP_X, &var|&AR[0-63][0-3], s1, e1, s2, e2, s3, e3, OP_Y, s4, OP_Z, s5)
ex4(OP_X, &var|&AR[0-63], s1, e1, s2, e2, s3, e3, OP_Y, s4, OP_Z, s5)
exe(OP_X, &var, INIT0?var:var, e1, s2, e2, s3, e3, OP_Y, s4, OP_Z, s5)
exe(OP_X, &var, var, e1, INIT0?s2:0, e2, s3, e3, OP_Y, s4, OP_Z, s5)
mex(OP MEX2, &s2, INIT0?s20:s2, INIT0?0:expr,
OP MEX1, &s1, INIT0?s10:s1, INIT0?0:expr, limit, BR[0-63][0-3][1], BR[0-63][0-3][0])
cex(OP_CEXE, &ex0-9, c3, c2, c1, c0, 16bit-pattern)
mop(OP_X, ex9-0, &src|&dst, base, offset, mask, top, len, block, force, ptop, plen)
mo4(OP_X, ex9-0, &src|&dst, base, offset, mask, top, len, block, force, ptop, plen)
DMA information

3. Original C C+IMAX-code Target
(A) Intel-PC Native Intel-emax6lib Intel-CC Intel-CC Algorithm
(B) Intel-PC Simulator ARM+emax6lib ARM-XCC ARM-XCC(cross compiler) Algorithm
(C) Intel-PC Simulator IMAX/PIO Conv-c2c + ARM-XCC IMAX-code
(D) Intel-PC Simulator IMAX/DMA Conv-c2c + ARM-XCC IMAX-code + testbench
(E) Verilog Simulator Vsim + Testbench Verilog
(F) FPGA+Chipscope Vivado + hw_server Real Hardware
(G) ARM-SoC ARM+emax6lib ARM-CC Conv-c2c + ARM-CC Algorithm
(H) ARM-SoC IMAX/PIO Conv-c2c + ARM-CC Hardware w/o DMA
(I) ARM-SoC IMAX/DMA Conv-c2c + ARM-CC Performance
Conv-c2c (IMAX-CC) runs on CentOS/FreeBSD/ARM-SoC
- IMAX-code is translated to IMAX-config + DMA sequence, and embedded in ARM binary.
Simulator (csim) runs on CentOS/FreeBSD
- Register transfer level simulator
- ARMv8, 64cores, 32threads/core, L1+L2cache/core, L2-directory
reorder-buffer, parameterized memory hierarchy
- 64 IMAX, AXI4-IF, test-bench generator
20220202
3
HW/SW codesign

4. /* SCREEN=WD*HT */
for (row=0; row<HT; row++) {
for (col=0; col<WD; col++) {
pix = in[row*WD+col];
r = t[ pix>>24 ];
g = t[256+((pix>>16)&255)];
b = t[512+((pix>> 8)&255)];
out[row*WD+col]=r<<24 | g<<16 | b<<8;
} }
20220202
4
簡単な tone_curveをC言語で書く
Load →
Store ←
Color map tables

5. /* SCREEN=WD*HT */
for (row=0; row<HT; row++) {
//EMAX5A begin tone_curve mapdist=0
for (LOOP0=WD, col=-4; LOOP0--;) {
col += 4;
pix = in[row*WD+col/4];
r = t[ pix>>24 ];
g = t[256+((pix>>16)&255)];
b = t[512+((pix>> 8)&255)];
out[row*WD+col/4]=r<<24 | g<<16 | b<<8;
}
//EMAX5A end
}
//EDMAX5A drain_dirty_lmm
20220202
5
IMAXのループ構造記述に合わせる
Load →
Store ←
Color map tables
/* SCREEN=WD*HT */
for (row=0; row<HT; row++) {
for (col=0; col<WD; col++) {
pix = in[row*WD+col];
r = t[ pix>>24 ];
g = t[256+((pix>>16)&255)];
b = t[512+((pix>> 8)&255)];
out[row*WD+col]=r<<24 | g<<16 | b<<8;
} }
Load →
Store ←
Color map tables

6. /* SCREEN=WD*HT */
for (row=0; row<HT; row++) {
//EMAX5A begin tone_curve mapdist=0
for (LOOP0=WD, col=-4; LOOP0--;) {
col += 4;
mop(OP_LDWR, &pix, in_row_WD, col, MSK_W0, in, WD); //pix = in[row*WD+col/4];
mop(OP_LDUBR, &r, t_r, pix, MSK_B3, t, 256*3/4); //r = t[ pix>>24 ];
mop(OP_LDUBR, &g, t_g, pix, MSK_B2, t, 256*3/4); //g = t[256+((pix>>16)&255)];
mop(OP_LDUBR, &b, t_b, pix, MSK_B1, t, 256*3/4); //b = t[512+((pix>> 8)&255)];
out[row*WD+col/4]=r<<24|g<<16|b<<8;
}
//EMAX5A end
}
//EDMAX5A drain_dirty_lmm
20220202
6
IMAXの高機能関数記述に書き換えながらデバッグする
Load →
Store ←
Color map tables

7. /* SCREEN=WD*HT */
for (row=0; row<HT; row++) {
//EMAX5A begin tone_curve mapdist=0
for (LOOP0=WD, col=-4; LOOP0--;) {
exe(OP_ADD, &col, col, EXP_H3210, 4, EXP_H3210, 0, EXP_H3210); //col += 4;
mop(OP_LDWR, &pix, in_row_WD, col, MSK_W0, in, WD); //pix = in[row*WD+col/4];
mop(OP_LDUBR, &r, t_r, pix, MSK_B3, t, 256*3/4); //r = t[ pix>>24 ];
mop(OP_LDUBR, &g, t_g, pix, MSK_B2, t, 256*3/4); //g = t[256+((pix>>16)&255)];
mop(OP_LDUBR, &b, t_b, pix, MSK_B1, t, 256*3/4); //b = t[512+((pix>> 8)&255)];
exe(OP_MMRG, &out, r, EXP_H3210, g, EXP_H3210, b, EXP_H3210);
mop(OP_STWR, &out, out_row_WD, col, MSK_W0); //out[row*WD+col/4]=r<<24|g<<16|b<<8;
}
//EMAX5A end
}
//EDMAX5A drain_dirty_lmm
20220202
7
全部書き換えたら逐次実行プログラムとしてデバッグ
Load →
Store ←
Color map tables

8. 20220202
8
データの配置と流れの観点から見直す
Load →
Store ←
Color map tables
/* SCREEN=WD*HT */
for (row=0; row<HT; row++) {
//EMAX5A begin tone_curve mapdist=0
for (LOOP0=WD, col=-4; LOOP0--;) {
exe(OP_ADD, &col, col, EXP_H3210, 4, EXP_H3210, 0, EXP_H3210); //col += 4;
mop(OP_LDWR, &pix, in_row_WD, col, MSK_W0, in, WD); //pix = in[row*WD+col/4];
mop(OP_LDUBR, &r, t_r, pix, MSK_B3, t, 256*3/4); //r = t[ pix>>24 ];
mop(OP_LDUBR, &g, t_g, pix, MSK_B2, t, 256*3/4); //g = t[256+((pix>>16)&255)];
mop(OP_LDUBR, &b, t_b, pix, MSK_B1, t, 256*3/4); //b = t[512+((pix>> 8)&255)];
exe(OP_MMRG, &out, r, EXP_H3210, g, EXP_H3210, b, EXP_H3210);
mop(OP_STWR, &out, out_row_WD, col, MSK_W0); //out[row*WD+col/4]=r<<24|g<<16|b<<8;
}
//EMAX5A end
}
//EDMAX5A drain_dirty_lmm

9. 20220202
9
コンパイル結果
Load →
Store ←
Color map tables

10. 20220202
10
アイマックスコンパイラとシミュレータのソースコードも公開