target/arm: Implement SVE2 FMMLA
PATCH v2 1/1 target/arm: Implement SVE2 FMMLA
Signed-off-by Stephen Long <[email protected]>
I'm guessing endianness doesn't matter because we are writing to the corresponding 32-bit/64-bit in the destination register.
--- target/arm/cpu.h | 10 +++++++++ target/arm/helper-sve.h | 3 +++ target/arm/sve.decode | 4 ++++ target/arm/sve_helper.c | 44 ++++++++++++++++++++++++++++++++++++++ target/arm/translate-sve.c | 29 +++++++++++++++++++++++++ 5 files changed, 90 insertions(+)
target/arm/cpu.h +10 -0
@@ -3870,6 +3870,16 @@ static inline bool isar_feature_aa64_sve2_bitperm(const ARMISARegisters *id)return FIELD_EX64 (id->id_aa64zfr0,ID_AA64ZFR0 ,BITPERM ) !=0 ; } +static inline bool isar_feature_aa64_sve2_f32mm (const ARMISARegisters *id) +{ +return FIELD_EX64 (id->id_aa64zfr0,ID_AA64ZFR0 ,F32MM ) !=0 ; +} + +static inline bool isar_feature_aa64_sve2_f64mm (const ARMISARegisters *id) +{ +return FIELD_EX64 (id->id_aa64zfr0,ID_AA64ZFR0 ,F64MM ) !=0 ; +} +/* * Feature tests for "does this exist in either 32-bit or 64-bit?" */
target/arm/helper-sve.h +3 -0
@@ -2683,3 +2683,6 @@ DEF_HELPER_FLAGS_5(sve2_sqrdcmlah_zzzz_s, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32) DEF_HELPER_FLAGS_5(sve2_sqrdcmlah_zzzz_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32) + +DEF_HELPER_FLAGS_6(fmmla_s, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_6(fmmla_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, ptr, i32)
target/arm/sve.decode +4 -0
@@ -1383,3 +1383,7 @@ UMLSLT_zzzw 01000100 .. 0 ..... 010 111 ..... ..... @rda_rn_rm CMLA_zzzz 01000100 esz:2 0 rm:5 0010 rot:2 rn:5 rd:5 ra=%reg_movprfx SQRDCMLAH_zzzz 01000100 esz:2 0 rm:5 0011 rot:2 rn:5 rd:5 ra=%reg_movprfx + +### SVE2 floating point matrix multiply accumulate + +FMMLA 01100100 .. 1 ..... 111001 ..... ..... @rda_rn_rm
target/arm/sve_helper.c +44 -0
47 lines not on this page · raw message
target/arm/translate-sve.c +29 -0
@@ -7615,6 +7615,35 @@ static bool do_sve2_zzzz_fn(DisasContext *s, int rd, int rn, int rm, int ra,return true; } +static bool trans_FMMLA (DisasContext *s,arg_rrrr_esz *a) +{ +if (a->esz <MO_32 ) { +return false; + } + +if (a->esz ==MO_32 && !dc_isar_feature (aa64_sve2_f32mm, s)) { +return false; + } + +if (a->esz ==MO_64 && !dc_isar_feature (aa64_sve2_f64mm, s)) { +return false; + } + +static gen_helper_gvec_4_ptr *const fns[2 ] = { + gen_helper_fmmla_s, gen_helper_fmmla_d + }; +if (sve_access_check (s)) { +unsigned vsz =vec_full_reg_size (s); +TCGv_ptr status =get_fpstatus_ptr (a->esz ==MO_16 ); +tcg_gen_gvec_4_ptr (vec_full_reg_offset (s, a->rd), +vec_full_reg_offset (s, a->ra), +vec_full_reg_offset (s, a->rn), +vec_full_reg_offset (s, a->rm), + status, vsz, vsz,0 , fns[a->esz -2 ]); + } +return true; +} +static bool do_sqdmlal_zzzw (DisasContext *s,arg_rrrr_esz *a,bool sel1,bool sel2) {
Signature · 1 line
2.17.1
Quoted 11 lines · Stephen Long
Signed-off-by Stephen Long <[email protected]>I'm guessing endianness doesn't matter because we are writing to the corresponding 32-bit/64-bit in the destination register.
--- target/arm/cpu.h | 10 +++++++++ target/arm/helper-sve.h | 3 +++ target/arm/sve.decode | 4 ++++ target/arm/sve_helper.c | 44 ++++++++++++++++++++++++++++++++++++++ target/arm/translate-sve.c | 29 +++++++++++++++++++++++++ 5 files changed, 90 insertions(+)
Endianness does matter for 32-bit, as we are writing into a host-endian 64-bit quantity. I was being over-brief in my earlier reply.
Quoted 13 lines
target/arm/sve_helper.c +13
-0+TYPE p0, p1, results[4 ]; \ + \ +/* i = 0, j = 0 */ \ + p0 =MUL (n00, m00, status); \ + p1 =MUL (n01, m01, status); \ + results[0 ] =ADD (a[0 ],ADD (p0, p1, status), status); \ + \ +/* i = 0, j = 1 */ \ + p0 =MUL (n00, m10, status); \ + p1 =MUL (n01, m11, status); \ + results[1 ] =ADD (a[1 ],ADD (p0, p1, status), status); \ + \ +/* i = 1, j = 0 */ \
target/arm/sve_helper.c void HELPER(sve2_histseg)(void *vd, void *vn, void *vm, uint32_t desc) · line 7430
Quoted 10 lines
target/arm/sve_helper.c +10
-0+ p0 =MUL (n10, m00, status); \ + p1 =MUL (n11, m01, status); \ + results[2 ] =ADD (a[2 ],ADD (p0, p1, status), status); \ + \ +/* i = 1, j = 1 */ \ + p0 =MUL (n10, m10, status); \ + p1 =MUL (n11, m11, status); \ + results[3 ] =ADD (a[3 ],ADD (p0, p1, status), status); \ + \ +memcpy (d, results,sizeof (TYPE ) *4 ); \
There's no need for the result array -- we have already read the inputs, so we can write back the result straight away.
r~