Merge pull request #23 from xianyi/develop

rebase
This commit is contained in:
Martin Kroeker 2021-04-19 22:24:12 +02:00 committed by GitHub
commit fc101b67e5
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
13 changed files with 1295 additions and 44 deletions

View File

@ -1,10 +1,21 @@
# COMPILER_PREFIX = mingw32- # COMPILER_PREFIX = mingw32-
ifdef HAVE_SSE ifndef DYNAMIC_ARCH
CCOMMON_OPT += -msse ADD_CPUFLAGS = 1
FCOMMON_OPT += -msse else
ifdef TARGET_CORE
ADD_CPUFLAGS = 1
endif
endif endif
ifdef ADD_CPUFLAGS
ifdef HAVE_SSE
CCOMMON_OPT += -msse
ifneq ($(F_COMPILER), NAG)
FCOMMON_OPT += -msse
endif
endif
endif
ifeq ($(OSNAME), Interix) ifeq ($(OSNAME), Interix)
ARFLAGS = -m x86 ARFLAGS = -m x86

View File

@ -8,6 +8,16 @@ endif
endif endif
endif endif
ifndef DYNAMIC_ARCH
ADD_CPUFLAGS = 1
else
ifdef TARGET_CORE
ADD_CPUFLAGS = 1
endif
endif
ifdef ADD_CPUFLAGS
ifdef HAVE_SSE3 ifdef HAVE_SSE3
CCOMMON_OPT += -msse3 CCOMMON_OPT += -msse3
ifneq ($(F_COMPILER), NAG) ifneq ($(F_COMPILER), NAG)
@ -44,7 +54,6 @@ endif
endif endif
ifeq ($(CORE), SKYLAKEX) ifeq ($(CORE), SKYLAKEX)
ifndef DYNAMIC_ARCH
ifndef NO_AVX512 ifndef NO_AVX512
CCOMMON_OPT += -march=skylake-avx512 CCOMMON_OPT += -march=skylake-avx512
ifneq ($(F_COMPILER), NAG) ifneq ($(F_COMPILER), NAG)
@ -62,10 +71,8 @@ endif
endif endif
endif endif
endif endif
endif
ifeq ($(CORE), COOPERLAKE) ifeq ($(CORE), COOPERLAKE)
ifndef DYNAMIC_ARCH
ifndef NO_AVX512 ifndef NO_AVX512
ifeq ($(C_COMPILER), GCC) ifeq ($(C_COMPILER), GCC)
# cooperlake support was added in 10.1 # cooperlake support was added in 10.1
@ -88,7 +95,6 @@ endif
endif endif
endif endif
endif endif
endif
ifdef HAVE_AVX2 ifdef HAVE_AVX2
ifndef NO_AVX2 ifndef NO_AVX2
@ -120,6 +126,7 @@ endif
endif endif
endif endif
endif
ifeq ($(OSNAME), Interix) ifeq ($(OSNAME), Interix)

View File

@ -299,6 +299,10 @@ if (NO_AVX2)
set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX2") set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX2")
endif () endif ()
if (NO_AVX512)
set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX512")
endif ()
if (USE_THREAD) if (USE_THREAD)
# USE_SIMPLE_THREADED_LEVEL3 = 1 # USE_SIMPLE_THREADED_LEVEL3 = 1
# NO_AFFINITY = 1 # NO_AFFINITY = 1

View File

@ -126,7 +126,7 @@ extern void openblas_warning(int verbose, const char * msg);
#endif #endif
#define get_cpu_ftr(id, var) ({ \ #define get_cpu_ftr(id, var) ({ \
__asm__ ("mrs %0, "#id : "=r" (var)); \ __asm__ __volatile__ ("mrs %0, "#id : "=r" (var)); \
}) })
static char *corename[] = { static char *corename[] = {

View File

@ -186,7 +186,7 @@ ZSWAPKERNEL = zswap.c
SGEMVNKERNEL = sgemv_n.c SGEMVNKERNEL = sgemv_n.c
DGEMVNKERNEL = dgemv_n_power10.c DGEMVNKERNEL = dgemv_n_power10.c
CGEMVNKERNEL = cgemv_n.c CGEMVNKERNEL = cgemv_n.c
ZGEMVNKERNEL = zgemv_n_4.c ZGEMVNKERNEL = zgemv_n_power10.c
# #
SGEMVTKERNEL = sgemv_t.c SGEMVTKERNEL = sgemv_t.c
DGEMVTKERNEL = dgemv_t_power10.c DGEMVTKERNEL = dgemv_t_power10.c

View File

@ -190,10 +190,9 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
__vector_quad acc0, acc1, acc2, acc3, acc4,acc5,acc6,acc7; __vector_quad acc0, acc1, acc2, acc3, acc4,acc5,acc6,acc7;
BLASLONG l = 0; BLASLONG l = 0;
vec_t *rowA = (vec_t *) & AO[0]; vec_t *rowA = (vec_t *) & AO[0];
vec_t *rb = (vec_t *) & BO[0];
__vector_pair rowB, rowB1; __vector_pair rowB, rowB1;
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]); rowB = *((__vector_pair *)((void *)&BO[0]));
__builtin_vsx_assemble_pair (&rowB1, rb[3], rb[2]); rowB1 = *((__vector_pair *)((void *)&BO[4]));
__builtin_mma_xvf64ger (&acc0, rowB, rowA[0]); __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64ger (&acc1, rowB1, rowA[0]); __builtin_mma_xvf64ger (&acc1, rowB1, rowA[0]);
__builtin_mma_xvf64ger (&acc2, rowB, rowA[1]); __builtin_mma_xvf64ger (&acc2, rowB, rowA[1]);
@ -205,9 +204,8 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
for (l = 1; l < temp; l++) for (l = 1; l < temp; l++)
{ {
rowA = (vec_t *) & AO[l << 3]; rowA = (vec_t *) & AO[l << 3];
rb = (vec_t *) & BO[l << 3]; rowB = *((__vector_pair *)((void *)&BO[l << 3]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]); rowB1 = *((__vector_pair *)((void *)&BO[(l << 3) + 4]));
__builtin_vsx_assemble_pair (&rowB1, rb[3], rb[2]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]); __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB1, rowA[0]); __builtin_mma_xvf64gerpp (&acc1, rowB1, rowA[0]);
__builtin_mma_xvf64gerpp (&acc2, rowB, rowA[1]); __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[1]);
@ -247,9 +245,8 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
BLASLONG l = 0; BLASLONG l = 0;
vec_t *rowA = (vec_t *) & AO[0]; vec_t *rowA = (vec_t *) & AO[0];
__vector_pair rowB, rowB1; __vector_pair rowB, rowB1;
vec_t *rb = (vec_t *) & BO[0]; rowB = *((__vector_pair *)((void *)&BO[0]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]); rowB1 = *((__vector_pair *)((void *)&BO[4]));
__builtin_vsx_assemble_pair (&rowB1, rb[3], rb[2]);
__builtin_mma_xvf64ger (&acc0, rowB, rowA[0]); __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64ger (&acc1, rowB1, rowA[0]); __builtin_mma_xvf64ger (&acc1, rowB1, rowA[0]);
__builtin_mma_xvf64ger (&acc2, rowB, rowA[1]); __builtin_mma_xvf64ger (&acc2, rowB, rowA[1]);
@ -257,9 +254,8 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
for (l = 1; l < temp; l++) for (l = 1; l < temp; l++)
{ {
rowA = (vec_t *) & AO[l << 2]; rowA = (vec_t *) & AO[l << 2];
rb = (vec_t *) & BO[l << 3]; rowB = *((__vector_pair *)((void *)&BO[l << 3]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]); rowB1 = *((__vector_pair *)((void *)&BO[(l << 3) + 4]));
__builtin_vsx_assemble_pair (&rowB1, rb[3], rb[2]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]); __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB1, rowA[0]); __builtin_mma_xvf64gerpp (&acc1, rowB1, rowA[0]);
__builtin_mma_xvf64gerpp (&acc2, rowB, rowA[1]); __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[1]);
@ -291,17 +287,15 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
BLASLONG l = 0; BLASLONG l = 0;
vec_t *rowA = (vec_t *) & AO[0]; vec_t *rowA = (vec_t *) & AO[0];
__vector_pair rowB, rowB1; __vector_pair rowB, rowB1;
vec_t *rb = (vec_t *) & BO[0]; rowB = *((__vector_pair *)((void *)&BO[0]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]); rowB1 = *((__vector_pair *)((void *)&BO[4]));
__builtin_vsx_assemble_pair (&rowB1, rb[3], rb[2]);
__builtin_mma_xvf64ger (&acc0, rowB, rowA[0]); __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64ger (&acc1, rowB1, rowA[0]); __builtin_mma_xvf64ger (&acc1, rowB1, rowA[0]);
for (l = 1; l < temp; l++) for (l = 1; l < temp; l++)
{ {
rowA = (vec_t *) & AO[l << 1]; rowA = (vec_t *) & AO[l << 1];
rb = (vec_t *) & BO[l << 3]; rowB = *((__vector_pair *)((void *)&BO[l << 3]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]); rowB1 = *((__vector_pair *)((void *)&BO[(l << 3) + 4]));
__builtin_vsx_assemble_pair (&rowB1, rb[3], rb[2]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]); __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB1, rowA[0]); __builtin_mma_xvf64gerpp (&acc1, rowB1, rowA[0]);
} }
@ -403,8 +397,7 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
BLASLONG l = 0; BLASLONG l = 0;
vec_t *rowA = (vec_t *) & AO[0]; vec_t *rowA = (vec_t *) & AO[0];
__vector_pair rowB; __vector_pair rowB;
vec_t *rb = (vec_t *) & BO[0]; rowB = *((__vector_pair *)((void *)&BO[0]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64ger (&acc0, rowB, rowA[0]); __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64ger (&acc1, rowB, rowA[1]); __builtin_mma_xvf64ger (&acc1, rowB, rowA[1]);
__builtin_mma_xvf64ger (&acc2, rowB, rowA[2]); __builtin_mma_xvf64ger (&acc2, rowB, rowA[2]);
@ -412,8 +405,7 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
for (l = 1; l < temp; l++) for (l = 1; l < temp; l++)
{ {
rowA = (vec_t *) & AO[l << 3]; rowA = (vec_t *) & AO[l << 3];
rb = (vec_t *) & BO[l << 2]; rowB = *((__vector_pair *)((void *)&BO[l << 2]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]); __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]); __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
__builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]); __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
@ -445,15 +437,13 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
BLASLONG l = 0; BLASLONG l = 0;
vec_t *rowA = (vec_t *) & AO[0]; vec_t *rowA = (vec_t *) & AO[0];
__vector_pair rowB; __vector_pair rowB;
vec_t *rb = (vec_t *) & BO[0]; rowB = *((__vector_pair *)((void *)&BO[0]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64ger (&acc0, rowB, rowA[0]); __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64ger (&acc1, rowB, rowA[1]); __builtin_mma_xvf64ger (&acc1, rowB, rowA[1]);
for (l = 1; l < temp; l++) for (l = 1; l < temp; l++)
{ {
rowA = (vec_t *) & AO[l << 2]; rowA = (vec_t *) & AO[l << 2];
rb = (vec_t *) & BO[l << 2]; rowB = *((__vector_pair *)((void *)&BO[l << 2]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]); __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]); __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
} }
@ -481,14 +471,12 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
BLASLONG l = 0; BLASLONG l = 0;
vec_t *rowA = (vec_t *) & AO[0]; vec_t *rowA = (vec_t *) & AO[0];
__vector_pair rowB; __vector_pair rowB;
vec_t *rb = (vec_t *) & BO[0]; rowB = *((__vector_pair *)((void *)&BO[0]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64ger (&acc0, rowB, rowA[0]); __builtin_mma_xvf64ger (&acc0, rowB, rowA[0]);
for (l = 1; l < temp; l++) for (l = 1; l < temp; l++)
{ {
rowA = (vec_t *) & AO[l << 1]; rowA = (vec_t *) & AO[l << 1];
rb = (vec_t *) & BO[l << 2]; rowB = *((__vector_pair *)((void *)&BO[l << 2]));
__builtin_vsx_assemble_pair (&rowB, rb[1], rb[0]);
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]); __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
} }
SAVE_ACC (&acc0, 0); SAVE_ACC (&acc0, 0);

File diff suppressed because it is too large Load Diff

View File

@ -43,6 +43,134 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#elif HAVE_KERNEL_4x4_VEC #elif HAVE_KERNEL_4x4_VEC
#if defined(POWER10) && (__BYTE_ORDER__ != __ORDER_BIG_ENDIAN__)
typedef __vector unsigned char vec_t;
typedef FLOAT v4sf_t __attribute__ ((vector_size (16)));
static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
BLASLONG i;
FLOAT *a0, *a1, *a2, *a3;
a0 = ap;
a1 = ap + lda;
a2 = a1 + lda;
a3 = a2 + lda;
__vector_quad acc0, acc1, acc2, acc3;;
__vector_quad acc4, acc5, acc6, acc7;
v4sf_t result[4];
__vector_pair *Va0, *Va1, *Va2, *Va3;
i = 0;
n = n << 1;
__builtin_mma_xxsetaccz (&acc0);
__builtin_mma_xxsetaccz (&acc1);
__builtin_mma_xxsetaccz (&acc2);
__builtin_mma_xxsetaccz (&acc3);
__builtin_mma_xxsetaccz (&acc4);
__builtin_mma_xxsetaccz (&acc5);
__builtin_mma_xxsetaccz (&acc6);
__builtin_mma_xxsetaccz (&acc7);
while (i < n) {
vec_t *rx = (vec_t *) & x[i];
Va0 = ((__vector_pair*)((void*)&a0[i]));
Va1 = ((__vector_pair*)((void*)&a1[i]));
Va2 = ((__vector_pair*)((void*)&a2[i]));
Va3 = ((__vector_pair*)((void*)&a3[i]));
__builtin_mma_xvf64gerpp (&acc0, Va0[0], rx[0]);
__builtin_mma_xvf64gerpp (&acc1, Va1[0], rx[0]);
__builtin_mma_xvf64gerpp (&acc2, Va2[0], rx[0]);
__builtin_mma_xvf64gerpp (&acc3, Va3[0], rx[0]);
__builtin_mma_xvf64gerpp (&acc4, Va0[0], rx[1]);
__builtin_mma_xvf64gerpp (&acc5, Va1[0], rx[1]);
__builtin_mma_xvf64gerpp (&acc6, Va2[0], rx[1]);
__builtin_mma_xvf64gerpp (&acc7, Va3[0], rx[1]);
__builtin_mma_xvf64gerpp (&acc0, Va0[1], rx[2]);
__builtin_mma_xvf64gerpp (&acc1, Va1[1], rx[2]);
__builtin_mma_xvf64gerpp (&acc2, Va2[1], rx[2]);
__builtin_mma_xvf64gerpp (&acc3, Va3[1], rx[2]);
__builtin_mma_xvf64gerpp (&acc4, Va0[1], rx[3]);
__builtin_mma_xvf64gerpp (&acc5, Va1[1], rx[3]);
__builtin_mma_xvf64gerpp (&acc6, Va2[1], rx[3]);
__builtin_mma_xvf64gerpp (&acc7, Va3[1], rx[3]);
i += 8;
}
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
__builtin_mma_disassemble_acc ((void *)result, &acc0);
register FLOAT temp_r0 = result[0][0] - result[1][1];
register FLOAT temp_i0 = result[0][1] + result[1][0];
__builtin_mma_disassemble_acc ((void *)result, &acc4);
temp_r0 += result[2][0] - result[3][1];
temp_i0 += result[2][1] + result[3][0];
__builtin_mma_disassemble_acc ((void *)result, &acc1);
register FLOAT temp_r1 = result[0][0] - result[1][1];
register FLOAT temp_i1 = result[0][1] + result[1][0];
__builtin_mma_disassemble_acc ((void *)result, &acc5);
temp_r1 += result[2][0] - result[3][1];
temp_i1 += result[2][1] + result[3][0];
__builtin_mma_disassemble_acc ((void *)result, &acc2);
register FLOAT temp_r2 = result[0][0] - result[1][1];
register FLOAT temp_i2 = result[0][1] + result[1][0];
__builtin_mma_disassemble_acc ((void *)result, &acc6);
temp_r2 += result[2][0] - result[3][1];
temp_i2 += result[2][1] + result[3][0];
__builtin_mma_disassemble_acc ((void *)result, &acc3);
register FLOAT temp_r3 = result[0][0] - result[1][1];
register FLOAT temp_i3 = result[0][1] + result[1][0];
__builtin_mma_disassemble_acc ((void *)result, &acc7);
temp_r3 += result[2][0] - result[3][1];
temp_i3 += result[2][1] + result[3][0];
#else
__builtin_mma_disassemble_acc ((void *)result, &acc0);
register FLOAT temp_r0 = result[0][0] + result[1][1];
register FLOAT temp_i0 = result[0][1] - result[1][0];
__builtin_mma_disassemble_acc ((void *)result, &acc4);
temp_r0 += result[2][0] + result[3][1];
temp_i0 += result[2][1] - result[3][0];
__builtin_mma_disassemble_acc ((void *)result, &acc1);
register FLOAT temp_r1 = result[0][0] + result[1][1];
register FLOAT temp_i1 = result[0][1] - result[1][0];
__builtin_mma_disassemble_acc ((void *)result, &acc5);
temp_r1 += result[2][0] + result[3][1];
temp_i1 += result[2][1] - result[3][0];
__builtin_mma_disassemble_acc ((void *)result, &acc2);
register FLOAT temp_r2 = result[0][0] + result[1][1];
register FLOAT temp_i2 = result[0][1] - result[1][0];
__builtin_mma_disassemble_acc ((void *)result, &acc6);
temp_r2 += result[2][0] + result[3][1];
temp_i2 += result[2][1] - result[3][0];
__builtin_mma_disassemble_acc ((void *)result, &acc3);
register FLOAT temp_r3 = result[0][0] + result[1][1];
register FLOAT temp_i3 = result[0][1] - result[1][0];
__builtin_mma_disassemble_acc ((void *)result, &acc7);
temp_r3 += result[2][0] + result[3][1];
temp_i3 += result[2][1] - result[3][0];
#endif
#if !defined(XCONJ)
y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
y[4] += alpha_r * temp_r2 - alpha_i * temp_i2;
y[5] += alpha_r * temp_i2 + alpha_i * temp_r2;
y[6] += alpha_r * temp_r3 - alpha_i * temp_i3;
y[7] += alpha_r * temp_i3 + alpha_i * temp_r3;
#else
y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
y[4] += alpha_r * temp_r2 + alpha_i * temp_i2;
y[5] -= alpha_r * temp_i2 - alpha_i * temp_r2;
y[6] += alpha_r * temp_r3 + alpha_i * temp_i3;
y[7] -= alpha_r * temp_i3 - alpha_i * temp_r3;
#endif
}
#else
static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) { static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
BLASLONG i; BLASLONG i;
FLOAT *a0, *a1, *a2, *a3; FLOAT *a0, *a1, *a2, *a3;
@ -198,6 +326,7 @@ static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
#endif #endif
} }
#endif
#else #else
static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) { static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {

View File

@ -501,7 +501,11 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, f
int32_t permil[16] = {0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3}; int32_t permil[16] = {0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3};
BLASLONG n_count = n; BLASLONG n_count = n;
float *a_pointer = A,*b_pointer = B,*c_pointer = C,*ctemp = C,*next_b = B; float *a_pointer = A,*b_pointer = B,*c_pointer = C,*ctemp = C,*next_b = B;
#if defined(__clang__)
for(;n_count>23;n_count-=24) COMPUTE(24)
#else
for(;n_count>23;n_count-=24) COMPUTE_n24 for(;n_count>23;n_count-=24) COMPUTE_n24
#endif
for(;n_count>19;n_count-=20) COMPUTE(20) for(;n_count>19;n_count-=20) COMPUTE(20)
for(;n_count>15;n_count-=16) COMPUTE(16) for(;n_count>15;n_count-=16) COMPUTE(16)
for(;n_count>11;n_count-=12) COMPUTE(12) for(;n_count>11;n_count-=12) COMPUTE(12)

View File

@ -319,14 +319,14 @@
REAL ABSB, ANORM, ASCALE, ATOL, BNORM, BSCALE, BTOL, REAL ABSB, ANORM, ASCALE, ATOL, BNORM, BSCALE, BTOL,
$ C, SAFMIN, TEMP, TEMP2, TEMPR, ULP $ C, SAFMIN, TEMP, TEMP2, TEMPR, ULP
COMPLEX ABI22, AD11, AD12, AD21, AD22, CTEMP, CTEMP2, COMPLEX ABI22, AD11, AD12, AD21, AD22, CTEMP, CTEMP2,
$ CTEMP3, ESHIFT, RTDISC, S, SHIFT, SIGNBC, T1, $ CTEMP3, ESHIFT, S, SHIFT, SIGNBC,
$ U12, X, ABI12, Y $ U12, X, ABI12, Y
* .. * ..
* .. External Functions .. * .. External Functions ..
COMPLEX CLADIV COMPLEX CLADIV
LOGICAL LSAME LOGICAL LSAME
REAL CLANHS, SLAMCH REAL CLANHS, SLAMCH
EXTERNAL CLADIV, LLSAME, CLANHS, SLAMCH EXTERNAL CLADIV, LSAME, CLANHS, SLAMCH
* .. * ..
* .. External Subroutines .. * .. External Subroutines ..
EXTERNAL CLARTG, CLASET, CROT, CSCAL, XERBLA EXTERNAL CLARTG, CLASET, CROT, CSCAL, XERBLA
@ -351,6 +351,7 @@
ILSCHR = .TRUE. ILSCHR = .TRUE.
ISCHUR = 2 ISCHUR = 2
ELSE ELSE
ILSCHR = .TRUE.
ISCHUR = 0 ISCHUR = 0
END IF END IF
* *
@ -364,6 +365,7 @@
ILQ = .TRUE. ILQ = .TRUE.
ICOMPQ = 3 ICOMPQ = 3
ELSE ELSE
ILQ = .TRUE.
ICOMPQ = 0 ICOMPQ = 0
END IF END IF
* *
@ -377,6 +379,7 @@
ILZ = .TRUE. ILZ = .TRUE.
ICOMPZ = 3 ICOMPZ = 3
ELSE ELSE
ILZ = .TRUE.
ICOMPZ = 0 ICOMPZ = 0
END IF END IF
* *

View File

@ -139,7 +139,7 @@
* ===================================================================== * =====================================================================
* *
* .. Parameters .. * .. Parameters ..
DOUBLE PRECISION ZERO, HALF, ONE DOUBLE PRECISION ZERO, HALF, ONE, TWO
PARAMETER ( ZERO = 0.0D+0, HALF = 0.5D+0, ONE = 1.0D+0, PARAMETER ( ZERO = 0.0D+0, HALF = 0.5D+0, ONE = 1.0D+0,
$ TWO = 2.0D0 ) $ TWO = 2.0D0 )
DOUBLE PRECISION MULTPL DOUBLE PRECISION MULTPL

View File

@ -139,7 +139,7 @@
* ===================================================================== * =====================================================================
* *
* .. Parameters .. * .. Parameters ..
REAL ZERO, HALF, ONE REAL ZERO, HALF, ONE, TWO
PARAMETER ( ZERO = 0.0E+0, HALF = 0.5E+0, ONE = 1.0E+0, PARAMETER ( ZERO = 0.0E+0, HALF = 0.5E+0, ONE = 1.0E+0,
$ TWO = 2.0E+0 ) $ TWO = 2.0E+0 )
REAL MULTPL REAL MULTPL

View File

@ -319,7 +319,7 @@
DOUBLE PRECISION ABSB, ANORM, ASCALE, ATOL, BNORM, BSCALE, BTOL, DOUBLE PRECISION ABSB, ANORM, ASCALE, ATOL, BNORM, BSCALE, BTOL,
$ C, SAFMIN, TEMP, TEMP2, TEMPR, ULP $ C, SAFMIN, TEMP, TEMP2, TEMPR, ULP
COMPLEX*16 ABI22, AD11, AD12, AD21, AD22, CTEMP, CTEMP2, COMPLEX*16 ABI22, AD11, AD12, AD21, AD22, CTEMP, CTEMP2,
$ CTEMP3, ESHIFT, RTDISC, S, SHIFT, SIGNBC, T1, $ CTEMP3, ESHIFT, S, SHIFT, SIGNBC,
$ U12, X, ABI12, Y $ U12, X, ABI12, Y
* .. * ..
* .. External Functions .. * .. External Functions ..
@ -352,6 +352,7 @@
ILSCHR = .TRUE. ILSCHR = .TRUE.
ISCHUR = 2 ISCHUR = 2
ELSE ELSE
ILSCHR = .TRUE.
ISCHUR = 0 ISCHUR = 0
END IF END IF
* *
@ -365,6 +366,7 @@
ILQ = .TRUE. ILQ = .TRUE.
ICOMPQ = 3 ICOMPQ = 3
ELSE ELSE
ILQ = .TRUE.
ICOMPQ = 0 ICOMPQ = 0
END IF END IF
* *
@ -378,6 +380,7 @@
ILZ = .TRUE. ILZ = .TRUE.
ICOMPZ = 3 ICOMPZ = 3
ELSE ELSE
ILZ = .TRUE.
ICOMPZ = 0 ICOMPZ = 0
END IF END IF
* *