Hi Tamar,
> On 31 Jul 2024, at 18:46, Tamar Christina <[email protected]> wrote:
>
> External email: Use caution opening links or attachments
>
>
> Hi Kyrill,
>
>>> /* True if the vector body contains a store to a decl and if the
>>> function is known to have a vld1 from the same decl.
>>>
>>> @@ -17291,6 +17297,17 @@ aarch64_vector_costs::add_stmt_cost (int count,
>> vect_cost_for_stmt kind,
>>> stmt_cost = aarch64_detect_vector_stmt_subtype (m_vinfo, kind,
>>> stmt_info, vectype,
>>> where, stmt_cost);
>>> +
>>> + /* Check if we've seen an SVE gather/scatter operation and which
>>> size. */
>>> + if (kind == scalar_load
>>> + && aarch64_sve_mode_p (TYPE_MODE (vectype))
>>> + && STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) ==
>> VMAT_GATHER_SCATTER)
>>> + {
>>> + if (GET_MODE_UNIT_BITSIZE (TYPE_MODE (vectype)) == 64)
>>> + m_sve_gather_scatter_x64 = true;
>>> + else
>>> + m_sve_gather_scatter_x32 = true;
>>
>> This is a bit academic at this stage but SVE2.1 adds quadword gather loads.
>> I know
>> we’re not vectoring for those yet, but maybe it’s worth explicitly checking
>> for 32-
>> bit size and gcc_unreachable () otherwise?
>
> To be honest I'm not quite sure how to detect it. If it just
> GET_MODE_UNIT_BITSIZE () == 128?
> But do we want an assert in the cost model? Happy to do so though but maybe a
> debug print is more
> appropriate? i.e. make it a missed optimization?
I don’t feel strongly about it. I’d drop my comment here. If and when we add
vectorization for 128-bit chunks I’m sure we’ll be revisiting this cost model
anyway.
>
>>
>>
>>> + }
>>> }
>>>
>>> /* Do any SVE-specific adjustments to the cost. */
>>> @@ -17676,6 +17693,18 @@ aarch64_vector_costs::finish_cost (const
>> vector_costs *uncast_scalar_costs)
>>> m_costs[vect_body] = adjust_body_cost (loop_vinfo, scalar_costs,
>>> m_costs[vect_body]);
>>> m_suggested_unroll_factor = determine_suggested_unroll_factor ();
>>> +
>>> + /* For gather and scatters there's an additional overhead for the
>>> first
>>> + iteration. For low count loops they're not beneficial so model the
>>> + overhead as loop prologue costs. */
>>> + if (m_sve_gather_scatter_x32 || m_sve_gather_scatter_x64)
>>> + {
>>> + const sve_vec_cost *sve_costs =
>>> aarch64_tune_params.vec_costs->sve;
>>> + if (m_sve_gather_scatter_x32)
>>> + m_costs[vect_prologue] += sve_costs->gather_load_x32_init_cost;
>>> + else
>>> + m_costs[vect_prologue] += sve_costs->gather_load_x64_init_cost;
>>
>> Shouldn’t this not be en else but rather:
>> If (m_sve_gather_scatter_x64)
>> m_costs[vect_prologue] += sve_costs->gather_load_x64_init_cost;
>>
>> In case the loop has both 32-bit and 64-bit gather/scatter?
>>
>
> This was an interesting comment. After some discussion and more benchmarking
> we've changed it to be an additive cost.
Ok.
>
>>
>>> + }
>>> }
>>>
>>> /* Apply the heuristic described above m_stp_sequence_cost. Prefer
>>> diff --git a/gcc/config/aarch64/tuning_models/a64fx.h
>> b/gcc/config/aarch64/tuning_models/a64fx.h
>>> index
>> 6091289d4c3c66f01d7e4dbf97a85c1f8c40bb0b..378a1b3889ee265859786c1
>> ff6525fce2305b615 100644
>>> --- a/gcc/config/aarch64/tuning_models/a64fx.h
>>> +++ b/gcc/config/aarch64/tuning_models/a64fx.h
>>> @@ -104,6 +104,8 @@ static const sve_vec_cost a64fx_sve_vector_cost =
>>> 13, /* fadda_f64_cost */
>>> 64, /* gather_load_x32_cost */
>>> 32, /* gather_load_x64_cost */
>>> + 0, /* gather_load_x32_init_cost */
>>> + 0, /* gather_load_x64_init_cost */
>>> 1 /* scatter_store_elt_cost */
>>> };
>>>
>>> diff --git a/gcc/config/aarch64/tuning_models/cortexx925.h
>> b/gcc/config/aarch64/tuning_models/cortexx925.h
>>> index
>> fb95e87526985b02410d54a5a3ec8539c1b0ba6d..c4206018a3ff707f89ff33007
>> 00ec7dc2a5bc6b0 100644
>>> --- a/gcc/config/aarch64/tuning_models/cortexx925.h
>>> +++ b/gcc/config/aarch64/tuning_models/cortexx925.h
>>> @@ -135,6 +135,8 @@ static const sve_vec_cost cortexx925_sve_vector_cost =
>>> operation more than a 64-bit gather. */
>>> 14, /* gather_load_x32_cost */
>>> 12, /* gather_load_x64_cost */
>>> + 42, /* gather_load_x32_init_cost */
>>> + 24, /* gather_load_x64_init_cost */
>>
>>
>> Can you comment on how these numbers are derived?
>
> They were derived essentially from benchmarking. I did a bunch of runs over
> various cores
> to determine at which iteration count they become profitable. From that as
> you can
> probably tell the costs are a multiple of the cost of the operations for the
> specific core.
>
> This because that cost already keeps in mind things like VL differences.
>
> Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
>
> Ok for master?
Ok with Richard’s comments addressed.
Thanks,
Kyrill
>
> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
> * config/aarch64/aarch64-protos.h (struct sve_vec_cost): Add
> gather_load_x32_init_cost and gather_load_x64_init_cost.
> * config/aarch64/aarch64.cc (aarch64_vector_costs): Add
> m_sve_gather_scatter_init_cost.
> (aarch64_vector_costs::add_stmt_cost): Use them.
> (aarch64_vector_costs::finish_cost): Likewise.
> * config/aarch64/tuning_models/a64fx.h: Update.
> * config/aarch64/tuning_models/cortexx925.h: Update.
> * config/aarch64/tuning_models/generic.h: Update.
> * config/aarch64/tuning_models/generic_armv8_a.h: Update.
> * config/aarch64/tuning_models/generic_armv9_a.h: Update.
> * config/aarch64/tuning_models/neoverse512tvb.h: Update.
> * config/aarch64/tuning_models/neoversen2.h: Update.
> * config/aarch64/tuning_models/neoversen3.h: Update.
> * config/aarch64/tuning_models/neoversev1.h: Update.
> * config/aarch64/tuning_models/neoversev2.h: Update.
> * config/aarch64/tuning_models/neoversev3.h: Update.
> * config/aarch64/tuning_models/neoversev3ae.h: Update.
>
> -- inline copy of patch --
>
> diff --git a/gcc/config/aarch64/aarch64-protos.h
> b/gcc/config/aarch64/aarch64-protos.h
> index
> 42639e9efcf1e0f9362f759ae63a31b8eeb0d581..16eb8edab4d9fdfc6e3672c56ef5c9f6962d0c0b
> 100644
> --- a/gcc/config/aarch64/aarch64-protos.h
> +++ b/gcc/config/aarch64/aarch64-protos.h
> @@ -262,6 +262,8 @@ struct sve_vec_cost : simd_vec_cost
> unsigned int fadda_f64_cost,
> unsigned int gather_load_x32_cost,
> unsigned int gather_load_x64_cost,
> + unsigned int gather_load_x32_init_cost,
> + unsigned int gather_load_x64_init_cost,
> unsigned int scatter_store_elt_cost)
> : simd_vec_cost (base),
> clast_cost (clast_cost),
> @@ -270,6 +272,8 @@ struct sve_vec_cost : simd_vec_cost
> fadda_f64_cost (fadda_f64_cost),
> gather_load_x32_cost (gather_load_x32_cost),
> gather_load_x64_cost (gather_load_x64_cost),
> + gather_load_x32_init_cost (gather_load_x32_init_cost),
> + gather_load_x64_init_cost (gather_load_x64_init_cost),
> scatter_store_elt_cost (scatter_store_elt_cost)
> {}
>
> @@ -289,6 +293,12 @@ struct sve_vec_cost : simd_vec_cost
> const int gather_load_x32_cost;
> const int gather_load_x64_cost;
>
> + /* Additional loop initialization cost of using a gather load instruction.
> The x32
> + value is for loads of 32-bit elements and the x64 value is for loads of
> + 64-bit elements. */
> + const int gather_load_x32_init_cost;
> + const int gather_load_x64_init_cost;
> +
> /* The per-element cost of a scatter store. */
> const int scatter_store_elt_cost;
> };
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index
> eafa377cb095f49408d8a926fb49ce13e2155ba2..da2feb54ddad9b39db92e0a9ec7c4e40cfa3e4e2
> 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -16227,6 +16227,10 @@ private:
> supported by Advanced SIMD and SVE2. */
> bool m_has_avg = false;
>
> + /* Additional initialization costs for using gather or scatter operation in
> + the current loop. */
> + unsigned int m_sve_gather_scatter_init_cost = 0;
> +
> /* True if the vector body contains a store to a decl and if the
> function is known to have a vld1 from the same decl.
>
> @@ -17291,6 +17295,20 @@ aarch64_vector_costs::add_stmt_cost (int count,
> vect_cost_for_stmt kind,
> stmt_cost = aarch64_detect_vector_stmt_subtype (m_vinfo, kind,
> stmt_info, vectype,
> where, stmt_cost);
> +
> + /* Check if we've seen an SVE gather/scatter operation and which size.
> */
> + if (kind == scalar_load
> + && aarch64_sve_mode_p (TYPE_MODE (vectype))
> + && STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) == VMAT_GATHER_SCATTER)
> + {
> + const sve_vec_cost *sve_costs = aarch64_tune_params.vec_costs->sve;
> + if (GET_MODE_UNIT_BITSIZE (TYPE_MODE (vectype)) == 64)
> + m_sve_gather_scatter_init_cost
> + += sve_costs->gather_load_x64_init_cost;
> + else
> + m_sve_gather_scatter_init_cost
> + += sve_costs->gather_load_x32_init_cost;
> + }
> }
>
> /* Do any SVE-specific adjustments to the cost. */
> @@ -17676,6 +17694,12 @@ aarch64_vector_costs::finish_cost (const
> vector_costs *uncast_scalar_costs)
> m_costs[vect_body] = adjust_body_cost (loop_vinfo, scalar_costs,
> m_costs[vect_body]);
> m_suggested_unroll_factor = determine_suggested_unroll_factor ();
> +
> + /* For gather and scatters there's an additional overhead for the first
> + iteration. For low count loops they're not beneficial so model the
> + overhead as loop prologue costs. */
> + if (m_sve_gather_scatter_init_cost)
> + m_costs[vect_prologue] += m_sve_gather_scatter_init_cost;
> }
>
> /* Apply the heuristic described above m_stp_sequence_cost. Prefer
> diff --git a/gcc/config/aarch64/tuning_models/a64fx.h
> b/gcc/config/aarch64/tuning_models/a64fx.h
> index
> 6091289d4c3c66f01d7e4dbf97a85c1f8c40bb0b..378a1b3889ee265859786c1ff6525fce2305b615
> 100644
> --- a/gcc/config/aarch64/tuning_models/a64fx.h
> +++ b/gcc/config/aarch64/tuning_models/a64fx.h
> @@ -104,6 +104,8 @@ static const sve_vec_cost a64fx_sve_vector_cost =
> 13, /* fadda_f64_cost */
> 64, /* gather_load_x32_cost */
> 32, /* gather_load_x64_cost */
> + 0, /* gather_load_x32_init_cost */
> + 0, /* gather_load_x64_init_cost */
> 1 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/cortexx925.h
> b/gcc/config/aarch64/tuning_models/cortexx925.h
> index
> 6cae5b7de5ca7ffad8a0f683e1285039bb55d159..b509cae758419a415d9067ec751ef1e6528eb09a
> 100644
> --- a/gcc/config/aarch64/tuning_models/cortexx925.h
> +++ b/gcc/config/aarch64/tuning_models/cortexx925.h
> @@ -135,6 +135,8 @@ static const sve_vec_cost cortexx925_sve_vector_cost =
> operation more than a 64-bit gather. */
> 14, /* gather_load_x32_cost */
> 12, /* gather_load_x64_cost */
> + 42, /* gather_load_x32_init_cost */
> + 24, /* gather_load_x64_init_cost */
> 1 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/generic.h
> b/gcc/config/aarch64/tuning_models/generic.h
> index
> 2b1f68b3052117814161a32f426422736ad6462b..101969bdbb9ccf7eafbd9a1cd6e25f0b584fb261
> 100644
> --- a/gcc/config/aarch64/tuning_models/generic.h
> +++ b/gcc/config/aarch64/tuning_models/generic.h
> @@ -105,6 +105,8 @@ static const sve_vec_cost generic_sve_vector_cost =
> 2, /* fadda_f64_cost */
> 4, /* gather_load_x32_cost */
> 2, /* gather_load_x64_cost */
> + 12, /* gather_load_x32_init_cost */
> + 4, /* gather_load_x64_init_cost */
> 1 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/generic_armv8_a.h
> b/gcc/config/aarch64/tuning_models/generic_armv8_a.h
> index
> b38b9a8c5cad7d12aa38afdb610a14a25e755010..b5088afe068aa4be7f9dd614cfdd2a51fa96e524
> 100644
> --- a/gcc/config/aarch64/tuning_models/generic_armv8_a.h
> +++ b/gcc/config/aarch64/tuning_models/generic_armv8_a.h
> @@ -106,6 +106,8 @@ static const sve_vec_cost generic_armv8_a_sve_vector_cost
> =
> 2, /* fadda_f64_cost */
> 4, /* gather_load_x32_cost */
> 2, /* gather_load_x64_cost */
> + 12, /* gather_load_x32_init_cost */
> + 4, /* gather_load_x64_init_cost */
> 1 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/generic_armv9_a.h
> b/gcc/config/aarch64/tuning_models/generic_armv9_a.h
> index
> 7156dbe5787e831bc4343deb7d7b88e9823fc1bc..999985ed40f694f2681779d940bdb282f289b8e3
> 100644
> --- a/gcc/config/aarch64/tuning_models/generic_armv9_a.h
> +++ b/gcc/config/aarch64/tuning_models/generic_armv9_a.h
> @@ -136,6 +136,8 @@ static const sve_vec_cost generic_armv9_a_sve_vector_cost
> =
> operation more than a 64-bit gather. */
> 14, /* gather_load_x32_cost */
> 12, /* gather_load_x64_cost */
> + 42, /* gather_load_x32_init_cost */
> + 24, /* gather_load_x64_init_cost */
> 3 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/neoverse512tvb.h
> b/gcc/config/aarch64/tuning_models/neoverse512tvb.h
> index
> 825c6a64990b72cda3641737957dc94d75db1509..d2a0b647791de8fca6d7684849d2ab1e9104b045
> 100644
> --- a/gcc/config/aarch64/tuning_models/neoverse512tvb.h
> +++ b/gcc/config/aarch64/tuning_models/neoverse512tvb.h
> @@ -79,6 +79,8 @@ static const sve_vec_cost neoverse512tvb_sve_vector_cost =
> operation more than a 64-bit gather. */
> 14, /* gather_load_x32_cost */
> 12, /* gather_load_x64_cost */
> + 42, /* gather_load_x32_init_cost */
> + 24, /* gather_load_x64_init_cost */
> 3 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/neoversen2.h
> b/gcc/config/aarch64/tuning_models/neoversen2.h
> index
> d41e714aa045266ecae62a36ed02dfbfb7597c3a..1a5b66901b5c3fb78f87fee40236957139644585
> 100644
> --- a/gcc/config/aarch64/tuning_models/neoversen2.h
> +++ b/gcc/config/aarch64/tuning_models/neoversen2.h
> @@ -135,6 +135,8 @@ static const sve_vec_cost neoversen2_sve_vector_cost =
> operation more than a 64-bit gather. */
> 14, /* gather_load_x32_cost */
> 12, /* gather_load_x64_cost */
> + 42, /* gather_load_x32_init_cost */
> + 24, /* gather_load_x64_init_cost */
> 3 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/neoversen3.h
> b/gcc/config/aarch64/tuning_models/neoversen3.h
> index
> 36f770c0a14fc127c75a60cd37048d46c3b069c7..cfd5060e6b64a0433de41b03cde886da119d9a1c
> 100644
> --- a/gcc/config/aarch64/tuning_models/neoversen3.h
> +++ b/gcc/config/aarch64/tuning_models/neoversen3.h
> @@ -135,6 +135,8 @@ static const sve_vec_cost neoversen3_sve_vector_cost =
> operation more than a 64-bit gather. */
> 14, /* gather_load_x32_cost */
> 12, /* gather_load_x64_cost */
> + 42, /* gather_load_x32_init_cost */
> + 24, /* gather_load_x64_init_cost */
> 1 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/neoversev1.h
> b/gcc/config/aarch64/tuning_models/neoversev1.h
> index
> 0fc41ce6a41b3135fa06d2bda1f517fdf4f8dbcf..705ed025730f6683109a4796c6eefa55b437cec9
> 100644
> --- a/gcc/config/aarch64/tuning_models/neoversev1.h
> +++ b/gcc/config/aarch64/tuning_models/neoversev1.h
> @@ -126,6 +126,8 @@ static const sve_vec_cost neoversev1_sve_vector_cost =
> 8, /* fadda_f64_cost */
> 32, /* gather_load_x32_cost */
> 16, /* gather_load_x64_cost */
> + 96, /* gather_load_x32_init_cost */
> + 32, /* gather_load_x64_init_cost */
> 3 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/neoversev2.h
> b/gcc/config/aarch64/tuning_models/neoversev2.h
> index
> c9c3019dd01a98bc20a76e8455fb59ff24a9ff6c..47908636b0f4c3eadd5848b590fd079c1c04aa10
> 100644
> --- a/gcc/config/aarch64/tuning_models/neoversev2.h
> +++ b/gcc/config/aarch64/tuning_models/neoversev2.h
> @@ -135,6 +135,8 @@ static const sve_vec_cost neoversev2_sve_vector_cost =
> operation more than a 64-bit gather. */
> 14, /* gather_load_x32_cost */
> 12, /* gather_load_x64_cost */
> + 42, /* gather_load_x32_init_cost */
> + 24, /* gather_load_x64_init_cost */
> 3 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/neoversev3.h
> b/gcc/config/aarch64/tuning_models/neoversev3.h
> index
> c602d067c7116cf6b081caeae8d36f9969e06d8d..c91e8c829532f9236de0102770e5c6b94e83da9a
> 100644
> --- a/gcc/config/aarch64/tuning_models/neoversev3.h
> +++ b/gcc/config/aarch64/tuning_models/neoversev3.h
> @@ -135,6 +135,8 @@ static const sve_vec_cost neoversev3_sve_vector_cost =
> operation more than a 64-bit gather. */
> 14, /* gather_load_x32_cost */
> 12, /* gather_load_x64_cost */
> + 42, /* gather_load_x32_init_cost */
> + 24, /* gather_load_x64_init_cost */
> 1 /* scatter_store_elt_cost */
> };
>
> diff --git a/gcc/config/aarch64/tuning_models/neoversev3ae.h
> b/gcc/config/aarch64/tuning_models/neoversev3ae.h
> index
> 96d7ccf03cd96056d09676d908c63a25e3da6765..61e439326eb6f983abf8574e657cfbb0c2f9bb33
> 100644
> --- a/gcc/config/aarch64/tuning_models/neoversev3ae.h
> +++ b/gcc/config/aarch64/tuning_models/neoversev3ae.h
> @@ -135,6 +135,8 @@ static const sve_vec_cost neoversev3ae_sve_vector_cost =
> operation more than a 64-bit gather. */
> 14, /* gather_load_x32_cost */
> 12, /* gather_load_x64_cost */
> + 42, /* gather_load_x32_init_cost */
> + 24, /* gather_load_x64_init_cost */
> 1 /* scatter_store_elt_cost */
> };
>
> <rb18671.patch>
