163 lines
5.1 KiB
Diff
163 lines
5.1 KiB
Diff
commit 30500125ff629ee83b856b246912408c33662a4b
|
|
Author: Nicolas Morey <nmorey@suse.com>
|
|
Date: Fri Jul 4 08:41:25 2025 +0200
|
|
|
|
libpsm2: disable AVX
|
|
|
|
PSM2 is built with AVX2 enable by default and if manually disables,
|
|
falls back to AVX.
|
|
Disable both for compat purposes.
|
|
|
|
Signed-off-by: Nicolas Morey <nmorey@suse.com>
|
|
|
|
diff --git buildflags.mak buildflags.mak
|
|
index 206223dbd0a0..24221f1531a8 100644
|
|
--- buildflags.mak
|
|
+++ buildflags.mak
|
|
@@ -98,54 +98,7 @@ INCLUDES += -I${IFS_HFI_HEADER_PATH}
|
|
|
|
BASECFLAGS +=-Wall $(WERROR)
|
|
|
|
-#
|
|
-# test if compiler supports 32B(AVX2)/64B(AVX512F) move instruction.
|
|
-#
|
|
-ifeq (${CC},icc)
|
|
- ifeq ($(PSM_DISABLE_AVX2),)
|
|
- MAVX2=-xATOM_SSE4.2 -DPSM_AVX512
|
|
- else
|
|
- MAVX2=-march=core-avx-i
|
|
- endif
|
|
-else
|
|
- ifeq ($(PSM_DISABLE_AVX2),)
|
|
- MAVX2=-mavx2
|
|
- else
|
|
- MAVX2=-mavx
|
|
- endif
|
|
-endif
|
|
-
|
|
-ifneq (icc,${CC})
|
|
- ifeq ($(PSM_DISABLE_AVX2),)
|
|
- RET := $(shell echo "int main() {}" | ${CC} ${MAVX2} -E -dM -xc - 2>&1 | grep -q AVX2 ; echo $$?)
|
|
- else
|
|
- RET := $(shell echo "int main() {}" | ${CC} ${MAVX2} -E -dM -xc - 2>&1 | grep -q AVX ; echo $$?)
|
|
- anerr := $(warning ***NOTE TO USER**** Disabling AVX2 will harm performance)
|
|
- endif
|
|
-
|
|
- ifeq (0,${RET})
|
|
- BASECFLAGS += ${MAVX2}
|
|
- else
|
|
- anerr := $(error Compiler does not support ${MAVX2} )
|
|
- endif
|
|
-else
|
|
- BASECFLAGS += ${MAVX2}
|
|
-endif
|
|
-
|
|
-# This support is dynamic at runtime, so is OK to enable as long as compiler can generate
|
|
-# the code.
|
|
-ifneq (,${PSM_AVX512})
|
|
- ifneq (icc,${CC})
|
|
- RET := $(shell echo "int main() {}" | ${CC} -mavx512f -E -dM -xc - 2>&1 | grep -q AVX512 ; echo $$?)
|
|
- ifeq (0,${RET})
|
|
- BASECFLAGS += -mavx512f
|
|
- else
|
|
- anerr := $(error Compiler does not support AVX512 )
|
|
- endif
|
|
- BASECFLAGS += -DPSM_AVX512
|
|
- endif
|
|
-endif
|
|
-
|
|
+BASECFLAGS += -msse4.2
|
|
#
|
|
# feature test macros for drand48_r
|
|
#
|
|
diff --git opa/opa_dwordcpy-generic.c opa/opa_dwordcpy-generic.c
|
|
index dfb7755d5fdf..e1313bc4f25c 100644
|
|
--- opa/opa_dwordcpy-generic.c
|
|
+++ opa/opa_dwordcpy-generic.c
|
|
@@ -192,33 +192,6 @@ void hfi_pio_blockcpy_512(volatile uint64_t *dest, const uint64_t *src, uint32_t
|
|
}
|
|
#endif
|
|
|
|
-void hfi_pio_blockcpy_256(volatile uint64_t *dest, const uint64_t *src, uint32_t nblock)
|
|
-{
|
|
- volatile __m256i *dp = (volatile __m256i *) dest;
|
|
- const __m256i *sp = (const __m256i *) src;
|
|
-
|
|
- psmi_assert((dp != NULL) && (sp != NULL));
|
|
- psmi_assert((((uintptr_t) dp) & 0x3f) == 0x0);
|
|
-
|
|
- if ((((uintptr_t) sp) & 0x1f) == 0x0) {
|
|
- /* source and destination are both 32 byte aligned */
|
|
- do {
|
|
- __m256i tmp0 = _mm256_load_si256(sp);
|
|
- __m256i tmp1 = _mm256_load_si256(sp + 1);
|
|
- _mm256_store_si256((__m256i *)dp, tmp0);
|
|
- _mm256_store_si256((__m256i *)(dp + 1), tmp1);
|
|
- } while ((--nblock) && (dp = dp+2) && (sp = sp+2));
|
|
- } else {
|
|
- /* only destination is 32 byte aligned - use unaligned loads */
|
|
- do {
|
|
- __m256i tmp0 = _mm256_loadu_si256(sp);
|
|
- __m256i tmp1 = _mm256_loadu_si256(sp + 1);
|
|
- _mm256_store_si256((__m256i *)dp, tmp0);
|
|
- _mm256_store_si256((__m256i *)(dp + 1), tmp1);
|
|
- } while ((--nblock) && (dp = dp+2) && (sp = sp+2));
|
|
- }
|
|
-}
|
|
-
|
|
void hfi_pio_blockcpy_128(volatile uint64_t *dest, const uint64_t *src, uint32_t nblock)
|
|
{
|
|
volatile __m128i *dp = (volatile __m128i *) dest;
|
|
diff --git opa/opa_dwordcpy-x86_64.c opa/opa_dwordcpy-x86_64.c
|
|
index dfb7755d5fdf..e1313bc4f25c 100644
|
|
--- opa/opa_dwordcpy-x86_64.c
|
|
+++ opa/opa_dwordcpy-x86_64.c
|
|
@@ -192,33 +192,6 @@ void hfi_pio_blockcpy_512(volatile uint64_t *dest, const uint64_t *src, uint32_t
|
|
}
|
|
#endif
|
|
|
|
-void hfi_pio_blockcpy_256(volatile uint64_t *dest, const uint64_t *src, uint32_t nblock)
|
|
-{
|
|
- volatile __m256i *dp = (volatile __m256i *) dest;
|
|
- const __m256i *sp = (const __m256i *) src;
|
|
-
|
|
- psmi_assert((dp != NULL) && (sp != NULL));
|
|
- psmi_assert((((uintptr_t) dp) & 0x3f) == 0x0);
|
|
-
|
|
- if ((((uintptr_t) sp) & 0x1f) == 0x0) {
|
|
- /* source and destination are both 32 byte aligned */
|
|
- do {
|
|
- __m256i tmp0 = _mm256_load_si256(sp);
|
|
- __m256i tmp1 = _mm256_load_si256(sp + 1);
|
|
- _mm256_store_si256((__m256i *)dp, tmp0);
|
|
- _mm256_store_si256((__m256i *)(dp + 1), tmp1);
|
|
- } while ((--nblock) && (dp = dp+2) && (sp = sp+2));
|
|
- } else {
|
|
- /* only destination is 32 byte aligned - use unaligned loads */
|
|
- do {
|
|
- __m256i tmp0 = _mm256_loadu_si256(sp);
|
|
- __m256i tmp1 = _mm256_loadu_si256(sp + 1);
|
|
- _mm256_store_si256((__m256i *)dp, tmp0);
|
|
- _mm256_store_si256((__m256i *)(dp + 1), tmp1);
|
|
- } while ((--nblock) && (dp = dp+2) && (sp = sp+2));
|
|
- }
|
|
-}
|
|
-
|
|
void hfi_pio_blockcpy_128(volatile uint64_t *dest, const uint64_t *src, uint32_t nblock)
|
|
{
|
|
volatile __m128i *dp = (volatile __m128i *) dest;
|
|
diff --git psm_hal_gen1/psm_hal_gen1_spio.c psm_hal_gen1/psm_hal_gen1_spio.c
|
|
index 5444897a3e44..9b11a4ec133b 100644
|
|
--- psm_hal_gen1/psm_hal_gen1_spio.c
|
|
+++ psm_hal_gen1/psm_hal_gen1_spio.c
|
|
@@ -171,8 +171,7 @@ ips_spio_init(const struct psmi_context *context, struct ptl *ptl,
|
|
get_cpuid(0x7, 0, &id);
|
|
|
|
/* 32B copying supported */
|
|
- ctrl->spio_blockcpy_large = (id.ebx & (1<<AVX2_BIT)) ?
|
|
- hfi_pio_blockcpy_256 : ctrl->spio_blockcpy_med;
|
|
+ ctrl->spio_blockcpy_large = ctrl->spio_blockcpy_med;
|
|
|
|
#ifdef PSM_AVX512
|
|
/* 64B copying supported */
|