It seems overkill to use LLVM IR for BFE. There are several ways to
express BFE(value, offset, bits) exactly, for example:
# if bits == 0:
# 0
# else if offset + bits < 32:
# (value << (32 - offset - bits)) >> (32 - bits)
# else:
# value >> offset
Which can be simplified to either:
#
On Wed, Mar 11, 2015 at 2:54 AM, Tom Stellard wrote:
> On Wed, Mar 11, 2015 at 01:27:32AM +0100, Marek Olšák wrote:
>> On Wed, Mar 11, 2015 at 12:09 AM, Tom Stellard wrote:
>> > On Tue, Mar 10, 2015 at 11:01:21PM +0100, Marek Olšák wrote:
>> >> I've looked into how to recognize BFM and BFI and di
On Wed, Mar 11, 2015 at 01:27:32AM +0100, Marek Olšák wrote:
> On Wed, Mar 11, 2015 at 12:09 AM, Tom Stellard wrote:
> > On Tue, Mar 10, 2015 at 11:01:21PM +0100, Marek Olšák wrote:
> >> I've looked into how to recognize BFM and BFI and discovered that if
> >> TGSI_OPCODE_BFI is expanded, it's _im
On Wed, Mar 11, 2015 at 12:09 AM, Tom Stellard wrote:
> On Tue, Mar 10, 2015 at 11:01:21PM +0100, Marek Olšák wrote:
>> I've looked into how to recognize BFM and BFI and discovered that if
>> TGSI_OPCODE_BFI is expanded, it's _impossible_ to recognize the
>> pattern in the backend due to LLVM tran
On Tue, Mar 10, 2015 at 11:01:21PM +0100, Marek Olšák wrote:
> I've looked into how to recognize BFM and BFI and discovered that if
> TGSI_OPCODE_BFI is expanded, it's _impossible_ to recognize the
> pattern in the backend due to LLVM transformations. The reason it's
> impossible is that one partic
I've looked into how to recognize BFM and BFI and discovered that if
TGSI_OPCODE_BFI is expanded, it's _impossible_ to recognize the
pattern in the backend due to LLVM transformations. The reason it's
impossible is that one particular simplification of the expanded IR
can always be done and it alwa
On Tue, Mar 10, 2015 at 12:42:38PM +0100, Marek Olšák wrote:
> OK. What about patches 8 an 9?
>
I think the intrinsics in 9 are OK, but 8 should be using LLVM IR.
-Tom
> Marek
>
> On Thu, Mar 5, 2015 at 8:30 PM, Tom Stellard wrote:
> > On Mon, Mar 02, 2015 at 12:54:21PM +0100, Marek Olšák wro
OK. What about patches 8 an 9?
Marek
On Thu, Mar 5, 2015 at 8:30 PM, Tom Stellard wrote:
> On Mon, Mar 02, 2015 at 12:54:21PM +0100, Marek Olšák wrote:
>> From: Marek Olšák
>>
>
> Hi Marek,
>
> After discussing with Matt, I think we should use LLVM IR rather than
> intrinsics for IBFE and UBFE
On Mon, Mar 02, 2015 at 12:54:21PM +0100, Marek Olšák wrote:
> From: Marek Olšák
>
Hi Marek,
After discussing with Matt, I think we should use LLVM IR rather than
intrinsics for IBFE and UBFE and then add patterns for them either in
the TableGen Files or AMDGPUISelDAGToDAG.cpp.
Using intrinsic
On Mon, Mar 02, 2015 at 12:54:21PM +0100, Marek Olšák wrote:
> From: Marek Olšák
>
I'm still unsure whether it's better to use intrinsics or LLVM IR
to implement these. I will think about this some more.
-Tom
> ---
> src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c | 8
> 1 file
From: Marek Olšák
---
src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c | 8
1 file changed, 8 insertions(+)
diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
index 385d3ad..034095f 100644
--- a/src/gallium/drivers
11 matches
Mail list logo
