2015-11-24 05:33:58 +08:00
|
|
|
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
2015-09-05 17:59:59 +08:00
|
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1
|
|
|
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2
|
2017-05-25 21:45:23 +08:00
|
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512VPOPCNTDQ
|
2015-05-29 06:46:48 +08:00
|
|
|
|
2015-09-05 17:59:59 +08:00
|
|
|
define <4 x i64> @testv4i64(<4 x i64> %in) nounwind {
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX1-LABEL: testv4i64:
|
|
|
|
|
; AVX1: # BB#0:
|
|
|
|
|
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
|
2017-02-11 13:32:57 +08:00
|
|
|
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm1, %xmm3
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm3, %xmm4, %xmm3
|
|
|
|
|
; AVX1-NEXT: vpsrlw $4, %xmm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm1, %xmm4, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3
|
2015-11-25 03:51:26 +08:00
|
|
|
; AVX1-NEXT: vpsadbw %xmm3, %xmm1, %xmm1
|
2017-02-11 13:32:57 +08:00
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm0, %xmm5
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vpshufb %xmm5, %xmm4, %xmm5
|
|
|
|
|
; AVX1-NEXT: vpsrlw $4, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm0, %xmm4, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpaddb %xmm5, %xmm0, %xmm0
|
2015-11-25 03:51:26 +08:00
|
|
|
; AVX1-NEXT: vpsadbw %xmm3, %xmm0, %xmm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
|
|
|
|
|
; AVX1-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX2-LABEL: testv4i64:
|
|
|
|
|
; AVX2: # BB#0:
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm2
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
|
|
|
|
|
; AVX2-NEXT: vpshufb %ymm2, %ymm3, %ymm2
|
|
|
|
|
; AVX2-NEXT: vpsrlw $4, %ymm0, %ymm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm0
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vpshufb %ymm0, %ymm3, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpxor %ymm1, %ymm1, %ymm1
|
2015-11-25 03:51:26 +08:00
|
|
|
; AVX2-NEXT: vpsadbw %ymm1, %ymm0, %ymm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX2-NEXT: retq
|
2017-05-25 21:45:23 +08:00
|
|
|
;
|
|
|
|
|
; AVX512VPOPCNTDQ-LABEL: testv4i64:
|
|
|
|
|
; AVX512VPOPCNTDQ: # BB#0:
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpopcntq %zmm0, %zmm0
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: retq
|
2015-05-29 06:46:48 +08:00
|
|
|
%out = call <4 x i64> @llvm.ctpop.v4i64(<4 x i64> %in)
|
|
|
|
|
ret <4 x i64> %out
|
|
|
|
|
}
|
|
|
|
|
|
2015-09-05 17:59:59 +08:00
|
|
|
define <8 x i32> @testv8i32(<8 x i32> %in) nounwind {
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX1-LABEL: testv8i32:
|
|
|
|
|
; AVX1: # BB#0:
|
|
|
|
|
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
|
2017-02-11 13:32:57 +08:00
|
|
|
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm1, %xmm3
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm3, %xmm4, %xmm3
|
|
|
|
|
; AVX1-NEXT: vpsrlw $4, %xmm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm1, %xmm4, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3
|
|
|
|
|
; AVX1-NEXT: vpunpckhdq {{.*#+}} xmm5 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]
|
2015-11-25 03:51:26 +08:00
|
|
|
; AVX1-NEXT: vpsadbw %xmm3, %xmm5, %xmm5
|
2016-11-29 22:18:51 +08:00
|
|
|
; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
|
2015-11-25 03:51:26 +08:00
|
|
|
; AVX1-NEXT: vpsadbw %xmm3, %xmm1, %xmm1
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vpackuswb %xmm5, %xmm1, %xmm1
|
2017-02-11 13:32:57 +08:00
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm0, %xmm5
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vpshufb %xmm5, %xmm4, %xmm5
|
|
|
|
|
; AVX1-NEXT: vpsrlw $4, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm0, %xmm4, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpaddb %xmm5, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm3[2],xmm0[3],xmm3[3]
|
2015-11-25 03:51:26 +08:00
|
|
|
; AVX1-NEXT: vpsadbw %xmm3, %xmm2, %xmm2
|
2016-11-29 22:18:51 +08:00
|
|
|
; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
|
2015-11-25 03:51:26 +08:00
|
|
|
; AVX1-NEXT: vpsadbw %xmm3, %xmm0, %xmm0
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
|
|
|
|
|
; AVX1-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX2-LABEL: testv8i32:
|
|
|
|
|
; AVX2: # BB#0:
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm2
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
|
|
|
|
|
; AVX2-NEXT: vpshufb %ymm2, %ymm3, %ymm2
|
|
|
|
|
; AVX2-NEXT: vpsrlw $4, %ymm0, %ymm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm0
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vpshufb %ymm0, %ymm3, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpxor %ymm1, %ymm1, %ymm1
|
|
|
|
|
; AVX2-NEXT: vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
|
2015-11-25 03:51:26 +08:00
|
|
|
; AVX2-NEXT: vpsadbw %ymm1, %ymm2, %ymm2
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
|
2015-11-25 03:51:26 +08:00
|
|
|
; AVX2-NEXT: vpsadbw %ymm1, %ymm0, %ymm0
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vpackuswb %ymm2, %ymm0, %ymm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX2-NEXT: retq
|
2017-05-25 21:45:23 +08:00
|
|
|
;
|
|
|
|
|
; AVX512VPOPCNTDQ-LABEL: testv8i32:
|
|
|
|
|
; AVX512VPOPCNTDQ: # BB#0:
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: retq
|
2015-05-29 06:46:48 +08:00
|
|
|
%out = call <8 x i32> @llvm.ctpop.v8i32(<8 x i32> %in)
|
|
|
|
|
ret <8 x i32> %out
|
|
|
|
|
}
|
|
|
|
|
|
2015-09-05 17:59:59 +08:00
|
|
|
define <16 x i16> @testv16i16(<16 x i16> %in) nounwind {
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX1-LABEL: testv16i16:
|
|
|
|
|
; AVX1: # BB#0:
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
|
|
|
|
|
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm2
|
|
|
|
|
; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm2, %xmm3, %xmm2
|
|
|
|
|
; AVX1-NEXT: vpsrlw $4, %xmm0, %xmm4
|
|
|
|
|
; AVX1-NEXT: vpand %xmm1, %xmm4, %xmm4
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm4, %xmm3, %xmm4
|
|
|
|
|
; AVX1-NEXT: vpaddb %xmm2, %xmm4, %xmm2
|
2015-05-30 18:35:03 +08:00
|
|
|
; AVX1-NEXT: vpsllw $8, %xmm2, %xmm4
|
|
|
|
|
; AVX1-NEXT: vpaddb %xmm2, %xmm4, %xmm2
|
|
|
|
|
; AVX1-NEXT: vpsrlw $8, %xmm2, %xmm2
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
|
2015-05-30 18:35:03 +08:00
|
|
|
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm4
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm4, %xmm3, %xmm4
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vpsrlw $4, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm0, %xmm3, %xmm0
|
2015-05-30 18:35:03 +08:00
|
|
|
; AVX1-NEXT: vpaddb %xmm4, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpsllw $8, %xmm0, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpaddb %xmm0, %xmm1, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpsrlw $8, %xmm0, %xmm0
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm2, %ymm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX1-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX2-LABEL: testv16i16:
|
|
|
|
|
; AVX2: # BB#0:
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
|
|
|
|
|
; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm2
|
|
|
|
|
; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
|
|
|
|
|
; AVX2-NEXT: vpshufb %ymm2, %ymm3, %ymm2
|
|
|
|
|
; AVX2-NEXT: vpsrlw $4, %ymm0, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpshufb %ymm0, %ymm3, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm0
|
2015-05-30 18:35:03 +08:00
|
|
|
; AVX2-NEXT: vpsllw $8, %ymm0, %ymm1
|
|
|
|
|
; AVX2-NEXT: vpaddb %ymm0, %ymm1, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpsrlw $8, %ymm0, %ymm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX2-NEXT: retq
|
2017-05-25 21:45:23 +08:00
|
|
|
;
|
|
|
|
|
; AVX512VPOPCNTDQ-LABEL: testv16i16:
|
|
|
|
|
; AVX512VPOPCNTDQ: # BB#0:
|
2017-07-03 03:32:37 +08:00
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpmovdw %zmm0, %ymm0
|
2017-05-25 21:45:23 +08:00
|
|
|
; AVX512VPOPCNTDQ-NEXT: retq
|
2015-05-29 06:46:48 +08:00
|
|
|
%out = call <16 x i16> @llvm.ctpop.v16i16(<16 x i16> %in)
|
|
|
|
|
ret <16 x i16> %out
|
|
|
|
|
}
|
|
|
|
|
|
2015-09-05 17:59:59 +08:00
|
|
|
define <32 x i8> @testv32i8(<32 x i8> %in) nounwind {
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX1-LABEL: testv32i8:
|
|
|
|
|
; AVX1: # BB#0:
|
|
|
|
|
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
|
2017-02-11 13:32:57 +08:00
|
|
|
; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm1, %xmm3
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm3, %xmm4, %xmm3
|
|
|
|
|
; AVX1-NEXT: vpsrlw $4, %xmm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm1, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm1, %xmm4, %xmm1
|
|
|
|
|
; AVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
|
2017-02-11 13:32:57 +08:00
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm0, %xmm3
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX1-NEXT: vpshufb %xmm3, %xmm4, %xmm3
|
|
|
|
|
; AVX1-NEXT: vpsrlw $4, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpand %xmm2, %xmm0, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpshufb %xmm0, %xmm4, %xmm0
|
|
|
|
|
; AVX1-NEXT: vpaddb %xmm3, %xmm0, %xmm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
|
|
|
|
|
; AVX1-NEXT: retq
|
|
|
|
|
;
|
|
|
|
|
; AVX2-LABEL: testv32i8:
|
|
|
|
|
; AVX2: # BB#0:
|
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
2015-05-30 11:20:59 +08:00
|
|
|
; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
|
|
|
|
|
; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm2
|
|
|
|
|
; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
|
|
|
|
|
; AVX2-NEXT: vpshufb %ymm2, %ymm3, %ymm2
|
|
|
|
|
; AVX2-NEXT: vpsrlw $4, %ymm0, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpshufb %ymm0, %ymm3, %ymm0
|
|
|
|
|
; AVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm0
|
2015-05-29 06:46:48 +08:00
|
|
|
; AVX2-NEXT: retq
|
2017-05-25 21:45:23 +08:00
|
|
|
;
|
|
|
|
|
; AVX512VPOPCNTDQ-LABEL: testv32i8:
|
|
|
|
|
; AVX512VPOPCNTDQ: # BB#0:
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpand %ymm1, %ymm0, %ymm2
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpshufb %ymm2, %ymm3, %ymm2
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpsrlw $4, %ymm0, %ymm0
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpand %ymm1, %ymm0, %ymm0
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpshufb %ymm0, %ymm3, %ymm0
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: vpaddb %ymm2, %ymm0, %ymm0
|
|
|
|
|
; AVX512VPOPCNTDQ-NEXT: retq
|
2015-05-29 06:46:48 +08:00
|
|
|
%out = call <32 x i8> @llvm.ctpop.v32i8(<32 x i8> %in)
|
|
|
|
|
ret <32 x i8> %out
|
|
|
|
|
}
|
|
|
|
|
|
2015-09-05 17:59:59 +08:00
|
|
|
define <4 x i64> @foldv4i64() nounwind {
|
|
|
|
|
; ALL-LABEL: foldv4i64:
|
|
|
|
|
; ALL: # BB#0:
|
|
|
|
|
; ALL-NEXT: vmovaps {{.*#+}} ymm0 = [1,64,0,8]
|
|
|
|
|
; ALL-NEXT: retq
|
2015-06-07 23:37:14 +08:00
|
|
|
%out = call <4 x i64> @llvm.ctpop.v4i64(<4 x i64> <i64 256, i64 -1, i64 0, i64 255>)
|
|
|
|
|
ret <4 x i64> %out
|
|
|
|
|
}
|
2015-06-08 00:09:48 +08:00
|
|
|
|
2015-09-05 17:59:59 +08:00
|
|
|
define <8 x i32> @foldv8i32() nounwind {
|
|
|
|
|
; ALL-LABEL: foldv8i32:
|
|
|
|
|
; ALL: # BB#0:
|
|
|
|
|
; ALL-NEXT: vmovaps {{.*#+}} ymm0 = [1,32,0,8,16,3,2,3]
|
|
|
|
|
; ALL-NEXT: retq
|
2015-06-07 23:37:14 +08:00
|
|
|
%out = call <8 x i32> @llvm.ctpop.v8i32(<8 x i32> <i32 256, i32 -1, i32 0, i32 255, i32 -65536, i32 7, i32 24, i32 88>)
|
|
|
|
|
ret <8 x i32> %out
|
|
|
|
|
}
|
2015-06-08 00:09:48 +08:00
|
|
|
|
2015-09-05 17:59:59 +08:00
|
|
|
define <16 x i16> @foldv16i16() nounwind {
|
|
|
|
|
; ALL-LABEL: foldv16i16:
|
|
|
|
|
; ALL: # BB#0:
|
|
|
|
|
; ALL-NEXT: vmovaps {{.*#+}} ymm0 = [1,16,0,8,0,3,2,3,15,7,1,1,1,1,1,1]
|
|
|
|
|
; ALL-NEXT: retq
|
2015-06-07 23:37:14 +08:00
|
|
|
%out = call <16 x i16> @llvm.ctpop.v16i16(<16 x i16> <i16 256, i16 -1, i16 0, i16 255, i16 -65536, i16 7, i16 24, i16 88, i16 -2, i16 254, i16 1, i16 2, i16 4, i16 8, i16 16, i16 32>)
|
|
|
|
|
ret <16 x i16> %out
|
|
|
|
|
}
|
2015-06-08 00:09:48 +08:00
|
|
|
|
2015-09-05 17:59:59 +08:00
|
|
|
define <32 x i8> @foldv32i8() nounwind {
|
|
|
|
|
; ALL-LABEL: foldv32i8:
|
|
|
|
|
; ALL: # BB#0:
|
|
|
|
|
; ALL-NEXT: vmovaps {{.*#+}} ymm0 = [0,8,0,8,0,3,2,3,7,7,1,1,1,1,1,1,1,1,0,0,1,2,3,4,5,6,7,8,2,2,3,7]
|
|
|
|
|
; ALL-NEXT: retq
|
2015-06-07 23:37:14 +08:00
|
|
|
%out = call <32 x i8> @llvm.ctpop.v32i8(<32 x i8> <i8 256, i8 -1, i8 0, i8 255, i8 -65536, i8 7, i8 24, i8 88, i8 -2, i8 254, i8 1, i8 2, i8 4, i8 8, i8 16, i8 32, i8 64, i8 128, i8 256, i8 -256, i8 -128, i8 -64, i8 -32, i8 -16, i8 -8, i8 -4, i8 -2, i8 -1, i8 3, i8 5, i8 7, i8 127>)
|
|
|
|
|
ret <32 x i8> %out
|
|
|
|
|
}
|
|
|
|
|
|
2015-05-29 06:46:48 +08:00
|
|
|
declare <4 x i64> @llvm.ctpop.v4i64(<4 x i64>)
|
|
|
|
|
declare <8 x i32> @llvm.ctpop.v8i32(<8 x i32>)
|
|
|
|
|
declare <16 x i16> @llvm.ctpop.v16i16(<16 x i16>)
|
|
|
|
|
declare <32 x i8> @llvm.ctpop.v32i8(<32 x i8>)
|