From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../generic_indirect.cpp | 1188 ++++++++++---------- 1 file changed, 594 insertions(+), 594 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp index 0b04ae064d..3db248924f 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp @@ -25,7 +25,7 @@ #include #include -#if __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) +#if defined(ARM_COMPUTE_ENABLE_SVE) namespace arm_conv { namespace depthwise { @@ -99,616 +99,616 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_indirect_impl( __asm__ __volatile__( "ptrue p3.b\n" - "ldr x17, [%x[params_struct], %[offsetof_args_params]]\n" - "add x16, %x[params_struct], %[offsetof_Args_inptrs]\n" - "ld1w { z15.s }, p3/Z, [x17]\n" - "cntw x15\n" - "mov x14, #0x0\n" - "ld1w { z0.s }, p3/Z, [x17, #1, MUL VL]\n" - "ld1w { z1.s }, p3/Z, [x17, #2, MUL VL]\n" + "ldr x7, [%x[params_struct], %[offsetof_args_params]]\n" + "add x8, %x[params_struct], %[offsetof_Args_inptrs]\n" + "ld1w { z17.s }, p3/Z, [x7]\n" + "cntw x17\n" + "mov x16, #0x0\n" + "ld1w { z0.s }, p3/Z, [x7, #1, MUL VL]\n" + "ld1w { z1.s }, p3/Z, [x7, #2, MUL VL]\n" "whilelt p2.s, XZR, %x[n_channels]\n" - "ld1w { z2.s }, p3/Z, [x17, #3, MUL VL]\n" - "ld1w { z3.s }, p3/Z, [x17, #4, MUL VL]\n" - "cmp x15, %x[n_channels]\n" - "ld1w { z4.s }, p3/Z, [x17, #5, MUL VL]\n" - "ld1w { z5.s }, p3/Z, [x17, #6, MUL VL]\n" - "sub x13, XZR, x15\n" - "ld1w { z6.s }, p3/Z, [x17, #7, MUL VL]\n" - "addvl x17, x17, #16\n" - "ldp x12, x11, [x16, #0x0]\n" - "ldp x10, x9, [x16, #0x10]\n" - "ldr x28, [%x[params_struct], %[offsetof_args_outptrs]]\n" - "ld1rw { z14.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" - "ld1rw { z13.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" - "ld1w { z7.s }, p3/Z, [x17, #-8, MUL VL]\n" - "ld1w { z8.s }, p3/Z, [x17, #-7, MUL VL]\n" - "addvl x17, x17, #-6\n" - "ld1w { z9.s }, p2/Z, [x12, x14, LSL #2]\n" - "ld1w { z10.s }, p2/Z, [x11, x14, LSL #2]\n" - "ld1w { z11.s }, p2/Z, [x10, x14, LSL #2]\n" - "ld1w { z12.s }, p2/Z, [x9, x14, LSL #2]\n" + "ld1w { z2.s }, p3/Z, [x7, #3, MUL VL]\n" + "ld1w { z3.s }, p3/Z, [x7, #4, MUL VL]\n" + "cmp x17, %x[n_channels]\n" + "ld1w { z4.s }, p3/Z, [x7, #5, MUL VL]\n" + "ld1w { z5.s }, p3/Z, [x7, #6, MUL VL]\n" + "sub x15, XZR, x17\n" + "ld1w { z6.s }, p3/Z, [x7, #7, MUL VL]\n" + "addvl x7, x7, #16\n" + "ldp x23, x22, [x8, #0x0]\n" + "ldp x21, x20, [x8, #0x10]\n" + "ldr x14, [%x[params_struct], %[offsetof_args_outptrs]]\n" + "ld1rw { z16.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" + "ld1rw { z19.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" + "ld1w { z7.s }, p3/Z, [x7, #-8, MUL VL]\n" + "ld1w { z8.s }, p3/Z, [x7, #-7, MUL VL]\n" + "addvl x7, x7, #-6\n" + "ld1w { z9.s }, p2/Z, [x23, x16, LSL #2]\n" + "ld1w { z10.s }, p2/Z, [x22, x16, LSL #2]\n" + "ld1w { z11.s }, p2/Z, [x21, x16, LSL #2]\n" + "ld1w { z12.s }, p2/Z, [x20, x16, LSL #2]\n" "bge 2f\n" "1:" // Channel loop - "movprfx z21, z15\n fmla z21.s, p3/M, z4.s, z9.s\n" - "movprfx z16, z15\n fmla z16.s, p3/M, z8.s, z9.s\n" - "ldr x27, [x16, #0x20]\n" - "ldr x26, [x16, #0x30]\n" - "movprfx z22, z15\n fmla z22.s, p3/M, z3.s, z9.s\n" - "movprfx z25, z15\n fmla z25.s, p3/M, z1.s, z9.s\n" - "ldr x25, [x16, #0x28]\n" - "ldr x24, [x16, #0x38]\n" - "movprfx z26, z15\n fmla z26.s, p3/M, z0.s, z9.s\n" - "movprfx z17, z15\n fmla z17.s, p3/M, z7.s, z9.s\n" - "ldr x12, [x16, #0x40]\n" - "ldr x11, [x16, #0x48]\n" - "movprfx z18, z15\n fmla z18.s, p3/M, z6.s, z9.s\n" - "fmla z21.s, p3/M, z5.s, z12.s\n" - "ldr x10, [x16, #0x50]\n" - "ldr x9, [x16, #0x58]\n" - "movprfx z20, z15\n fmla z20.s, p3/M, z5.s, z9.s\n" - "movprfx z24, z15\n fmla z24.s, p3/M, z2.s, z9.s\n" - "ld1w { z9.s }, p2/Z, [x26, x14, LSL #2]\n" - "ldr x26, [x16, #0x70]\n" - "fmla z16.s, p3/M, z0.s, z10.s\n" - "movprfx z19, z15\n fmla z19.s, p3/M, z2.s, z11.s\n" - "ld1w { z10.s }, p2/Z, [x27, x14, LSL #2]\n" - "ld1w { z11.s }, p2/Z, [x25, x14, LSL #2]\n" - "fmla z22.s, p3/M, z4.s, z12.s\n" - "fmla z25.s, p3/M, z2.s, z12.s\n" - "ldr x27, [x16, #0x60]\n" - "ldr x25, [x16, #0x68]\n" - "fmla z26.s, p3/M, z1.s, z12.s\n" - "fmla z17.s, p3/M, z8.s, z12.s\n" - "incw x13\n" + "movprfx z20, z17\n fmla z20.s, p3/M, z4.s, z9.s\n" + "movprfx z26, z17\n fmla z26.s, p3/M, z8.s, z9.s\n" + "ldr x27, [x8, #0x20]\n" + "ldr x24, [x8, #0x30]\n" + "movprfx z24, z17\n fmla z24.s, p3/M, z3.s, z9.s\n" + "movprfx z30, z17\n fmla z30.s, p3/M, z1.s, z9.s\n" + "ldr x23, [x8, #0x28]\n" + "ldr x22, [x8, #0x38]\n" + "movprfx z31, z17\n fmla z31.s, p3/M, z0.s, z9.s\n" + "movprfx z22, z17\n fmla z22.s, p3/M, z7.s, z9.s\n" + "ldr x26, [x8, #0x40]\n" + "ldr x21, [x8, #0x48]\n" + "movprfx z27, z17\n fmla z27.s, p3/M, z6.s, z9.s\n" + "fmla z20.s, p3/M, z5.s, z12.s\n" + "ldr x25, [x8, #0x50]\n" + "ldr x20, [x8, #0x58]\n" + "movprfx z14, z17\n fmla z14.s, p3/M, z5.s, z9.s\n" + "movprfx z23, z17\n fmla z23.s, p3/M, z2.s, z9.s\n" + "ld1w { z25.s }, p2/Z, [x24, x16, LSL #2]\n" + "ldr x13, [x8, #0x70]\n" + "fmla z26.s, p3/M, z0.s, z10.s\n" + "movprfx z9, z17\n fmla z9.s, p3/M, z2.s, z11.s\n" + "ld1w { z28.s }, p2/Z, [x27, x16, LSL #2]\n" + "ld1w { z21.s }, p2/Z, [x23, x16, LSL #2]\n" + "fmla z24.s, p3/M, z4.s, z12.s\n" + "fmla z30.s, p3/M, z2.s, z12.s\n" + "ldr x24, [x8, #0x60]\n" + "ldr x23, [x8, #0x68]\n" + "fmla z31.s, p3/M, z1.s, z12.s\n" + "fmla z22.s, p3/M, z8.s, z12.s\n" + "incw x15\n" "mov p1.b, p2.b\n" - "fmla z18.s, p3/M, z7.s, z12.s\n" - "movprfx z28, z15\n fmla z28.s, p3/M, z6.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x11, x14, LSL #2]\n" - "ldr x11, [x16, #0x88]\n" - "fmla z21.s, p3/M, z7.s, z9.s\n" - "fmla z19.s, p3/M, z6.s, z12.s\n" - "ldr x23, [x28, #0x0]\n" - "ldr x22, [x28, #0x8]\n" - "movprfx z23, z15\n fmla z23.s, p3/M, z3.s, z12.s\n" - "movprfx z27, z15\n fmla z27.s, p3/M, z0.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n" - "ldr x24, [x16, #0x78]\n" - "movprfx z31, z15\n fmla z31.s, p3/M, z8.s, z11.s\n" - "fmla z22.s, p3/M, z6.s, z9.s\n" - "ld1w { z11.s }, p2/Z, [x12, x14, LSL #2]\n" - "ldr x12, [x16, #0x80]\n" - "fmla z25.s, p3/M, z4.s, z9.s\n" - "fmla z26.s, p3/M, z3.s, z9.s\n" - "ldr x21, [x28, #0x10]\n" - "ldr x20, [x28, #0x18]\n" - "fmla z20.s, p3/M, z8.s, z9.s\n" - "fmla z24.s, p3/M, z5.s, z9.s\n" - "whilelt p0.s, x15, %x[n_channels]\n" - "fmla z28.s, p3/M, z2.s, z9.s\n" - "fmla z16.s, p3/M, z1.s, z12.s\n" - "fmla z17.s, p3/M, z0.s, z12.s\n" - "movprfx z29, z15\n fmla z29.s, p3/M, z1.s, z9.s\n" - "movprfx z30, z15\n fmla z30.s, p3/M, z0.s, z9.s\n" - "fmla z18.s, p3/M, z2.s, z11.s\n" - "ld1w { z9.s }, p2/Z, [x10, x14, LSL #2]\n" - "ldr x10, [x16, #0x90]\n" - "fmla z21.s, p3/M, z8.s, z10.s\n" - "fmla z19.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x27, x14, LSL #2]\n" - "ldr x27, [x16, #0xa0]\n" - "fmla z22.s, p3/M, z7.s, z10.s\n" - "fmla z23.s, p3/M, z6.s, z10.s\n" - "fmla z25.s, p3/M, z5.s, z10.s\n" - "fmla z26.s, p3/M, z4.s, z10.s\n" - "fmla z27.s, p3/M, z3.s, z10.s\n" - "fmla z29.s, p3/M, z2.s, z10.s\n" - "fmla z30.s, p3/M, z1.s, z10.s\n" - "fmla z31.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x25, x14, LSL #2]\n" - "ldr x25, [x16, #0xa8]\n" - "fmla z16.s, p3/M, z3.s, z9.s\n" - "fmla z20.s, p3/M, z0.s, z9.s\n" - "ld1w { z12.s }, p2/Z, [x9, x14, LSL #2]\n" - "ldr x9, [x16, #0x98]\n" - "fmla z24.s, p3/M, z6.s, z11.s\n" - "fmla z28.s, p3/M, z3.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x26, x14, LSL #2]\n" - "ldr x26, [x16, #0xb0]\n" - "fmla z17.s, p3/M, z4.s, z10.s\n" - "fmla z18.s, p3/M, z3.s, z10.s\n" + "fmla z27.s, p3/M, z7.s, z12.s\n" + "movprfx z15, z17\n fmla z15.s, p3/M, z6.s, z28.s\n" + "ld1w { z10.s }, p2/Z, [x21, x16, LSL #2]\n" + "ldr x28, [x8, #0x88]\n" + "fmla z20.s, p3/M, z7.s, z25.s\n" + "fmla z9.s, p3/M, z6.s, z12.s\n" + "ldr x12, [x14, #0x0]\n" + "ldr x11, [x14, #0x8]\n" + "movprfx z11, z17\n fmla z11.s, p3/M, z3.s, z12.s\n" + "movprfx z13, z17\n fmla z13.s, p3/M, z0.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x22, x16, LSL #2]\n" + "ldr x22, [x8, #0x78]\n" + "movprfx z28, z17\n fmla z28.s, p3/M, z8.s, z21.s\n" + "fmla z24.s, p3/M, z6.s, z25.s\n" + "ld1w { z29.s }, p2/Z, [x26, x16, LSL #2]\n" + "ldr x21, [x8, #0x80]\n" + "fmla z30.s, p3/M, z4.s, z25.s\n" + "fmla z31.s, p3/M, z3.s, z25.s\n" + "ldr x10, [x14, #0x10]\n" + "ldr x9, [x14, #0x18]\n" + "movprfx z18, z17\n fmla z18.s, p3/M, z1.s, z25.s\n" + "movprfx z21, z17\n fmla z21.s, p3/M, z0.s, z25.s\n" + "whilelt p0.s, x17, %x[n_channels]\n" + "ld1w { z17.s }, p3/Z, [x7]\n" + "fmla z14.s, p3/M, z8.s, z25.s\n" + "fmla z23.s, p3/M, z5.s, z25.s\n" + "fmla z15.s, p3/M, z2.s, z25.s\n" + "fmla z26.s, p3/M, z1.s, z12.s\n" + "ld1w { z25.s }, p2/Z, [x25, x16, LSL #2]\n" + "ldr x27, [x8, #0x90]\n" + "fmla z22.s, p3/M, z0.s, z12.s\n" + "fmla z27.s, p3/M, z2.s, z29.s\n" + "ld1w { z12.s }, p2/Z, [x20, x16, LSL #2]\n" + "ldr x20, [x8, #0x98]\n" + "fmla z20.s, p3/M, z8.s, z10.s\n" + "fmla z9.s, p3/M, z1.s, z29.s\n" + "ld1w { z29.s }, p2/Z, [x24, x16, LSL #2]\n" + "ldr x26, [x8, #0xa0]\n" + "fmla z24.s, p3/M, z7.s, z10.s\n" + "fmla z11.s, p3/M, z6.s, z10.s\n" + "fmla z30.s, p3/M, z5.s, z10.s\n" + "fmla z31.s, p3/M, z4.s, z10.s\n" + "fmla z13.s, p3/M, z3.s, z10.s\n" + "fmla z18.s, p3/M, z2.s, z10.s\n" "fmla z21.s, p3/M, z1.s, z10.s\n" - "fmla z19.s, p3/M, z5.s, z12.s\n" - "fmla z23.s, p3/M, z2.s, z12.s\n" - "fmla z22.s, p3/M, z0.s, z10.s\n" - "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n" - "ldr x24, [x16, #0xb8]\n" - "fmla z27.s, p3/M, z8.s, z11.s\n" - "fmla z31.s, p3/M, z5.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x12, x14, LSL #2]\n" - "ldr x12, [x16, #0xc0]\n" - "fmla z16.s, p3/M, z5.s, z10.s\n" - "fmla z20.s, p3/M, z2.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x11, x14, LSL #2]\n" - "ldr x11, [x16, #0xc8]\n" - "fmla z17.s, p3/M, z5.s, z12.s\n" - "fmla z18.s, p3/M, z4.s, z12.s\n" - "fmla z21.s, p3/M, z2.s, z12.s\n" - "fmla z19.s, p3/M, z3.s, z12.s\n" - "fmla z22.s, p3/M, z1.s, z12.s\n" - "fmla z23.s, p3/M, z0.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x9, x14, LSL #2]\n" - "ldr x9, [x16, #0xd8]\n" - "fmla z28.s, p3/M, z7.s, z11.s\n" - "fmla z29.s, p3/M, z6.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x10, x14, LSL #2]\n" - "ldr x10, [x16, #0xd0]\n" - "fmla z16.s, p3/M, z7.s, z10.s\n" - "fmla z17.s, p3/M, z6.s, z10.s\n" - "fmla z20.s, p3/M, z4.s, z10.s\n" - "fmla z21.s, p3/M, z3.s, z10.s\n" - "fmla z24.s, p3/M, z1.s, z10.s\n" - "fmla z25.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x27, x14, LSL #2]\n" - "ldr x27, [x16, #0xe0]\n" - "fmla z18.s, p3/M, z8.s, z12.s\n" - "fmla z30.s, p3/M, z8.s, z11.s\n" - "fmla z31.s, p3/M, z7.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x25, x14, LSL #2]\n" - "fmla z27.s, p3/M, z1.s, z12.s\n" - "ldr x25, [x16, #0xe8]\n" - "fmla z19.s, p3/M, z7.s, z12.s\n" - "fmla z22.s, p3/M, z5.s, z12.s\n" - "fmla z23.s, p3/M, z4.s, z12.s\n" - "fmla z26.s, p3/M, z2.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x26, x14, LSL #2]\n" - "ldr x26, [x16, #0xf0]\n" - "fmla z16.s, p3/M, z2.s, z10.s\n" - "fmla z17.s, p3/M, z1.s, z10.s\n" - "fmla z18.s, p3/M, z0.s, z10.s\n" - "fmla z20.s, p3/M, z7.s, z11.s\n" - "ld1w { z10.s }, p2/Z, [x24, x14, LSL #2]\n" - "ldr x24, [x16, #0xf8]\n" - "fmla z21.s, p3/M, z6.s, z11.s\n" - "fmla z24.s, p3/M, z4.s, z11.s\n" - "fmla z25.s, p3/M, z3.s, z11.s\n" - "fmla z28.s, p3/M, z1.s, z11.s\n" - "fmla z29.s, p3/M, z0.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x12, x14, LSL #2]\n" - "fmla z27.s, p3/M, z4.s, z11.s\n" - "ldr x12, [x16, #0x100]\n" - "fmla z30.s, p3/M, z2.s, z11.s\n" - "fmla z17.s, p3/M, z2.s, z12.s\n" - "fmla z18.s, p3/M, z1.s, z12.s\n" - "fmla z19.s, p3/M, z0.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x11, x14, LSL #2]\n" - "ldr x11, [x16, #0x108]\n" - "fmla z16.s, p3/M, z6.s, z10.s\n" - "fmla z20.s, p3/M, z3.s, z10.s\n" - "fmla z24.s, p3/M, z0.s, z10.s\n" - "fmla z22.s, p3/M, z8.s, z11.s\n" - "ld1w { z10.s }, p2/Z, [x10, x14, LSL #2]\n" - "ldr x10, [x16, #0x110]\n" - "fmla z23.s, p3/M, z7.s, z11.s\n" - "fmla z26.s, p3/M, z5.s, z11.s\n" - "fmla z31.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x9, x14, LSL #2]\n" - "fmla z27.s, p3/M, z2.s, z12.s\n" - "ldr x9, [x16, #0x118]\n" "fmla z28.s, p3/M, z0.s, z10.s\n" - "fmla z29.s, p3/M, z4.s, z11.s\n" - "fmla z30.s, p3/M, z3.s, z11.s\n" - "fmla z19.s, p3/M, z8.s, z12.s\n" - "fmla z23.s, p3/M, z5.s, z12.s\n" - "fmla z20.s, p3/M, z6.s, z10.s\n" - "ld1w { z12.s }, p2/Z, [x27, x14, LSL #2]\n" - "fmla z24.s, p3/M, z3.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x25, x14, LSL #2]\n" - "fmla z25.s, p3/M, z7.s, z11.s\n" - "fmla z26.s, p3/M, z6.s, z11.s\n" - "fmla z28.s, p3/M, z5.s, z11.s\n" - "fmla z27.s, p3/M, z5.s, z12.s\n" - "fmla z31.s, p3/M, z2.s, z12.s\n" - "fmla z29.s, p3/M, z7.s, z10.s\n" - "fmla z30.s, p3/M, z6.s, z10.s\n" - "fmla z24.s, p3/M, z8.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x26, x14, LSL #2]\n" - "fmla z28.s, p3/M, z8.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x12, x14, LSL #2]\n" - "fmla z25.s, p3/M, z8.s, z11.s\n" - "fmla z26.s, p3/M, z7.s, z11.s\n" - "fmla z27.s, p3/M, z6.s, z11.s\n" - "fmla z29.s, p3/M, z5.s, z11.s\n" - "fmla z30.s, p3/M, z4.s, z11.s\n" - "fmla z31.s, p3/M, z3.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x11, x14, LSL #2]\n" - "ldp x12, x11, [x16, #0x0]\n" - "fmla z23.s, p3/M, z8.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n" - "fmla z16.s, p3/M, z4.s, z10.s\n" - "fmax z16.s, p3/M, z16.s, z14.s\n" - "fmla z17.s, p3/M, z3.s, z10.s\n" - "fmla z18.s, p3/M, z5.s, z11.s\n" - "fmax z17.s, p3/M, z17.s, z14.s\n" - "fmax z18.s, p3/M, z18.s, z14.s\n" - "fmla z19.s, p3/M, z4.s, z11.s\n" - "fmla z29.s, p3/M, z8.s, z12.s\n" - "fmax z19.s, p3/M, z19.s, z14.s\n" - "fmin z16.s, p3/M, z16.s, z13.s\n" - "fmla z30.s, p3/M, z7.s, z12.s\n" - "fmla z31.s, p3/M, z6.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x10, x14, LSL #2]\n" - "fmin z17.s, p3/M, z17.s, z13.s\n" + "ld1w { z10.s }, p2/Z, [x23, x16, LSL #2]\n" + "ldr x25, [x8, #0xa8]\n" + "fmla z26.s, p3/M, z3.s, z25.s\n" + "fmla z14.s, p3/M, z0.s, z25.s\n" + "fmla z23.s, p3/M, z6.s, z29.s\n" + "fmla z15.s, p3/M, z3.s, z29.s\n" + "ld1w { z25.s }, p2/Z, [x13, x16, LSL #2]\n" + "ldr x24, [x8, #0xb0]\n" + "fmla z22.s, p3/M, z4.s, z10.s\n" + "fmla z27.s, p3/M, z3.s, z10.s\n" "fmla z20.s, p3/M, z1.s, z10.s\n" - "fmla z21.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x9, x14, LSL #2]\n" - "fmin z18.s, p3/M, z18.s, z13.s\n" - "fmla z22.s, p3/M, z2.s, z11.s\n" - "fmla z23.s, p3/M, z1.s, z11.s\n" - "fmin z19.s, p3/M, z19.s, z13.s\n" - "fmax z20.s, p3/M, z20.s, z14.s\n" - "fmla z24.s, p3/M, z7.s, z12.s\n" - "fmla z25.s, p3/M, z6.s, z12.s\n" - "fmax z21.s, p3/M, z21.s, z14.s\n" - "fmax z22.s, p3/M, z22.s, z14.s\n" - "fmla z26.s, p3/M, z8.s, z10.s\n" - "fmla z27.s, p3/M, z7.s, z10.s\n" - "fmax z23.s, p3/M, z23.s, z14.s\n" - "st1w { z16.s }, p1, [x23, x13, LSL #2]\n" - "st1w { z17.s }, p1, [x22, x13, LSL #2]\n" - "ldr x23, [x28, #0x20]\n" - "ldr x22, [x28, #0x28]\n" + "fmla z9.s, p3/M, z5.s, z12.s\n" + "fmla z11.s, p3/M, z2.s, z12.s\n" + "fmla z24.s, p3/M, z0.s, z10.s\n" + "ld1w { z12.s }, p2/Z, [x22, x16, LSL #2]\n" + "ldr x23, [x8, #0xb8]\n" + "fmla z13.s, p3/M, z8.s, z25.s\n" + "fmla z28.s, p3/M, z5.s, z25.s\n" + "ld1w { z25.s }, p2/Z, [x21, x16, LSL #2]\n" + "ldr x22, [x8, #0xc0]\n" + "fmla z26.s, p3/M, z5.s, z10.s\n" + "fmla z14.s, p3/M, z2.s, z10.s\n" + "ld1w { z29.s }, p2/Z, [x28, x16, LSL #2]\n" + "ldr x21, [x8, #0xc8]\n" + "fmla z22.s, p3/M, z5.s, z12.s\n" + "fmla z27.s, p3/M, z4.s, z12.s\n" + "fmla z20.s, p3/M, z2.s, z12.s\n" + "fmla z9.s, p3/M, z3.s, z12.s\n" + "fmla z24.s, p3/M, z1.s, z12.s\n" + "fmla z11.s, p3/M, z0.s, z12.s\n" + "ld1w { z10.s }, p2/Z, [x20, x16, LSL #2]\n" + "ldr x28, [x8, #0xd8]\n" + "fmla z15.s, p3/M, z7.s, z25.s\n" + "fmla z18.s, p3/M, z6.s, z25.s\n" + "ld1w { z25.s }, p2/Z, [x27, x16, LSL #2]\n" + "ldr x20, [x8, #0xd0]\n" + "fmla z26.s, p3/M, z7.s, z29.s\n" + "fmla z22.s, p3/M, z6.s, z29.s\n" + "fmla z14.s, p3/M, z4.s, z29.s\n" + "fmla z20.s, p3/M, z3.s, z29.s\n" + "fmla z23.s, p3/M, z1.s, z29.s\n" + "fmla z30.s, p3/M, z0.s, z29.s\n" + "ld1w { z29.s }, p2/Z, [x26, x16, LSL #2]\n" + "ldr x27, [x8, #0xe0]\n" + "fmla z27.s, p3/M, z8.s, z10.s\n" + "fmla z21.s, p3/M, z8.s, z25.s\n" + "fmla z28.s, p3/M, z7.s, z25.s\n" + "ld1w { z25.s }, p2/Z, [x25, x16, LSL #2]\n" + "fmla z13.s, p3/M, z1.s, z10.s\n" + "ldr x26, [x8, #0xe8]\n" + "fmla z9.s, p3/M, z7.s, z10.s\n" + "fmla z24.s, p3/M, z5.s, z10.s\n" + "fmla z11.s, p3/M, z4.s, z10.s\n" + "fmla z31.s, p3/M, z2.s, z10.s\n" + "ld1w { z10.s }, p2/Z, [x24, x16, LSL #2]\n" + "ldr x25, [x8, #0xf0]\n" + "fmla z26.s, p3/M, z2.s, z29.s\n" + "fmla z22.s, p3/M, z1.s, z29.s\n" + "fmla z27.s, p3/M, z0.s, z29.s\n" + "fmla z14.s, p3/M, z7.s, z25.s\n" + "ld1w { z29.s }, p2/Z, [x23, x16, LSL #2]\n" + "ldr x24, [x8, #0xf8]\n" + "fmla z20.s, p3/M, z6.s, z25.s\n" + "fmla z23.s, p3/M, z4.s, z25.s\n" + "fmla z30.s, p3/M, z3.s, z25.s\n" + "fmla z15.s, p3/M, z1.s, z25.s\n" + "fmla z18.s, p3/M, z0.s, z25.s\n" + "ld1w { z25.s }, p2/Z, [x22, x16, LSL #2]\n" + "fmla z13.s, p3/M, z4.s, z25.s\n" + "ldr x23, [x8, #0x100]\n" + "fmla z21.s, p3/M, z2.s, z25.s\n" + "fmla z22.s, p3/M, z2.s, z10.s\n" + "fmla z27.s, p3/M, z1.s, z10.s\n" + "fmla z9.s, p3/M, z0.s, z10.s\n" + "ld1w { z12.s }, p2/Z, [x21, x16, LSL #2]\n" + "ldr x22, [x8, #0x108]\n" + "fmla z26.s, p3/M, z6.s, z29.s\n" + "fmla z14.s, p3/M, z3.s, z29.s\n" + "fmla z23.s, p3/M, z0.s, z29.s\n" + "fmla z24.s, p3/M, z8.s, z25.s\n" + "ld1w { z10.s }, p2/Z, [x20, x16, LSL #2]\n" + "ldr x21, [x8, #0x110]\n" + "fmla z11.s, p3/M, z7.s, z25.s\n" + "fmla z31.s, p3/M, z5.s, z25.s\n" + "fmla z28.s, p3/M, z1.s, z25.s\n" + "ld1w { z25.s }, p2/Z, [x28, x16, LSL #2]\n" + "fmla z13.s, p3/M, z2.s, z12.s\n" + "ldr x20, [x8, #0x118]\n" + "fmla z15.s, p3/M, z0.s, z10.s\n" + "fmla z18.s, p3/M, z4.s, z25.s\n" + "fmla z21.s, p3/M, z3.s, z25.s\n" + "fmla z9.s, p3/M, z8.s, z12.s\n" + "fmla z11.s, p3/M, z5.s, z12.s\n" + "fmla z14.s, p3/M, z6.s, z10.s\n" + "ld1w { z12.s }, p2/Z, [x27, x16, LSL #2]\n" + "fmla z23.s, p3/M, z3.s, z10.s\n" + "ld1w { z29.s }, p2/Z, [x26, x16, LSL #2]\n" + "fmla z30.s, p3/M, z7.s, z25.s\n" + "fmla z31.s, p3/M, z6.s, z25.s\n" + "fmla z15.s, p3/M, z5.s, z25.s\n" + "fmla z13.s, p3/M, z5.s, z12.s\n" + "fmla z28.s, p3/M, z2.s, z12.s\n" + "fmla z18.s, p3/M, z7.s, z29.s\n" + "fmla z21.s, p3/M, z6.s, z29.s\n" + "fmla z23.s, p3/M, z8.s, z25.s\n" + "ld1w { z25.s }, p2/Z, [x25, x16, LSL #2]\n" + "fmla z15.s, p3/M, z8.s, z29.s\n" + "ld1w { z29.s }, p2/Z, [x23, x16, LSL #2]\n" + "fmla z30.s, p3/M, z8.s, z25.s\n" + "fmla z31.s, p3/M, z7.s, z25.s\n" + "fmla z13.s, p3/M, z6.s, z25.s\n" + "fmla z18.s, p3/M, z5.s, z25.s\n" + "fmla z21.s, p3/M, z4.s, z25.s\n" + "fmla z28.s, p3/M, z3.s, z25.s\n" + "ld1w { z25.s }, p2/Z, [x22, x16, LSL #2]\n" + "ldp x27, x26, [x8, #0x0]\n" + "fmla z11.s, p3/M, z8.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x24, x16, LSL #2]\n" + "fmla z26.s, p3/M, z4.s, z29.s\n" + "fmax z26.s, p3/M, z26.s, z16.s\n" + "fmla z22.s, p3/M, z3.s, z29.s\n" + "fmla z27.s, p3/M, z5.s, z25.s\n" + "fmax z22.s, p3/M, z22.s, z16.s\n" + "fmax z27.s, p3/M, z27.s, z16.s\n" + "fmla z9.s, p3/M, z4.s, z25.s\n" + "fmla z18.s, p3/M, z8.s, z12.s\n" + "fmax z9.s, p3/M, z9.s, z16.s\n" + "fmin z26.s, p3/M, z26.s, z19.s\n" + "fmla z21.s, p3/M, z7.s, z12.s\n" + "fmla z28.s, p3/M, z6.s, z12.s\n" + "ld1w { z10.s }, p2/Z, [x21, x16, LSL #2]\n" + "fmin z22.s, p3/M, z22.s, z19.s\n" + "fmla z14.s, p3/M, z1.s, z29.s\n" + "fmla z20.s, p3/M, z0.s, z29.s\n" + "ld1w { z12.s }, p2/Z, [x20, x16, LSL #2]\n" + "fmin z27.s, p3/M, z27.s, z19.s\n" + "fmla z24.s, p3/M, z2.s, z25.s\n" + "fmla z11.s, p3/M, z1.s, z25.s\n" + "fmin z9.s, p3/M, z9.s, z19.s\n" + "fmax z14.s, p3/M, z14.s, z16.s\n" + "fmla z23.s, p3/M, z7.s, z10.s\n" + "fmla z30.s, p3/M, z6.s, z10.s\n" + "fmax z20.s, p3/M, z20.s, z16.s\n" + "fmax z24.s, p3/M, z24.s, z16.s\n" + "fmla z31.s, p3/M, z8.s, z12.s\n" + "fmla z13.s, p3/M, z7.s, z12.s\n" + "fmax z11.s, p3/M, z11.s, z16.s\n" + "st1w { z26.s }, p1, [x12, x15, LSL #2]\n" + "st1w { z22.s }, p1, [x11, x15, LSL #2]\n" + "ldr x23, [x14, #0x20]\n" + "ldr x22, [x14, #0x28]\n" + "fmla z15.s, p3/M, z4.s, z10.s\n" + "st1w { z27.s }, p1, [x10, x15, LSL #2]\n" + "ldr x21, [x14, #0x30]\n" + "fmla z18.s, p3/M, z3.s, z10.s\n" + "fmla z21.s, p3/M, z5.s, z12.s\n" + "st1w { z9.s }, p1, [x9, x15, LSL #2]\n" + "ldr x20, [x14, #0x38]\n" "fmla z28.s, p3/M, z4.s, z12.s\n" - "st1w { z18.s }, p1, [x21, x13, LSL #2]\n" - "ldr x21, [x28, #0x30]\n" - "fmla z29.s, p3/M, z3.s, z12.s\n" - "fmla z30.s, p3/M, z5.s, z10.s\n" - "st1w { z19.s }, p1, [x20, x13, LSL #2]\n" - "ldr x20, [x28, #0x38]\n" - "fmla z31.s, p3/M, z4.s, z10.s\n" - "ldp x10, x9, [x16, #0x10]\n" - "fmin z20.s, p3/M, z20.s, z13.s\n" - "fmin z21.s, p3/M, z21.s, z13.s\n" - "st1w { z20.s }, p1, [x23, x13, LSL #2]\n" - "ldr x23, [x28, #0x40]\n" - "fmin z22.s, p3/M, z22.s, z13.s\n" - "fmin z23.s, p3/M, z23.s, z13.s\n" - "st1w { z21.s }, p1, [x22, x13, LSL #2]\n" - "ldr x22, [x28, #0x48]\n" - "fmax z24.s, p3/M, z24.s, z14.s\n" - "fmax z25.s, p3/M, z25.s, z14.s\n" - "st1w { z22.s }, p1, [x21, x13, LSL #2]\n" - "ldr x21, [x28, #0x50]\n" - "fmax z26.s, p3/M, z26.s, z14.s\n" - "fmax z27.s, p3/M, z27.s, z14.s\n" - "st1w { z23.s }, p1, [x20, x13, LSL #2]\n" - "ldr x20, [x28, #0x58]\n" - "incw x14\n" - "ld1w { z9.s }, p0/Z, [x12, x15, LSL #2]\n" - "ld1w { z10.s }, p0/Z, [x11, x15, LSL #2]\n" - "fmin z24.s, p3/M, z24.s, z13.s\n" - "ld1w { z11.s }, p0/Z, [x10, x15, LSL #2]\n" - "ld1w { z12.s }, p0/Z, [x9, x15, LSL #2]\n" - "incw x15\n" - "fmin z25.s, p3/M, z25.s, z13.s\n" - "fmin z26.s, p3/M, z26.s, z13.s\n" - "fmin z27.s, p3/M, z27.s, z13.s\n" - "st1w { z24.s }, p1, [x23, x13, LSL #2]\n" - "ldr x23, [x28, #0x60]\n" - "fmax z28.s, p3/M, z28.s, z14.s\n" - "fmax z29.s, p3/M, z29.s, z14.s\n" - "st1w { z25.s }, p1, [x22, x13, LSL #2]\n" - "ldr x22, [x28, #0x68]\n" - "fmax z30.s, p3/M, z30.s, z14.s\n" - "fmax z31.s, p3/M, z31.s, z14.s\n" - "st1w { z26.s }, p1, [x21, x13, LSL #2]\n" - "ldr x21, [x28, #0x70]\n" - "st1w { z27.s }, p1, [x20, x13, LSL #2]\n" - "ldr x20, [x28, #0x78]\n" - "ld1w { z15.s }, p3/Z, [x17]\n" - "whilelt p2.s, x14, %x[n_channels]\n" - "ld1w { z0.s }, p3/Z, [x17, #1, MUL VL]\n" - "ld1w { z1.s }, p3/Z, [x17, #2, MUL VL]\n" - "cmp x15, %x[n_channels]\n" - "fmin z28.s, p3/M, z28.s, z13.s\n" - "ld1w { z2.s }, p3/Z, [x17, #3, MUL VL]\n" - "ld1w { z3.s }, p3/Z, [x17, #4, MUL VL]\n" - "fmin z29.s, p3/M, z29.s, z13.s\n" - "fmin z30.s, p3/M, z30.s, z13.s\n" - "ld1w { z4.s }, p3/Z, [x17, #5, MUL VL]\n" - "ld1w { z5.s }, p3/Z, [x17, #6, MUL VL]\n" - "fmin z31.s, p3/M, z31.s, z13.s\n" - "st1w { z28.s }, p1, [x23, x13, LSL #2]\n" - "ld1w { z6.s }, p3/Z, [x17, #7, MUL VL]\n" - "addvl x17, x17, #16\n" - "st1w { z29.s }, p1, [x22, x13, LSL #2]\n" - "ld1w { z7.s }, p3/Z, [x17, #-8, MUL VL]\n" - "st1w { z30.s }, p1, [x21, x13, LSL #2]\n" - "ld1w { z8.s }, p3/Z, [x17, #-7, MUL VL]\n" - "addvl x17, x17, #-6\n" - "st1w { z31.s }, p1, [x20, x13, LSL #2]\n" + "ldp x25, x24, [x8, #0x10]\n" + "fmin z14.s, p3/M, z14.s, z19.s\n" + "fmin z20.s, p3/M, z20.s, z19.s\n" + "st1w { z14.s }, p1, [x23, x15, LSL #2]\n" + "ldr x23, [x14, #0x40]\n" + "fmin z24.s, p3/M, z24.s, z19.s\n" + "fmin z11.s, p3/M, z11.s, z19.s\n" + "st1w { z20.s }, p1, [x22, x15, LSL #2]\n" + "ldr x22, [x14, #0x48]\n" + "fmax z23.s, p3/M, z23.s, z16.s\n" + "fmax z30.s, p3/M, z30.s, z16.s\n" + "st1w { z24.s }, p1, [x21, x15, LSL #2]\n" + "ldr x21, [x14, #0x50]\n" + "fmax z31.s, p3/M, z31.s, z16.s\n" + "fmax z13.s, p3/M, z13.s, z16.s\n" + "st1w { z11.s }, p1, [x20, x15, LSL #2]\n" + "ldr x20, [x14, #0x58]\n" + "incw x16\n" + "ld1w { z9.s }, p0/Z, [x27, x17, LSL #2]\n" + "ld1w { z10.s }, p0/Z, [x26, x17, LSL #2]\n" + "fmin z23.s, p3/M, z23.s, z19.s\n" + "ld1w { z11.s }, p0/Z, [x25, x17, LSL #2]\n" + "ld1w { z12.s }, p0/Z, [x24, x17, LSL #2]\n" + "incw x17\n" + "fmin z30.s, p3/M, z30.s, z19.s\n" + "fmin z31.s, p3/M, z31.s, z19.s\n" + "fmin z13.s, p3/M, z13.s, z19.s\n" + "st1w { z23.s }, p1, [x23, x15, LSL #2]\n" + "ldr x23, [x14, #0x60]\n" + "fmax z15.s, p3/M, z15.s, z16.s\n" + "fmax z18.s, p3/M, z18.s, z16.s\n" + "st1w { z30.s }, p1, [x22, x15, LSL #2]\n" + "ldr x22, [x14, #0x68]\n" + "fmax z21.s, p3/M, z21.s, z16.s\n" + "fmax z28.s, p3/M, z28.s, z16.s\n" + "st1w { z31.s }, p1, [x21, x15, LSL #2]\n" + "ldr x21, [x14, #0x70]\n" + "st1w { z13.s }, p1, [x20, x15, LSL #2]\n" + "ldr x20, [x14, #0x78]\n" + "ld1w { z0.s }, p3/Z, [x7, #1, MUL VL]\n" + "whilelt p2.s, x16, %x[n_channels]\n" + "ld1w { z1.s }, p3/Z, [x7, #2, MUL VL]\n" + "ld1w { z2.s }, p3/Z, [x7, #3, MUL VL]\n" + "cmp x17, %x[n_channels]\n" + "fmin z15.s, p3/M, z15.s, z19.s\n" + "ld1w { z3.s }, p3/Z, [x7, #4, MUL VL]\n" + "ld1w { z4.s }, p3/Z, [x7, #5, MUL VL]\n" + "fmin z18.s, p3/M, z18.s, z19.s\n" + "fmin z21.s, p3/M, z21.s, z19.s\n" + "ld1w { z5.s }, p3/Z, [x7, #6, MUL VL]\n" + "ld1w { z6.s }, p3/Z, [x7, #7, MUL VL]\n" + "addvl x7, x7, #16\n" + "fmin z28.s, p3/M, z28.s, z19.s\n" + "st1w { z15.s }, p1, [x23, x15, LSL #2]\n" + "ld1w { z7.s }, p3/Z, [x7, #-8, MUL VL]\n" + "ld1w { z8.s }, p3/Z, [x7, #-7, MUL VL]\n" + "addvl x7, x7, #-6\n" + "st1w { z18.s }, p1, [x22, x15, LSL #2]\n" + "st1w { z21.s }, p1, [x21, x15, LSL #2]\n" + "st1w { z28.s }, p1, [x20, x15, LSL #2]\n" "blt 1b\n" "2:" // Channel tail - "movprfx z21, z15\n fmla z21.s, p3/M, z4.s, z9.s\n" - "movprfx z16, z15\n fmla z16.s, p3/M, z8.s, z9.s\n" - "ldr x27, [x16, #0x20]\n" - "ldr x26, [x16, #0x30]\n" - "movprfx z22, z15\n fmla z22.s, p3/M, z3.s, z9.s\n" - "movprfx z25, z15\n fmla z25.s, p3/M, z1.s, z9.s\n" - "ldr x25, [x16, #0x28]\n" - "ldr x24, [x16, #0x38]\n" - "movprfx z26, z15\n fmla z26.s, p3/M, z0.s, z9.s\n" - "movprfx z17, z15\n fmla z17.s, p3/M, z7.s, z9.s\n" - "ldr x12, [x16, #0x40]\n" - "ldr x11, [x16, #0x48]\n" - "movprfx z18, z15\n fmla z18.s, p3/M, z6.s, z9.s\n" - "fmla z21.s, p3/M, z5.s, z12.s\n" - "ldr x10, [x16, #0x50]\n" - "ldr x9, [x16, #0x58]\n" - "movprfx z20, z15\n fmla z20.s, p3/M, z5.s, z9.s\n" - "movprfx z24, z15\n fmla z24.s, p3/M, z2.s, z9.s\n" - "ld1w { z9.s }, p2/Z, [x26, x14, LSL #2]\n" - "ldr x26, [x16, #0x70]\n" - "fmla z16.s, p3/M, z0.s, z10.s\n" - "movprfx z19, z15\n fmla z19.s, p3/M, z2.s, z11.s\n" - "ld1w { z10.s }, p2/Z, [x27, x14, LSL #2]\n" - "ld1w { z11.s }, p2/Z, [x25, x14, LSL #2]\n" - "fmla z22.s, p3/M, z4.s, z12.s\n" - "fmla z25.s, p3/M, z2.s, z12.s\n" - "ldr x27, [x16, #0x60]\n" - "ldr x25, [x16, #0x68]\n" - "fmla z26.s, p3/M, z1.s, z12.s\n" - "fmla z17.s, p3/M, z8.s, z12.s\n" - "incw x13\n" - "mov p1.b, p2.b\n" - "fmla z18.s, p3/M, z7.s, z12.s\n" - "movprfx z28, z15\n fmla z28.s, p3/M, z6.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x11, x14, LSL #2]\n" - "ldr x11, [x16, #0x88]\n" - "fmla z21.s, p3/M, z7.s, z9.s\n" - "fmla z19.s, p3/M, z6.s, z12.s\n" - "ldr x23, [x28, #0x0]\n" - "ldr x22, [x28, #0x8]\n" - "movprfx z23, z15\n fmla z23.s, p3/M, z3.s, z12.s\n" - "movprfx z27, z15\n fmla z27.s, p3/M, z0.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n" - "ldr x24, [x16, #0x78]\n" - "movprfx z31, z15\n fmla z31.s, p3/M, z8.s, z11.s\n" - "fmla z22.s, p3/M, z6.s, z9.s\n" - "ld1w { z11.s }, p2/Z, [x12, x14, LSL #2]\n" - "ldr x12, [x16, #0x80]\n" - "fmla z25.s, p3/M, z4.s, z9.s\n" - "fmla z26.s, p3/M, z3.s, z9.s\n" - "ldr x21, [x28, #0x10]\n" - "ldr x20, [x28, #0x18]\n" - "fmla z20.s, p3/M, z8.s, z9.s\n" - "fmla z24.s, p3/M, z5.s, z9.s\n" - "fmla z28.s, p3/M, z2.s, z9.s\n" - "fmla z16.s, p3/M, z1.s, z12.s\n" - "fmla z17.s, p3/M, z0.s, z12.s\n" - "movprfx z29, z15\n fmla z29.s, p3/M, z1.s, z9.s\n" - "movprfx z30, z15\n fmla z30.s, p3/M, z0.s, z9.s\n" - "fmla z18.s, p3/M, z2.s, z11.s\n" - "ld1w { z9.s }, p2/Z, [x10, x14, LSL #2]\n" - "ldr x10, [x16, #0x90]\n" - "fmla z21.s, p3/M, z8.s, z10.s\n" - "fmla z19.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x27, x14, LSL #2]\n" - "ldr x27, [x16, #0xa0]\n" - "fmla z22.s, p3/M, z7.s, z10.s\n" - "fmla z23.s, p3/M, z6.s, z10.s\n" - "fmla z25.s, p3/M, z5.s, z10.s\n" - "fmla z26.s, p3/M, z4.s, z10.s\n" - "fmla z27.s, p3/M, z3.s, z10.s\n" - "fmla z29.s, p3/M, z2.s, z10.s\n" - "fmla z30.s, p3/M, z1.s, z10.s\n" - "fmla z31.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x25, x14, LSL #2]\n" - "ldr x25, [x16, #0xa8]\n" - "fmla z16.s, p3/M, z3.s, z9.s\n" - "fmla z20.s, p3/M, z0.s, z9.s\n" - "ld1w { z12.s }, p2/Z, [x9, x14, LSL #2]\n" - "ldr x9, [x16, #0x98]\n" - "fmla z24.s, p3/M, z6.s, z11.s\n" - "fmla z28.s, p3/M, z3.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x26, x14, LSL #2]\n" - "ldr x26, [x16, #0xb0]\n" - "fmla z17.s, p3/M, z4.s, z10.s\n" - "fmla z18.s, p3/M, z3.s, z10.s\n" - "fmla z21.s, p3/M, z1.s, z10.s\n" - "fmla z19.s, p3/M, z5.s, z12.s\n" - "fmla z23.s, p3/M, z2.s, z12.s\n" - "fmla z22.s, p3/M, z0.s, z10.s\n" - "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n" - "ldr x24, [x16, #0xb8]\n" - "fmla z27.s, p3/M, z8.s, z11.s\n" - "fmla z31.s, p3/M, z5.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x12, x14, LSL #2]\n" - "ldr x12, [x16, #0xc0]\n" - "fmla z16.s, p3/M, z5.s, z10.s\n" - "fmla z20.s, p3/M, z2.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x11, x14, LSL #2]\n" - "ldr x11, [x16, #0xc8]\n" - "fmla z17.s, p3/M, z5.s, z12.s\n" - "fmla z18.s, p3/M, z4.s, z12.s\n" - "fmla z21.s, p3/M, z2.s, z12.s\n" - "fmla z19.s, p3/M, z3.s, z12.s\n" - "fmla z22.s, p3/M, z1.s, z12.s\n" - "fmla z23.s, p3/M, z0.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x9, x14, LSL #2]\n" - "ldr x9, [x16, #0xd8]\n" - "fmla z28.s, p3/M, z7.s, z11.s\n" - "fmla z29.s, p3/M, z6.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x10, x14, LSL #2]\n" - "ldr x10, [x16, #0xd0]\n" - "fmla z16.s, p3/M, z7.s, z10.s\n" - "fmla z17.s, p3/M, z6.s, z10.s\n" - "fmla z20.s, p3/M, z4.s, z10.s\n" - "fmla z21.s, p3/M, z3.s, z10.s\n" - "fmla z24.s, p3/M, z1.s, z10.s\n" - "fmla z25.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x27, x14, LSL #2]\n" - "ldr x27, [x16, #0xe0]\n" - "fmla z18.s, p3/M, z8.s, z12.s\n" - "fmla z30.s, p3/M, z8.s, z11.s\n" - "fmla z31.s, p3/M, z7.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x25, x14, LSL #2]\n" - "fmla z27.s, p3/M, z1.s, z12.s\n" - "ldr x25, [x16, #0xe8]\n" - "fmla z19.s, p3/M, z7.s, z12.s\n" - "fmla z22.s, p3/M, z5.s, z12.s\n" - "fmla z23.s, p3/M, z4.s, z12.s\n" - "fmla z26.s, p3/M, z2.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x26, x14, LSL #2]\n" - "ldr x26, [x16, #0xf0]\n" - "fmla z16.s, p3/M, z2.s, z10.s\n" - "fmla z17.s, p3/M, z1.s, z10.s\n" + "movprfx z14, z17\n fmla z14.s, p3/M, z4.s, z9.s\n" + "movprfx z18, z17\n fmla z18.s, p3/M, z8.s, z9.s\n" + "ldr x27, [x8, #0x20]\n" + "ldr x24, [x8, #0x30]\n" + "movprfx z15, z17\n fmla z15.s, p3/M, z3.s, z9.s\n" + "movprfx z30, z17\n fmla z30.s, p3/M, z1.s, z9.s\n" + "ldr x23, [x8, #0x28]\n" + "ldr x22, [x8, #0x38]\n" + "movprfx z20, z17\n fmla z20.s, p3/M, z0.s, z9.s\n" + "movprfx z13, z17\n fmla z13.s, p3/M, z7.s, z9.s\n" + "ldr x26, [x8, #0x40]\n" + "ldr x21, [x8, #0x48]\n" + "movprfx z22, z17\n fmla z22.s, p3/M, z6.s, z9.s\n" + "fmla z14.s, p3/M, z5.s, z12.s\n" + "ldr x25, [x8, #0x50]\n" + "ldr x20, [x8, #0x58]\n" + "movprfx z27, z17\n fmla z27.s, p3/M, z5.s, z9.s\n" + "movprfx z31, z17\n fmla z31.s, p3/M, z2.s, z9.s\n" + "ld1w { z23.s }, p2/Z, [x24, x16, LSL #2]\n" + "ldr x13, [x8, #0x70]\n" "fmla z18.s, p3/M, z0.s, z10.s\n" - "fmla z20.s, p3/M, z7.s, z11.s\n" - "ld1w { z10.s }, p2/Z, [x24, x14, LSL #2]\n" - "ldr x24, [x16, #0xf8]\n" - "fmla z21.s, p3/M, z6.s, z11.s\n" - "fmla z24.s, p3/M, z4.s, z11.s\n" - "fmla z25.s, p3/M, z3.s, z11.s\n" - "fmla z28.s, p3/M, z1.s, z11.s\n" - "fmla z29.s, p3/M, z0.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x12, x14, LSL #2]\n" - "fmla z27.s, p3/M, z4.s, z11.s\n" - "ldr x12, [x16, #0x100]\n" - "fmla z30.s, p3/M, z2.s, z11.s\n" - "fmla z17.s, p3/M, z2.s, z12.s\n" + "movprfx z9, z17\n fmla z9.s, p3/M, z2.s, z11.s\n" + "ld1w { z21.s }, p2/Z, [x27, x16, LSL #2]\n" + "ld1w { z25.s }, p2/Z, [x23, x16, LSL #2]\n" + "fmla z15.s, p3/M, z4.s, z12.s\n" + "fmla z30.s, p3/M, z2.s, z12.s\n" + "ldr x24, [x8, #0x60]\n" + "ldr x23, [x8, #0x68]\n" + "fmla z20.s, p3/M, z1.s, z12.s\n" + "fmla z13.s, p3/M, z8.s, z12.s\n" + "incw x15\n" + "mov p0.b, p2.b\n" + "fmla z22.s, p3/M, z7.s, z12.s\n" + "movprfx z28, z17\n fmla z28.s, p3/M, z6.s, z21.s\n" + "ld1w { z29.s }, p2/Z, [x21, x16, LSL #2]\n" + "ldr x28, [x8, #0x88]\n" + "fmla z14.s, p3/M, z7.s, z23.s\n" + "fmla z9.s, p3/M, z6.s, z12.s\n" + "ldr x12, [x14, #0x0]\n" + "ldr x11, [x14, #0x8]\n" + "movprfx z11, z17\n fmla z11.s, p3/M, z3.s, z12.s\n" + "movprfx z10, z17\n fmla z10.s, p3/M, z0.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x22, x16, LSL #2]\n" + "ldr x22, [x8, #0x78]\n" + "movprfx z26, z17\n fmla z26.s, p3/M, z8.s, z25.s\n" + "fmla z15.s, p3/M, z6.s, z23.s\n" + "ld1w { z21.s }, p2/Z, [x26, x16, LSL #2]\n" + "ldr x21, [x8, #0x80]\n" + "fmla z30.s, p3/M, z4.s, z23.s\n" + "fmla z20.s, p3/M, z3.s, z23.s\n" + "ldr x10, [x14, #0x10]\n" + "ldr x9, [x14, #0x18]\n" + "movprfx z25, z17\n fmla z25.s, p3/M, z1.s, z23.s\n" + "movprfx z24, z17\n fmla z24.s, p3/M, z0.s, z23.s\n" + "fmla z27.s, p3/M, z8.s, z23.s\n" + "fmla z31.s, p3/M, z5.s, z23.s\n" + "fmla z28.s, p3/M, z2.s, z23.s\n" "fmla z18.s, p3/M, z1.s, z12.s\n" - "fmla z19.s, p3/M, z0.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x11, x14, LSL #2]\n" - "ldr x11, [x16, #0x108]\n" - "fmla z16.s, p3/M, z6.s, z10.s\n" - "fmla z20.s, p3/M, z3.s, z10.s\n" - "fmla z24.s, p3/M, z0.s, z10.s\n" - "fmla z22.s, p3/M, z8.s, z11.s\n" - "ld1w { z10.s }, p2/Z, [x10, x14, LSL #2]\n" - "ldr x10, [x16, #0x110]\n" - "fmla z23.s, p3/M, z7.s, z11.s\n" - "fmla z26.s, p3/M, z5.s, z11.s\n" - "fmla z31.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x9, x14, LSL #2]\n" - "fmla z27.s, p3/M, z2.s, z12.s\n" - "ldr x9, [x16, #0x118]\n" - "fmla z28.s, p3/M, z0.s, z10.s\n" - "fmla z29.s, p3/M, z4.s, z11.s\n" - "fmla z30.s, p3/M, z3.s, z11.s\n" - "fmla z19.s, p3/M, z8.s, z12.s\n" - "fmla z23.s, p3/M, z5.s, z12.s\n" - "fmla z20.s, p3/M, z6.s, z10.s\n" - "ld1w { z12.s }, p2/Z, [x27, x14, LSL #2]\n" - "fmla z24.s, p3/M, z3.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x25, x14, LSL #2]\n" - "fmla z25.s, p3/M, z7.s, z11.s\n" - "fmla z26.s, p3/M, z6.s, z11.s\n" - "fmla z28.s, p3/M, z5.s, z11.s\n" - "fmla z27.s, p3/M, z5.s, z12.s\n" - "fmla z31.s, p3/M, z2.s, z12.s\n" - "fmla z29.s, p3/M, z7.s, z10.s\n" - "fmla z30.s, p3/M, z6.s, z10.s\n" - "fmla z24.s, p3/M, z8.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x26, x14, LSL #2]\n" - "fmla z28.s, p3/M, z8.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x12, x14, LSL #2]\n" - "fmla z25.s, p3/M, z8.s, z11.s\n" - "fmla z26.s, p3/M, z7.s, z11.s\n" - "fmla z27.s, p3/M, z6.s, z11.s\n" - "fmla z29.s, p3/M, z5.s, z11.s\n" - "fmla z30.s, p3/M, z4.s, z11.s\n" - "fmla z31.s, p3/M, z3.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x11, x14, LSL #2]\n" - "fmla z23.s, p3/M, z8.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n" - "fmla z16.s, p3/M, z4.s, z10.s\n" - "fmax z16.s, p3/M, z16.s, z14.s\n" - "fmla z17.s, p3/M, z3.s, z10.s\n" - "fmla z18.s, p3/M, z5.s, z11.s\n" - "fmax z17.s, p3/M, z17.s, z14.s\n" - "fmax z18.s, p3/M, z18.s, z14.s\n" - "fmla z19.s, p3/M, z4.s, z11.s\n" - "fmla z29.s, p3/M, z8.s, z12.s\n" - "fmax z19.s, p3/M, z19.s, z14.s\n" - "fmin z16.s, p3/M, z16.s, z13.s\n" - "fmla z30.s, p3/M, z7.s, z12.s\n" - "fmla z31.s, p3/M, z6.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x10, x14, LSL #2]\n" - "fmin z17.s, p3/M, z17.s, z13.s\n" - "fmla z20.s, p3/M, z1.s, z10.s\n" - "fmla z21.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x9, x14, LSL #2]\n" - "fmin z18.s, p3/M, z18.s, z13.s\n" - "fmla z22.s, p3/M, z2.s, z11.s\n" - "fmla z23.s, p3/M, z1.s, z11.s\n" - "fmin z19.s, p3/M, z19.s, z13.s\n" - "fmax z20.s, p3/M, z20.s, z14.s\n" - "fmla z24.s, p3/M, z7.s, z12.s\n" - "fmla z25.s, p3/M, z6.s, z12.s\n" - "fmax z21.s, p3/M, z21.s, z14.s\n" - "fmax z22.s, p3/M, z22.s, z14.s\n" - "fmla z26.s, p3/M, z8.s, z10.s\n" - "fmla z27.s, p3/M, z7.s, z10.s\n" - "fmax z23.s, p3/M, z23.s, z14.s\n" - "st1w { z16.s }, p1, [x23, x13, LSL #2]\n" - "st1w { z17.s }, p1, [x22, x13, LSL #2]\n" - "ldr x23, [x28, #0x20]\n" - "ldr x22, [x28, #0x28]\n" - "fmla z28.s, p3/M, z4.s, z12.s\n" - "st1w { z18.s }, p1, [x21, x13, LSL #2]\n" - "ldr x21, [x28, #0x30]\n" - "fmla z29.s, p3/M, z3.s, z12.s\n" - "fmla z30.s, p3/M, z5.s, z10.s\n" - "st1w { z19.s }, p1, [x20, x13, LSL #2]\n" - "ldr x20, [x28, #0x38]\n" - "fmla z31.s, p3/M, z4.s, z10.s\n" - "fmin z20.s, p3/M, z20.s, z13.s\n" - "fmin z21.s, p3/M, z21.s, z13.s\n" - "fmin z22.s, p3/M, z22.s, z13.s\n" - "st1w { z20.s }, p1, [x23, x13, LSL #2]\n" - "ldr x23, [x28, #0x40]\n" - "fmin z23.s, p3/M, z23.s, z13.s\n" - "fmax z24.s, p3/M, z24.s, z14.s\n" - "st1w { z21.s }, p1, [x22, x13, LSL #2]\n" - "ldr x22, [x28, #0x48]\n" - "fmax z25.s, p3/M, z25.s, z14.s\n" - "fmax z26.s, p3/M, z26.s, z14.s\n" - "st1w { z22.s }, p1, [x21, x13, LSL #2]\n" - "ldr x21, [x28, #0x50]\n" - "fmax z27.s, p3/M, z27.s, z14.s\n" - "st1w { z23.s }, p1, [x20, x13, LSL #2]\n" - "ldr x20, [x28, #0x58]\n" - "fmin z24.s, p3/M, z24.s, z13.s\n" - "fmin z25.s, p3/M, z25.s, z13.s\n" - "fmin z26.s, p3/M, z26.s, z13.s\n" - "st1w { z24.s }, p1, [x23, x13, LSL #2]\n" - "ldr x23, [x28, #0x60]\n" - "fmin z27.s, p3/M, z27.s, z13.s\n" - "fmax z28.s, p3/M, z28.s, z14.s\n" - "st1w { z25.s }, p1, [x22, x13, LSL #2]\n" - "ldr x22, [x28, #0x68]\n" - "fmax z29.s, p3/M, z29.s, z14.s\n" - "fmax z30.s, p3/M, z30.s, z14.s\n" - "st1w { z26.s }, p1, [x21, x13, LSL #2]\n" - "ldr x21, [x28, #0x70]\n" - "fmax z31.s, p3/M, z31.s, z14.s\n" - "st1w { z27.s }, p1, [x20, x13, LSL #2]\n" - "ldr x20, [x28, #0x78]\n" - "fmin z28.s, p3/M, z28.s, z13.s\n" - "fmin z29.s, p3/M, z29.s, z13.s\n" - "fmin z30.s, p3/M, z30.s, z13.s\n" - "st1w { z28.s }, p1, [x23, x13, LSL #2]\n" - "fmin z31.s, p3/M, z31.s, z13.s\n" - "st1w { z29.s }, p1, [x22, x13, LSL #2]\n" - "st1w { z30.s }, p1, [x21, x13, LSL #2]\n" - "st1w { z31.s }, p1, [x20, x13, LSL #2]\n" + "ld1w { z23.s }, p2/Z, [x25, x16, LSL #2]\n" + "ldr x27, [x8, #0x90]\n" + "fmla z13.s, p3/M, z0.s, z12.s\n" + "fmla z22.s, p3/M, z2.s, z21.s\n" + "ld1w { z12.s }, p2/Z, [x20, x16, LSL #2]\n" + "ldr x20, [x8, #0x98]\n" + "fmla z14.s, p3/M, z8.s, z29.s\n" + "fmla z9.s, p3/M, z1.s, z21.s\n" + "ld1w { z21.s }, p2/Z, [x24, x16, LSL #2]\n" + "ldr x26, [x8, #0xa0]\n" + "fmla z15.s, p3/M, z7.s, z29.s\n" + "fmla z11.s, p3/M, z6.s, z29.s\n" + "fmla z30.s, p3/M, z5.s, z29.s\n" + "fmla z20.s, p3/M, z4.s, z29.s\n" + "fmla z10.s, p3/M, z3.s, z29.s\n" + "fmla z25.s, p3/M, z2.s, z29.s\n" + "fmla z24.s, p3/M, z1.s, z29.s\n" + "fmla z26.s, p3/M, z0.s, z29.s\n" + "ld1w { z29.s }, p2/Z, [x23, x16, LSL #2]\n" + "ldr x25, [x8, #0xa8]\n" + "fmla z18.s, p3/M, z3.s, z23.s\n" + "fmla z27.s, p3/M, z0.s, z23.s\n" + "fmla z31.s, p3/M, z6.s, z21.s\n" + "fmla z28.s, p3/M, z3.s, z21.s\n" + "ld1w { z21.s }, p2/Z, [x13, x16, LSL #2]\n" + "ldr x24, [x8, #0xb0]\n" + "fmla z13.s, p3/M, z4.s, z29.s\n" + "fmla z22.s, p3/M, z3.s, z29.s\n" + "fmla z14.s, p3/M, z1.s, z29.s\n" + "fmla z9.s, p3/M, z5.s, z12.s\n" + "fmla z11.s, p3/M, z2.s, z12.s\n" + "fmla z15.s, p3/M, z0.s, z29.s\n" + "ld1w { z17.s }, p2/Z, [x22, x16, LSL #2]\n" + "ldr x23, [x8, #0xb8]\n" + "fmla z10.s, p3/M, z8.s, z21.s\n" + "fmla z26.s, p3/M, z5.s, z21.s\n" + "ld1w { z23.s }, p2/Z, [x21, x16, LSL #2]\n" + "ldr x22, [x8, #0xc0]\n" + "fmla z18.s, p3/M, z5.s, z29.s\n" + "fmla z27.s, p3/M, z2.s, z29.s\n" + "ld1w { z21.s }, p2/Z, [x28, x16, LSL #2]\n" + "ldr x21, [x8, #0xc8]\n" + "fmla z13.s, p3/M, z5.s, z17.s\n" + "fmla z22.s, p3/M, z4.s, z17.s\n" + "fmla z14.s, p3/M, z2.s, z17.s\n" + "fmla z9.s, p3/M, z3.s, z17.s\n" + "fmla z15.s, p3/M, z1.s, z17.s\n" + "fmla z11.s, p3/M, z0.s, z17.s\n" + "ld1w { z29.s }, p2/Z, [x20, x16, LSL #2]\n" + "ldr x28, [x8, #0xd8]\n" + "fmla z28.s, p3/M, z7.s, z23.s\n" + "fmla z25.s, p3/M, z6.s, z23.s\n" + "ld1w { z23.s }, p2/Z, [x27, x16, LSL #2]\n" + "ldr x20, [x8, #0xd0]\n" + "fmla z18.s, p3/M, z7.s, z21.s\n" + "fmla z13.s, p3/M, z6.s, z21.s\n" + "fmla z27.s, p3/M, z4.s, z21.s\n" + "fmla z14.s, p3/M, z3.s, z21.s\n" + "fmla z31.s, p3/M, z1.s, z21.s\n" + "fmla z30.s, p3/M, z0.s, z21.s\n" + "ld1w { z21.s }, p2/Z, [x26, x16, LSL #2]\n" + "ldr x27, [x8, #0xe0]\n" + "fmla z22.s, p3/M, z8.s, z29.s\n" + "fmla z24.s, p3/M, z8.s, z23.s\n" + "fmla z26.s, p3/M, z7.s, z23.s\n" + "ld1w { z23.s }, p2/Z, [x25, x16, LSL #2]\n" + "fmla z10.s, p3/M, z1.s, z29.s\n" + "ldr x26, [x8, #0xe8]\n" + "fmla z9.s, p3/M, z7.s, z29.s\n" + "fmla z15.s, p3/M, z5.s, z29.s\n" + "fmla z11.s, p3/M, z4.s, z29.s\n" + "fmla z20.s, p3/M, z2.s, z29.s\n" + "ld1w { z29.s }, p2/Z, [x24, x16, LSL #2]\n" + "ldr x25, [x8, #0xf0]\n" + "fmla z18.s, p3/M, z2.s, z21.s\n" + "fmla z13.s, p3/M, z1.s, z21.s\n" + "fmla z22.s, p3/M, z0.s, z21.s\n" + "fmla z27.s, p3/M, z7.s, z23.s\n" + "ld1w { z21.s }, p2/Z, [x23, x16, LSL #2]\n" + "ldr x24, [x8, #0xf8]\n" + "fmla z14.s, p3/M, z6.s, z23.s\n" + "fmla z31.s, p3/M, z4.s, z23.s\n" + "fmla z30.s, p3/M, z3.s, z23.s\n" + "fmla z28.s, p3/M, z1.s, z23.s\n" + "fmla z25.s, p3/M, z0.s, z23.s\n" + "ld1w { z17.s }, p2/Z, [x22, x16, LSL #2]\n" + "fmla z10.s, p3/M, z4.s, z17.s\n" + "ldr x23, [x8, #0x100]\n" + "fmla z24.s, p3/M, z2.s, z17.s\n" + "fmla z13.s, p3/M, z2.s, z29.s\n" + "fmla z22.s, p3/M, z1.s, z29.s\n" + "fmla z9.s, p3/M, z0.s, z29.s\n" + "ld1w { z23.s }, p2/Z, [x21, x16, LSL #2]\n" + "ldr x22, [x8, #0x108]\n" + "fmla z18.s, p3/M, z6.s, z21.s\n" + "fmla z27.s, p3/M, z3.s, z21.s\n" + "fmla z31.s, p3/M, z0.s, z21.s\n" + "fmla z15.s, p3/M, z8.s, z17.s\n" + "ld1w { z29.s }, p2/Z, [x20, x16, LSL #2]\n" + "ldr x21, [x8, #0x110]\n" + "fmla z11.s, p3/M, z7.s, z17.s\n" + "fmla z20.s, p3/M, z5.s, z17.s\n" + "fmla z26.s, p3/M, z1.s, z17.s\n" + "ld1w { z21.s }, p2/Z, [x28, x16, LSL #2]\n" + "fmla z10.s, p3/M, z2.s, z23.s\n" + "ldr x20, [x8, #0x118]\n" + "fmla z28.s, p3/M, z0.s, z29.s\n" + "fmla z25.s, p3/M, z4.s, z21.s\n" + "fmla z24.s, p3/M, z3.s, z21.s\n" + "fmla z9.s, p3/M, z8.s, z23.s\n" + "fmla z11.s, p3/M, z5.s, z23.s\n" + "fmla z27.s, p3/M, z6.s, z29.s\n" + "ld1w { z23.s }, p2/Z, [x27, x16, LSL #2]\n" + "fmla z31.s, p3/M, z3.s, z29.s\n" + "ld1w { z17.s }, p2/Z, [x26, x16, LSL #2]\n" + "fmla z30.s, p3/M, z7.s, z21.s\n" + "fmla z20.s, p3/M, z6.s, z21.s\n" + "fmla z28.s, p3/M, z5.s, z21.s\n" + "fmla z10.s, p3/M, z5.s, z23.s\n" + "fmla z26.s, p3/M, z2.s, z23.s\n" + "fmla z25.s, p3/M, z7.s, z17.s\n" + "fmla z24.s, p3/M, z6.s, z17.s\n" + "fmla z31.s, p3/M, z8.s, z21.s\n" + "ld1w { z21.s }, p2/Z, [x25, x16, LSL #2]\n" + "fmla z28.s, p3/M, z8.s, z17.s\n" + "ld1w { z12.s }, p2/Z, [x23, x16, LSL #2]\n" + "fmla z30.s, p3/M, z8.s, z21.s\n" + "fmla z20.s, p3/M, z7.s, z21.s\n" + "fmla z10.s, p3/M, z6.s, z21.s\n" + "fmla z25.s, p3/M, z5.s, z21.s\n" + "fmla z24.s, p3/M, z4.s, z21.s\n" + "fmla z26.s, p3/M, z3.s, z21.s\n" + "ld1w { z21.s }, p2/Z, [x22, x16, LSL #2]\n" + "fmla z11.s, p3/M, z8.s, z23.s\n" + "ld1w { z29.s }, p2/Z, [x24, x16, LSL #2]\n" + "fmla z18.s, p3/M, z4.s, z12.s\n" + "fmax z18.s, p3/M, z18.s, z16.s\n" + "fmla z13.s, p3/M, z3.s, z12.s\n" + "fmla z22.s, p3/M, z5.s, z21.s\n" + "fmax z13.s, p3/M, z13.s, z16.s\n" + "fmax z22.s, p3/M, z22.s, z16.s\n" + "fmla z9.s, p3/M, z4.s, z21.s\n" + "fmla z25.s, p3/M, z8.s, z29.s\n" + "fmax z9.s, p3/M, z9.s, z16.s\n" + "fmin z18.s, p3/M, z18.s, z19.s\n" + "fmla z24.s, p3/M, z7.s, z29.s\n" + "fmla z26.s, p3/M, z6.s, z29.s\n" + "ld1w { z23.s }, p2/Z, [x21, x16, LSL #2]\n" + "fmin z13.s, p3/M, z13.s, z19.s\n" + "fmla z27.s, p3/M, z1.s, z12.s\n" + "fmla z14.s, p3/M, z0.s, z12.s\n" + "ld1w { z29.s }, p2/Z, [x20, x16, LSL #2]\n" + "fmin z22.s, p3/M, z22.s, z19.s\n" + "fmla z15.s, p3/M, z2.s, z21.s\n" + "fmla z11.s, p3/M, z1.s, z21.s\n" + "fmin z9.s, p3/M, z9.s, z19.s\n" + "fmax z27.s, p3/M, z27.s, z16.s\n" + "fmla z31.s, p3/M, z7.s, z23.s\n" + "fmla z30.s, p3/M, z6.s, z23.s\n" + "fmax z14.s, p3/M, z14.s, z16.s\n" + "fmax z15.s, p3/M, z15.s, z16.s\n" + "fmla z20.s, p3/M, z8.s, z29.s\n" + "fmla z10.s, p3/M, z7.s, z29.s\n" + "fmax z11.s, p3/M, z11.s, z16.s\n" + "st1w { z18.s }, p0, [x12, x15, LSL #2]\n" + "st1w { z13.s }, p0, [x11, x15, LSL #2]\n" + "ldr x23, [x14, #0x20]\n" + "ldr x22, [x14, #0x28]\n" + "fmla z28.s, p3/M, z4.s, z23.s\n" + "st1w { z22.s }, p0, [x10, x15, LSL #2]\n" + "ldr x21, [x14, #0x30]\n" + "fmla z25.s, p3/M, z3.s, z23.s\n" + "fmla z24.s, p3/M, z5.s, z29.s\n" + "st1w { z9.s }, p0, [x9, x15, LSL #2]\n" + "ldr x20, [x14, #0x38]\n" + "fmla z26.s, p3/M, z4.s, z29.s\n" + "fmin z27.s, p3/M, z27.s, z19.s\n" + "fmin z14.s, p3/M, z14.s, z19.s\n" + "fmin z15.s, p3/M, z15.s, z19.s\n" + "st1w { z27.s }, p0, [x23, x15, LSL #2]\n" + "ldr x23, [x14, #0x40]\n" + "fmin z11.s, p3/M, z11.s, z19.s\n" + "fmax z31.s, p3/M, z31.s, z16.s\n" + "st1w { z14.s }, p0, [x22, x15, LSL #2]\n" + "ldr x22, [x14, #0x48]\n" + "fmax z30.s, p3/M, z30.s, z16.s\n" + "fmax z20.s, p3/M, z20.s, z16.s\n" + "st1w { z15.s }, p0, [x21, x15, LSL #2]\n" + "ldr x21, [x14, #0x50]\n" + "fmax z10.s, p3/M, z10.s, z16.s\n" + "st1w { z11.s }, p0, [x20, x15, LSL #2]\n" + "ldr x20, [x14, #0x58]\n" + "fmin z31.s, p3/M, z31.s, z19.s\n" + "fmin z30.s, p3/M, z30.s, z19.s\n" + "fmin z20.s, p3/M, z20.s, z19.s\n" + "st1w { z31.s }, p0, [x23, x15, LSL #2]\n" + "ldr x23, [x14, #0x60]\n" + "fmin z10.s, p3/M, z10.s, z19.s\n" + "fmax z28.s, p3/M, z28.s, z16.s\n" + "st1w { z30.s }, p0, [x22, x15, LSL #2]\n" + "ldr x22, [x14, #0x68]\n" + "fmax z25.s, p3/M, z25.s, z16.s\n" + "fmax z24.s, p3/M, z24.s, z16.s\n" + "st1w { z20.s }, p0, [x21, x15, LSL #2]\n" + "ldr x21, [x14, #0x70]\n" + "fmax z26.s, p3/M, z26.s, z16.s\n" + "st1w { z10.s }, p0, [x20, x15, LSL #2]\n" + "ldr x20, [x14, #0x78]\n" + "fmin z28.s, p3/M, z28.s, z19.s\n" + "fmin z25.s, p3/M, z25.s, z19.s\n" + "fmin z24.s, p3/M, z24.s, z19.s\n" + "st1w { z28.s }, p0, [x23, x15, LSL #2]\n" + "fmin z26.s, p3/M, z26.s, z19.s\n" + "st1w { z25.s }, p0, [x22, x15, LSL #2]\n" + "st1w { z24.s }, p0, [x21, x15, LSL #2]\n" + "st1w { z26.s }, p0, [x20, x15, LSL #2]\n" : : [n_channels] "r" ((unsigned long) n_channels), [offsetof_Args_inptrs] "I" (offsetof(Args, inptrs)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_args_params] "I" (offsetof(Args, params)), [params_struct] "r" (¶ms_struct) - : "cc", "memory", "p0", "p1", "p2", "p3", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } } // namespace depthwise } // namespace arm_conv -#endif // __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) +#endif // defined(ARM_COMPUTE_ENABLE_SVE) -- cgit v1.2.1