[gcc(refs/vendors/ibm/heads/dmr-mainline)] rs6000: Builtins for MMA+ float16 outer product instructions
Surya Kumari Jangala via Gcc-cvs <[email protected]>
| Newsgroups | gmane.comp.gcc.cvs |
|---|---|
| Message-ID | <[email protected]> |
https://gcc.gnu.org/g:0a643f03c6ea131c417290344f455185100e497a commit 0a643f03c6ea131c417290344f455185100e497a Author: Surya Kumari Jangala <[email protected]> Date: Fri Jul 17 00:14:15 2026 -0500 rs6000: Builtins for MMA+ float16 outer product instructions This patch adds builtins for the Matrix Multiply Assist Plus (MMA+) float16 GER instructions. These instructions may or may not be supported in a future Power processor. Specifically, builtins have been added for the following instructions: Float16 GER operations: - dmxvf16gerx2 and its variants (pp, pn, np, nn) - pmdmxvf16gerx2 and its variants (pp, pn, np, nn) Note, the names of the builtins may change in the future. 2026-06-25 Surya Kumari Jangala <[email protected]> gcc: * config/rs6000/mma.md (UNSPEC_DMF_DMXVF16GERX2): New UNSPEC entry. (UNSPEC_DMF_DMXVF16GERX2PP): Likewise. (UNSPEC_DMF_DMXVF16GERX2PN): Likewise. (UNSPEC_DMF_DMXVF16GERX2NP): Likewise. (UNSPEC_DMF_DMXVF16GERX2NN): Likewise. (UNSPEC_DMF_PMDMXVF16GERX2): Likewise. (UNSPEC_DMF_PMDMXVF16GERX2PP): Likewise. (UNSPEC_DMF_PMDMXVF16GERX2PN): Likewise. (UNSPEC_DMF_PMDMXVF16GERX2NP): Likewise. (UNSPEC_DMF_PMDMXVF16GERX2NN): Likewise. (DMF_PV): Add UNSPEC_DMF_DMXVF16GERX2. (DMF_DPV): Add f16 variants. (DMF_PVI8I4I2): Add UNSPEC_DMF_PMDMXVF16GERX2. (DMF_DPVI8I4I2): Add f16 prefixed variants. (pv): Add dmxvf16gerx2 attribute. (dpv): Add f16 attributes. (pvi8i4i2): Add pmdmxvf16gerx2 attribute. (dpvi8i4i2): Add f16 prefixed attributes. * config/rs6000/rs6000-builtins.def (__builtin_mma_dmxvf16gerx2): New builtin. (__builtin_mma_dmxvf16gerx2pp): Likewise. (__builtin_mma_dmxvf16gerx2pn): Likewise. (__builtin_mma_dmxvf16gerx2np): Likewise. (__builtin_mma_dmxvf16gerx2nn): Likewise. (__builtin_mma_pmdmxvf16gerx2): Likewise. (__builtin_mma_pmdmxvf16gerx2pp): Likewise. (__builtin_mma_pmdmxvf16gerx2pn): Likewise. (__builtin_mma_pmdmxvf16gerx2np): Likewise. (__builtin_mma_pmdmxvf16gerx2nn): Likewise. (__builtin_mma_dmxvf16gerx2_internal): New internal builtin. (__builtin_mma_dmxvf16gerx2pp_internal): Likewise. (__builtin_mma_dmxvf16gerx2pn_internal): Likewise. (__builtin_mma_dmxvf16gerx2np_internal): Likewise. (__builtin_mma_dmxvf16gerx2nn_internal): Likewise. (__builtin_mma_pmdmxvf16gerx2_internal): Likewise. (__builtin_mma_pmdmxvf16gerx2pp_internal): Likewise. (__builtin_mma_pmdmxvf16gerx2pn_internal): Likewise. (__builtin_mma_pmdmxvf16gerx2np_internal): Likewise. (__builtin_mma_pmdmxvf16gerx2nn_internal): Likewise. * doc/extend.texi (PowerPC Matrix-Multiply Assist Built-in Functions): Document new MMA+ builtins for f16 operations. gcc/testsuite: * gcc.target/powerpc/dmf-builtin-2.c: New test. Diff: --- gcc/config/rs6000/mma.md | 48 +++++-- gcc/config/rs6000/rs6000-builtins.def | 70 ++++++++++ gcc/testsuite/gcc.target/powerpc/dmf-builtin-2.c | 170 +++++++++++++++++++++++ 3 files changed, 279 insertions(+), 9 deletions(-) diff --git a/gcc/config/rs6000/mma.md b/gcc/config/rs6000/mma.md index 890166642101..5740fb01dad6 100644 --- a/gcc/config/rs6000/mma.md +++ b/gcc/config/rs6000/mma.md @@ -115,6 +115,16 @@ UNSPEC_DMF_PMDMXVBF16GERX2PN UNSPEC_DMF_PMDMXVBF16GERX2NP UNSPEC_DMF_PMDMXVBF16GERX2NN + UNSPEC_DMF_DMXVF16GERX2 + UNSPEC_DMF_DMXVF16GERX2PP + UNSPEC_DMF_DMXVF16GERX2PN + UNSPEC_DMF_DMXVF16GERX2NP + UNSPEC_DMF_DMXVF16GERX2NN + UNSPEC_DMF_PMDMXVF16GERX2 + UNSPEC_DMF_PMDMXVF16GERX2PP + UNSPEC_DMF_PMDMXVF16GERX2PN + UNSPEC_DMF_PMDMXVF16GERX2NP + UNSPEC_DMF_PMDMXVF16GERX2NN ]) (define_c_enum "unspecv" @@ -159,7 +169,8 @@ ; DMF instructions with 1 vector pair and 1 vector arguments (define_int_iterator DMF_PV [UNSPEC_DMF_DMXVI8GERX4 - UNSPEC_DMF_DMXVBF16GERX2]) + UNSPEC_DMF_DMXVBF16GERX2 + UNSPEC_DMF_DMXVF16GERX2]) ;; MMA instructions with 1 accumulator, 1 vector pair and 1 vector arguments (define_int_iterator MMA_APV [UNSPEC_MMA_XVF64GERPP @@ -173,7 +184,11 @@ UNSPEC_DMF_DMXVBF16GERX2PP UNSPEC_DMF_DMXVBF16GERX2PN UNSPEC_DMF_DMXVBF16GERX2NP - UNSPEC_DMF_DMXVBF16GERX2NN]) + UNSPEC_DMF_DMXVBF16GERX2NN + UNSPEC_DMF_DMXVF16GERX2PP + UNSPEC_DMF_DMXVF16GERX2PN + UNSPEC_DMF_DMXVF16GERX2NP + UNSPEC_DMF_DMXVF16GERX2NN]) ;; MMA instructions with 2 vector, 2 4-bit and 1 8-bit arguments (define_int_iterator MMA_VVI4I4I8 [UNSPEC_MMA_PMXVI4GER8]) @@ -235,14 +250,19 @@ ;; DMF instructions with 1 vector pair, 1 vector, 1 8-bit, 1 4-bit ;; and 1 2-bit arguments -(define_int_iterator DMF_PVI8I4I2 [UNSPEC_DMF_PMDMXVBF16GERX2]) +(define_int_iterator DMF_PVI8I4I2 [UNSPEC_DMF_PMDMXVBF16GERX2 + UNSPEC_DMF_PMDMXVF16GERX2]) ;; DMF instructions with 1dmr, 1 vector pair, 1 vector, 1 8-bit, ;; 1 4-bit and 1 2-bit arguments (define_int_iterator DMF_DPVI8I4I2 [UNSPEC_DMF_PMDMXVBF16GERX2PP UNSPEC_DMF_PMDMXVBF16GERX2PN UNSPEC_DMF_PMDMXVBF16GERX2NP - UNSPEC_DMF_PMDMXVBF16GERX2NN]) + UNSPEC_DMF_PMDMXVBF16GERX2NN + UNSPEC_DMF_PMDMXVF16GERX2PP + UNSPEC_DMF_PMDMXVF16GERX2PN + UNSPEC_DMF_PMDMXVF16GERX2NP + UNSPEC_DMF_PMDMXVF16GERX2NN]) (define_int_attr acc [(UNSPEC_MMA_XXMFACC "xxmfacc") (UNSPEC_MMA_XXMTACC "xxmtacc")]) @@ -275,7 +295,8 @@ (define_int_attr pv [(UNSPEC_MMA_XVF64GER "xvf64ger") (UNSPEC_DMF_DMXVI8GERX4 "dmxvi8gerx4") - (UNSPEC_DMF_DMXVBF16GERX2 "dmxvbf16gerx2")]) + (UNSPEC_DMF_DMXVBF16GERX2 "dmxvbf16gerx2") + (UNSPEC_DMF_DMXVF16GERX2 "dmxvf16gerx2")]) (define_int_attr apv [(UNSPEC_MMA_XVF64GERPP "xvf64gerpp") (UNSPEC_MMA_XVF64GERPN "xvf64gerpn") @@ -287,7 +308,11 @@ (UNSPEC_DMF_DMXVBF16GERX2PP "dmxvbf16gerx2pp") (UNSPEC_DMF_DMXVBF16GERX2PN "dmxvbf16gerx2pn") (UNSPEC_DMF_DMXVBF16GERX2NP "dmxvbf16gerx2np") - (UNSPEC_DMF_DMXVBF16GERX2NN "dmxvbf16gerx2nn")]) + (UNSPEC_DMF_DMXVBF16GERX2NN "dmxvbf16gerx2nn") + (UNSPEC_DMF_DMXVF16GERX2PP "dmxvf16gerx2pp") + (UNSPEC_DMF_DMXVF16GERX2PN "dmxvf16gerx2pn") + (UNSPEC_DMF_DMXVF16GERX2NP "dmxvf16gerx2np") + (UNSPEC_DMF_DMXVF16GERX2NN "dmxvf16gerx2nn")]) (define_int_attr vvi4i4i8 [(UNSPEC_MMA_PMXVI4GER8 "pmxvi4ger8")]) @@ -333,12 +358,17 @@ (define_int_attr dpvi8i4i4 [(UNSPEC_DMF_PMDMXVI8GERX4PP "pmdmxvi8gerx4pp") (UNSPEC_DMF_PMDMXVI8GERX4SPP "pmdmxvi8gerx4spp")]) -(define_int_attr pvi8i4i2 [(UNSPEC_DMF_PMDMXVBF16GERX2 "pmdmxvbf16gerx2")]) +(define_int_attr pvi8i4i2 [(UNSPEC_DMF_PMDMXVBF16GERX2 "pmdmxvbf16gerx2") + (UNSPEC_DMF_PMDMXVF16GERX2 "pmdmxvf16gerx2")]) (define_int_attr dpvi8i4i2 [(UNSPEC_DMF_PMDMXVBF16GERX2PP "pmdmxvbf16gerx2pp") (UNSPEC_DMF_PMDMXVBF16GERX2PN "pmdmxvbf16gerx2pn") (UNSPEC_DMF_PMDMXVBF16GERX2NP "pmdmxvbf16gerx2np") - (UNSPEC_DMF_PMDMXVBF16GERX2NN "pmdmxvbf16gerx2nn")]) + (UNSPEC_DMF_PMDMXVBF16GERX2NN "pmdmxvbf16gerx2nn") + (UNSPEC_DMF_PMDMXVF16GERX2PP "pmdmxvf16gerx2pp") + (UNSPEC_DMF_PMDMXVF16GERX2PN "pmdmxvf16gerx2pn") + (UNSPEC_DMF_PMDMXVF16GERX2NP "pmdmxvf16gerx2np") + (UNSPEC_DMF_PMDMXVF16GERX2NN "pmdmxvf16gerx2nn")]) ;; Vector pair support. OOmode can only live in VSRs. (define_expand "movoo" @@ -1083,7 +1113,7 @@ (set_attr "type" "vecload")]) (define_insn "dmf_<pv>" - [(set (match_operand:TDO 0 "accumulator_operand" "=wD") + [(set (match_operand:TDO 0 "dmr_register_operand" "=wD") (unspec:TDO [(match_operand:OO 1 "vsx_register_operand" "wa") (match_operand:V16QI 2 "vsx_register_operand" "wa")] DMF_PV))] diff --git a/gcc/config/rs6000/rs6000-builtins.def b/gcc/config/rs6000/rs6000-builtins.def index 6787b0a9b802..428a01ea2e39 100644 --- a/gcc/config/rs6000/rs6000-builtins.def +++ b/gcc/config/rs6000/rs6000-builtins.def @@ -4243,3 +4243,73 @@ dmr1024 __builtin_mma_pmdmxvbf16gerx2nn_internal (dmr1024, v256, vuc, const int<8>, \ const int<4>, const int<2>); PMDMXVBF16GERX2NN_INTERNAL dmf_pmdmxvbf16gerx2nn {dm,pair} + + void __builtin_mma_dmxvf16gerx2 (dmr1024 *, v256, vuc); + DMXVF16GERX2 nothing {dm,dmint} + + dmr1024 __builtin_mma_dmxvf16gerx2_internal (v256, vuc); + DMXVF16GERX2_INTERNAL dmf_dmxvf16gerx2 {dm} + + void __builtin_mma_dmxvf16gerx2pp (dmr1024 *, v256, vuc); + DMXVF16GERX2PP nothing {dm,dmint,dmr} + + dmr1024 __builtin_mma_dmxvf16gerx2pp_internal (dmr1024, v256, vuc); + DMXVF16GERX2PP_INTERNAL dmf_dmxvf16gerx2pp {dm} + + void __builtin_mma_dmxvf16gerx2pn (dmr1024 *, v256, vuc); + DMXVF16GERX2PN nothing {dm,dmint,dmr} + + dmr1024 __builtin_mma_dmxvf16gerx2pn_internal (dmr1024, v256, vuc); + DMXVF16GERX2PN_INTERNAL dmf_dmxvf16gerx2pn {dm} + + void __builtin_mma_dmxvf16gerx2np (dmr1024 *, v256, vuc); + DMXVF16GERX2NP nothing {dm,dmint,dmr} + + dmr1024 __builtin_mma_dmxvf16gerx2np_internal (dmr1024, v256, vuc); + DMXVF16GERX2NP_INTERNAL dmf_dmxvf16gerx2np {dm} + + void __builtin_mma_dmxvf16gerx2nn (dmr1024 *, v256, vuc); + DMXVF16GERX2NN nothing {dm,dmint,dmr} + + dmr1024 __builtin_mma_dmxvf16gerx2nn_internal (dmr1024, v256, vuc); + DMXVF16GERX2NN_INTERNAL dmf_dmxvf16gerx2nn {dm} + + void __builtin_mma_pmdmxvf16gerx2 (dmr1024 *, v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2 nothing {dm,pair,dmint} + + dmr1024 __builtin_mma_pmdmxvf16gerx2_internal (v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2_INTERNAL dmf_pmdmxvf16gerx2 {dm,pair} + + void __builtin_mma_pmdmxvf16gerx2pp (dmr1024 *, v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2PP nothing {dm,pair,dmint,dmr} + + dmr1024 __builtin_mma_pmdmxvf16gerx2pp_internal (dmr1024, v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2PP_INTERNAL dmf_pmdmxvf16gerx2pp {dm,pair} + + void __builtin_mma_pmdmxvf16gerx2pn (dmr1024 *, v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2PN nothing {dm,pair,dmint,dmr} + + dmr1024 __builtin_mma_pmdmxvf16gerx2pn_internal (dmr1024, v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2PN_INTERNAL dmf_pmdmxvf16gerx2pn {dm,pair} + + void __builtin_mma_pmdmxvf16gerx2np (dmr1024 *, v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2NP nothing {dm,pair,dmint,dmr} + + dmr1024 __builtin_mma_pmdmxvf16gerx2np_internal (dmr1024, v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2NP_INTERNAL dmf_pmdmxvf16gerx2np {dm,pair} + + void __builtin_mma_pmdmxvf16gerx2nn (dmr1024 *, v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2NN nothing {dm,pair,dmint,dmr} + + dmr1024 __builtin_mma_pmdmxvf16gerx2nn_internal (dmr1024, v256, vuc, const int<8>, \ + const int<4>, const int<2>); + PMDMXVF16GERX2NN_INTERNAL dmf_pmdmxvf16gerx2nn {dm,pair} diff --git a/gcc/testsuite/gcc.target/powerpc/dmf-builtin-2.c b/gcc/testsuite/gcc.target/powerpc/dmf-builtin-2.c new file mode 100644 index 000000000000..2074edc7aed9 --- /dev/null +++ b/gcc/testsuite/gcc.target/powerpc/dmf-builtin-2.c @@ -0,0 +1,170 @@ +/* { dg-do compile } */ +/* { dg-require-effective-target powerpc_future_compile_ok } */ +/* { dg-options "-mdejagnu-cpu=future -O2" } */ + +typedef unsigned char vec_t __attribute__((vector_size(16))); + +void +foo (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_dmsetdmrz (&dmr); + __builtin_mma_dmxvf16gerx2 (&dmr, vp, vec); + *dst = dmr; +} + +void +bar (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr = dst[0];; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_dmxvf16gerx2 (&dmr, vp, vec); + dst[1] = dmr; +} + +/* { dg-final { scan-assembler-times {\mdmxvf16gerx2\M} 2 } } */ + +void +foo_1 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_dmsetdmrz (&dmr); + __builtin_mma_dmxvf16gerx2nn (&dmr, vp, vec); + *dst = dmr; +} + +void +bar_1 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr = dst[0];; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_dmxvf16gerx2nn (&dmr, vp, vec); + dst[1] = dmr; +} + +/* { dg-final { scan-assembler-times {\mdmxvf16gerx2nn\M} 2 } } */ + +void +foo_2 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_dmsetdmrz (&dmr); + __builtin_mma_dmxvf16gerx2np (&dmr, vp, vec); + *dst = dmr; +} + +void +bar_2 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr = dst[0];; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_dmxvf16gerx2np (&dmr, vp, vec); + dst[1] = dmr; +} + +/* { dg-final { scan-assembler-times {\mdmxvf16gerx2np\M} 2 } } */ + +void +foo_3 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_dmsetdmrz (&dmr); + __builtin_mma_dmxvf16gerx2pn (&dmr, vp, vec); + *dst = dmr; +} + +void +bar_3 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr = dst[0];; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_dmxvf16gerx2pn (&dmr, vp, vec); + dst[1] = dmr; +} + +/* { dg-final { scan-assembler-times {\mdmxvf16gerx2pn\M} 2 } } */ + +void +foo_4 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_dmsetdmrz (&dmr); + __builtin_mma_dmxvf16gerx2pp (&dmr, vp, vec); + *dst = dmr; +} + +void +bar_4 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __dmr1024 dmr = dst[0];; + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_dmxvf16gerx2pp (&dmr, vp, vec); + dst[1] = dmr; +} + +/* { dg-final { scan-assembler-times {\mdmxvf16gerx2pp\M} 2 } } */ + +void +foo_5 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_pmdmxvf16gerx2 (dst, vp, vec, 255, 15, 2); +} + +/* { dg-final { scan-assembler-times {\mpmdmxvf16gerx2\M} 1 } } */ + +void +foo_6 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_pmdmxvf16gerx2nn (dst, vp, vec, 255, 15, 2); +} + +/* { dg-final { scan-assembler-times {\mpmdmxvf16gerx2nn\M} 1 } } */ + +void +foo_7 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_pmdmxvf16gerx2np (dst, vp, vec, 255, 15, 2); +} + +/* { dg-final { scan-assembler-times {\mpmdmxvf16gerx2np\M} 1 } } */ + +void +foo_8 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_pmdmxvf16gerx2pn (dst, vp, vec, 255, 15, 2); +} + +/* { dg-final { scan-assembler-times {\mpmdmxvf16gerx2pn\M} 1 } } */ + +void +foo_9 (__dmr1024 *dst, __vector_pair *vpp, vec_t *src) +{ + __vector_pair vp = *vpp; + vec_t vec = *src; + __builtin_mma_pmdmxvf16gerx2pp (dst, vp, vec, 255, 15, 2); +} + +/* { dg-final { scan-assembler-times {\mpmdmxvf16gerx2pp\M} 1 } } */