From 7cd26d4a1b14bc4bf7c61496803416ab3d84791f Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Wed, 9 Jan 2019 18:35:17 +0000 Subject: COMPMID-1867: Add NEON/SVE GEMM Hybrid kernels. Change-Id: Ib40a9921e7f9a6a8be6c38872d6b3a0f24ed0cd3 Reviewed-on: https://review.mlplatform.org/515 Reviewed-by: Anthony Barbier Tested-by: Arm Jenkins --- .../sve_interleaved_fp16_mla_3VLx8/generic.cpp | 48 ++++++++++------------ 1 file changed, 21 insertions(+), 27 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_3VLx8') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_3VLx8/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_3VLx8/generic.cpp index 92ec888244..517895ca7f 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_3VLx8/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_3VLx8/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 Arm Limited. + * Copyright (c) 2018-2019 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -48,24 +48,24 @@ void sve_interleaved_fp16_mla_3VLx8(const __fp16 *Apanel, const __fp16 *Bpanel, "mov z8.h, #0\n" "ptrue p0.h\n" "mov z9.h, #0\n" - "ld1rqh z0.h, p0/z, [%[a_ptr]]\n" "mov z10.h, #0\n" - "ld1h z2.h, p0/z, [%[b_ptr]]\n" "mov z11.h, #0\n" - "ld1h z3.h, p0/z, [%[b_ptr], #1, MUL VL]\n" "mov z12.h, #0\n" - "ld1h z4.h, p0/z, [%[b_ptr], #2, MUL VL]\n" + "ld1rqh z0.h, p0/z, [%[a_ptr]]\n" "mov z13.h, #0\n" - "ld1h z5.h, p0/z, [%[b_ptr], #3, MUL VL]\n" + "ld1h z2.h, p0/z, [%[b_ptr]]\n" "mov z14.h, #0\n" - "ld1h z6.h, p0/z, [%[b_ptr], #4, MUL VL]\n" + "ld1h z3.h, p0/z, [%[b_ptr], #1, MUL VL]\n" "mov z15.h, #0\n" - "add %[a_ptr], %[a_ptr], #0x20\n" + "ld1h z4.h, p0/z, [%[b_ptr], #2, MUL VL]\n" "mov z16.h, #0\n" - "addvl %[b_ptr], %[b_ptr], #6\n" + "ld1h z5.h, p0/z, [%[b_ptr], #3, MUL VL]\n" "mov z17.h, #0\n" + "ld1h z6.h, p0/z, [%[b_ptr], #4, MUL VL]\n" "mov z18.h, #0\n" + "add %[a_ptr], %[a_ptr], #0x20\n" "mov z19.h, #0\n" + "addvl %[b_ptr], %[b_ptr], #6\n" "mov z20.h, #0\n" "mov z21.h, #0\n" "mov z22.h, #0\n" @@ -199,37 +199,31 @@ void sve_interleaved_fp16_mla_3VLx8(const __fp16 *Apanel, const __fp16 *Bpanel, "fmla z30.h, z7.h, z1.h[6]\n" "fmla z31.h, z7.h, z1.h[7]\n" "fmla z8.h, z2.h, z0.h[0]\n" - "st1h z8.h, p0, [%[c_ptr]]\n" "fmla z9.h, z2.h, z0.h[1]\n" "fmla z10.h, z2.h, z0.h[2]\n" "fmla z11.h, z2.h, z0.h[3]\n" "fmla z12.h, z2.h, z0.h[4]\n" + "st1h z8.h, p0, [%[c_ptr]]\n" "fmla z13.h, z2.h, z0.h[5]\n" "fmla z14.h, z2.h, z0.h[6]\n" "fmla z15.h, z2.h, z0.h[7]\n" "fmla z16.h, z3.h, z0.h[0]\n" - "st1h z16.h, p0, [%[c_ptr], #1, MUL VL]\n" "fmla z17.h, z3.h, z0.h[1]\n" "fmla z18.h, z3.h, z0.h[2]\n" "fmla z19.h, z3.h, z0.h[3]\n" "fmla z20.h, z3.h, z0.h[4]\n" + "st1h z16.h, p0, [%[c_ptr], #1, MUL VL]\n" "fmla z21.h, z3.h, z0.h[5]\n" "fmla z22.h, z3.h, z0.h[6]\n" "fmla z23.h, z3.h, z0.h[7]\n" "fmla z24.h, z4.h, z0.h[0]\n" - "st1h z24.h, p0, [%[c_ptr], #2, MUL VL]\n" "fmla z25.h, z4.h, z0.h[1]\n" - "st1h z9.h, p0, [%[c_ptr], #3, MUL VL]\n" "fmla z26.h, z4.h, z0.h[2]\n" - "st1h z17.h, p0, [%[c_ptr], #4, MUL VL]\n" "fmla z27.h, z4.h, z0.h[3]\n" - "st1h z25.h, p0, [%[c_ptr], #5, MUL VL]\n" "fmla z28.h, z4.h, z0.h[4]\n" - "st1h z10.h, p0, [%[c_ptr], #6, MUL VL]\n" + "st1h z24.h, p0, [%[c_ptr], #2, MUL VL]\n" "fmla z29.h, z4.h, z0.h[5]\n" - "st1h z18.h, p0, [%[c_ptr], #7, MUL VL]\n" "fmla z30.h, z4.h, z0.h[6]\n" - "addvl %[c_ptr], %[c_ptr], #16\n" "fmla z31.h, z4.h, z0.h[7]\n" "b 4f\n" "3:\n" @@ -260,39 +254,39 @@ void sve_interleaved_fp16_mla_3VLx8(const __fp16 *Apanel, const __fp16 *Bpanel, "fmla z30.h, z4.h, z0.h[6]\n" "fmla z31.h, z4.h, z0.h[7]\n" "fmla z8.h, z5.h, z1.h[0]\n" - "st1h z8.h, p0, [%[c_ptr]]\n" "fmla z9.h, z5.h, z1.h[1]\n" "fmla z10.h, z5.h, z1.h[2]\n" "fmla z11.h, z5.h, z1.h[3]\n" "fmla z12.h, z5.h, z1.h[4]\n" + "st1h z8.h, p0, [%[c_ptr]]\n" "fmla z13.h, z5.h, z1.h[5]\n" "fmla z14.h, z5.h, z1.h[6]\n" "fmla z15.h, z5.h, z1.h[7]\n" "fmla z16.h, z6.h, z1.h[0]\n" - "st1h z16.h, p0, [%[c_ptr], #1, MUL VL]\n" "fmla z17.h, z6.h, z1.h[1]\n" "fmla z18.h, z6.h, z1.h[2]\n" "fmla z19.h, z6.h, z1.h[3]\n" "fmla z20.h, z6.h, z1.h[4]\n" + "st1h z16.h, p0, [%[c_ptr], #1, MUL VL]\n" "fmla z21.h, z6.h, z1.h[5]\n" "fmla z22.h, z6.h, z1.h[6]\n" "fmla z23.h, z6.h, z1.h[7]\n" "fmla z24.h, z7.h, z1.h[0]\n" - "st1h z24.h, p0, [%[c_ptr], #2, MUL VL]\n" "fmla z25.h, z7.h, z1.h[1]\n" - "st1h z9.h, p0, [%[c_ptr], #3, MUL VL]\n" "fmla z26.h, z7.h, z1.h[2]\n" - "st1h z17.h, p0, [%[c_ptr], #4, MUL VL]\n" "fmla z27.h, z7.h, z1.h[3]\n" - "st1h z25.h, p0, [%[c_ptr], #5, MUL VL]\n" "fmla z28.h, z7.h, z1.h[4]\n" - "st1h z10.h, p0, [%[c_ptr], #6, MUL VL]\n" + "st1h z24.h, p0, [%[c_ptr], #2, MUL VL]\n" "fmla z29.h, z7.h, z1.h[5]\n" - "st1h z18.h, p0, [%[c_ptr], #7, MUL VL]\n" "fmla z30.h, z7.h, z1.h[6]\n" - "addvl %[c_ptr], %[c_ptr], #16\n" "fmla z31.h, z7.h, z1.h[7]\n" "4:\n" + "st1h z9.h, p0, [%[c_ptr], #3, MUL VL]\n" + "st1h z17.h, p0, [%[c_ptr], #4, MUL VL]\n" + "st1h z25.h, p0, [%[c_ptr], #5, MUL VL]\n" + "st1h z10.h, p0, [%[c_ptr], #6, MUL VL]\n" + "st1h z18.h, p0, [%[c_ptr], #7, MUL VL]\n" + "addvl %[c_ptr], %[c_ptr], #16\n" "st1h z26.h, p0, [%[c_ptr], #-8, MUL VL]\n" "st1h z11.h, p0, [%[c_ptr], #-7, MUL VL]\n" "st1h z19.h, p0, [%[c_ptr], #-6, MUL VL]\n" -- cgit v1.2.1