[PATCH 2/2] arm: Auto-vectorization for MVE: add pack/unpack patterns
Christophe Lyon
christophe.lyon@linaro.org
Thu Jun 10 13:46:52 GMT 2021
On Thu, 10 Jun 2021 at 11:50, Christophe Lyon
<christophe.lyon@linaro.org> wrote:
>
> On Tue, 8 Jun 2021 at 14:10, Richard Sandiford
> <richard.sandiford@arm.com> wrote:
> >
> > Christophe Lyon <christophe.lyon@linaro.org> writes:
> > > This patch adds vec_unpack<US>_hi_<mode>, vec_unpack<US>_lo_<mode>,
> > > vec_pack_trunc_<mode> patterns for MVE.
> > >
> > > It does so by moving the unpack patterns from neon.md to
> > > vec-common.md, while adding them support for MVE. The pack expander is
> > > derived from the Neon one (which in turn is renamed into
> > > neon_quad_vec_pack_trunc_<mode>).
> > >
> > > The patch introduces mve_vec_pack_trunc_<mode> to avoid the need for a
> > > zero-initialized temporary, which is needed if the
> > > vec_pack_trunc_<mode> expander calls @mve_vmovn[bt]q_<supf><mode>
> > > instead.
> > >
> > > With this patch, we can now vectorize the 16 and 8-bit versions of
> > > vclz and vshl, although the generated code could still be improved.
> > > For test_clz_s16, we now generate
> > > vldrh.16 q3, [r1]
> > > vmovlb.s16 q2, q3
> > > vmovlt.s16 q3, q3
> > > vclz.i32 q2, q2
> > > vclz.i32 q3, q3
> > > vmovnb.i32 q1, q2
> > > vmovnt.i32 q1, q3
> > > vstrh.16 q1, [r0]
> > > which could be improved to
> > > vldrh.16 q3, [r1]
> > > vclz.i16 q1, q3
> > > vstrh.16 q1, [r0]
> > > if we could avoid the need for unpack/pack steps.
> >
> > Yeah, there was a PR about fixing this for popcount. I guess the same
> > approach would apply here too.
> >
> > > For reference, clang-12 generates:
> > > vldrh.s32 q0, [r1]
> > > vldrh.s32 q1, [r1, #8]
> > > vclz.i32 q0, q0
> > > vstrh.32 q0, [r0]
> > > vclz.i32 q0, q1
> > > vstrh.32 q0, [r0, #8]
> > >
> > > 2021-06-03 Christophe Lyon <christophe.lyon@linaro.org>
> > >
> > > gcc/
> > > * config/arm/mve.md (mve_vmovltq_<supf><mode>): Prefix with '@'.
> > > (mve_vmovlbq_<supf><mode>): Likewise.
> > > (mve_vmovnbq_<supf><mode>): Likewise.
> > > (mve_vmovntq_<supf><mode>): Likewise.
> > > (@mve_vec_pack_trunc_<mode>): New pattern.
> > > * config/arm/neon.md (vec_unpack<US>_hi_<mode>): Move to
> > > vec-common.md.
> > > (vec_unpack<US>_lo_<mode>): Likewise.
> > > (vec_pack_trunc_<mode>): Rename to
> > > neon_quad_vec_pack_trunc_<mode>.
> > > * config/arm/vec-common.md (vec_unpack<US>_hi_<mode>): New
> > > pattern.
> > > (vec_unpack<US>_lo_<mode>): New.
> > > (vec_pack_trunc_<mode>): New.
> > >
> > > gcc/testsuite/
> > > * gcc.target/arm/simd/mve-vclz.c: Update expected results.
> > > * gcc.target/arm/simd/mve-vshl.c: Likewise.
> > > ---
> > > gcc/config/arm/mve.md | 20 ++++-
> > > gcc/config/arm/neon.md | 39 +--------
> > > gcc/config/arm/vec-common.md | 89 ++++++++++++++++++++
> > > gcc/testsuite/gcc.target/arm/simd/mve-vclz.c | 7 +-
> > > gcc/testsuite/gcc.target/arm/simd/mve-vshl.c | 5 +-
> > > 5 files changed, 114 insertions(+), 46 deletions(-)
> > >
> > > diff --git a/gcc/config/arm/mve.md b/gcc/config/arm/mve.md
> > > index 99e46d0bc69..b18292c07d3 100644
> > > --- a/gcc/config/arm/mve.md
> > > +++ b/gcc/config/arm/mve.md
> > > @@ -510,7 +510,7 @@ (define_insn "mve_vrev32q_<supf><mode>"
> > > ;;
> > > ;; [vmovltq_u, vmovltq_s])
> > > ;;
> > > -(define_insn "mve_vmovltq_<supf><mode>"
> > > +(define_insn "@mve_vmovltq_<supf><mode>"
> > > [
> > > (set (match_operand:<V_double_width> 0 "s_register_operand" "=w")
> > > (unspec:<V_double_width> [(match_operand:MVE_3 1 "s_register_operand" "w")]
> > > @@ -524,7 +524,7 @@ (define_insn "mve_vmovltq_<supf><mode>"
> > > ;;
> > > ;; [vmovlbq_s, vmovlbq_u])
> > > ;;
> > > -(define_insn "mve_vmovlbq_<supf><mode>"
> > > +(define_insn "@mve_vmovlbq_<supf><mode>"
> > > [
> > > (set (match_operand:<V_double_width> 0 "s_register_operand" "=w")
> > > (unspec:<V_double_width> [(match_operand:MVE_3 1 "s_register_operand" "w")]
> > > @@ -2187,7 +2187,7 @@ (define_insn "mve_vmlsldavxq_s<mode>"
> > > ;;
> > > ;; [vmovnbq_u, vmovnbq_s])
> > > ;;
> > > -(define_insn "mve_vmovnbq_<supf><mode>"
> > > +(define_insn "@mve_vmovnbq_<supf><mode>"
> > > [
> > > (set (match_operand:<V_narrow_pack> 0 "s_register_operand" "=w")
> > > (unspec:<V_narrow_pack> [(match_operand:<V_narrow_pack> 1 "s_register_operand" "0")
> > > @@ -2202,7 +2202,7 @@ (define_insn "mve_vmovnbq_<supf><mode>"
> > > ;;
> > > ;; [vmovntq_s, vmovntq_u])
> > > ;;
> > > -(define_insn "mve_vmovntq_<supf><mode>"
> > > +(define_insn "@mve_vmovntq_<supf><mode>"
> > > [
> > > (set (match_operand:<V_narrow_pack> 0 "s_register_operand" "=w")
> > > (unspec:<V_narrow_pack> [(match_operand:<V_narrow_pack> 1 "s_register_operand" "0")
> > > @@ -2214,6 +2214,18 @@ (define_insn "mve_vmovntq_<supf><mode>"
> > > [(set_attr "type" "mve_move")
> > > ])
> > >
> > > +(define_insn "@mve_vec_pack_trunc_<mode>"
> > > + [(set (match_operand:<V_narrow_pack> 0 "register_operand" "=&w")
> > > + (vec_concat:<V_narrow_pack>
> > > + (truncate:<V_narrow>
> > > + (match_operand:MVE_5 1 "register_operand" "w"))
> > > + (truncate:<V_narrow>
> > > + (match_operand:MVE_5 2 "register_operand" "w"))))]
> > > + "TARGET_HAVE_MVE"
> > > + "vmovnb.i<V_sz_elem> %q0, %q1\;vmovnt.i<V_sz_elem> %q0, %q2"
> > > + [(set_attr "type" "mve_move")]
> > > +)
> > > +
> >
> > I realise this is (like you say) based on the neon.md pattern,
> > but we should use separate vmovnb and vmovnt instructions instead
> > of putting two instructions into a single pattern.
> >
> > One specific advantage to using separate patterns is that it would
> > avoid the imprecision of the earlyclobber: the output only conflicts
> > with operand 1 and can be tied to operand 2.
> >
> I'm not sure to follow: with the current patch, for test_clz_s16 we generate:
> vldrh.16 q3, [r1]
> vmovlb.s16 q2, q3
> vmovlt.s16 q3, q3
> vclz.i32 q2, q2
> vclz.i32 q3, q3
> vmovnb.i32 q1, q2
> vmovnt.i32 q1, q3
> vstrh.16 q1, [r0]
>
> If I remove the define_insn "@mve_vec_pack_trunc_<mode>" pattern
> and update define_expand "vec_pack_trunc_<mode>" to call:
> emit_insn (gen_mve_vmovnbq (VMOVNBQ_S, <MODE>mode, operands[0],
> operands[0], operands[1]));
> emit_insn (gen_mve_vmovntq (VMOVNTQ_S, <MODE>mode, operands[0],
> operands[0], operands[2]));
> instead, we now generate:
> vldrh.16 q3, [r1]
> vmovlb.s16 q1, q3
> vldr.64 d4, .L5
> vldr.64 d5, .L5+8
> vmovlt.s16 q3, q3
> vclz.i32 q1, q1
> vclz.i32 q3, q3
> vmovnb.i32 q2, q1
> vmovnt.i32 q2, q3
> vstrh.16 q2, [r0]
> bx lr
> .L6:
> .align 3
> .L5:
> .short 0
> .short 0
> .short 0
> .short 0
> .short 0
> .short 0
> .short 0
> .short 0
>
> That's what I tried to explain in the commit message.
>
> I guess I've misunderstood your comment?
>
> Thanks,
>
> Christophe
>
> > > ;;
> > > ;; [vmulq_f])
> > > ;;
> > > diff --git a/gcc/config/arm/neon.md b/gcc/config/arm/neon.md
> > > index 0fdffaf4ec4..392d9607919 100644
> > > --- a/gcc/config/arm/neon.md
> > > +++ b/gcc/config/arm/neon.md
> > > @@ -5924,43 +5924,6 @@ (define_insn "neon_vec_unpack<US>_hi_<mode>"
> > > [(set_attr "type" "neon_shift_imm_long")]
> > > )
> > >
> > > -(define_expand "vec_unpack<US>_hi_<mode>"
> > > - [(match_operand:<V_unpack> 0 "register_operand")
> > > - (SE:<V_unpack> (match_operand:VU 1 "register_operand"))]
> > > - "TARGET_NEON && !BYTES_BIG_ENDIAN"
> > > - {
> > > - rtvec v = rtvec_alloc (<V_mode_nunits>/2) ;
> > > - rtx t1;
> > > - int i;
> > > - for (i = 0; i < (<V_mode_nunits>/2); i++)
> > > - RTVEC_ELT (v, i) = GEN_INT ((<V_mode_nunits>/2) + i);
> > > -
> > > - t1 = gen_rtx_PARALLEL (<MODE>mode, v);
> > > - emit_insn (gen_neon_vec_unpack<US>_hi_<mode> (operands[0],
> > > - operands[1],
> > > - t1));
> > > - DONE;
> > > - }
> > > -)
> > > -
> > > -(define_expand "vec_unpack<US>_lo_<mode>"
> > > - [(match_operand:<V_unpack> 0 "register_operand")
> > > - (SE:<V_unpack> (match_operand:VU 1 "register_operand"))]
> > > - "TARGET_NEON && !BYTES_BIG_ENDIAN"
> > > - {
> > > - rtvec v = rtvec_alloc (<V_mode_nunits>/2) ;
> > > - rtx t1;
> > > - int i;
> > > - for (i = 0; i < (<V_mode_nunits>/2) ; i++)
> > > - RTVEC_ELT (v, i) = GEN_INT (i);
> > > - t1 = gen_rtx_PARALLEL (<MODE>mode, v);
> > > - emit_insn (gen_neon_vec_unpack<US>_lo_<mode> (operands[0],
> > > - operands[1],
> > > - t1));
> > > - DONE;
> > > - }
> > > -)
> > > -
> > > (define_insn "neon_vec_<US>mult_lo_<mode>"
> > > [(set (match_operand:<V_unpack> 0 "register_operand" "=w")
> > > (mult:<V_unpack> (SE:<V_unpack> (vec_select:<V_HALF>
> > > @@ -6176,7 +6139,7 @@ (define_expand "vec_widen_<US>shiftl_lo_<mode>"
> > > ; because the ordering of vector elements in Q registers is different from what
> > > ; the semantics of the instructions require.
> > >
> > > -(define_insn "vec_pack_trunc_<mode>"
> > > +(define_insn "neon_quad_vec_pack_trunc_<mode>"
> > > [(set (match_operand:<V_narrow_pack> 0 "register_operand" "=&w")
> > > (vec_concat:<V_narrow_pack>
> > > (truncate:<V_narrow>
> > > diff --git a/gcc/config/arm/vec-common.md b/gcc/config/arm/vec-common.md
> > > index 1ba1e5eb008..0ffc7a9322c 100644
> > > --- a/gcc/config/arm/vec-common.md
> > > +++ b/gcc/config/arm/vec-common.md
> > > @@ -638,3 +638,92 @@ (define_expand "clz<mode>2"
> > > emit_insn (gen_mve_vclzq_s (<MODE>mode, operands[0], operands[1]));
> > > DONE;
> > > })
> > > +
> > > +;; vmovl[tb] are not available for V4SI on MVE
> > > +(define_expand "vec_unpack<US>_hi_<mode>"
> > > + [(match_operand:<V_unpack> 0 "register_operand")
> > > + (SE:<V_unpack> (match_operand:VU 1 "register_operand"))]
> > > + "ARM_HAVE_<MODE>_ARITH
> > > + && !TARGET_REALLY_IWMMXT
> > > + && ! (<MODE>mode == V4SImode && TARGET_HAVE_MVE)
> > > + && !BYTES_BIG_ENDIAN"
> > > + {
> > > + if (TARGET_NEON)
> > > + {
> > > + rtvec v = rtvec_alloc (<V_mode_nunits>/2);
> > > + rtx t1;
> > > + int i;
> > > + for (i = 0; i < (<V_mode_nunits>/2); i++)
> > > + RTVEC_ELT (v, i) = GEN_INT ((<V_mode_nunits>/2) + i);
> > > +
> > > + t1 = gen_rtx_PARALLEL (<MODE>mode, v);
> > > + emit_insn (gen_neon_vec_unpack<US>_hi_<mode> (operands[0],
> > > + operands[1],
> > > + t1));
> > > + }
> > > + else
> > > + {
> > > + emit_insn (gen_mve_vmovltq (VMOVLTQ_S, <MODE>mode, operands[0],
> > > + operands[1]));
> >
> > It would be good to use the same rtx pattern for the MVE instructions,
> > since the compiler can optimise it better than an unspec.
> >
> > It would be good to have separate tests for the packs and unpacks
> > as well, in case the vclz thing is fixed in future.
> >
And while writing new tests, I noticed that the same problem happens
with unpack as I proposed it:
test_pack_s32:
mov r3, r1
vldr.64 d6, .L3
vldr.64 d7, .L3+8
vldrw.32 q1, [r3]
adds r1, r1, #16
vldrw.32 q2, [r1]
vmovnb.i32 q3, q1
vmovnt.i32 q3, q2
vstrh.16 q3, [r0]
bx lr
.L4:
.align 3
.L3:
.short 0
.short 0
.short 0
.short 0
.short 0
.short 0
.short 0
.short 0
We shouldn't need to initialize d6/d7 (q3)
> > Thanks,
> > Richard
> >
> > > + }
> > > + DONE;
> > > + }
> > > +)
> > > +
> > > +;; vmovl[tb] are not available for V4SI on MVE
> > > +(define_expand "vec_unpack<US>_lo_<mode>"
> > > + [(match_operand:<V_unpack> 0 "register_operand")
> > > + (SE:<V_unpack> (match_operand:VU 1 "register_operand"))]
> > > + "ARM_HAVE_<MODE>_ARITH
> > > + && !TARGET_REALLY_IWMMXT
> > > + && ! (<MODE>mode == V4SImode && TARGET_HAVE_MVE)
> > > + && !BYTES_BIG_ENDIAN"
> > > + {
> > > + if (TARGET_NEON)
> > > + {
> > > + rtvec v = rtvec_alloc (<V_mode_nunits>/2);
> > > + rtx t1;
> > > + int i;
> > > + for (i = 0; i < (<V_mode_nunits>/2) ; i++)
> > > + RTVEC_ELT (v, i) = GEN_INT (i);
> > > +
> > > + t1 = gen_rtx_PARALLEL (<MODE>mode, v);
> > > + emit_insn (gen_neon_vec_unpack<US>_lo_<mode> (operands[0],
> > > + operands[1],
> > > + t1));
> > > + }
> > > + else
> > > + {
> > > + emit_insn (gen_mve_vmovlbq (VMOVLBQ_S, <MODE>mode, operands[0],
> > > + operands[1]));
> > > + }
> > > + DONE;
> > > + }
> > > +)
> > > +
> > > +;; vmovn[tb] are not available for V2DI on MVE
> > > +(define_expand "vec_pack_trunc_<mode>"
> > > + [(set (match_operand:<V_narrow_pack> 0 "register_operand" "=&w")
> > > + (vec_concat:<V_narrow_pack>
> > > + (truncate:<V_narrow>
> > > + (match_operand:VN 1 "register_operand" "w"))
> > > + (truncate:<V_narrow>
> > > + (match_operand:VN 2 "register_operand" "w"))))]
> > > + "ARM_HAVE_<MODE>_ARITH
> > > + && !TARGET_REALLY_IWMMXT
> > > + && ! (<MODE>mode == V2DImode && TARGET_HAVE_MVE)
> > > + && !BYTES_BIG_ENDIAN"
> > > + {
> > > + if (TARGET_NEON)
> > > + {
> > > + emit_insn (gen_neon_quad_vec_pack_trunc_<mode> (operands[0], operands[1],
> > > + operands[2]));
> > > + }
> > > + else
> > > + {
> > > + emit_insn (gen_mve_vec_pack_trunc (<MODE>mode, operands[0], operands[1],
> > > + operands[2]));
> > > + }
> > > + DONE;
> > > + }
> > > +)
> > > diff --git a/gcc/testsuite/gcc.target/arm/simd/mve-vclz.c b/gcc/testsuite/gcc.target/arm/simd/mve-vclz.c
> > > index 7068736bc28..5d6e991cfc6 100644
> > > --- a/gcc/testsuite/gcc.target/arm/simd/mve-vclz.c
> > > +++ b/gcc/testsuite/gcc.target/arm/simd/mve-vclz.c
> > > @@ -21,8 +21,9 @@ FUNC(u, uint, 16, clz)
> > > FUNC(s, int, 8, clz)
> > > FUNC(u, uint, 8, clz)
> > >
> > > -/* 16 and 8-bit versions are not vectorized because they need pack/unpack
> > > - patterns since __builtin_clz uses 32-bit parameter and return value. */
> > > -/* { dg-final { scan-assembler-times {vclz\.i32 q[0-9]+, q[0-9]+} 2 } } */
> > > +/* 16 and 8-bit versions still use 32-bit intermediate temporaries, so for
> > > + instance instead of using vclz.i8, we need 4 vclz.i32, leading to a total of
> > > + 14 vclz.i32 expected in this testcase. */
> > > +/* { dg-final { scan-assembler-times {vclz\.i32 q[0-9]+, q[0-9]+} 14 } } */
> > > /* { dg-final { scan-assembler-times {vclz\.i16 q[0-9]+, q[0-9]+} 2 { xfail *-*-* } } } */
> > > /* { dg-final { scan-assembler-times {vclz\.i8 q[0-9]+, q[0-9]+} 2 { xfail *-*-* } } } */
> > > diff --git a/gcc/testsuite/gcc.target/arm/simd/mve-vshl.c b/gcc/testsuite/gcc.target/arm/simd/mve-vshl.c
> > > index 7a0644997c8..91dd942d818 100644
> > > --- a/gcc/testsuite/gcc.target/arm/simd/mve-vshl.c
> > > +++ b/gcc/testsuite/gcc.target/arm/simd/mve-vshl.c
> > > @@ -56,7 +56,10 @@ FUNC_IMM(u, uint, 8, 16, <<, vshlimm)
> > > /* MVE has only 128-bit vectors, so we can vectorize only half of the
> > > functions above. */
> > > /* We only emit vshl.u, which is equivalent to vshl.s anyway. */
> > > -/* { dg-final { scan-assembler-times {vshl.u[0-9]+\tq[0-9]+, q[0-9]+} 2 } } */
> > > +/* 16 and 8-bit versions still use 32-bit intermediate temporaries, so for
> > > + instance instead of using vshl.u8, we need 4 vshl.i32, leading to a total of
> > > + 14 vshl.i32 expected in this testcase. */
> > > +/* { dg-final { scan-assembler-times {vshl.u[0-9]+\tq[0-9]+, q[0-9]+} 14 } } */
> > >
> > > /* We emit vshl.i when the shift amount is an immediate. */
> > > /* { dg-final { scan-assembler-times {vshl.i[0-9]+\tq[0-9]+, q[0-9]+} 6 } } */
More information about the Gcc-patches
mailing list