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 | 710 ++++++++++----------- 1 file changed, 355 insertions(+), 355 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp index 5380567d36..efd37c38ec 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp @@ -22,11 +22,11 @@ * SOFTWARE. */ -#if defined(ARM_COMPUTE_ENABLE_SME2) - #include #include +#if defined(ARM_COMPUTE_ENABLE_SME2) + namespace arm_conv { namespace depthwise { @@ -193,18 +193,18 @@ void sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl( "ldr x22, [%x[params_struct], %[offsetof_args_ld_output_row]]\n" "mul x21, x2, x22\n" // offset = tile_i * ld_output_row "mov x20, #0x4\n" - "ld1w { z15.s }, p3/Z, [x17]\n" + "ld1w { z14.s }, p3/Z, [x17]\n" "ldr x9, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" "madd x21, x3, x9, x21\n" // offset += tile_j * ld_output_col "mul x21, x21, x20\n" // offset *= output_tile_size - "ld1rw { z14.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" + "ld1rw { z13.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" "ldr x28, [%x[params_struct], %[offsetof_args_outptr]]\n" "add x28, x28, x21, LSL #2\n" // outptrs[0] += offset * sizeof(float) "addvl x17, x17, #1\n" ".inst 0xa040c220 // ld1w { z0.s-z3.s }, pn8.b/Z, [x17]\n" "add x27, x28, x22, LSL #2\n" "cntw x26\n" - "ld1rw { z13.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" + "ld1rw { z15.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" "addvl x17, x17, #4\n" "add x25, x27, x22, LSL #2\n" ".inst 0xa040c224 // ld1w { z4.s-z7.s }, pn8.b/Z, [x17]\n" @@ -224,440 +224,440 @@ void sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl( "addvl x17, x17, #1\n" "bge 4f\n" "3:" // 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 z25, z14\n fmla z25.s, p3/M, z4.s, z9.s\n" + "movprfx z28, z14\n fmla z28.s, p3/M, z8.s, z9.s\n" "whilelt p1.s, x26, %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 z26, z14\n fmla z26.s, p3/M, z3.s, z9.s\n" + "movprfx z17, z14\n fmla z17.s, p3/M, z1.s, z9.s\n" "incw x26\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 z18, z14\n fmla z18.s, p3/M, z0.s, z9.s\n" + "fmla z25.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 z29, z14\n fmla z29.s, p3/M, z7.s, z9.s\n" + "movprfx z30, z14\n fmla z30.s, p3/M, z6.s, z9.s\n" + "movprfx z24, z14\n fmla z24.s, p3/M, z5.s, z9.s\n" + "movprfx z16, z14\n fmla z16.s, p3/M, z2.s, z9.s\n" "ld1w { z9.s }, p2/Z, [x16, x8, 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, [x12]\n" - "fmla z22.s, p3/M, z4.s, z12.s\n" - "fmla z25.s, p3/M, z2.s, z12.s\n" - "ld1w { z11.s }, p2/Z, [x12, x11, LSL #2]\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, [x16, x15, 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, [x5, x4, 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, [x5, x13, 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, [x6, x11, 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, [x14]\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, [x14, x11, 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, [x6, x15, 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" + "fmla z28.s, p3/M, z0.s, z10.s\n" + "movprfx z31, z14\n fmla z31.s, p3/M, z2.s, z11.s\n" + "ld1w { z19.s }, p2/Z, [x12]\n" + "fmla z26.s, p3/M, z4.s, z12.s\n" + "fmla z17.s, p3/M, z2.s, z12.s\n" + "ld1w { z22.s }, p2/Z, [x12, x11, LSL #2]\n" + "fmla z18.s, p3/M, z1.s, z12.s\n" + "movprfx z20, z14\n fmla z20.s, p3/M, z6.s, z19.s\n" + "ld1w { z11.s }, p2/Z, [x16, x15, LSL #2]\n" + "fmla z25.s, p3/M, z7.s, z9.s\n" + "fmla z29.s, p3/M, z8.s, z12.s\n" + "fmla z30.s, p3/M, z7.s, z12.s\n" + "fmla z31.s, p3/M, z6.s, z12.s\n" + "movprfx z27, z14\n fmla z27.s, p3/M, z3.s, z12.s\n" + "movprfx z19, z14\n fmla z19.s, p3/M, z0.s, z12.s\n" + "ld1w { z10.s }, p2/Z, [x5, x4, LSL #2]\n" + "movprfx z23, z14\n fmla z23.s, p3/M, z8.s, z22.s\n" + "fmla z26.s, p3/M, z6.s, z9.s\n" + "ld1w { z12.s }, p2/Z, [x5, x13, LSL #2]\n" + "fmla z17.s, p3/M, z4.s, z9.s\n" + "fmla z18.s, p3/M, z3.s, z9.s\n" + "movprfx z21, z14\n fmla z21.s, p3/M, z1.s, z9.s\n" + "movprfx z22, z14\n fmla z22.s, p3/M, z0.s, z9.s\n" + "ld1w { z14.s }, p3/Z, [x17]\n" + "addvl x17, x17, #1\n" + "fmla z24.s, p3/M, z8.s, z9.s\n" + "fmla z16.s, p3/M, z5.s, z9.s\n" + "fmla z20.s, p3/M, z2.s, z9.s\n" + "fmla z25.s, p3/M, z8.s, z11.s\n" "ld1w { z9.s }, p2/Z, [x6]\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, [x6, x8, 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, [x12, x4, 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, [x7, x4, 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, [x7, x13, LSL #2]\n" - "fmla z28.s, p3/M, z7.s, z11.s\n" - "fmla z29.s, p3/M, z6.s, z11.s\n" + "fmla z28.s, p3/M, z1.s, z10.s\n" + "fmla z29.s, p3/M, z0.s, z10.s\n" + "ld1w { z10.s }, p2/Z, [x6, x11, LSL #2]\n" + "fmla z30.s, p3/M, z2.s, z12.s\n" + "fmla z31.s, p3/M, z1.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x14]\n" + "fmla z26.s, p3/M, z7.s, z11.s\n" + "fmla z27.s, p3/M, z6.s, z11.s\n" + "fmla z17.s, p3/M, z5.s, z11.s\n" + "fmla z18.s, p3/M, z4.s, z11.s\n" + "fmla z19.s, p3/M, z3.s, z11.s\n" + "fmla z21.s, p3/M, z2.s, z11.s\n" + "fmla z22.s, p3/M, z1.s, z11.s\n" + "fmla z23.s, p3/M, z0.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x6, x8, LSL #2]\n" + "fmla z24.s, p3/M, z0.s, z9.s\n" + "fmla z16.s, p3/M, z6.s, z12.s\n" + "fmla z20.s, p3/M, z3.s, z12.s\n" + "fmla z25.s, p3/M, z1.s, z11.s\n" + "ld1w { z12.s }, p2/Z, [x14, x11, LSL #2]\n" + "fmla z28.s, p3/M, z3.s, z9.s\n" + "fmla z31.s, p3/M, z5.s, z10.s\n" + "fmla z27.s, p3/M, z2.s, z10.s\n" + "fmla z29.s, p3/M, z4.s, z11.s\n" + "ld1w { z10.s }, p2/Z, [x6, x15, LSL #2]\n" + "fmla z30.s, p3/M, z3.s, z11.s\n" + "fmla z26.s, p3/M, z0.s, z11.s\n" + "fmla z19.s, p3/M, z8.s, z12.s\n" + "fmla z23.s, p3/M, z5.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x12, x4, LSL #2]\n" + "fmla z24.s, p3/M, z2.s, z11.s\n" + "fmla z25.s, p3/M, z2.s, z10.s\n" + "fmla z28.s, p3/M, z5.s, z11.s\n" + "fmla z29.s, p3/M, z5.s, z10.s\n" + "ld1w { z9.s }, p2/Z, [x7, x4, LSL #2]\n" + "fmla z30.s, p3/M, z4.s, z10.s\n" + "fmla z31.s, p3/M, z3.s, z10.s\n" + "fmla z26.s, p3/M, z1.s, z10.s\n" + "fmla z27.s, p3/M, z0.s, z10.s\n" + "ld1w { z10.s }, p2/Z, [x7, x13, LSL #2]\n" + "fmla z20.s, p3/M, z7.s, z12.s\n" + "fmla z21.s, p3/M, z6.s, z12.s\n" "ld1w { z11.s }, p2/Z, [x12, x13, 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, [x5, x8, LSL #2]\n" - "fmla z30.s, p3/M, z8.s, z11.s\n" - "fmla z31.s, p3/M, z7.s, z11.s\n" + "fmla z24.s, p3/M, z4.s, z9.s\n" + "fmla z25.s, p3/M, z3.s, z9.s\n" + "fmla z16.s, p3/M, z1.s, z9.s\n" + "fmla z17.s, p3/M, z0.s, z9.s\n" + "fmla z28.s, p3/M, z7.s, z9.s\n" + "fmla z29.s, p3/M, z6.s, z9.s\n" + "ld1w { z12.s }, p2/Z, [x5, x8, LSL #2]\n" + "fmla z22.s, p3/M, z8.s, z11.s\n" + "fmla z23.s, p3/M, z7.s, z11.s\n" "ld1w { z11.s }, p2/Z, [x16, x4, 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, [x5, x15, LSL #2]\n" + "fmla z30.s, p3/M, z8.s, z10.s\n" + "fmla z31.s, p3/M, z7.s, z10.s\n" + "fmla z26.s, p3/M, z5.s, z10.s\n" + "fmla z27.s, p3/M, z4.s, z10.s\n" + "fmla z18.s, p3/M, z2.s, z10.s\n" + "fmla z19.s, p3/M, z1.s, z10.s\n" + "ld1w { z9.s }, p2/Z, [x5, x15, LSL #2]\n" "addvl x5, x5, #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" + "fmla z24.s, p3/M, z7.s, z11.s\n" + "fmla z25.s, p3/M, z6.s, z11.s\n" + "fmla z16.s, p3/M, z4.s, z11.s\n" + "fmla z17.s, p3/M, z3.s, z11.s\n" + "fmla z20.s, p3/M, z1.s, z11.s\n" + "fmla z21.s, p3/M, z0.s, z11.s\n" "ld1w { z11.s }, p2/Z, [x16, x13, 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" + "fmla z28.s, p3/M, z2.s, z12.s\n" + "fmla z29.s, p3/M, z1.s, z12.s\n" + "fmla z30.s, p3/M, z0.s, z12.s\n" "ld1w { z10.s }, p2/Z, [x7]\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" + "fmla z22.s, p3/M, z2.s, z11.s\n" + "fmla z31.s, p3/M, z0.s, z9.s\n" + "fmla z24.s, p3/M, z3.s, z10.s\n" + "fmla z16.s, p3/M, z0.s, z10.s\n" + "fmla z26.s, p3/M, z8.s, z11.s\n" + "fmla z27.s, p3/M, z7.s, z11.s\n" + "fmla z18.s, p3/M, z5.s, z11.s\n" + "fmla z19.s, p3/M, z4.s, z11.s\n" + "fmla z23.s, p3/M, z1.s, z11.s\n" "ld1w { z11.s }, p2/Z, [x14, x8, LSL #2]\n" - "fmla z17.s, p3/M, z2.s, z12.s\n" - "fmla z18.s, p3/M, z1.s, z12.s\n" + "fmla z29.s, p3/M, z2.s, z9.s\n" + "fmla z30.s, p3/M, z1.s, z9.s\n" "ld1w { z12.s }, p2/Z, [x7, x11, LSL #2]\n" "addvl x7, x7, #1\n" - "fmla z16.s, p3/M, z6.s, z10.s\n" + "fmla z28.s, p3/M, z6.s, z10.s\n" "ld1w { z10.s }, p2/Z, [x16]\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" + "fmla z21.s, p3/M, z4.s, z11.s\n" + "fmla z22.s, p3/M, z3.s, z11.s\n" + "fmla z31.s, p3/M, z8.s, z12.s\n" + "ld1w { z9.s }, p1/Z, [x7, x8, LSL #2]\n" + "fmla z27.s, p3/M, z5.s, z12.s\n" + "fmla z19.s, p3/M, z2.s, z12.s\n" "ld1w { z12.s }, p2/Z, [x16, x11, LSL #2]\n" "addvl x16, x16, #1\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" + "fmla z24.s, p3/M, z6.s, z10.s\n" + "fmla z16.s, p3/M, z3.s, z10.s\n" + "fmla z20.s, p3/M, z0.s, z10.s\n" "ld1w { z10.s }, p2/Z, [x12, x8, 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" + "fmla z23.s, p3/M, z2.s, z12.s\n" + "fmla z21.s, p3/M, z7.s, z10.s\n" + "fmla z22.s, p3/M, z6.s, z10.s\n" + "fmla z16.s, p3/M, z8.s, z11.s\n" + "fmla z17.s, p3/M, z7.s, z11.s\n" + "fmla z18.s, p3/M, z6.s, z11.s\n" + "fmla z20.s, p3/M, z5.s, z11.s\n" "ld1w { z11.s }, p2/Z, [x14, x15, 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" + "fmla z19.s, p3/M, z5.s, z12.s\n" + "fmla z21.s, p3/M, z5.s, z11.s\n" + "fmla z22.s, p3/M, z4.s, z11.s\n" + "fmla z23.s, p3/M, z3.s, z11.s\n" + "fmla z27.s, p3/M, z8.s, z12.s\n" "ld1w { z12.s }, p2/Z, [x12, x15, LSL #2]\n" - "fmla z28.s, p3/M, z8.s, z10.s\n" + "fmla z20.s, p3/M, z8.s, z10.s\n" "addvl x12, x12, #1\n" "ld1w { z10.s }, p2/Z, [x6, x4, 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" + "fmla z17.s, p3/M, z8.s, z11.s\n" + "fmla z18.s, p3/M, z7.s, z11.s\n" + "fmla z19.s, p3/M, z6.s, z11.s\n" + "fmla z21.s, p3/M, z8.s, z12.s\n" "ld1w { z11.s }, p2/Z, [x6, x13, LSL #2]\n" "addvl x6, x6, #1\n" - "fmla z30.s, p3/M, z7.s, z12.s\n" - "fmla z31.s, p3/M, z6.s, z12.s\n" + "fmla z22.s, p3/M, z7.s, z12.s\n" + "fmla z23.s, p3/M, z6.s, z12.s\n" "ld1w { z12.s }, p2/Z, [x14, x4, LSL #2]\n" - "fmla z16.s, p3/M, z4.s, z10.s\n" - "fmla z17.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" + "fmla z28.s, p3/M, z4.s, z10.s\n" + "fmla z29.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, [x14, x13, LSL #2]\n" "whilelt p2.s, x21, %x[n_channels]\n" - "fmla z18.s, p3/M, z5.s, z11.s\n" - "fmla z19.s, p3/M, z4.s, z11.s\n" - "ld1w { z15.s }, p3/Z, [x17]\n" - "addvl x17, x17, #1\n" - "fmla z22.s, p3/M, z2.s, z11.s\n" - "fmla z23.s, p3/M, z1.s, z11.s\n" + "fmla z30.s, p3/M, z5.s, z11.s\n" + "fmla z31.s, p3/M, z4.s, z11.s\n" "cmp x26, %x[n_channels]\n" "addvl x14, x14, #1\n" - "fmla z24.s, p3/M, z7.s, z12.s\n" - "fmla z25.s, p3/M, z6.s, z12.s\n" - "ld1w { z9.s }, p1/Z, [x7, x8, LSL #2]\n" - "fmla z28.s, p3/M, z4.s, z12.s\n" - "fmla z29.s, p3/M, z3.s, z12.s\n" + "fmla z26.s, p3/M, z2.s, z11.s\n" + "fmla z27.s, p3/M, z1.s, z11.s\n" + "ld1w { z11.s }, p1/Z, [x5, x11, LSL #2]\n" + "fmla z16.s, p3/M, z7.s, z12.s\n" + "fmla z17.s, p3/M, z6.s, z12.s\n" + "fmla z20.s, p3/M, z4.s, z12.s\n" + "fmla z21.s, p3/M, z3.s, z12.s\n" ".inst 0xa040c220 // ld1w { z0.s-z3.s }, pn8.b/Z, [x17]\n" "addvl x17, x17, #4\n" - "fmla z26.s, p3/M, z8.s, z10.s\n" - "fmla z27.s, p3/M, z7.s, z10.s\n" - "ld1w { z11.s }, p1/Z, [x5, x11, LSL #2]\n" - "fmla z30.s, p3/M, z5.s, z10.s\n" - "fmla z31.s, p3/M, z4.s, z10.s\n" + "fmla z18.s, p3/M, z8.s, z10.s\n" + "fmla z19.s, p3/M, z7.s, z10.s\n" + "ld1w { z12.s }, p1/Z, [x7, x15, LSL #2]\n" + "fmla z22.s, p3/M, z5.s, z10.s\n" + "fmla z23.s, p3/M, z4.s, z10.s\n" ".inst 0xa040c224 // ld1w { z4.s-z7.s }, pn8.b/Z, [x17]\n" "addvl x17, x17, #4\n" - ".inst 0xc1adc9d0 // fclamp { z16.s-z19.s }, z14.s, z13.s\n" - ".inst 0xc1adc9d4 // fclamp { z20.s-z23.s }, z14.s, z13.s\n" + ".inst 0xc1afc9bc // fclamp { z28.s-z31.s }, z13.s, z15.s\n" + ".inst 0xc1afc9b8 // fclamp { z24.s-z27.s }, z13.s, z15.s\n" "ld1w { z10.s }, p1/Z, [x5]\n" - ".inst 0xc1adc9d8 // fclamp { z24.s-z27.s }, z14.s, z13.s\n" - ".inst 0xc1adc9dc // fclamp { z28.s-z31.s }, z14.s, z13.s\n" - "st1w { z16.s }, p0, [x28]\n" - "ld1w { z12.s }, p1/Z, [x7, x15, LSL #2]\n" - "st1w { z17.s }, p0, [x28, x9, LSL #2]\n" - "st1w { z18.s }, p0, [x28, x24, LSL #2]\n" + ".inst 0xc1afc9b0 // fclamp { z16.s-z19.s }, z13.s, z15.s\n" + ".inst 0xc1afc9b4 // fclamp { z20.s-z23.s }, z13.s, z15.s\n" + "st1w { z28.s }, p0, [x28]\n" + "st1w { z29.s }, p0, [x28, x9, LSL #2]\n" "ld1w { z8.s }, p3/Z, [x17]\n" "addvl x17, x17, #1\n" - "st1w { z19.s }, p0, [x28, x22, LSL #2]\n" + "st1w { z30.s }, p0, [x28, x24, LSL #2]\n" + "st1w { z31.s }, p0, [x28, x22, LSL #2]\n" "addvl x28, x28, #1\n" - "st1w { z20.s }, p0, [x27]\n" - "st1w { z21.s }, p0, [x27, x9, LSL #2]\n" - "st1w { z22.s }, p0, [x27, x24, LSL #2]\n" - "st1w { z23.s }, p0, [x27, x22, LSL #2]\n" + "st1w { z24.s }, p0, [x27]\n" + "st1w { z25.s }, p0, [x27, x9, LSL #2]\n" + "st1w { z26.s }, p0, [x27, x24, LSL #2]\n" + "st1w { z27.s }, p0, [x27, x22, LSL #2]\n" "addvl x27, x27, #1\n" - "st1w { z24.s }, p0, [x25]\n" - "st1w { z25.s }, p0, [x25, x9, LSL #2]\n" - "st1w { z26.s }, p0, [x25, x24, LSL #2]\n" - "st1w { z27.s }, p0, [x25, x22, LSL #2]\n" + "st1w { z16.s }, p0, [x25]\n" + "st1w { z17.s }, p0, [x25, x9, LSL #2]\n" + "st1w { z18.s }, p0, [x25, x24, LSL #2]\n" + "st1w { z19.s }, p0, [x25, x22, LSL #2]\n" "addvl x25, x25, #1\n" - "st1w { z28.s }, p0, [x23]\n" - "st1w { z29.s }, p0, [x23, x9, LSL #2]\n" - "st1w { z30.s }, p0, [x23, x24, LSL #2]\n" - "st1w { z31.s }, p0, [x23, x22, LSL #2]\n" + "st1w { z20.s }, p0, [x23]\n" + "st1w { z21.s }, p0, [x23, x9, LSL #2]\n" + "st1w { z22.s }, p0, [x23, x24, LSL #2]\n" + "st1w { z23.s }, p0, [x23, x22, LSL #2]\n" "addvl x23, x23, #1\n" "blt 3b\n" "4:" // 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 z21, z14\n fmla z21.s, p3/M, z4.s, z9.s\n" + "movprfx z24, z14\n fmla z24.s, p3/M, z8.s, z9.s\n" "ldr x3, [%x[params_struct], %[offsetof_args_tile_j]]\n" "add x3, x3, #0x1\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 z22, z14\n fmla z22.s, p3/M, z3.s, z9.s\n" + "movprfx z29, z14\n fmla z29.s, p3/M, z1.s, z9.s\n" "ldr x2, [%x[params_struct], %[offsetof_args_tile_i]]\n" "add x21, x2, #0x1\n" - "movprfx z26, z15\n fmla z26.s, p3/M, z0.s, z9.s\n" + "movprfx z30, z14\n fmla z30.s, p3/M, z0.s, z9.s\n" "fmla z21.s, p3/M, z5.s, z12.s\n" "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" "cmp x3, 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 z25, z14\n fmla z25.s, p3/M, z7.s, z9.s\n" + "movprfx z26, z14\n fmla z26.s, p3/M, z6.s, z9.s\n" "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" "csel x2, x2, 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 z20, z14\n fmla z20.s, p3/M, z5.s, z9.s\n" + "movprfx z28, z14\n fmla z28.s, p3/M, z2.s, z9.s\n" "ld1w { z9.s }, p2/Z, [x16, x8, 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, [x12]\n" + "fmla z24.s, p3/M, z0.s, z10.s\n" + "movprfx z27, z14\n fmla z27.s, p3/M, z2.s, z11.s\n" + "ld1w { z17.s }, p2/Z, [x12]\n" "csel x3, x3, XZR, LT\n" "fmla z22.s, p3/M, z4.s, z12.s\n" - "fmla z25.s, p3/M, z2.s, z12.s\n" - "ld1w { z11.s }, p2/Z, [x12, x11, LSL #2]\n" + "fmla z29.s, p3/M, z2.s, z12.s\n" + "ld1w { z18.s }, p2/Z, [x12, x11, LSL #2]\n" "cmp x2, 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, [x16, x15, LSL #2]\n" + "fmla z30.s, p3/M, z1.s, z12.s\n" + "movprfx z16, z14\n fmla z16.s, p3/M, z6.s, z17.s\n" + "ld1w { z11.s }, p2/Z, [x16, x15, 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, [x5, x4, LSL #2]\n" - "movprfx z31, z15\n fmla z31.s, p3/M, z8.s, z11.s\n" + "fmla z25.s, p3/M, z8.s, z12.s\n" + "fmla z26.s, p3/M, z7.s, z12.s\n" + "fmla z27.s, p3/M, z6.s, z12.s\n" + "movprfx z23, z14\n fmla z23.s, p3/M, z3.s, z12.s\n" + "movprfx z31, z14\n fmla z31.s, p3/M, z0.s, z12.s\n" + "ld1w { z10.s }, p2/Z, [x5, x4, LSL #2]\n" + "movprfx z19, z14\n fmla z19.s, p3/M, z8.s, z18.s\n" "fmla z22.s, p3/M, z6.s, z9.s\n" - "ld1w { z11.s }, p2/Z, [x5, x13, LSL #2]\n" - "fmla z25.s, p3/M, z4.s, z9.s\n" - "fmla z26.s, p3/M, z3.s, z9.s\n" + "ld1w { z12.s }, p2/Z, [x5, x13, LSL #2]\n" + "fmla z29.s, p3/M, z4.s, z9.s\n" + "fmla z30.s, p3/M, z3.s, z9.s\n" + "movprfx z17, z14\n fmla z17.s, p3/M, z1.s, z9.s\n" + "movprfx z18, z14\n fmla z18.s, p3/M, z0.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, [x6, x11, 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, [x14]\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" + "fmla z28.s, p3/M, z5.s, z9.s\n" + "fmla z16.s, p3/M, z2.s, z9.s\n" + "fmla z21.s, p3/M, z8.s, z11.s\n" + "ld1w { z14.s }, p2/Z, [x6]\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, [x6, x11, LSL #2]\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, [x14]\n" + "fmla z22.s, p3/M, z7.s, z11.s\n" + "fmla z23.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" + "fmla z17.s, p3/M, z2.s, z11.s\n" + "fmla z18.s, p3/M, z1.s, z11.s\n" + "fmla z19.s, p3/M, z0.s, z11.s\n" + "ld1w { z9.s }, p2/Z, [x6, x8, LSL #2]\n" + "fmla z20.s, p3/M, z0.s, z14.s\n" + "fmla z28.s, p3/M, z6.s, z12.s\n" + "fmla z16.s, p3/M, z3.s, z12.s\n" + "fmla z21.s, p3/M, z1.s, z9.s\n" "ld1w { z11.s }, p2/Z, [x14, x11, LSL #2]\n" - "fmla z19.s, p3/M, z5.s, z12.s\n" - "fmla z23.s, p3/M, z2.s, z12.s\n" + "fmla z24.s, p3/M, z3.s, z14.s\n" + "fmla z27.s, p3/M, z5.s, z10.s\n" + "fmla z23.s, p3/M, z2.s, z10.s\n" + "fmla z25.s, p3/M, z4.s, z9.s\n" "ld1w { z12.s }, p2/Z, [x6, x15, 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, [x6]\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, [x6, x8, 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, [x12, x4, 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 z26.s, p3/M, z3.s, z9.s\n" + "fmla z22.s, p3/M, z0.s, z9.s\n" + "fmla z31.s, p3/M, z8.s, z11.s\n" + "fmla z19.s, p3/M, z5.s, z11.s\n" + "ld1w { z10.s }, p2/Z, [x12, x4, LSL #2]\n" + "fmla z20.s, p3/M, z2.s, z9.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, [x7, x4, LSL #2]\n" - "fmla z18.s, p3/M, z4.s, z12.s\n" - "fmla z19.s, p3/M, z3.s, z12.s\n" + "fmla z24.s, p3/M, z5.s, z9.s\n" + "fmla z25.s, p3/M, z5.s, z12.s\n" + "ld1w { z9.s }, p2/Z, [x7, x4, LSL #2]\n" + "fmla z26.s, p3/M, z4.s, z12.s\n" + "fmla z27.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, [x7, x13, 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, [x12, x13, 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 { z11.s }, p2/Z, [x12, x13, LSL #2]\n" + "fmla z20.s, p3/M, z4.s, z9.s\n" + "fmla z21.s, p3/M, z3.s, z9.s\n" + "fmla z28.s, p3/M, z1.s, z9.s\n" + "fmla z29.s, p3/M, z0.s, z9.s\n" + "fmla z24.s, p3/M, z7.s, z9.s\n" + "fmla z25.s, p3/M, z6.s, z9.s\n" "ld1w { z10.s }, p2/Z, [x5, x8, 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, [x16, x4, LSL #2]\n" - "fmla z18.s, p3/M, z8.s, z12.s\n" - "fmla z19.s, p3/M, z7.s, z12.s\n" + "fmla z18.s, p3/M, z8.s, z11.s\n" + "fmla z19.s, p3/M, z7.s, z11.s\n" + "ld1w { z14.s }, p2/Z, [x16, x4, LSL #2]\n" + "fmla z26.s, p3/M, z8.s, z12.s\n" + "fmla z27.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, [x5, x15, 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, [x16, x13, 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" + "fmla z30.s, p3/M, z2.s, z12.s\n" + "fmla z31.s, p3/M, z1.s, z12.s\n" + "ld1w { z9.s }, p2/Z, [x5, x15, LSL #2]\n" + "fmla z20.s, p3/M, z7.s, z14.s\n" + "fmla z21.s, p3/M, z6.s, z14.s\n" + "fmla z28.s, p3/M, z4.s, z14.s\n" + "fmla z29.s, p3/M, z3.s, z14.s\n" + "fmla z16.s, p3/M, z1.s, z14.s\n" + "fmla z17.s, p3/M, z0.s, z14.s\n" + "ld1w { z14.s }, p2/Z, [x16, x13, LSL #2]\n" + "fmla z24.s, p3/M, z2.s, z10.s\n" + "fmla z25.s, p3/M, z1.s, z10.s\n" + "fmla z26.s, p3/M, z0.s, z10.s\n" "ld1w { z10.s }, p2/Z, [x7]\n" - "fmla z30.s, p3/M, z2.s, z11.s\n" - "fmla z19.s, p3/M, z0.s, z12.s\n" + "fmla z18.s, p3/M, z2.s, z14.s\n" + "fmla z27.s, p3/M, z0.s, z9.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" + "fmla z28.s, p3/M, z0.s, z10.s\n" + "fmla z22.s, p3/M, z8.s, z14.s\n" + "fmla z23.s, p3/M, z7.s, z14.s\n" + "fmla z30.s, p3/M, z5.s, z14.s\n" + "fmla z31.s, p3/M, z4.s, z14.s\n" + "fmla z19.s, p3/M, z1.s, z14.s\n" "ld1w { z11.s }, p2/Z, [x14, x8, LSL #2]\n" - "fmla z17.s, p3/M, z2.s, z12.s\n" - "fmla z18.s, p3/M, z1.s, z12.s\n" + "fmla z25.s, p3/M, z2.s, z9.s\n" + "fmla z26.s, p3/M, z1.s, z9.s\n" "ld1w { z12.s }, p2/Z, [x7, x11, LSL #2]\n" - "fmla z16.s, p3/M, z6.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x16]\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 z24.s, p3/M, z6.s, z10.s\n" + "ld1w { z14.s }, p2/Z, [x16]\n" + "fmla z17.s, p3/M, z4.s, z11.s\n" + "fmla z18.s, p3/M, z3.s, z11.s\n" + "fmla z27.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, [x16, x11, 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, [x12, x8, 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, [x14, x15, 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, [x12, x15, LSL #2]\n" - "fmla z28.s, p3/M, z8.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x6, x4, 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 { z9.s }, p2/Z, [x16, x11, LSL #2]\n" + "fmla z20.s, p3/M, z6.s, z14.s\n" + "fmla z28.s, p3/M, z3.s, z14.s\n" + "fmla z16.s, p3/M, z0.s, z14.s\n" + "ld1w { z12.s }, p2/Z, [x12, x8, LSL #2]\n" + "fmla z19.s, p3/M, z2.s, z9.s\n" + "fmla z17.s, p3/M, z7.s, z12.s\n" + "fmla z18.s, p3/M, z6.s, z12.s\n" + "fmla z28.s, p3/M, z8.s, z11.s\n" + "fmla z29.s, p3/M, z7.s, z11.s\n" + "fmla z30.s, p3/M, z6.s, z11.s\n" + "fmla z16.s, p3/M, z5.s, z11.s\n" + "ld1w { z10.s }, p2/Z, [x14, x15, LSL #2]\n" + "fmla z31.s, p3/M, z5.s, z9.s\n" + "fmla z17.s, p3/M, z5.s, z10.s\n" + "fmla z18.s, p3/M, z4.s, z10.s\n" + "fmla z19.s, p3/M, z3.s, z10.s\n" + "fmla z23.s, p3/M, z8.s, z9.s\n" + "ld1w { z14.s }, p2/Z, [x12, x15, LSL #2]\n" + "fmla z16.s, p3/M, z8.s, z12.s\n" + "ld1w { z9.s }, p2/Z, [x6, x4, LSL #2]\n" + "fmla z29.s, p3/M, z8.s, z10.s\n" + "fmla z30.s, p3/M, z7.s, z10.s\n" + "fmla z31.s, p3/M, z6.s, z10.s\n" + "fmla z17.s, p3/M, z8.s, z14.s\n" "ld1w { z11.s }, p2/Z, [x6, x13, 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, [x14, x4, LSL #2]\n" - "fmla z16.s, p3/M, z4.s, z10.s\n" - "fmla z17.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, [x14, x13, LSL #2]\n" - "fmla z18.s, p3/M, z5.s, z11.s\n" - "fmla z19.s, p3/M, z4.s, z11.s\n" + "fmla z18.s, p3/M, z7.s, z14.s\n" + "fmla z19.s, p3/M, z6.s, z14.s\n" + "ld1w { z10.s }, p2/Z, [x14, x4, LSL #2]\n" + "fmla z24.s, p3/M, z4.s, z9.s\n" + "fmla z25.s, p3/M, z3.s, z9.s\n" + "fmla z20.s, p3/M, z1.s, z9.s\n" + "fmla z21.s, p3/M, z0.s, z9.s\n" + "ld1w { z12.s }, p2/Z, [x14, x13, LSL #2]\n" + "fmla z26.s, p3/M, z5.s, z11.s\n" + "fmla z27.s, p3/M, z4.s, z11.s\n" "fmla z22.s, p3/M, z2.s, z11.s\n" "fmla z23.s, p3/M, z1.s, z11.s\n" - "fmla z24.s, p3/M, z7.s, z12.s\n" - "fmla z25.s, p3/M, z6.s, z12.s\n" - "fmla z28.s, p3/M, z4.s, z12.s\n" - "fmla z29.s, p3/M, z3.s, z12.s\n" - "fmla z26.s, p3/M, z8.s, z10.s\n" - "fmla z27.s, p3/M, z7.s, z10.s\n" - "fmla z30.s, p3/M, z5.s, z10.s\n" - "fmla z31.s, p3/M, z4.s, z10.s\n" - ".inst 0xc1adc9d0 // fclamp { z16.s-z19.s }, z14.s, z13.s\n" - ".inst 0xc1adc9d4 // fclamp { z20.s-z23.s }, z14.s, z13.s\n" - "st1w { z16.s }, p0, [x28]\n" - ".inst 0xc1adc9d8 // fclamp { z24.s-z27.s }, z14.s, z13.s\n" - ".inst 0xc1adc9dc // fclamp { z28.s-z31.s }, z14.s, z13.s\n" - "st1w { z17.s }, p0, [x28, x9, LSL #2]\n" - "st1w { z18.s }, p0, [x28, x24, LSL #2]\n" - "st1w { z19.s }, p0, [x28, x22, LSL #2]\n" + "fmla z28.s, p3/M, z7.s, z10.s\n" + "fmla z29.s, p3/M, z6.s, z10.s\n" + "fmla z16.s, p3/M, z4.s, z10.s\n" + "fmla z17.s, p3/M, z3.s, z10.s\n" + "fmla z30.s, p3/M, z8.s, z12.s\n" + "fmla z31.s, p3/M, z7.s, z12.s\n" + "fmla z18.s, p3/M, z5.s, z12.s\n" + "fmla z19.s, p3/M, z4.s, z12.s\n" + ".inst 0xc1afc9b8 // fclamp { z24.s-z27.s }, z13.s, z15.s\n" + ".inst 0xc1afc9b4 // fclamp { z20.s-z23.s }, z13.s, z15.s\n" + "st1w { z24.s }, p0, [x28]\n" + ".inst 0xc1afc9bc // fclamp { z28.s-z31.s }, z13.s, z15.s\n" + ".inst 0xc1afc9b0 // fclamp { z16.s-z19.s }, z13.s, z15.s\n" + "st1w { z25.s }, p0, [x28, x9, LSL #2]\n" + "st1w { z26.s }, p0, [x28, x24, LSL #2]\n" + "st1w { z27.s }, p0, [x28, x22, LSL #2]\n" "st1w { z20.s }, p0, [x27]\n" "st1w { z21.s }, p0, [x27, x9, LSL #2]\n" "st1w { z22.s }, p0, [x27, x24, LSL #2]\n" "st1w { z23.s }, p0, [x27, x22, LSL #2]\n" - "st1w { z24.s }, p0, [x25]\n" - "st1w { z25.s }, p0, [x25, x9, LSL #2]\n" - "st1w { z26.s }, p0, [x25, x24, LSL #2]\n" - "st1w { z27.s }, p0, [x25, x22, LSL #2]\n" - "st1w { z28.s }, p0, [x23]\n" - "st1w { z29.s }, p0, [x23, x9, LSL #2]\n" - "st1w { z30.s }, p0, [x23, x24, LSL #2]\n" - "st1w { z31.s }, p0, [x23, x22, LSL #2]\n" + "st1w { z28.s }, p0, [x25]\n" + "st1w { z29.s }, p0, [x25, x9, LSL #2]\n" + "st1w { z30.s }, p0, [x25, x24, LSL #2]\n" + "st1w { z31.s }, p0, [x25, x22, LSL #2]\n" + "st1w { z16.s }, p0, [x23]\n" + "st1w { z17.s }, p0, [x23, x9, LSL #2]\n" + "st1w { z18.s }, p0, [x23, x24, LSL #2]\n" + "st1w { z19.s }, p0, [x23, x22, LSL #2]\n" "blt 1b\n" ".inst 0xd503467f // SMSTOP\n" : -- cgit v1.2.1