From e28cf395b57a091d0850cd28cecc81046153b843 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Sun, 31 Jan 2021 05:18:43 +0000 Subject: Regenerate kernels and update A55 versions Signed-off-by: Georgios Pinitas Change-Id: I9eae76c77db03b8806af65729da34ab2d77f95f2 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4965 Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- .../kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp | 1903 ++++++++++---------- 1 file changed, 914 insertions(+), 989 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp index 495637bcdd..f901a814f9 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp @@ -88,182 +88,176 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ptrue p2.b\n" "1:" // Row loop "cmp %x[M], #0x6\n" - "bge 71f\n" + "bge 66f\n" "cmp %x[M], #0x4\n" - "bgt 57f\n" - "beq 43f\n" + "bgt 53f\n" + "beq 40f\n" "cmp %x[M], #0x2\n" - "bgt 29f\n" - "beq 15f\n" - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 2f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "b 3f\n" - "2:" // Height 1: setup direct output - "mov x13, %x[output_ptr]\n" - "3:" // Height 1: Column loop + "bgt 27f\n" + "beq 14f\n" + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "2:" // Height 1: Column loop "mov z8.s, #0x0\n" "mov x19, #0x0\n" "mov z9.s, #0x0\n" - "whilelt p1.b, x19, x15\n" + "whilelt p1.b, x19, x10\n" "mov z10.s, #0x0\n" "mov z11.s, #0x0\n" - "4:" // Height 1: setup done - "mov x12, #0x0\n" - "5:" // Height 1: String loop + "3:" // Height 1: setup done + "mov x27, #0x0\n" + "4:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 6f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 5f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "cbnz x12, 7f\n" + "ldr x25, [x20, #0x0]\n" + "cbnz x27, 6f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "b 7f\n" - "6:" // Height 1: setup direct input - "mov x10, %x[input_ptr]\n" - "7:" // Height 1: input setup done - "cmp x11, #0x10\n" - "ble 9f\n" - "8:" // Height 1: Multiply loop: Main loop head - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "sub x11, x11, #0x10\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "add x25, x25, x19\n" + "b 6f\n" + "5:" // Height 1: setup direct input + "mov x25, %x[input_ptr]\n" + "6:" // Height 1: input setup done + "cmp x26, #0x10\n" + "ble 8f\n" + "7:" // Height 1: Multiply loop: Main loop head + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "sub x26, x26, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" - "add x10, x10, #0x10\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "cmp x11, #0x10\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "cmp x26, #0x10\n" "sdot z10.s, z6.b, z0.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #4, MUL VL]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "ld1b { z6.b }, p2/Z, [x28, #4, MUL VL]\n" + "prfm pldl1keep, [x25, #0x80]\n" "sdot z11.s, z7.b, z0.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" "sdot z8.s, z6.b, z0.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #7, MUL VL]\n" - "addvl x14, x14, #16\n" + "ld1b { z7.b }, p2/Z, [x28, #7, MUL VL]\n" + "addvl x28, x28, #16\n" "sdot z10.s, z6.b, z0.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #-8, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-8, MUL VL]\n" "sdot z11.s, z7.b, z0.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #-7, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-7, MUL VL]\n" "sdot z8.s, z6.b, z0.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-5, MUL VL]\n" "sdot z10.s, z6.b, z0.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-3, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-3, MUL VL]\n" "sdot z8.s, z6.b, z0.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #-2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #-1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-1, MUL VL]\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z11.s, z7.b, z0.b[3]\n" - "bgt 8b\n" - "9:" // Height 1: Multiply loop: Single iteration only - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "bgt 7b\n" + "8:" // Height 1: Multiply loop: Single iteration only + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" - "add x10, x10, #0x10\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z11.s, z7.b, z0.b[0]\n" - "ble 10f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 9f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z9.s, z7.b, z0.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "sdot z10.s, z6.b, z0.b[1]\n" - "addvl x14, x14, #4\n" + "addvl x28, x28, #4\n" "sdot z11.s, z7.b, z0.b[1]\n" - "ble 10f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 9f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z9.s, z7.b, z0.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "sdot z10.s, z6.b, z0.b[2]\n" - "addvl x14, x14, #4\n" + "addvl x28, x28, #4\n" "sdot z11.s, z7.b, z0.b[2]\n" - "ble 10f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 9f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z11.s, z7.b, z0.b[3]\n" - "10:" // Height 1: Multiply loop: multiply skip - "prfm pldl1keep, [x10, #0x80]\n" - "add x12, x12, #0x1\n" + "9:" // Height 1: Multiply loop: multiply skip + "prfm pldl1keep, [x25, #0x80]\n" + "add x27, x27, #0x1\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x12, x19\n" - "bne 5b\n" - "ld1w { z0.s }, p2/Z, [x16]\n" + "cmp x27, x19\n" + "bne 4b\n" + "ld1w { z0.s }, p2/Z, [x11]\n" "add z8.s, z8.s, z0.s\n" - "ld1w { z1.s }, p2/Z, [x16, #1, MUL VL]\n" - "ld1w { z2.s }, p2/Z, [x16, #2, MUL VL]\n" + "ld1w { z1.s }, p2/Z, [x11, #1, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x11, #2, MUL VL]\n" "add z9.s, z9.s, z1.s\n" - "ld1w { z3.s }, p2/Z, [x16, #3, MUL VL]\n" - "addvl x16, x16, #4\n" + "ld1w { z3.s }, p2/Z, [x11, #3, MUL VL]\n" + "addvl x11, x11, #4\n" "add z10.s, z10.s, z2.s\n" "add z11.s, z11.s, z3.s\n" - "tbz %x[flags], #4, 11f\n" - "ld1w { z0.s }, p2/Z, [x17]\n" - "ld1w { z4.s }, p2/Z, [x8]\n" - "ld1w { z1.s }, p2/Z, [x17, #1, MUL VL]\n" - "ld1w { z5.s }, p2/Z, [x8, #1, MUL VL]\n" - "ld1w { z2.s }, p2/Z, [x17, #2, MUL VL]\n" - "ld1w { z6.s }, p2/Z, [x8, #2, MUL VL]\n" - "ld1w { z3.s }, p2/Z, [x17, #3, MUL VL]\n" - "addvl x17, x17, #4\n" - "ld1w { z7.s }, p2/Z, [x8, #3, MUL VL]\n" - "addvl x8, x8, #4\n" - "b 12f\n" - "11:" // Height 1: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1rw { z0.s }, p2/Z, [x19]\n" + "tbz %x[flags], #4, 10f\n" + "ld1w { z0.s }, p2/Z, [x12]\n" + "ld1w { z4.s }, p2/Z, [x13]\n" + "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n" + "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n" + "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "b 11f\n" + "10:" // Height 1: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1rw { z0.s }, p2/Z, [x24]\n" "mov z1.d, z0.d\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" "mov z2.d, z0.d\n" "mov z3.d, z0.d\n" "mov z5.d, z4.d\n" "mov z6.d, z4.d\n" "mov z7.d, z4.d\n" - "12:" // Height 1: parameters loaded + "11:" // Height 1: parameters loaded ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n" ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n" ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n" ".inst 0x04a7756b // sqrdmulh z11.s, z11.s, z7.s\n" - "tbz %x[flags], #5, 13f\n" + "tbz %x[flags], #5, 12f\n" "and z4.d, z8.d, z0.d\n" "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" @@ -276,17 +270,17 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sqadd z9.s, z9.s, z5.s\n" "sqadd z10.s, z10.s, z6.s\n" "sqadd z11.s, z11.s, z7.s\n" - "13:" // Height 1: no shift correction + "12:" // Height 1: no shift correction ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n" - "ld1rw { z5.s }, p2/Z, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1rw { z5.s }, p2/Z, [x24]\n" + "add x24, %x[qp], %[maxval]\n" ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n" - "ld1rw { z6.s }, p2/Z, [x19]\n" + "ld1rw { z6.s }, p2/Z, [x24]\n" "add z8.s, z8.s, z4.s\n" "add z9.s, z9.s, z4.s\n" "add z10.s, z10.s, z4.s\n" @@ -302,230 +296,223 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z8.h, z8.h, z9.h\n" "uzp1 z9.h, z10.h, z11.h\n" "uzp1 z8.b, z8.b, z9.b\n" - "st1b { z8.b }, p1, [x13]\n" - "addvl x13, x13, #1\n" - "14:" // Height 1: Writeback done - "decw x15, ALL, MUL #4\n" - "cmp x15, XZR\n" - "bgt 3b\n" - "b 86f\n" - "15:" // Height 2 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 16f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "add x9, x9, x19\n" - "b 17f\n" - "16:" // Height 2: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "17:" // Height 2: Column loop + "st1b { z8.b }, p1, [x9]\n" + "addvl x9, x9, #1\n" + "13:" // Height 1: Writeback done + "decw x10, ALL, MUL #4\n" + "cmp x10, XZR\n" + "bgt 2b\n" + "b 80f\n" + "14:" // Height 2 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "15:" // Height 2: Column loop "mov z8.s, #0x0\n" "mov x19, #0x0\n" "mov z9.s, #0x0\n" - "whilelt p1.b, x19, x15\n" + "whilelt p1.b, x19, x10\n" "mov z10.s, #0x0\n" "mov z11.s, #0x0\n" "mov z12.s, #0x0\n" "mov z13.s, #0x0\n" "mov z14.s, #0x0\n" "mov z15.s, #0x0\n" - "18:" // Height 2: setup done - "mov x12, #0x0\n" - "19:" // Height 2: String loop + "16:" // Height 2: setup done + "mov x27, #0x0\n" + "17:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 20f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 18f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "cbnz x12, 21f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "cbnz x27, 19f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "b 21f\n" - "20:" // Height 2: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "21:" // Height 2: input setup done - "cmp x11, #0x10\n" - "ble 23f\n" - "22:" // Height 2: Multiply loop: Main loop head - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "sub x11, x11, #0x10\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "add x25, x25, x19\n" + "add x24, x24, x19\n" + "b 19f\n" + "18:" // Height 2: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "19:" // Height 2: input setup done + "cmp x26, #0x10\n" + "ble 21f\n" + "20:" // Height 2: Multiply loop: Main loop head + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "sub x26, x26, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" - "cmp x11, #0x10\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" + "cmp x26, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "prfm pldl1keep, [x24, #0x80]\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[0]\n" "sdot z15.s, z7.b, z1.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" "sdot z8.s, z6.b, z0.b[1]\n" "sdot z12.s, z6.b, z1.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #7, MUL VL]\n" - "addvl x14, x14, #16\n" + "ld1b { z7.b }, p2/Z, [x28, #7, MUL VL]\n" + "addvl x28, x28, #16\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #-8, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-8, MUL VL]\n" "sdot z11.s, z7.b, z0.b[1]\n" "sdot z15.s, z7.b, z1.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #-7, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-7, MUL VL]\n" "sdot z8.s, z6.b, z0.b[2]\n" "sdot z12.s, z6.b, z1.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-5, MUL VL]\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[2]\n" "sdot z15.s, z7.b, z1.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-3, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-3, MUL VL]\n" "sdot z8.s, z6.b, z0.b[3]\n" "sdot z12.s, z6.b, z1.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #-2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #-1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-1, MUL VL]\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z11.s, z7.b, z0.b[3]\n" "sdot z15.s, z7.b, z1.b[3]\n" - "bgt 22b\n" - "23:" // Height 2: Multiply loop: Single iteration only - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "bgt 20b\n" + "21:" // Height 2: Multiply loop: Single iteration only + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "add x28, x28, #0x10\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z13.s, z7.b, z1.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z11.s, z7.b, z0.b[0]\n" "sdot z15.s, z7.b, z1.b[0]\n" - "ble 24f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 22f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" "sdot z11.s, z7.b, z0.b[1]\n" "sdot z15.s, z7.b, z1.b[1]\n" - "ble 24f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 22f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" "sdot z11.s, z7.b, z0.b[2]\n" "sdot z15.s, z7.b, z1.b[2]\n" - "ble 24f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 22f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" "sdot z12.s, z6.b, z1.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z11.s, z7.b, z0.b[3]\n" "sdot z15.s, z7.b, z1.b[3]\n" - "24:" // Height 2: Multiply loop: multiply skip - "prfm pldl1keep, [x10, #0x80]\n" - "add x12, x12, #0x1\n" - "prfm pldl1keep, [x28, #0x80]\n" + "22:" // Height 2: Multiply loop: multiply skip + "prfm pldl1keep, [x25, #0x80]\n" + "add x27, x27, #0x1\n" + "prfm pldl1keep, [x24, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x12, x19\n" - "bne 19b\n" - "ld1w { z0.s }, p2/Z, [x16]\n" + "cmp x27, x19\n" + "bne 17b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z0.s }, p2/Z, [x11]\n" "add z8.s, z8.s, z0.s\n" - "ld1w { z1.s }, p2/Z, [x16, #1, MUL VL]\n" + "ld1w { z1.s }, p2/Z, [x11, #1, MUL VL]\n" + "add x23, x9, x19\n" "add z12.s, z12.s, z0.s\n" - "ld1w { z2.s }, p2/Z, [x16, #2, MUL VL]\n" - "ld1w { z3.s }, p2/Z, [x16, #3, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x11, #2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x11, #3, MUL VL]\n" "add z9.s, z9.s, z1.s\n" - "addvl x16, x16, #4\n" + "addvl x11, x11, #4\n" "add z13.s, z13.s, z1.s\n" "add z10.s, z10.s, z2.s\n" "add z11.s, z11.s, z3.s\n" "add z14.s, z14.s, z2.s\n" "add z15.s, z15.s, z3.s\n" - "tbz %x[flags], #4, 25f\n" - "ld1w { z0.s }, p2/Z, [x17]\n" - "ld1w { z4.s }, p2/Z, [x8]\n" - "ld1w { z1.s }, p2/Z, [x17, #1, MUL VL]\n" - "ld1w { z5.s }, p2/Z, [x8, #1, MUL VL]\n" - "ld1w { z2.s }, p2/Z, [x17, #2, MUL VL]\n" - "ld1w { z6.s }, p2/Z, [x8, #2, MUL VL]\n" - "ld1w { z3.s }, p2/Z, [x17, #3, MUL VL]\n" - "addvl x17, x17, #4\n" - "ld1w { z7.s }, p2/Z, [x8, #3, MUL VL]\n" - "addvl x8, x8, #4\n" - "b 26f\n" - "25:" // Height 2: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1rw { z0.s }, p2/Z, [x19]\n" + "tbz %x[flags], #4, 23f\n" + "ld1w { z0.s }, p2/Z, [x12]\n" + "ld1w { z4.s }, p2/Z, [x13]\n" + "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n" + "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n" + "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "b 24f\n" + "23:" // Height 2: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1rw { z0.s }, p2/Z, [x24]\n" "mov z1.d, z0.d\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" "mov z2.d, z0.d\n" "mov z3.d, z0.d\n" "mov z5.d, z4.d\n" "mov z6.d, z4.d\n" "mov z7.d, z4.d\n" - "26:" // Height 2: parameters loaded + "24:" // Height 2: parameters loaded ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n" ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n" ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n" @@ -534,7 +521,7 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a575ad // sqrdmulh z13.s, z13.s, z5.s\n" ".inst 0x04a675ce // sqrdmulh z14.s, z14.s, z6.s\n" ".inst 0x04a775ef // sqrdmulh z15.s, z15.s, z7.s\n" - "tbz %x[flags], #5, 27f\n" + "tbz %x[flags], #5, 25f\n" "and z4.d, z8.d, z0.d\n" "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" @@ -559,17 +546,17 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sqadd z13.s, z13.s, z5.s\n" "sqadd z14.s, z14.s, z6.s\n" "sqadd z15.s, z15.s, z7.s\n" - "27:" // Height 2: no shift correction + "25:" // Height 2: no shift correction ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n" - "ld1rw { z5.s }, p2/Z, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1rw { z5.s }, p2/Z, [x24]\n" + "add x24, %x[qp], %[maxval]\n" ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n" - "ld1rw { z6.s }, p2/Z, [x19]\n" + "ld1rw { z6.s }, p2/Z, [x24]\n" ".inst 0x4482880c // srshl z12.s, p2/M, z12.s, z0.s\n" "add z8.s, z8.s, z4.s\n" "add z9.s, z9.s, z4.s\n" @@ -590,9 +577,9 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z9.h, z10.h, z11.h\n" "smax z12.s, p2/M, z12.s, z5.s\n" "uzp1 z8.b, z8.b, z9.b\n" - "st1b { z8.b }, p1, [x13]\n" + "st1b { z8.b }, p1, [x9]\n" "add z13.s, z13.s, z4.s\n" - "addvl x13, x13, #1\n" + "addvl x9, x9, #1\n" ".inst 0x4482884e // srshl z14.s, p2/M, z14.s, z2.s\n" ".inst 0x4482886f // srshl z15.s, p2/M, z15.s, z3.s\n" "smin z13.s, p2/M, z13.s, z6.s\n" @@ -606,37 +593,24 @@ void sve_hybrid_s8qs_dot_6x4VL ( "smax z15.s, p2/M, z15.s, z5.s\n" "uzp1 z13.h, z14.h, z15.h\n" "uzp1 z12.b, z12.b, z13.b\n" - "st1b { z12.b }, p1, [x9]\n" - "addvl x9, x9, #1\n" - "28:" // Height 2: Writeback done - "decw x15, ALL, MUL #4\n" - "cmp x15, XZR\n" - "bgt 17b\n" - "b 86f\n" - "29:" // Height 3 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 30f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19\n" - "add x27, x27, x19\n" - "b 31f\n" - "30:" // Height 3: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "add x27, x9, x19\n" - "31:" // Height 3: Column loop + "st1b { z12.b }, p1, [x23]\n" + "26:" // Height 2: Writeback done + "decw x10, ALL, MUL #4\n" + "cmp x10, XZR\n" + "bgt 15b\n" + "b 80f\n" + "27:" // Height 3 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "28:" // Height 3: Column loop "mov z8.s, #0x0\n" "mov x19, #0x0\n" "mov z9.s, #0x0\n" - "whilelt p1.b, x19, x15\n" + "whilelt p1.b, x19, x10\n" "mov z10.s, #0x0\n" "mov z11.s, #0x0\n" "mov z12.s, #0x0\n" @@ -647,208 +621,211 @@ void sve_hybrid_s8qs_dot_6x4VL ( "mov z17.s, #0x0\n" "mov z18.s, #0x0\n" "mov z19.s, #0x0\n" - "32:" // Height 3: setup done - "mov x12, #0x0\n" - "33:" // Height 3: String loop + "29:" // Height 3: setup done + "mov x27, #0x0\n" + "30:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 34f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 31f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x26, [x20, #0x10]\n" - "cbnz x12, 35f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "cbnz x27, 32f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "add x26, x26, x19\n" - "b 35f\n" - "34:" // Height 3: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "add x26, x28, x19\n" - "35:" // Height 3: input setup done - "cmp x11, #0x10\n" - "ble 37f\n" - "36:" // Height 3: Multiply loop: Main loop head - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "sub x11, x11, #0x10\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "add x25, x25, x19\n" + "add x24, x24, x19\n" + "add x23, x23, x19\n" + "b 32f\n" + "31:" // Height 3: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "add x23, x24, x19\n" + "32:" // Height 3: input setup done + "cmp x26, #0x10\n" + "ble 34f\n" + "33:" // Height 3: Multiply loop: Main loop head + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "sub x26, x26, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1rqb { z2.b }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x23, x23, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" - "cmp x11, #0x10\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" + "cmp x26, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "sdot z17.s, z7.b, z2.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "prfm pldl1keep, [x23, #0x80]\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[0]\n" "sdot z15.s, z7.b, z1.b[0]\n" "sdot z19.s, z7.b, z2.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" "sdot z8.s, z6.b, z0.b[1]\n" "sdot z12.s, z6.b, z1.b[1]\n" "sdot z16.s, z6.b, z2.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" "sdot z17.s, z7.b, z2.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #7, MUL VL]\n" - "addvl x14, x14, #16\n" + "ld1b { z7.b }, p2/Z, [x28, #7, MUL VL]\n" + "addvl x28, x28, #16\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" "sdot z18.s, z6.b, z2.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #-8, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-8, MUL VL]\n" "sdot z11.s, z7.b, z0.b[1]\n" "sdot z15.s, z7.b, z1.b[1]\n" "sdot z19.s, z7.b, z2.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #-7, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-7, MUL VL]\n" "sdot z8.s, z6.b, z0.b[2]\n" "sdot z12.s, z6.b, z1.b[2]\n" "sdot z16.s, z6.b, z2.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" "sdot z17.s, z7.b, z2.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-5, MUL VL]\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" "sdot z18.s, z6.b, z2.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[2]\n" "sdot z15.s, z7.b, z1.b[2]\n" "sdot z19.s, z7.b, z2.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-3, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-3, MUL VL]\n" "sdot z8.s, z6.b, z0.b[3]\n" "sdot z12.s, z6.b, z1.b[3]\n" "sdot z16.s, z6.b, z2.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #-2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" "sdot z17.s, z7.b, z2.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #-1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-1, MUL VL]\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z18.s, z6.b, z2.b[3]\n" "sdot z11.s, z7.b, z0.b[3]\n" "sdot z15.s, z7.b, z1.b[3]\n" "sdot z19.s, z7.b, z2.b[3]\n" - "bgt 36b\n" - "37:" // Height 3: Multiply loop: Single iteration only - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "bgt 33b\n" + "34:" // Height 3: Multiply loop: Single iteration only + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1rqb { z2.b }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "add x26, x26, #0x10\n" + "add x23, x23, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" "sdot z16.s, z6.b, z2.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z17.s, z7.b, z2.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" "sdot z11.s, z7.b, z0.b[0]\n" "sdot z15.s, z7.b, z1.b[0]\n" "sdot z19.s, z7.b, z2.b[0]\n" - "ble 38f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 35f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[1]\n" "sdot z16.s, z6.b, z2.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" "sdot z17.s, z7.b, z2.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" "sdot z18.s, z6.b, z2.b[1]\n" "sdot z11.s, z7.b, z0.b[1]\n" "sdot z15.s, z7.b, z1.b[1]\n" "sdot z19.s, z7.b, z2.b[1]\n" - "ble 38f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 35f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[2]\n" "sdot z16.s, z6.b, z2.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" "sdot z17.s, z7.b, z2.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" "sdot z18.s, z6.b, z2.b[2]\n" "sdot z11.s, z7.b, z0.b[2]\n" "sdot z15.s, z7.b, z1.b[2]\n" "sdot z19.s, z7.b, z2.b[2]\n" - "ble 38f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 35f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" "sdot z12.s, z6.b, z1.b[3]\n" "sdot z16.s, z6.b, z2.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" "sdot z17.s, z7.b, z2.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z18.s, z6.b, z2.b[3]\n" "sdot z11.s, z7.b, z0.b[3]\n" "sdot z15.s, z7.b, z1.b[3]\n" "sdot z19.s, z7.b, z2.b[3]\n" - "38:" // Height 3: Multiply loop: multiply skip - "prfm pldl1keep, [x10, #0x80]\n" - "add x12, x12, #0x1\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "35:" // Height 3: Multiply loop: multiply skip + "prfm pldl1keep, [x25, #0x80]\n" + "add x27, x27, #0x1\n" + "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x12, x19\n" - "bne 33b\n" - "ld1w { z0.s }, p2/Z, [x16]\n" + "cmp x27, x19\n" + "bne 30b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z0.s }, p2/Z, [x11]\n" "add z8.s, z8.s, z0.s\n" - "ld1w { z1.s }, p2/Z, [x16, #1, MUL VL]\n" + "ld1w { z1.s }, p2/Z, [x11, #1, MUL VL]\n" + "add x23, x9, x19\n" "add z12.s, z12.s, z0.s\n" - "ld1w { z2.s }, p2/Z, [x16, #2, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x11, #2, MUL VL]\n" + "add x22, x23, x19\n" "add z16.s, z16.s, z0.s\n" - "ld1w { z3.s }, p2/Z, [x16, #3, MUL VL]\n" - "addvl x16, x16, #4\n" + "ld1w { z3.s }, p2/Z, [x11, #3, MUL VL]\n" + "addvl x11, x11, #4\n" "add z9.s, z9.s, z1.s\n" "add z13.s, z13.s, z1.s\n" "add z10.s, z10.s, z2.s\n" @@ -858,30 +835,30 @@ void sve_hybrid_s8qs_dot_6x4VL ( "add z17.s, z17.s, z1.s\n" "add z18.s, z18.s, z2.s\n" "add z19.s, z19.s, z3.s\n" - "tbz %x[flags], #4, 39f\n" - "ld1w { z0.s }, p2/Z, [x17]\n" - "ld1w { z4.s }, p2/Z, [x8]\n" - "ld1w { z1.s }, p2/Z, [x17, #1, MUL VL]\n" - "ld1w { z5.s }, p2/Z, [x8, #1, MUL VL]\n" - "ld1w { z2.s }, p2/Z, [x17, #2, MUL VL]\n" - "ld1w { z6.s }, p2/Z, [x8, #2, MUL VL]\n" - "ld1w { z3.s }, p2/Z, [x17, #3, MUL VL]\n" - "addvl x17, x17, #4\n" - "ld1w { z7.s }, p2/Z, [x8, #3, MUL VL]\n" - "addvl x8, x8, #4\n" - "b 40f\n" - "39:" // Height 3: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1rw { z0.s }, p2/Z, [x19]\n" + "tbz %x[flags], #4, 36f\n" + "ld1w { z0.s }, p2/Z, [x12]\n" + "ld1w { z4.s }, p2/Z, [x13]\n" + "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n" + "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n" + "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "b 37f\n" + "36:" // Height 3: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1rw { z0.s }, p2/Z, [x24]\n" "mov z1.d, z0.d\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" "mov z2.d, z0.d\n" "mov z3.d, z0.d\n" "mov z5.d, z4.d\n" "mov z6.d, z4.d\n" "mov z7.d, z4.d\n" - "40:" // Height 3: parameters loaded + "37:" // Height 3: parameters loaded ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n" ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n" ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n" @@ -894,7 +871,7 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a57631 // sqrdmulh z17.s, z17.s, z5.s\n" ".inst 0x04a67652 // sqrdmulh z18.s, z18.s, z6.s\n" ".inst 0x04a77673 // sqrdmulh z19.s, z19.s, z7.s\n" - "tbz %x[flags], #5, 41f\n" + "tbz %x[flags], #5, 38f\n" "and z4.d, z8.d, z0.d\n" "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" @@ -931,17 +908,17 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sqadd z17.s, z17.s, z5.s\n" "sqadd z18.s, z18.s, z6.s\n" "sqadd z19.s, z19.s, z7.s\n" - "41:" // Height 3: no shift correction + "38:" // Height 3: no shift correction ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n" - "ld1rw { z5.s }, p2/Z, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1rw { z5.s }, p2/Z, [x24]\n" + "add x24, %x[qp], %[maxval]\n" ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n" - "ld1rw { z6.s }, p2/Z, [x19]\n" + "ld1rw { z6.s }, p2/Z, [x24]\n" ".inst 0x4482880c // srshl z12.s, p2/M, z12.s, z0.s\n" "add z8.s, z8.s, z4.s\n" "add z9.s, z9.s, z4.s\n" @@ -962,9 +939,9 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z9.h, z10.h, z11.h\n" "smax z12.s, p2/M, z12.s, z5.s\n" "uzp1 z8.b, z8.b, z9.b\n" - "st1b { z8.b }, p1, [x13]\n" + "st1b { z8.b }, p1, [x9]\n" "add z13.s, z13.s, z4.s\n" - "addvl x13, x13, #1\n" + "addvl x9, x9, #1\n" ".inst 0x4482884e // srshl z14.s, p2/M, z14.s, z2.s\n" ".inst 0x4482886f // srshl z15.s, p2/M, z15.s, z3.s\n" ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n" @@ -987,52 +964,35 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z13.h, z14.h, z15.h\n" ".inst 0x44828873 // srshl z19.s, p2/M, z19.s, z3.s\n" "uzp1 z12.b, z12.b, z13.b\n" - "st1b { z12.b }, p1, [x9]\n" + "st1b { z12.b }, p1, [x23]\n" "add z18.s, z18.s, z4.s\n" - "addvl x9, x9, #1\n" - "add z19.s, z19.s, z4.s\n" "smax z17.s, p2/M, z17.s, z5.s\n" + "add z19.s, z19.s, z4.s\n" "smin z18.s, p2/M, z18.s, z6.s\n" - "smin z19.s, p2/M, z19.s, z6.s\n" "uzp1 z16.h, z16.h, z17.h\n" + "smin z19.s, p2/M, z19.s, z6.s\n" "smax z18.s, p2/M, z18.s, z5.s\n" "smax z19.s, p2/M, z19.s, z5.s\n" "uzp1 z17.h, z18.h, z19.h\n" "uzp1 z16.b, z16.b, z17.b\n" - "st1b { z16.b }, p1, [x27]\n" - "addvl x27, x27, #1\n" - "42:" // Height 3: Writeback done - "decw x15, ALL, MUL #4\n" - "cmp x15, XZR\n" - "bgt 31b\n" - "b 86f\n" - "43:" // Height 4 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 44f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19\n" - "ldr x25, [%x[output_ptr], #0x18]\n" - "add x27, x27, x19\n" - "add x25, x25, x19\n" - "b 45f\n" - "44:" // Height 4: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "add x27, x9, x19\n" - "add x25, x27, x19\n" - "45:" // Height 4: Column loop + "st1b { z16.b }, p1, [x22]\n" + "39:" // Height 3: Writeback done + "decw x10, ALL, MUL #4\n" + "cmp x10, XZR\n" + "bgt 28b\n" + "b 80f\n" + "40:" // Height 4 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "41:" // Height 4: Column loop "mov z8.s, #0x0\n" "mov x19, #0x0\n" "mov z9.s, #0x0\n" - "whilelt p1.b, x19, x15\n" + "whilelt p1.b, x19, x10\n" "mov z10.s, #0x0\n" "mov z11.s, #0x0\n" "mov z12.s, #0x0\n" @@ -1047,123 +1007,123 @@ void sve_hybrid_s8qs_dot_6x4VL ( "mov z21.s, #0x0\n" "mov z22.s, #0x0\n" "mov z23.s, #0x0\n" - "46:" // Height 4: setup done - "mov x12, #0x0\n" - "47:" // Height 4: String loop + "42:" // Height 4: setup done + "mov x27, #0x0\n" + "43:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 48f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 44f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x26, [x20, #0x10]\n" - "ldr x24, [x20, #0x18]\n" - "cbnz x12, 49f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "ldr x22, [x20, #0x18]\n" + "cbnz x27, 45f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "add x26, x26, x19\n" + "add x25, x25, x19\n" "add x24, x24, x19\n" - "b 49f\n" - "48:" // Height 4: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "add x26, x28, x19\n" - "add x24, x26, x19\n" - "49:" // Height 4: input setup done - "cmp x11, #0x10\n" - "ble 51f\n" - "50:" // Height 4: Multiply loop: Main loop head - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "sub x11, x11, #0x10\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "add x23, x23, x19\n" + "add x22, x22, x19\n" + "b 45f\n" + "44:" // Height 4: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "add x23, x24, x19\n" + "add x22, x23, x19\n" + "45:" // Height 4: input setup done + "cmp x26, #0x10\n" + "ble 47f\n" + "46:" // Height 4: Multiply loop: Main loop head + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "sub x26, x26, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1rqb { z2.b }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "add x26, x26, #0x10\n" + "ld1rqb { z3.b }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x22, x22, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "cmp x11, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "cmp x26, #0x10\n" "sdot z20.s, z6.b, z3.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z17.s, z7.b, z2.b[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "sdot z21.s, z7.b, z3.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "prfm pldl1keep, [x22, #0x80]\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" "sdot z22.s, z6.b, z3.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[0]\n" "sdot z15.s, z7.b, z1.b[0]\n" "sdot z19.s, z7.b, z2.b[0]\n" "sdot z23.s, z7.b, z3.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" "sdot z8.s, z6.b, z0.b[1]\n" "sdot z12.s, z6.b, z1.b[1]\n" "sdot z16.s, z6.b, z2.b[1]\n" "sdot z20.s, z6.b, z3.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" "sdot z17.s, z7.b, z2.b[1]\n" "sdot z21.s, z7.b, z3.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #7, MUL VL]\n" - "addvl x14, x14, #16\n" + "ld1b { z7.b }, p2/Z, [x28, #7, MUL VL]\n" + "addvl x28, x28, #16\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" "sdot z18.s, z6.b, z2.b[1]\n" "sdot z22.s, z6.b, z3.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #-8, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-8, MUL VL]\n" "sdot z11.s, z7.b, z0.b[1]\n" "sdot z15.s, z7.b, z1.b[1]\n" "sdot z19.s, z7.b, z2.b[1]\n" "sdot z23.s, z7.b, z3.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #-7, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-7, MUL VL]\n" "sdot z8.s, z6.b, z0.b[2]\n" "sdot z12.s, z6.b, z1.b[2]\n" "sdot z16.s, z6.b, z2.b[2]\n" "sdot z20.s, z6.b, z3.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" "sdot z17.s, z7.b, z2.b[2]\n" "sdot z21.s, z7.b, z3.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-5, MUL VL]\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" "sdot z18.s, z6.b, z2.b[2]\n" "sdot z22.s, z6.b, z3.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[2]\n" "sdot z15.s, z7.b, z1.b[2]\n" "sdot z19.s, z7.b, z2.b[2]\n" "sdot z23.s, z7.b, z3.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-3, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-3, MUL VL]\n" "sdot z8.s, z6.b, z0.b[3]\n" "sdot z12.s, z6.b, z1.b[3]\n" "sdot z16.s, z6.b, z2.b[3]\n" "sdot z20.s, z6.b, z3.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #-2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" "sdot z17.s, z7.b, z2.b[3]\n" "sdot z21.s, z7.b, z3.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #-1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-1, MUL VL]\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z18.s, z6.b, z2.b[3]\n" @@ -1172,31 +1132,31 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z15.s, z7.b, z1.b[3]\n" "sdot z19.s, z7.b, z2.b[3]\n" "sdot z23.s, z7.b, z3.b[3]\n" - "bgt 50b\n" - "51:" // Height 4: Multiply loop: Single iteration only - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "bgt 46b\n" + "47:" // Height 4: Multiply loop: Single iteration only + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1rqb { z2.b }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "add x26, x26, #0x10\n" + "ld1rqb { z3.b }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" - "add x24, x24, #0x10\n" + "add x22, x22, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" "sdot z17.s, z7.b, z2.b[0]\n" "sdot z20.s, z6.b, z3.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z21.s, z7.b, z3.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" @@ -1205,21 +1165,21 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z15.s, z7.b, z1.b[0]\n" "sdot z19.s, z7.b, z2.b[0]\n" "sdot z23.s, z7.b, z3.b[0]\n" - "ble 52f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 48f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[1]\n" "sdot z16.s, z6.b, z2.b[1]\n" "sdot z20.s, z6.b, z3.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" "sdot z17.s, z7.b, z2.b[1]\n" "sdot z21.s, z7.b, z3.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" "sdot z18.s, z6.b, z2.b[1]\n" @@ -1228,21 +1188,21 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z15.s, z7.b, z1.b[1]\n" "sdot z19.s, z7.b, z2.b[1]\n" "sdot z23.s, z7.b, z3.b[1]\n" - "ble 52f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 48f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[2]\n" "sdot z16.s, z6.b, z2.b[2]\n" "sdot z20.s, z6.b, z3.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" "sdot z17.s, z7.b, z2.b[2]\n" "sdot z21.s, z7.b, z3.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" "sdot z18.s, z6.b, z2.b[2]\n" @@ -1251,20 +1211,20 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z15.s, z7.b, z1.b[2]\n" "sdot z19.s, z7.b, z2.b[2]\n" "sdot z23.s, z7.b, z3.b[2]\n" - "ble 52f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 48f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" "sdot z12.s, z6.b, z1.b[3]\n" "sdot z16.s, z6.b, z2.b[3]\n" "sdot z20.s, z6.b, z3.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" "sdot z17.s, z7.b, z2.b[3]\n" "sdot z21.s, z7.b, z3.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z18.s, z6.b, z2.b[3]\n" @@ -1273,24 +1233,28 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z15.s, z7.b, z1.b[3]\n" "sdot z19.s, z7.b, z2.b[3]\n" "sdot z23.s, z7.b, z3.b[3]\n" - "52:" // Height 4: Multiply loop: multiply skip - "prfm pldl1keep, [x10, #0x80]\n" - "add x12, x12, #0x1\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "48:" // Height 4: Multiply loop: multiply skip + "prfm pldl1keep, [x25, #0x80]\n" + "add x27, x27, #0x1\n" "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x12, x19\n" - "bne 47b\n" - "ld1w { z0.s }, p2/Z, [x16]\n" + "cmp x27, x19\n" + "bne 43b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z0.s }, p2/Z, [x11]\n" "add z8.s, z8.s, z0.s\n" - "ld1w { z1.s }, p2/Z, [x16, #1, MUL VL]\n" + "ld1w { z1.s }, p2/Z, [x11, #1, MUL VL]\n" + "add x23, x9, x19\n" "add z12.s, z12.s, z0.s\n" - "ld1w { z2.s }, p2/Z, [x16, #2, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x11, #2, MUL VL]\n" + "add x22, x23, x19\n" "add z16.s, z16.s, z0.s\n" - "ld1w { z3.s }, p2/Z, [x16, #3, MUL VL]\n" - "addvl x16, x16, #4\n" + "ld1w { z3.s }, p2/Z, [x11, #3, MUL VL]\n" + "add x21, x22, x19\n" "add z9.s, z9.s, z1.s\n" + "addvl x11, x11, #4\n" "add z13.s, z13.s, z1.s\n" "add z10.s, z10.s, z2.s\n" "add z11.s, z11.s, z3.s\n" @@ -1303,30 +1267,30 @@ void sve_hybrid_s8qs_dot_6x4VL ( "add z21.s, z21.s, z1.s\n" "add z22.s, z22.s, z2.s\n" "add z23.s, z23.s, z3.s\n" - "tbz %x[flags], #4, 53f\n" - "ld1w { z0.s }, p2/Z, [x17]\n" - "ld1w { z4.s }, p2/Z, [x8]\n" - "ld1w { z1.s }, p2/Z, [x17, #1, MUL VL]\n" - "ld1w { z5.s }, p2/Z, [x8, #1, MUL VL]\n" - "ld1w { z2.s }, p2/Z, [x17, #2, MUL VL]\n" - "ld1w { z6.s }, p2/Z, [x8, #2, MUL VL]\n" - "ld1w { z3.s }, p2/Z, [x17, #3, MUL VL]\n" - "addvl x17, x17, #4\n" - "ld1w { z7.s }, p2/Z, [x8, #3, MUL VL]\n" - "addvl x8, x8, #4\n" - "b 54f\n" - "53:" // Height 4: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1rw { z0.s }, p2/Z, [x19]\n" + "tbz %x[flags], #4, 49f\n" + "ld1w { z0.s }, p2/Z, [x12]\n" + "ld1w { z4.s }, p2/Z, [x13]\n" + "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n" + "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n" + "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "b 50f\n" + "49:" // Height 4: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1rw { z0.s }, p2/Z, [x24]\n" "mov z1.d, z0.d\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" "mov z2.d, z0.d\n" "mov z3.d, z0.d\n" "mov z5.d, z4.d\n" "mov z6.d, z4.d\n" "mov z7.d, z4.d\n" - "54:" // Height 4: parameters loaded + "50:" // Height 4: parameters loaded ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n" ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n" ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n" @@ -1343,7 +1307,7 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a576b5 // sqrdmulh z21.s, z21.s, z5.s\n" ".inst 0x04a676d6 // sqrdmulh z22.s, z22.s, z6.s\n" ".inst 0x04a776f7 // sqrdmulh z23.s, z23.s, z7.s\n" - "tbz %x[flags], #5, 55f\n" + "tbz %x[flags], #5, 51f\n" "and z4.d, z8.d, z0.d\n" "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" @@ -1392,17 +1356,17 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sqadd z21.s, z21.s, z5.s\n" "sqadd z22.s, z22.s, z6.s\n" "sqadd z23.s, z23.s, z7.s\n" - "55:" // Height 4: no shift correction + "51:" // Height 4: no shift correction ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n" - "ld1rw { z5.s }, p2/Z, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1rw { z5.s }, p2/Z, [x24]\n" + "add x24, %x[qp], %[maxval]\n" ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n" - "ld1rw { z6.s }, p2/Z, [x19]\n" + "ld1rw { z6.s }, p2/Z, [x24]\n" ".inst 0x4482880c // srshl z12.s, p2/M, z12.s, z0.s\n" "add z8.s, z8.s, z4.s\n" "add z9.s, z9.s, z4.s\n" @@ -1423,9 +1387,9 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z9.h, z10.h, z11.h\n" "smax z12.s, p2/M, z12.s, z5.s\n" "uzp1 z8.b, z8.b, z9.b\n" - "st1b { z8.b }, p1, [x13]\n" + "st1b { z8.b }, p1, [x9]\n" "add z13.s, z13.s, z4.s\n" - "addvl x13, x13, #1\n" + "addvl x9, x9, #1\n" ".inst 0x4482884e // srshl z14.s, p2/M, z14.s, z2.s\n" ".inst 0x4482886f // srshl z15.s, p2/M, z15.s, z3.s\n" ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n" @@ -1448,76 +1412,55 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z13.h, z14.h, z15.h\n" ".inst 0x44828873 // srshl z19.s, p2/M, z19.s, z3.s\n" "uzp1 z12.b, z12.b, z13.b\n" - "st1b { z12.b }, p1, [x9]\n" + "st1b { z12.b }, p1, [x23]\n" "add z18.s, z18.s, z4.s\n" - "addvl x9, x9, #1\n" - "add z19.s, z19.s, z4.s\n" "smax z17.s, p2/M, z17.s, z5.s\n" + "add z19.s, z19.s, z4.s\n" ".inst 0x44828814 // srshl z20.s, p2/M, z20.s, z0.s\n" "smin z18.s, p2/M, z18.s, z6.s\n" - "smin z19.s, p2/M, z19.s, z6.s\n" "uzp1 z16.h, z16.h, z17.h\n" + "smin z19.s, p2/M, z19.s, z6.s\n" "add z20.s, z20.s, z4.s\n" "smax z18.s, p2/M, z18.s, z5.s\n" + ".inst 0x44828835 // srshl z21.s, p2/M, z21.s, z1.s\n" "smax z19.s, p2/M, z19.s, z5.s\n" "smin z20.s, p2/M, z20.s, z6.s\n" - ".inst 0x44828835 // srshl z21.s, p2/M, z21.s, z1.s\n" ".inst 0x44828856 // srshl z22.s, p2/M, z22.s, z2.s\n" + "add z21.s, z21.s, z4.s\n" "uzp1 z17.h, z18.h, z19.h\n" "smax z20.s, p2/M, z20.s, z5.s\n" - "add z21.s, z21.s, z4.s\n" "add z22.s, z22.s, z4.s\n" "uzp1 z16.b, z16.b, z17.b\n" - "st1b { z16.b }, p1, [x27]\n" + "st1b { z16.b }, p1, [x22]\n" "smin z21.s, p2/M, z21.s, z6.s\n" - "addvl x27, x27, #1\n" "smin z22.s, p2/M, z22.s, z6.s\n" ".inst 0x44828877 // srshl z23.s, p2/M, z23.s, z3.s\n" "smax z21.s, p2/M, z21.s, z5.s\n" - "add z23.s, z23.s, z4.s\n" "smax z22.s, p2/M, z22.s, z5.s\n" + "add z23.s, z23.s, z4.s\n" "uzp1 z20.h, z20.h, z21.h\n" "smin z23.s, p2/M, z23.s, z6.s\n" "smax z23.s, p2/M, z23.s, z5.s\n" "uzp1 z21.h, z22.h, z23.h\n" "uzp1 z20.b, z20.b, z21.b\n" - "st1b { z20.b }, p1, [x25]\n" - "addvl x25, x25, #1\n" - "56:" // Height 4: Writeback done - "decw x15, ALL, MUL #4\n" - "cmp x15, XZR\n" - "bgt 45b\n" - "b 86f\n" - "57:" // Height 5 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 58f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19\n" - "ldr x25, [%x[output_ptr], #0x18]\n" - "ldr x23, [%x[output_ptr], #0x20]\n" - "add x27, x27, x19\n" - "add x25, x25, x19\n" - "add x23, x23, x19\n" - "b 59f\n" - "58:" // Height 5: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "add x27, x9, x19\n" - "add x25, x27, x19\n" - "add x23, x25, x19\n" - "59:" // Height 5: Column loop + "st1b { z20.b }, p1, [x21]\n" + "52:" // Height 4: Writeback done + "decw x10, ALL, MUL #4\n" + "cmp x10, XZR\n" + "bgt 41b\n" + "b 80f\n" + "53:" // Height 5 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "54:" // Height 5: Column loop "mov z8.s, #0x0\n" "mov x19, #0x0\n" "mov z9.s, #0x0\n" - "whilelt p1.b, x19, x15\n" + "whilelt p1.b, x19, x10\n" "mov z10.s, #0x0\n" "mov z11.s, #0x0\n" "mov z12.s, #0x0\n" @@ -1536,143 +1479,143 @@ void sve_hybrid_s8qs_dot_6x4VL ( "mov z25.s, #0x0\n" "mov z26.s, #0x0\n" "mov z27.s, #0x0\n" - "60:" // Height 5: setup done - "mov x12, #0x0\n" - "61:" // Height 5: String loop + "55:" // Height 5: setup done + "mov x27, #0x0\n" + "56:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 62f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 57f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x26, [x20, #0x10]\n" - "ldr x24, [x20, #0x18]\n" - "ldr x22, [x20, #0x20]\n" - "cbnz x12, 63f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "ldr x22, [x20, #0x18]\n" + "ldr x21, [x20, #0x20]\n" + "cbnz x27, 58f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "add x26, x26, x19\n" + "add x25, x25, x19\n" "add x24, x24, x19\n" + "add x23, x23, x19\n" "add x22, x22, x19\n" - "b 63f\n" - "62:" // Height 5: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "add x26, x28, x19\n" - "add x24, x26, x19\n" - "add x22, x24, x19\n" - "63:" // Height 5: input setup done - "cmp x11, #0x10\n" - "ble 65f\n" - "64:" // Height 5: Multiply loop: Main loop head - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "sub x11, x11, #0x10\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "add x21, x21, x19\n" + "b 58f\n" + "57:" // Height 5: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "add x23, x24, x19\n" + "add x22, x23, x19\n" + "add x21, x22, x19\n" + "58:" // Height 5: input setup done + "cmp x26, #0x10\n" + "ble 60f\n" + "59:" // Height 5: Multiply loop: Main loop head + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "sub x26, x26, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1rqb { z2.b }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "add x26, x26, #0x10\n" + "ld1rqb { z3.b }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" - "ld1rqb { z4.b }, p0/Z, [x22]\n" - "add x24, x24, #0x10\n" - "sdot z13.s, z7.b, z1.b[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "ld1rqb { z4.b }, p0/Z, [x21]\n" "add x22, x22, #0x10\n" + "sdot z13.s, z7.b, z1.b[0]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x21, x21, #0x10\n" "sdot z20.s, z6.b, z3.b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "cmp x11, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "cmp x26, #0x10\n" "sdot z24.s, z6.b, z4.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z17.s, z7.b, z2.b[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "sdot z21.s, z7.b, z3.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" "sdot z25.s, z7.b, z4.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "sdot z10.s, z6.b, z0.b[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x21, #0x80]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" "sdot z22.s, z6.b, z3.b[0]\n" "sdot z26.s, z6.b, z4.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[0]\n" "sdot z15.s, z7.b, z1.b[0]\n" "sdot z19.s, z7.b, z2.b[0]\n" "sdot z23.s, z7.b, z3.b[0]\n" "sdot z27.s, z7.b, z4.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" "sdot z8.s, z6.b, z0.b[1]\n" "sdot z12.s, z6.b, z1.b[1]\n" "sdot z16.s, z6.b, z2.b[1]\n" "sdot z20.s, z6.b, z3.b[1]\n" "sdot z24.s, z6.b, z4.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" "sdot z17.s, z7.b, z2.b[1]\n" "sdot z21.s, z7.b, z3.b[1]\n" "sdot z25.s, z7.b, z4.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #7, MUL VL]\n" - "addvl x14, x14, #16\n" + "ld1b { z7.b }, p2/Z, [x28, #7, MUL VL]\n" + "addvl x28, x28, #16\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" "sdot z18.s, z6.b, z2.b[1]\n" "sdot z22.s, z6.b, z3.b[1]\n" "sdot z26.s, z6.b, z4.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #-8, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-8, MUL VL]\n" "sdot z11.s, z7.b, z0.b[1]\n" "sdot z15.s, z7.b, z1.b[1]\n" "sdot z19.s, z7.b, z2.b[1]\n" "sdot z23.s, z7.b, z3.b[1]\n" "sdot z27.s, z7.b, z4.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #-7, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-7, MUL VL]\n" "sdot z8.s, z6.b, z0.b[2]\n" "sdot z12.s, z6.b, z1.b[2]\n" "sdot z16.s, z6.b, z2.b[2]\n" "sdot z20.s, z6.b, z3.b[2]\n" "sdot z24.s, z6.b, z4.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" "sdot z17.s, z7.b, z2.b[2]\n" "sdot z21.s, z7.b, z3.b[2]\n" "sdot z25.s, z7.b, z4.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-5, MUL VL]\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" "sdot z18.s, z6.b, z2.b[2]\n" "sdot z22.s, z6.b, z3.b[2]\n" "sdot z26.s, z6.b, z4.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[2]\n" "sdot z15.s, z7.b, z1.b[2]\n" "sdot z19.s, z7.b, z2.b[2]\n" "sdot z23.s, z7.b, z3.b[2]\n" "sdot z27.s, z7.b, z4.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-3, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-3, MUL VL]\n" "sdot z8.s, z6.b, z0.b[3]\n" "sdot z12.s, z6.b, z1.b[3]\n" "sdot z16.s, z6.b, z2.b[3]\n" "sdot z20.s, z6.b, z3.b[3]\n" "sdot z24.s, z6.b, z4.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #-2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" "sdot z17.s, z7.b, z2.b[3]\n" "sdot z21.s, z7.b, z3.b[3]\n" "sdot z25.s, z7.b, z4.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #-1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-1, MUL VL]\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z18.s, z6.b, z2.b[3]\n" @@ -1683,35 +1626,35 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z19.s, z7.b, z2.b[3]\n" "sdot z23.s, z7.b, z3.b[3]\n" "sdot z27.s, z7.b, z4.b[3]\n" - "bgt 64b\n" - "65:" // Height 5: Multiply loop: Single iteration only - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "bgt 59b\n" + "60:" // Height 5: Multiply loop: Single iteration only + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1rqb { z2.b }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "add x26, x26, #0x10\n" + "ld1rqb { z3.b }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" - "ld1rqb { z4.b }, p0/Z, [x22]\n" - "add x24, x24, #0x10\n" - "sdot z13.s, z7.b, z1.b[0]\n" + "ld1rqb { z4.b }, p0/Z, [x21]\n" "add x22, x22, #0x10\n" + "sdot z13.s, z7.b, z1.b[0]\n" + "add x21, x21, #0x10\n" "sdot z17.s, z7.b, z2.b[0]\n" "sdot z20.s, z6.b, z3.b[0]\n" "sdot z24.s, z6.b, z4.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z21.s, z7.b, z3.b[0]\n" "sdot z25.s, z7.b, z4.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" @@ -1722,23 +1665,23 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z19.s, z7.b, z2.b[0]\n" "sdot z23.s, z7.b, z3.b[0]\n" "sdot z27.s, z7.b, z4.b[0]\n" - "ble 66f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 61f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[1]\n" "sdot z16.s, z6.b, z2.b[1]\n" "sdot z20.s, z6.b, z3.b[1]\n" "sdot z24.s, z6.b, z4.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" "sdot z17.s, z7.b, z2.b[1]\n" "sdot z21.s, z7.b, z3.b[1]\n" "sdot z25.s, z7.b, z4.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" "sdot z18.s, z6.b, z2.b[1]\n" @@ -1749,23 +1692,23 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z19.s, z7.b, z2.b[1]\n" "sdot z23.s, z7.b, z3.b[1]\n" "sdot z27.s, z7.b, z4.b[1]\n" - "ble 66f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 61f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[2]\n" "sdot z16.s, z6.b, z2.b[2]\n" "sdot z20.s, z6.b, z3.b[2]\n" "sdot z24.s, z6.b, z4.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" "sdot z17.s, z7.b, z2.b[2]\n" "sdot z21.s, z7.b, z3.b[2]\n" "sdot z25.s, z7.b, z4.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" "sdot z18.s, z6.b, z2.b[2]\n" @@ -1776,22 +1719,22 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z19.s, z7.b, z2.b[2]\n" "sdot z23.s, z7.b, z3.b[2]\n" "sdot z27.s, z7.b, z4.b[2]\n" - "ble 66f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 61f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" "sdot z12.s, z6.b, z1.b[3]\n" "sdot z16.s, z6.b, z2.b[3]\n" "sdot z20.s, z6.b, z3.b[3]\n" "sdot z24.s, z6.b, z4.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" "sdot z17.s, z7.b, z2.b[3]\n" "sdot z21.s, z7.b, z3.b[3]\n" "sdot z25.s, z7.b, z4.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z18.s, z6.b, z2.b[3]\n" @@ -1802,26 +1745,31 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z19.s, z7.b, z2.b[3]\n" "sdot z23.s, z7.b, z3.b[3]\n" "sdot z27.s, z7.b, z4.b[3]\n" - "66:" // Height 5: Multiply loop: multiply skip - "prfm pldl1keep, [x10, #0x80]\n" - "add x12, x12, #0x1\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "61:" // Height 5: Multiply loop: multiply skip + "prfm pldl1keep, [x25, #0x80]\n" + "add x27, x27, #0x1\n" "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x21, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x12, x19\n" - "bne 61b\n" - "ld1w { z0.s }, p2/Z, [x16]\n" + "cmp x27, x19\n" + "bne 56b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z0.s }, p2/Z, [x11]\n" "add z8.s, z8.s, z0.s\n" - "ld1w { z1.s }, p2/Z, [x16, #1, MUL VL]\n" + "ld1w { z1.s }, p2/Z, [x11, #1, MUL VL]\n" + "add x23, x9, x19\n" "add z12.s, z12.s, z0.s\n" - "ld1w { z2.s }, p2/Z, [x16, #2, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x11, #2, MUL VL]\n" + "add x22, x23, x19\n" "add z16.s, z16.s, z0.s\n" - "ld1w { z3.s }, p2/Z, [x16, #3, MUL VL]\n" - "addvl x16, x16, #4\n" + "ld1w { z3.s }, p2/Z, [x11, #3, MUL VL]\n" + "add x21, x22, x19\n" "add z9.s, z9.s, z1.s\n" + "add x20, x21, x19\n" "add z13.s, z13.s, z1.s\n" + "addvl x11, x11, #4\n" "add z10.s, z10.s, z2.s\n" "add z11.s, z11.s, z3.s\n" "add z14.s, z14.s, z2.s\n" @@ -1837,30 +1785,30 @@ void sve_hybrid_s8qs_dot_6x4VL ( "add z25.s, z25.s, z1.s\n" "add z26.s, z26.s, z2.s\n" "add z27.s, z27.s, z3.s\n" - "tbz %x[flags], #4, 67f\n" - "ld1w { z0.s }, p2/Z, [x17]\n" - "ld1w { z4.s }, p2/Z, [x8]\n" - "ld1w { z1.s }, p2/Z, [x17, #1, MUL VL]\n" - "ld1w { z5.s }, p2/Z, [x8, #1, MUL VL]\n" - "ld1w { z2.s }, p2/Z, [x17, #2, MUL VL]\n" - "ld1w { z6.s }, p2/Z, [x8, #2, MUL VL]\n" - "ld1w { z3.s }, p2/Z, [x17, #3, MUL VL]\n" - "addvl x17, x17, #4\n" - "ld1w { z7.s }, p2/Z, [x8, #3, MUL VL]\n" - "addvl x8, x8, #4\n" - "b 68f\n" - "67:" // Height 5: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1rw { z0.s }, p2/Z, [x19]\n" + "tbz %x[flags], #4, 62f\n" + "ld1w { z0.s }, p2/Z, [x12]\n" + "ld1w { z4.s }, p2/Z, [x13]\n" + "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n" + "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n" + "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "b 63f\n" + "62:" // Height 5: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1rw { z0.s }, p2/Z, [x24]\n" "mov z1.d, z0.d\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" "mov z2.d, z0.d\n" "mov z3.d, z0.d\n" "mov z5.d, z4.d\n" "mov z6.d, z4.d\n" "mov z7.d, z4.d\n" - "68:" // Height 5: parameters loaded + "63:" // Height 5: parameters loaded ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n" ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n" ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n" @@ -1881,7 +1829,7 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a57739 // sqrdmulh z25.s, z25.s, z5.s\n" ".inst 0x04a6775a // sqrdmulh z26.s, z26.s, z6.s\n" ".inst 0x04a7777b // sqrdmulh z27.s, z27.s, z7.s\n" - "tbz %x[flags], #5, 69f\n" + "tbz %x[flags], #5, 64f\n" "and z4.d, z8.d, z0.d\n" "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" @@ -1942,17 +1890,17 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sqadd z25.s, z25.s, z5.s\n" "sqadd z26.s, z26.s, z6.s\n" "sqadd z27.s, z27.s, z7.s\n" - "69:" // Height 5: no shift correction + "64:" // Height 5: no shift correction ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n" - "ld1rw { z5.s }, p2/Z, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1rw { z5.s }, p2/Z, [x24]\n" + "add x24, %x[qp], %[maxval]\n" ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n" - "ld1rw { z6.s }, p2/Z, [x19]\n" + "ld1rw { z6.s }, p2/Z, [x24]\n" ".inst 0x4482880c // srshl z12.s, p2/M, z12.s, z0.s\n" "add z8.s, z8.s, z4.s\n" "add z9.s, z9.s, z4.s\n" @@ -1973,9 +1921,9 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z9.h, z10.h, z11.h\n" "smax z12.s, p2/M, z12.s, z5.s\n" "uzp1 z8.b, z8.b, z9.b\n" - "st1b { z8.b }, p1, [x13]\n" + "st1b { z8.b }, p1, [x9]\n" "add z13.s, z13.s, z4.s\n" - "addvl x13, x13, #1\n" + "addvl x9, x9, #1\n" ".inst 0x4482884e // srshl z14.s, p2/M, z14.s, z2.s\n" ".inst 0x4482886f // srshl z15.s, p2/M, z15.s, z3.s\n" ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n" @@ -1998,29 +1946,27 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z13.h, z14.h, z15.h\n" ".inst 0x44828873 // srshl z19.s, p2/M, z19.s, z3.s\n" "uzp1 z12.b, z12.b, z13.b\n" - "st1b { z12.b }, p1, [x9]\n" + "st1b { z12.b }, p1, [x23]\n" "add z18.s, z18.s, z4.s\n" - "addvl x9, x9, #1\n" - "add z19.s, z19.s, z4.s\n" "smax z17.s, p2/M, z17.s, z5.s\n" + "add z19.s, z19.s, z4.s\n" ".inst 0x44828814 // srshl z20.s, p2/M, z20.s, z0.s\n" "smin z18.s, p2/M, z18.s, z6.s\n" - "smin z19.s, p2/M, z19.s, z6.s\n" "uzp1 z16.h, z16.h, z17.h\n" + "smin z19.s, p2/M, z19.s, z6.s\n" "add z20.s, z20.s, z4.s\n" "smax z18.s, p2/M, z18.s, z5.s\n" + ".inst 0x44828835 // srshl z21.s, p2/M, z21.s, z1.s\n" "smax z19.s, p2/M, z19.s, z5.s\n" "smin z20.s, p2/M, z20.s, z6.s\n" - ".inst 0x44828835 // srshl z21.s, p2/M, z21.s, z1.s\n" ".inst 0x44828856 // srshl z22.s, p2/M, z22.s, z2.s\n" + "add z21.s, z21.s, z4.s\n" "uzp1 z17.h, z18.h, z19.h\n" "smax z20.s, p2/M, z20.s, z5.s\n" - "add z21.s, z21.s, z4.s\n" "add z22.s, z22.s, z4.s\n" "uzp1 z16.b, z16.b, z17.b\n" - "st1b { z16.b }, p1, [x27]\n" + "st1b { z16.b }, p1, [x22]\n" "smin z21.s, p2/M, z21.s, z6.s\n" - "addvl x27, x27, #1\n" "smin z22.s, p2/M, z22.s, z6.s\n" ".inst 0x44828877 // srshl z23.s, p2/M, z23.s, z3.s\n" ".inst 0x44828818 // srshl z24.s, p2/M, z24.s, z0.s\n" @@ -2043,57 +1989,35 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x4482887b // srshl z27.s, p2/M, z27.s, z3.s\n" "uzp1 z24.h, z24.h, z25.h\n" "uzp1 z20.b, z20.b, z21.b\n" - "st1b { z20.b }, p1, [x25]\n" + "st1b { z20.b }, p1, [x21]\n" "add z27.s, z27.s, z4.s\n" - "addvl x25, x25, #1\n" "smin z26.s, p2/M, z26.s, z6.s\n" "smin z27.s, p2/M, z27.s, z6.s\n" "smax z26.s, p2/M, z26.s, z5.s\n" "smax z27.s, p2/M, z27.s, z5.s\n" "uzp1 z25.h, z26.h, z27.h\n" "uzp1 z24.b, z24.b, z25.b\n" - "st1b { z24.b }, p1, [x23]\n" - "addvl x23, x23, #1\n" - "70:" // Height 5: Writeback done - "decw x15, ALL, MUL #4\n" - "cmp x15, XZR\n" - "bgt 59b\n" - "b 86f\n" - "71:" // Height 6 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "st1b { z24.b }, p1, [x20]\n" + "65:" // Height 5: Writeback done + "decw x10, ALL, MUL #4\n" + "cmp x10, XZR\n" + "bgt 54b\n" + "b 80f\n" + "66:" // Height 6 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "mov x20, #0x6\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 72f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19\n" - "ldr x25, [%x[output_ptr], #0x18]\n" - "ldr x23, [%x[output_ptr], #0x20]\n" - "add x27, x27, x19\n" - "ldr x21, [%x[output_ptr], #0x28]\n" - "add %x[output_ptr], %x[output_ptr], #0x30\n" - "add x25, x25, x19\n" - "add x23, x23, x19\n" - "add x21, x21, x19\n" - "b 73f\n" - "72:" // Height 6: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "add x27, x9, x19\n" - "add x25, x27, x19\n" - "add x23, x25, x19\n" - "add x21, x23, x19\n" - "add %x[output_ptr], x21, x19\n" - "73:" // Height 6: Column loop + "madd %x[output_ptr], x19, x20, %x[output_ptr]\n" + "67:" // Height 6: Column loop "mov z8.s, #0x0\n" "mov x19, #0x0\n" "mov z9.s, #0x0\n" - "whilelt p1.b, x19, x15\n" + "whilelt p1.b, x19, x10\n" "mov z10.s, #0x0\n" "mov z11.s, #0x0\n" "mov z12.s, #0x0\n" @@ -2116,77 +2040,77 @@ void sve_hybrid_s8qs_dot_6x4VL ( "mov z29.s, #0x0\n" "mov z30.s, #0x0\n" "mov z31.s, #0x0\n" - "74:" // Height 6: setup done - "mov x12, #0x0\n" - "75:" // Height 6: String loop + "68:" // Height 6: setup done + "mov x27, #0x0\n" + "69:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 76f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 70f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x26, [x20, #0x10]\n" - "ldr x24, [x20, #0x18]\n" - "ldr x22, [x20, #0x20]\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "ldr x22, [x20, #0x18]\n" + "ldr x21, [x20, #0x20]\n" "ldr x20, [x20, #0x28]\n" - "cbnz x12, 77f\n" + "cbnz x27, 71f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "add x26, x26, x19\n" + "add x25, x25, x19\n" "add x24, x24, x19\n" + "add x23, x23, x19\n" "add x22, x22, x19\n" + "add x21, x21, x19\n" "add x20, x20, x19\n" - "b 77f\n" - "76:" // Height 6: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "add x26, x28, x19\n" - "add x24, x26, x19\n" - "add x22, x24, x19\n" - "add x20, x22, x19\n" - "77:" // Height 6: input setup done - "cmp x11, #0x10\n" - "ble 79f\n" - "78:" // Height 6: Multiply loop: Main loop head - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "sub x11, x11, #0x10\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "b 71f\n" + "70:" // Height 6: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "add x23, x24, x19\n" + "add x22, x23, x19\n" + "add x21, x22, x19\n" + "add x20, x21, x19\n" + "71:" // Height 6: input setup done + "cmp x26, #0x10\n" + "ble 73f\n" + "72:" // Height 6: Multiply loop: Main loop head + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "sub x26, x26, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1rqb { z2.b }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "add x26, x26, #0x10\n" + "ld1rqb { z3.b }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" - "ld1rqb { z4.b }, p0/Z, [x22]\n" - "add x24, x24, #0x10\n" + "ld1rqb { z4.b }, p0/Z, [x21]\n" + "add x22, x22, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" "ld1rqb { z5.b }, p0/Z, [x20]\n" - "add x22, x22, #0x10\n" + "add x21, x21, #0x10\n" "sdot z20.s, z6.b, z3.b[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "add x20, x20, #0x10\n" "sdot z24.s, z6.b, z4.b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "cmp x11, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "cmp x26, #0x10\n" "sdot z28.s, z6.b, z5.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z17.s, z7.b, z2.b[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "sdot z21.s, z7.b, z3.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "sdot z25.s, z7.b, z4.b[0]\n" "prfm pldl1keep, [x22, #0x80]\n" + "sdot z25.s, z7.b, z4.b[0]\n" + "prfm pldl1keep, [x21, #0x80]\n" "sdot z29.s, z7.b, z5.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "sdot z10.s, z6.b, z0.b[0]\n" "prfm pldl1keep, [x20, #0x80]\n" "sdot z14.s, z6.b, z1.b[0]\n" @@ -2194,85 +2118,85 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z22.s, z6.b, z3.b[0]\n" "sdot z26.s, z6.b, z4.b[0]\n" "sdot z30.s, z6.b, z5.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[0]\n" "sdot z15.s, z7.b, z1.b[0]\n" "sdot z19.s, z7.b, z2.b[0]\n" "sdot z23.s, z7.b, z3.b[0]\n" "sdot z27.s, z7.b, z4.b[0]\n" "sdot z31.s, z7.b, z5.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" "sdot z8.s, z6.b, z0.b[1]\n" "sdot z12.s, z6.b, z1.b[1]\n" "sdot z16.s, z6.b, z2.b[1]\n" "sdot z20.s, z6.b, z3.b[1]\n" "sdot z24.s, z6.b, z4.b[1]\n" "sdot z28.s, z6.b, z5.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" "sdot z17.s, z7.b, z2.b[1]\n" "sdot z21.s, z7.b, z3.b[1]\n" "sdot z25.s, z7.b, z4.b[1]\n" "sdot z29.s, z7.b, z5.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #7, MUL VL]\n" - "addvl x14, x14, #16\n" + "ld1b { z7.b }, p2/Z, [x28, #7, MUL VL]\n" + "addvl x28, x28, #16\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" "sdot z18.s, z6.b, z2.b[1]\n" "sdot z22.s, z6.b, z3.b[1]\n" "sdot z26.s, z6.b, z4.b[1]\n" "sdot z30.s, z6.b, z5.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #-8, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-8, MUL VL]\n" "sdot z11.s, z7.b, z0.b[1]\n" "sdot z15.s, z7.b, z1.b[1]\n" "sdot z19.s, z7.b, z2.b[1]\n" "sdot z23.s, z7.b, z3.b[1]\n" "sdot z27.s, z7.b, z4.b[1]\n" "sdot z31.s, z7.b, z5.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #-7, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-7, MUL VL]\n" "sdot z8.s, z6.b, z0.b[2]\n" "sdot z12.s, z6.b, z1.b[2]\n" "sdot z16.s, z6.b, z2.b[2]\n" "sdot z20.s, z6.b, z3.b[2]\n" "sdot z24.s, z6.b, z4.b[2]\n" "sdot z28.s, z6.b, z5.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-6, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-6, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" "sdot z17.s, z7.b, z2.b[2]\n" "sdot z21.s, z7.b, z3.b[2]\n" "sdot z25.s, z7.b, z4.b[2]\n" "sdot z29.s, z7.b, z5.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-5, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-5, MUL VL]\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" "sdot z18.s, z6.b, z2.b[2]\n" "sdot z22.s, z6.b, z3.b[2]\n" "sdot z26.s, z6.b, z4.b[2]\n" "sdot z30.s, z6.b, z5.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #-4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-4, MUL VL]\n" "sdot z11.s, z7.b, z0.b[2]\n" "sdot z15.s, z7.b, z1.b[2]\n" "sdot z19.s, z7.b, z2.b[2]\n" "sdot z23.s, z7.b, z3.b[2]\n" "sdot z27.s, z7.b, z4.b[2]\n" "sdot z31.s, z7.b, z5.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #-3, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-3, MUL VL]\n" "sdot z8.s, z6.b, z0.b[3]\n" "sdot z12.s, z6.b, z1.b[3]\n" "sdot z16.s, z6.b, z2.b[3]\n" "sdot z20.s, z6.b, z3.b[3]\n" "sdot z24.s, z6.b, z4.b[3]\n" "sdot z28.s, z6.b, z5.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #-2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #-2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" "sdot z17.s, z7.b, z2.b[3]\n" "sdot z21.s, z7.b, z3.b[3]\n" "sdot z25.s, z7.b, z4.b[3]\n" "sdot z29.s, z7.b, z5.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #-1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #-1, MUL VL]\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z18.s, z6.b, z2.b[3]\n" @@ -2285,39 +2209,39 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z23.s, z7.b, z3.b[3]\n" "sdot z27.s, z7.b, z4.b[3]\n" "sdot z31.s, z7.b, z5.b[3]\n" - "bgt 78b\n" - "79:" // Height 6: Multiply loop: Single iteration only - "ld1b { z6.b }, p2/Z, [x14]\n" - "whilelt p0.b, XZR, x11\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" - "ld1rqb { z0.b }, p0/Z, [x10]\n" + "bgt 72b\n" + "73:" // Height 6: Multiply loop: Single iteration only + "ld1b { z6.b }, p2/Z, [x28]\n" + "whilelt p0.b, XZR, x26\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" - "ld1rqb { z1.b }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "ld1rqb { z2.b }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "add x26, x26, #0x10\n" + "ld1rqb { z3.b }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" - "ld1rqb { z4.b }, p0/Z, [x22]\n" - "add x24, x24, #0x10\n" + "ld1rqb { z4.b }, p0/Z, [x21]\n" + "add x22, x22, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" "ld1rqb { z5.b }, p0/Z, [x20]\n" - "add x22, x22, #0x10\n" + "add x21, x21, #0x10\n" "sdot z20.s, z6.b, z3.b[0]\n" "add x20, x20, #0x10\n" "sdot z17.s, z7.b, z2.b[0]\n" "sdot z24.s, z6.b, z4.b[0]\n" "sdot z28.s, z6.b, z5.b[0]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z21.s, z7.b, z3.b[0]\n" "sdot z25.s, z7.b, z4.b[0]\n" "sdot z29.s, z7.b, z5.b[0]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" @@ -2330,25 +2254,25 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z23.s, z7.b, z3.b[0]\n" "sdot z27.s, z7.b, z4.b[0]\n" "sdot z31.s, z7.b, z5.b[0]\n" - "ble 80f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 74f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[1]\n" "sdot z16.s, z6.b, z2.b[1]\n" "sdot z20.s, z6.b, z3.b[1]\n" "sdot z24.s, z6.b, z4.b[1]\n" "sdot z28.s, z6.b, z5.b[1]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[1]\n" "sdot z13.s, z7.b, z1.b[1]\n" "sdot z17.s, z7.b, z2.b[1]\n" "sdot z21.s, z7.b, z3.b[1]\n" "sdot z25.s, z7.b, z4.b[1]\n" "sdot z29.s, z7.b, z5.b[1]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[1]\n" "sdot z14.s, z6.b, z1.b[1]\n" "sdot z18.s, z6.b, z2.b[1]\n" @@ -2361,25 +2285,25 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z23.s, z7.b, z3.b[1]\n" "sdot z27.s, z7.b, z4.b[1]\n" "sdot z31.s, z7.b, z5.b[1]\n" - "ble 80f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 74f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" - "subs x11, x11, #0x4\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x26, x26, #0x4\n" "sdot z12.s, z6.b, z1.b[2]\n" "sdot z16.s, z6.b, z2.b[2]\n" "sdot z20.s, z6.b, z3.b[2]\n" "sdot z24.s, z6.b, z4.b[2]\n" "sdot z28.s, z6.b, z5.b[2]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[2]\n" "sdot z13.s, z7.b, z1.b[2]\n" "sdot z17.s, z7.b, z2.b[2]\n" "sdot z21.s, z7.b, z3.b[2]\n" "sdot z25.s, z7.b, z4.b[2]\n" "sdot z29.s, z7.b, z5.b[2]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[2]\n" "sdot z14.s, z6.b, z1.b[2]\n" "sdot z18.s, z6.b, z2.b[2]\n" @@ -2392,24 +2316,24 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z23.s, z7.b, z3.b[2]\n" "sdot z27.s, z7.b, z4.b[2]\n" "sdot z31.s, z7.b, z5.b[2]\n" - "ble 80f\n" - "ld1b { z6.b }, p2/Z, [x14]\n" + "ble 74f\n" + "ld1b { z6.b }, p2/Z, [x28]\n" "sdot z8.s, z6.b, z0.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #1, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #1, MUL VL]\n" "sdot z12.s, z6.b, z1.b[3]\n" "sdot z16.s, z6.b, z2.b[3]\n" "sdot z20.s, z6.b, z3.b[3]\n" "sdot z24.s, z6.b, z4.b[3]\n" "sdot z28.s, z6.b, z5.b[3]\n" - "ld1b { z6.b }, p2/Z, [x14, #2, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b[3]\n" "sdot z13.s, z7.b, z1.b[3]\n" "sdot z17.s, z7.b, z2.b[3]\n" "sdot z21.s, z7.b, z3.b[3]\n" "sdot z25.s, z7.b, z4.b[3]\n" "sdot z29.s, z7.b, z5.b[3]\n" - "ld1b { z7.b }, p2/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "addvl x28, x28, #4\n" "sdot z10.s, z6.b, z0.b[3]\n" "sdot z14.s, z6.b, z1.b[3]\n" "sdot z18.s, z6.b, z2.b[3]\n" @@ -2422,28 +2346,34 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z23.s, z7.b, z3.b[3]\n" "sdot z27.s, z7.b, z4.b[3]\n" "sdot z31.s, z7.b, z5.b[3]\n" - "80:" // Height 6: Multiply loop: multiply skip - "prfm pldl1keep, [x10, #0x80]\n" - "add x12, x12, #0x1\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "74:" // Height 6: Multiply loop: multiply skip + "prfm pldl1keep, [x25, #0x80]\n" + "add x27, x27, #0x1\n" "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x21, #0x80]\n" "prfm pldl1keep, [x20, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x12, x19\n" - "bne 75b\n" - "ld1w { z0.s }, p2/Z, [x16]\n" + "cmp x27, x19\n" + "bne 69b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z0.s }, p2/Z, [x11]\n" "add z8.s, z8.s, z0.s\n" - "ld1w { z1.s }, p2/Z, [x16, #1, MUL VL]\n" + "ld1w { z1.s }, p2/Z, [x11, #1, MUL VL]\n" + "add x23, x9, x19\n" "add z12.s, z12.s, z0.s\n" - "ld1w { z2.s }, p2/Z, [x16, #2, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x11, #2, MUL VL]\n" + "add x22, x23, x19\n" "add z16.s, z16.s, z0.s\n" - "ld1w { z3.s }, p2/Z, [x16, #3, MUL VL]\n" - "addvl x16, x16, #4\n" + "ld1w { z3.s }, p2/Z, [x11, #3, MUL VL]\n" + "add x21, x22, x19\n" "add z9.s, z9.s, z1.s\n" + "add x20, x21, x19\n" "add z13.s, z13.s, z1.s\n" + "add x19, x20, x19\n" "add z10.s, z10.s, z2.s\n" + "addvl x11, x11, #4\n" "add z11.s, z11.s, z3.s\n" "add z14.s, z14.s, z2.s\n" "add z15.s, z15.s, z3.s\n" @@ -2462,30 +2392,30 @@ void sve_hybrid_s8qs_dot_6x4VL ( "add z29.s, z29.s, z1.s\n" "add z30.s, z30.s, z2.s\n" "add z31.s, z31.s, z3.s\n" - "tbz %x[flags], #4, 81f\n" - "ld1w { z0.s }, p2/Z, [x17]\n" - "ld1w { z4.s }, p2/Z, [x8]\n" - "ld1w { z1.s }, p2/Z, [x17, #1, MUL VL]\n" - "ld1w { z5.s }, p2/Z, [x8, #1, MUL VL]\n" - "ld1w { z2.s }, p2/Z, [x17, #2, MUL VL]\n" - "ld1w { z6.s }, p2/Z, [x8, #2, MUL VL]\n" - "ld1w { z3.s }, p2/Z, [x17, #3, MUL VL]\n" - "addvl x17, x17, #4\n" - "ld1w { z7.s }, p2/Z, [x8, #3, MUL VL]\n" - "addvl x8, x8, #4\n" - "b 82f\n" - "81:" // Height 6: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1rw { z0.s }, p2/Z, [x19]\n" + "tbz %x[flags], #4, 75f\n" + "ld1w { z0.s }, p2/Z, [x12]\n" + "ld1w { z4.s }, p2/Z, [x13]\n" + "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n" + "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n" + "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n" + "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "b 76f\n" + "75:" // Height 6: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1rw { z0.s }, p2/Z, [x24]\n" "mov z1.d, z0.d\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" "mov z2.d, z0.d\n" "mov z3.d, z0.d\n" "mov z5.d, z4.d\n" "mov z6.d, z4.d\n" "mov z7.d, z4.d\n" - "82:" // Height 6: parameters loaded + "76:" // Height 6: parameters loaded ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n" ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n" ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n" @@ -2510,7 +2440,7 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a577bd // sqrdmulh z29.s, z29.s, z5.s\n" ".inst 0x04a677de // sqrdmulh z30.s, z30.s, z6.s\n" ".inst 0x04a777ff // sqrdmulh z31.s, z31.s, z7.s\n" - "tbz %x[flags], #5, 83f\n" + "tbz %x[flags], #5, 77f\n" "and z4.d, z8.d, z0.d\n" "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" @@ -2583,17 +2513,17 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sqadd z29.s, z29.s, z5.s\n" "sqadd z30.s, z30.s, z6.s\n" "sqadd z31.s, z31.s, z7.s\n" - "83:" // Height 6: no shift correction + "77:" // Height 6: no shift correction ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1rw { z4.s }, p2/Z, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1rw { z4.s }, p2/Z, [x24]\n" ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n" - "ld1rw { z5.s }, p2/Z, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1rw { z5.s }, p2/Z, [x24]\n" + "add x24, %x[qp], %[maxval]\n" ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n" - "ld1rw { z6.s }, p2/Z, [x19]\n" + "ld1rw { z6.s }, p2/Z, [x24]\n" ".inst 0x4482880c // srshl z12.s, p2/M, z12.s, z0.s\n" "add z8.s, z8.s, z4.s\n" "add z9.s, z9.s, z4.s\n" @@ -2614,9 +2544,9 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z9.h, z10.h, z11.h\n" "smax z12.s, p2/M, z12.s, z5.s\n" "uzp1 z8.b, z8.b, z9.b\n" - "st1b { z8.b }, p1, [x13]\n" + "st1b { z8.b }, p1, [x9]\n" "add z13.s, z13.s, z4.s\n" - "addvl x13, x13, #1\n" + "addvl x9, x9, #1\n" ".inst 0x4482884e // srshl z14.s, p2/M, z14.s, z2.s\n" ".inst 0x4482886f // srshl z15.s, p2/M, z15.s, z3.s\n" ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n" @@ -2639,29 +2569,27 @@ void sve_hybrid_s8qs_dot_6x4VL ( "uzp1 z13.h, z14.h, z15.h\n" ".inst 0x44828873 // srshl z19.s, p2/M, z19.s, z3.s\n" "uzp1 z12.b, z12.b, z13.b\n" - "st1b { z12.b }, p1, [x9]\n" + "st1b { z12.b }, p1, [x23]\n" "add z18.s, z18.s, z4.s\n" - "addvl x9, x9, #1\n" - "add z19.s, z19.s, z4.s\n" "smax z17.s, p2/M, z17.s, z5.s\n" + "add z19.s, z19.s, z4.s\n" ".inst 0x44828814 // srshl z20.s, p2/M, z20.s, z0.s\n" "smin z18.s, p2/M, z18.s, z6.s\n" - "smin z19.s, p2/M, z19.s, z6.s\n" "uzp1 z16.h, z16.h, z17.h\n" + "smin z19.s, p2/M, z19.s, z6.s\n" "add z20.s, z20.s, z4.s\n" "smax z18.s, p2/M, z18.s, z5.s\n" + ".inst 0x44828835 // srshl z21.s, p2/M, z21.s, z1.s\n" "smax z19.s, p2/M, z19.s, z5.s\n" "smin z20.s, p2/M, z20.s, z6.s\n" - ".inst 0x44828835 // srshl z21.s, p2/M, z21.s, z1.s\n" ".inst 0x44828856 // srshl z22.s, p2/M, z22.s, z2.s\n" + "add z21.s, z21.s, z4.s\n" "uzp1 z17.h, z18.h, z19.h\n" "smax z20.s, p2/M, z20.s, z5.s\n" - "add z21.s, z21.s, z4.s\n" "add z22.s, z22.s, z4.s\n" "uzp1 z16.b, z16.b, z17.b\n" - "st1b { z16.b }, p1, [x27]\n" + "st1b { z16.b }, p1, [x22]\n" "smin z21.s, p2/M, z21.s, z6.s\n" - "addvl x27, x27, #1\n" "smin z22.s, p2/M, z22.s, z6.s\n" ".inst 0x44828877 // srshl z23.s, p2/M, z23.s, z3.s\n" ".inst 0x44828818 // srshl z24.s, p2/M, z24.s, z0.s\n" @@ -2684,58 +2612,55 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x4482887b // srshl z27.s, p2/M, z27.s, z3.s\n" "uzp1 z24.h, z24.h, z25.h\n" "uzp1 z20.b, z20.b, z21.b\n" - "st1b { z20.b }, p1, [x25]\n" + "st1b { z20.b }, p1, [x21]\n" "add z27.s, z27.s, z4.s\n" - "addvl x25, x25, #1\n" "smin z26.s, p2/M, z26.s, z6.s\n" ".inst 0x4482881c // srshl z28.s, p2/M, z28.s, z0.s\n" ".inst 0x4482883d // srshl z29.s, p2/M, z29.s, z1.s\n" "smin z27.s, p2/M, z27.s, z6.s\n" - ".inst 0x4482885e // srshl z30.s, p2/M, z30.s, z2.s\n" + "smax z26.s, p2/M, z26.s, z5.s\n" "add z28.s, z28.s, z4.s\n" "add z29.s, z29.s, z4.s\n" - "smax z26.s, p2/M, z26.s, z5.s\n" - "add z30.s, z30.s, z4.s\n" "smax z27.s, p2/M, z27.s, z5.s\n" "smin z28.s, p2/M, z28.s, z6.s\n" "smin z29.s, p2/M, z29.s, z6.s\n" - "smin z30.s, p2/M, z30.s, z6.s\n" + ".inst 0x4482885e // srshl z30.s, p2/M, z30.s, z2.s\n" "uzp1 z25.h, z26.h, z27.h\n" "smax z28.s, p2/M, z28.s, z5.s\n" "uzp1 z24.b, z24.b, z25.b\n" - "st1b { z24.b }, p1, [x23]\n" + "st1b { z24.b }, p1, [x20]\n" + "add z30.s, z30.s, z4.s\n" "smax z29.s, p2/M, z29.s, z5.s\n" - "addvl x23, x23, #1\n" - "smax z30.s, p2/M, z30.s, z5.s\n" ".inst 0x4482887f // srshl z31.s, p2/M, z31.s, z3.s\n" + "smin z30.s, p2/M, z30.s, z6.s\n" "uzp1 z28.h, z28.h, z29.h\n" "add z31.s, z31.s, z4.s\n" + "smax z30.s, p2/M, z30.s, z5.s\n" "smin z31.s, p2/M, z31.s, z6.s\n" "smax z31.s, p2/M, z31.s, z5.s\n" "uzp1 z29.h, z30.h, z31.h\n" "uzp1 z28.b, z28.b, z29.b\n" - "st1b { z28.b }, p1, [x21]\n" - "addvl x21, x21, #1\n" - "84:" // Height 6: Writeback done - "decw x15, ALL, MUL #4\n" - "cmp x15, XZR\n" - "bgt 73b\n" + "st1b { z28.b }, p1, [x19]\n" + "78:" // Height 6: Writeback done + "decw x10, ALL, MUL #4\n" + "cmp x10, XZR\n" + "bgt 67b\n" "subs %x[M], %x[M], #0x6\n" - "beq 86f\n" + "beq 80f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" - "tbz %x[flags], #3, 85f\n" + "tbz %x[flags], #3, 79f\n" "add x20, x20, #0x6\n" "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "b 1b\n" - "85:" // Update direct input + "79:" // Update direct input "mov x19, #0x6\n" "madd %x[input_ptr], x19, x20, %x[input_ptr]\n" "b 1b\n" - "86:" // Exit + "80:" // Exit : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp) - : "cc", "memory", "p0", "p1", "p2", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x13", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } -- cgit v1.2.1