generic-library/isa-l
1
0
Fork 0
mirror of https://github.com/intel/isa-l.git synced 2024-12-12 17:33:50 +01:00
Code Issues Projects Releases Wiki Activity
2ca781df19
isa-l/erasure_code/aarch64/gf_6vect_dot_prod_sve.S

259 lines
7.1 KiB
ArmAsm
Raw Normal View History

Enable SVE in ISA-L erasure code for aarch64 This patch adds Arm (aarch64) SVE [1] variable-length vector assembly support into ISA-L erasure code library. "Arm designed the Scalable Vector Extension (SVE) as a next-generation SIMD extension to AArch64. SVE allows flexible vector length implementations with a range of possible values in CPU implementations. The vector length can vary from a minimum of 128 bits up to a maximum of 2048 bits, at 128-bit increments. The SVE design guarantees that the same application can run on different implementations that support SVE, without the need to recompile the code. " [3] Test method: - This patch was tested on Fujitsu's A64FX [2], and it passed all erasure code related test cases, including "make checks" , "make test", and "make perf". - To ensure code testing coverage, parameters in files (erasure_code/ erasure_code_test.c , erasure_code_update_test.c and gf_vect_mad_test.c) are modified to cover all _vect versions of _mad_sve() / _dot_prod_sve() rutines. Performance improvements over NEON: In general, SVE benchmarks (bandwidth in MB/s) are 40% ~ 100% higher than NEON when running _cold style (data uncached and pulled from memory) perfs. This includes routines of dot_prod, mad, and mul. Optimization points: This patch was tuned for the best performance on A64FX. Tuning points being touched in this patch include: 1) Data prefetch into L2 cache before loading. See _sve.S files. 2) Instruction sequence orchestration. Such as interleaving every two 'ld1b/st1b' instructions with other instructions. See _sve.S files. 3) To improve dest vectors parallelism, in highlevel, running gf_4vect_dot_prod_sve twice is better than running gf_8vect_dot_prod_sve() once, and it's also better than running _7vect + _vect, _6vect + _2vect, and _5vect + _3vect. The similar idea is applied to improve 11 ~ 9 dest vectors dot product computing as well. The related change can be found in ec_encode_data_sve() of file: erasure_code/aarch64/ec_aarch64_highlevel_func.c Notes: 1) About vector length: A64FX has a vector register length of 512bit. However, this patchset was written with variable length assembly so it work automatically on aarch64 machines with any types of SVE vector length, such as SVE-128, SVE-256, etc.. 2) About optimization: Due to differences in microarchitecture and cache/memory design, to achieve optimum performance on SVE capable CPUs other than A64FX, it is considered necessary to do microarchitecture-level tunings on these CPUs. [1] Introduction to SVE - Arm Developer. https://developer.arm.com/documentation/102476/latest/ [2] FUJITSU Processor A64FX. https://www.fujitsu.com/global/products/computing/servers/supercomputer/a64fx/ [3] Introducing SVE. https://developer.arm.com/documentation/102476/0001/Introducing-SVE Change-Id: If49eb8a956154d799dcda0ba4c9c6d979f5064a9 Signed-off-by: Guodong Xu <guodong.xu@linaro.org>
2021-12-28 10:32:39 +01:00
/*************************************************************
Copyright (c) 2021 Linaro Ltd.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
* Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
* Neither the name of Huawei Corporation nor the names of its
contributors may be used to endorse or promote products derived
from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
**********************************************************************/
.text
.align 6
.arch armv8-a+sve
Fixes for aarch64 mac - It should be fine to enable pmull always on Apple Silicon - macOS 12+ is required for PMULL instruction. - Changed the conditional macro to __APPLE__ - Rewritten dispatcher using sysctlbyname - Use __USER_LABEL_PREFIX__ - Use __TEXT,__const as readonly section - use ASM_DEF_RODATA macro - fix func decl Change-Id: I800593f21085d8187b480c8bb3ab2bd70c4a6974 Signed-off-by: Taiju Yamada <tyamada@bi.a.u-tokyo.ac.jp>
2020-11-21 17:51:37 +01:00
#include "../include/aarch64_label.h"
.global cdecl(gf_6vect_dot_prod_sve)
#ifndef __APPLE__
Enable SVE in ISA-L erasure code for aarch64 This patch adds Arm (aarch64) SVE [1] variable-length vector assembly support into ISA-L erasure code library. "Arm designed the Scalable Vector Extension (SVE) as a next-generation SIMD extension to AArch64. SVE allows flexible vector length implementations with a range of possible values in CPU implementations. The vector length can vary from a minimum of 128 bits up to a maximum of 2048 bits, at 128-bit increments. The SVE design guarantees that the same application can run on different implementations that support SVE, without the need to recompile the code. " [3] Test method: - This patch was tested on Fujitsu's A64FX [2], and it passed all erasure code related test cases, including "make checks" , "make test", and "make perf". - To ensure code testing coverage, parameters in files (erasure_code/ erasure_code_test.c , erasure_code_update_test.c and gf_vect_mad_test.c) are modified to cover all _vect versions of _mad_sve() / _dot_prod_sve() rutines. Performance improvements over NEON: In general, SVE benchmarks (bandwidth in MB/s) are 40% ~ 100% higher than NEON when running _cold style (data uncached and pulled from memory) perfs. This includes routines of dot_prod, mad, and mul. Optimization points: This patch was tuned for the best performance on A64FX. Tuning points being touched in this patch include: 1) Data prefetch into L2 cache before loading. See _sve.S files. 2) Instruction sequence orchestration. Such as interleaving every two 'ld1b/st1b' instructions with other instructions. See _sve.S files. 3) To improve dest vectors parallelism, in highlevel, running gf_4vect_dot_prod_sve twice is better than running gf_8vect_dot_prod_sve() once, and it's also better than running _7vect + _vect, _6vect + _2vect, and _5vect + _3vect. The similar idea is applied to improve 11 ~ 9 dest vectors dot product computing as well. The related change can be found in ec_encode_data_sve() of file: erasure_code/aarch64/ec_aarch64_highlevel_func.c Notes: 1) About vector length: A64FX has a vector register length of 512bit. However, this patchset was written with variable length assembly so it work automatically on aarch64 machines with any types of SVE vector length, such as SVE-128, SVE-256, etc.. 2) About optimization: Due to differences in microarchitecture and cache/memory design, to achieve optimum performance on SVE capable CPUs other than A64FX, it is considered necessary to do microarchitecture-level tunings on these CPUs. [1] Introduction to SVE - Arm Developer. https://developer.arm.com/documentation/102476/latest/ [2] FUJITSU Processor A64FX. https://www.fujitsu.com/global/products/computing/servers/supercomputer/a64fx/ [3] Introducing SVE. https://developer.arm.com/documentation/102476/0001/Introducing-SVE Change-Id: If49eb8a956154d799dcda0ba4c9c6d979f5064a9 Signed-off-by: Guodong Xu <guodong.xu@linaro.org>
2021-12-28 10:32:39 +01:00
.type gf_6vect_dot_prod_sve, %function
Fixes for aarch64 mac - It should be fine to enable pmull always on Apple Silicon - macOS 12+ is required for PMULL instruction. - Changed the conditional macro to __APPLE__ - Rewritten dispatcher using sysctlbyname - Use __USER_LABEL_PREFIX__ - Use __TEXT,__const as readonly section - use ASM_DEF_RODATA macro - fix func decl Change-Id: I800593f21085d8187b480c8bb3ab2bd70c4a6974 Signed-off-by: Taiju Yamada <tyamada@bi.a.u-tokyo.ac.jp>
2020-11-21 17:51:37 +01:00
#endif
Enable SVE in ISA-L erasure code for aarch64 This patch adds Arm (aarch64) SVE [1] variable-length vector assembly support into ISA-L erasure code library. "Arm designed the Scalable Vector Extension (SVE) as a next-generation SIMD extension to AArch64. SVE allows flexible vector length implementations with a range of possible values in CPU implementations. The vector length can vary from a minimum of 128 bits up to a maximum of 2048 bits, at 128-bit increments. The SVE design guarantees that the same application can run on different implementations that support SVE, without the need to recompile the code. " [3] Test method: - This patch was tested on Fujitsu's A64FX [2], and it passed all erasure code related test cases, including "make checks" , "make test", and "make perf". - To ensure code testing coverage, parameters in files (erasure_code/ erasure_code_test.c , erasure_code_update_test.c and gf_vect_mad_test.c) are modified to cover all _vect versions of _mad_sve() / _dot_prod_sve() rutines. Performance improvements over NEON: In general, SVE benchmarks (bandwidth in MB/s) are 40% ~ 100% higher than NEON when running _cold style (data uncached and pulled from memory) perfs. This includes routines of dot_prod, mad, and mul. Optimization points: This patch was tuned for the best performance on A64FX. Tuning points being touched in this patch include: 1) Data prefetch into L2 cache before loading. See _sve.S files. 2) Instruction sequence orchestration. Such as interleaving every two 'ld1b/st1b' instructions with other instructions. See _sve.S files. 3) To improve dest vectors parallelism, in highlevel, running gf_4vect_dot_prod_sve twice is better than running gf_8vect_dot_prod_sve() once, and it's also better than running _7vect + _vect, _6vect + _2vect, and _5vect + _3vect. The similar idea is applied to improve 11 ~ 9 dest vectors dot product computing as well. The related change can be found in ec_encode_data_sve() of file: erasure_code/aarch64/ec_aarch64_highlevel_func.c Notes: 1) About vector length: A64FX has a vector register length of 512bit. However, this patchset was written with variable length assembly so it work automatically on aarch64 machines with any types of SVE vector length, such as SVE-128, SVE-256, etc.. 2) About optimization: Due to differences in microarchitecture and cache/memory design, to achieve optimum performance on SVE capable CPUs other than A64FX, it is considered necessary to do microarchitecture-level tunings on these CPUs. [1] Introduction to SVE - Arm Developer. https://developer.arm.com/documentation/102476/latest/ [2] FUJITSU Processor A64FX. https://www.fujitsu.com/global/products/computing/servers/supercomputer/a64fx/ [3] Introducing SVE. https://developer.arm.com/documentation/102476/0001/Introducing-SVE Change-Id: If49eb8a956154d799dcda0ba4c9c6d979f5064a9 Signed-off-by: Guodong Xu <guodong.xu@linaro.org>
2021-12-28 10:32:39 +01:00
/* void gf_6vect_dot_prod_sve(int len, int vlen, unsigned char *gftbls,
unsigned char **src, unsigned char **dest);
*/
/* arguments */
x_len .req x0 /* vector length */
x_vec .req x1 /* number of source vectors (ie. data blocks) */
x_tbl .req x2
x_src .req x3
x_dest .req x4
/* returns */
w_ret .req w0
/* local variables */
x_vec_i .req x5
x_ptr .req x6
x_pos .req x7
x_tbl1 .req x8
x_tbl2 .req x9
x_tbl3 .req x10
x_tbl4 .req x11
x_tbl5 .req x12
x_tbl6 .req x13
x_dest1 .req x14
x_dest2 .req x15
x_dest6 .req x_dest /* reused */
/* r16,r17,r18,r29,r30: special role registers, avoided */
/* r19..r29 and SP must be preserved */
x_dest3 .req x19
x_dest4 .req x20
x_dest5 .req x21
/* vectors */
z_mask0f .req z0
z_src .req z1
z_src_lo .req z2
z_src_hi .req z_src
z_dest1 .req z3
z_gft1_lo .req z4
z_gft1_hi .req z5
q_gft1_lo .req q4
q_gft1_hi .req q5
/* bottom 64-bit of v8..v15 must be preserved if used */
z_gft2_lo .req z17
z_gft2_hi .req z18
q_gft2_lo .req q17
q_gft2_hi .req q18
z_gft3_lo .req z19
z_gft3_hi .req z20
q_gft3_lo .req q19
q_gft3_hi .req q20
z_gft4_lo .req z21
z_gft4_hi .req z22
q_gft4_lo .req q21
q_gft4_hi .req q22
z_gft5_lo .req z23
z_gft5_hi .req z24
q_gft5_lo .req q23
q_gft5_hi .req q24
z_gft6_lo .req z25
z_gft6_hi .req z26
q_gft6_lo .req q25
q_gft6_hi .req q26
z_dest2 .req z27
z_dest3 .req z28
z_dest4 .req z29
z_dest5 .req z30
z_dest6 .req z31
Fixes for aarch64 mac - It should be fine to enable pmull always on Apple Silicon - macOS 12+ is required for PMULL instruction. - Changed the conditional macro to __APPLE__ - Rewritten dispatcher using sysctlbyname - Use __USER_LABEL_PREFIX__ - Use __TEXT,__const as readonly section - use ASM_DEF_RODATA macro - fix func decl Change-Id: I800593f21085d8187b480c8bb3ab2bd70c4a6974 Signed-off-by: Taiju Yamada <tyamada@bi.a.u-tokyo.ac.jp>
2020-11-21 17:51:37 +01:00
cdecl(gf_6vect_dot_prod_sve):
Enable SVE in ISA-L erasure code for aarch64 This patch adds Arm (aarch64) SVE [1] variable-length vector assembly support into ISA-L erasure code library. "Arm designed the Scalable Vector Extension (SVE) as a next-generation SIMD extension to AArch64. SVE allows flexible vector length implementations with a range of possible values in CPU implementations. The vector length can vary from a minimum of 128 bits up to a maximum of 2048 bits, at 128-bit increments. The SVE design guarantees that the same application can run on different implementations that support SVE, without the need to recompile the code. " [3] Test method: - This patch was tested on Fujitsu's A64FX [2], and it passed all erasure code related test cases, including "make checks" , "make test", and "make perf". - To ensure code testing coverage, parameters in files (erasure_code/ erasure_code_test.c , erasure_code_update_test.c and gf_vect_mad_test.c) are modified to cover all _vect versions of _mad_sve() / _dot_prod_sve() rutines. Performance improvements over NEON: In general, SVE benchmarks (bandwidth in MB/s) are 40% ~ 100% higher than NEON when running _cold style (data uncached and pulled from memory) perfs. This includes routines of dot_prod, mad, and mul. Optimization points: This patch was tuned for the best performance on A64FX. Tuning points being touched in this patch include: 1) Data prefetch into L2 cache before loading. See _sve.S files. 2) Instruction sequence orchestration. Such as interleaving every two 'ld1b/st1b' instructions with other instructions. See _sve.S files. 3) To improve dest vectors parallelism, in highlevel, running gf_4vect_dot_prod_sve twice is better than running gf_8vect_dot_prod_sve() once, and it's also better than running _7vect + _vect, _6vect + _2vect, and _5vect + _3vect. The similar idea is applied to improve 11 ~ 9 dest vectors dot product computing as well. The related change can be found in ec_encode_data_sve() of file: erasure_code/aarch64/ec_aarch64_highlevel_func.c Notes: 1) About vector length: A64FX has a vector register length of 512bit. However, this patchset was written with variable length assembly so it work automatically on aarch64 machines with any types of SVE vector length, such as SVE-128, SVE-256, etc.. 2) About optimization: Due to differences in microarchitecture and cache/memory design, to achieve optimum performance on SVE capable CPUs other than A64FX, it is considered necessary to do microarchitecture-level tunings on these CPUs. [1] Introduction to SVE - Arm Developer. https://developer.arm.com/documentation/102476/latest/ [2] FUJITSU Processor A64FX. https://www.fujitsu.com/global/products/computing/servers/supercomputer/a64fx/ [3] Introducing SVE. https://developer.arm.com/documentation/102476/0001/Introducing-SVE Change-Id: If49eb8a956154d799dcda0ba4c9c6d979f5064a9 Signed-off-by: Guodong Xu <guodong.xu@linaro.org>
2021-12-28 10:32:39 +01:00
/* less than 16 bytes, return_fail */
cmp x_len, #16
blt .return_fail
/* save r19..r29 */
sub sp, sp, #32 /* alignment */
stp x19, x20, [sp]
str x21, [sp, #16]
mov z_mask0f.b, #0x0f /* z_mask0f = 0x0F0F...0F */
mov x_pos, #0
lsl x_vec, x_vec, #3
ldp x_dest1, x_dest2, [x_dest, #8*0]
ldp x_dest3, x_dest4, [x_dest, #8*2]
ldp x_dest5, x_dest6, [x_dest, #8*4] /* x_dest6 reuses x_dest */
/* Loop 1: x_len, vector length */
.Lloopsve_vl:
whilelo p0.b, x_pos, x_len
b.none .return_pass
mov x_vec_i, #0 /* clear x_vec_i */
ldr x_ptr, [x_src, x_vec_i] /* x_ptr: src base addr. */
mov z_dest1.b, #0 /* clear z_dest1 */
mov z_dest2.b, #0 /* clear z_dest2 */
mov z_dest3.b, #0 /* clear z_dest3 */
mov z_dest4.b, #0 /* clear z_dest4 */
mov z_dest5.b, #0 /* clear z_dest5 */
mov z_dest6.b, #0 /* clear z_dest6 */
/* gf_tbl base = (x_tbl + dest_idx * x_vec * 32) */
mov x_tbl1, x_tbl /* reset x_tbl1 */
add x_tbl2, x_tbl1, x_vec, LSL #2 /* reset x_tbl2 */
add x_tbl3, x_tbl2, x_vec, LSL #2 /* reset x_tbl3 */
add x_tbl4, x_tbl3, x_vec, LSL #2 /* reset x_tbl4 */
add x_tbl5, x_tbl4, x_vec, LSL #2 /* reset x_tbl5 */
add x_tbl6, x_tbl5, x_vec, LSL #2 /* reset x_tbl6 */
/* Loop 2: x_vec, number of source vectors (ie. data blocks) */
.Lloopsve_vl_vects:
/* load src data, governed by p0 */
ld1b z_src.b, p0/z, [x_ptr, x_pos] /* load from: src base + pos offset */
/* split 4-bit lo; 4-bit hi */
and z_src_lo.d, z_src.d, z_mask0f.d
lsr z_src_hi.b, z_src.b, #4
/* gf_tbl addr: (x_tbl + dest_idx * x_vec * 32) + src_vec_idx * 32 */
/* load gf_table's */
ldp q_gft1_lo, q_gft1_hi, [x_tbl1], #32 /* x_tbl1 is post-added by #32 for each src vect */
ldp q_gft2_lo, q_gft2_hi, [x_tbl2], #32
/* prefetch */
prfb pldl2keep, p0, [x_tbl1]
prfb pldl2keep, p0, [x_tbl2]
/* calc for next and prefetch */
add x_vec_i, x_vec_i, #8 /* move x_vec_i to next */
ldr x_ptr, [x_src, x_vec_i] /* x_ptr: src base addr. */
/* dest 1 */
/* table indexing, ie. gf(2^8) multiplication */
tbl z_gft1_lo.b, {z_gft1_lo.b}, z_src_lo.b
tbl z_gft1_hi.b, {z_gft1_hi.b}, z_src_hi.b
/* exclusive or, ie. gf(2^8) add */
eor z_dest1.d, z_gft1_lo.d, z_dest1.d
eor z_dest1.d, z_dest1.d, z_gft1_hi.d
ldp q_gft3_lo, q_gft3_hi, [x_tbl3], #32
ldp q_gft4_lo, q_gft4_hi, [x_tbl4], #32
prfb pldl2keep, p0, [x_tbl3]
prfb pldl2keep, p0, [x_tbl4]
/* dest 2 */
tbl z_gft2_lo.b, {z_gft2_lo.b}, z_src_lo.b
tbl z_gft2_hi.b, {z_gft2_hi.b}, z_src_hi.b
eor z_dest2.d, z_gft2_lo.d, z_dest2.d
eor z_dest2.d, z_dest2.d, z_gft2_hi.d
/* dest 3 */
tbl z_gft3_lo.b, {z_gft3_lo.b}, z_src_lo.b
tbl z_gft3_hi.b, {z_gft3_hi.b}, z_src_hi.b
eor z_dest3.d, z_gft3_lo.d, z_dest3.d
eor z_dest3.d, z_dest3.d, z_gft3_hi.d
ldp q_gft5_lo, q_gft5_hi, [x_tbl5], #32
ldp q_gft6_lo, q_gft6_hi, [x_tbl6], #32
prfb pldl2keep, p0, [x_tbl5]
prfb pldl2keep, p0, [x_tbl6]
/* dest 4 */
tbl z_gft4_lo.b, {z_gft4_lo.b}, z_src_lo.b
tbl z_gft4_hi.b, {z_gft4_hi.b}, z_src_hi.b
eor z_dest4.d, z_gft4_lo.d, z_dest4.d
eor z_dest4.d, z_dest4.d, z_gft4_hi.d
/* dest 5 */
tbl z_gft5_lo.b, {z_gft5_lo.b}, z_src_lo.b
tbl z_gft5_hi.b, {z_gft5_hi.b}, z_src_hi.b
eor z_dest5.d, z_gft5_lo.d, z_dest5.d
eor z_dest5.d, z_dest5.d, z_gft5_hi.d
/* dest 6 */
tbl z_gft6_lo.b, {z_gft6_lo.b}, z_src_lo.b
tbl z_gft6_hi.b, {z_gft6_hi.b}, z_src_hi.b
eor z_dest6.d, z_gft6_lo.d, z_dest6.d
eor z_dest6.d, z_dest6.d, z_gft6_hi.d
cmp x_vec_i, x_vec
blt .Lloopsve_vl_vects
/* end of Loop 2 */
/* store dest data, governed by p0 */
st1b z_dest1.b, p0, [x_dest1, x_pos]
st1b z_dest2.b, p0, [x_dest2, x_pos]
st1b z_dest3.b, p0, [x_dest3, x_pos]
st1b z_dest4.b, p0, [x_dest4, x_pos]
st1b z_dest5.b, p0, [x_dest5, x_pos]
st1b z_dest6.b, p0, [x_dest6, x_pos]
/* increment one vector length */
incb x_pos
b .Lloopsve_vl
/* end of Loop 1 */
.return_pass:
/* restore r19..r29 */
ldr x21, [sp, #16]
ldp x19, x20, [sp]
add sp, sp, #32
mov w_ret, #0
ret
.return_fail:
mov w_ret, #1
ret
Reference in New Issue Copy Permalink