diff options
author | Michael Tyler <michael.tyler@arm.com> | 2023-04-12 17:43:17 +0100 |
---|---|---|
committer | michael.tyler <michael.tyler@arm.com> | 2023-06-05 15:57:58 +0000 |
commit | 74921eee924625426429044decefe3673561b174 (patch) | |
tree | 654da1a95e3d42d6af8ad1ff27bb40d77b1fd8c5 /src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp | |
parent | df5d9878008be9b60586df97ebfff197abb5195e (diff) | |
download | ComputeLibrary-74921eee924625426429044decefe3673561b174.tar.gz |
Update CPU kernel implementations and guard directives
Resolves COMPMID-6023
Change-Id: I868975d14c4f98af6716726feda22405a6a4c891
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp | 120 |
1 files changed, 62 insertions, 58 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp index e5f59d220b..f6e1a75c15 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp @@ -29,8 +29,12 @@ namespace arm_gemm { void sve_interleaved_s8s32_dot_8x3VL( - const int8_t *Apanel, const int8_t *Bpanel, - int32_t *Cpanel, int ablocks, int bblocks, int K) { + const int8_t *Apanel, + const int8_t *Bpanel, + int32_t *Cpanel, + int ablocks, + int bblocks, + int K) { struct KernelArgs { size_t K = {}; @@ -85,10 +89,10 @@ void sve_interleaved_s8s32_dot_8x3VL( "3:" // main loop head "sdot z8.s, z4.b, z0.b[0]\n" "sdot z11.s, z4.b, z0.b[1]\n" - "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #32]\n" + "ld1rqb { z3.b }, p0/Z, [%x[Apanel], #32]\n" "sdot z14.s, z4.b, z0.b[2]\n" "sdot z17.s, z4.b, z0.b[3]\n" - "ld1rqb { z3.b }, p0/Z, [%x[Apanel], #48]\n" + "ld1rqb { z7.b }, p0/Z, [%x[Apanel], #48]\n" "sdot z20.s, z4.b, z1.b[0]\n" "sdot z23.s, z4.b, z1.b[1]\n" "sub x20, x20, #0x2\n" @@ -115,35 +119,35 @@ void sve_interleaved_s8s32_dot_8x3VL( "sdot z25.s, z6.b, z1.b[1]\n" "sdot z28.s, z6.b, z1.b[2]\n" "sdot z31.s, z6.b, z1.b[3]\n" - "ld1b { z6.b }, p0/Z, [x22, #5, MUL VL]\n" + "ld1b { z2.b }, p0/Z, [x22, #5, MUL VL]\n" "addvl x22, x22, #6\n" - "sdot z8.s, z4.b, z2.b[0]\n" - "sdot z11.s, z4.b, z2.b[1]\n" + "sdot z8.s, z4.b, z3.b[0]\n" + "sdot z11.s, z4.b, z3.b[1]\n" "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n" - "sdot z14.s, z4.b, z2.b[2]\n" - "sdot z17.s, z4.b, z2.b[3]\n" - "sdot z20.s, z4.b, z3.b[0]\n" - "sdot z23.s, z4.b, z3.b[1]\n" - "sdot z26.s, z4.b, z3.b[2]\n" - "sdot z29.s, z4.b, z3.b[3]\n" + "sdot z14.s, z4.b, z3.b[2]\n" + "sdot z17.s, z4.b, z3.b[3]\n" + "sdot z20.s, z4.b, z7.b[0]\n" + "sdot z23.s, z4.b, z7.b[1]\n" + "sdot z26.s, z4.b, z7.b[2]\n" + "sdot z29.s, z4.b, z7.b[3]\n" "ld1b { z4.b }, p0/Z, [x22]\n" - "sdot z9.s, z5.b, z2.b[0]\n" - "sdot z12.s, z5.b, z2.b[1]\n" - "sdot z15.s, z5.b, z2.b[2]\n" - "sdot z18.s, z5.b, z2.b[3]\n" - "sdot z21.s, z5.b, z3.b[0]\n" - "sdot z24.s, z5.b, z3.b[1]\n" - "sdot z27.s, z5.b, z3.b[2]\n" - "sdot z30.s, z5.b, z3.b[3]\n" + "sdot z9.s, z5.b, z3.b[0]\n" + "sdot z12.s, z5.b, z3.b[1]\n" + "sdot z15.s, z5.b, z3.b[2]\n" + "sdot z18.s, z5.b, z3.b[3]\n" + "sdot z21.s, z5.b, z7.b[0]\n" + "sdot z24.s, z5.b, z7.b[1]\n" + "sdot z27.s, z5.b, z7.b[2]\n" + "sdot z30.s, z5.b, z7.b[3]\n" "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n" - "sdot z10.s, z6.b, z2.b[0]\n" - "sdot z13.s, z6.b, z2.b[1]\n" - "sdot z16.s, z6.b, z2.b[2]\n" - "sdot z19.s, z6.b, z2.b[3]\n" - "sdot z22.s, z6.b, z3.b[0]\n" - "sdot z25.s, z6.b, z3.b[1]\n" - "sdot z28.s, z6.b, z3.b[2]\n" - "sdot z31.s, z6.b, z3.b[3]\n" + "sdot z10.s, z2.b, z3.b[0]\n" + "sdot z13.s, z2.b, z3.b[1]\n" + "sdot z16.s, z2.b, z3.b[2]\n" + "sdot z19.s, z2.b, z3.b[3]\n" + "sdot z22.s, z2.b, z7.b[0]\n" + "sdot z25.s, z2.b, z7.b[1]\n" + "sdot z28.s, z2.b, z7.b[2]\n" + "sdot z31.s, z2.b, z7.b[3]\n" "ld1b { z6.b }, p0/Z, [x22, #2, MUL VL]\n" "bge 3b\n" "4:" // main loop skip @@ -174,37 +178,37 @@ void sve_interleaved_s8s32_dot_8x3VL( "sdot z28.s, z6.b, z1.b[2]\n" "sdot z31.s, z6.b, z1.b[3]\n" "cbz x20, 5f\n" - "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n" - "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n" + "ld1rqb { z4.b }, p0/Z, [%x[Apanel]]\n" + "ld1rqb { z3.b }, p0/Z, [%x[Apanel], #16]\n" "add %x[Apanel], %x[Apanel], #0x20\n" - "ld1b { z7.b }, p0/Z, [x22]\n" - "ld1b { z4.b }, p0/Z, [x22, #1, MUL VL]\n" - "sdot z8.s, z7.b, z0.b[0]\n" - "ld1b { z5.b }, p0/Z, [x22, #2, MUL VL]\n" - "sdot z11.s, z7.b, z0.b[1]\n" - "sdot z14.s, z7.b, z0.b[2]\n" - "sdot z17.s, z7.b, z0.b[3]\n" - "sdot z20.s, z7.b, z1.b[0]\n" + "ld1b { z2.b }, p0/Z, [x22]\n" + "ld1b { z1.b }, p0/Z, [x22, #1, MUL VL]\n" + "sdot z8.s, z2.b, z4.b[0]\n" + "ld1b { z0.b }, p0/Z, [x22, #2, MUL VL]\n" + "sdot z11.s, z2.b, z4.b[1]\n" + "sdot z14.s, z2.b, z4.b[2]\n" + "sdot z17.s, z2.b, z4.b[3]\n" + "sdot z20.s, z2.b, z3.b[0]\n" "addvl x22, x22, #3\n" - "sdot z23.s, z7.b, z1.b[1]\n" - "sdot z26.s, z7.b, z1.b[2]\n" - "sdot z29.s, z7.b, z1.b[3]\n" - "sdot z9.s, z4.b, z0.b[0]\n" - "sdot z12.s, z4.b, z0.b[1]\n" - "sdot z15.s, z4.b, z0.b[2]\n" - "sdot z18.s, z4.b, z0.b[3]\n" - "sdot z21.s, z4.b, z1.b[0]\n" - "sdot z24.s, z4.b, z1.b[1]\n" - "sdot z27.s, z4.b, z1.b[2]\n" - "sdot z30.s, z4.b, z1.b[3]\n" - "sdot z10.s, z5.b, z0.b[0]\n" - "sdot z13.s, z5.b, z0.b[1]\n" - "sdot z16.s, z5.b, z0.b[2]\n" - "sdot z19.s, z5.b, z0.b[3]\n" - "sdot z22.s, z5.b, z1.b[0]\n" - "sdot z25.s, z5.b, z1.b[1]\n" - "sdot z28.s, z5.b, z1.b[2]\n" - "sdot z31.s, z5.b, z1.b[3]\n" + "sdot z23.s, z2.b, z3.b[1]\n" + "sdot z26.s, z2.b, z3.b[2]\n" + "sdot z29.s, z2.b, z3.b[3]\n" + "sdot z9.s, z1.b, z4.b[0]\n" + "sdot z12.s, z1.b, z4.b[1]\n" + "sdot z15.s, z1.b, z4.b[2]\n" + "sdot z18.s, z1.b, z4.b[3]\n" + "sdot z21.s, z1.b, z3.b[0]\n" + "sdot z24.s, z1.b, z3.b[1]\n" + "sdot z27.s, z1.b, z3.b[2]\n" + "sdot z30.s, z1.b, z3.b[3]\n" + "sdot z10.s, z0.b, z4.b[0]\n" + "sdot z13.s, z0.b, z4.b[1]\n" + "sdot z16.s, z0.b, z4.b[2]\n" + "sdot z19.s, z0.b, z4.b[3]\n" + "sdot z22.s, z0.b, z3.b[0]\n" + "sdot z25.s, z0.b, z3.b[1]\n" + "sdot z28.s, z0.b, z3.b[2]\n" + "sdot z31.s, z0.b, z3.b[3]\n" "5:" // multiply loop done "st1w { z8.s }, p0, [%x[Cpanel]]\n" "subs x23, x23, #0x1\n" |