ShuriZma
/
xemu
mirror of https://github.com/xemu-project/xemu.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

target/mips: Unroll loops for MSA float max/min instructions 

Slight preformance improvement for MSA float max/min instructions.

Signed-off-by: Aleksandar Markovic <amarkovic@wavecomp.com>
Reviewed-by: Aleksandar Rikalo <arikalo@wavecomp.com>
Message-Id: <1562068213-11307-7-git-send-email-aleksandar.markovic@rt-rk.com>
This commit is contained in:
Aleksandar Markovic 2019-07-02 13:50:12 +02:00
parent 44da090ba0
commit 807e6773a5
1 changed files with 125 additions and 73 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

198
target/mips/msa_helper.c
View File

@ -3850,35 +3850,65 @@ void helper_msa_fmin_df(CPUMIPSState *env, uint32_t df, uint32_t wd,
wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
wr_t *pws = &(env->active_fpu.fpr[ws].wr);
wr_t *pwt = &(env->active_fpu.fpr[wt].wr);
uint32_t i;
clear_msacsr_cause(env);
switch (df) {
case DF_WORD:
for (i = 0; i < DF_ELEMENTS(DF_WORD); i++) {
if (NUMBER_QNAN_PAIR(pws->w[i], pwt->w[i], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[i], min, pws->w[i], pws->w[i], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[i], pws->w[i], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[i], min, pwt->w[i], pwt->w[i], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[i], min, pws->w[i], pwt->w[i], 32);
}
if (df == DF_WORD) {
if (NUMBER_QNAN_PAIR(pws->w[0], pwt->w[0], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[0], min, pws->w[0], pws->w[0], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[0], pws->w[0], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[0], min, pwt->w[0], pwt->w[0], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[0], min, pws->w[0], pwt->w[0], 32);
}
break;
case DF_DOUBLE:
for (i = 0; i < DF_ELEMENTS(DF_DOUBLE); i++) {
if (NUMBER_QNAN_PAIR(pws->d[i], pwt->d[i], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[i], min, pws->d[i], pws->d[i], 64);
} else if (NUMBER_QNAN_PAIR(pwt->d[i], pws->d[i], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[i], min, pwt->d[i], pwt->d[i], 64);
} else {
MSA_FLOAT_MAXOP(pwx->d[i], min, pws->d[i], pwt->d[i], 64);
}
if (NUMBER_QNAN_PAIR(pws->w[1], pwt->w[1], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[1], min, pws->w[1], pws->w[1], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[1], pws->w[1], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[1], min, pwt->w[1], pwt->w[1], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[1], min, pws->w[1], pwt->w[1], 32);
}
break;
default:
if (NUMBER_QNAN_PAIR(pws->w[2], pwt->w[2], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[2], min, pws->w[2], pws->w[2], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[2], pws->w[2], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[2], min, pwt->w[2], pwt->w[2], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[2], min, pws->w[2], pwt->w[2], 32);
}
if (NUMBER_QNAN_PAIR(pws->w[3], pwt->w[3], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[3], min, pws->w[3], pws->w[3], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[3], pws->w[3], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[3], min, pwt->w[3], pwt->w[3], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[3], min, pws->w[3], pwt->w[3], 32);
}
} else if (df == DF_DOUBLE) {
if (NUMBER_QNAN_PAIR(pws->d[0], pwt->d[0], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[0], min, pws->d[0], pws->d[0], 64);
} else if (NUMBER_QNAN_PAIR(pwt->d[0], pws->d[0], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[0], min, pwt->d[0], pwt->d[0], 64);
} else {
MSA_FLOAT_MAXOP(pwx->d[0], min, pws->d[0], pwt->d[0], 64);
}
if (NUMBER_QNAN_PAIR(pws->d[1], pwt->d[1], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[1], min, pws->d[1], pws->d[1], 64);
} else if (NUMBER_QNAN_PAIR(pwt->d[1], pws->d[1], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[1], min, pwt->d[1], pwt->d[1], 64);
} else {
MSA_FLOAT_MAXOP(pwx->d[1], min, pws->d[1], pwt->d[1], 64);
}
} else {
assert(0);
}
check_msacsr_cause(env, GETPC());
@ -3894,22 +3924,18 @@ void helper_msa_fmin_a_df(CPUMIPSState *env, uint32_t df, uint32_t wd,
wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
wr_t *pws = &(env->active_fpu.fpr[ws].wr);
wr_t *pwt = &(env->active_fpu.fpr[wt].wr);
uint32_t i;
clear_msacsr_cause(env);
switch (df) {
case DF_WORD:
for (i = 0; i < DF_ELEMENTS(DF_WORD); i++) {
FMAXMIN_A(min, max, pwx->w[i], pws->w[i], pwt->w[i], 32, status);
}
break;
case DF_DOUBLE:
for (i = 0; i < DF_ELEMENTS(DF_DOUBLE); i++) {
FMAXMIN_A(min, max, pwx->d[i], pws->d[i], pwt->d[i], 64, status);
}
break;
default:
if (df == DF_WORD) {
FMAXMIN_A(min, max, pwx->w[0], pws->w[0], pwt->w[0], 32, status);
FMAXMIN_A(min, max, pwx->w[1], pws->w[1], pwt->w[1], 32, status);
FMAXMIN_A(min, max, pwx->w[2], pws->w[2], pwt->w[2], 32, status);
FMAXMIN_A(min, max, pwx->w[3], pws->w[3], pwt->w[3], 32, status);
} else if (df == DF_DOUBLE) {
FMAXMIN_A(min, max, pwx->d[0], pws->d[0], pwt->d[0], 64, status);
FMAXMIN_A(min, max, pwx->d[1], pws->d[1], pwt->d[1], 64, status);
} else {
assert(0);
}
@ -3921,40 +3947,70 @@ void helper_msa_fmin_a_df(CPUMIPSState *env, uint32_t df, uint32_t wd,
void helper_msa_fmax_df(CPUMIPSState *env, uint32_t df, uint32_t wd,
uint32_t ws, uint32_t wt)
{
float_status *status = &env->active_tc.msa_fp_status;
float_status *status = &env->active_tc.msa_fp_status;
wr_t wx, *pwx = &wx;
wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
wr_t *pws = &(env->active_fpu.fpr[ws].wr);
wr_t *pwt = &(env->active_fpu.fpr[wt].wr);
uint32_t i;
clear_msacsr_cause(env);
switch (df) {
case DF_WORD:
for (i = 0; i < DF_ELEMENTS(DF_WORD); i++) {
if (NUMBER_QNAN_PAIR(pws->w[i], pwt->w[i], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[i], max, pws->w[i], pws->w[i], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[i], pws->w[i], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[i], max, pwt->w[i], pwt->w[i], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[i], max, pws->w[i], pwt->w[i], 32);
}
if (df == DF_WORD) {
if (NUMBER_QNAN_PAIR(pws->w[0], pwt->w[0], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[0], max, pws->w[0], pws->w[0], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[0], pws->w[0], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[0], max, pwt->w[0], pwt->w[0], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[0], max, pws->w[0], pwt->w[0], 32);
}
break;
case DF_DOUBLE:
for (i = 0; i < DF_ELEMENTS(DF_DOUBLE); i++) {
if (NUMBER_QNAN_PAIR(pws->d[i], pwt->d[i], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[i], max, pws->d[i], pws->d[i], 64);
} else if (NUMBER_QNAN_PAIR(pwt->d[i], pws->d[i], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[i], max, pwt->d[i], pwt->d[i], 64);
} else {
MSA_FLOAT_MAXOP(pwx->d[i], max, pws->d[i], pwt->d[i], 64);
}
if (NUMBER_QNAN_PAIR(pws->w[1], pwt->w[1], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[1], max, pws->w[1], pws->w[1], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[1], pws->w[1], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[1], max, pwt->w[1], pwt->w[1], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[1], max, pws->w[1], pwt->w[1], 32);
}
break;
default:
if (NUMBER_QNAN_PAIR(pws->w[2], pwt->w[2], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[2], max, pws->w[2], pws->w[2], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[2], pws->w[2], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[2], max, pwt->w[2], pwt->w[2], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[2], max, pws->w[2], pwt->w[2], 32);
}
if (NUMBER_QNAN_PAIR(pws->w[3], pwt->w[3], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[3], max, pws->w[3], pws->w[3], 32);
} else if (NUMBER_QNAN_PAIR(pwt->w[3], pws->w[3], 32, status)) {
MSA_FLOAT_MAXOP(pwx->w[3], max, pwt->w[3], pwt->w[3], 32);
} else {
MSA_FLOAT_MAXOP(pwx->w[3], max, pws->w[3], pwt->w[3], 32);
}
} else if (df == DF_DOUBLE) {
if (NUMBER_QNAN_PAIR(pws->d[0], pwt->d[0], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[0], max, pws->d[0], pws->d[0], 64);
} else if (NUMBER_QNAN_PAIR(pwt->d[0], pws->d[0], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[0], max, pwt->d[0], pwt->d[0], 64);
} else {
MSA_FLOAT_MAXOP(pwx->d[0], max, pws->d[0], pwt->d[0], 64);
}
if (NUMBER_QNAN_PAIR(pws->d[1], pwt->d[1], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[1], max, pws->d[1], pws->d[1], 64);
} else if (NUMBER_QNAN_PAIR(pwt->d[1], pws->d[1], 64, status)) {
MSA_FLOAT_MAXOP(pwx->d[1], max, pwt->d[1], pwt->d[1], 64);
} else {
MSA_FLOAT_MAXOP(pwx->d[1], max, pws->d[1], pwt->d[1], 64);
}
} else {
assert(0);
}
check_msacsr_cause(env, GETPC());
@ -3970,22 +4026,18 @@ void helper_msa_fmax_a_df(CPUMIPSState *env, uint32_t df, uint32_t wd,
wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
wr_t *pws = &(env->active_fpu.fpr[ws].wr);
wr_t *pwt = &(env->active_fpu.fpr[wt].wr);
uint32_t i;
clear_msacsr_cause(env);
switch (df) {
case DF_WORD:
for (i = 0; i < DF_ELEMENTS(DF_WORD); i++) {
FMAXMIN_A(max, min, pwx->w[i], pws->w[i], pwt->w[i], 32, status);
}
break;
case DF_DOUBLE:
for (i = 0; i < DF_ELEMENTS(DF_DOUBLE); i++) {
FMAXMIN_A(max, min, pwx->d[i], pws->d[i], pwt->d[i], 64, status);
}
break;
default:
if (df == DF_WORD) {
FMAXMIN_A(max, min, pwx->w[0], pws->w[0], pwt->w[0], 32, status);
FMAXMIN_A(max, min, pwx->w[1], pws->w[1], pwt->w[1], 32, status);
FMAXMIN_A(max, min, pwx->w[2], pws->w[2], pwt->w[2], 32, status);
FMAXMIN_A(max, min, pwx->w[3], pws->w[3], pwt->w[3], 32, status);
} else if (df == DF_DOUBLE) {
FMAXMIN_A(max, min, pwx->d[0], pws->d[0], pwt->d[0], 64, status);
FMAXMIN_A(max, min, pwx->d[1], pws->d[1], pwt->d[1], 64, status);
} else {
assert(0);
}