Message ID | 20240215081449.848220-10-richard.henderson@linaro.org |
---|---|
State | Superseded |
Headers | show |
Series | Optimize buffer_is_zero | expand |
On Wed, 14 Feb 2024, Richard Henderson wrote: > Because non-embedded aarch64 is expected to have AdvSIMD enabled, merely > double-check with the compiler flags for __ARM_NEON and don't bother with > a runtime check. Otherwise, model the loop after the x86 SSE2 function, > and use VADDV to reduce the four vector comparisons. I am not very familiar with Neon but I wonder if this couldn't use SHRN for the final 128b->64b reduction similar to 2022 Glibc optimizations: https://inbox.sourceware.org/libc-alpha/20220620174628.2820531-1-danilak@google.com/ In git history I see the previous Neon buffer_is_zero was removed because it was not faster. Is it because integer LDP was as good as vector loads at saturating load bandwidth on older cores, and things are different now? Alexander > > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> > --- > util/bufferiszero.c | 74 +++++++++++++++++++++++++++++++++++++++++++++ > 1 file changed, 74 insertions(+) > > diff --git a/util/bufferiszero.c b/util/bufferiszero.c > index 4eef6d47bc..2809b09225 100644 > --- a/util/bufferiszero.c > +++ b/util/bufferiszero.c > @@ -214,7 +214,81 @@ bool test_buffer_is_zero_next_accel(void) > } > return false; > } > + > +#elif defined(__aarch64__) && defined(__ARM_NEON) > +#include <arm_neon.h> > + > +#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 { > + /* Each comparison is [-1,0], so reduction is in [-4..0]. */ > + if (unlikely(vaddvq_u32(vceqzq_u32(t0)) != -4)) { > + 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 vaddvq_u32(vceqzq_u32(t0)) == -4; > +} > + > +static biz_accel_fn const accel_table[] = { > + buffer_is_zero_int_ge256, > + buffer_is_zero_simd, > +}; > + > +static unsigned accel_index = 1; > +#define INIT_ACCEL buffer_is_zero_simd > + > +bool test_buffer_is_zero_next_accel(void) > +{ > + if (accel_index != 0) { > + buffer_is_zero_accel = accel_table[--accel_index]; > + return true; > + } > + return false; > +} > + > #else > + > bool test_buffer_is_zero_next_accel(void) > { > return false; >
On 2/14/24 22:47, Alexander Monakov wrote: > > On Wed, 14 Feb 2024, Richard Henderson wrote: > >> Because non-embedded aarch64 is expected to have AdvSIMD enabled, merely >> double-check with the compiler flags for __ARM_NEON and don't bother with >> a runtime check. Otherwise, model the loop after the x86 SSE2 function, >> and use VADDV to reduce the four vector comparisons. > > I am not very familiar with Neon but I wonder if this couldn't use SHRN > for the final 128b->64b reduction similar to 2022 Glibc optimizations: > https://inbox.sourceware.org/libc-alpha/20220620174628.2820531-1-danilak@google.com/ The reason they use SHRN for memchr is that they have also applied a mask to the comparison so that they can identify which byte contained the match. That is not required here, so any reduction will do. > In git history I see the previous Neon buffer_is_zero was removed because > it was not faster. Is it because integer LDP was as good as vector loads > at saturating load bandwidth on older cores, and things are different now? The old reduction was a bit silly, -#define DO_NONZERO(X) (vgetq_lane_u64((X), 0) | vgetq_lane_u64((X), 1)) performing two cross-register-set fetches. It's also possible that we were saturating the load bandwidth on the old mustang. This time I'm testing on a neoverse-n1, which is quite a few years newer. The loop kernel compiles to this: 19c: ad401c20 ldp q0, q7, [x1] 1a0: ad411823 ldp q3, q6, [x1, #32] 1a4: ad421421 ldp q1, q5, [x1, #64] 1a8: ad431022 ldp q2, q4, [x1, #96] 1ac: 91020021 add x1, x1, #0x80 1b0: 4ea71c00 orr v0.16b, v0.16b, v7.16b 1b4: 4ea61c63 orr v3.16b, v3.16b, v6.16b 1b8: 4ea51c21 orr v1.16b, v1.16b, v5.16b 1bc: 4ea41c42 orr v2.16b, v2.16b, v4.16b 1c0: 4ea31c00 orr v0.16b, v0.16b, v3.16b 1c4: 4ea21c21 orr v1.16b, v1.16b, v2.16b 1c8: 4ea11c00 orr v0.16b, v0.16b, v1.16b 1cc: eb03003f cmp x1, x3 1d0: 54000162 b.cs 1fc <buffer_is_zero_simd+0xb8> // b.hs, b.nlast 1d4: 4ea09800 cmeq v0.4s, v0.4s, #0 1d8: 4eb1b800 addv s0, v0.4s 1dc: 1e260000 fmov w0, s0 1e0: 3100101f cmn w0, #0x4 1e4: 54fffdc0 b.eq 19c <buffer_is_zero_simd+0x58> // b.none r~ > > Alexander > >> >> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> >> --- >> util/bufferiszero.c | 74 +++++++++++++++++++++++++++++++++++++++++++++ >> 1 file changed, 74 insertions(+) >> >> diff --git a/util/bufferiszero.c b/util/bufferiszero.c >> index 4eef6d47bc..2809b09225 100644 >> --- a/util/bufferiszero.c >> +++ b/util/bufferiszero.c >> @@ -214,7 +214,81 @@ bool test_buffer_is_zero_next_accel(void) >> } >> return false; >> } >> + >> +#elif defined(__aarch64__) && defined(__ARM_NEON) >> +#include <arm_neon.h> >> + >> +#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 { >> + /* Each comparison is [-1,0], so reduction is in [-4..0]. */ >> + if (unlikely(vaddvq_u32(vceqzq_u32(t0)) != -4)) { >> + 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 vaddvq_u32(vceqzq_u32(t0)) == -4; >> +} >> + >> +static biz_accel_fn const accel_table[] = { >> + buffer_is_zero_int_ge256, >> + buffer_is_zero_simd, >> +}; >> + >> +static unsigned accel_index = 1; >> +#define INIT_ACCEL buffer_is_zero_simd >> + >> +bool test_buffer_is_zero_next_accel(void) >> +{ >> + if (accel_index != 0) { >> + buffer_is_zero_accel = accel_table[--accel_index]; >> + return true; >> + } >> + return false; >> +} >> + >> #else >> + >> bool test_buffer_is_zero_next_accel(void) >> { >> return false; >>
On Thu, 15 Feb 2024, Richard Henderson wrote: > On 2/14/24 22:47, Alexander Monakov wrote: > > > > On Wed, 14 Feb 2024, Richard Henderson wrote: > > > >> Because non-embedded aarch64 is expected to have AdvSIMD enabled, merely > >> double-check with the compiler flags for __ARM_NEON and don't bother with > >> a runtime check. Otherwise, model the loop after the x86 SSE2 function, > >> and use VADDV to reduce the four vector comparisons. > > > > I am not very familiar with Neon but I wonder if this couldn't use SHRN > > for the final 128b->64b reduction similar to 2022 Glibc optimizations: > > https://inbox.sourceware.org/libc-alpha/20220620174628.2820531-1-danilak@google.com/ > > The reason they use SHRN for memchr is that they have also applied a mask > to the comparison so that they can identify which byte contained the match. > That is not required here, so any reduction will do. Right, so we can pick the cheapest reduction method, and if I'm reading Neoverse-N1 SOG right, SHRN is marginally cheaper than ADDV (latency 2 instead of 3), and it should be generally preferable on other cores, no? For that matter, cannot UQXTN (unsigned saturating extract narrow) be used in place of CMEQ+ADDV here? Alexander
On 2/15/24 08:46, Alexander Monakov wrote: > Right, so we can pick the cheapest reduction method, and if I'm reading > Neoverse-N1 SOG right, SHRN is marginally cheaper than ADDV (latency 2 > instead of 3), and it should be generally preferable on other cores, no? Fair. > For that matter, cannot UQXTN (unsigned saturating extract narrow) be > used in place of CMEQ+ADDV here? Interesting. I hadn't thought about using saturation to preserve non-zeroness like that. Using 1 4-cycle insn instead of 2 2-cycle insns is interesting as well. I suppose, since it's at the end of the dependency chain, the fact that it is restricted to the V1 pipe matters not at all. r~
diff --git a/util/bufferiszero.c b/util/bufferiszero.c index 4eef6d47bc..2809b09225 100644 --- a/util/bufferiszero.c +++ b/util/bufferiszero.c @@ -214,7 +214,81 @@ bool test_buffer_is_zero_next_accel(void) } return false; } + +#elif defined(__aarch64__) && defined(__ARM_NEON) +#include <arm_neon.h> + +#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 { + /* Each comparison is [-1,0], so reduction is in [-4..0]. */ + if (unlikely(vaddvq_u32(vceqzq_u32(t0)) != -4)) { + 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 vaddvq_u32(vceqzq_u32(t0)) == -4; +} + +static biz_accel_fn const accel_table[] = { + buffer_is_zero_int_ge256, + buffer_is_zero_simd, +}; + +static unsigned accel_index = 1; +#define INIT_ACCEL buffer_is_zero_simd + +bool test_buffer_is_zero_next_accel(void) +{ + if (accel_index != 0) { + buffer_is_zero_accel = accel_table[--accel_index]; + return true; + } + return false; +} + #else + bool test_buffer_is_zero_next_accel(void) { return false;
Because non-embedded aarch64 is expected to have AdvSIMD enabled, merely double-check with the compiler flags for __ARM_NEON and don't bother with a runtime check. Otherwise, model the loop after the x86 SSE2 function, and use VADDV to reduce the four vector comparisons. Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- util/bufferiszero.c | 74 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 74 insertions(+)