From ba209750abc1ac7e42bab9fef5db284384d70fb3 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Thu, 15 Dec 2022 12:39:29 +0000 Subject: Update CPU kernels to remove x19 Resolves: COMPMID-5805 Signed-off-by: Michael Tyler Change-Id: I250f64531e209625e4ff176dd5a552c1c34bc484 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8909 Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Reviewed-by: Gunes Bayir Reviewed-by: Viet-Hoa Do Benchmark: Arm Jenkins --- .../generic.cpp | 120 ++++++++++----------- 1 file changed, 60 insertions(+), 60 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp index cda3d4248a..838cd3406c 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2022 Arm Limited. + * Copyright (c) 2021-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -63,80 +63,80 @@ void sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst_impl( pad_left, pad_top, pad_right, pad_bottom); __asm__ __volatile__( - "ldr x14, [%x[args], %[offsetof_n_channels]]\n" - "ldr x20, [%x[args], %[offsetof_outptrs]]\n" - "mov x13, #0x0\n" - "whilelt p2.h, x13, x14\n" - "ldr x19, [%x[args], %[offsetof_inptrs]]\n" - "ldp x12, x11, [x20, #0x0]\n" + "ldr x15, [%x[args], %[offsetof_n_channels]]\n" + "ldr x21, [%x[args], %[offsetof_outptrs]]\n" + "mov x14, #0x0\n" + "whilelt p2.h, x14, x15\n" + "ldr x20, [%x[args], %[offsetof_inptrs]]\n" + "ldp x13, x12, [x21, #0x0]\n" "ptrue p1.b\n" - "mov x10, #0x0\n" - "ldp x9, x28, [x20, #0x10]\n" - "ldp x27, x26, [x19, #0x0]\n" - "ldp x25, x24, [x19, #0x10]\n" - "ldp x23, x22, [x19, #0x20]\n" - "ldp x21, x20, [x19, #0x30]\n" - "ldr x19, [x19, #0x40]\n" - "ld1h { z31.h }, p2/Z, [x26, x13, LSL #1]\n" - "ld1h { z30.h }, p2/Z, [x23, x13, LSL #1]\n" - "ld1h { z29.h }, p2/Z, [x20, x13, LSL #1]\n" - "ld1h { z28.h }, p2/Z, [x24, x13, LSL #1]\n" - "ld1h { z27.h }, p2/Z, [x27, x13, LSL #1]\n" - "ld1h { z26.h }, p2/Z, [x22, x13, LSL #1]\n" - "ld1h { z25.h }, p2/Z, [x25, x13, LSL #1]\n" - "ld1h { z24.h }, p2/Z, [x21, x13, LSL #1]\n" - "ld1h { z23.h }, p2/Z, [x19, x13, LSL #1]\n" - "incw x13\n" - "whilelt p2.h, x13, x14\n" + "mov x11, #0x0\n" + "ldp x10, x9, [x21, #0x10]\n" + "ldp x28, x27, [x20, #0x0]\n" + "ldp x26, x25, [x20, #0x10]\n" + "ldp x24, x23, [x20, #0x20]\n" + "ldp x22, x21, [x20, #0x30]\n" + "ldr x20, [x20, #0x40]\n" + "ld1h { z31.h }, p2/Z, [x27, x14, LSL #1]\n" + "ld1h { z30.h }, p2/Z, [x24, x14, LSL #1]\n" + "ld1h { z29.h }, p2/Z, [x21, x14, LSL #1]\n" + "ld1h { z28.h }, p2/Z, [x25, x14, LSL #1]\n" + "ld1h { z27.h }, p2/Z, [x28, x14, LSL #1]\n" + "ld1h { z26.h }, p2/Z, [x26, x14, LSL #1]\n" + "ld1h { z25.h }, p2/Z, [x23, x14, LSL #1]\n" + "ld1h { z24.h }, p2/Z, [x22, x14, LSL #1]\n" + "ld1h { z23.h }, p2/Z, [x20, x14, LSL #1]\n" + "incw x14\n" + "whilelt p2.h, x14, x15\n" "b.none 2f\n" "1:" // Vector: Loop "movprfx z22, z31\n fmax z22.h, p1/M, z22.h, z30.h\n" "movprfx z21, z30\n fmax z21.h, p1/M, z21.h, z29.h\n" - "ld1h { z31.h }, p2/Z, [x26, x13, LSL #1]\n" - "ld1h { z30.h }, p2/Z, [x23, x13, LSL #1]\n" + "ld1h { z31.h }, p2/Z, [x27, x14, LSL #1]\n" + "ld1h { z30.h }, p2/Z, [x24, x14, LSL #1]\n" "movprfx z20, z28\n fmax z20.h, p1/M, z20.h, z27.h\n" - "movprfx z17, z26\n fmax z17.h, p1/M, z17.h, z25.h\n" - "ld1h { z29.h }, p2/Z, [x20, x13, LSL #1]\n" - "ld1h { z27.h }, p2/Z, [x27, x13, LSL #1]\n" - "movprfx z19, z24\n fmax z19.h, p1/M, z19.h, z28.h\n" - "movprfx z18, z26\n fmax z18.h, p1/M, z18.h, z23.h\n" - "ld1h { z28.h }, p2/Z, [x24, x13, LSL #1]\n" - "ld1h { z26.h }, p2/Z, [x22, x13, LSL #1]\n" - "ld1h { z25.h }, p2/Z, [x25, x13, LSL #1]\n" - "ld1h { z24.h }, p2/Z, [x21, x13, LSL #1]\n" - "whilelt p0.h, x10, x14\n" + "movprfx z19, z26\n fmax z19.h, p1/M, z19.h, z25.h\n" + "ld1h { z29.h }, p2/Z, [x21, x14, LSL #1]\n" + "ld1h { z27.h }, p2/Z, [x28, x14, LSL #1]\n" + "movprfx z17, z28\n fmax z17.h, p1/M, z17.h, z24.h\n" + "movprfx z18, z25\n fmax z18.h, p1/M, z18.h, z23.h\n" + "ld1h { z28.h }, p2/Z, [x25, x14, LSL #1]\n" + "ld1h { z26.h }, p2/Z, [x26, x14, LSL #1]\n" + "ld1h { z25.h }, p2/Z, [x23, x14, LSL #1]\n" + "ld1h { z24.h }, p2/Z, [x22, x14, LSL #1]\n" + "whilelt p0.h, x11, x15\n" "movprfx z16, z22\n fmax z16.h, p1/M, z16.h, z20.h\n" - "ld1h { z23.h }, p2/Z, [x19, x13, LSL #1]\n" - "incw x13\n" - "whilelt p2.h, x13, x14\n" - "st1h { z16.h }, p0, [x12, x10, LSL #1]\n" - "movprfx z16, z17\n fmax z16.h, p1/M, z16.h, z22.h\n" - "movprfx z17, z21\n fmax z17.h, p1/M, z17.h, z19.h\n" - "st1h { z16.h }, p0, [x11, x10, LSL #1]\n" - "movprfx z16, z21\n fmax z16.h, p1/M, z16.h, z18.h\n" - "st1h { z17.h }, p0, [x9, x10, LSL #1]\n" - "st1h { z16.h }, p0, [x28, x10, LSL #1]\n" - "incw x10\n" + "ld1h { z23.h }, p2/Z, [x20, x14, LSL #1]\n" + "incw x14\n" + "whilelt p2.h, x14, x15\n" + "st1h { z16.h }, p0, [x13, x11, LSL #1]\n" + "movprfx z16, z19\n fmax z16.h, p1/M, z16.h, z22.h\n" + "fmax z17.h, p1/M, z17.h, z21.h\n" + "st1h { z16.h }, p0, [x12, x11, LSL #1]\n" + "movprfx z16, z18\n fmax z16.h, p1/M, z16.h, z21.h\n" + "st1h { z17.h }, p0, [x10, x11, LSL #1]\n" + "st1h { z16.h }, p0, [x9, x11, LSL #1]\n" + "incw x11\n" "b.any 1b\n" "2:" // Vector: Tail "movprfx z22, z31\n fmax z22.h, p1/M, z22.h, z30.h\n" "movprfx z21, z30\n fmax z21.h, p1/M, z21.h, z29.h\n" "movprfx z20, z28\n fmax z20.h, p1/M, z20.h, z27.h\n" - "movprfx z17, z26\n fmax z17.h, p1/M, z17.h, z25.h\n" - "movprfx z19, z24\n fmax z19.h, p1/M, z19.h, z28.h\n" - "movprfx z18, z26\n fmax z18.h, p1/M, z18.h, z23.h\n" - "whilelt p0.h, x10, x14\n" + "movprfx z19, z26\n fmax z19.h, p1/M, z19.h, z25.h\n" + "movprfx z17, z28\n fmax z17.h, p1/M, z17.h, z24.h\n" + "movprfx z18, z25\n fmax z18.h, p1/M, z18.h, z23.h\n" + "whilelt p0.h, x11, x15\n" "movprfx z16, z22\n fmax z16.h, p1/M, z16.h, z20.h\n" - "st1h { z16.h }, p0, [x12, x10, LSL #1]\n" - "movprfx z16, z17\n fmax z16.h, p1/M, z16.h, z22.h\n" - "movprfx z17, z21\n fmax z17.h, p1/M, z17.h, z19.h\n" - "st1h { z16.h }, p0, [x11, x10, LSL #1]\n" - "movprfx z16, z21\n fmax z16.h, p1/M, z16.h, z18.h\n" - "st1h { z17.h }, p0, [x9, x10, LSL #1]\n" - "st1h { z16.h }, p0, [x28, x10, LSL #1]\n" + "st1h { z16.h }, p0, [x13, x11, LSL #1]\n" + "movprfx z16, z19\n fmax z16.h, p1/M, z16.h, z22.h\n" + "fmax z17.h, p1/M, z17.h, z21.h\n" + "st1h { z16.h }, p0, [x12, x11, LSL #1]\n" + "movprfx z16, z18\n fmax z16.h, p1/M, z16.h, z21.h\n" + "st1h { z17.h }, p0, [x10, x11, LSL #1]\n" + "st1h { z16.h }, p0, [x9, x11, LSL #1]\n" : : [args] "r" (&args), [offsetof_inptrs] "I" (offsetof(KernelArgs, inptrs)), [offsetof_n_channels] "I" (offsetof(KernelArgs, n_channels)), [offsetof_outptrs] "I" (offsetof(KernelArgs, outptrs)) - : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x13", "x14", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } -- cgit v1.2.1