Skip to content

Add little-endian PowerPC (ppc64le) support to VM JIT compiler #860

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
runlevel5 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Add little-endian PowerPC (ppc64le) support to VM JIT compiler #860

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ade8709
Add little-endian PowerPC (ppc64le) support to VM JIT compiler
runlevel5 Aug 14, 2026
4d16669
Use POWER8 GPR-FPR moves for VM float conversions
runlevel5 Aug 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 16 additions & 0 deletions cmake/compilers/gnu.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,3 +25,19 @@ add_compile_options(-fno-strict-aliasing)
# This is necessary to hide all symbols unless explicitly exported
# via the Q_EXPORT macro
add_compile_options(-fvisibility=hidden)

# ppc64le: target POWER8 as baseline (the ppc64le minimum), which enables
# VSX/Altivec. PPC64LE_CPU can raise the baseline (power9, power10,
# native) for machine-specific builds. Big-endian ppc64 is left at the
# compiler default so builds keep working on pre-POWER8 CPUs
# (970/G5, POWER5-7).
include(utils/arch)
if(ARCH STREQUAL "ppc64" AND CMAKE_C_BYTE_ORDER STREQUAL "LITTLE_ENDIAN")
set(PPC64LE_CPU "power8" CACHE STRING
"-mcpu baseline for ppc64le builds (power8, power9, power10, native)")
add_compile_options(-mcpu=${PPC64LE_CPU})
if(PPC64LE_CPU STREQUAL "power8")
# keep POWER8 compatibility but schedule for POWER9 cores
add_compile_options(-mtune=power9)
endif()
endif()
122 changes: 101 additions & 21 deletions code/client/snd_altivec.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,12 @@ Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA

#if !defined(__APPLE__)
#include <altivec.h>
#undef bool
#undef pixel
#else
/* Apple's -faltivec provides the AltiVec keywords without altivec.h,
but only the unprefixed spellings */
#define __vector vector
#endif

void S_PaintChannelFrom16_altivec( portable_samplepair_t paintbuffer[PAINTBUFFER_SIZE], int snd_vol, channel_t *ch, const sfx_t *sc, int count, int sampleOffset, int bufferOffset ) {
Expand Down Expand Up @@ -72,12 +78,25 @@ void S_PaintChannelFrom16_altivec( portable_samplepair_t paintbuffer[PAINTBUFFER
}

if (!ch->doppler || ch->dopplerScale==1.0f) {
vector signed short volume_vec;
vector unsigned int volume_shift;
__vector signed short volume_vec;
__vector unsigned int volume_shift;
int vectorCount, samplesLeft, chunkSamplesLeft;
leftvol = ch->leftvol*snd_vol;
rightvol = ch->rightvol*snd_vol;
samples = chunk->sndChunk;
#if defined(__VSX__)
/* Build volume vector: {L, L, R, R, L, L, R, R}
Use vec_perm to interleave left and right volume splats
in an endian-safe way (BE-style indices). */
{
__vector signed short vl = vec_splats((short)leftvol);
__vector signed short vr = vec_splats((short)rightvol);
__vector unsigned char volume_perm = VECCONST_UINT8(
0, 1, 2, 3, 16, 17, 18, 19,
4, 5, 6, 7, 20, 21, 22, 23);
volume_vec = (__vector signed short)vec_perm(vl, vr, volume_perm);
}
#else
((short *)&volume_vec)[0] = leftvol;
((short *)&volume_vec)[1] = leftvol;
((short *)&volume_vec)[4] = leftvol;
Expand All @@ -86,6 +105,7 @@ void S_PaintChannelFrom16_altivec( portable_samplepair_t paintbuffer[PAINTBUFFER
((short *)&volume_vec)[3] = rightvol;
((short *)&volume_vec)[6] = rightvol;
((short *)&volume_vec)[7] = rightvol;
#endif
volume_shift = vec_splat_u32(8);
i = 0;

Expand Down Expand Up @@ -119,15 +139,70 @@ void S_PaintChannelFrom16_altivec( portable_samplepair_t paintbuffer[PAINTBUFFER

if(vectorCount)
{
vector unsigned char tmp;
vector short s0, s1, sampleData0, sampleData1;
vector signed int merge0, merge1;
vector signed int d0, d1, d2, d3;
vector unsigned char samplePermute0 =
#if defined(__VSX__)
__vector short s0, sampleData0, sampleData1;
__vector signed int merge0, merge1;
__vector signed int d0, d1, d2, d3;
__vector unsigned char samplePermute0 =
VECCONST_UINT8(0, 1, 4, 5, 0, 1, 4, 5, 2, 3, 6, 7, 2, 3, 6, 7);
vector unsigned char samplePermute1 =
__vector unsigned char samplePermute1 =
VECCONST_UINT8(8, 9, 12, 13, 8, 9, 12, 13, 10, 11, 14, 15, 10, 11, 14, 15);
vector unsigned char loadPermute0, loadPermute1;

while(vectorCount)
{
/* Load up source (16-bit) sample data */
s0 = vec_xl(0, &samples[sampleOffset]);

/* Load up destination sample data */
d0 = vec_xl(0, (int *)&samp[i]);
d1 = vec_xl(0, (int *)&samp[i+2]);
d2 = vec_xl(0, (int *)&samp[i+4]);
d3 = vec_xl(0, (int *)&samp[i+6]);

/* Rearrange samples: duplicate each sample for L/R processing.
samplePermute0 selects first 4 samples (each duplicated twice),
samplePermute1 selects second 4 samples (each duplicated twice). */
sampleData0 = vec_perm(s0, s0, samplePermute0);
sampleData1 = vec_perm(s0, s0, samplePermute1);

merge0 = vec_mule(sampleData0, volume_vec);
merge0 = vec_sra(merge0, volume_shift);

merge1 = vec_mulo(sampleData0, volume_vec);
merge1 = vec_sra(merge1, volume_shift);

d0 = vec_add(merge0, d0);
d1 = vec_add(merge1, d1);

merge0 = vec_mule(sampleData1, volume_vec);
merge0 = vec_sra(merge0, volume_shift);

merge1 = vec_mulo(sampleData1, volume_vec);
merge1 = vec_sra(merge1, volume_shift);

d2 = vec_add(merge0, d2);
d3 = vec_add(merge1, d3);

/* Store destination sample data */
vec_xst(d0, 0, (int *)&samp[i]);
vec_xst(d1, 0, (int *)&samp[i+2]);
vec_xst(d2, 0, (int *)&samp[i+4]);
vec_xst(d3, 0, (int *)&samp[i+6]);

i += 8;
vectorCount--;
sampleOffset += 8;
}
#else
__vector unsigned char tmp;
__vector short s0, s1, sampleData0, sampleData1;
__vector signed int merge0, merge1;
__vector signed int d0, d1, d2, d3;
__vector unsigned char samplePermute0 =
VECCONST_UINT8(0, 1, 4, 5, 0, 1, 4, 5, 2, 3, 6, 7, 2, 3, 6, 7);
__vector unsigned char samplePermute1 =
VECCONST_UINT8(8, 9, 12, 13, 8, 9, 12, 13, 10, 11, 14, 15, 10, 11, 14, 15);
__vector unsigned char loadPermute0, loadPermute1;

// Rather than permute the vectors after we load them to do the sample
// replication and rearrangement, we permute the alignment vector so
Expand All @@ -136,17 +211,22 @@ void S_PaintChannelFrom16_altivec( portable_samplepair_t paintbuffer[PAINTBUFFER
loadPermute0 = vec_perm(tmp,tmp,samplePermute0);
loadPermute1 = vec_perm(tmp,tmp,samplePermute1);

s0 = *(vector short *)&samples[sampleOffset];
s0 = *(__vector short *)&samples[sampleOffset];
while(vectorCount)
{
/* Load up source (16-bit) sample data */
s1 = *(vector short *)&samples[sampleOffset+7];
/* Load up source (16-bit) sample data. The offset
must reach the next 16-byte block for every source
alignment: +8 shorts (+16 bytes) always does, while
the old +7 stayed in the current block when
sampleOffset was 16-byte aligned, going stale after
the s0 = s1 carry below */
s1 = *(__vector short *)&samples[sampleOffset+8];

/* Load up destination sample data */
d0 = *(vector signed int *)&samp[i];
d1 = *(vector signed int *)&samp[i+2];
d2 = *(vector signed int *)&samp[i+4];
d3 = *(vector signed int *)&samp[i+6];
d0 = *(__vector signed int *)&samp[i];
d1 = *(__vector signed int *)&samp[i+2];
d2 = *(__vector signed int *)&samp[i+4];
d3 = *(__vector signed int *)&samp[i+6];

sampleData0 = vec_perm(s0,s1,loadPermute0);
sampleData1 = vec_perm(s0,s1,loadPermute1);
Expand All @@ -170,16 +250,17 @@ void S_PaintChannelFrom16_altivec( portable_samplepair_t paintbuffer[PAINTBUFFER
d3 = vec_add(merge1,d3);

/* Store destination sample data */
*(vector signed int *)&samp[i] = d0;
*(vector signed int *)&samp[i+2] = d1;
*(vector signed int *)&samp[i+4] = d2;
*(vector signed int *)&samp[i+6] = d3;
*(__vector signed int *)&samp[i] = d0;
*(__vector signed int *)&samp[i+2] = d1;
*(__vector signed int *)&samp[i+4] = d2;
*(__vector signed int *)&samp[i+6] = d3;

i += 8;
vectorCount--;
s0 = s1;
sampleOffset += 8;
}
#endif
if (sampleOffset == SND_CHUNK_SIZE) {
chunk = chunk->next;
samples = chunk->sndChunk;
Expand Down Expand Up @@ -226,4 +307,3 @@ void S_PaintChannelFrom16_altivec( portable_samplepair_t paintbuffer[PAINTBUFFER


#endif

5 changes: 3 additions & 2 deletions code/qcommon/q_platform.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,9 +42,10 @@ Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
#endif

#if (defined(powerc) || defined(powerpc) || defined(ppc) || \
defined(__ppc) || defined(__ppc__)) && !defined(C_ONLY)
defined(__ppc) || defined(__ppc__) || defined(__ppc64__) || \
defined(__powerpc64__)) && !defined(C_ONLY)
#define idppc 1
#if defined(__VEC__)
#if defined(__VEC__) || defined(__ALTIVEC__)
#define idppc_altivec 1
#ifdef __APPLE__ // Apple's GCC does this differently than the FSF.
#define VECCONST_UINT8(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p) \
Expand Down
Loading
Loading