Revert "Update CPU kernels and add mixed sign GEMM support"

This reverts commit fc94f4d23abd4bc427b701f54ad85282e9ec7872 and 5d6fff041ade7eb44af0945867212f3979be3d3e (because the latter fixes a build failure caused by the former)

Change-Id: I7d07fea8307e9a7033b30874bbb14ba9202b23d8
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11815
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Adnan AlSinan <adnan.alsinan@arm.com>

1 files changed, 58 insertions, 58 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL.hpp
index e10f818ba2..0e138e4422 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL.hpp
@@ -34,57 +34,57 @@ void sve_transpose_interleave_3VL(uint16_t *out, const uint16_t *in, size_t widt
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "ptrue p2.b\n"
+      "ptrue p3.b\n"
       "blt 4f\n"
       "1:"  // Main row loop: Head
-      "mov x27, %x[in]\n"
-      "mov x26, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "mov x25, %x[width]\n"
-      "add x24, x27, %x[in_stride]\n"
+      "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "mov x21, x25\n"
-      "mov x20, x26\n"
-      "dech x25, ALL, MUL #3\n"
-      "add x26, x26, %x[out_stride]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z27.h }, p0/Z, [x27]\n"
-      "ld1h { z26.h }, p0/Z, [x24]\n"
-      "ld1h { z25.h }, p0/Z, [x23]\n"
-      "ld1h { z24.h }, p0/Z, [x22]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "cmp x25, #0x0\n"
-      "ld1h { z23.h }, p1/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z22.h }, p1/Z, [x24, #1, MUL VL]\n"
-      "ld1h { z21.h }, p1/Z, [x23, #1, MUL VL]\n"
-      "ld1h { z20.h }, p1/Z, [x22, #1, MUL VL]\n"
-      "ld1h { z19.h }, p0/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z18.h }, p0/Z, [x24, #2, MUL VL]\n"
-      "addvl x27, x27, #3\n"
+      "mov x20, x21\n"
+      "whilelt p2.h, XZR, x20\n"
+      "ld1h { z27.h }, p2/Z, [x26]\n"
+      "ld1h { z26.h }, p2/Z, [x25]\n"
+      "dech x20\n"
+      "whilelt p1.h, XZR, x20\n"
+      "ld1h { z25.h }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1h { z24.h }, p1/Z, [x25, #1, MUL VL]\n"
+      "dech x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z23.h }, p0/Z, [x26, #2, MUL VL]\n"
+      "ld1h { z22.h }, p0/Z, [x25, #2, MUL VL]\n"
+      "mov x20, x22\n"
+      "dech x21, ALL, MUL #3\n"
+      "ld1h { z21.h }, p2/Z, [x24]\n"
+      "ld1h { z20.h }, p1/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z19.h }, p0/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z18.h }, p2/Z, [x23]\n"
+      "cmp x21, #0x0\n"
+      "addvl x26, x26, #3\n"
+      "ld1h { z17.h }, p1/Z, [x23, #1, MUL VL]\n"
+      "ld1h { z16.h }, p0/Z, [x23, #2, MUL VL]\n"
+      "st1h { z27.h }, p3, [x20]\n"
+      "addvl x25, x25, #3\n"
+      "st1h { z25.h }, p3, [x20, #1, MUL VL]\n"
       "addvl x24, x24, #3\n"
-      "ld1h { z17.h }, p0/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x22, #2, MUL VL]\n"
-      "st1h { z27.h }, p2, [x20]\n"
       "addvl x23, x23, #3\n"
-      "st1h { z23.h }, p2, [x20, #1, MUL VL]\n"
-      "addvl x22, x22, #3\n"
-      "st1h { z19.h }, p2, [x20, #2, MUL VL]\n"
-      "st1h { z26.h }, p2, [x20, #3, MUL VL]\n"
-      "st1h { z22.h }, p2, [x20, #4, MUL VL]\n"
-      "st1h { z18.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z25.h }, p2, [x20, #6, MUL VL]\n"
-      "st1h { z21.h }, p2, [x20, #7, MUL VL]\n"
+      "st1h { z23.h }, p3, [x20, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z26.h }, p3, [x20, #3, MUL VL]\n"
+      "st1h { z24.h }, p3, [x20, #4, MUL VL]\n"
+      "st1h { z22.h }, p3, [x20, #5, MUL VL]\n"
+      "st1h { z21.h }, p3, [x20, #6, MUL VL]\n"
+      "st1h { z20.h }, p3, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #12\n"
-      "st1h { z17.h }, p2, [x20, #-4, MUL VL]\n"
-      "st1h { z24.h }, p2, [x20, #-3, MUL VL]\n"
-      "st1h { z20.h }, p2, [x20, #-2, MUL VL]\n"
-      "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n"
+      "st1h { z19.h }, p3, [x20, #-4, MUL VL]\n"
+      "st1h { z18.h }, p3, [x20, #-3, MUL VL]\n"
+      "st1h { z17.h }, p3, [x20, #-2, MUL VL]\n"
+      "st1h { z16.h }, p3, [x20, #-1, MUL VL]\n"
       "bgt 2b\n"
       "3:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x4\n"
@@ -93,28 +93,28 @@ void sve_transpose_interleave_3VL(uint16_t *out, const uint16_t *in, size_t widt
       "cbz %x[height], 8f\n"
       "4:"  // Main loop skip
       "5:"  // Tail row loop: Head
-      "mov x27, %x[in]\n"
-      "mov x26, %x[out]\n"
+      "mov x26, %x[in]\n"
+      "add %x[in], x26, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x1\n"
       "mov x21, %x[width]\n"
-      "add %x[in], x27, %x[in_stride]\n"
       "6:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "dech x21, ALL, MUL #3\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z18.h }, p0/Z, [x26]\n"
       "dech x20\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z17.h }, p0/Z, [x26, #1, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z18.h }, p0/Z, [x27]\n"
+      "dech x21, ALL, MUL #3\n"
       "whilelt p0.h, XZR, x20\n"
       "cmp x21, #0x0\n"
-      "ld1h { z17.h }, p1/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x27, #2, MUL VL]\n"
-      "addvl x27, x27, #3\n"
-      "st1h { z18.h }, p2, [x26]\n"
-      "st1h { z17.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z16.h }, p2, [x26, #2, MUL VL]\n"
-      "add x26, x26, %x[out_stride]\n"
+      "ld1h { z16.h }, p0/Z, [x26, #2, MUL VL]\n"
+      "st1h { z18.h }, p3, [x22]\n"
+      "addvl x26, x26, #3\n"
+      "st1h { z17.h }, p3, [x22, #1, MUL VL]\n"
+      "st1h { z16.h }, p3, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 6b\n"
       "7:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -123,7 +123,7 @@ void sve_transpose_interleave_3VL(uint16_t *out, const uint16_t *in, size_t widt
       "8:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27"
     );
 }