[Mesa-dev] [PATCH 07/19] tgsi: add BALLOT/READ_* opcodes
Ilia Mirkin
imirkin at alum.mit.edu
Tue Apr 4 02:08:44 UTC 2017
On Mon, Apr 3, 2017 at 10:01 PM, Boyan Ding <boyan.j.ding at gmail.com> wrote:
> 2017-04-01 1:14 GMT+08:00 Nicolai Hähnle <nhaehnle at gmail.com>:
>> From: Ilia Mirkin <imirkin at alum.mit.edu>
>>
>> v2 (Nicolai):
>> - BALLOT isn't per-channel
>> - expand the documentation (also for VOTE_*)
>>
>> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
>> Signed-off-by: Nicolai Hähnle <nicolai.haehnle at amd.com>
>> ---
>> src/gallium/auxiliary/tgsi/tgsi_info.c | 6 +--
>> src/gallium/docs/source/tgsi.rst | 67 +++++++++++++++++++++++++-----
>> src/gallium/include/pipe/p_shader_tokens.h | 6 +--
>> 3 files changed, 62 insertions(+), 17 deletions(-)
>>
>> diff --git a/src/gallium/auxiliary/tgsi/tgsi_info.c b/src/gallium/auxiliary/tgsi/tgsi_info.c
>> index 5a6a9bc..30bad6d 100644
>> --- a/src/gallium/auxiliary/tgsi/tgsi_info.c
>> +++ b/src/gallium/auxiliary/tgsi/tgsi_info.c
>> @@ -106,51 +106,51 @@ static const struct tgsi_opcode_info opcode_info[TGSI_OPCODE_LAST] =
>> { 1, 3, 0, 0, 0, 0, 0, COMP, "CMP", TGSI_OPCODE_CMP },
>> { 1, 1, 0, 0, 0, 0, 0, CHAN, "SCS", TGSI_OPCODE_SCS },
>> { 1, 2, 1, 0, 0, 0, 0, OTHR, "TXB", TGSI_OPCODE_TXB },
>> { 1, 1, 0, 0, 0, 0, 0, OTHR, "FBFETCH", TGSI_OPCODE_FBFETCH },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "DIV", TGSI_OPCODE_DIV },
>> { 1, 2, 0, 0, 0, 0, 0, REPL, "DP2", TGSI_OPCODE_DP2 },
>> { 1, 2, 1, 0, 0, 0, 0, OTHR, "TXL", TGSI_OPCODE_TXL },
>> { 0, 0, 0, 0, 0, 0, 0, NONE, "BRK", TGSI_OPCODE_BRK },
>> { 0, 1, 0, 0, 1, 0, 1, NONE, "IF", TGSI_OPCODE_IF },
>> { 0, 1, 0, 0, 1, 0, 1, NONE, "UIF", TGSI_OPCODE_UIF },
>> - { 0, 1, 0, 0, 0, 0, 1, NONE, "", 76 }, /* removed */
>> + { 1, 2, 0, 0, 0, 0, 0, COMP, "READ_INVOC", TGSI_OPCODE_READ_INVOC },
>> { 0, 0, 0, 0, 1, 1, 1, NONE, "ELSE", TGSI_OPCODE_ELSE },
>> { 0, 0, 0, 0, 0, 1, 0, NONE, "ENDIF", TGSI_OPCODE_ENDIF },
>> { 1, 1, 0, 0, 0, 0, 0, COMP, "DDX_FINE", TGSI_OPCODE_DDX_FINE },
>> { 1, 1, 0, 0, 0, 0, 0, COMP, "DDY_FINE", TGSI_OPCODE_DDY_FINE },
>> { 0, 1, 0, 0, 0, 0, 0, NONE, "PUSHA", TGSI_OPCODE_PUSHA },
>> { 1, 0, 0, 0, 0, 0, 0, NONE, "POPA", TGSI_OPCODE_POPA },
>> { 1, 1, 0, 0, 0, 0, 0, COMP, "CEIL", TGSI_OPCODE_CEIL },
>> { 1, 1, 0, 0, 0, 0, 0, COMP, "I2F", TGSI_OPCODE_I2F },
>> { 1, 1, 0, 0, 0, 0, 0, COMP, "NOT", TGSI_OPCODE_NOT },
>> { 1, 1, 0, 0, 0, 0, 0, COMP, "TRUNC", TGSI_OPCODE_TRUNC },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "SHL", TGSI_OPCODE_SHL },
>> - { 0, 0, 0, 0, 0, 0, 0, NONE, "", 88 }, /* removed */
>> + { 1, 1, 0, 0, 0, 0, 0, OTHR, "BALLOT", TGSI_OPCODE_BALLOT },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "AND", TGSI_OPCODE_AND },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "OR", TGSI_OPCODE_OR },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "MOD", TGSI_OPCODE_MOD },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "XOR", TGSI_OPCODE_XOR },
>> { 1, 3, 0, 0, 0, 0, 0, COMP, "SAD", TGSI_OPCODE_SAD },
>> { 1, 2, 1, 0, 0, 0, 0, OTHR, "TXF", TGSI_OPCODE_TXF },
>> { 1, 2, 1, 0, 0, 0, 0, OTHR, "TXQ", TGSI_OPCODE_TXQ },
>> { 0, 0, 0, 0, 0, 0, 0, NONE, "CONT", TGSI_OPCODE_CONT },
>> { 0, 1, 0, 0, 0, 0, 0, NONE, "EMIT", TGSI_OPCODE_EMIT },
>> { 0, 1, 0, 0, 0, 0, 0, NONE, "ENDPRIM", TGSI_OPCODE_ENDPRIM },
>> { 0, 0, 0, 0, 1, 0, 1, NONE, "BGNLOOP", TGSI_OPCODE_BGNLOOP },
>> { 0, 0, 0, 0, 0, 0, 1, NONE, "BGNSUB", TGSI_OPCODE_BGNSUB },
>> { 0, 0, 0, 0, 1, 1, 0, NONE, "ENDLOOP", TGSI_OPCODE_ENDLOOP },
>> { 0, 0, 0, 0, 0, 1, 0, NONE, "ENDSUB", TGSI_OPCODE_ENDSUB },
>> { 1, 1, 1, 0, 0, 0, 0, OTHR, "TXQ_LZ", TGSI_OPCODE_TXQ_LZ },
>> { 1, 1, 1, 0, 0, 0, 0, OTHR, "TXQS", TGSI_OPCODE_TXQS },
>> { 1, 1, 0, 0, 0, 0, 0, OTHR, "RESQ", TGSI_OPCODE_RESQ },
>> - { 0, 0, 0, 0, 0, 0, 0, NONE, "", 106 }, /* removed */
>> + { 1, 1, 0, 0, 0, 0, 0, COMP, "READ_FIRST", TGSI_OPCODE_READ_FIRST },
>> { 0, 0, 0, 0, 0, 0, 0, NONE, "NOP", TGSI_OPCODE_NOP },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "FSEQ", TGSI_OPCODE_FSEQ },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "FSGE", TGSI_OPCODE_FSGE },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "FSLT", TGSI_OPCODE_FSLT },
>> { 1, 2, 0, 0, 0, 0, 0, COMP, "FSNE", TGSI_OPCODE_FSNE },
>> { 0, 1, 0, 0, 0, 0, 0, OTHR, "MEMBAR", TGSI_OPCODE_MEMBAR },
>> { 0, 1, 0, 0, 0, 0, 0, NONE, "CALLNZ", TGSI_OPCODE_CALLNZ },
>> { 0, 1, 0, 0, 0, 0, 0, NONE, "", 114 }, /* removed */
>> { 0, 1, 0, 0, 0, 0, 0, NONE, "BREAKC", TGSI_OPCODE_BREAKC },
>> { 0, 1, 0, 0, 0, 0, 0, NONE, "KILL_IF", TGSI_OPCODE_KILL_IF },
>> diff --git a/src/gallium/docs/source/tgsi.rst b/src/gallium/docs/source/tgsi.rst
>> index 05b06ce..7e9b47c 100644
>> --- a/src/gallium/docs/source/tgsi.rst
>> +++ b/src/gallium/docs/source/tgsi.rst
>> @@ -2852,36 +2852,81 @@ only be used with 32-bit integer image formats.
>>
>> The following operation is performed atomically:
>>
>> .. math::
>>
>> dst_x = resource[offset]
>>
>> resource[offset] = (dst_x > src_x ? dst_x : src_x)
>>
>>
>> -.. _voteopcodes:
>> +.. _interlaneopcodes:
>> +
>> +Inter-lane opcodes
>> +^^^^^^^^^^^^^^^^^^
>> +
>> +These opcodes reduce the given value across the shader invocations
>> +running in the current SIMD group. For all but the READ_INVOC operations,
>> +every thread in the subgroup will receive the same result. The VOTE_* and
>> +BALLOT operations accept a single-channel argument that is treated as a
>> +boolean and produce a 64-bit value.
>
> Really? Aren't VOTE_* opcodes producing bool (32-bit) values?
Yeah, they should be producing (32-bit) bools. At least how they're
implemented in nouveau is also per-channel, i.e.
case TGSI_OPCODE_VOTE_ALL:
case TGSI_OPCODE_VOTE_ANY:
case TGSI_OPCODE_VOTE_EQ:
val0 = new_LValue(func, FILE_PREDICATE);
FOR_EACH_DST_ENABLED_CHANNEL(0, c, tgsi) {
mkCmp(OP_SET, CC_NE, TYPE_U32, val0, TYPE_U32, fetchSrc(0, c), zero);
mkOp1(op, dstTy, val0, val0)
->subOp = tgsi::opcodeToSubOp(tgsi.getOpcode());
mkCvt(OP_CVT, TYPE_U32, dst0[c], TYPE_U8, val0);
}
break;
although that per-channel-ness isn't used by st/mesa.
-ilia
More information about the mesa-dev
mailing list