From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp | 2283 ++++++++++---------- 1 file changed, 1141 insertions(+), 1142 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp index 71c6afba42..e1581f2026 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp @@ -139,11 +139,11 @@ void sve_hybrid_fp32_mla_6x4VL ( "6:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 7f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" "cbnz x28, 8f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -156,87 +156,87 @@ void sve_hybrid_fp32_mla_6x4VL ( "9:" // Height 1: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #4, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #5, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #6, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #7, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10]\n" + "fmla z8.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z9.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z10.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z11.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x10, #4, MUL VL]\n" + "fmla z8.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x10, #5, MUL VL]\n" + "fmla z9.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x10, #6, MUL VL]\n" + "fmla z10.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #-8, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #-7, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-6, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #-5, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-4, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #-3, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #-2, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #-1, MUL VL]\n" + "fmla z11.s, z16.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10, #-8, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #-7, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[2]\n" + "fmla z9.s, z16.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10, #-6, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #-5, MUL VL]\n" + "fmla z10.s, z17.s, z0.s[2]\n" + "fmla z11.s, z16.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10, #-4, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #-3, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[3]\n" + "fmla z9.s, z16.s, z0.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10, #-2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #-1, MUL VL]\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" + "fmla z10.s, z17.s, z0.s[3]\n" + "fmla z11.s, z16.s, z0.s[3]\n" "add x26, x26, #0x10\n" "bgt 9b\n" "10:" // Height 1: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10]\n" + "fmla z8.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z9.s, z16.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" "subs x27, x27, #0x1\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z10.s, z17.s, z0.s[0]\n" + "fmla z11.s, z16.s, z0.s[0]\n" "addvl x10, x10, #4\n" "ble 11f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[1]\n" + "fmla z9.s, z16.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" + "fmla z10.s, z17.s, z0.s[1]\n" + "fmla z11.s, z16.s, z0.s[1]\n" "addvl x10, x10, #4\n" "ble 11f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[2]\n" + "fmla z9.s, z16.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" + "fmla z10.s, z17.s, z0.s[2]\n" + "fmla z11.s, z16.s, z0.s[2]\n" "addvl x10, x10, #4\n" "ble 11f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[3]\n" + "fmla z9.s, z16.s, z0.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z10.s, z17.s, z0.s[3]\n" + "fmla z11.s, z16.s, z0.s[3]\n" "addvl x10, x10, #4\n" "11:" // Height 1: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -245,17 +245,17 @@ void sve_hybrid_fp32_mla_6x4VL ( "bne 6b\n" "tbz %x[flags], #1, 12f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z17.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" + "ld1rw { z16.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z17.s\n" + "fmin z9.s, p5/M, z9.s, z17.s\n" + "fmin z10.s, p5/M, z10.s, z17.s\n" + "fmin z11.s, p5/M, z11.s, z17.s\n" + "fmax z8.s, p5/M, z8.s, z16.s\n" + "fmax z9.s, p5/M, z9.s, z16.s\n" + "fmax z10.s, p5/M, z10.s, z16.s\n" + "fmax z11.s, p5/M, z11.s, z16.s\n" "12:" // Height 1: No activation "st1w { z8.s }, p4, [x9]\n" "st1w { z9.s }, p3, [x9, #1, MUL VL]\n" @@ -295,15 +295,15 @@ void sve_hybrid_fp32_mla_6x4VL ( "16:" // Height 2: no bias "tbz %x[flags], #0, 17f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" + "add x20, x9, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x9]\n" "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x20]\n" + "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n" "b 18f\n" "17:" // Height 2: no accumulate "mov z8.b, #0x0\n" @@ -319,12 +319,12 @@ void sve_hybrid_fp32_mla_6x4VL ( "19:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 20f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" "cbnz x28, 21f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -332,130 +332,130 @@ void sve_hybrid_fp32_mla_6x4VL ( "b 21f\n" "20:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" "21:" // Height 2: input setup done "cmp x27, #0x4\n" "ble 23f\n" "22:" // Height 2: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "ld1rqw { z0.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #4, MUL VL]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z17.s, z1.s[0]\n" + "fmla z12.s, z17.s, z0.s[0]\n" + "fmla z9.s, z16.s, z1.s[0]\n" + "fmla z13.s, z16.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z10.s, z17.s, z1.s[0]\n" + "fmla z14.s, z17.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10, #4, MUL VL]\n" "cmp x27, #0x4\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #5, MUL VL]\n" + "fmla z11.s, z16.s, z1.s[0]\n" + "fmla z15.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x10, #5, MUL VL]\n" "add x26, x26, #0x10\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #6, MUL VL]\n" + "fmla z8.s, z17.s, z1.s[1]\n" + "fmla z12.s, z17.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10, #6, MUL VL]\n" "add x25, x25, #0x10\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #7, MUL VL]\n" + "fmla z9.s, z16.s, z1.s[1]\n" + "fmla z13.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #-8, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #-7, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-6, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-5, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-4, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-3, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #-2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "ld1w { z7.s }, p5/Z, [x10, #-1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" + "fmla z10.s, z17.s, z1.s[1]\n" + "fmla z14.s, z17.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10, #-8, MUL VL]\n" + "fmla z11.s, z16.s, z1.s[1]\n" + "fmla z15.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x10, #-7, MUL VL]\n" + "fmla z8.s, z17.s, z1.s[2]\n" + "fmla z12.s, z17.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10, #-6, MUL VL]\n" + "fmla z9.s, z16.s, z1.s[2]\n" + "fmla z13.s, z16.s, z0.s[2]\n" + "ld1w { z16.s }, p5/Z, [x10, #-5, MUL VL]\n" + "fmla z10.s, z17.s, z1.s[2]\n" + "fmla z14.s, z17.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10, #-4, MUL VL]\n" + "fmla z11.s, z16.s, z1.s[2]\n" + "fmla z15.s, z16.s, z0.s[2]\n" + "ld1w { z16.s }, p5/Z, [x10, #-3, MUL VL]\n" + "fmla z8.s, z17.s, z1.s[3]\n" + "fmla z12.s, z17.s, z0.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10, #-2, MUL VL]\n" + "fmla z9.s, z16.s, z1.s[3]\n" + "fmla z13.s, z16.s, z0.s[3]\n" + "ld1w { z16.s }, p5/Z, [x10, #-1, MUL VL]\n" + "fmla z10.s, z17.s, z1.s[3]\n" + "fmla z14.s, z17.s, z0.s[3]\n" + "fmla z11.s, z16.s, z1.s[3]\n" + "fmla z15.s, z16.s, z0.s[3]\n" "bgt 22b\n" "23:" // Height 2: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" "ld1rqw { z1.s }, p0/Z, [x25]\n" "subs x27, x27, #0x1\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[0]\n" + "fmla z12.s, z17.s, z1.s[0]\n" + "fmla z9.s, z16.s, z0.s[0]\n" + "fmla z13.s, z16.s, z1.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z10.s, z17.s, z0.s[0]\n" + "fmla z14.s, z17.s, z1.s[0]\n" "addvl x10, x10, #4\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" + "fmla z11.s, z16.s, z0.s[0]\n" + "fmla z15.s, z16.s, z1.s[0]\n" "ble 24f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[1]\n" + "fmla z12.s, z17.s, z1.s[1]\n" + "fmla z9.s, z16.s, z0.s[1]\n" + "fmla z13.s, z16.s, z1.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" + "fmla z10.s, z17.s, z0.s[1]\n" + "fmla z14.s, z17.s, z1.s[1]\n" "addvl x10, x10, #4\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" + "fmla z11.s, z16.s, z0.s[1]\n" + "fmla z15.s, z16.s, z1.s[1]\n" "ble 24f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[2]\n" + "fmla z12.s, z17.s, z1.s[2]\n" + "fmla z9.s, z16.s, z0.s[2]\n" + "fmla z13.s, z16.s, z1.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" + "fmla z10.s, z17.s, z0.s[2]\n" + "fmla z14.s, z17.s, z1.s[2]\n" "addvl x10, x10, #4\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" + "fmla z11.s, z16.s, z0.s[2]\n" + "fmla z15.s, z16.s, z1.s[2]\n" "ble 24f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[3]\n" + "fmla z12.s, z17.s, z1.s[3]\n" + "fmla z9.s, z16.s, z0.s[3]\n" + "fmla z13.s, z16.s, z1.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z10.s, z17.s, z0.s[3]\n" + "fmla z14.s, z17.s, z1.s[3]\n" "addvl x10, x10, #4\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" + "fmla z11.s, z16.s, z0.s[3]\n" + "fmla z15.s, z16.s, z1.s[3]\n" "24:" // Height 2: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -465,25 +465,25 @@ void sve_hybrid_fp32_mla_6x4VL ( "add x25, x9, x20, LSL #2\n" "tbz %x[flags], #1, 25f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z17.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" + "ld1rw { z16.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z17.s\n" + "fmin z9.s, p5/M, z9.s, z17.s\n" + "fmin z10.s, p5/M, z10.s, z17.s\n" + "fmin z11.s, p5/M, z11.s, z17.s\n" + "fmin z12.s, p5/M, z12.s, z17.s\n" + "fmin z13.s, p5/M, z13.s, z17.s\n" + "fmin z14.s, p5/M, z14.s, z17.s\n" + "fmin z15.s, p5/M, z15.s, z17.s\n" + "fmax z8.s, p5/M, z8.s, z16.s\n" + "fmax z9.s, p5/M, z9.s, z16.s\n" + "fmax z10.s, p5/M, z10.s, z16.s\n" + "fmax z11.s, p5/M, z11.s, z16.s\n" + "fmax z12.s, p5/M, z12.s, z16.s\n" + "fmax z13.s, p5/M, z13.s, z16.s\n" + "fmax z14.s, p5/M, z14.s, z16.s\n" + "fmax z15.s, p5/M, z15.s, z16.s\n" "25:" // Height 2: No activation "st1w { z8.s }, p4, [x9]\n" "st1w { z9.s }, p3, [x9, #1, MUL VL]\n" @@ -531,20 +531,20 @@ void sve_hybrid_fp32_mla_6x4VL ( "29:" // Height 3: no bias "tbz %x[flags], #0, 30f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x21, x9, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x9]\n" "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x24]\n" - "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x21]\n" + "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x20]\n" + "ld1w { z17.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x20, #3, MUL VL]\n" "b 31f\n" "30:" // Height 3: no accumulate "mov z8.b, #0x0\n" @@ -564,13 +564,13 @@ void sve_hybrid_fp32_mla_6x4VL ( "32:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 33f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" "cbnz x28, 34f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -579,86 +579,86 @@ void sve_hybrid_fp32_mla_6x4VL ( "b 34f\n" "33:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" "34:" // Height 3: input setup done "cmp x27, #0x4\n" "ble 36f\n" "35:" // Height 3: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" + "ld1rqw { z2.s }, p0/Z, [x26]\n" "ld1rqw { z1.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1rqw { z0.s }, p0/Z, [x24]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" + "fmla z8.s, z21.s, z2.s[0]\n" + "fmla z12.s, z21.s, z1.s[0]\n" + "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z16.s, z21.s, z0.s[0]\n" + "fmla z9.s, z20.s, z2.s[0]\n" + "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[0]\n" + "fmla z17.s, z20.s, z0.s[0]\n" + "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n" "cmp x27, #0x4\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" + "fmla z10.s, z21.s, z2.s[0]\n" + "fmla z14.s, z21.s, z1.s[0]\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #4, MUL VL]\n" + "fmla z18.s, z21.s, z0.s[0]\n" + "fmla z11.s, z20.s, z2.s[0]\n" + "ld1w { z21.s }, p5/Z, [x10, #4, MUL VL]\n" "add x24, x24, #0x10\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #5, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #6, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #7, MUL VL]\n" + "fmla z15.s, z20.s, z1.s[0]\n" + "fmla z19.s, z20.s, z0.s[0]\n" + "ld1w { z20.s }, p5/Z, [x10, #5, MUL VL]\n" + "fmla z8.s, z21.s, z2.s[1]\n" + "fmla z12.s, z21.s, z1.s[1]\n" + "fmla z16.s, z21.s, z0.s[1]\n" + "fmla z9.s, z20.s, z2.s[1]\n" + "ld1w { z21.s }, p5/Z, [x10, #6, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[1]\n" + "fmla z17.s, z20.s, z0.s[1]\n" + "ld1w { z20.s }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #-8, MUL VL]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #-7, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-6, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-5, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-4, MUL VL]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-3, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #-2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "ld1w { z7.s }, p5/Z, [x10, #-1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" + "fmla z10.s, z21.s, z2.s[1]\n" + "fmla z14.s, z21.s, z1.s[1]\n" + "fmla z18.s, z21.s, z0.s[1]\n" + "fmla z11.s, z20.s, z2.s[1]\n" + "ld1w { z21.s }, p5/Z, [x10, #-8, MUL VL]\n" + "fmla z15.s, z20.s, z1.s[1]\n" + "fmla z19.s, z20.s, z0.s[1]\n" + "ld1w { z20.s }, p5/Z, [x10, #-7, MUL VL]\n" + "fmla z8.s, z21.s, z2.s[2]\n" + "fmla z12.s, z21.s, z1.s[2]\n" + "fmla z16.s, z21.s, z0.s[2]\n" + "fmla z9.s, z20.s, z2.s[2]\n" + "ld1w { z21.s }, p5/Z, [x10, #-6, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[2]\n" + "fmla z17.s, z20.s, z0.s[2]\n" + "ld1w { z20.s }, p5/Z, [x10, #-5, MUL VL]\n" + "fmla z10.s, z21.s, z2.s[2]\n" + "fmla z14.s, z21.s, z1.s[2]\n" + "fmla z18.s, z21.s, z0.s[2]\n" + "fmla z11.s, z20.s, z2.s[2]\n" + "ld1w { z21.s }, p5/Z, [x10, #-4, MUL VL]\n" + "fmla z15.s, z20.s, z1.s[2]\n" + "fmla z19.s, z20.s, z0.s[2]\n" + "ld1w { z20.s }, p5/Z, [x10, #-3, MUL VL]\n" + "fmla z8.s, z21.s, z2.s[3]\n" + "fmla z12.s, z21.s, z1.s[3]\n" + "fmla z16.s, z21.s, z0.s[3]\n" + "fmla z9.s, z20.s, z2.s[3]\n" + "ld1w { z21.s }, p5/Z, [x10, #-2, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[3]\n" + "fmla z17.s, z20.s, z0.s[3]\n" + "ld1w { z20.s }, p5/Z, [x10, #-1, MUL VL]\n" + "fmla z10.s, z21.s, z2.s[3]\n" + "fmla z14.s, z21.s, z1.s[3]\n" + "fmla z18.s, z21.s, z0.s[3]\n" + "fmla z11.s, z20.s, z2.s[3]\n" + "fmla z15.s, z20.s, z1.s[3]\n" + "fmla z19.s, z20.s, z0.s[3]\n" "bgt 35b\n" "36:" // Height 3: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" @@ -666,79 +666,79 @@ void sve_hybrid_fp32_mla_6x4VL ( "ld1rqw { z1.s }, p0/Z, [x25]\n" "subs x27, x27, #0x1\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" + "fmla z8.s, z21.s, z0.s[0]\n" + "fmla z12.s, z21.s, z1.s[0]\n" + "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z16.s, z21.s, z2.s[0]\n" + "fmla z9.s, z20.s, z0.s[0]\n" + "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[0]\n" + "fmla z17.s, z20.s, z2.s[0]\n" + "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" + "fmla z10.s, z21.s, z0.s[0]\n" + "fmla z14.s, z21.s, z1.s[0]\n" + "fmla z18.s, z21.s, z2.s[0]\n" + "fmla z11.s, z20.s, z0.s[0]\n" + "fmla z15.s, z20.s, z1.s[0]\n" + "fmla z19.s, z20.s, z2.s[0]\n" "ble 37f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" + "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z21.s, z0.s[1]\n" + "fmla z12.s, z21.s, z1.s[1]\n" + "fmla z16.s, z21.s, z2.s[1]\n" + "fmla z9.s, z20.s, z0.s[1]\n" + "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n" "subs x27, x27, #0x1\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[1]\n" + "fmla z17.s, z20.s, z2.s[1]\n" + "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" + "fmla z10.s, z21.s, z0.s[1]\n" + "fmla z14.s, z21.s, z1.s[1]\n" + "fmla z18.s, z21.s, z2.s[1]\n" + "fmla z11.s, z20.s, z0.s[1]\n" + "fmla z15.s, z20.s, z1.s[1]\n" + "fmla z19.s, z20.s, z2.s[1]\n" "ble 37f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" + "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z21.s, z0.s[2]\n" + "fmla z12.s, z21.s, z1.s[2]\n" + "fmla z16.s, z21.s, z2.s[2]\n" + "fmla z9.s, z20.s, z0.s[2]\n" + "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n" "subs x27, x27, #0x1\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[2]\n" + "fmla z17.s, z20.s, z2.s[2]\n" + "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" + "fmla z10.s, z21.s, z0.s[2]\n" + "fmla z14.s, z21.s, z1.s[2]\n" + "fmla z18.s, z21.s, z2.s[2]\n" + "fmla z11.s, z20.s, z0.s[2]\n" + "fmla z15.s, z20.s, z1.s[2]\n" + "fmla z19.s, z20.s, z2.s[2]\n" "ble 37f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" + "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z21.s, z0.s[3]\n" + "fmla z12.s, z21.s, z1.s[3]\n" + "fmla z16.s, z21.s, z2.s[3]\n" + "fmla z9.s, z20.s, z0.s[3]\n" + "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[3]\n" + "fmla z17.s, z20.s, z2.s[3]\n" + "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" + "fmla z10.s, z21.s, z0.s[3]\n" + "fmla z14.s, z21.s, z1.s[3]\n" + "fmla z18.s, z21.s, z2.s[3]\n" + "fmla z11.s, z20.s, z0.s[3]\n" + "fmla z15.s, z20.s, z1.s[3]\n" + "fmla z19.s, z20.s, z2.s[3]\n" "37:" // Height 3: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -749,33 +749,33 @@ void sve_hybrid_fp32_mla_6x4VL ( "add x24, x25, x20, LSL #2\n" "tbz %x[flags], #1, 38f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z21.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmin z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z1.s\n" - "fmin z18.s, p5/M, z18.s, z1.s\n" - "fmin z19.s, p5/M, z19.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" - "fmax z16.s, p5/M, z16.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z0.s\n" - "fmax z18.s, p5/M, z18.s, z0.s\n" - "fmax z19.s, p5/M, z19.s, z0.s\n" + "ld1rw { z20.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z21.s\n" + "fmin z9.s, p5/M, z9.s, z21.s\n" + "fmin z10.s, p5/M, z10.s, z21.s\n" + "fmin z11.s, p5/M, z11.s, z21.s\n" + "fmin z12.s, p5/M, z12.s, z21.s\n" + "fmin z13.s, p5/M, z13.s, z21.s\n" + "fmin z14.s, p5/M, z14.s, z21.s\n" + "fmin z15.s, p5/M, z15.s, z21.s\n" + "fmin z16.s, p5/M, z16.s, z21.s\n" + "fmin z17.s, p5/M, z17.s, z21.s\n" + "fmin z18.s, p5/M, z18.s, z21.s\n" + "fmin z19.s, p5/M, z19.s, z21.s\n" + "fmax z8.s, p5/M, z8.s, z20.s\n" + "fmax z9.s, p5/M, z9.s, z20.s\n" + "fmax z10.s, p5/M, z10.s, z20.s\n" + "fmax z11.s, p5/M, z11.s, z20.s\n" + "fmax z12.s, p5/M, z12.s, z20.s\n" + "fmax z13.s, p5/M, z13.s, z20.s\n" + "fmax z14.s, p5/M, z14.s, z20.s\n" + "fmax z15.s, p5/M, z15.s, z20.s\n" + "fmax z16.s, p5/M, z16.s, z20.s\n" + "fmax z17.s, p5/M, z17.s, z20.s\n" + "fmax z18.s, p5/M, z18.s, z20.s\n" + "fmax z19.s, p5/M, z19.s, z20.s\n" "38:" // Height 3: No activation "st1w { z8.s }, p4, [x9]\n" "st1w { z9.s }, p3, [x9, #1, MUL VL]\n" @@ -831,25 +831,25 @@ void sve_hybrid_fp32_mla_6x4VL ( "42:" // Height 4: no bias "tbz %x[flags], #0, 43f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x22, x9, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x9]\n" - "add x23, x24, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x24]\n" - "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x22]\n" + "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x21]\n" + "ld1w { z17.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x20]\n" + "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n" "b 44f\n" "43:" // Height 4: no accumulate "mov z8.b, #0x0\n" @@ -873,14 +873,14 @@ void sve_hybrid_fp32_mla_6x4VL ( "45:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 46f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" "cbnz x28, 47f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -890,105 +890,105 @@ void sve_hybrid_fp32_mla_6x4VL ( "b 47f\n" "46:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "47:" // Height 4: input setup done "cmp x27, #0x4\n" "ble 49f\n" "48:" // Height 4: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" + "ld1rqw { z3.s }, p0/Z, [x26]\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" + "ld1rqw { z1.s }, p0/Z, [x24]\n" + "ld1rqw { z0.s }, p0/Z, [x23]\n" "cmp x27, #0x4\n" "add x26, x26, #0x10\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" + "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z25.s, z3.s[0]\n" + "fmla z12.s, z25.s, z2.s[0]\n" + "fmla z16.s, z25.s, z1.s[0]\n" + "fmla z20.s, z25.s, z0.s[0]\n" + "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n" "add x25, x25, #0x10\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" + "fmla z9.s, z24.s, z3.s[0]\n" + "fmla z13.s, z24.s, z2.s[0]\n" "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #4, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #5, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #6, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #7, MUL VL]\n" + "fmla z17.s, z24.s, z1.s[0]\n" + "fmla z21.s, z24.s, z0.s[0]\n" + "ld1w { z24.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z10.s, z25.s, z3.s[0]\n" + "fmla z14.s, z25.s, z2.s[0]\n" + "fmla z18.s, z25.s, z1.s[0]\n" + "fmla z22.s, z25.s, z0.s[0]\n" + "ld1w { z25.s }, p5/Z, [x10, #4, MUL VL]\n" + "fmla z11.s, z24.s, z3.s[0]\n" + "fmla z15.s, z24.s, z2.s[0]\n" + "fmla z19.s, z24.s, z1.s[0]\n" + "fmla z23.s, z24.s, z0.s[0]\n" + "ld1w { z24.s }, p5/Z, [x10, #5, MUL VL]\n" + "fmla z8.s, z25.s, z3.s[1]\n" + "fmla z12.s, z25.s, z2.s[1]\n" + "fmla z16.s, z25.s, z1.s[1]\n" + "fmla z20.s, z25.s, z0.s[1]\n" + "ld1w { z25.s }, p5/Z, [x10, #6, MUL VL]\n" + "fmla z9.s, z24.s, z3.s[1]\n" + "fmla z13.s, z24.s, z2.s[1]\n" + "fmla z17.s, z24.s, z1.s[1]\n" + "fmla z21.s, z24.s, z0.s[1]\n" + "ld1w { z24.s }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #-8, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #-7, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-6, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-5, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-4, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-3, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #-2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "ld1w { z7.s }, p5/Z, [x10, #-1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" + "fmla z10.s, z25.s, z3.s[1]\n" + "fmla z14.s, z25.s, z2.s[1]\n" + "fmla z18.s, z25.s, z1.s[1]\n" + "fmla z22.s, z25.s, z0.s[1]\n" + "ld1w { z25.s }, p5/Z, [x10, #-8, MUL VL]\n" + "fmla z11.s, z24.s, z3.s[1]\n" + "fmla z15.s, z24.s, z2.s[1]\n" + "fmla z19.s, z24.s, z1.s[1]\n" + "fmla z23.s, z24.s, z0.s[1]\n" + "ld1w { z24.s }, p5/Z, [x10, #-7, MUL VL]\n" + "fmla z8.s, z25.s, z3.s[2]\n" + "fmla z12.s, z25.s, z2.s[2]\n" + "fmla z16.s, z25.s, z1.s[2]\n" + "fmla z20.s, z25.s, z0.s[2]\n" + "ld1w { z25.s }, p5/Z, [x10, #-6, MUL VL]\n" + "fmla z9.s, z24.s, z3.s[2]\n" + "fmla z13.s, z24.s, z2.s[2]\n" + "fmla z17.s, z24.s, z1.s[2]\n" + "fmla z21.s, z24.s, z0.s[2]\n" + "ld1w { z24.s }, p5/Z, [x10, #-5, MUL VL]\n" + "fmla z10.s, z25.s, z3.s[2]\n" + "fmla z14.s, z25.s, z2.s[2]\n" + "fmla z18.s, z25.s, z1.s[2]\n" + "fmla z22.s, z25.s, z0.s[2]\n" + "ld1w { z25.s }, p5/Z, [x10, #-4, MUL VL]\n" + "fmla z11.s, z24.s, z3.s[2]\n" + "fmla z15.s, z24.s, z2.s[2]\n" + "fmla z19.s, z24.s, z1.s[2]\n" + "fmla z23.s, z24.s, z0.s[2]\n" + "ld1w { z24.s }, p5/Z, [x10, #-3, MUL VL]\n" + "fmla z8.s, z25.s, z3.s[3]\n" + "fmla z12.s, z25.s, z2.s[3]\n" + "fmla z16.s, z25.s, z1.s[3]\n" + "fmla z20.s, z25.s, z0.s[3]\n" + "ld1w { z25.s }, p5/Z, [x10, #-2, MUL VL]\n" + "fmla z9.s, z24.s, z3.s[3]\n" + "fmla z13.s, z24.s, z2.s[3]\n" + "fmla z17.s, z24.s, z1.s[3]\n" + "fmla z21.s, z24.s, z0.s[3]\n" + "ld1w { z24.s }, p5/Z, [x10, #-1, MUL VL]\n" + "fmla z10.s, z25.s, z3.s[3]\n" + "fmla z14.s, z25.s, z2.s[3]\n" + "fmla z18.s, z25.s, z1.s[3]\n" + "fmla z22.s, z25.s, z0.s[3]\n" + "fmla z11.s, z24.s, z3.s[3]\n" + "fmla z15.s, z24.s, z2.s[3]\n" + "fmla z19.s, z24.s, z1.s[3]\n" + "fmla z23.s, z24.s, z0.s[3]\n" "bgt 48b\n" "49:" // Height 4: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" @@ -997,95 +997,95 @@ void sve_hybrid_fp32_mla_6x4VL ( "subs x27, x27, #0x1\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" "ld1rqw { z3.s }, p0/Z, [x23]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" + "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z25.s, z0.s[0]\n" + "fmla z12.s, z25.s, z1.s[0]\n" + "fmla z16.s, z25.s, z2.s[0]\n" + "fmla z20.s, z25.s, z3.s[0]\n" + "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z24.s, z0.s[0]\n" + "fmla z13.s, z24.s, z1.s[0]\n" + "fmla z17.s, z24.s, z2.s[0]\n" + "fmla z21.s, z24.s, z3.s[0]\n" + "ld1w { z24.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" + "fmla z10.s, z25.s, z0.s[0]\n" + "fmla z14.s, z25.s, z1.s[0]\n" + "fmla z18.s, z25.s, z2.s[0]\n" + "fmla z22.s, z25.s, z3.s[0]\n" + "fmla z11.s, z24.s, z0.s[0]\n" + "fmla z15.s, z24.s, z1.s[0]\n" + "fmla z19.s, z24.s, z2.s[0]\n" + "fmla z23.s, z24.s, z3.s[0]\n" "ble 50f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" + "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z25.s, z0.s[1]\n" + "fmla z12.s, z25.s, z1.s[1]\n" + "fmla z16.s, z25.s, z2.s[1]\n" + "fmla z20.s, z25.s, z3.s[1]\n" + "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n" "subs x27, x27, #0x1\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z9.s, z24.s, z0.s[1]\n" + "fmla z13.s, z24.s, z1.s[1]\n" + "fmla z17.s, z24.s, z2.s[1]\n" + "fmla z21.s, z24.s, z3.s[1]\n" + "ld1w { z24.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" + "fmla z10.s, z25.s, z0.s[1]\n" + "fmla z14.s, z25.s, z1.s[1]\n" + "fmla z18.s, z25.s, z2.s[1]\n" + "fmla z22.s, z25.s, z3.s[1]\n" + "fmla z11.s, z24.s, z0.s[1]\n" + "fmla z15.s, z24.s, z1.s[1]\n" + "fmla z19.s, z24.s, z2.s[1]\n" + "fmla z23.s, z24.s, z3.s[1]\n" "ble 50f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" + "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z25.s, z0.s[2]\n" + "fmla z12.s, z25.s, z1.s[2]\n" + "fmla z16.s, z25.s, z2.s[2]\n" + "fmla z20.s, z25.s, z3.s[2]\n" + "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n" "subs x27, x27, #0x1\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z9.s, z24.s, z0.s[2]\n" + "fmla z13.s, z24.s, z1.s[2]\n" + "fmla z17.s, z24.s, z2.s[2]\n" + "fmla z21.s, z24.s, z3.s[2]\n" + "ld1w { z24.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" + "fmla z10.s, z25.s, z0.s[2]\n" + "fmla z14.s, z25.s, z1.s[2]\n" + "fmla z18.s, z25.s, z2.s[2]\n" + "fmla z22.s, z25.s, z3.s[2]\n" + "fmla z11.s, z24.s, z0.s[2]\n" + "fmla z15.s, z24.s, z1.s[2]\n" + "fmla z19.s, z24.s, z2.s[2]\n" + "fmla z23.s, z24.s, z3.s[2]\n" "ble 50f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" + "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z25.s, z0.s[3]\n" + "fmla z12.s, z25.s, z1.s[3]\n" + "fmla z16.s, z25.s, z2.s[3]\n" + "fmla z20.s, z25.s, z3.s[3]\n" + "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z24.s, z0.s[3]\n" + "fmla z13.s, z24.s, z1.s[3]\n" + "fmla z17.s, z24.s, z2.s[3]\n" + "fmla z21.s, z24.s, z3.s[3]\n" + "ld1w { z24.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" + "fmla z10.s, z25.s, z0.s[3]\n" + "fmla z14.s, z25.s, z1.s[3]\n" + "fmla z18.s, z25.s, z2.s[3]\n" + "fmla z22.s, z25.s, z3.s[3]\n" + "fmla z11.s, z24.s, z0.s[3]\n" + "fmla z15.s, z24.s, z1.s[3]\n" + "fmla z19.s, z24.s, z2.s[3]\n" + "fmla z23.s, z24.s, z3.s[3]\n" "50:" // Height 4: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1097,41 +1097,41 @@ void sve_hybrid_fp32_mla_6x4VL ( "add x23, x24, x20, LSL #2\n" "tbz %x[flags], #1, 51f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z25.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmin z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z1.s\n" - "fmin z18.s, p5/M, z18.s, z1.s\n" - "fmin z19.s, p5/M, z19.s, z1.s\n" - "fmin z20.s, p5/M, z20.s, z1.s\n" - "fmin z21.s, p5/M, z21.s, z1.s\n" - "fmin z22.s, p5/M, z22.s, z1.s\n" - "fmin z23.s, p5/M, z23.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" - "fmax z16.s, p5/M, z16.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z0.s\n" - "fmax z18.s, p5/M, z18.s, z0.s\n" - "fmax z19.s, p5/M, z19.s, z0.s\n" - "fmax z20.s, p5/M, z20.s, z0.s\n" - "fmax z21.s, p5/M, z21.s, z0.s\n" - "fmax z22.s, p5/M, z22.s, z0.s\n" - "fmax z23.s, p5/M, z23.s, z0.s\n" + "ld1rw { z24.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z25.s\n" + "fmin z9.s, p5/M, z9.s, z25.s\n" + "fmin z10.s, p5/M, z10.s, z25.s\n" + "fmin z11.s, p5/M, z11.s, z25.s\n" + "fmin z12.s, p5/M, z12.s, z25.s\n" + "fmin z13.s, p5/M, z13.s, z25.s\n" + "fmin z14.s, p5/M, z14.s, z25.s\n" + "fmin z15.s, p5/M, z15.s, z25.s\n" + "fmin z16.s, p5/M, z16.s, z25.s\n" + "fmin z17.s, p5/M, z17.s, z25.s\n" + "fmin z18.s, p5/M, z18.s, z25.s\n" + "fmin z19.s, p5/M, z19.s, z25.s\n" + "fmin z20.s, p5/M, z20.s, z25.s\n" + "fmin z21.s, p5/M, z21.s, z25.s\n" + "fmin z22.s, p5/M, z22.s, z25.s\n" + "fmin z23.s, p5/M, z23.s, z25.s\n" + "fmax z8.s, p5/M, z8.s, z24.s\n" + "fmax z9.s, p5/M, z9.s, z24.s\n" + "fmax z10.s, p5/M, z10.s, z24.s\n" + "fmax z11.s, p5/M, z11.s, z24.s\n" + "fmax z12.s, p5/M, z12.s, z24.s\n" + "fmax z13.s, p5/M, z13.s, z24.s\n" + "fmax z14.s, p5/M, z14.s, z24.s\n" + "fmax z15.s, p5/M, z15.s, z24.s\n" + "fmax z16.s, p5/M, z16.s, z24.s\n" + "fmax z17.s, p5/M, z17.s, z24.s\n" + "fmax z18.s, p5/M, z18.s, z24.s\n" + "fmax z19.s, p5/M, z19.s, z24.s\n" + "fmax z20.s, p5/M, z20.s, z24.s\n" + "fmax z21.s, p5/M, z21.s, z24.s\n" + "fmax z22.s, p5/M, z22.s, z24.s\n" + "fmax z23.s, p5/M, z23.s, z24.s\n" "51:" // Height 4: No activation "st1w { z8.s }, p4, [x9]\n" "st1w { z9.s }, p3, [x9, #1, MUL VL]\n" @@ -1195,30 +1195,30 @@ void sve_hybrid_fp32_mla_6x4VL ( "55:" // Height 5: no bias "tbz %x[flags], #0, 56f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z8.s }, p4/Z, [x9]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x9, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" + "ld1w { z8.s }, p4/Z, [x9]\n" + "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x24]\n" - "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z24.s }, p4/Z, [x22]\n" - "ld1w { z25.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z26.s }, p2/Z, [x22, #2, MUL VL]\n" - "ld1w { z27.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x23]\n" + "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x22]\n" + "ld1w { z17.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x21]\n" + "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z24.s }, p4/Z, [x20]\n" + "ld1w { z25.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z26.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z27.s }, p1/Z, [x20, #3, MUL VL]\n" "b 57f\n" "56:" // Height 5: no accumulate "mov z8.b, #0x0\n" @@ -1246,15 +1246,15 @@ void sve_hybrid_fp32_mla_6x4VL ( "58:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 59f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" "cbnz x28, 60f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1265,124 +1265,124 @@ void sve_hybrid_fp32_mla_6x4VL ( "b 60f\n" "59:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "60:" // Height 5: input setup done "cmp x27, #0x4\n" "ble 62f\n" "61:" // Height 5: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" + "ld1rqw { z4.s }, p0/Z, [x26]\n" + "ld1rqw { z3.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" + "ld1rqw { z1.s }, p0/Z, [x23]\n" "cmp x27, #0x4\n" "add x26, x26, #0x10\n" - "ld1rqw { z4.s }, p0/Z, [x22]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" + "ld1rqw { z0.s }, p0/Z, [x22]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" + "fmla z8.s, z29.s, z4.s[0]\n" + "fmla z12.s, z29.s, z3.s[0]\n" + "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z16.s, z29.s, z2.s[0]\n" + "fmla z20.s, z29.s, z1.s[0]\n" "add x25, x25, #0x10\n" - "fmla z24.s, z6.s, z4.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z24.s, z29.s, z0.s[0]\n" + "fmla z9.s, z28.s, z4.s[0]\n" + "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n" "add x24, x24, #0x10\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" + "fmla z13.s, z28.s, z3.s[0]\n" + "fmla z17.s, z28.s, z2.s[0]\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "fmla z25.s, z7.s, z4.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z26.s, z6.s, z4.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #4, MUL VL]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "fmla z27.s, z7.s, z4.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #5, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" - "fmla z24.s, z6.s, z4.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #6, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "fmla z25.s, z7.s, z4.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #7, MUL VL]\n" + "fmla z21.s, z28.s, z1.s[0]\n" + "fmla z25.s, z28.s, z0.s[0]\n" + "ld1w { z28.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z10.s, z29.s, z4.s[0]\n" + "fmla z14.s, z29.s, z3.s[0]\n" + "fmla z18.s, z29.s, z2.s[0]\n" + "fmla z22.s, z29.s, z1.s[0]\n" + "fmla z26.s, z29.s, z0.s[0]\n" + "fmla z11.s, z28.s, z4.s[0]\n" + "ld1w { z29.s }, p5/Z, [x10, #4, MUL VL]\n" + "fmla z15.s, z28.s, z3.s[0]\n" + "fmla z19.s, z28.s, z2.s[0]\n" + "fmla z23.s, z28.s, z1.s[0]\n" + "fmla z27.s, z28.s, z0.s[0]\n" + "ld1w { z28.s }, p5/Z, [x10, #5, MUL VL]\n" + "fmla z8.s, z29.s, z4.s[1]\n" + "fmla z12.s, z29.s, z3.s[1]\n" + "fmla z16.s, z29.s, z2.s[1]\n" + "fmla z20.s, z29.s, z1.s[1]\n" + "fmla z24.s, z29.s, z0.s[1]\n" + "fmla z9.s, z28.s, z4.s[1]\n" + "ld1w { z29.s }, p5/Z, [x10, #6, MUL VL]\n" + "fmla z13.s, z28.s, z3.s[1]\n" + "fmla z17.s, z28.s, z2.s[1]\n" + "fmla z21.s, z28.s, z1.s[1]\n" + "fmla z25.s, z28.s, z0.s[1]\n" + "ld1w { z28.s }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z26.s, z6.s, z4.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #-8, MUL VL]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "fmla z27.s, z7.s, z4.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #-7, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" - "fmla z24.s, z6.s, z4.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-6, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "fmla z25.s, z7.s, z4.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-5, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z26.s, z6.s, z4.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-4, MUL VL]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "fmla z27.s, z7.s, z4.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-3, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "fmla z24.s, z6.s, z4.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #-2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "fmla z25.s, z7.s, z4.s[3]\n" - "ld1w { z7.s }, p5/Z, [x10, #-1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z26.s, z6.s, z4.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" - "fmla z27.s, z7.s, z4.s[3]\n" + "fmla z10.s, z29.s, z4.s[1]\n" + "fmla z14.s, z29.s, z3.s[1]\n" + "fmla z18.s, z29.s, z2.s[1]\n" + "fmla z22.s, z29.s, z1.s[1]\n" + "fmla z26.s, z29.s, z0.s[1]\n" + "fmla z11.s, z28.s, z4.s[1]\n" + "ld1w { z29.s }, p5/Z, [x10, #-8, MUL VL]\n" + "fmla z15.s, z28.s, z3.s[1]\n" + "fmla z19.s, z28.s, z2.s[1]\n" + "fmla z23.s, z28.s, z1.s[1]\n" + "fmla z27.s, z28.s, z0.s[1]\n" + "ld1w { z28.s }, p5/Z, [x10, #-7, MUL VL]\n" + "fmla z8.s, z29.s, z4.s[2]\n" + "fmla z12.s, z29.s, z3.s[2]\n" + "fmla z16.s, z29.s, z2.s[2]\n" + "fmla z20.s, z29.s, z1.s[2]\n" + "fmla z24.s, z29.s, z0.s[2]\n" + "fmla z9.s, z28.s, z4.s[2]\n" + "ld1w { z29.s }, p5/Z, [x10, #-6, MUL VL]\n" + "fmla z13.s, z28.s, z3.s[2]\n" + "fmla z17.s, z28.s, z2.s[2]\n" + "fmla z21.s, z28.s, z1.s[2]\n" + "fmla z25.s, z28.s, z0.s[2]\n" + "ld1w { z28.s }, p5/Z, [x10, #-5, MUL VL]\n" + "fmla z10.s, z29.s, z4.s[2]\n" + "fmla z14.s, z29.s, z3.s[2]\n" + "fmla z18.s, z29.s, z2.s[2]\n" + "fmla z22.s, z29.s, z1.s[2]\n" + "fmla z26.s, z29.s, z0.s[2]\n" + "fmla z11.s, z28.s, z4.s[2]\n" + "ld1w { z29.s }, p5/Z, [x10, #-4, MUL VL]\n" + "fmla z15.s, z28.s, z3.s[2]\n" + "fmla z19.s, z28.s, z2.s[2]\n" + "fmla z23.s, z28.s, z1.s[2]\n" + "fmla z27.s, z28.s, z0.s[2]\n" + "ld1w { z28.s }, p5/Z, [x10, #-3, MUL VL]\n" + "fmla z8.s, z29.s, z4.s[3]\n" + "fmla z12.s, z29.s, z3.s[3]\n" + "fmla z16.s, z29.s, z2.s[3]\n" + "fmla z20.s, z29.s, z1.s[3]\n" + "fmla z24.s, z29.s, z0.s[3]\n" + "fmla z9.s, z28.s, z4.s[3]\n" + "ld1w { z29.s }, p5/Z, [x10, #-2, MUL VL]\n" + "fmla z13.s, z28.s, z3.s[3]\n" + "fmla z17.s, z28.s, z2.s[3]\n" + "fmla z21.s, z28.s, z1.s[3]\n" + "fmla z25.s, z28.s, z0.s[3]\n" + "ld1w { z28.s }, p5/Z, [x10, #-1, MUL VL]\n" + "fmla z10.s, z29.s, z4.s[3]\n" + "fmla z14.s, z29.s, z3.s[3]\n" + "fmla z18.s, z29.s, z2.s[3]\n" + "fmla z22.s, z29.s, z1.s[3]\n" + "fmla z26.s, z29.s, z0.s[3]\n" + "fmla z11.s, z28.s, z4.s[3]\n" + "fmla z15.s, z28.s, z3.s[3]\n" + "fmla z19.s, z28.s, z2.s[3]\n" + "fmla z23.s, z28.s, z1.s[3]\n" + "fmla z27.s, z28.s, z0.s[3]\n" "bgt 61b\n" "62:" // Height 5: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" @@ -1392,111 +1392,111 @@ void sve_hybrid_fp32_mla_6x4VL ( "ld1rqw { z2.s }, p0/Z, [x24]\n" "ld1rqw { z3.s }, p0/Z, [x23]\n" "ld1rqw { z4.s }, p0/Z, [x22]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" - "fmla z24.s, z6.s, z4.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "fmla z25.s, z7.s, z4.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" + "fmla z8.s, z29.s, z0.s[0]\n" + "fmla z12.s, z29.s, z1.s[0]\n" + "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z16.s, z29.s, z2.s[0]\n" + "fmla z20.s, z29.s, z3.s[0]\n" + "fmla z24.s, z29.s, z4.s[0]\n" + "fmla z9.s, z28.s, z0.s[0]\n" + "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z13.s, z28.s, z1.s[0]\n" + "fmla z17.s, z28.s, z2.s[0]\n" + "fmla z21.s, z28.s, z3.s[0]\n" + "fmla z25.s, z28.s, z4.s[0]\n" + "ld1w { z28.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z26.s, z6.s, z4.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "fmla z27.s, z7.s, z4.s[0]\n" + "fmla z10.s, z29.s, z0.s[0]\n" + "fmla z14.s, z29.s, z1.s[0]\n" + "fmla z18.s, z29.s, z2.s[0]\n" + "fmla z22.s, z29.s, z3.s[0]\n" + "fmla z26.s, z29.s, z4.s[0]\n" + "fmla z11.s, z28.s, z0.s[0]\n" + "fmla z15.s, z28.s, z1.s[0]\n" + "fmla z19.s, z28.s, z2.s[0]\n" + "fmla z23.s, z28.s, z3.s[0]\n" + "fmla z27.s, z28.s, z4.s[0]\n" "ble 63f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" + "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z29.s, z0.s[1]\n" + "fmla z12.s, z29.s, z1.s[1]\n" + "fmla z16.s, z29.s, z2.s[1]\n" + "fmla z20.s, z29.s, z3.s[1]\n" "subs x27, x27, #0x1\n" - "fmla z24.s, z6.s, z4.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "fmla z25.s, z7.s, z4.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z24.s, z29.s, z4.s[1]\n" + "fmla z9.s, z28.s, z0.s[1]\n" + "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z13.s, z28.s, z1.s[1]\n" + "fmla z17.s, z28.s, z2.s[1]\n" + "fmla z21.s, z28.s, z3.s[1]\n" + "fmla z25.s, z28.s, z4.s[1]\n" + "ld1w { z28.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z26.s, z6.s, z4.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "fmla z27.s, z7.s, z4.s[1]\n" + "fmla z10.s, z29.s, z0.s[1]\n" + "fmla z14.s, z29.s, z1.s[1]\n" + "fmla z18.s, z29.s, z2.s[1]\n" + "fmla z22.s, z29.s, z3.s[1]\n" + "fmla z26.s, z29.s, z4.s[1]\n" + "fmla z11.s, z28.s, z0.s[1]\n" + "fmla z15.s, z28.s, z1.s[1]\n" + "fmla z19.s, z28.s, z2.s[1]\n" + "fmla z23.s, z28.s, z3.s[1]\n" + "fmla z27.s, z28.s, z4.s[1]\n" "ble 63f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" + "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z29.s, z0.s[2]\n" + "fmla z12.s, z29.s, z1.s[2]\n" + "fmla z16.s, z29.s, z2.s[2]\n" + "fmla z20.s, z29.s, z3.s[2]\n" "subs x27, x27, #0x1\n" - "fmla z24.s, z6.s, z4.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "fmla z25.s, z7.s, z4.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z24.s, z29.s, z4.s[2]\n" + "fmla z9.s, z28.s, z0.s[2]\n" + "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z13.s, z28.s, z1.s[2]\n" + "fmla z17.s, z28.s, z2.s[2]\n" + "fmla z21.s, z28.s, z3.s[2]\n" + "fmla z25.s, z28.s, z4.s[2]\n" + "ld1w { z28.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z26.s, z6.s, z4.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "fmla z27.s, z7.s, z4.s[2]\n" + "fmla z10.s, z29.s, z0.s[2]\n" + "fmla z14.s, z29.s, z1.s[2]\n" + "fmla z18.s, z29.s, z2.s[2]\n" + "fmla z22.s, z29.s, z3.s[2]\n" + "fmla z26.s, z29.s, z4.s[2]\n" + "fmla z11.s, z28.s, z0.s[2]\n" + "fmla z15.s, z28.s, z1.s[2]\n" + "fmla z19.s, z28.s, z2.s[2]\n" + "fmla z23.s, z28.s, z3.s[2]\n" + "fmla z27.s, z28.s, z4.s[2]\n" "ble 63f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "fmla z24.s, z6.s, z4.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "fmla z25.s, z7.s, z4.s[3]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" + "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z29.s, z0.s[3]\n" + "fmla z12.s, z29.s, z1.s[3]\n" + "fmla z16.s, z29.s, z2.s[3]\n" + "fmla z20.s, z29.s, z3.s[3]\n" + "fmla z24.s, z29.s, z4.s[3]\n" + "fmla z9.s, z28.s, z0.s[3]\n" + "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z13.s, z28.s, z1.s[3]\n" + "fmla z17.s, z28.s, z2.s[3]\n" + "fmla z21.s, z28.s, z3.s[3]\n" + "fmla z25.s, z28.s, z4.s[3]\n" + "ld1w { z28.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z26.s, z6.s, z4.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" - "fmla z27.s, z7.s, z4.s[3]\n" + "fmla z10.s, z29.s, z0.s[3]\n" + "fmla z14.s, z29.s, z1.s[3]\n" + "fmla z18.s, z29.s, z2.s[3]\n" + "fmla z22.s, z29.s, z3.s[3]\n" + "fmla z26.s, z29.s, z4.s[3]\n" + "fmla z11.s, z28.s, z0.s[3]\n" + "fmla z15.s, z28.s, z1.s[3]\n" + "fmla z19.s, z28.s, z2.s[3]\n" + "fmla z23.s, z28.s, z3.s[3]\n" + "fmla z27.s, z28.s, z4.s[3]\n" "63:" // Height 5: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1509,49 +1509,49 @@ void sve_hybrid_fp32_mla_6x4VL ( "add x22, x23, x20, LSL #2\n" "tbz %x[flags], #1, 64f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z29.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmin z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z1.s\n" - "fmin z18.s, p5/M, z18.s, z1.s\n" - "fmin z19.s, p5/M, z19.s, z1.s\n" - "fmin z20.s, p5/M, z20.s, z1.s\n" - "fmin z21.s, p5/M, z21.s, z1.s\n" - "fmin z22.s, p5/M, z22.s, z1.s\n" - "fmin z23.s, p5/M, z23.s, z1.s\n" - "fmin z24.s, p5/M, z24.s, z1.s\n" - "fmin z25.s, p5/M, z25.s, z1.s\n" - "fmin z26.s, p5/M, z26.s, z1.s\n" - "fmin z27.s, p5/M, z27.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" - "fmax z16.s, p5/M, z16.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z0.s\n" - "fmax z18.s, p5/M, z18.s, z0.s\n" - "fmax z19.s, p5/M, z19.s, z0.s\n" - "fmax z20.s, p5/M, z20.s, z0.s\n" - "fmax z21.s, p5/M, z21.s, z0.s\n" - "fmax z22.s, p5/M, z22.s, z0.s\n" - "fmax z23.s, p5/M, z23.s, z0.s\n" - "fmax z24.s, p5/M, z24.s, z0.s\n" - "fmax z25.s, p5/M, z25.s, z0.s\n" - "fmax z26.s, p5/M, z26.s, z0.s\n" - "fmax z27.s, p5/M, z27.s, z0.s\n" + "ld1rw { z28.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z29.s\n" + "fmin z9.s, p5/M, z9.s, z29.s\n" + "fmin z10.s, p5/M, z10.s, z29.s\n" + "fmin z11.s, p5/M, z11.s, z29.s\n" + "fmin z12.s, p5/M, z12.s, z29.s\n" + "fmin z13.s, p5/M, z13.s, z29.s\n" + "fmin z14.s, p5/M, z14.s, z29.s\n" + "fmin z15.s, p5/M, z15.s, z29.s\n" + "fmin z16.s, p5/M, z16.s, z29.s\n" + "fmin z17.s, p5/M, z17.s, z29.s\n" + "fmin z18.s, p5/M, z18.s, z29.s\n" + "fmin z19.s, p5/M, z19.s, z29.s\n" + "fmin z20.s, p5/M, z20.s, z29.s\n" + "fmin z21.s, p5/M, z21.s, z29.s\n" + "fmin z22.s, p5/M, z22.s, z29.s\n" + "fmin z23.s, p5/M, z23.s, z29.s\n" + "fmin z24.s, p5/M, z24.s, z29.s\n" + "fmin z25.s, p5/M, z25.s, z29.s\n" + "fmin z26.s, p5/M, z26.s, z29.s\n" + "fmin z27.s, p5/M, z27.s, z29.s\n" + "fmax z8.s, p5/M, z8.s, z28.s\n" + "fmax z9.s, p5/M, z9.s, z28.s\n" + "fmax z10.s, p5/M, z10.s, z28.s\n" + "fmax z11.s, p5/M, z11.s, z28.s\n" + "fmax z12.s, p5/M, z12.s, z28.s\n" + "fmax z13.s, p5/M, z13.s, z28.s\n" + "fmax z14.s, p5/M, z14.s, z28.s\n" + "fmax z15.s, p5/M, z15.s, z28.s\n" + "fmax z16.s, p5/M, z16.s, z28.s\n" + "fmax z17.s, p5/M, z17.s, z28.s\n" + "fmax z18.s, p5/M, z18.s, z28.s\n" + "fmax z19.s, p5/M, z19.s, z28.s\n" + "fmax z20.s, p5/M, z20.s, z28.s\n" + "fmax z21.s, p5/M, z21.s, z28.s\n" + "fmax z22.s, p5/M, z22.s, z28.s\n" + "fmax z23.s, p5/M, z23.s, z28.s\n" + "fmax z24.s, p5/M, z24.s, z28.s\n" + "fmax z25.s, p5/M, z25.s, z28.s\n" + "fmax z26.s, p5/M, z26.s, z28.s\n" + "fmax z27.s, p5/M, z27.s, z28.s\n" "64:" // Height 5: No activation "st1w { z8.s }, p4, [x9]\n" "st1w { z9.s }, p3, [x9, #1, MUL VL]\n" @@ -1626,35 +1626,35 @@ void sve_hybrid_fp32_mla_6x4VL ( "68:" // Height 6: no bias "tbz %x[flags], #0, 69f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z8.s }, p4/Z, [x9]\n" + "add x24, x9, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" + "ld1w { z8.s }, p4/Z, [x9]\n" "add x22, x23, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n" - "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x24]\n" - "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z24.s }, p4/Z, [x22]\n" - "ld1w { z25.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z26.s }, p2/Z, [x22, #2, MUL VL]\n" - "ld1w { z27.s }, p1/Z, [x22, #3, MUL VL]\n" - "ld1w { z28.s }, p4/Z, [x21]\n" - "ld1w { z29.s }, p3/Z, [x21, #1, MUL VL]\n" - "ld1w { z30.s }, p2/Z, [x21, #2, MUL VL]\n" - "ld1w { z31.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x24]\n" + "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x23]\n" + "ld1w { z17.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x22]\n" + "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z24.s }, p4/Z, [x21]\n" + "ld1w { z25.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z26.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z27.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z28.s }, p4/Z, [x20]\n" + "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n" "b 70f\n" "69:" // Height 6: no accumulate "mov z8.b, #0x0\n" @@ -1686,16 +1686,16 @@ void sve_hybrid_fp32_mla_6x4VL ( "71:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 72f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "ldr x21, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "ldr x21, [x20, #0x28]\n" "cbnz x28, 73f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1707,143 +1707,143 @@ void sve_hybrid_fp32_mla_6x4VL ( "b 73f\n" "72:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "73:" // Height 6: input setup done "cmp x27, #0x4\n" "ble 75f\n" "74:" // Height 6: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" + "ld1rqw { z7.s }, p0/Z, [x26]\n" + "ld1rqw { z6.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" + "ld1rqw { z5.s }, p0/Z, [x24]\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" "cmp x27, #0x4\n" "add x26, x26, #0x10\n" - "ld1rqw { z4.s }, p0/Z, [x22]\n" - "ld1rqw { z5.s }, p0/Z, [x21]\n" + "ld1rqw { z3.s }, p0/Z, [x22]\n" + "ld1rqw { z2.s }, p0/Z, [x21]\n" "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" + "ld1w { z1.s }, p5/Z, [x10]\n" + "ld1w { z0.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z1.s, z7.s[0]\n" + "fmla z12.s, z1.s, z6.s[0]\n" + "fmla z16.s, z1.s, z5.s[0]\n" + "fmla z20.s, z1.s, z4.s[0]\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - "fmla z24.s, z6.s, z4.s[0]\n" - "fmla z28.s, z6.s, z5.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z24.s, z1.s, z3.s[0]\n" + "fmla z28.s, z1.s, z2.s[0]\n" + "ld1w { z1.s }, p5/Z, [x10, #2, MUL VL]\n" "add x21, x21, #0x10\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "fmla z25.s, z7.s, z4.s[0]\n" - "fmla z29.s, z7.s, z5.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z26.s, z6.s, z4.s[0]\n" - "fmla z30.s, z6.s, z5.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #4, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "fmla z27.s, z7.s, z4.s[0]\n" - "fmla z31.s, z7.s, z5.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #5, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" - "fmla z24.s, z6.s, z4.s[1]\n" - "fmla z28.s, z6.s, z5.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #6, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "fmla z25.s, z7.s, z4.s[1]\n" - "fmla z29.s, z7.s, z5.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #7, MUL VL]\n" + "fmla z9.s, z0.s, z7.s[0]\n" + "fmla z13.s, z0.s, z6.s[0]\n" + "fmla z17.s, z0.s, z5.s[0]\n" + "fmla z21.s, z0.s, z4.s[0]\n" + "fmla z25.s, z0.s, z3.s[0]\n" + "fmla z29.s, z0.s, z2.s[0]\n" + "ld1w { z0.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z10.s, z1.s, z7.s[0]\n" + "fmla z14.s, z1.s, z6.s[0]\n" + "fmla z18.s, z1.s, z5.s[0]\n" + "fmla z22.s, z1.s, z4.s[0]\n" + "fmla z26.s, z1.s, z3.s[0]\n" + "fmla z30.s, z1.s, z2.s[0]\n" + "ld1w { z1.s }, p5/Z, [x10, #4, MUL VL]\n" + "fmla z11.s, z0.s, z7.s[0]\n" + "fmla z15.s, z0.s, z6.s[0]\n" + "fmla z19.s, z0.s, z5.s[0]\n" + "fmla z23.s, z0.s, z4.s[0]\n" + "fmla z27.s, z0.s, z3.s[0]\n" + "fmla z31.s, z0.s, z2.s[0]\n" + "ld1w { z0.s }, p5/Z, [x10, #5, MUL VL]\n" + "fmla z8.s, z1.s, z7.s[1]\n" + "fmla z12.s, z1.s, z6.s[1]\n" + "fmla z16.s, z1.s, z5.s[1]\n" + "fmla z20.s, z1.s, z4.s[1]\n" + "fmla z24.s, z1.s, z3.s[1]\n" + "fmla z28.s, z1.s, z2.s[1]\n" + "ld1w { z1.s }, p5/Z, [x10, #6, MUL VL]\n" + "fmla z9.s, z0.s, z7.s[1]\n" + "fmla z13.s, z0.s, z6.s[1]\n" + "fmla z17.s, z0.s, z5.s[1]\n" + "fmla z21.s, z0.s, z4.s[1]\n" + "fmla z25.s, z0.s, z3.s[1]\n" + "fmla z29.s, z0.s, z2.s[1]\n" + "ld1w { z0.s }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z26.s, z6.s, z4.s[1]\n" - "fmla z30.s, z6.s, z5.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #-8, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "fmla z27.s, z7.s, z4.s[1]\n" - "fmla z31.s, z7.s, z5.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #-7, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" - "fmla z24.s, z6.s, z4.s[2]\n" - "fmla z28.s, z6.s, z5.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-6, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "fmla z25.s, z7.s, z4.s[2]\n" - "fmla z29.s, z7.s, z5.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-5, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z26.s, z6.s, z4.s[2]\n" - "fmla z30.s, z6.s, z5.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #-4, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "fmla z27.s, z7.s, z4.s[2]\n" - "fmla z31.s, z7.s, z5.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #-3, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "fmla z24.s, z6.s, z4.s[3]\n" - "fmla z28.s, z6.s, z5.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #-2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "fmla z25.s, z7.s, z4.s[3]\n" - "fmla z29.s, z7.s, z5.s[3]\n" - "ld1w { z7.s }, p5/Z, [x10, #-1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z26.s, z6.s, z4.s[3]\n" - "fmla z30.s, z6.s, z5.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" - "fmla z27.s, z7.s, z4.s[3]\n" - "fmla z31.s, z7.s, z5.s[3]\n" + "fmla z10.s, z1.s, z7.s[1]\n" + "fmla z14.s, z1.s, z6.s[1]\n" + "fmla z18.s, z1.s, z5.s[1]\n" + "fmla z22.s, z1.s, z4.s[1]\n" + "fmla z26.s, z1.s, z3.s[1]\n" + "fmla z30.s, z1.s, z2.s[1]\n" + "ld1w { z1.s }, p5/Z, [x10, #-8, MUL VL]\n" + "fmla z11.s, z0.s, z7.s[1]\n" + "fmla z15.s, z0.s, z6.s[1]\n" + "fmla z19.s, z0.s, z5.s[1]\n" + "fmla z23.s, z0.s, z4.s[1]\n" + "fmla z27.s, z0.s, z3.s[1]\n" + "fmla z31.s, z0.s, z2.s[1]\n" + "ld1w { z0.s }, p5/Z, [x10, #-7, MUL VL]\n" + "fmla z8.s, z1.s, z7.s[2]\n" + "fmla z12.s, z1.s, z6.s[2]\n" + "fmla z16.s, z1.s, z5.s[2]\n" + "fmla z20.s, z1.s, z4.s[2]\n" + "fmla z24.s, z1.s, z3.s[2]\n" + "fmla z28.s, z1.s, z2.s[2]\n" + "ld1w { z1.s }, p5/Z, [x10, #-6, MUL VL]\n" + "fmla z9.s, z0.s, z7.s[2]\n" + "fmla z13.s, z0.s, z6.s[2]\n" + "fmla z17.s, z0.s, z5.s[2]\n" + "fmla z21.s, z0.s, z4.s[2]\n" + "fmla z25.s, z0.s, z3.s[2]\n" + "fmla z29.s, z0.s, z2.s[2]\n" + "ld1w { z0.s }, p5/Z, [x10, #-5, MUL VL]\n" + "fmla z10.s, z1.s, z7.s[2]\n" + "fmla z14.s, z1.s, z6.s[2]\n" + "fmla z18.s, z1.s, z5.s[2]\n" + "fmla z22.s, z1.s, z4.s[2]\n" + "fmla z26.s, z1.s, z3.s[2]\n" + "fmla z30.s, z1.s, z2.s[2]\n" + "ld1w { z1.s }, p5/Z, [x10, #-4, MUL VL]\n" + "fmla z11.s, z0.s, z7.s[2]\n" + "fmla z15.s, z0.s, z6.s[2]\n" + "fmla z19.s, z0.s, z5.s[2]\n" + "fmla z23.s, z0.s, z4.s[2]\n" + "fmla z27.s, z0.s, z3.s[2]\n" + "fmla z31.s, z0.s, z2.s[2]\n" + "ld1w { z0.s }, p5/Z, [x10, #-3, MUL VL]\n" + "fmla z8.s, z1.s, z7.s[3]\n" + "fmla z12.s, z1.s, z6.s[3]\n" + "fmla z16.s, z1.s, z5.s[3]\n" + "fmla z20.s, z1.s, z4.s[3]\n" + "fmla z24.s, z1.s, z3.s[3]\n" + "fmla z28.s, z1.s, z2.s[3]\n" + "ld1w { z1.s }, p5/Z, [x10, #-2, MUL VL]\n" + "fmla z9.s, z0.s, z7.s[3]\n" + "fmla z13.s, z0.s, z6.s[3]\n" + "fmla z17.s, z0.s, z5.s[3]\n" + "fmla z21.s, z0.s, z4.s[3]\n" + "fmla z25.s, z0.s, z3.s[3]\n" + "fmla z29.s, z0.s, z2.s[3]\n" + "ld1w { z0.s }, p5/Z, [x10, #-1, MUL VL]\n" + "fmla z10.s, z1.s, z7.s[3]\n" + "fmla z14.s, z1.s, z6.s[3]\n" + "fmla z18.s, z1.s, z5.s[3]\n" + "fmla z22.s, z1.s, z4.s[3]\n" + "fmla z26.s, z1.s, z3.s[3]\n" + "fmla z30.s, z1.s, z2.s[3]\n" + "fmla z11.s, z0.s, z7.s[3]\n" + "fmla z15.s, z0.s, z6.s[3]\n" + "fmla z19.s, z0.s, z5.s[3]\n" + "fmla z23.s, z0.s, z4.s[3]\n" + "fmla z27.s, z0.s, z3.s[3]\n" + "fmla z31.s, z0.s, z2.s[3]\n" "bgt 74b\n" "75:" // Height 6: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" @@ -1854,127 +1854,127 @@ void sve_hybrid_fp32_mla_6x4VL ( "ld1rqw { z3.s }, p0/Z, [x23]\n" "ld1rqw { z4.s }, p0/Z, [x22]\n" "ld1rqw { z5.s }, p0/Z, [x21]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" - "fmla z24.s, z6.s, z4.s[0]\n" - "fmla z28.s, z6.s, z5.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "fmla z25.s, z7.s, z4.s[0]\n" - "fmla z29.s, z7.s, z5.s[0]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z7.s }, p5/Z, [x10]\n" + "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z7.s, z0.s[0]\n" + "fmla z12.s, z7.s, z1.s[0]\n" + "fmla z16.s, z7.s, z2.s[0]\n" + "fmla z20.s, z7.s, z3.s[0]\n" + "fmla z24.s, z7.s, z4.s[0]\n" + "fmla z28.s, z7.s, z5.s[0]\n" + "ld1w { z7.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z6.s, z0.s[0]\n" + "fmla z13.s, z6.s, z1.s[0]\n" + "fmla z17.s, z6.s, z2.s[0]\n" + "fmla z21.s, z6.s, z3.s[0]\n" + "fmla z25.s, z6.s, z4.s[0]\n" + "fmla z29.s, z6.s, z5.s[0]\n" + "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z26.s, z6.s, z4.s[0]\n" - "fmla z30.s, z6.s, z5.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "fmla z27.s, z7.s, z4.s[0]\n" - "fmla z31.s, z7.s, z5.s[0]\n" + "fmla z10.s, z7.s, z0.s[0]\n" + "fmla z14.s, z7.s, z1.s[0]\n" + "fmla z18.s, z7.s, z2.s[0]\n" + "fmla z22.s, z7.s, z3.s[0]\n" + "fmla z26.s, z7.s, z4.s[0]\n" + "fmla z30.s, z7.s, z5.s[0]\n" + "fmla z11.s, z6.s, z0.s[0]\n" + "fmla z15.s, z6.s, z1.s[0]\n" + "fmla z19.s, z6.s, z2.s[0]\n" + "fmla z23.s, z6.s, z3.s[0]\n" + "fmla z27.s, z6.s, z4.s[0]\n" + "fmla z31.s, z6.s, z5.s[0]\n" "ble 76f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" + "ld1w { z7.s }, p5/Z, [x10]\n" + "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z7.s, z0.s[1]\n" + "fmla z12.s, z7.s, z1.s[1]\n" + "fmla z16.s, z7.s, z2.s[1]\n" + "fmla z20.s, z7.s, z3.s[1]\n" "subs x27, x27, #0x1\n" - "fmla z24.s, z6.s, z4.s[1]\n" - "fmla z28.s, z6.s, z5.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "fmla z25.s, z7.s, z4.s[1]\n" - "fmla z29.s, z7.s, z5.s[1]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z24.s, z7.s, z4.s[1]\n" + "fmla z28.s, z7.s, z5.s[1]\n" + "ld1w { z7.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z6.s, z0.s[1]\n" + "fmla z13.s, z6.s, z1.s[1]\n" + "fmla z17.s, z6.s, z2.s[1]\n" + "fmla z21.s, z6.s, z3.s[1]\n" + "fmla z25.s, z6.s, z4.s[1]\n" + "fmla z29.s, z6.s, z5.s[1]\n" + "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z26.s, z6.s, z4.s[1]\n" - "fmla z30.s, z6.s, z5.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "fmla z27.s, z7.s, z4.s[1]\n" - "fmla z31.s, z7.s, z5.s[1]\n" + "fmla z10.s, z7.s, z0.s[1]\n" + "fmla z14.s, z7.s, z1.s[1]\n" + "fmla z18.s, z7.s, z2.s[1]\n" + "fmla z22.s, z7.s, z3.s[1]\n" + "fmla z26.s, z7.s, z4.s[1]\n" + "fmla z30.s, z7.s, z5.s[1]\n" + "fmla z11.s, z6.s, z0.s[1]\n" + "fmla z15.s, z6.s, z1.s[1]\n" + "fmla z19.s, z6.s, z2.s[1]\n" + "fmla z23.s, z6.s, z3.s[1]\n" + "fmla z27.s, z6.s, z4.s[1]\n" + "fmla z31.s, z6.s, z5.s[1]\n" "ble 76f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" + "ld1w { z7.s }, p5/Z, [x10]\n" + "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z7.s, z0.s[2]\n" + "fmla z12.s, z7.s, z1.s[2]\n" + "fmla z16.s, z7.s, z2.s[2]\n" + "fmla z20.s, z7.s, z3.s[2]\n" "subs x27, x27, #0x1\n" - "fmla z24.s, z6.s, z4.s[2]\n" - "fmla z28.s, z6.s, z5.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "fmla z25.s, z7.s, z4.s[2]\n" - "fmla z29.s, z7.s, z5.s[2]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z24.s, z7.s, z4.s[2]\n" + "fmla z28.s, z7.s, z5.s[2]\n" + "ld1w { z7.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z6.s, z0.s[2]\n" + "fmla z13.s, z6.s, z1.s[2]\n" + "fmla z17.s, z6.s, z2.s[2]\n" + "fmla z21.s, z6.s, z3.s[2]\n" + "fmla z25.s, z6.s, z4.s[2]\n" + "fmla z29.s, z6.s, z5.s[2]\n" + "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z26.s, z6.s, z4.s[2]\n" - "fmla z30.s, z6.s, z5.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "fmla z27.s, z7.s, z4.s[2]\n" - "fmla z31.s, z7.s, z5.s[2]\n" + "fmla z10.s, z7.s, z0.s[2]\n" + "fmla z14.s, z7.s, z1.s[2]\n" + "fmla z18.s, z7.s, z2.s[2]\n" + "fmla z22.s, z7.s, z3.s[2]\n" + "fmla z26.s, z7.s, z4.s[2]\n" + "fmla z30.s, z7.s, z5.s[2]\n" + "fmla z11.s, z6.s, z0.s[2]\n" + "fmla z15.s, z6.s, z1.s[2]\n" + "fmla z19.s, z6.s, z2.s[2]\n" + "fmla z23.s, z6.s, z3.s[2]\n" + "fmla z27.s, z6.s, z4.s[2]\n" + "fmla z31.s, z6.s, z5.s[2]\n" "ble 76f\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "fmla z24.s, z6.s, z4.s[3]\n" - "fmla z28.s, z6.s, z5.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "fmla z25.s, z7.s, z4.s[3]\n" - "fmla z29.s, z7.s, z5.s[3]\n" - "ld1w { z7.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z7.s }, p5/Z, [x10]\n" + "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z7.s, z0.s[3]\n" + "fmla z12.s, z7.s, z1.s[3]\n" + "fmla z16.s, z7.s, z2.s[3]\n" + "fmla z20.s, z7.s, z3.s[3]\n" + "fmla z24.s, z7.s, z4.s[3]\n" + "fmla z28.s, z7.s, z5.s[3]\n" + "ld1w { z7.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z6.s, z0.s[3]\n" + "fmla z13.s, z6.s, z1.s[3]\n" + "fmla z17.s, z6.s, z2.s[3]\n" + "fmla z21.s, z6.s, z3.s[3]\n" + "fmla z25.s, z6.s, z4.s[3]\n" + "fmla z29.s, z6.s, z5.s[3]\n" + "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z26.s, z6.s, z4.s[3]\n" - "fmla z30.s, z6.s, z5.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" - "fmla z27.s, z7.s, z4.s[3]\n" - "fmla z31.s, z7.s, z5.s[3]\n" + "fmla z10.s, z7.s, z0.s[3]\n" + "fmla z14.s, z7.s, z1.s[3]\n" + "fmla z18.s, z7.s, z2.s[3]\n" + "fmla z22.s, z7.s, z3.s[3]\n" + "fmla z26.s, z7.s, z4.s[3]\n" + "fmla z30.s, z7.s, z5.s[3]\n" + "fmla z11.s, z6.s, z0.s[3]\n" + "fmla z15.s, z6.s, z1.s[3]\n" + "fmla z19.s, z6.s, z2.s[3]\n" + "fmla z23.s, z6.s, z3.s[3]\n" + "fmla z27.s, z6.s, z4.s[3]\n" + "fmla z31.s, z6.s, z5.s[3]\n" "76:" // Height 6: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -2081,7 +2081,6 @@ void sve_hybrid_fp32_mla_6x4VL ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "80:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, 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_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "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" @@ -2089,4 +2088,4 @@ void sve_hybrid_fp32_mla_6x4VL ( } } // namespace arm_gemm -#endif // ARM_COMPUTE_ENABLE_SVE +#endif // ARM_COMPUTE_ENABLE_SVE -- cgit v1.2.1