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 | 186 ++++++++++----------- 1 file changed, 93 insertions(+), 93 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_generic_depthfirst/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_generic_depthfirst/generic.cpp index 3b07befc23..9f1f9e7377 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_generic_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2022 Arm Limited. + * Copyright (c) 2021-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -40,82 +40,82 @@ void sve_fp16_nhwc_max_generic_depthfirst_impl( ) { __asm__ __volatile__( - "mov x28, #0x0\n" - "cnth x27\n" - "cnth x26, ALL, MUL #2\n" - "cnth x25, ALL, MUL #3\n" - "whilelt p4.h, x28, %x[n_channels]\n" - "whilelt p3.h, x27, %x[n_channels]\n" - "whilelt p2.h, x26, %x[n_channels]\n" - "whilelt p1.h, x25, %x[n_channels]\n" + "mov x9, #0x0\n" + "cnth x28\n" + "cnth x27, ALL, MUL #2\n" + "cnth x26, ALL, MUL #3\n" + "whilelt p4.h, x9, %x[n_channels]\n" + "whilelt p3.h, x28, %x[n_channels]\n" + "whilelt p2.h, x27, %x[n_channels]\n" + "whilelt p1.h, x26, %x[n_channels]\n" "ptrue p0.b\n" "b.none 7f\n" "1:" // 4-vectors of channels - "lsr x24, %x[n_valid_cells], #0x2\n" + "lsr x25, %x[n_valid_cells], #0x2\n" "mov z8.h, #0xfc00\n" "mov z7.h, #0xfc00\n" - "mov x19, %x[inptrs]\n" + "mov x20, %x[inptrs]\n" "mov z6.h, #0xfc00\n" "mov z5.h, #0xfc00\n" - "cbz x24, 4f\n" - "ldp x23, x22, [x19, #0x0]\n" - "ldp x21, x20, [x19, #0x10]\n" - "subs x24, x24, #0x1\n" - "add x19, x19, #0x20\n" - "ld1h { z4.h }, p4/Z, [x23, x28, LSL #1]\n" - "ld1h { z3.h }, p4/Z, [x22, x28, LSL #1]\n" - "ld1h { z2.h }, p4/Z, [x21, x28, LSL #1]\n" - "ld1h { z1.h }, p4/Z, [x20, x28, LSL #1]\n" - "ld1h { z0.h }, p3/Z, [x23, x27, LSL #1]\n" - "ld1h { z31.h }, p3/Z, [x22, x27, LSL #1]\n" - "ld1h { z22.h }, p3/Z, [x21, x27, LSL #1]\n" - "ld1h { z30.h }, p3/Z, [x20, x27, LSL #1]\n" - "ld1h { z29.h }, p2/Z, [x23, x26, LSL #1]\n" - "ld1h { z28.h }, p2/Z, [x22, x26, LSL #1]\n" - "ld1h { z21.h }, p2/Z, [x21, x26, LSL #1]\n" - "ld1h { z27.h }, p2/Z, [x20, x26, LSL #1]\n" - "ld1h { z26.h }, p1/Z, [x23, x25, LSL #1]\n" - "ld1h { z25.h }, p1/Z, [x22, x25, LSL #1]\n" - "ld1h { z20.h }, p1/Z, [x21, x25, LSL #1]\n" - "ld1h { z24.h }, p1/Z, [x20, x25, LSL #1]\n" + "cbz x25, 4f\n" + "ldp x24, x23, [x20, #0x0]\n" + "ldp x22, x21, [x20, #0x10]\n" + "subs x25, x25, #0x1\n" + "add x20, x20, #0x20\n" + "ld1h { z4.h }, p4/Z, [x24, x9, LSL #1]\n" + "ld1h { z3.h }, p4/Z, [x23, x9, LSL #1]\n" + "ld1h { z2.h }, p4/Z, [x22, x9, LSL #1]\n" + "ld1h { z1.h }, p4/Z, [x21, x9, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x24, x28, LSL #1]\n" + "ld1h { z31.h }, p3/Z, [x23, x28, LSL #1]\n" + "ld1h { z22.h }, p3/Z, [x22, x28, LSL #1]\n" + "ld1h { z30.h }, p3/Z, [x21, x28, LSL #1]\n" + "ld1h { z29.h }, p2/Z, [x24, x27, LSL #1]\n" + "ld1h { z28.h }, p2/Z, [x23, x27, LSL #1]\n" + "ld1h { z21.h }, p2/Z, [x22, x27, LSL #1]\n" + "ld1h { z27.h }, p2/Z, [x21, x27, LSL #1]\n" + "ld1h { z26.h }, p1/Z, [x24, x26, LSL #1]\n" + "ld1h { z25.h }, p1/Z, [x23, x26, LSL #1]\n" + "ld1h { z20.h }, p1/Z, [x22, x26, LSL #1]\n" + "ld1h { z24.h }, p1/Z, [x21, x26, LSL #1]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop "movprfx z19, z4\n fmax z19.h, p0/M, z19.h, z3.h\n" "movprfx z23, z2\n fmax z23.h, p0/M, z23.h, z1.h\n" - "ldp x23, x22, [x19, #0x0]\n" - "ldp x21, x20, [x19, #0x10]\n" + "ldp x24, x23, [x20, #0x0]\n" + "ldp x22, x21, [x20, #0x10]\n" "movprfx z18, z0\n fmax z18.h, p0/M, z18.h, z31.h\n" "fmax z22.h, p0/M, z22.h, z30.h\n" - "ld1h { z4.h }, p4/Z, [x23, x28, LSL #1]\n" - "ld1h { z3.h }, p4/Z, [x22, x28, LSL #1]\n" + "ld1h { z4.h }, p4/Z, [x24, x9, LSL #1]\n" + "ld1h { z3.h }, p4/Z, [x23, x9, LSL #1]\n" "movprfx z17, z29\n fmax z17.h, p0/M, z17.h, z28.h\n" "fmax z21.h, p0/M, z21.h, z27.h\n" - "ld1h { z2.h }, p4/Z, [x21, x28, LSL #1]\n" - "ld1h { z1.h }, p4/Z, [x20, x28, LSL #1]\n" + "ld1h { z2.h }, p4/Z, [x22, x9, LSL #1]\n" + "ld1h { z1.h }, p4/Z, [x21, x9, LSL #1]\n" "movprfx z16, z26\n fmax z16.h, p0/M, z16.h, z25.h\n" "fmax z20.h, p0/M, z20.h, z24.h\n" - "ld1h { z0.h }, p3/Z, [x23, x27, LSL #1]\n" - "ld1h { z31.h }, p3/Z, [x22, x27, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x24, x28, LSL #1]\n" + "ld1h { z31.h }, p3/Z, [x23, x28, LSL #1]\n" "fmax z19.h, p0/M, z19.h, z23.h\n" "fmax z18.h, p0/M, z18.h, z22.h\n" - "ld1h { z22.h }, p3/Z, [x21, x27, LSL #1]\n" - "ld1h { z30.h }, p3/Z, [x20, x27, LSL #1]\n" + "ld1h { z22.h }, p3/Z, [x22, x28, LSL #1]\n" + "ld1h { z30.h }, p3/Z, [x21, x28, LSL #1]\n" "fmax z17.h, p0/M, z17.h, z21.h\n" "fmax z16.h, p0/M, z16.h, z20.h\n" - "ld1h { z29.h }, p2/Z, [x23, x26, LSL #1]\n" - "ld1h { z28.h }, p2/Z, [x22, x26, LSL #1]\n" - "subs x24, x24, #0x1\n" + "ld1h { z29.h }, p2/Z, [x24, x27, LSL #1]\n" + "ld1h { z28.h }, p2/Z, [x23, x27, LSL #1]\n" + "subs x25, x25, #0x1\n" "fmax z8.h, p0/M, z8.h, z19.h\n" - "ld1h { z21.h }, p2/Z, [x21, x26, LSL #1]\n" - "ld1h { z27.h }, p2/Z, [x20, x26, LSL #1]\n" + "ld1h { z21.h }, p2/Z, [x22, x27, LSL #1]\n" + "ld1h { z27.h }, p2/Z, [x21, x27, LSL #1]\n" "fmax z7.h, p0/M, z7.h, z18.h\n" "fmax z6.h, p0/M, z6.h, z17.h\n" - "ld1h { z26.h }, p1/Z, [x23, x25, LSL #1]\n" - "ld1h { z25.h }, p1/Z, [x22, x25, LSL #1]\n" + "ld1h { z26.h }, p1/Z, [x24, x26, LSL #1]\n" + "ld1h { z25.h }, p1/Z, [x23, x26, LSL #1]\n" "fmax z5.h, p0/M, z5.h, z16.h\n" - "add x19, x19, #0x20\n" - "ld1h { z20.h }, p1/Z, [x21, x25, LSL #1]\n" - "ld1h { z24.h }, p1/Z, [x20, x25, LSL #1]\n" + "add x20, x20, #0x20\n" + "ld1h { z20.h }, p1/Z, [x22, x26, LSL #1]\n" + "ld1h { z24.h }, p1/Z, [x21, x26, LSL #1]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail "movprfx z19, z4\n fmax z19.h, p0/M, z19.h, z3.h\n" @@ -135,61 +135,61 @@ void sve_fp16_nhwc_max_generic_depthfirst_impl( "fmax z6.h, p0/M, z6.h, z17.h\n" "fmax z5.h, p0/M, z5.h, z16.h\n" "4:" // 4-vectors of channels: After loop - "ands x20, %x[n_valid_cells], #0x3\n" + "ands x21, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x23, [x19], #0x8\n" - "ld1h { z4.h }, p4/Z, [x23, x28, LSL #1]\n" - "subs x20, x20, #0x1\n" + "ldr x24, [x20], #0x8\n" + "ld1h { z4.h }, p4/Z, [x24, x9, LSL #1]\n" + "subs x21, x21, #0x1\n" "fmax z8.h, p0/M, z8.h, z4.h\n" - "ld1h { z0.h }, p3/Z, [x23, x27, LSL #1]\n" - "ld1h { z29.h }, p2/Z, [x23, x26, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x24, x28, LSL #1]\n" + "ld1h { z29.h }, p2/Z, [x24, x27, LSL #1]\n" "fmax z7.h, p0/M, z7.h, z0.h\n" "fmax z6.h, p0/M, z6.h, z29.h\n" - "ld1h { z26.h }, p1/Z, [x23, x25, LSL #1]\n" + "ld1h { z26.h }, p1/Z, [x24, x26, LSL #1]\n" "fmax z5.h, p0/M, z5.h, z26.h\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - "st1h { z8.h }, p4, [%x[outptr], x28, LSL #1]\n" + "st1h { z8.h }, p4, [%x[outptr], x9, LSL #1]\n" + "inch x9, ALL, MUL #4\n" + "st1h { z7.h }, p3, [%x[outptr], x28, LSL #1]\n" "inch x28, ALL, MUL #4\n" - "st1h { z7.h }, p3, [%x[outptr], x27, LSL #1]\n" + "st1h { z6.h }, p2, [%x[outptr], x27, LSL #1]\n" "inch x27, ALL, MUL #4\n" - "st1h { z6.h }, p2, [%x[outptr], x26, LSL #1]\n" + "st1h { z5.h }, p1, [%x[outptr], x26, LSL #1]\n" "inch x26, ALL, MUL #4\n" - "st1h { z5.h }, p1, [%x[outptr], x25, LSL #1]\n" - "inch x25, ALL, MUL #4\n" - "whilelt p1.h, x25, %x[n_channels]\n" + "whilelt p1.h, x26, %x[n_channels]\n" "b.any 1b\n" "7:" // Single vector of channels - "whilelt p4.h, x28, %x[n_channels]\n" + "whilelt p4.h, x9, %x[n_channels]\n" "b.none 14f\n" "8:" // Single vector of channels: Loop - "lsr x24, %x[n_valid_cells], #0x2\n" + "lsr x25, %x[n_valid_cells], #0x2\n" "mov z8.h, #0xfc00\n" - "mov x19, %x[inptrs]\n" - "cbz x24, 11f\n" - "ldp x23, x22, [x19, #0x0]\n" - "ldp x21, x20, [x19, #0x10]\n" - "subs x24, x24, #0x1\n" - "add x19, x19, #0x20\n" - "ld1h { z4.h }, p4/Z, [x23, x28, LSL #1]\n" - "ld1h { z3.h }, p4/Z, [x22, x28, LSL #1]\n" - "ld1h { z2.h }, p4/Z, [x21, x28, LSL #1]\n" - "ld1h { z1.h }, p4/Z, [x20, x28, LSL #1]\n" + "mov x20, %x[inptrs]\n" + "cbz x25, 11f\n" + "ldp x24, x23, [x20, #0x0]\n" + "ldp x22, x21, [x20, #0x10]\n" + "subs x25, x25, #0x1\n" + "add x20, x20, #0x20\n" + "ld1h { z4.h }, p4/Z, [x24, x9, LSL #1]\n" + "ld1h { z3.h }, p4/Z, [x23, x9, LSL #1]\n" + "ld1h { z2.h }, p4/Z, [x22, x9, LSL #1]\n" + "ld1h { z1.h }, p4/Z, [x21, x9, LSL #1]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop "movprfx z19, z4\n fmax z19.h, p0/M, z19.h, z3.h\n" "movprfx z23, z2\n fmax z23.h, p0/M, z23.h, z1.h\n" - "ldp x23, x22, [x19, #0x0]\n" - "ldp x21, x20, [x19, #0x10]\n" + "ldp x24, x23, [x20, #0x0]\n" + "ldp x22, x21, [x20, #0x10]\n" "fmax z19.h, p0/M, z19.h, z23.h\n" - "subs x24, x24, #0x1\n" - "ld1h { z4.h }, p4/Z, [x23, x28, LSL #1]\n" - "ld1h { z3.h }, p4/Z, [x22, x28, LSL #1]\n" + "subs x25, x25, #0x1\n" + "ld1h { z4.h }, p4/Z, [x24, x9, LSL #1]\n" + "ld1h { z3.h }, p4/Z, [x23, x9, LSL #1]\n" "fmax z8.h, p0/M, z8.h, z19.h\n" - "add x19, x19, #0x20\n" - "ld1h { z2.h }, p4/Z, [x21, x28, LSL #1]\n" - "ld1h { z1.h }, p4/Z, [x20, x28, LSL #1]\n" + "add x20, x20, #0x20\n" + "ld1h { z2.h }, p4/Z, [x22, x9, LSL #1]\n" + "ld1h { z1.h }, p4/Z, [x21, x9, LSL #1]\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail "movprfx z19, z4\n fmax z19.h, p0/M, z19.h, z3.h\n" @@ -197,23 +197,23 @@ void sve_fp16_nhwc_max_generic_depthfirst_impl( "fmax z19.h, p0/M, z19.h, z23.h\n" "fmax z8.h, p0/M, z8.h, z19.h\n" "11:" // Single vector of channels: Loop: After loop - "ands x20, %x[n_valid_cells], #0x3\n" + "ands x21, %x[n_valid_cells], #0x3\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop - "ldr x23, [x19], #0x8\n" - "ld1h { z4.h }, p4/Z, [x23, x28, LSL #1]\n" - "subs x20, x20, #0x1\n" + "ldr x24, [x20], #0x8\n" + "ld1h { z4.h }, p4/Z, [x24, x9, LSL #1]\n" + "subs x21, x21, #0x1\n" "fmax z8.h, p0/M, z8.h, z4.h\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - "st1h { z8.h }, p4, [%x[outptr], x28, LSL #1]\n" - "inch x28\n" - "whilelt p4.h, x28, %x[n_channels]\n" + "st1h { z8.h }, p4, [%x[outptr], x9, LSL #1]\n" + "inch x9\n" + "whilelt p4.h, x9, %x[n_channels]\n" "b.any 8b\n" "14:" // End : : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [n_valid_cells] "r" (n_valid_cells), [outptr] "r" (outptr) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } -- cgit v1.2.1