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_direct.cpp | 884 ++++++++++----------- 1 file changed, 442 insertions(+), 442 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp index cdf77a1cf0..35445595f8 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.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 { @@ -113,7 +113,7 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl( "mul x21, x21, x25\n" // offset *= kernel_stride * output_size "add x8, x8, x21, LSL #2\n" // inptr[0] += offset * sizeof(float) "add x13, x8, x23, LSL #2\n" - "ld1w { z15.s }, p3/Z, [x17]\n" + "ld1w { z19.s }, p3/Z, [x17]\n" "mul x20, x20, x24\n" // offset *= output_tile_size "add x12, x13, x23, LSL #2\n" "add x15, x15, x20, LSL #2\n" // outptrs[0] += offset * sizeof(float) @@ -132,8 +132,8 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl( "add x27, x10, x5\n" "add x26, x9, x22, LSL #2\n" "add x25, x6, x6\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" + "ld1rw { z15.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" + "ld1rw { z16.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" "cmp x16, %x[n_channels]\n" "add x24, x28, x23, LSL #2\n" "ld1w { z7.s }, p3/Z, [x17, #-8, MUL VL]\n" @@ -149,500 +149,500 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl( "addvl x17, x17, #-6\n" "bge 3f\n" "2:" // Tile loop: 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" + "movprfx z14, z19\n fmla z14.s, p3/M, z4.s, z9.s\n" + "movprfx z31, z19\n fmla z31.s, p3/M, z8.s, z9.s\n" "whilelt p1.s, x16, %x[n_channels]\n" "incw x21\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" + "movprfx z21, z19\n fmla z21.s, p3/M, z3.s, z9.s\n" + "movprfx z22, z19\n fmla z22.s, p3/M, z1.s, z9.s\n" "incw x16\n" "mov p0.b, p2.b\n" - "movprfx z26, z15\n fmla z26.s, p3/M, z0.s, z9.s\n" - "fmla z21.s, p3/M, z5.s, z12.s\n" + "movprfx z20, z19\n fmla z20.s, p3/M, z0.s, z9.s\n" + "fmla z14.s, p3/M, z5.s, z12.s\n" "incw x20\n" - "movprfx z17, z15\n fmla z17.s, p3/M, z7.s, z9.s\n" - "movprfx z18, z15\n fmla z18.s, p3/M, z6.s, z9.s\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" + "movprfx z13, z19\n fmla z13.s, p3/M, z7.s, z9.s\n" + "movprfx z17, z19\n fmla z17.s, p3/M, z6.s, z9.s\n" + "movprfx z27, z19\n fmla z27.s, p3/M, z5.s, z9.s\n" + "movprfx z18, z19\n fmla z18.s, p3/M, z2.s, z9.s\n" "ld1w { z9.s }, p2/Z, [x11, x7, LSL #2]\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, [x24]\n" + "fmla z31.s, p3/M, z0.s, z10.s\n" + "movprfx z30, z19\n fmla z30.s, p3/M, z2.s, z11.s\n" + "ld1w { z29.s }, p2/Z, [x24]\n" "ld1w { z11.s }, p2/Z, [x24, x27, LSL #2]\n" - "fmla z22.s, p3/M, z4.s, z12.s\n" - "fmla z25.s, p3/M, z2.s, z12.s\n" - "fmla z26.s, p3/M, z1.s, z12.s\n" - "movprfx z28, z15\n fmla z28.s, p3/M, z6.s, z10.s\n" + "fmla z21.s, p3/M, z4.s, z12.s\n" + "fmla z22.s, p3/M, z2.s, z12.s\n" + "fmla z20.s, p3/M, z1.s, z12.s\n" + "movprfx z23, z19\n fmla z23.s, p3/M, z6.s, z29.s\n" "ld1w { z10.s }, p2/Z, [x11, x14, LSL #2]\n" - "fmla z21.s, p3/M, z7.s, z9.s\n" - "fmla z17.s, p3/M, z8.s, z12.s\n" - "fmla z18.s, p3/M, z7.s, z12.s\n" - "fmla z19.s, p3/M, z6.s, z12.s\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" + "fmla z14.s, p3/M, z7.s, z9.s\n" + "fmla z13.s, p3/M, z8.s, z12.s\n" + "fmla z17.s, p3/M, z7.s, z12.s\n" + "fmla z30.s, p3/M, z6.s, z12.s\n" + "movprfx z26, z19\n fmla z26.s, p3/M, z3.s, z12.s\n" + "movprfx z28, z19\n fmla z28.s, p3/M, z0.s, z12.s\n" "ld1w { z12.s }, p2/Z, [x8, x5, LSL #2]\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" + "movprfx z24, z19\n fmla z24.s, p3/M, z8.s, z11.s\n" + "fmla z21.s, p3/M, z6.s, z9.s\n" "ld1w { z11.s }, p2/Z, [x8, x10, LSL #2]\n" - "fmla z25.s, p3/M, z4.s, z9.s\n" - "fmla z26.s, p3/M, z3.s, z9.s\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 z21.s, p3/M, z8.s, z10.s\n" - "fmla z16.s, p3/M, z1.s, z12.s\n" - "fmla z17.s, p3/M, z0.s, z12.s\n" + "fmla z22.s, p3/M, z4.s, z9.s\n" + "fmla z20.s, p3/M, z3.s, z9.s\n" + "movprfx z25, z19\n fmla z25.s, p3/M, z1.s, z9.s\n" + "movprfx z29, z19\n fmla z29.s, p3/M, z0.s, z9.s\n" + "ld1w { z19.s }, p3/Z, [x17]\n" + "fmla z27.s, p3/M, z8.s, z9.s\n" + "fmla z18.s, p3/M, z5.s, z9.s\n" + "fmla z23.s, p3/M, z2.s, z9.s\n" + "fmla z14.s, p3/M, z8.s, z10.s\n" + "ld1w { z9.s }, p2/Z, [x13]\n" + "fmla z31.s, p3/M, z1.s, z12.s\n" + "fmla z13.s, p3/M, z0.s, z12.s\n" "ld1w { z12.s }, p2/Z, [x13, x27, LSL #2]\n" - "fmla z18.s, p3/M, z2.s, z11.s\n" - "fmla z19.s, p3/M, z1.s, z11.s\n" + "fmla z17.s, p3/M, z2.s, z11.s\n" + "fmla z30.s, p3/M, z1.s, z11.s\n" "ld1w { z11.s }, p2/Z, [x28]\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 z31.s, p3/M, z0.s, z10.s\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, [x28, x27, LSL #2]\n" - "fmla z19.s, p3/M, z5.s, z12.s\n" - "fmla z23.s, p3/M, z2.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x13, x14, LSL #2]\n" - "fmla z27.s, p3/M, z8.s, z11.s\n" - "fmla z31.s, p3/M, z5.s, z11.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" - "ld1w { z9.s }, p2/Z, [x13]\n" - "fmla z29.s, p3/M, z2.s, z10.s\n" - "fmla z30.s, p3/M, z1.s, z10.s\n" + "fmla z21.s, p3/M, z7.s, z10.s\n" + "fmla z26.s, p3/M, z6.s, z10.s\n" + "fmla z22.s, p3/M, z5.s, z10.s\n" + "fmla z20.s, p3/M, z4.s, z10.s\n" + "fmla z28.s, p3/M, z3.s, z10.s\n" + "fmla z25.s, p3/M, z2.s, z10.s\n" + "fmla z29.s, p3/M, z1.s, z10.s\n" + "fmla z24.s, p3/M, z0.s, z10.s\n" "ld1w { z10.s }, p2/Z, [x13, x7, LSL #2]\n" - "fmla z20.s, p3/M, z0.s, z9.s\n" - "fmla z21.s, p3/M, z1.s, z10.s\n" - "fmla z16.s, p3/M, z3.s, z9.s\n" - "fmla z17.s, p3/M, z4.s, z10.s\n" - "ld1w { z11.s }, p2/Z, [x24, x5, LSL #2]\n" - "fmla z18.s, p3/M, z3.s, z10.s\n" - "fmla z22.s, p3/M, z0.s, z10.s\n" + "fmla z27.s, p3/M, z0.s, z9.s\n" + "fmla z18.s, p3/M, z6.s, z11.s\n" + "fmla z23.s, p3/M, z3.s, z11.s\n" + "fmla z14.s, p3/M, z1.s, z10.s\n" + "ld1w { z11.s }, p2/Z, [x28, x27, LSL #2]\n" + "fmla z31.s, p3/M, z3.s, z9.s\n" + "fmla z30.s, p3/M, z5.s, z12.s\n" + "fmla z26.s, p3/M, z2.s, z12.s\n" + "fmla z13.s, p3/M, z4.s, z10.s\n" + "ld1w { z9.s }, p2/Z, [x13, x14, LSL #2]\n" + "fmla z17.s, p3/M, z3.s, z10.s\n" + "fmla z21.s, p3/M, z0.s, z10.s\n" + "fmla z28.s, p3/M, z8.s, z11.s\n" + "fmla z24.s, p3/M, z5.s, z11.s\n" + "ld1w { z12.s }, p2/Z, [x24, x5, LSL #2]\n" + "fmla z27.s, p3/M, z2.s, z10.s\n" + "fmla z14.s, p3/M, z2.s, z9.s\n" + "fmla z31.s, p3/M, z5.s, z10.s\n" + "fmla z13.s, p3/M, z5.s, z9.s\n" + "ld1w { z11.s }, p2/Z, [x12, x5, LSL #2]\n" + "fmla z17.s, p3/M, z4.s, z9.s\n" + "fmla z30.s, p3/M, z3.s, z9.s\n" + "fmla z21.s, p3/M, z1.s, z9.s\n" + "fmla z26.s, p3/M, z0.s, z9.s\n" + "ld1w { z10.s }, p2/Z, [x12, x10, LSL #2]\n" + "fmla z23.s, p3/M, z7.s, z12.s\n" + "fmla z25.s, p3/M, z6.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x24, x10, LSL #2]\n" + "fmla z27.s, p3/M, z4.s, z11.s\n" + "fmla z14.s, p3/M, z3.s, z11.s\n" + "fmla z18.s, p3/M, z1.s, z11.s\n" + "fmla z22.s, p3/M, z0.s, z11.s\n" + "fmla z31.s, p3/M, z7.s, z11.s\n" + "fmla z13.s, p3/M, z6.s, z11.s\n" + "ld1w { z9.s }, p2/Z, [x8, x7, LSL #2]\n" + "fmla z29.s, p3/M, z8.s, z12.s\n" + "fmla z24.s, p3/M, z7.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x11, x5, LSL #2]\n" + "fmla z17.s, p3/M, z8.s, z10.s\n" + "fmla z30.s, p3/M, z7.s, z10.s\n" + "fmla z21.s, p3/M, z5.s, z10.s\n" + "fmla z26.s, p3/M, z4.s, z10.s\n" "fmla z20.s, p3/M, z2.s, z10.s\n" - "fmla z21.s, p3/M, z2.s, z12.s\n" - "fmla z16.s, p3/M, z5.s, z10.s\n" - "fmla z17.s, p3/M, z5.s, z12.s\n" - "ld1w { z10.s }, p2/Z, [x12, x5, LSL #2]\n" + "fmla z28.s, p3/M, z1.s, z10.s\n" + "ld1w { z11.s }, p2/Z, [x8, x14, LSL #2]\n" + "addvl x8, x8, #1\n" + "fmla z27.s, p3/M, z7.s, z12.s\n" + "fmla z14.s, p3/M, z6.s, z12.s\n" "fmla z18.s, p3/M, z4.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, [x12, x10, LSL #2]\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, [x24, x10, LSL #2]\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" - "fmla z16.s, p3/M, z7.s, z10.s\n" - "fmla z17.s, p3/M, z6.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x8, x7, LSL #2]\n" + "fmla z22.s, p3/M, z3.s, z12.s\n" + "fmla z23.s, p3/M, z1.s, z12.s\n" + "fmla z25.s, p3/M, z0.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x11, x10, LSL #2]\n" + "fmla z31.s, p3/M, z2.s, z9.s\n" + "fmla z13.s, p3/M, z1.s, z9.s\n" + "fmla z17.s, p3/M, z0.s, z9.s\n" + "ld1w { z9.s }, p2/Z, [x12]\n" + "fmla z29.s, p3/M, z2.s, z12.s\n" + "fmla z30.s, p3/M, z0.s, z11.s\n" + "fmla z27.s, p3/M, z3.s, z9.s\n" + "fmla z18.s, p3/M, z0.s, z9.s\n" + "fmla z21.s, p3/M, z8.s, z12.s\n" + "fmla z26.s, p3/M, z7.s, z12.s\n" + "fmla z20.s, p3/M, z5.s, z12.s\n" + "fmla z28.s, p3/M, z4.s, z12.s\n" + "fmla z24.s, p3/M, z1.s, z12.s\n" + "ld1w { z10.s }, p2/Z, [x28, x7, LSL #2]\n" + "fmla z13.s, p3/M, z2.s, z11.s\n" + "fmla z17.s, p3/M, z1.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x12, x27, LSL #2]\n" + "addvl x12, x12, #1\n" + "fmla z31.s, p3/M, z6.s, z9.s\n" + "ld1w { z12.s }, p2/Z, [x11]\n" + "fmla z25.s, p3/M, z4.s, z10.s\n" + "ld1w { z9.s }, p1/Z, [x12, x7, LSL #2]\n" + "fmla z29.s, p3/M, z3.s, z10.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, [x11, x5, LSL #2]\n" - "fmla z18.s, p3/M, z8.s, z12.s\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" - "fmla z27.s, p3/M, z1.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x8, x14, LSL #2]\n" - "addvl x8, x8, #1\n" - "fmla z20.s, p3/M, z7.s, z11.s\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, [x11, x10, LSL #2]\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" - "ld1w { z10.s }, p2/Z, [x12]\n" - "fmla z30.s, p3/M, z2.s, z11.s\n" - "fmla z19.s, p3/M, z0.s, z12.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" - "fmla z23.s, p3/M, z7.s, z11.s\n" "fmla z26.s, p3/M, z5.s, z11.s\n" - "fmla z27.s, p3/M, z4.s, z11.s\n" - "fmla z31.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x28, x7, LSL #2]\n" - "fmla z17.s, p3/M, z2.s, z12.s\n" - "fmla z18.s, p3/M, z1.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x12, x27, LSL #2]\n" - "addvl x12, x12, #1\n" - "fmla z16.s, p3/M, z6.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x11]\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 z27.s, p3/M, z2.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x11, x27, LSL #2]\n" + "fmla z28.s, p3/M, z2.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x11, x27, LSL #2]\n" "addvl x11, x11, #1\n" + "fmla z27.s, p3/M, z6.s, z12.s\n" + "fmla z18.s, p3/M, z3.s, z12.s\n" + "fmla z23.s, p3/M, z0.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x24, x7, LSL #2]\n" + "fmla z24.s, p3/M, z2.s, z11.s\n" + "fmla z25.s, p3/M, z7.s, z12.s\n" + "fmla z29.s, p3/M, z6.s, z12.s\n" + "fmla z18.s, p3/M, z8.s, z10.s\n" + "fmla z22.s, p3/M, z7.s, z10.s\n" "fmla z20.s, p3/M, z6.s, z10.s\n" - "fmla z24.s, p3/M, z3.s, z10.s\n" - "fmla z28.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x24, x7, LSL #2]\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" - "fmla z25.s, p3/M, z7.s, z11.s\n" - "fmla z26.s, p3/M, z6.s, z11.s\n" + "fmla z23.s, p3/M, z5.s, z10.s\n" + "ld1w { z10.s }, p2/Z, [x28, x14, LSL #2]\n" "fmla z28.s, p3/M, z5.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x28, x14, LSL #2]\n" - "fmla z27.s, p3/M, z5.s, z12.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" + "fmla z25.s, p3/M, z5.s, z10.s\n" + "fmla z29.s, p3/M, z4.s, z10.s\n" + "fmla z24.s, p3/M, z3.s, z10.s\n" + "fmla z26.s, p3/M, z8.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x24, 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 z28.s, p3/M, z8.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x13, x5, LSL #2]\n" - "fmla z25.s, p3/M, z8.s, z11.s\n" - "fmla z26.s, p3/M, z7.s, z11.s\n" + "ld1w { z12.s }, p2/Z, [x13, x5, LSL #2]\n" + "fmla z22.s, p3/M, z8.s, z10.s\n" + "fmla z20.s, p3/M, z7.s, z10.s\n" "addvl x24, x24, #1\n" - "fmla z27.s, p3/M, z6.s, z11.s\n" - "fmla z29.s, p3/M, z8.s, z12.s\n" - "ld1w { z11.s }, p2/Z, [x13, x10, LSL #2]\n" + "fmla z28.s, p3/M, z6.s, z10.s\n" + "fmla z25.s, p3/M, z8.s, z11.s\n" + "ld1w { z10.s }, p2/Z, [x13, x10, LSL #2]\n" "addvl x13, x13, #1\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, [x28, x5, LSL #2]\n" - "fmla z16.s, p3/M, z4.s, z10.s\n" - "fmla z17.s, p3/M, z3.s, z10.s\n" - "fmax z16.s, p3/M, z16.s, z14.s\n" - "fmax z17.s, p3/M, z17.s, z14.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, [x28, x10, LSL #2]\n" - "fmax z20.s, p3/M, z20.s, z14.s\n" - "fmla z18.s, p3/M, z5.s, z11.s\n" - "fmla z19.s, p3/M, z4.s, z11.s\n" - "fmax z18.s, p3/M, z18.s, z14.s\n" - "fmax z19.s, p3/M, z19.s, z14.s\n" - "fmla z22.s, p3/M, z2.s, z11.s\n" - "fmla z23.s, p3/M, z1.s, z11.s\n" - "fmax z21.s, p3/M, z21.s, z14.s\n" - "fmax z22.s, p3/M, z22.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 z23.s, p3/M, z23.s, z14.s\n" - "fmax z24.s, p3/M, z24.s, z14.s\n" - "fmla z28.s, p3/M, z4.s, z12.s\n" - "fmla z29.s, p3/M, z3.s, z12.s\n" - "fmax z25.s, p3/M, z25.s, z14.s\n" - "fmax z28.s, p3/M, z28.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 z26.s, p3/M, z26.s, z14.s\n" - "fmax z27.s, p3/M, z27.s, z14.s\n" - "fmla z30.s, p3/M, z5.s, z10.s\n" - "fmla z31.s, p3/M, z4.s, z10.s\n" - "fmax z29.s, p3/M, z29.s, z14.s\n" - "fmax z30.s, p3/M, z30.s, z14.s\n" - "fmax z31.s, p3/M, z31.s, z14.s\n" - "ld1w { z15.s }, p3/Z, [x17]\n" + "fmla z29.s, p3/M, z7.s, z11.s\n" + "fmla z24.s, p3/M, z6.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x28, x5, LSL #2]\n" + "fmla z31.s, p3/M, z4.s, z12.s\n" + "fmla z13.s, p3/M, z3.s, z12.s\n" + "fmax z31.s, p3/M, z31.s, z15.s\n" + "fmax z13.s, p3/M, z13.s, z15.s\n" + "fmla z27.s, p3/M, z1.s, z12.s\n" + "fmla z14.s, p3/M, z0.s, z12.s\n" + "ld1w { z0.s }, p2/Z, [x28, x10, LSL #2]\n" + "fmax z27.s, p3/M, z27.s, z15.s\n" + "fmla z17.s, p3/M, z5.s, z10.s\n" + "fmla z30.s, p3/M, z4.s, z10.s\n" + "fmax z17.s, p3/M, z17.s, z15.s\n" + "fmax z30.s, p3/M, z30.s, z15.s\n" + "fmla z21.s, p3/M, z2.s, z10.s\n" + "fmla z26.s, p3/M, z1.s, z10.s\n" + "fmax z14.s, p3/M, z14.s, z15.s\n" + "fmax z21.s, p3/M, z21.s, z15.s\n" + "fmla z18.s, p3/M, z7.s, z11.s\n" + "fmla z22.s, p3/M, z6.s, z11.s\n" + "fmax z26.s, p3/M, z26.s, z15.s\n" + "fmax z18.s, p3/M, z18.s, z15.s\n" + "fmla z23.s, p3/M, z4.s, z11.s\n" + "fmla z25.s, p3/M, z3.s, z11.s\n" + "fmax z22.s, p3/M, z22.s, z15.s\n" + "fmax z23.s, p3/M, z23.s, z15.s\n" + "fmla z20.s, p3/M, z8.s, z0.s\n" + "fmla z28.s, p3/M, z7.s, z0.s\n" + "fmax z20.s, p3/M, z20.s, z15.s\n" + "fmax z28.s, p3/M, z28.s, z15.s\n" + "fmla z29.s, p3/M, z5.s, z0.s\n" + "fmla z24.s, p3/M, z4.s, z0.s\n" + "fmax z25.s, p3/M, z25.s, z15.s\n" + "fmax z29.s, p3/M, z29.s, z15.s\n" + "fmax z24.s, p3/M, z24.s, z15.s\n" "ld1w { z0.s }, p3/Z, [x17, #1, MUL VL]\n" - "whilelt p2.s, x21, %x[n_channels]\n" "ld1w { z1.s }, p3/Z, [x17, #2, MUL VL]\n" + "whilelt p2.s, x21, %x[n_channels]\n" "ld1w { z2.s }, p3/Z, [x17, #3, MUL VL]\n" - "cmp x16, %x[n_channels]\n" - "fmin z16.s, p3/M, z16.s, z13.s\n" "ld1w { z3.s }, p3/Z, [x17, #4, MUL VL]\n" + "cmp x16, %x[n_channels]\n" + "fmin z31.s, p3/M, z31.s, z16.s\n" "ld1w { z4.s }, p3/Z, [x17, #5, MUL VL]\n" - "fmin z17.s, p3/M, z17.s, z13.s\n" - "fmin z18.s, p3/M, z18.s, z13.s\n" "ld1w { z5.s }, p3/Z, [x17, #6, MUL VL]\n" + "fmin z13.s, p3/M, z13.s, z16.s\n" + "fmin z17.s, p3/M, z17.s, z16.s\n" "ld1w { z6.s }, p3/Z, [x17, #7, MUL VL]\n" "addvl x17, x17, #16\n" - "fmin z19.s, p3/M, z19.s, z13.s\n" - "fmin z20.s, p3/M, z20.s, z13.s\n" - "fmin z21.s, p3/M, z21.s, z13.s\n" - "ld1w { z9.s }, p1/Z, [x12, x7, LSL #2]\n" + "fmin z30.s, p3/M, z30.s, z16.s\n" "ld1w { z10.s }, p1/Z, [x8]\n" - "fmin z22.s, p3/M, z22.s, z13.s\n" - "fmin z23.s, p3/M, z23.s, z13.s\n" + "fmin z27.s, p3/M, z27.s, z16.s\n" + "fmin z14.s, p3/M, z14.s, z16.s\n" "ld1w { z11.s }, p1/Z, [x8, x27, LSL #2]\n" "ld1w { z12.s }, p1/Z, [x12, x14, LSL #2]\n" - "fmin z24.s, p3/M, z24.s, z13.s\n" - "fmin z25.s, p3/M, z25.s, z13.s\n" - "st1w { z16.s }, p0, [x15]\n" + "fmin z21.s, p3/M, z21.s, z16.s\n" + "fmin z26.s, p3/M, z26.s, z16.s\n" + "st1w { z31.s }, p0, [x15]\n" "ld1w { z7.s }, p3/Z, [x17, #-8, MUL VL]\n" - "fmin z26.s, p3/M, z26.s, z13.s\n" - "fmin z27.s, p3/M, z27.s, z13.s\n" - "st1w { z17.s }, p0, [x15, x6, LSL #2]\n" + "fmin z18.s, p3/M, z18.s, z16.s\n" + "fmin z22.s, p3/M, z22.s, z16.s\n" + "st1w { z13.s }, p0, [x15, x6, LSL #2]\n" "ld1w { z8.s }, p3/Z, [x17, #-7, MUL VL]\n" - "fmin z28.s, p3/M, z28.s, z13.s\n" - "fmin z29.s, p3/M, z29.s, z13.s\n" - "st1w { z18.s }, p0, [x15, x25, LSL #2]\n" - "fmin z30.s, p3/M, z30.s, z13.s\n" - "fmin z31.s, p3/M, z31.s, z13.s\n" - "st1w { z19.s }, p0, [x15, x22, LSL #2]\n" + "fmin z20.s, p3/M, z20.s, z16.s\n" + "fmin z28.s, p3/M, z28.s, z16.s\n" + "st1w { z17.s }, p0, [x15, x25, LSL #2]\n" + "fmin z23.s, p3/M, z23.s, z16.s\n" + "fmin z25.s, p3/M, z25.s, z16.s\n" + "st1w { z30.s }, p0, [x15, x22, LSL #2]\n" + "fmin z29.s, p3/M, z29.s, z16.s\n" + "fmin z24.s, p3/M, z24.s, z16.s\n" + "st1w { z27.s }, p0, [x9]\n" "addvl x28, x28, #1\n" - "st1w { z20.s }, p0, [x9]\n" + "st1w { z14.s }, p0, [x9, x6, LSL #2]\n" "addvl x15, x15, #1\n" - "st1w { z21.s }, p0, [x9, x6, LSL #2]\n" + "st1w { z21.s }, p0, [x9, x25, LSL #2]\n" "addvl x17, x17, #-6\n" - "st1w { z22.s }, p0, [x9, x25, LSL #2]\n" - "st1w { z23.s }, p0, [x9, x22, LSL #2]\n" + "st1w { z26.s }, p0, [x9, x22, LSL #2]\n" "addvl x9, x9, #1\n" - "st1w { z24.s }, p0, [x26]\n" - "st1w { z25.s }, p0, [x26, x6, LSL #2]\n" - "st1w { z26.s }, p0, [x26, x25, LSL #2]\n" - "st1w { z27.s }, p0, [x26, x22, LSL #2]\n" + "st1w { z18.s }, p0, [x26]\n" + "st1w { z22.s }, p0, [x26, x6, LSL #2]\n" + "st1w { z20.s }, p0, [x26, x25, LSL #2]\n" + "st1w { z28.s }, p0, [x26, x22, LSL #2]\n" "addvl x26, x26, #1\n" - "st1w { z28.s }, p0, [x23]\n" - "st1w { z29.s }, p0, [x23, x6, LSL #2]\n" - "st1w { z30.s }, p0, [x23, x25, LSL #2]\n" - "st1w { z31.s }, p0, [x23, x22, LSL #2]\n" + "st1w { z23.s }, p0, [x23]\n" + "st1w { z25.s }, p0, [x23, x6, LSL #2]\n" + "st1w { z29.s }, p0, [x23, x25, LSL #2]\n" + "st1w { z24.s }, p0, [x23, x22, LSL #2]\n" "addvl x23, x23, #1\n" "blt 2b\n" "3:" // Tile loop: 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" + "movprfx z14, z19\n fmla z14.s, p3/M, z4.s, z9.s\n" + "movprfx z31, z19\n fmla z31.s, p3/M, z8.s, z9.s\n" "ldr x4, [%x[params_struct], %[offsetof_args_tile_j]]\n" "ldr x16, [%x[params_struct], %[offsetof_args_tile_i]]\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" + "movprfx z30, z19\n fmla z30.s, p3/M, z3.s, z9.s\n" + "movprfx z13, z19\n fmla z13.s, p3/M, z1.s, z9.s\n" "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" "add x4, x4, #0x1\n" - "movprfx z26, z15\n fmla z26.s, p3/M, z0.s, z9.s\n" - "fmla z21.s, p3/M, z5.s, z12.s\n" + "movprfx z20, z19\n fmla z20.s, p3/M, z0.s, z9.s\n" + "fmla z14.s, p3/M, z5.s, z12.s\n" "cmp x4, x20\n" "add x21, x16, #0x1\n" - "movprfx z17, z15\n fmla z17.s, p3/M, z7.s, z9.s\n" - "movprfx z18, z15\n fmla z18.s, p3/M, z6.s, z9.s\n" + "movprfx z18, z19\n fmla z18.s, p3/M, z7.s, z9.s\n" + "movprfx z28, z19\n fmla z28.s, p3/M, z6.s, z9.s\n" "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" "csel x16, x16, x21, LT\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" + "movprfx z17, z19\n fmla z17.s, p3/M, z5.s, z9.s\n" + "movprfx z26, z19\n fmla z26.s, p3/M, z2.s, z9.s\n" "ld1w { z9.s }, p2/Z, [x11, x7, LSL #2]\n" "mov p0.b, p2.b\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, [x24]\n" - "ld1w { z11.s }, p2/Z, [x24, x27, LSL #2]\n" - "fmla z22.s, p3/M, z4.s, z12.s\n" - "fmla z25.s, p3/M, z2.s, z12.s\n" + "fmla z31.s, p3/M, z0.s, z10.s\n" + "movprfx z27, z19\n fmla z27.s, p3/M, z2.s, z11.s\n" + "ld1w { z29.s }, p2/Z, [x24]\n" + "ld1w { z21.s }, p2/Z, [x24, x27, LSL #2]\n" + "fmla z30.s, p3/M, z4.s, z12.s\n" + "fmla z13.s, p3/M, z2.s, z12.s\n" "csel x4, x4, XZR, LT\n" "cmp x16, x20\n" - "fmla z26.s, p3/M, z1.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" - "fmla z21.s, p3/M, z7.s, z9.s\n" - "fmla z17.s, p3/M, z8.s, z12.s\n" - "fmla z18.s, p3/M, z7.s, z12.s\n" - "fmla z19.s, p3/M, z6.s, z12.s\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, [x8, x5, LSL #2]\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, [x8, x10, LSL #2]\n" - "fmla z25.s, p3/M, z4.s, z9.s\n" - "fmla z26.s, p3/M, z3.s, z9.s\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 z21.s, p3/M, z8.s, z10.s\n" - "fmla z16.s, p3/M, z1.s, z12.s\n" - "fmla z17.s, p3/M, z0.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x13, x27, LSL #2]\n" - "fmla z18.s, p3/M, z2.s, z11.s\n" - "fmla z19.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x28]\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 z31.s, p3/M, z0.s, z10.s\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, [x28, x27, LSL #2]\n" - "fmla z19.s, p3/M, z5.s, z12.s\n" - "fmla z23.s, p3/M, z2.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x13, x14, LSL #2]\n" - "fmla z27.s, p3/M, z8.s, z11.s\n" - "fmla z31.s, p3/M, z5.s, z11.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" - "ld1w { z9.s }, p2/Z, [x13]\n" - "fmla z29.s, p3/M, z2.s, z10.s\n" - "fmla z30.s, p3/M, z1.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x13, x7, LSL #2]\n" - "fmla z20.s, p3/M, z0.s, z9.s\n" - "fmla z21.s, p3/M, z1.s, z10.s\n" - "fmla z16.s, p3/M, z3.s, z9.s\n" - "fmla z17.s, p3/M, z4.s, z10.s\n" - "ld1w { z11.s }, p2/Z, [x24, x5, LSL #2]\n" - "fmla z18.s, p3/M, z3.s, z10.s\n" - "fmla z22.s, p3/M, z0.s, z10.s\n" - "fmla z20.s, p3/M, z2.s, z10.s\n" - "fmla z21.s, p3/M, z2.s, z12.s\n" - "fmla z16.s, p3/M, z5.s, z10.s\n" - "fmla z17.s, p3/M, z5.s, z12.s\n" - "ld1w { z10.s }, p2/Z, [x12, x5, LSL #2]\n" - "fmla z18.s, p3/M, z4.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, [x12, x10, LSL #2]\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, [x24, x10, LSL #2]\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" - "fmla z16.s, p3/M, z7.s, z10.s\n" - "fmla z17.s, p3/M, z6.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x8, x7, LSL #2]\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, [x11, x5, LSL #2]\n" + "fmla z20.s, p3/M, z1.s, z12.s\n" + "movprfx z10, z19\n fmla z10.s, p3/M, z6.s, z29.s\n" + "ld1w { z29.s }, p2/Z, [x11, x14, LSL #2]\n" + "fmla z14.s, p3/M, z7.s, z9.s\n" "fmla z18.s, p3/M, z8.s, z12.s\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" - "fmla z27.s, p3/M, z1.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x8, x14, LSL #2]\n" - "fmla z20.s, p3/M, z7.s, z11.s\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, [x11, x10, LSL #2]\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" - "ld1w { z10.s }, p2/Z, [x12]\n" - "fmla z30.s, p3/M, z2.s, z11.s\n" - "fmla z19.s, p3/M, z0.s, z12.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" - "fmla z23.s, p3/M, z7.s, z11.s\n" - "fmla z26.s, p3/M, z5.s, z11.s\n" - "fmla z27.s, p3/M, z4.s, z11.s\n" - "fmla z31.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x28, x7, LSL #2]\n" - "fmla z17.s, p3/M, z2.s, z12.s\n" - "fmla z18.s, p3/M, z1.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x12, x27, LSL #2]\n" - "fmla z16.s, p3/M, z6.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x11]\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 z27.s, p3/M, z2.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x11, x27, LSL #2]\n" - "fmla z20.s, p3/M, z6.s, z10.s\n" - "fmla z24.s, p3/M, z3.s, z10.s\n" - "fmla z28.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x24, x7, LSL #2]\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" - "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" - "ld1w { z11.s }, p2/Z, [x28, x14, LSL #2]\n" - "fmla z27.s, p3/M, z5.s, z12.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" - "fmla z23.s, p3/M, z8.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n" - "fmla z28.s, p3/M, z8.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x13, x5, 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, z8.s, z12.s\n" - "ld1w { z11.s }, p2/Z, [x13, x10, LSL #2]\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, [x28, x5, LSL #2]\n" - "fmla z16.s, p3/M, z4.s, z10.s\n" - "fmla z17.s, p3/M, z3.s, z10.s\n" - "fmax z16.s, p3/M, z16.s, z14.s\n" - "fmax z17.s, p3/M, z17.s, z14.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, [x28, x10, LSL #2]\n" - "fmax z20.s, p3/M, z20.s, z14.s\n" - "fmla z18.s, p3/M, z5.s, z11.s\n" - "fmla z19.s, p3/M, z4.s, z11.s\n" - "fmax z18.s, p3/M, z18.s, z14.s\n" - "fmax z19.s, p3/M, z19.s, z14.s\n" - "fmla z22.s, p3/M, z2.s, z11.s\n" - "fmla z23.s, p3/M, z1.s, z11.s\n" - "fmax z21.s, p3/M, z21.s, z14.s\n" - "fmax z22.s, p3/M, z22.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 z23.s, p3/M, z23.s, z14.s\n" - "fmax z24.s, p3/M, z24.s, z14.s\n" - "fmla z28.s, p3/M, z4.s, z12.s\n" - "fmla z29.s, p3/M, z3.s, z12.s\n" - "fmax z25.s, p3/M, z25.s, z14.s\n" - "fmax z28.s, p3/M, z28.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 z26.s, p3/M, z26.s, z14.s\n" - "fmax z27.s, p3/M, z27.s, z14.s\n" - "fmla z30.s, p3/M, z5.s, z10.s\n" - "fmla z31.s, p3/M, z4.s, z10.s\n" - "fmax z29.s, p3/M, z29.s, z14.s\n" - "fmax z30.s, p3/M, z30.s, z14.s\n" - "fmax z31.s, p3/M, z31.s, z14.s\n" - "fmin z16.s, p3/M, z16.s, z13.s\n" - "st1w { z16.s }, p0, [x15]\n" - "fmin z17.s, p3/M, z17.s, z13.s\n" - "fmin z18.s, p3/M, z18.s, z13.s\n" - "st1w { z17.s }, p0, [x15, x6, LSL #2]\n" - "fmin z19.s, p3/M, z19.s, z13.s\n" - "fmin z20.s, p3/M, z20.s, z13.s\n" - "st1w { z18.s }, p0, [x15, x25, LSL #2]\n" - "fmin z21.s, p3/M, z21.s, z13.s\n" - "fmin z22.s, p3/M, z22.s, z13.s\n" - "st1w { z19.s }, p0, [x15, x22, LSL #2]\n" - "fmin z23.s, p3/M, z23.s, z13.s\n" - "fmin z24.s, p3/M, z24.s, z13.s\n" - "st1w { z20.s }, p0, [x9]\n" - "fmin z25.s, p3/M, z25.s, z13.s\n" - "fmin z26.s, p3/M, z26.s, z13.s\n" - "st1w { z21.s }, p0, [x9, x6, LSL #2]\n" - "fmin z27.s, p3/M, z27.s, z13.s\n" - "fmin z28.s, p3/M, z28.s, z13.s\n" - "st1w { z22.s }, p0, [x9, x25, LSL #2]\n" - "fmin z29.s, p3/M, z29.s, z13.s\n" - "fmin z30.s, p3/M, z30.s, z13.s\n" - "st1w { z23.s }, p0, [x9, x22, LSL #2]\n" - "fmin z31.s, p3/M, z31.s, z13.s\n" - "st1w { z24.s }, p0, [x26]\n" - "st1w { z25.s }, p0, [x26, x6, LSL #2]\n" - "st1w { z26.s }, p0, [x26, x25, LSL #2]\n" - "st1w { z27.s }, p0, [x26, x22, LSL #2]\n" - "st1w { z28.s }, p0, [x23]\n" - "st1w { z29.s }, p0, [x23, x6, LSL #2]\n" - "st1w { z30.s }, p0, [x23, x25, LSL #2]\n" - "st1w { z31.s }, p0, [x23, x22, LSL #2]\n" + "fmla z28.s, p3/M, z7.s, z12.s\n" + "fmla z27.s, p3/M, z6.s, z12.s\n" + "movprfx z11, z19\n fmla z11.s, p3/M, z3.s, z12.s\n" + "movprfx z25, z19\n fmla z25.s, p3/M, z0.s, z12.s\n" + "ld1w { z22.s }, p2/Z, [x8, x5, LSL #2]\n" + "movprfx z24, z19\n fmla z24.s, p3/M, z8.s, z21.s\n" + "fmla z30.s, p3/M, z6.s, z9.s\n" + "ld1w { z21.s }, p2/Z, [x8, x10, LSL #2]\n" + "fmla z13.s, p3/M, z4.s, z9.s\n" + "fmla z20.s, p3/M, z3.s, z9.s\n" + "movprfx z12, z19\n fmla z12.s, p3/M, z1.s, z9.s\n" + "movprfx z23, z19\n fmla z23.s, p3/M, z0.s, z9.s\n" + "fmla z17.s, p3/M, z8.s, z9.s\n" + "fmla z26.s, p3/M, z5.s, z9.s\n" + "fmla z10.s, p3/M, z2.s, z9.s\n" + "fmla z14.s, p3/M, z8.s, z29.s\n" + "ld1w { z9.s }, p2/Z, [x13]\n" + "fmla z31.s, p3/M, z1.s, z22.s\n" + "fmla z18.s, p3/M, z0.s, z22.s\n" + "ld1w { z22.s }, p2/Z, [x13, x27, LSL #2]\n" + "fmla z28.s, p3/M, z2.s, z21.s\n" + "fmla z27.s, p3/M, z1.s, z21.s\n" + "ld1w { z19.s }, p2/Z, [x28]\n" + "fmla z30.s, p3/M, z7.s, z29.s\n" + "fmla z11.s, p3/M, z6.s, z29.s\n" + "fmla z13.s, p3/M, z5.s, z29.s\n" + "fmla z20.s, p3/M, z4.s, z29.s\n" + "fmla z25.s, p3/M, z3.s, z29.s\n" + "fmla z12.s, p3/M, z2.s, z29.s\n" + "fmla z23.s, p3/M, z1.s, z29.s\n" + "fmla z24.s, p3/M, z0.s, z29.s\n" + "ld1w { z21.s }, p2/Z, [x13, x7, LSL #2]\n" + "fmla z17.s, p3/M, z0.s, z9.s\n" + "fmla z26.s, p3/M, z6.s, z19.s\n" + "fmla z10.s, p3/M, z3.s, z19.s\n" + "fmla z14.s, p3/M, z1.s, z21.s\n" + "ld1w { z19.s }, p2/Z, [x28, x27, LSL #2]\n" + "fmla z31.s, p3/M, z3.s, z9.s\n" + "fmla z27.s, p3/M, z5.s, z22.s\n" + "fmla z11.s, p3/M, z2.s, z22.s\n" + "fmla z18.s, p3/M, z4.s, z21.s\n" + "ld1w { z29.s }, p2/Z, [x13, x14, LSL #2]\n" + "fmla z28.s, p3/M, z3.s, z21.s\n" + "fmla z30.s, p3/M, z0.s, z21.s\n" + "fmla z25.s, p3/M, z8.s, z19.s\n" + "fmla z24.s, p3/M, z5.s, z19.s\n" + "ld1w { z19.s }, p2/Z, [x24, x5, LSL #2]\n" + "fmla z17.s, p3/M, z2.s, z21.s\n" + "fmla z14.s, p3/M, z2.s, z29.s\n" + "fmla z31.s, p3/M, z5.s, z21.s\n" + "fmla z18.s, p3/M, z5.s, z29.s\n" + "ld1w { z22.s }, p2/Z, [x12, x5, LSL #2]\n" + "fmla z28.s, p3/M, z4.s, z29.s\n" + "fmla z27.s, p3/M, z3.s, z29.s\n" + "fmla z30.s, p3/M, z1.s, z29.s\n" + "fmla z11.s, p3/M, z0.s, z29.s\n" + "ld1w { z21.s }, p2/Z, [x12, x10, LSL #2]\n" + "fmla z10.s, p3/M, z7.s, z19.s\n" + "fmla z12.s, p3/M, z6.s, z19.s\n" + "ld1w { z19.s }, p2/Z, [x24, x10, LSL #2]\n" + "fmla z17.s, p3/M, z4.s, z22.s\n" + "fmla z14.s, p3/M, z3.s, z22.s\n" + "fmla z26.s, p3/M, z1.s, z22.s\n" + "fmla z13.s, p3/M, z0.s, z22.s\n" + "fmla z31.s, p3/M, z7.s, z22.s\n" + "fmla z18.s, p3/M, z6.s, z22.s\n" + "ld1w { z29.s }, p2/Z, [x8, x7, LSL #2]\n" + "fmla z23.s, p3/M, z8.s, z19.s\n" + "fmla z24.s, p3/M, z7.s, z19.s\n" + "ld1w { z19.s }, p2/Z, [x11, x5, LSL #2]\n" + "fmla z28.s, p3/M, z8.s, z21.s\n" + "fmla z27.s, p3/M, z7.s, z21.s\n" + "fmla z30.s, p3/M, z5.s, z21.s\n" + "fmla z11.s, p3/M, z4.s, z21.s\n" + "fmla z20.s, p3/M, z2.s, z21.s\n" + "fmla z25.s, p3/M, z1.s, z21.s\n" + "ld1w { z22.s }, p2/Z, [x8, x14, LSL #2]\n" + "fmla z17.s, p3/M, z7.s, z19.s\n" + "fmla z14.s, p3/M, z6.s, z19.s\n" + "fmla z26.s, p3/M, z4.s, z19.s\n" + "fmla z13.s, p3/M, z3.s, z19.s\n" + "fmla z10.s, p3/M, z1.s, z19.s\n" + "fmla z12.s, p3/M, z0.s, z19.s\n" + "ld1w { z21.s }, p2/Z, [x11, x10, LSL #2]\n" + "fmla z31.s, p3/M, z2.s, z29.s\n" + "fmla z18.s, p3/M, z1.s, z29.s\n" + "fmla z28.s, p3/M, z0.s, z29.s\n" + "ld1w { z29.s }, p2/Z, [x12]\n" + "fmla z23.s, p3/M, z2.s, z21.s\n" + "fmla z27.s, p3/M, z0.s, z22.s\n" + "fmla z17.s, p3/M, z3.s, z29.s\n" + "fmla z26.s, p3/M, z0.s, z29.s\n" + "fmla z30.s, p3/M, z8.s, z21.s\n" + "fmla z11.s, p3/M, z7.s, z21.s\n" + "fmla z20.s, p3/M, z5.s, z21.s\n" + "fmla z25.s, p3/M, z4.s, z21.s\n" + "fmla z24.s, p3/M, z1.s, z21.s\n" + "ld1w { z19.s }, p2/Z, [x28, x7, LSL #2]\n" + "fmla z18.s, p3/M, z2.s, z22.s\n" + "fmla z28.s, p3/M, z1.s, z22.s\n" + "ld1w { z21.s }, p2/Z, [x12, x27, LSL #2]\n" + "fmla z31.s, p3/M, z6.s, z29.s\n" + "ld1w { z29.s }, p2/Z, [x11]\n" + "fmla z12.s, p3/M, z4.s, z19.s\n" + "fmla z23.s, p3/M, z3.s, z19.s\n" + "fmla z27.s, p3/M, z8.s, z21.s\n" + "fmla z11.s, p3/M, z5.s, z21.s\n" + "fmla z25.s, p3/M, z2.s, z21.s\n" + "ld1w { z9.s }, p2/Z, [x11, x27, LSL #2]\n" + "fmla z17.s, p3/M, z6.s, z29.s\n" + "fmla z26.s, p3/M, z3.s, z29.s\n" + "fmla z10.s, p3/M, z0.s, z29.s\n" + "ld1w { z22.s }, p2/Z, [x24, x7, LSL #2]\n" + "fmla z24.s, p3/M, z2.s, z9.s\n" + "fmla z12.s, p3/M, z7.s, z22.s\n" + "fmla z23.s, p3/M, z6.s, z22.s\n" + "fmla z26.s, p3/M, z8.s, z19.s\n" + "fmla z13.s, p3/M, z7.s, z19.s\n" + "fmla z20.s, p3/M, z6.s, z19.s\n" + "fmla z10.s, p3/M, z5.s, z19.s\n" + "ld1w { z21.s }, p2/Z, [x28, x14, LSL #2]\n" + "fmla z25.s, p3/M, z5.s, z9.s\n" + "fmla z12.s, p3/M, z5.s, z21.s\n" + "fmla z23.s, p3/M, z4.s, z21.s\n" + "fmla z24.s, p3/M, z3.s, z21.s\n" + "fmla z11.s, p3/M, z8.s, z9.s\n" + "ld1w { z19.s }, p2/Z, [x24, x14, LSL #2]\n" + "fmla z10.s, p3/M, z8.s, z22.s\n" + "ld1w { z22.s }, p2/Z, [x13, x5, LSL #2]\n" + "fmla z13.s, p3/M, z8.s, z21.s\n" + "fmla z20.s, p3/M, z7.s, z21.s\n" + "fmla z25.s, p3/M, z6.s, z21.s\n" + "fmla z12.s, p3/M, z8.s, z19.s\n" + "ld1w { z29.s }, p2/Z, [x13, x10, LSL #2]\n" + "fmla z23.s, p3/M, z7.s, z19.s\n" + "fmla z24.s, p3/M, z6.s, z19.s\n" + "ld1w { z21.s }, p2/Z, [x28, x5, LSL #2]\n" + "fmla z31.s, p3/M, z4.s, z22.s\n" + "fmla z18.s, p3/M, z3.s, z22.s\n" + "fmax z31.s, p3/M, z31.s, z15.s\n" + "fmax z18.s, p3/M, z18.s, z15.s\n" + "fmla z17.s, p3/M, z1.s, z22.s\n" + "fmla z14.s, p3/M, z0.s, z22.s\n" + "ld1w { z9.s }, p2/Z, [x28, x10, LSL #2]\n" + "fmax z17.s, p3/M, z17.s, z15.s\n" + "fmla z28.s, p3/M, z5.s, z29.s\n" + "fmla z27.s, p3/M, z4.s, z29.s\n" + "fmax z28.s, p3/M, z28.s, z15.s\n" + "fmax z27.s, p3/M, z27.s, z15.s\n" + "fmla z30.s, p3/M, z2.s, z29.s\n" + "fmla z11.s, p3/M, z1.s, z29.s\n" + "fmax z14.s, p3/M, z14.s, z15.s\n" + "fmax z30.s, p3/M, z30.s, z15.s\n" + "fmla z26.s, p3/M, z7.s, z21.s\n" + "fmla z13.s, p3/M, z6.s, z21.s\n" + "fmax z11.s, p3/M, z11.s, z15.s\n" + "fmax z26.s, p3/M, z26.s, z15.s\n" + "fmla z10.s, p3/M, z4.s, z21.s\n" + "fmla z12.s, p3/M, z3.s, z21.s\n" + "fmax z13.s, p3/M, z13.s, z15.s\n" + "fmax z10.s, p3/M, z10.s, z15.s\n" + "fmla z20.s, p3/M, z8.s, z9.s\n" + "fmla z25.s, p3/M, z7.s, z9.s\n" + "fmax z20.s, p3/M, z20.s, z15.s\n" + "fmax z25.s, p3/M, z25.s, z15.s\n" + "fmla z23.s, p3/M, z5.s, z9.s\n" + "fmla z24.s, p3/M, z4.s, z9.s\n" + "fmax z12.s, p3/M, z12.s, z15.s\n" + "fmax z23.s, p3/M, z23.s, z15.s\n" + "fmax z24.s, p3/M, z24.s, z15.s\n" + "fmin z31.s, p3/M, z31.s, z16.s\n" + "st1w { z31.s }, p0, [x15]\n" + "fmin z18.s, p3/M, z18.s, z16.s\n" + "fmin z28.s, p3/M, z28.s, z16.s\n" + "st1w { z18.s }, p0, [x15, x6, LSL #2]\n" + "fmin z27.s, p3/M, z27.s, z16.s\n" + "fmin z17.s, p3/M, z17.s, z16.s\n" + "st1w { z28.s }, p0, [x15, x25, LSL #2]\n" + "fmin z14.s, p3/M, z14.s, z16.s\n" + "fmin z30.s, p3/M, z30.s, z16.s\n" + "st1w { z27.s }, p0, [x15, x22, LSL #2]\n" + "fmin z11.s, p3/M, z11.s, z16.s\n" + "fmin z26.s, p3/M, z26.s, z16.s\n" + "st1w { z17.s }, p0, [x9]\n" + "fmin z13.s, p3/M, z13.s, z16.s\n" + "fmin z20.s, p3/M, z20.s, z16.s\n" + "st1w { z14.s }, p0, [x9, x6, LSL #2]\n" + "fmin z25.s, p3/M, z25.s, z16.s\n" + "fmin z10.s, p3/M, z10.s, z16.s\n" + "st1w { z30.s }, p0, [x9, x25, LSL #2]\n" + "fmin z12.s, p3/M, z12.s, z16.s\n" + "fmin z23.s, p3/M, z23.s, z16.s\n" + "st1w { z11.s }, p0, [x9, x22, LSL #2]\n" + "fmin z24.s, p3/M, z24.s, z16.s\n" + "st1w { z26.s }, p0, [x26]\n" + "st1w { z13.s }, p0, [x26, x6, LSL #2]\n" + "st1w { z20.s }, p0, [x26, x25, LSL #2]\n" + "st1w { z25.s }, p0, [x26, x22, LSL #2]\n" + "st1w { z10.s }, p0, [x23]\n" + "st1w { z12.s }, p0, [x23, x6, LSL #2]\n" + "st1w { z23.s }, p0, [x23, x25, LSL #2]\n" + "st1w { z24.s }, p0, [x23, x22, LSL #2]\n" "blt 1b\n" : : [n_channels] "r" ((unsigned long) n_channels), [offsetof_args_inptr] "I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col] "I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row] "I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col] "I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row] "I" (offsetof(Args, ld_output_row)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_n_tile_cols] "I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows] "I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr] "I" (offsetof(Args, outptr)), [offsetof_args_params] "I" (offsetof(Args, params)), [offsetof_args_tile_i] "I" (offsetof(Args, tile_i)), [offsetof_args_tile_j] "I" (offsetof(Args, tile_j)), [params_struct] "r" (¶ms_struct) @@ -653,4 +653,4 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl( } // namespace depthwise } // namespace arm_conv -#endif // __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) +#endif // defined(ARM_COMPUTE_ENABLE_SVE) -- cgit v1.2.1