* [PATCH v2 0/2] util/bufferiszero: Split out hosts, add loongarch64
@ 2024-06-07 0:24 Richard Henderson
2024-06-07 0:24 ` [PATCH v2 1/2] util/bufferiszero: Split out host include files Richard Henderson
2024-06-07 0:24 ` [PATCH v2 2/2] util/bufferiszero: Add loongarch64 vector acceleration Richard Henderson
0 siblings, 2 replies; 4+ messages in thread
From: Richard Henderson @ 2024-06-07 0:24 UTC (permalink / raw)
To: qemu-devel; +Cc: maobibo
Based-on: 20240527211912.14060-1-richard.henderson@linaro.org
("[PATCH 00/18] tcg/loongarch64: Support v64 and v256")
For "util/loongarch64: Detect LASX vector support"
For v2:
* Rename to bufferiszero.c.inc (philmd).
* Add inline assembly for loongarch64.
On cfarm400.cfarm.net (Loongson-3C5000L-LL @ 2.0GHz):
# Start of bufferiszero tests
# buffer_is_zero #0: 1KB 11021 MB/sec
# buffer_is_zero #0: 4KB 32107 MB/sec
# buffer_is_zero #0: 16KB 59118 MB/sec
# buffer_is_zero #0: 64KB 67940 MB/sec
#
# buffer_is_zero #1: 1KB 9540 MB/sec
# buffer_is_zero #1: 4KB 24050 MB/sec
# buffer_is_zero #1: 16KB 38082 MB/sec
# buffer_is_zero #1: 64KB 36399 MB/sec
#
# buffer_is_zero #2: 1KB 8026 MB/sec
# buffer_is_zero #2: 4KB 15493 MB/sec
# buffer_is_zero #2: 16KB 20865 MB/sec
# buffer_is_zero #2: 64KB 19694 MB/sec
r~
Richard Henderson (2):
util/bufferiszero: Split out host include files
util/bufferiszero: Add loongarch64 vector acceleration
util/bufferiszero.c | 191 +-----------------
host/include/aarch64/host/bufferiszero.c.inc | 76 +++++++
host/include/generic/host/bufferiszero.c.inc | 10 +
host/include/i386/host/bufferiszero.c.inc | 124 ++++++++++++
.../loongarch64/host/bufferiszero.c.inc | 143 +++++++++++++
host/include/x86_64/host/bufferiszero.c.inc | 1 +
6 files changed, 355 insertions(+), 190 deletions(-)
create mode 100644 host/include/aarch64/host/bufferiszero.c.inc
create mode 100644 host/include/generic/host/bufferiszero.c.inc
create mode 100644 host/include/i386/host/bufferiszero.c.inc
create mode 100644 host/include/loongarch64/host/bufferiszero.c.inc
create mode 100644 host/include/x86_64/host/bufferiszero.c.inc
--
2.34.1
^ permalink raw reply [flat|nested] 4+ messages in thread
* [PATCH v2 1/2] util/bufferiszero: Split out host include files
2024-06-07 0:24 [PATCH v2 0/2] util/bufferiszero: Split out hosts, add loongarch64 Richard Henderson
@ 2024-06-07 0:24 ` Richard Henderson
2024-06-07 0:24 ` [PATCH v2 2/2] util/bufferiszero: Add loongarch64 vector acceleration Richard Henderson
1 sibling, 0 replies; 4+ messages in thread
From: Richard Henderson @ 2024-06-07 0:24 UTC (permalink / raw)
To: qemu-devel; +Cc: maobibo, Philippe Mathieu-Daudé
Split out host/bufferiszero.h.inc for x86, aarch64 and generic
in order to avoid an overlong ifdef ladder.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
util/bufferiszero.c | 191 +------------------
host/include/aarch64/host/bufferiszero.c.inc | 76 ++++++++
host/include/generic/host/bufferiszero.c.inc | 10 +
host/include/i386/host/bufferiszero.c.inc | 124 ++++++++++++
host/include/x86_64/host/bufferiszero.c.inc | 1 +
5 files changed, 212 insertions(+), 190 deletions(-)
create mode 100644 host/include/aarch64/host/bufferiszero.c.inc
create mode 100644 host/include/generic/host/bufferiszero.c.inc
create mode 100644 host/include/i386/host/bufferiszero.c.inc
create mode 100644 host/include/x86_64/host/bufferiszero.c.inc
diff --git a/util/bufferiszero.c b/util/bufferiszero.c
index 11c080e02c..522146dab9 100644
--- a/util/bufferiszero.c
+++ b/util/bufferiszero.c
@@ -81,196 +81,7 @@ static bool buffer_is_zero_int_ge256(const void *buf, size_t len)
return t == 0;
}
-#if defined(CONFIG_AVX2_OPT) || defined(__SSE2__)
-#include <immintrin.h>
-
-/* Helper for preventing the compiler from reassociating
- chains of binary vector operations. */
-#define SSE_REASSOC_BARRIER(vec0, vec1) asm("" : "+x"(vec0), "+x"(vec1))
-
-/* Note that these vectorized functions may assume len >= 256. */
-
-static bool __attribute__((target("sse2")))
-buffer_zero_sse2(const void *buf, size_t len)
-{
- /* Unaligned loads at head/tail. */
- __m128i v = *(__m128i_u *)(buf);
- __m128i w = *(__m128i_u *)(buf + len - 16);
- /* Align head/tail to 16-byte boundaries. */
- const __m128i *p = QEMU_ALIGN_PTR_DOWN(buf + 16, 16);
- const __m128i *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 16);
- __m128i zero = { 0 };
-
- /* Collect a partial block at tail end. */
- v |= e[-1]; w |= e[-2];
- SSE_REASSOC_BARRIER(v, w);
- v |= e[-3]; w |= e[-4];
- SSE_REASSOC_BARRIER(v, w);
- v |= e[-5]; w |= e[-6];
- SSE_REASSOC_BARRIER(v, w);
- v |= e[-7]; v |= w;
-
- /*
- * Loop over complete 128-byte blocks.
- * With the head and tail removed, e - p >= 14, so the loop
- * must iterate at least once.
- */
- do {
- v = _mm_cmpeq_epi8(v, zero);
- if (unlikely(_mm_movemask_epi8(v) != 0xFFFF)) {
- return false;
- }
- v = p[0]; w = p[1];
- SSE_REASSOC_BARRIER(v, w);
- v |= p[2]; w |= p[3];
- SSE_REASSOC_BARRIER(v, w);
- v |= p[4]; w |= p[5];
- SSE_REASSOC_BARRIER(v, w);
- v |= p[6]; w |= p[7];
- SSE_REASSOC_BARRIER(v, w);
- v |= w;
- p += 8;
- } while (p < e - 7);
-
- return _mm_movemask_epi8(_mm_cmpeq_epi8(v, zero)) == 0xFFFF;
-}
-
-#ifdef CONFIG_AVX2_OPT
-static bool __attribute__((target("avx2")))
-buffer_zero_avx2(const void *buf, size_t len)
-{
- /* Unaligned loads at head/tail. */
- __m256i v = *(__m256i_u *)(buf);
- __m256i w = *(__m256i_u *)(buf + len - 32);
- /* Align head/tail to 32-byte boundaries. */
- const __m256i *p = QEMU_ALIGN_PTR_DOWN(buf + 32, 32);
- const __m256i *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 32);
- __m256i zero = { 0 };
-
- /* Collect a partial block at tail end. */
- v |= e[-1]; w |= e[-2];
- SSE_REASSOC_BARRIER(v, w);
- v |= e[-3]; w |= e[-4];
- SSE_REASSOC_BARRIER(v, w);
- v |= e[-5]; w |= e[-6];
- SSE_REASSOC_BARRIER(v, w);
- v |= e[-7]; v |= w;
-
- /* Loop over complete 256-byte blocks. */
- for (; p < e - 7; p += 8) {
- /* PTEST is not profitable here. */
- v = _mm256_cmpeq_epi8(v, zero);
- if (unlikely(_mm256_movemask_epi8(v) != 0xFFFFFFFF)) {
- return false;
- }
- v = p[0]; w = p[1];
- SSE_REASSOC_BARRIER(v, w);
- v |= p[2]; w |= p[3];
- SSE_REASSOC_BARRIER(v, w);
- v |= p[4]; w |= p[5];
- SSE_REASSOC_BARRIER(v, w);
- v |= p[6]; w |= p[7];
- SSE_REASSOC_BARRIER(v, w);
- v |= w;
- }
-
- return _mm256_movemask_epi8(_mm256_cmpeq_epi8(v, zero)) == 0xFFFFFFFF;
-}
-#endif /* CONFIG_AVX2_OPT */
-
-static biz_accel_fn const accel_table[] = {
- buffer_is_zero_int_ge256,
- buffer_zero_sse2,
-#ifdef CONFIG_AVX2_OPT
- buffer_zero_avx2,
-#endif
-};
-
-static unsigned best_accel(void)
-{
-#ifdef CONFIG_AVX2_OPT
- unsigned info = cpuinfo_init();
-
- if (info & CPUINFO_AVX2) {
- return 2;
- }
-#endif
- return 1;
-}
-
-#elif defined(__aarch64__) && defined(__ARM_NEON)
-#include <arm_neon.h>
-
-/*
- * Helper for preventing the compiler from reassociating
- * chains of binary vector operations.
- */
-#define REASSOC_BARRIER(vec0, vec1) asm("" : "+w"(vec0), "+w"(vec1))
-
-static bool buffer_is_zero_simd(const void *buf, size_t len)
-{
- uint32x4_t t0, t1, t2, t3;
-
- /* Align head/tail to 16-byte boundaries. */
- const uint32x4_t *p = QEMU_ALIGN_PTR_DOWN(buf + 16, 16);
- const uint32x4_t *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 16);
-
- /* Unaligned loads at head/tail. */
- t0 = vld1q_u32(buf) | vld1q_u32(buf + len - 16);
-
- /* Collect a partial block at tail end. */
- t1 = e[-7] | e[-6];
- t2 = e[-5] | e[-4];
- t3 = e[-3] | e[-2];
- t0 |= e[-1];
- REASSOC_BARRIER(t0, t1);
- REASSOC_BARRIER(t2, t3);
- t0 |= t1;
- t2 |= t3;
- REASSOC_BARRIER(t0, t2);
- t0 |= t2;
-
- /*
- * Loop over complete 128-byte blocks.
- * With the head and tail removed, e - p >= 14, so the loop
- * must iterate at least once.
- */
- do {
- /*
- * Reduce via UMAXV. Whatever the actual result,
- * it will only be zero if all input bytes are zero.
- */
- if (unlikely(vmaxvq_u32(t0) != 0)) {
- return false;
- }
-
- t0 = p[0] | p[1];
- t1 = p[2] | p[3];
- t2 = p[4] | p[5];
- t3 = p[6] | p[7];
- REASSOC_BARRIER(t0, t1);
- REASSOC_BARRIER(t2, t3);
- t0 |= t1;
- t2 |= t3;
- REASSOC_BARRIER(t0, t2);
- t0 |= t2;
- p += 8;
- } while (p < e - 7);
-
- return vmaxvq_u32(t0) == 0;
-}
-
-#define best_accel() 1
-static biz_accel_fn const accel_table[] = {
- buffer_is_zero_int_ge256,
- buffer_is_zero_simd,
-};
-#else
-#define best_accel() 0
-static biz_accel_fn const accel_table[1] = {
- buffer_is_zero_int_ge256
-};
-#endif
+#include "host/bufferiszero.c.inc"
static biz_accel_fn buffer_is_zero_accel;
static unsigned accel_index;
diff --git a/host/include/aarch64/host/bufferiszero.c.inc b/host/include/aarch64/host/bufferiszero.c.inc
new file mode 100644
index 0000000000..947ee7ca1f
--- /dev/null
+++ b/host/include/aarch64/host/bufferiszero.c.inc
@@ -0,0 +1,76 @@
+/*
+ * SPDX-License-Identifier: GPL-2.0-or-later
+ * buffer_is_zero acceleration, aarch64 version.
+ */
+
+#ifdef __ARM_NEON
+#include <arm_neon.h>
+
+/*
+ * Helper for preventing the compiler from reassociating
+ * chains of binary vector operations.
+ */
+#define REASSOC_BARRIER(vec0, vec1) asm("" : "+w"(vec0), "+w"(vec1))
+
+static bool buffer_is_zero_simd(const void *buf, size_t len)
+{
+ uint32x4_t t0, t1, t2, t3;
+
+ /* Align head/tail to 16-byte boundaries. */
+ const uint32x4_t *p = QEMU_ALIGN_PTR_DOWN(buf + 16, 16);
+ const uint32x4_t *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 16);
+
+ /* Unaligned loads at head/tail. */
+ t0 = vld1q_u32(buf) | vld1q_u32(buf + len - 16);
+
+ /* Collect a partial block at tail end. */
+ t1 = e[-7] | e[-6];
+ t2 = e[-5] | e[-4];
+ t3 = e[-3] | e[-2];
+ t0 |= e[-1];
+ REASSOC_BARRIER(t0, t1);
+ REASSOC_BARRIER(t2, t3);
+ t0 |= t1;
+ t2 |= t3;
+ REASSOC_BARRIER(t0, t2);
+ t0 |= t2;
+
+ /*
+ * Loop over complete 128-byte blocks.
+ * With the head and tail removed, e - p >= 14, so the loop
+ * must iterate at least once.
+ */
+ do {
+ /*
+ * Reduce via UMAXV. Whatever the actual result,
+ * it will only be zero if all input bytes are zero.
+ */
+ if (unlikely(vmaxvq_u32(t0) != 0)) {
+ return false;
+ }
+
+ t0 = p[0] | p[1];
+ t1 = p[2] | p[3];
+ t2 = p[4] | p[5];
+ t3 = p[6] | p[7];
+ REASSOC_BARRIER(t0, t1);
+ REASSOC_BARRIER(t2, t3);
+ t0 |= t1;
+ t2 |= t3;
+ REASSOC_BARRIER(t0, t2);
+ t0 |= t2;
+ p += 8;
+ } while (p < e - 7);
+
+ return vmaxvq_u32(t0) == 0;
+}
+
+static biz_accel_fn const accel_table[] = {
+ buffer_is_zero_int_ge256,
+ buffer_is_zero_simd,
+};
+
+#define best_accel() 1
+#else
+# include "host/include/generic/host/bufferiszero.c.inc"
+#endif
diff --git a/host/include/generic/host/bufferiszero.c.inc b/host/include/generic/host/bufferiszero.c.inc
new file mode 100644
index 0000000000..ea0875c24a
--- /dev/null
+++ b/host/include/generic/host/bufferiszero.c.inc
@@ -0,0 +1,10 @@
+/*
+ * SPDX-License-Identifier: GPL-2.0-or-later
+ * buffer_is_zero acceleration, generic version.
+ */
+
+static biz_accel_fn const accel_table[1] = {
+ buffer_is_zero_int_ge256
+};
+
+#define best_accel() 0
diff --git a/host/include/i386/host/bufferiszero.c.inc b/host/include/i386/host/bufferiszero.c.inc
new file mode 100644
index 0000000000..3b9605d806
--- /dev/null
+++ b/host/include/i386/host/bufferiszero.c.inc
@@ -0,0 +1,124 @@
+/*
+ * SPDX-License-Identifier: GPL-2.0-or-later
+ * buffer_is_zero acceleration, x86 version.
+ */
+
+#if defined(CONFIG_AVX2_OPT) || defined(__SSE2__)
+#include <immintrin.h>
+
+/* Helper for preventing the compiler from reassociating
+ chains of binary vector operations. */
+#define SSE_REASSOC_BARRIER(vec0, vec1) asm("" : "+x"(vec0), "+x"(vec1))
+
+/* Note that these vectorized functions may assume len >= 256. */
+
+static bool __attribute__((target("sse2")))
+buffer_zero_sse2(const void *buf, size_t len)
+{
+ /* Unaligned loads at head/tail. */
+ __m128i v = *(__m128i_u *)(buf);
+ __m128i w = *(__m128i_u *)(buf + len - 16);
+ /* Align head/tail to 16-byte boundaries. */
+ const __m128i *p = QEMU_ALIGN_PTR_DOWN(buf + 16, 16);
+ const __m128i *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 16);
+ __m128i zero = { 0 };
+
+ /* Collect a partial block at tail end. */
+ v |= e[-1]; w |= e[-2];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= e[-3]; w |= e[-4];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= e[-5]; w |= e[-6];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= e[-7]; v |= w;
+
+ /*
+ * Loop over complete 128-byte blocks.
+ * With the head and tail removed, e - p >= 14, so the loop
+ * must iterate at least once.
+ */
+ do {
+ v = _mm_cmpeq_epi8(v, zero);
+ if (unlikely(_mm_movemask_epi8(v) != 0xFFFF)) {
+ return false;
+ }
+ v = p[0]; w = p[1];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= p[2]; w |= p[3];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= p[4]; w |= p[5];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= p[6]; w |= p[7];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= w;
+ p += 8;
+ } while (p < e - 7);
+
+ return _mm_movemask_epi8(_mm_cmpeq_epi8(v, zero)) == 0xFFFF;
+}
+
+#ifdef CONFIG_AVX2_OPT
+static bool __attribute__((target("avx2")))
+buffer_zero_avx2(const void *buf, size_t len)
+{
+ /* Unaligned loads at head/tail. */
+ __m256i v = *(__m256i_u *)(buf);
+ __m256i w = *(__m256i_u *)(buf + len - 32);
+ /* Align head/tail to 32-byte boundaries. */
+ const __m256i *p = QEMU_ALIGN_PTR_DOWN(buf + 32, 32);
+ const __m256i *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 32);
+ __m256i zero = { 0 };
+
+ /* Collect a partial block at tail end. */
+ v |= e[-1]; w |= e[-2];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= e[-3]; w |= e[-4];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= e[-5]; w |= e[-6];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= e[-7]; v |= w;
+
+ /* Loop over complete 256-byte blocks. */
+ for (; p < e - 7; p += 8) {
+ /* PTEST is not profitable here. */
+ v = _mm256_cmpeq_epi8(v, zero);
+ if (unlikely(_mm256_movemask_epi8(v) != 0xFFFFFFFF)) {
+ return false;
+ }
+ v = p[0]; w = p[1];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= p[2]; w |= p[3];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= p[4]; w |= p[5];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= p[6]; w |= p[7];
+ SSE_REASSOC_BARRIER(v, w);
+ v |= w;
+ }
+
+ return _mm256_movemask_epi8(_mm256_cmpeq_epi8(v, zero)) == 0xFFFFFFFF;
+}
+#endif /* CONFIG_AVX2_OPT */
+
+static biz_accel_fn const accel_table[] = {
+ buffer_is_zero_int_ge256,
+ buffer_zero_sse2,
+#ifdef CONFIG_AVX2_OPT
+ buffer_zero_avx2,
+#endif
+};
+
+static unsigned best_accel(void)
+{
+#ifdef CONFIG_AVX2_OPT
+ unsigned info = cpuinfo_init();
+ if (info & CPUINFO_AVX2) {
+ return 2;
+ }
+#endif
+ return 1;
+}
+
+#else
+# include "host/include/generic/host/bufferiszero.c.inc"
+#endif
diff --git a/host/include/x86_64/host/bufferiszero.c.inc b/host/include/x86_64/host/bufferiszero.c.inc
new file mode 100644
index 0000000000..1d3f1fd6f5
--- /dev/null
+++ b/host/include/x86_64/host/bufferiszero.c.inc
@@ -0,0 +1 @@
+#include "host/include/i386/host/bufferiszero.c.inc"
--
2.34.1
^ permalink raw reply related [flat|nested] 4+ messages in thread
* [PATCH v2 2/2] util/bufferiszero: Add loongarch64 vector acceleration
2024-06-07 0:24 [PATCH v2 0/2] util/bufferiszero: Split out hosts, add loongarch64 Richard Henderson
2024-06-07 0:24 ` [PATCH v2 1/2] util/bufferiszero: Split out host include files Richard Henderson
@ 2024-06-07 0:24 ` Richard Henderson
2024-06-07 1:52 ` maobibo
1 sibling, 1 reply; 4+ messages in thread
From: Richard Henderson @ 2024-06-07 0:24 UTC (permalink / raw)
To: qemu-devel; +Cc: maobibo
Use inline assembly because no release compiler allows
per-function selection of the ISA.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
.../loongarch64/host/bufferiszero.c.inc | 143 ++++++++++++++++++
1 file changed, 143 insertions(+)
create mode 100644 host/include/loongarch64/host/bufferiszero.c.inc
diff --git a/host/include/loongarch64/host/bufferiszero.c.inc b/host/include/loongarch64/host/bufferiszero.c.inc
new file mode 100644
index 0000000000..69891eac80
--- /dev/null
+++ b/host/include/loongarch64/host/bufferiszero.c.inc
@@ -0,0 +1,143 @@
+/*
+ * SPDX-License-Identifier: GPL-2.0-or-later
+ * buffer_is_zero acceleration, loongarch64 version.
+ */
+
+/*
+ * Builtins for LSX and LASX are introduced by gcc 14 and llvm 18,
+ * but as yet neither has support for attribute target, so neither
+ * is able to enable the optimization without globally enabling
+ * vector support. Since we want runtime detection, use assembly.
+ */
+
+static bool buffer_is_zero_lsx(const void *buf, size_t len)
+{
+ const void *p = QEMU_ALIGN_PTR_DOWN(buf + 16, 16);
+ const void *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 16) - (7 * 16);
+ const void *l = buf + len;
+ bool ret;
+
+ asm("vld $vr0,%2,0\n\t" /* first: buf + 0 */
+ "vld $vr1,%4,-16\n\t" /* last: buf + len - 16 */
+ "vld $vr2,%3,0\n\t" /* e[0] */
+ "vld $vr3,%3,16\n\t" /* e[1] */
+ "vld $vr4,%3,32\n\t" /* e[2] */
+ "vld $vr5,%3,48\n\t" /* e[3] */
+ "vld $vr6,%3,64\n\t" /* e[4] */
+ "vld $vr7,%3,80\n\t" /* e[5] */
+ "vld $vr8,%3,96\n\t" /* e[6] */
+ "vor.v $vr0,$vr0,$vr1\n\t"
+ "vor.v $vr2,$vr2,$vr3\n\t"
+ "vor.v $vr4,$vr4,$vr5\n\t"
+ "vor.v $vr6,$vr6,$vr7\n\t"
+ "vor.v $vr0,$vr0,$vr2\n\t"
+ "vor.v $vr4,$vr4,$vr6\n\t"
+ "vor.v $vr0,$vr0,$vr4\n\t"
+ "vor.v $vr0,$vr0,$vr8\n\t"
+ "or %0,$r0,$r0\n" /* prepare return false */
+ "1:\n\t"
+ "vsetnez.v $fcc0,$vr0\n\t"
+ "bcnez $fcc0,2f\n\t"
+ "vld $vr0,%1,0\n\t" /* p[0] */
+ "vld $vr1,%1,16\n\t" /* p[1] */
+ "vld $vr2,%1,32\n\t" /* p[2] */
+ "vld $vr3,%1,48\n\t" /* p[3] */
+ "vld $vr4,%1,64\n\t" /* p[4] */
+ "vld $vr5,%1,80\n\t" /* p[5] */
+ "vld $vr6,%1,96\n\t" /* p[6] */
+ "vld $vr7,%1,112\n\t" /* p[7] */
+ "addi.d %1,%1,128\n\t"
+ "vor.v $vr0,$vr0,$vr1\n\t"
+ "vor.v $vr2,$vr2,$vr3\n\t"
+ "vor.v $vr4,$vr4,$vr5\n\t"
+ "vor.v $vr6,$vr6,$vr7\n\t"
+ "vor.v $vr0,$vr0,$vr2\n\t"
+ "vor.v $vr4,$vr4,$vr6\n\t"
+ "vor.v $vr0,$vr0,$vr4\n\t"
+ "bltu %1,%3,1b\n\t"
+ "vsetnez.v $fcc0,$vr0\n\t"
+ "bcnez $fcc0,2f\n\t"
+ "ori %0,$r0,1\n"
+ "2:"
+ : "=&r"(ret), "+r"(p)
+ : "r"(buf), "r"(e), "r"(l)
+ : "f0", "f1", "f2", "f3", "f4", "f5", "f6", "f7", "f8", "fcc0");
+
+ return ret;
+}
+
+static bool buffer_is_zero_lasx(const void *buf, size_t len)
+{
+ const void *p = QEMU_ALIGN_PTR_DOWN(buf + 32, 32);
+ const void *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 32) - (7 * 32);
+ const void *l = buf + len;
+ bool ret;
+
+ asm("xvld $xr0,%2,0\n\t" /* first: buf + 0 */
+ "xvld $xr1,%4,-32\n\t" /* last: buf + len - 32 */
+ "xvld $xr2,%3,0\n\t" /* e[0] */
+ "xvld $xr3,%3,32\n\t" /* e[1] */
+ "xvld $xr4,%3,64\n\t" /* e[2] */
+ "xvld $xr5,%3,96\n\t" /* e[3] */
+ "xvld $xr6,%3,128\n\t" /* e[4] */
+ "xvld $xr7,%3,160\n\t" /* e[5] */
+ "xvld $xr8,%3,192\n\t" /* e[6] */
+ "xvor.v $xr0,$xr0,$xr1\n\t"
+ "xvor.v $xr2,$xr2,$xr3\n\t"
+ "xvor.v $xr4,$xr4,$xr5\n\t"
+ "xvor.v $xr6,$xr6,$xr7\n\t"
+ "xvor.v $xr0,$xr0,$xr2\n\t"
+ "xvor.v $xr4,$xr4,$xr6\n\t"
+ "xvor.v $xr0,$xr0,$xr4\n\t"
+ "xvor.v $xr0,$xr0,$xr8\n\t"
+ "or %0,$r0,$r0\n\t" /* prepare return false */
+ "bgeu %1,%3,2f\n"
+ "1:\n\t"
+ "xvsetnez.v $fcc0,$xr0\n\t"
+ "bcnez $fcc0,3f\n\t"
+ "xvld $xr0,%1,0\n\t" /* p[0] */
+ "xvld $xr1,%1,32\n\t" /* p[1] */
+ "xvld $xr2,%1,64\n\t" /* p[2] */
+ "xvld $xr3,%1,96\n\t" /* p[3] */
+ "xvld $xr4,%1,128\n\t" /* p[4] */
+ "xvld $xr5,%1,160\n\t" /* p[5] */
+ "xvld $xr6,%1,192\n\t" /* p[6] */
+ "xvld $xr7,%1,224\n\t" /* p[7] */
+ "addi.d %1,%1,256\n\t"
+ "xvor.v $xr0,$xr0,$xr1\n\t"
+ "xvor.v $xr2,$xr2,$xr3\n\t"
+ "xvor.v $xr4,$xr4,$xr5\n\t"
+ "xvor.v $xr6,$xr6,$xr7\n\t"
+ "xvor.v $xr0,$xr0,$xr2\n\t"
+ "xvor.v $xr4,$xr4,$xr6\n\t"
+ "xvor.v $xr0,$xr0,$xr4\n\t"
+ "bltu %1,%3,1b\n"
+ "2:\n\t"
+ "xvsetnez.v $fcc0,$xr0\n\t"
+ "bcnez $fcc0,3f\n\t"
+ "ori %0,$r0,1\n"
+ "3:"
+ : "=&r"(ret), "+r"(p)
+ : "r"(buf), "r"(e), "r"(l)
+ : "f0", "f1", "f2", "f3", "f4", "f5", "f6", "f7", "f8", "fcc0");
+
+ return ret;
+}
+
+static biz_accel_fn const accel_table[] = {
+ buffer_is_zero_int_ge256,
+ buffer_is_zero_lsx,
+ buffer_is_zero_lasx,
+};
+
+static unsigned best_accel(void)
+{
+ unsigned info = cpuinfo_init();
+ if (info & CPUINFO_LASX) {
+ return 2;
+ }
+ if (info & CPUINFO_LSX) {
+ return 1;
+ }
+ return 0;
+}
--
2.34.1
^ permalink raw reply related [flat|nested] 4+ messages in thread
* Re: [PATCH v2 2/2] util/bufferiszero: Add loongarch64 vector acceleration
2024-06-07 0:24 ` [PATCH v2 2/2] util/bufferiszero: Add loongarch64 vector acceleration Richard Henderson
@ 2024-06-07 1:52 ` maobibo
0 siblings, 0 replies; 4+ messages in thread
From: maobibo @ 2024-06-07 1:52 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 2024/6/7 上午8:24, Richard Henderson wrote:
> Use inline assembly because no release compiler allows
> per-function selection of the ISA.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> .../loongarch64/host/bufferiszero.c.inc | 143 ++++++++++++++++++
> 1 file changed, 143 insertions(+)
> create mode 100644 host/include/loongarch64/host/bufferiszero.c.inc
>
> diff --git a/host/include/loongarch64/host/bufferiszero.c.inc b/host/include/loongarch64/host/bufferiszero.c.inc
> new file mode 100644
> index 0000000000..69891eac80
> --- /dev/null
> +++ b/host/include/loongarch64/host/bufferiszero.c.inc
> @@ -0,0 +1,143 @@
> +/*
> + * SPDX-License-Identifier: GPL-2.0-or-later
> + * buffer_is_zero acceleration, loongarch64 version.
> + */
> +
> +/*
> + * Builtins for LSX and LASX are introduced by gcc 14 and llvm 18,
> + * but as yet neither has support for attribute target, so neither
> + * is able to enable the optimization without globally enabling
> + * vector support. Since we want runtime detection, use assembly.
> + */
> +
> +static bool buffer_is_zero_lsx(const void *buf, size_t len)
> +{
> + const void *p = QEMU_ALIGN_PTR_DOWN(buf + 16, 16);
> + const void *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 16) - (7 * 16);
> + const void *l = buf + len;
> + bool ret;
> +
> + asm("vld $vr0,%2,0\n\t" /* first: buf + 0 */
> + "vld $vr1,%4,-16\n\t" /* last: buf + len - 16 */
> + "vld $vr2,%3,0\n\t" /* e[0] */
> + "vld $vr3,%3,16\n\t" /* e[1] */
> + "vld $vr4,%3,32\n\t" /* e[2] */
> + "vld $vr5,%3,48\n\t" /* e[3] */
> + "vld $vr6,%3,64\n\t" /* e[4] */
> + "vld $vr7,%3,80\n\t" /* e[5] */
> + "vld $vr8,%3,96\n\t" /* e[6] */
> + "vor.v $vr0,$vr0,$vr1\n\t"
> + "vor.v $vr2,$vr2,$vr3\n\t"
> + "vor.v $vr4,$vr4,$vr5\n\t"
> + "vor.v $vr6,$vr6,$vr7\n\t"
> + "vor.v $vr0,$vr0,$vr2\n\t"
> + "vor.v $vr4,$vr4,$vr6\n\t"
> + "vor.v $vr0,$vr0,$vr4\n\t"
> + "vor.v $vr0,$vr0,$vr8\n\t"
> + "or %0,$r0,$r0\n" /* prepare return false */
> + "1:\n\t"
> + "vsetnez.v $fcc0,$vr0\n\t"
> + "bcnez $fcc0,2f\n\t"
> + "vld $vr0,%1,0\n\t" /* p[0] */
> + "vld $vr1,%1,16\n\t" /* p[1] */
> + "vld $vr2,%1,32\n\t" /* p[2] */
> + "vld $vr3,%1,48\n\t" /* p[3] */
> + "vld $vr4,%1,64\n\t" /* p[4] */
> + "vld $vr5,%1,80\n\t" /* p[5] */
> + "vld $vr6,%1,96\n\t" /* p[6] */
> + "vld $vr7,%1,112\n\t" /* p[7] */
> + "addi.d %1,%1,128\n\t"
> + "vor.v $vr0,$vr0,$vr1\n\t"
> + "vor.v $vr2,$vr2,$vr3\n\t"
> + "vor.v $vr4,$vr4,$vr5\n\t"
> + "vor.v $vr6,$vr6,$vr7\n\t"
> + "vor.v $vr0,$vr0,$vr2\n\t"
> + "vor.v $vr4,$vr4,$vr6\n\t"
> + "vor.v $vr0,$vr0,$vr4\n\t"
> + "bltu %1,%3,1b\n\t"
> + "vsetnez.v $fcc0,$vr0\n\t"
> + "bcnez $fcc0,2f\n\t"
> + "ori %0,$r0,1\n"
> + "2:"
> + : "=&r"(ret), "+r"(p)
> + : "r"(buf), "r"(e), "r"(l)
> + : "f0", "f1", "f2", "f3", "f4", "f5", "f6", "f7", "f8", "fcc0");
> +
> + return ret;
> +}
> +
> +static bool buffer_is_zero_lasx(const void *buf, size_t len)
> +{
> + const void *p = QEMU_ALIGN_PTR_DOWN(buf + 32, 32);
> + const void *e = QEMU_ALIGN_PTR_DOWN(buf + len - 1, 32) - (7 * 32);
> + const void *l = buf + len;
> + bool ret;
> +
> + asm("xvld $xr0,%2,0\n\t" /* first: buf + 0 */
> + "xvld $xr1,%4,-32\n\t" /* last: buf + len - 32 */
> + "xvld $xr2,%3,0\n\t" /* e[0] */
> + "xvld $xr3,%3,32\n\t" /* e[1] */
> + "xvld $xr4,%3,64\n\t" /* e[2] */
> + "xvld $xr5,%3,96\n\t" /* e[3] */
> + "xvld $xr6,%3,128\n\t" /* e[4] */
> + "xvld $xr7,%3,160\n\t" /* e[5] */
> + "xvld $xr8,%3,192\n\t" /* e[6] */
> + "xvor.v $xr0,$xr0,$xr1\n\t"
> + "xvor.v $xr2,$xr2,$xr3\n\t"
> + "xvor.v $xr4,$xr4,$xr5\n\t"
> + "xvor.v $xr6,$xr6,$xr7\n\t"
> + "xvor.v $xr0,$xr0,$xr2\n\t"
> + "xvor.v $xr4,$xr4,$xr6\n\t"
> + "xvor.v $xr0,$xr0,$xr4\n\t"
> + "xvor.v $xr0,$xr0,$xr8\n\t"
> + "or %0,$r0,$r0\n\t" /* prepare return false */
> + "bgeu %1,%3,2f\n"
> + "1:\n\t"
> + "xvsetnez.v $fcc0,$xr0\n\t"
> + "bcnez $fcc0,3f\n\t"
> + "xvld $xr0,%1,0\n\t" /* p[0] */
> + "xvld $xr1,%1,32\n\t" /* p[1] */
> + "xvld $xr2,%1,64\n\t" /* p[2] */
> + "xvld $xr3,%1,96\n\t" /* p[3] */
> + "xvld $xr4,%1,128\n\t" /* p[4] */
> + "xvld $xr5,%1,160\n\t" /* p[5] */
> + "xvld $xr6,%1,192\n\t" /* p[6] */
> + "xvld $xr7,%1,224\n\t" /* p[7] */
> + "addi.d %1,%1,256\n\t"
> + "xvor.v $xr0,$xr0,$xr1\n\t"
> + "xvor.v $xr2,$xr2,$xr3\n\t"
> + "xvor.v $xr4,$xr4,$xr5\n\t"
> + "xvor.v $xr6,$xr6,$xr7\n\t"
> + "xvor.v $xr0,$xr0,$xr2\n\t"
> + "xvor.v $xr4,$xr4,$xr6\n\t"
> + "xvor.v $xr0,$xr0,$xr4\n\t"
> + "bltu %1,%3,1b\n"
> + "2:\n\t"
> + "xvsetnez.v $fcc0,$xr0\n\t"
> + "bcnez $fcc0,3f\n\t"
> + "ori %0,$r0,1\n"
> + "3:"
> + : "=&r"(ret), "+r"(p)
> + : "r"(buf), "r"(e), "r"(l)
> + : "f0", "f1", "f2", "f3", "f4", "f5", "f6", "f7", "f8", "fcc0");
> +
> + return ret;
> +}
> +
> +static biz_accel_fn const accel_table[] = {
> + buffer_is_zero_int_ge256,
> + buffer_is_zero_lsx,
> + buffer_is_zero_lasx,
> +};
> +
> +static unsigned best_accel(void)
> +{
> + unsigned info = cpuinfo_init();
> + if (info & CPUINFO_LASX) {
> + return 2;
> + }
> + if (info & CPUINFO_LSX) {
> + return 1;
> + }
> + return 0;
> +}
>
This is better than mine, there is no dependency on compiler option any
more.
Tested-by: Bibo Mao <maobibo@loongson.cn>
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2024-06-07 1:53 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2024-06-07 0:24 [PATCH v2 0/2] util/bufferiszero: Split out hosts, add loongarch64 Richard Henderson
2024-06-07 0:24 ` [PATCH v2 1/2] util/bufferiszero: Split out host include files Richard Henderson
2024-06-07 0:24 ` [PATCH v2 2/2] util/bufferiszero: Add loongarch64 vector acceleration Richard Henderson
2024-06-07 1:52 ` maobibo
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).