Re: [PATCH v6 0/8] VSX MMA Implementation

[Daniel Henrique Barboza]

Queued in gitlab.com/danielhb/qemu/tree/ppc-next. Thanks,


Daniel

On 5/24/22 11:05, Lucas Mateus Castro(alqotel) wrote:
> From: "Lucas Mateus Castro (alqotel)" <lucas.araujo@eldorado.org.br>
>
> Based-on: https://gitlab.com/danielhb/qemu/-/tree/ppc-next
>
> This patch series is a patch series of the Matrix-Multiply Assist (MMA)
> instructions implementation from the PowerISA 3.1
>
> This patch series was created based on Victor's target/ppc: Fix FPSCR.FI
> bit patch series changes as that series changed do_check_float_status,
> which is called by the GER helper functions.
>
> These and the VDIV/VMOD implementation are the last new PowerISA 3.1
> instructions left to be implemented.
>
> The XVFGER instructions accumulate the exception status and at the end
> set the FPSCR and take a Program interrupt on a trap-enabled exception,
> previous versions were based on Victor's rework of FPU exceptions, but
> as that patch was rejected this version worked around the fact that
> OX/UX/XX and invalid instructions were handled in different functions
> by disabling all enable bits then re-enabling them and calling the mtfsf
> deferred exception helper.
>
> v6 changes:
>      - Rebased on ppc-next
>      - Wrapped lines to stay <= 80 characters
>
> v5 changes:
>      - Changed VSXGER16 accumulation to negate the multiplication and
>        accumulation in independent if's (if necessary) and sum their
>        values.
>
> v4 changes:
>      - Changed VSXGER16 accumulation to always use float32_sum and negate
>        the elements according to the type of accumulation
>
> v3 changes:
>      - GER helpers now use ppc_acc_t instead of ppc_vsr_t for passing acc
>      - Removed do_ger_XX3 and updated the decodetree to pass the masks in
>        32 bits instructions
>      - Removed unnecessary rounding mode function
>      - Moved float32_neg to fpu_helper.c and renamed it bfp32_negate to
>        make it clearer that it's a 32 bit version of the PowerISA
>        bfp_NEGATE
>      - Negated accumulation now a subtraction
>      - Changed exception handling by disabling all enable FPSCR enable
>        bits to set all FPSCR bits (except FEX) correctly, then re-enable
>        them and call do_fpscr_check_status to raise the exception
>        accordingly and set FEX if necessary
>
> v2 changes:
>      - Changed VSXGER, VSXGER16 and XVIGER macros to functions
>      - Set rounding mode in floating-point instructions based on RN
>        before operations
>      - Separated accumulate and with saturation instructions in
>        different helpers
>      - Used FIELD, FIELD_EX32 and FIELD_DP32 for packing/unpacking masks
>
>
> Joel Stanley (1):
>    linux-user: Add PowerPC ISA 3.1 and MMA to hwcap
>
> Lucas Mateus Castro (alqotel) (7):
>    target/ppc: Implement xxm[tf]acc and xxsetaccz
>    target/ppc: Implemented xvi*ger* instructions
>    target/ppc: Implemented pmxvi*ger* instructions
>    target/ppc: Implemented xvf*ger*
>    target/ppc: Implemented xvf16ger*
>    target/ppc: Implemented pmxvf*ger*
>    target/ppc: Implemented [pm]xvbf16ger2*
>
>   linux-user/elfload.c                |   4 +
>   target/ppc/cpu.h                    |  13 ++
>   target/ppc/fpu_helper.c             | 329 +++++++++++++++++++++++++++-
>   target/ppc/helper.h                 |  33 +++
>   target/ppc/insn32.decode            |  52 +++++
>   target/ppc/insn64.decode            |  79 +++++++
>   target/ppc/int_helper.c             | 130 +++++++++++
>   target/ppc/internal.h               |  15 ++
>   target/ppc/translate/vsx-impl.c.inc | 130 +++++++++++
>   9 files changed, 783 insertions(+), 2 deletions(-)