Revert "Update CPU kernels to remove x19"

This reverts commit 3c59f01c209d2732a15d97d65565ead964787a8b.

Resolves: COMPMID-5817
Change-Id: Ie2443a21854a95db1e3d0cafa2121c0187a5e237
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8974
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 58 insertions, 58 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp
index 0cf8044733..dfb6457ed9 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, 2023 Arm Limited.
+ * Copyright (c) 2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -51,82 +51,82 @@ size_t interleave_sve_s8q_3x3_dot::get_packed_size(const DepthwiseArgs &args)
 void interleave_sve_s8q_3x3_dot::pack_parameters(unsigned int n_channels, void *outptr, const int32_t *bias, const int8_t *weights, const arm_gemm::Requantize32 &qp, size_t ld_weight_col, size_t ld_weight_row)
 {
   __asm__ __volatile__(
+    "mov z30.b, #0x0\n"
+    "ptrue p2.b\n"
+    "ld1rw { z29.s }, p2/Z, [%x[qp], %[offsetof_input_offset]]\n"
+    "mov z28.b, #0x1\n"
     "cmp %x[ld_weight_col], XZR\n"
-    "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n"
     "mov z16.s, #0x9\n"
-    "mov z28.b, #0x0\n"
-    "mov x20, #0x3\n"
-    "ptrue p2.b\n"
-    "mul x20, %x[ld_weight_col], x20\n"
-    "ld1rw { z27.s }, p2/Z, [%x[qp], %[offsetof_input_offset]]\n"
-    "ld1rw { z26.s }, p2/Z, [%x[qp], %[offsetof_weights_offset]]\n"
+    "ld1rw { z27.s }, p2/Z, [%x[qp], %[offsetof_weights_offset]]\n"
+    "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n"
+    "mul z27.s, p2/M, z27.s, z29.s\n"
+    "ld1rw { z26.s }, p2/Z, [%x[qp], %[offsetof_per_layer_mul]]\n"
+    "mov x19, #0x3\n"
+    "mul z27.s, p2/M, z27.s, z16.s\n"
+    "ld1rw { z25.s }, p2/Z, [%x[qp], %[offsetof_per_layer_right_shift]]\n"
+    "mul x19, %x[ld_weight_col], x19\n"
     "cmp %x[ld_weight_row], XZR\n"
-    "csel %x[ld_weight_row], %x[ld_weight_row], x20, NE\n"
-    "mov z25.b, #0x1\n"
-    "mul z26.s, p2/M, z26.s, z27.s\n"
-    "add x24, %x[weights], %x[ld_weight_row]\n"
-    "ld1rw { z24.s }, p2/Z, [%x[qp], %[offsetof_per_layer_mul]]\n"
-    "ld1rw { z23.s }, p2/Z, [%x[qp], %[offsetof_per_layer_right_shift]]\n"
-    "add x23, x24, %x[ld_weight_row]\n"
-    "add x22, %x[ld_weight_col], %x[ld_weight_col]\n"
+    "add x23, %x[ld_weight_col], %x[ld_weight_col]\n"
+    "csel %x[ld_weight_row], %x[ld_weight_row], x19, NE\n"
+    "add x22, %x[weights], %x[ld_weight_row]\n"
+    "add x21, x22, %x[ld_weight_row]\n"
     "whilelt p1.s, XZR, %x[n_channels]\n"
-    "mov x21, #0x0\n"
-    "mul z26.s, p2/M, z26.s, z16.s\n"
+    "mov x20, #0x0\n"
     "pfalse p8.b\n"
     "cbz %x[bias], 1f\n"
     "ptrue p8.s\n"
     "1:"  // No bias
 
     "2:"  // Loop
-    "cntp x20, p2, p1.s\n"
-    "whilelt p0.b, XZR, x20\n"
-    "ld1b { z18.b }, p0/Z, [%x[weights]]\n"
-    "ld1b { z17.b }, p0/Z, [%x[weights], %x[ld_weight_col]]\n"
-    "ld1b { z16.b }, p0/Z, [%x[weights], x22]\n"
-    "zip1 z20.b, z18.b, z16.b\n"
-    "zip1 z19.b, z17.b, z28.b\n"
-    "ld1b { z18.b }, p0/Z, [x24]\n"
-    "ld1b { z17.b }, p0/Z, [x24, %x[ld_weight_col]]\n"
-    "ld1b { z16.b }, p0/Z, [x24, x22]\n"
-    "zip1 z22.b, z20.b, z19.b\n"
-    "zip1 z21.b, z18.b, z16.b\n"
-    "zip1 z19.b, z17.b, z28.b\n"
-    "mov z20.s, #0x0\n"
-    "ld1b { z18.b }, p0/Z, [x23]\n"
-    "ld1b { z17.b }, p0/Z, [x23, %x[ld_weight_col]]\n"
-    "ld1b { z16.b }, p0/Z, [x23, x22]\n"
-    "sdot z20.s, z25.b, z22.b\n"
-    "zip1 z19.b, z21.b, z19.b\n"
-    "sdot z20.s, z25.b, z19.b\n"
-    "zip1 z18.b, z18.b, z16.b\n"
-    "zip1 z16.b, z17.b, z28.b\n"
+    "mov z24.s, #0x0\n"
+    "cntp x19, p2, p1.s\n"
     "and p0.b, p2/Z, p8.b, p1.b\n"
-    "ld1w { z17.s }, p0/Z, [%x[bias], x21, LSL #2]\n"
-    "zip1 z16.b, z18.b, z16.b\n"
-    "sdot z20.s, z25.b, z16.b\n"
-    "mls z17.s, p2/M, z20.s, z27.s\n"
-    "add %x[weights], %x[weights], x20\n"
-    "add x24, x24, x20\n"
-    "add x23, x23, x20\n"
-    "add z17.s, z17.s, z26.s\n"
-    "st1w { z17.s }, p2, [%x[outptr]]\n"
-    "st1b { z22.b }, p2, [%x[outptr], #1, MUL VL]\n"
-    "st1b { z19.b }, p2, [%x[outptr], #2, MUL VL]\n"
+    "ld1w { z23.s }, p0/Z, [%x[bias], x20, LSL #2]\n"
+    "whilelt p0.b, XZR, x19\n"
+    "ld1b { z17.b }, p0/Z, [%x[weights]]\n"
+    "ld1b { z16.b }, p0/Z, [%x[weights], %x[ld_weight_col]]\n"
+    "zip1 z18.b, z16.b, z30.b\n"
+    "ld1b { z16.b }, p0/Z, [%x[weights], x23]\n"
+    "add %x[weights], %x[weights], x19\n"
+    "zip1 z16.b, z17.b, z16.b\n"
+    "ld1b { z22.b }, p0/Z, [x22]\n"
+    "ld1b { z17.b }, p0/Z, [x22, %x[ld_weight_col]]\n"
+    "zip1 z21.b, z16.b, z18.b\n"
+    "ld1b { z16.b }, p0/Z, [x22, x23]\n"
+    "sdot z24.s, z28.b, z21.b\n"
+    "add x22, x22, x19\n"
+    "zip1 z18.b, z17.b, z30.b\n"
+    "ld1b { z20.b }, p0/Z, [x21]\n"
+    "ld1b { z19.b }, p0/Z, [x21, %x[ld_weight_col]]\n"
+    "zip1 z17.b, z22.b, z16.b\n"
+    "ld1b { z16.b }, p0/Z, [x21, x23]\n"
+    "zip1 z18.b, z17.b, z18.b\n"
+    "add x21, x21, x19\n"
+    "zip1 z17.b, z19.b, z30.b\n"
+    "sdot z24.s, z28.b, z18.b\n"
+    "zip1 z16.b, z20.b, z16.b\n"
+    "zip1 z16.b, z16.b, z17.b\n"
+    "sdot z24.s, z28.b, z16.b\n"
+    "mls z23.s, p2/M, z24.s, z29.s\n"
+    "add z23.s, z23.s, z27.s\n"
+    "st1w { z23.s }, p2, [%x[outptr]]\n"
+    "st1b { z21.b }, p2, [%x[outptr], #1, MUL VL]\n"
+    "st1b { z18.b }, p2, [%x[outptr], #2, MUL VL]\n"
     "st1b { z16.b }, p2, [%x[outptr], #3, MUL VL]\n"
     "addvl %x[outptr], %x[outptr], #4\n"
     "cbz %x[rq_mul_perchannel], 3f\n"
-    "ld1w { z24.s }, p1/Z, [%x[rq_mul_perchannel], x21, LSL #2]\n"
-    "ld1w { z23.s }, p1/Z, [%x[rq_shift_perchannel], x21, LSL #2]\n"
+    "ld1w { z26.s }, p1/Z, [%x[rq_mul_perchannel], x20, LSL #2]\n"
+    "ld1w { z25.s }, p1/Z, [%x[rq_shift_perchannel], x20, LSL #2]\n"
     "3:"  // Loop: Quantisation parameters: Store
-    "incw x21\n"
-    "whilelt p1.s, x21, %x[n_channels]\n"
-    "st1w { z24.s }, p2, [%x[outptr]]\n"
-    "st1w { z23.s }, p2, [%x[outptr], #1, MUL VL]\n"
+    "st1w { z26.s }, p2, [%x[outptr]]\n"
+    "incw x20\n"
+    "st1w { z25.s }, p2, [%x[outptr], #1, MUL VL]\n"
+    "whilelt p1.s, x20, %x[n_channels]\n"
     "addvl %x[outptr], %x[outptr], #2\n"
     "b.any 2b\n"
     : [ld_weight_col] "+&r" (ld_weight_col), [ld_weight_row] "+&r" (ld_weight_row), [outptr] "+&r" (outptr), [weights] "+&r" (weights)
     : [bias] "r" (bias), [n_channels] "r" (n_channels), [offsetof_input_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_per_layer_mul] "I" (offsetof(arm_gemm::Requantize32, per_layer_mul)), [offsetof_per_layer_right_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_right_shift)), [offsetof_weights_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [qp] "r" (&qp), [rq_mul_perchannel] "r" (qp.per_channel_muls), [rq_shift_perchannel] "r" (qp.per_channel_right_shifts)
-    : "cc", "memory", "p0", "p1", "p2", "p8", "x20", "x21", "x22", "x23", "x24", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28"
+    : "cc", "memory", "p0", "p1", "p2", "p8", "x19", "x20", "x21", "x22", "x23", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30"
   );
 }