Update CPU kernel implementations and guard directives
Resolves COMPMID-6023
Change-Id: I868975d14c4f98af6716726feda22405a6a4c891
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_u8_u8_summing.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_u8_u8_summing.hpp
index 6c4a5fa..56d34a8 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_u8_u8_summing.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_u8_u8_summing.hpp
@@ -156,182 +156,182 @@
"cbz %x[width], 14f\n"
"tbz %x[width], #3, 9f\n"
"ldr d27, [x28], #0x8\n"
- "ldr d19, [x27], #0x8\n"
+ "ldr d26, [x27], #0x8\n"
"ldr d25, [x26], #0x8\n"
- "ldr d18, [x25], #0x8\n"
+ "ldr d24, [x25], #0x8\n"
"ldr d23, [x24], #0x8\n"
- "ldr d17, [x23], #0x8\n"
+ "ldr d22, [x23], #0x8\n"
"ldr d21, [x22], #0x8\n"
- "ldr d16, [x21], #0x8\n"
+ "ldr d20, [x21], #0x8\n"
"tbz %x[width], #2, 7f\n"
"ld1 { v27.s }[2], [x28], #0x4\n"
- "ld1 { v19.s }[2], [x27], #0x4\n"
+ "ld1 { v26.s }[2], [x27], #0x4\n"
"ld1 { v25.s }[2], [x26], #0x4\n"
- "ld1 { v18.s }[2], [x25], #0x4\n"
+ "ld1 { v24.s }[2], [x25], #0x4\n"
"ld1 { v23.s }[2], [x24], #0x4\n"
- "ld1 { v17.s }[2], [x23], #0x4\n"
+ "ld1 { v22.s }[2], [x23], #0x4\n"
"ld1 { v21.s }[2], [x22], #0x4\n"
- "ld1 { v16.s }[2], [x21], #0x4\n"
+ "ld1 { v20.s }[2], [x21], #0x4\n"
"tbz %x[width], #1, 6f\n"
"ld1 { v27.h }[6], [x28], #0x2\n"
- "ld1 { v19.h }[6], [x27], #0x2\n"
+ "ld1 { v26.h }[6], [x27], #0x2\n"
"mov x20, #0x2\n"
"ld1 { v25.h }[6], [x26], #0x2\n"
- "ld1 { v18.h }[6], [x25], #0x2\n"
+ "ld1 { v24.h }[6], [x25], #0x2\n"
"ld1 { v23.h }[6], [x24], #0x2\n"
- "ld1 { v17.h }[6], [x23], #0x2\n"
+ "ld1 { v22.h }[6], [x23], #0x2\n"
"ld1 { v21.h }[6], [x22], #0x2\n"
- "ld1 { v16.h }[6], [x21], #0x2\n"
+ "ld1 { v20.h }[6], [x21], #0x2\n"
"tbz %x[width], #0, 13f\n"
"ld1 { v27.b }[14], [x28]\n"
- "ld1 { v19.b }[14], [x27]\n"
+ "ld1 { v26.b }[14], [x27]\n"
"ld1 { v25.b }[14], [x26]\n"
- "ld1 { v18.b }[14], [x25]\n"
+ "ld1 { v24.b }[14], [x25]\n"
"ld1 { v23.b }[14], [x24]\n"
- "ld1 { v17.b }[14], [x23]\n"
+ "ld1 { v22.b }[14], [x23]\n"
"ld1 { v21.b }[14], [x22]\n"
- "ld1 { v16.b }[14], [x21]\n"
+ "ld1 { v20.b }[14], [x21]\n"
"b 13f\n"
"6:" // odd_loads_1_12
"mov x20, #0x2\n"
"tbz %x[width], #0, 13f\n"
"ld1 { v27.b }[12], [x28]\n"
- "ld1 { v19.b }[12], [x27]\n"
+ "ld1 { v26.b }[12], [x27]\n"
"ld1 { v25.b }[12], [x26]\n"
- "ld1 { v18.b }[12], [x25]\n"
+ "ld1 { v24.b }[12], [x25]\n"
"ld1 { v23.b }[12], [x24]\n"
- "ld1 { v17.b }[12], [x23]\n"
+ "ld1 { v22.b }[12], [x23]\n"
"ld1 { v21.b }[12], [x22]\n"
- "ld1 { v16.b }[12], [x21]\n"
+ "ld1 { v20.b }[12], [x21]\n"
"b 13f\n"
"7:" // odd_loads_2_8
"tbz %x[width], #1, 8f\n"
"ld1 { v27.h }[4], [x28], #0x2\n"
- "ld1 { v19.h }[4], [x27], #0x2\n"
+ "ld1 { v26.h }[4], [x27], #0x2\n"
"mov x20, #0x2\n"
"ld1 { v25.h }[4], [x26], #0x2\n"
- "ld1 { v18.h }[4], [x25], #0x2\n"
+ "ld1 { v24.h }[4], [x25], #0x2\n"
"ld1 { v23.h }[4], [x24], #0x2\n"
- "ld1 { v17.h }[4], [x23], #0x2\n"
+ "ld1 { v22.h }[4], [x23], #0x2\n"
"ld1 { v21.h }[4], [x22], #0x2\n"
- "ld1 { v16.h }[4], [x21], #0x2\n"
+ "ld1 { v20.h }[4], [x21], #0x2\n"
"tbz %x[width], #0, 13f\n"
"ld1 { v27.b }[10], [x28]\n"
- "ld1 { v19.b }[10], [x27]\n"
+ "ld1 { v26.b }[10], [x27]\n"
"ld1 { v25.b }[10], [x26]\n"
- "ld1 { v18.b }[10], [x25]\n"
+ "ld1 { v24.b }[10], [x25]\n"
"ld1 { v23.b }[10], [x24]\n"
- "ld1 { v17.b }[10], [x23]\n"
+ "ld1 { v22.b }[10], [x23]\n"
"ld1 { v21.b }[10], [x22]\n"
- "ld1 { v16.b }[10], [x21]\n"
+ "ld1 { v20.b }[10], [x21]\n"
"b 13f\n"
"8:" // odd_loads_1_8
"mov x20, #0x1\n"
"tbz %x[width], #0, 13f\n"
"ld1 { v27.b }[8], [x28]\n"
- "ld1 { v19.b }[8], [x27]\n"
+ "ld1 { v26.b }[8], [x27]\n"
"mov x20, #0x2\n"
"ld1 { v25.b }[8], [x26]\n"
- "ld1 { v18.b }[8], [x25]\n"
+ "ld1 { v24.b }[8], [x25]\n"
"ld1 { v23.b }[8], [x24]\n"
- "ld1 { v17.b }[8], [x23]\n"
+ "ld1 { v22.b }[8], [x23]\n"
"ld1 { v21.b }[8], [x22]\n"
- "ld1 { v16.b }[8], [x21]\n"
+ "ld1 { v20.b }[8], [x21]\n"
"b 13f\n"
"9:" // odd_loads_4_0
"tbz %x[width], #2, 11f\n"
"ldr s27, [x28], #0x4\n"
- "ldr s19, [x27], #0x4\n"
+ "ldr s26, [x27], #0x4\n"
"ldr s25, [x26], #0x4\n"
- "ldr s18, [x25], #0x4\n"
+ "ldr s24, [x25], #0x4\n"
"ldr s23, [x24], #0x4\n"
- "ldr s17, [x23], #0x4\n"
+ "ldr s22, [x23], #0x4\n"
"ldr s21, [x22], #0x4\n"
- "ldr s16, [x21], #0x4\n"
+ "ldr s20, [x21], #0x4\n"
"tbz %x[width], #1, 10f\n"
"ld1 { v27.h }[2], [x28], #0x2\n"
- "ld1 { v19.h }[2], [x27], #0x2\n"
+ "ld1 { v26.h }[2], [x27], #0x2\n"
"mov x20, #0x1\n"
"ld1 { v25.h }[2], [x26], #0x2\n"
- "ld1 { v18.h }[2], [x25], #0x2\n"
+ "ld1 { v24.h }[2], [x25], #0x2\n"
"ld1 { v23.h }[2], [x24], #0x2\n"
- "ld1 { v17.h }[2], [x23], #0x2\n"
+ "ld1 { v22.h }[2], [x23], #0x2\n"
"ld1 { v21.h }[2], [x22], #0x2\n"
- "ld1 { v16.h }[2], [x21], #0x2\n"
+ "ld1 { v20.h }[2], [x21], #0x2\n"
"tbz %x[width], #0, 13f\n"
"ld1 { v27.b }[6], [x28]\n"
- "ld1 { v19.b }[6], [x27]\n"
+ "ld1 { v26.b }[6], [x27]\n"
"ld1 { v25.b }[6], [x26]\n"
- "ld1 { v18.b }[6], [x25]\n"
+ "ld1 { v24.b }[6], [x25]\n"
"ld1 { v23.b }[6], [x24]\n"
- "ld1 { v17.b }[6], [x23]\n"
+ "ld1 { v22.b }[6], [x23]\n"
"ld1 { v21.b }[6], [x22]\n"
- "ld1 { v16.b }[6], [x21]\n"
+ "ld1 { v20.b }[6], [x21]\n"
"b 13f\n"
"10:" // odd_loads_1_4
"mov x20, #0x1\n"
"tbz %x[width], #0, 13f\n"
"ld1 { v27.b }[4], [x28]\n"
- "ld1 { v19.b }[4], [x27]\n"
+ "ld1 { v26.b }[4], [x27]\n"
"ld1 { v25.b }[4], [x26]\n"
- "ld1 { v18.b }[4], [x25]\n"
+ "ld1 { v24.b }[4], [x25]\n"
"ld1 { v23.b }[4], [x24]\n"
- "ld1 { v17.b }[4], [x23]\n"
+ "ld1 { v22.b }[4], [x23]\n"
"ld1 { v21.b }[4], [x22]\n"
- "ld1 { v16.b }[4], [x21]\n"
+ "ld1 { v20.b }[4], [x21]\n"
"b 13f\n"
"11:" // odd_loads_2_0
"tbz %x[width], #1, 12f\n"
"ldr h27, [x28], #0x2\n"
- "ldr h19, [x27], #0x2\n"
+ "ldr h26, [x27], #0x2\n"
"mov x20, #0x1\n"
"ldr h25, [x26], #0x2\n"
- "ldr h18, [x25], #0x2\n"
+ "ldr h24, [x25], #0x2\n"
"ldr h23, [x24], #0x2\n"
- "ldr h17, [x23], #0x2\n"
+ "ldr h22, [x23], #0x2\n"
"ldr h21, [x22], #0x2\n"
- "ldr h16, [x21], #0x2\n"
+ "ldr h20, [x21], #0x2\n"
"tbz %x[width], #0, 13f\n"
"ld1 { v27.b }[2], [x28]\n"
- "ld1 { v19.b }[2], [x27]\n"
+ "ld1 { v26.b }[2], [x27]\n"
"ld1 { v25.b }[2], [x26]\n"
- "ld1 { v18.b }[2], [x25]\n"
+ "ld1 { v24.b }[2], [x25]\n"
"ld1 { v23.b }[2], [x24]\n"
- "ld1 { v17.b }[2], [x23]\n"
+ "ld1 { v22.b }[2], [x23]\n"
"ld1 { v21.b }[2], [x22]\n"
- "ld1 { v16.b }[2], [x21]\n"
+ "ld1 { v20.b }[2], [x21]\n"
"b 13f\n"
"12:" // odd_loads_1_0
"ldr b27, [x28, #0x0]\n"
- "ldr b19, [x27, #0x0]\n"
+ "ldr b26, [x27, #0x0]\n"
"mov x20, #0x1\n"
"ldr b25, [x26, #0x0]\n"
- "ldr b18, [x25, #0x0]\n"
+ "ldr b24, [x25, #0x0]\n"
"ldr b23, [x24, #0x0]\n"
- "ldr b17, [x23, #0x0]\n"
+ "ldr b22, [x23, #0x0]\n"
"ldr b21, [x22, #0x0]\n"
- "ldr b16, [x21, #0x0]\n"
+ "ldr b20, [x21, #0x0]\n"
"13:" // Odd load end
- "zip1 v26.2d, v27.2d, v19.2d\n"
- "zip1 v24.2d, v25.2d, v18.2d\n"
+ "zip1 v19.2d, v27.2d, v26.2d\n"
+ "zip1 v18.2d, v25.2d, v24.2d\n"
"subs x20, x20, #0x1\n"
- "str q26, [%x[out_ptr], #0x0]\n"
- "zip1 v22.2d, v23.2d, v17.2d\n"
- "zip1 v20.2d, v21.2d, v16.2d\n"
- "str q24, [%x[out_ptr], #0x10]\n"
- "uadalp v5.8h, v26.16b\n"
- "uadalp v4.8h, v24.16b\n"
- "str q22, [%x[out_ptr], #0x20]\n"
- "uadalp v3.8h, v22.16b\n"
- "str q20, [%x[out_ptr], #0x30]\n"
- "uadalp v2.8h, v20.16b\n"
+ "str q19, [%x[out_ptr], #0x0]\n"
+ "zip1 v17.2d, v23.2d, v22.2d\n"
+ "zip1 v16.2d, v21.2d, v20.2d\n"
+ "str q18, [%x[out_ptr], #0x10]\n"
+ "uadalp v5.8h, v19.16b\n"
+ "uadalp v4.8h, v18.16b\n"
+ "str q17, [%x[out_ptr], #0x20]\n"
+ "uadalp v3.8h, v17.16b\n"
+ "str q16, [%x[out_ptr], #0x30]\n"
+ "uadalp v2.8h, v16.16b\n"
"add %x[out_ptr], %x[out_ptr], #0x40\n"
"beq 14f\n"
- "zip2 v19.2d, v27.2d, v19.2d\n"
- "zip2 v18.2d, v25.2d, v18.2d\n"
+ "zip2 v19.2d, v27.2d, v26.2d\n"
+ "zip2 v18.2d, v25.2d, v24.2d\n"
"str q19, [%x[out_ptr], #0x0]\n"
- "zip2 v17.2d, v23.2d, v17.2d\n"
- "zip2 v16.2d, v21.2d, v16.2d\n"
+ "zip2 v17.2d, v23.2d, v22.2d\n"
+ "zip2 v16.2d, v21.2d, v20.2d\n"
"str q18, [%x[out_ptr], #0x10]\n"
"uadalp v5.8h, v19.16b\n"
"uadalp v4.8h, v18.16b\n"
@@ -346,11 +346,11 @@
"uadalp v31.4s, v3.8h\n"
"uadalp v30.4s, v2.8h\n"
"addp v1.4s, v1.4s, v0.4s\n"
- "addp v0.4s, v31.4s, v30.4s\n"
+ "addp v16.4s, v31.4s, v30.4s\n"
"add v1.4s, v1.4s, v29.4s\n"
- "add v0.4s, v0.4s, v28.4s\n"
+ "add v16.4s, v16.4s, v28.4s\n"
"str q1, [%x[out_ptr], #0x0]\n"
- "str q0, [%x[out_ptr], #0x10]\n"
+ "str q16, [%x[out_ptr], #0x10]\n"
"add %x[out_ptr], %x[out_ptr], #0x20\n"
: [out_ptr] "+&r" (out_ptr), [width] "+&r" (width)
: [first] "r" (first), [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)