Skip to content

Commit

Permalink
Improve the tuned STREAM example program
Browse files Browse the repository at this point in the history
  • Loading branch information
@fwsGonzo
fwsGonzo committed Nov 24, 2023
1 parent 3fbd8ca commit 4260d73
Show file tree
Hide file tree
Showing 2 changed files with 77 additions and 64 deletions.
6 changes: 3 additions & 3 deletions binaries/STREAM/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,17 +11,17 @@ if (GCC_TRIPLE STREQUAL "riscv32-unknown-elf")
if (TUNED)
set(RISCV_ABI "-march=rv32gv_zba -mabi=ilp32d")
else()
set(RISCV_ABI "-march=rv32g -mabi=ilp32d")
set(RISCV_ABI "")
endif()
else()
if (TUNED)
set(RISCV_ABI "-march=rv64gv_zba_zbb_zbc_zbs -mabi=lp64d")
else()
set(RISCV_ABI "-march=rv64g -mabi=lp64d")
set(RISCV_ABI "")
endif()
endif()
set(WARNINGS "-Wall -Wextra")
set(COMMON "-O3 -ffast-math -fno-math-errno -fno-stack-protector")
set(COMMON "-O3 -g -ffast-math -fno-math-errno -fno-stack-protector")
set(CMAKE_C_FLAGS "${WARNINGS} ${RISCV_ABI} -std=gnu99 ${COMMON}")

if (DEBUGGING)
Expand Down
135 changes: 74 additions & 61 deletions binaries/STREAM/src/stream.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -547,24 +547,27 @@ void checkSTREAMresults ()
#endif
}

#define VLOAD(elem, vec) asm("vle32.v "#vec", %1" : : "r"(elem), "m"(elem))
#define VSTORE(elem, vec) asm("vse32.v "#vec", %1" : "=m"(elem) : "m"(elem))
#define VLOAD(elem, vec) asm("vle32.v "#vec", %1" : : "r"(&elem), "m"(elem))
#define VSTORE(elem, vec) asm("vse32.v "#vec", (%1)" : "=m"(elem) : "r"(&elem))
#define FLOAD(reg, scalar) asm("fmv.s.x "#reg", %0" : : "r"(scalar) : #reg)
#ifdef TUNED

void tuned_STREAM_Copy()
{
ssize_t j;
for (j=0; j<STREAM_ARRAY_SIZE; j += 32) {
VLOAD(a[j + 0], v1);
VLOAD(a[j + 8], v2);
VLOAD(a[j + 16], v3);
VLOAD(a[j + 24], v4);

VSTORE(c[j + 0], v1);
VSTORE(c[j + 8], v2);
VSTORE(c[j + 16], v3);
VSTORE(c[j + 24], v4);
float* ap = &a[j];
float* cp = &c[j];

VLOAD(ap[ 0], v1);
VLOAD(ap[ 8], v2);
VLOAD(ap[16], v3);
VLOAD(ap[24], v4);

VSTORE(cp[ 0], v1);
VSTORE(cp[ 8], v2);
VSTORE(cp[16], v3);
VSTORE(cp[24], v4);
}
}

Expand All @@ -575,69 +578,75 @@ void tuned_STREAM_Scale(STREAM_TYPE scalar)
for (j=0; j<STREAM_ARRAY_SIZE; j += 32) {
//b[j] = scalar*c[j];

VLOAD(c[j + 0], v1);
VLOAD(c[j + 8], v2);
VLOAD(c[j + 16], v3);
VLOAD(c[j + 24], v4);
float* cp = &c[j];
VLOAD(cp[ 0], v1);
VLOAD(cp[ 8], v2);
VLOAD(cp[16], v3);
VLOAD(cp[24], v4);

asm("vfmul.vf v1, v1, fa0");
asm("vfmul.vf v2, v2, fa0");
asm("vfmul.vf v3, v3, fa0");
asm("vfmul.vf v4, v4, fa0");

VSTORE(b[j + 0], v1);
VSTORE(b[j + 8], v2);
VSTORE(b[j + 16], v3);
VSTORE(b[j + 24], v4);
float* bp = &b[j];
VSTORE(bp[ 0], v1);
VSTORE(bp[ 8], v2);
VSTORE(bp[16], v3);
VSTORE(bp[24], v4);
}
}

void tuned_STREAM_Add()
{
ssize_t j;
for (j=0; j<STREAM_ARRAY_SIZE; j += 32) {
float* ap = &a[j];
float* bp = &b[j];

asm("vle32.v v1, %1"
:
: "r"(a[j + 0]), "m"(a[j + 0]));
: "r"(&ap[0]), "m"(ap[0]));
asm("vle32.v v2, %1"
:
: "r"(a[j + 8]), "m"(a[j + 8]));
: "r"(&ap[8]), "m"(ap[8]));
asm("vle32.v v3, %1"
:
: "r"(a[j + 16]), "m"(a[j + 16]));
: "r"(&ap[16]), "m"(ap[16]));
asm("vle32.v v4, %1"
:
: "r"(a[j + 24]), "m"(a[j + 24]));
: "r"(&ap[24]), "m"(ap[24]));
asm("vle32.v v5, %1"
:
: "r"(b[j + 0]), "m"(b[j + 0]));
: "r"(&bp[0]), "m"(bp[0]));
asm("vle32.v v6, %1"
:
: "r"(b[j + 8]), "m"(b[j + 8]));
: "r"(&bp[8]), "m"(bp[8]));
asm("vle32.v v7, %1"
:
: "r"(b[j + 16]), "m"(b[j + 16]));
: "r"(&bp[16]), "m"(bp[16]));
asm("vle32.v v8, %1"
:
: "r"(b[j + 24]), "m"(b[j + 24]));
: "r"(&bp[24]), "m"(bp[24]));

asm("vfadd.vv v1, v1, v5");
asm("vfadd.vv v2, v2, v6");
asm("vfadd.vv v3, v3, v7");
asm("vfadd.vv v4, v4, v8");

asm("vse32.v v1, %1"
: "=m"(c[j + 0])
: "m"(c[j + 0]));
asm("vse32.v v2, %1"
: "=m"(c[j + 8])
: "m"(c[j + 8]));
asm("vse32.v v3, %1"
: "=m"(c[j + 16])
: "m"(c[j + 16]));
asm("vse32.v v4, %1"
: "=m"(c[j + 24])
: "m"(c[j + 24]));
float* cp = &c[j];
asm("vse32.v v1, (%1)"
: "=m"(cp[0])
: "r"(&cp[0]));
asm("vse32.v v2, (%1)"
: "=m"(cp[8])
: "r"(&cp[8]));
asm("vse32.v v3, (%1)"
: "=m"(cp[16])
: "r"(&cp[16]));
asm("vse32.v v4, (%1)"
: "=m"(cp[24])
: "r"(&cp[24]));
}
}

Expand All @@ -646,31 +655,35 @@ void tuned_STREAM_Triad(STREAM_TYPE scalar)
FLOAD(fa0, scalar);
ssize_t j;
for (j=0; j<STREAM_ARRAY_SIZE; j += 32) {
VLOAD(b[j + 0], v1);
VLOAD(b[j + 8], v2);
VLOAD(b[j + 16], v3);
VLOAD(b[j + 24], v4);

VLOAD(c[j + 0], v5);
VLOAD(c[j + 8], v6);
VLOAD(c[j + 16], v7);
VLOAD(c[j + 24], v8);
float* bp = &b[j];
float* cp = &c[j];

// a[j] = b[j]+scalar*c[j];
asm("vfmul.vf v5, v5, fa0");
asm("vfmul.vf v6, v6, fa0");
asm("vfmul.vf v7, v7, fa0");
asm("vfmul.vf v8, v8, fa0");
VLOAD(cp[ 0], v5);
VLOAD(cp[ 8], v6);
VLOAD(cp[16], v7);
VLOAD(cp[24], v8);

asm("vfadd.vv v1, v1, v5");
asm("vfadd.vv v2, v2, v6");
asm("vfadd.vv v3, v3, v7");
asm("vfadd.vv v4, v4, v8");
VLOAD(bp[ 0], v1);
VLOAD(bp[ 8], v2);
VLOAD(bp[16], v3);
VLOAD(bp[24], v4);

VSTORE(a[j + 0], v1);
VSTORE(a[j + 8], v2);
VSTORE(a[j + 16], v3);
VSTORE(a[j + 24], v4);
// a[j] = b[j]+scalar*c[j];
asm("vfmul.vf v5, v5, fa0\n"
"vfmul.vf v6, v6, fa0\n"
"vfmul.vf v7, v7, fa0\n"
"vfmul.vf v8, v8, fa0\n"

"vfadd.vv v1, v1, v5\n"
"vfadd.vv v2, v2, v6\n"
"vfadd.vv v3, v3, v7\n"
"vfadd.vv v4, v4, v8\n");

float* ap = &a[j];
VSTORE(ap[ 0], v1);
VSTORE(ap[ 8], v2);
VSTORE(ap[16], v3);
VSTORE(ap[24], v4);
}
}

Expand Down

0 comments on commit 4260d73

Please sign in to comment.