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.cpp | 62 +++++++++++----------- 1 file changed, 31 insertions(+), 31 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp index 39197aa04d..c5ea5adea0 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp @@ -88,8 +88,8 @@ void sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst_impl( "mov x20, #0x4\n" "ldr x4, [%x[args], %[offsetof_inptrs]]\n" "ldp x5, x6, [x21, #0x0]\n" - "whilelt p0.s, XZR, x20\n" - "whilelt p1.s, x3, x2\n" + "whilelt p2.s, XZR, x20\n" + "whilelt p0.s, x3, x2\n" "ldp x7, x8, [x21, #0x10]\n" "ldp x17, x16, [x4, #0x0]\n" "add x15, %x[args], %[offsetof_rescale]\n" @@ -101,25 +101,25 @@ void sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst_impl( "ldp x25, x24, [x4, #0x50]\n" "ldp x23, x22, [x4, #0x60]\n" "ldp x21, x20, [x4, #0x70]\n" - "ld1w { z7.s }, p1/Z, [x10, x3, LSL #2]\n" - "ld1w { z6.s }, p1/Z, [x9, x3, LSL #2]\n" - "ld1w { z5.s }, p1/Z, [x26, x3, LSL #2]\n" - "ld1w { z4.s }, p1/Z, [x25, x3, LSL #2]\n" - "ld1w { z3.s }, p1/Z, [x16, x3, LSL #2]\n" - "ld1w { z2.s }, p1/Z, [x13, x3, LSL #2]\n" - "ld1w { z1.s }, p1/Z, [x11, x3, LSL #2]\n" - "ld1w { z31.s }, p1/Z, [x27, x3, LSL #2]\n" - "ld1w { z30.s }, p1/Z, [x28, x3, LSL #2]\n" - "ld1w { z29.s }, p1/Z, [x24, x3, LSL #2]\n" - "ld1w { z28.s }, p1/Z, [x22, x3, LSL #2]\n" - "ld1w { z27.s }, p1/Z, [x21, x3, LSL #2]\n" - "ld1w { z26.s }, p1/Z, [x17, x3, LSL #2]\n" - "ld1w { z25.s }, p1/Z, [x12, x3, LSL #2]\n" - "ld1w { z24.s }, p1/Z, [x23, x3, LSL #2]\n" - "ld1w { z23.s }, p1/Z, [x20, x3, LSL #2]\n" + "ld1w { z7.s }, p0/Z, [x10, x3, LSL #2]\n" + "ld1w { z6.s }, p0/Z, [x9, x3, LSL #2]\n" + "ld1w { z5.s }, p0/Z, [x26, x3, LSL #2]\n" + "ld1w { z4.s }, p0/Z, [x25, x3, LSL #2]\n" + "ld1w { z3.s }, p0/Z, [x16, x3, LSL #2]\n" + "ld1w { z2.s }, p0/Z, [x13, x3, LSL #2]\n" + "ld1w { z1.s }, p0/Z, [x11, x3, LSL #2]\n" + "ld1w { z31.s }, p0/Z, [x27, x3, LSL #2]\n" + "ld1w { z30.s }, p0/Z, [x28, x3, LSL #2]\n" + "ld1w { z29.s }, p0/Z, [x24, x3, LSL #2]\n" + "ld1w { z28.s }, p0/Z, [x22, x3, LSL #2]\n" + "ld1w { z27.s }, p0/Z, [x21, x3, LSL #2]\n" + "ld1w { z26.s }, p0/Z, [x17, x3, LSL #2]\n" + "ld1w { z25.s }, p0/Z, [x12, x3, LSL #2]\n" + "ld1w { z24.s }, p0/Z, [x23, x3, LSL #2]\n" + "ld1w { z23.s }, p0/Z, [x20, x3, LSL #2]\n" "incw x3\n" "whilelt p1.s, x3, x2\n" - "ld1rqw { z0.s }, p0/Z, [x15]\n" + "ld1rqw { z0.s }, p2/Z, [x15]\n" "b.none 2f\n" "1:" // Vector: Loop "fadd z17.s, z7.s, z6.s\n" @@ -172,32 +172,32 @@ void sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst_impl( "fadd z17.s, z7.s, z6.s\n" "fadd z16.s, z5.s, z4.s\n" "whilelt p0.s, x14, x2\n" - "fadd z19.s, z17.s, z16.s\n" + "fadd z20.s, z17.s, z16.s\n" "fadd z18.s, z3.s, z2.s\n" "fadd z17.s, z1.s, z31.s\n" - "fadd z22.s, z30.s, z29.s\n" + "fadd z19.s, z30.s, z29.s\n" "fadd z16.s, z28.s, z27.s\n" - "fadd z21.s, z18.s, z19.s\n" - "fadd z20.s, z16.s, z19.s\n" - "fadd z19.s, z26.s, z17.s\n" - "fadd z18.s, z25.s, z22.s\n" + "fadd z21.s, z18.s, z20.s\n" + "fadd z20.s, z16.s, z20.s\n" + "fadd z16.s, z26.s, z17.s\n" + "fadd z18.s, z25.s, z19.s\n" "fadd z17.s, z24.s, z17.s\n" - "fadd z16.s, z23.s, z22.s\n" - "fadd z19.s, z21.s, z19.s\n" - "fmul z19.s, z19.s, z0.s[0]\n" - "st1w { z19.s }, p0, [x5, x14, LSL #2]\n" + "fadd z19.s, z23.s, z19.s\n" + "fadd z16.s, z21.s, z16.s\n" + "fmul z16.s, z16.s, z0.s[0]\n" + "st1w { z16.s }, p0, [x5, x14, LSL #2]\n" "fadd z18.s, z21.s, z18.s\n" "fadd z17.s, z17.s, z20.s\n" "fmul z18.s, z18.s, z0.s[1]\n" "fmul z17.s, z17.s, z0.s[2]\n" - "fadd z16.s, z16.s, z20.s\n" + "fadd z16.s, z19.s, z20.s\n" "fmul z16.s, z16.s, z0.s[3]\n" "st1w { z18.s }, p0, [x6, x14, LSL #2]\n" "st1w { z17.s }, p0, [x7, x14, LSL #2]\n" "st1w { z16.s }, p0, [x8, x14, LSL #2]\n" : : [args] "r" (&args), [offsetof_inptrs] "I" (offsetof(KernelArgs, inptrs)), [offsetof_n_channels] "I" (offsetof(KernelArgs, n_channels)), [offsetof_outptrs] "I" (offsetof(KernelArgs, outptrs)), [offsetof_rescale] "I" (offsetof(KernelArgs, rescale_vals)) - : "cc", "memory", "p0", "p1", "x2", "x3", "x4", "x5", "x6", "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", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "x2", "x3", "x4", "x5", "x6", "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", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } -- cgit v1.2.1