diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp | 1934 |
1 files changed, 967 insertions, 967 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp index 13f2e488dd..9136e32567 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp @@ -157,16 +157,16 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "b 6f\n" "4:" // Height 1: no bias "tbz %x[flags], #0, 5f\n" - "ld1w { z9.s }, p4/Z, [x13]\n" - "ld1w { z10.s }, p3/Z, [x13, #1, MUL VL]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z11.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x13]\n" + "ld1w { z18.s }, p3/Z, [x13, #1, MUL VL]\n" + "zip1 z8.d, z16.d, z12.d\n" + "zip2 z12.d, z16.d, z12.d\n" + "ld1w { z17.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" + "zip1 z9.d, z18.d, z13.d\n" + "zip2 z13.d, z18.d, z13.d\n" + "zip1 z10.d, z17.d, z14.d\n" + "zip2 z14.d, z17.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" "b 6f\n" @@ -184,11 +184,11 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "7:" // 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, 8f\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, 9f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -200,43 +200,43 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "ble 11f\n" "10:" // Height 1: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1rqh { z1.h }, p0/Z, [x26]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" - "trn2 z1.d, z1.d, z2.d\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x12, #2, MUL VL]\n" - "ld1h { z6.h }, p5/Z, [x12, #3, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11, #2, MUL VL]\n" - "ld1h { z6.h }, p5/Z, [x11, #3, MUL VL]\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" - "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9, #2, MUL VL]\n" - "ld1h { z6.h }, p5/Z, [x9, #3, MUL VL]\n" + "ld1rqh { z20.h }, p0/Z, [x26]\n" + "trn1 z18.d, z20.d, z19.d\n" + "ld1h { z17.h }, p5/Z, [x12]\n" + "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6471e648 // bfmmla z8.s, z18.h, z17.h\n" + ".inst 0x6470e64c // bfmmla z12.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x11]\n" + "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6471e649 // bfmmla z9.s, z18.h, z17.h\n" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6471e64a // bfmmla z10.s, z18.h, z17.h\n" + ".inst 0x6470e64e // bfmmla z14.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x9]\n" + "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n" + "trn2 z20.d, z20.d, z19.d\n" + ".inst 0x6471e64b // bfmmla z11.s, z18.h, z17.h\n" + ".inst 0x6470e64f // bfmmla z15.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x12, #2, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x12, #3, MUL VL]\n" + ".inst 0x6471e688 // bfmmla z8.s, z20.h, z17.h\n" + ".inst 0x6470e68c // bfmmla z12.s, z20.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x11, #2, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x11, #3, MUL VL]\n" + ".inst 0x6471e689 // bfmmla z9.s, z20.h, z17.h\n" + ".inst 0x6470e68d // bfmmla z13.s, z20.h, z16.h\n" + "ld1h { z16.h }, p5/Z, [x10, #2, MUL VL]\n" + "ld1h { z17.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6470e68a // bfmmla z10.s, z20.h, z16.h\n" + ".inst 0x6471e68e // bfmmla z14.s, z20.h, z17.h\n" + "ld1h { z17.h }, p5/Z, [x9, #2, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x9, #3, MUL VL]\n" "sub x27, x27, #0x8\n" "cmp x27, #0x8\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" + ".inst 0x6471e68b // bfmmla z11.s, z20.h, z17.h\n" + ".inst 0x6470e68f // bfmmla z15.s, z20.h, z16.h\n" "add x26, x26, #0x10\n" "addvl x12, x12, #4\n" "addvl x11, x11, #4\n" @@ -246,46 +246,46 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "11:" // Height 1: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" "ld1rqh { z1.h }, p0/Z, [x26]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + "trn1 z18.d, z1.d, z19.d\n" + "ld1h { z17.h }, p5/Z, [x12]\n" + "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6471e648 // bfmmla z8.s, z18.h, z17.h\n" + ".inst 0x6470e64c // bfmmla z12.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x11]\n" + "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6471e649 // bfmmla z9.s, z18.h, z17.h\n" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6471e64a // bfmmla z10.s, z18.h, z17.h\n" + ".inst 0x6470e64e // bfmmla z14.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x9]\n" + "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n" "subs x27, x27, #0x4\n" - "trn2 z1.d, z1.d, z2.d\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" + "trn2 z1.d, z1.d, z19.d\n" + ".inst 0x6471e64b // bfmmla z11.s, z18.h, z17.h\n" + ".inst 0x6470e64f // bfmmla z15.s, z18.h, z16.h\n" "addvl x12, x12, #2\n" "addvl x11, x11, #2\n" "addvl x10, x10, #2\n" "addvl x9, x9, #2\n" "ble 12f\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" + "ld1h { z17.h }, p5/Z, [x12]\n" + "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6471e428 // bfmmla z8.s, z1.h, z17.h\n" + ".inst 0x6470e42c // bfmmla z12.s, z1.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x11]\n" + "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6471e429 // bfmmla z9.s, z1.h, z17.h\n" + ".inst 0x6470e42d // bfmmla z13.s, z1.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6471e42a // bfmmla z10.s, z1.h, z17.h\n" + ".inst 0x6470e42e // bfmmla z14.s, z1.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x9]\n" + "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6471e42b // bfmmla z11.s, z1.h, z17.h\n" + ".inst 0x6470e42f // bfmmla z15.s, z1.h, z16.h\n" "addvl x12, x12, #2\n" "addvl x11, x11, #2\n" "addvl x10, x10, #2\n" @@ -301,17 +301,17 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "uzp1 z11.d, z11.d, z15.d\n" "tbz %x[flags], #1, 13f\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" - "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, 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" + "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" "13:" // Height 1: No activation "st1w { z8.s }, p4, [x13]\n" "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" @@ -376,21 +376,21 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "18:" // Height 2: no bias "tbz %x[flags], #0, 19f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x13]\n" - "ld1w { z10.s }, p3/Z, [x13, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x13, #2, MUL VL]\n" + "add x20, x13, x20, LSL #2\n" + "ld1w { z19.s }, p4/Z, [x13]\n" + "ld1w { z18.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z17.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" + "ld1w { z12.s }, p4/Z, [x20]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n" + "zip1 z9.d, z18.d, z13.d\n" + "zip2 z13.d, z18.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n" + "zip1 z10.d, z17.d, z14.d\n" + "zip2 z14.d, z17.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" "b 20f\n" @@ -408,12 +408,12 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "21:" // 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, 22f\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, 23f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -421,50 +421,50 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "b 23f\n" "22:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" "23:" // Height 2: input setup done "cmp x27, #0x8\n" "ble 25f\n" "24:" // Height 2: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" - "trn2 z1.d, z1.d, z2.d\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x12, #2, MUL VL]\n" - "ld1h { z6.h }, p5/Z, [x12, #3, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11, #2, MUL VL]\n" - "ld1h { z6.h }, p5/Z, [x11, #3, MUL VL]\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" - "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9, #2, MUL VL]\n" - "ld1h { z6.h }, p5/Z, [x9, #3, MUL VL]\n" + "ld1rqh { z20.h }, p0/Z, [x26]\n" + "ld1rqh { z19.h }, p0/Z, [x25]\n" + "trn1 z18.d, z20.d, z19.d\n" + "ld1h { z17.h }, p5/Z, [x12]\n" + "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6471e648 // bfmmla z8.s, z18.h, z17.h\n" + ".inst 0x6470e64c // bfmmla z12.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x11]\n" + "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6471e649 // bfmmla z9.s, z18.h, z17.h\n" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6471e64a // bfmmla z10.s, z18.h, z17.h\n" + ".inst 0x6470e64e // bfmmla z14.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x9]\n" + "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n" + "trn2 z20.d, z20.d, z19.d\n" + ".inst 0x6471e64b // bfmmla z11.s, z18.h, z17.h\n" + ".inst 0x6470e64f // bfmmla z15.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x12, #2, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x12, #3, MUL VL]\n" + ".inst 0x6471e688 // bfmmla z8.s, z20.h, z17.h\n" + ".inst 0x6470e68c // bfmmla z12.s, z20.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x11, #2, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x11, #3, MUL VL]\n" + ".inst 0x6471e689 // bfmmla z9.s, z20.h, z17.h\n" + ".inst 0x6470e68d // bfmmla z13.s, z20.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6471e68a // bfmmla z10.s, z20.h, z17.h\n" + ".inst 0x6470e68e // bfmmla z14.s, z20.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x9, #2, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x9, #3, MUL VL]\n" "sub x27, x27, #0x8\n" "cmp x27, #0x8\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" + ".inst 0x6471e68b // bfmmla z11.s, z20.h, z17.h\n" + ".inst 0x6470e68f // bfmmla z15.s, z20.h, z16.h\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "addvl x12, x12, #4\n" @@ -475,47 +475,47 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "25:" // Height 2: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + "ld1rqh { z19.h }, p0/Z, [x25]\n" + "trn1 z18.d, z1.d, z19.d\n" + "ld1h { z17.h }, p5/Z, [x12]\n" + "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6471e648 // bfmmla z8.s, z18.h, z17.h\n" + ".inst 0x6470e64c // bfmmla z12.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x11]\n" + "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6471e649 // bfmmla z9.s, z18.h, z17.h\n" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6471e64a // bfmmla z10.s, z18.h, z17.h\n" + ".inst 0x6470e64e // bfmmla z14.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x9]\n" + "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n" "subs x27, x27, #0x4\n" - "trn2 z1.d, z1.d, z2.d\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" + "trn2 z1.d, z1.d, z19.d\n" + ".inst 0x6471e64b // bfmmla z11.s, z18.h, z17.h\n" + ".inst 0x6470e64f // bfmmla z15.s, z18.h, z16.h\n" "addvl x12, x12, #2\n" "addvl x11, x11, #2\n" "addvl x10, x10, #2\n" "addvl x9, x9, #2\n" "ble 26f\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" + "ld1h { z17.h }, p5/Z, [x12]\n" + "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6471e428 // bfmmla z8.s, z1.h, z17.h\n" + ".inst 0x6470e42c // bfmmla z12.s, z1.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x11]\n" + "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6471e429 // bfmmla z9.s, z1.h, z17.h\n" + ".inst 0x6470e42d // bfmmla z13.s, z1.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6471e42a // bfmmla z10.s, z1.h, z17.h\n" + ".inst 0x6470e42e // bfmmla z14.s, z1.h, z16.h\n" + "ld1h { z22.h }, p5/Z, [x9]\n" + "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6476e42b // bfmmla z11.s, z1.h, z22.h\n" + ".inst 0x6470e42f // bfmmla z15.s, z1.h, z16.h\n" "addvl x12, x12, #2\n" "addvl x11, x11, #2\n" "addvl x10, x10, #2\n" @@ -537,25 +537,25 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "uzp2 z11.d, z11.d, z15.d\n" "tbz %x[flags], #1, 27f\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 z7.s, p5/M, z7.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 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 z7.s, p5/M, z7.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 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 z7.s, p5/M, z7.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 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 z7.s, p5/M, z7.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 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" "27:" // Height 2: No activation "st1w { z7.s }, p4, [x13]\n" "st1w { z12.s }, p3, [x13, #1, MUL VL]\n" @@ -632,28 +632,28 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "32:" // Height 3: no bias "tbz %x[flags], #0, 33f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x13]\n" - "ld1w { z10.s }, p3/Z, [x13, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x13, #2, MUL VL]\n" + "add x21, x13, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z19.s }, p4/Z, [x13]\n" + "ld1w { z17.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x24]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x21]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z13.d\n" + "zip2 z13.d, z17.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x20]\n" + "zip1 z10.d, z18.d, z14.d\n" + "zip2 z14.d, z18.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" "zip1 z17.d, z18.d, z21.d\n" @@ -685,13 +685,13 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "35:" // 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, 36f\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, 37f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -700,145 +700,145 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "b 37f\n" "36:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" "37:" // Height 3: input setup done "cmp x27, #0x8\n" "ble 39f\n" "38:" // Height 3: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" - "ld1rqh { z3.h }, p0/Z, [x24]\n" - "trn1 z0.d, z1.d, z2.d\n" - "trn2 z1.d, z1.d, z2.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - "trn2 z3.d, z3.d, z4.d\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + "ld1rqh { z30.h }, p0/Z, [x26]\n" + "ld1rqh { z24.h }, p0/Z, [x25]\n" + "ld1rqh { z28.h }, p0/Z, [x24]\n" + "trn1 z27.d, z30.d, z24.d\n" + "trn2 z30.d, z30.d, z24.d\n" + "ld1h { z25.h }, p5/Z, [x12]\n" + "trn1 z26.d, z28.d, z29.d\n" + "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6479e768 // bfmmla z8.s, z27.h, z25.h\n" + ".inst 0x6479e750 // bfmmla z16.s, z26.h, z25.h\n" + ".inst 0x6478e76c // bfmmla z12.s, z27.h, z24.h\n" + ".inst 0x6478e754 // bfmmla z20.s, z26.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x11]\n" + "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6479e769 // bfmmla z9.s, z27.h, z25.h\n" + ".inst 0x6479e751 // bfmmla z17.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" + "trn2 z28.d, z28.d, z29.d\n" + ".inst 0x6478e76d // bfmmla z13.s, z27.h, z24.h\n" + ".inst 0x6478e755 // bfmmla z21.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" "sub x27, x27, #0x8\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" + ".inst 0x6479e76a // bfmmla z10.s, z27.h, z25.h\n" + ".inst 0x6479e752 // bfmmla z18.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x9]\n" "cmp x27, #0x8\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6478e76e // bfmmla z14.s, z27.h, z24.h\n" + ".inst 0x6478e756 // bfmmla z22.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n" "add x26, x26, #0x10\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x12, #2, MUL VL]\n" + ".inst 0x6479e76b // bfmmla z11.s, z27.h, z25.h\n" + ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x12, #2, MUL VL]\n" "add x25, x25, #0x10\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x12, #3, MUL VL]\n" + ".inst 0x6478e76f // bfmmla z15.s, z27.h, z24.h\n" + ".inst 0x6478e757 // bfmmla z23.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x12, #3, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x11, #2, MUL VL]\n" + ".inst 0x6479e7c8 // bfmmla z8.s, z30.h, z25.h\n" + ".inst 0x6479e790 // bfmmla z16.s, z28.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x11, #2, MUL VL]\n" "addvl x12, x12, #4\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #3, MUL VL]\n" + ".inst 0x6478e7cc // bfmmla z12.s, z30.h, z24.h\n" + ".inst 0x6478e794 // bfmmla z20.s, z28.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x11, #3, MUL VL]\n" "addvl x11, x11, #4\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6479e7c9 // bfmmla z9.s, z30.h, z25.h\n" + ".inst 0x6479e791 // bfmmla z17.s, z28.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6478e7cd // bfmmla z13.s, z30.h, z24.h\n" + ".inst 0x6478e795 // bfmmla z21.s, z28.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9, #2, MUL VL]\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #3, MUL VL]\n" + ".inst 0x6479e7ca // bfmmla z10.s, z30.h, z25.h\n" + ".inst 0x6479e792 // bfmmla z18.s, z28.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x9, #2, MUL VL]\n" + ".inst 0x6478e7ce // bfmmla z14.s, z30.h, z24.h\n" + ".inst 0x6478e796 // bfmmla z22.s, z28.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6467e473 // bfmmla z19.s, z3.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" - ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" + ".inst 0x6479e7cb // bfmmla z11.s, z30.h, z25.h\n" + ".inst 0x6479e793 // bfmmla z19.s, z28.h, z25.h\n" + ".inst 0x6478e7cf // bfmmla z15.s, z30.h, z24.h\n" + ".inst 0x6478e797 // bfmmla z23.s, z28.h, z24.h\n" "bgt 38b\n" "39:" // Height 3: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" + "ld1rqh { z24.h }, p0/Z, [x25]\n" "ld1rqh { z3.h }, p0/Z, [x24]\n" - "trn1 z0.d, z1.d, z2.d\n" - "trn2 z1.d, z1.d, z2.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" + "trn1 z27.d, z1.d, z24.d\n" + "trn2 z1.d, z1.d, z24.d\n" + "ld1h { z25.h }, p5/Z, [x12]\n" + "trn1 z26.d, z3.d, z28.d\n" + "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6479e768 // bfmmla z8.s, z27.h, z25.h\n" + ".inst 0x6479e750 // bfmmla z16.s, z26.h, z25.h\n" + ".inst 0x6478e76c // bfmmla z12.s, z27.h, z24.h\n" + ".inst 0x6478e754 // bfmmla z20.s, z26.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x11]\n" + "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6479e769 // bfmmla z9.s, z27.h, z25.h\n" + ".inst 0x6479e751 // bfmmla z17.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" "subs x27, x27, #0x4\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "trn2 z3.d, z3.d, z4.d\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" + ".inst 0x6478e76d // bfmmla z13.s, z27.h, z24.h\n" + ".inst 0x6478e755 // bfmmla z21.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" + "trn2 z3.d, z3.d, z28.d\n" + ".inst 0x6479e76a // bfmmla z10.s, z27.h, z25.h\n" + ".inst 0x6479e752 // bfmmla z18.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x9]\n" "addvl x12, x12, #2\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6478e76e // bfmmla z14.s, z27.h, z24.h\n" + ".inst 0x6478e756 // bfmmla z22.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n" "addvl x11, x11, #2\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" + ".inst 0x6479e76b // bfmmla z11.s, z27.h, z25.h\n" + ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" "addvl x10, x10, #2\n" "addvl x9, x9, #2\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" + ".inst 0x6478e76f // bfmmla z15.s, z27.h, z24.h\n" + ".inst 0x6478e757 // bfmmla z23.s, z26.h, z24.h\n" "ble 40f\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" + "ld1h { z25.h }, p5/Z, [x12]\n" + "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6479e428 // bfmmla z8.s, z1.h, z25.h\n" + ".inst 0x6479e470 // bfmmla z16.s, z3.h, z25.h\n" + ".inst 0x6478e42c // bfmmla z12.s, z1.h, z24.h\n" + ".inst 0x6478e474 // bfmmla z20.s, z3.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x11]\n" + "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6479e429 // bfmmla z9.s, z1.h, z25.h\n" + ".inst 0x6479e471 // bfmmla z17.s, z3.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" "addvl x12, x12, #2\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6478e42d // bfmmla z13.s, z1.h, z24.h\n" + ".inst 0x6478e475 // bfmmla z21.s, z3.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" "addvl x11, x11, #2\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" + ".inst 0x6479e42a // bfmmla z10.s, z1.h, z25.h\n" + ".inst 0x6479e472 // bfmmla z18.s, z3.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x9]\n" "addvl x10, x10, #2\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6478e42e // bfmmla z14.s, z1.h, z24.h\n" + ".inst 0x6478e476 // bfmmla z22.s, z3.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n" "addvl x9, x9, #2\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6467e473 // bfmmla z19.s, z3.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" - ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" + ".inst 0x6479e42b // bfmmla z11.s, z1.h, z25.h\n" + ".inst 0x6479e473 // bfmmla z19.s, z3.h, z25.h\n" + ".inst 0x6478e42f // bfmmla z15.s, z1.h, z24.h\n" + ".inst 0x6478e477 // bfmmla z23.s, z3.h, z24.h\n" "40:" // Height 3: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -861,33 +861,33 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "uzp1 z19.d, z19.d, z23.d\n" "tbz %x[flags], #1, 41f\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 z7.s, p5/M, z7.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 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 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 z7.s, p5/M, z7.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 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 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 { z24.s }, p5/Z, [x20]\n" + "fmin z7.s, p5/M, z7.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 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 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" + "fmax z7.s, p5/M, z7.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 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 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" "41:" // Height 3: No activation "st1w { z7.s }, p4, [x13]\n" "st1w { z12.s }, p3, [x13, #1, MUL VL]\n" @@ -968,37 +968,37 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "46:" // Height 4: no bias "tbz %x[flags], #0, 47f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x13]\n" - "add x23, x24, x20, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x13, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x13, #2, MUL VL]\n" + "add x22, x13, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" + "ld1w { z19.s }, p4/Z, [x13]\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z17.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x24]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x22]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z13.d\n" + "zip2 z13.d, z17.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x21]\n" + "zip1 z10.d, z18.d, z14.d\n" + "zip2 z14.d, z18.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x21, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" + "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x20]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z17.d, z18.d, z21.d\n" "zip2 z21.d, z18.d, z21.d\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n" "zip1 z18.d, z19.d, z22.d\n" "zip2 z22.d, z19.d, z22.d\n" "zip1 z19.d, z24.d, z23.d\n" @@ -1026,14 +1026,14 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "49:" // 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, 50f\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, 51f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -1043,149 +1043,149 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "b 51f\n" "50:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" "51:" // Height 4: input setup done "cmp x27, #0x8\n" "ble 53f\n" "52:" // Height 4: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x24]\n" - "ld1rqh { z4.h }, p0/Z, [x23]\n" - "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - "trn2 z3.d, z3.d, z4.d\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + "ld1rqh { z30.h }, p0/Z, [x26]\n" + "ld1rqh { z24.h }, p0/Z, [x25]\n" + "trn1 z29.d, z30.d, z24.d\n" + "ld1rqh { z28.h }, p0/Z, [x24]\n" + "ld1rqh { z27.h }, p0/Z, [x23]\n" + "trn2 z30.d, z30.d, z24.d\n" + "trn1 z26.d, z28.d, z27.d\n" + "ld1h { z25.h }, p5/Z, [x12]\n" + "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6479e7a8 // bfmmla z8.s, z29.h, z25.h\n" + ".inst 0x6479e750 // bfmmla z16.s, z26.h, z25.h\n" + ".inst 0x6478e7ac // bfmmla z12.s, z29.h, z24.h\n" + ".inst 0x6478e754 // bfmmla z20.s, z26.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x11]\n" + "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6479e7a9 // bfmmla z9.s, z29.h, z25.h\n" + ".inst 0x6479e751 // bfmmla z17.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" + "trn2 z28.d, z28.d, z27.d\n" + ".inst 0x6478e7ad // bfmmla z13.s, z29.h, z24.h\n" + ".inst 0x6478e755 // bfmmla z21.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" "sub x27, x27, #0x8\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" + ".inst 0x6479e7aa // bfmmla z10.s, z29.h, z25.h\n" + ".inst 0x6479e752 // bfmmla z18.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x9]\n" "cmp x27, #0x8\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6478e7ae // bfmmla z14.s, z29.h, z24.h\n" + ".inst 0x6478e756 // bfmmla z22.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n" "add x26, x26, #0x10\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x12, #2, MUL VL]\n" + ".inst 0x6479e7ab // bfmmla z11.s, z29.h, z25.h\n" + ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x12, #2, MUL VL]\n" "add x25, x25, #0x10\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x12, #3, MUL VL]\n" + ".inst 0x6478e7af // bfmmla z15.s, z29.h, z24.h\n" + ".inst 0x6478e757 // bfmmla z23.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x12, #3, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x11, #2, MUL VL]\n" + ".inst 0x6479e7c8 // bfmmla z8.s, z30.h, z25.h\n" + ".inst 0x6479e790 // bfmmla z16.s, z28.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x11, #2, MUL VL]\n" "add x23, x23, #0x10\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #3, MUL VL]\n" + ".inst 0x6478e7cc // bfmmla z12.s, z30.h, z24.h\n" + ".inst 0x6478e794 // bfmmla z20.s, z28.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x11, #3, MUL VL]\n" "addvl x12, x12, #4\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6479e7c9 // bfmmla z9.s, z30.h, z25.h\n" + ".inst 0x6479e791 // bfmmla z17.s, z28.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n" "addvl x11, x11, #4\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6478e7cd // bfmmla z13.s, z30.h, z24.h\n" + ".inst 0x6478e795 // bfmmla z21.s, z28.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9, #2, MUL VL]\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #3, MUL VL]\n" + ".inst 0x6479e7ca // bfmmla z10.s, z30.h, z25.h\n" + ".inst 0x6479e792 // bfmmla z18.s, z28.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x9, #2, MUL VL]\n" + ".inst 0x6478e7ce // bfmmla z14.s, z30.h, z24.h\n" + ".inst 0x6478e796 // bfmmla z22.s, z28.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6467e473 // bfmmla z19.s, z3.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" - ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" + ".inst 0x6479e7cb // bfmmla z11.s, z30.h, z25.h\n" + ".inst 0x6479e793 // bfmmla z19.s, z28.h, z25.h\n" + ".inst 0x6478e7cf // bfmmla z15.s, z30.h, z24.h\n" + ".inst 0x6478e797 // bfmmla z23.s, z28.h, z24.h\n" "bgt 52b\n" "53:" // Height 4: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" + "ld1rqh { z24.h }, p0/Z, [x25]\n" + "trn1 z28.d, z1.d, z24.d\n" "ld1rqh { z3.h }, p0/Z, [x24]\n" - "ld1rqh { z4.h }, p0/Z, [x23]\n" - "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" + "ld1rqh { z27.h }, p0/Z, [x23]\n" + "trn2 z1.d, z1.d, z24.d\n" + "trn1 z26.d, z3.d, z27.d\n" + "ld1h { z25.h }, p5/Z, [x12]\n" + "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6479e788 // bfmmla z8.s, z28.h, z25.h\n" + ".inst 0x6479e750 // bfmmla z16.s, z26.h, z25.h\n" + ".inst 0x6478e78c // bfmmla z12.s, z28.h, z24.h\n" + ".inst 0x6478e754 // bfmmla z20.s, z26.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x11]\n" + "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6479e789 // bfmmla z9.s, z28.h, z25.h\n" + ".inst 0x6479e751 // bfmmla z17.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" "subs x27, x27, #0x4\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "trn2 z3.d, z3.d, z4.d\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" + ".inst 0x6478e78d // bfmmla z13.s, z28.h, z24.h\n" + ".inst 0x6478e755 // bfmmla z21.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" + "trn2 z3.d, z3.d, z27.d\n" + ".inst 0x6479e78a // bfmmla z10.s, z28.h, z25.h\n" + ".inst 0x6479e752 // bfmmla z18.s, z26.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x9]\n" "addvl x12, x12, #2\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6478e78e // bfmmla z14.s, z28.h, z24.h\n" + ".inst 0x6478e756 // bfmmla z22.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n" "addvl x11, x11, #2\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" + ".inst 0x6479e78b // bfmmla z11.s, z28.h, z25.h\n" + ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" "addvl x10, x10, #2\n" "addvl x9, x9, #2\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" + ".inst 0x6478e78f // bfmmla z15.s, z28.h, z24.h\n" + ".inst 0x6478e757 // bfmmla z23.s, z26.h, z24.h\n" "ble 54f\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" + "ld1h { z25.h }, p5/Z, [x12]\n" + "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6479e428 // bfmmla z8.s, z1.h, z25.h\n" + ".inst 0x6479e470 // bfmmla z16.s, z3.h, z25.h\n" + ".inst 0x6478e42c // bfmmla z12.s, z1.h, z24.h\n" + ".inst 0x6478e474 // bfmmla z20.s, z3.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x11]\n" + "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6479e429 // bfmmla z9.s, z1.h, z25.h\n" + ".inst 0x6479e471 // bfmmla z17.s, z3.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" "addvl x12, x12, #2\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6478e42d // bfmmla z13.s, z1.h, z24.h\n" + ".inst 0x6478e475 // bfmmla z21.s, z3.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" "addvl x11, x11, #2\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" + ".inst 0x6479e42a // bfmmla z10.s, z1.h, z25.h\n" + ".inst 0x6479e472 // bfmmla z18.s, z3.h, z25.h\n" + "ld1h { z25.h }, p5/Z, [x9]\n" "addvl x10, x10, #2\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6478e42e // bfmmla z14.s, z1.h, z24.h\n" + ".inst 0x6478e476 // bfmmla z22.s, z3.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n" "addvl x9, x9, #2\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6467e473 // bfmmla z19.s, z3.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" - ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" + ".inst 0x6479e42b // bfmmla z11.s, z1.h, z25.h\n" + ".inst 0x6479e473 // bfmmla z19.s, z3.h, z25.h\n" + ".inst 0x6478e42f // bfmmla z15.s, z1.h, z24.h\n" + ".inst 0x6478e477 // bfmmla z23.s, z3.h, z24.h\n" "54:" // Height 4: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1213,41 +1213,41 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "uzp2 z19.d, z19.d, z23.d\n" "tbz %x[flags], #1, 55f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z24.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z7.s, p5/M, z7.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 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 z15.s, p5/M, z15.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 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 z7.s, p5/M, z7.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 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 z15.s, p5/M, z15.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 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 { z23.s }, p5/Z, [x20]\n" + "fmin z7.s, p5/M, z7.s, z24.s\n" + "fmin z12.s, p5/M, z12.s, z24.s\n" + "fmin z13.s, p5/M, z13.s, z24.s\n" + "fmin z14.s, p5/M, z14.s, z24.s\n" + "fmin z8.s, p5/M, z8.s, z24.s\n" + "fmin z9.s, p5/M, z9.s, z24.s\n" + "fmin z10.s, p5/M, z10.s, z24.s\n" + "fmin z11.s, p5/M, z11.s, z24.s\n" + "fmin z15.s, p5/M, z15.s, z24.s\n" + "fmin z20.s, p5/M, z20.s, z24.s\n" + "fmin z21.s, p5/M, z21.s, z24.s\n" + "fmin z22.s, p5/M, z22.s, z24.s\n" + "fmin z16.s, p5/M, z16.s, z24.s\n" + "fmin z17.s, p5/M, z17.s, z24.s\n" + "fmin z18.s, p5/M, z18.s, z24.s\n" + "fmin z19.s, p5/M, z19.s, z24.s\n" + "fmax z7.s, p5/M, z7.s, z23.s\n" + "fmax z12.s, p5/M, z12.s, z23.s\n" + "fmax z13.s, p5/M, z13.s, z23.s\n" + "fmax z14.s, p5/M, z14.s, z23.s\n" + "fmax z8.s, p5/M, z8.s, z23.s\n" + "fmax z9.s, p5/M, z9.s, z23.s\n" + "fmax z10.s, p5/M, z10.s, z23.s\n" + "fmax z11.s, p5/M, z11.s, z23.s\n" + "fmax z15.s, p5/M, z15.s, z23.s\n" + "fmax z20.s, p5/M, z20.s, z23.s\n" + "fmax z21.s, p5/M, z21.s, z23.s\n" + "fmax z22.s, p5/M, z22.s, z23.s\n" + "fmax z16.s, p5/M, z16.s, z23.s\n" + "fmax z17.s, p5/M, z17.s, z23.s\n" + "fmax z18.s, p5/M, z18.s, z23.s\n" + "fmax z19.s, p5/M, z19.s, z23.s\n" "55:" // Height 4: No activation "st1w { z7.s }, p4, [x13]\n" "st1w { z12.s }, p3, [x13, #1, MUL VL]\n" @@ -1340,54 +1340,54 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "60:" // Height 5: no bias "tbz %x[flags], #0, 61f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x13]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x13, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x13, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z19.s }, p4/Z, [x13]\n" + "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z17.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x24]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x23]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z13.d\n" + "zip2 z13.d, z17.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x22]\n" + "zip1 z10.d, z18.d, z14.d\n" + "zip2 z14.d, z18.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" + "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x21]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n" "zip1 z17.d, z18.d, z21.d\n" "zip2 z21.d, z18.d, z21.d\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z25.s }, p4/Z, [x22]\n" + "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z25.s }, p4/Z, [x20]\n" "zip1 z18.d, z19.d, z22.d\n" "zip2 z22.d, z19.d, z22.d\n" - "ld1w { z26.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z27.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z19.d, z24.d, z23.d\n" "zip2 z23.d, z24.d, z23.d\n" - "ld1w { z6.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n" "zip1 z24.d, z25.d, z28.d\n" "zip2 z28.d, z25.d, z28.d\n" "zip1 z25.d, z26.d, z29.d\n" "zip2 z29.d, z26.d, z29.d\n" "zip1 z26.d, z27.d, z30.d\n" "zip2 z30.d, z27.d, z30.d\n" - "zip1 z27.d, z6.d, z31.d\n" - "zip2 z31.d, z6.d, z31.d\n" + "zip1 z27.d, z0.d, z31.d\n" + "zip2 z31.d, z0.d, z31.d\n" "b 62f\n" "61:" // Height 5: no accumulate "mov z8.b, #0x0\n" @@ -1419,15 +1419,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "63:" // 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, 64f\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, 65f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -1438,189 +1438,189 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "b 65f\n" "64:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" + "add x22, x23, x21, LSL #1\n" "65:" // Height 5: input setup done "cmp x27, #0x8\n" "ble 67f\n" "66:" // Height 5: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" - "ld1rqh { z3.h }, p0/Z, [x24]\n" - "ld1rqh { z4.h }, p0/Z, [x23]\n" - "trn1 z0.d, z1.d, z2.d\n" - "trn2 z1.d, z1.d, z2.d\n" - "ld1rqh { z5.h }, p0/Z, [x22]\n" - "trn1 z2.d, z3.d, z4.d\n" - "trn2 z3.d, z3.d, z4.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "trn1 z4.d, z5.d, z6.d\n" - "trn2 z5.d, z5.d, z6.d\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - ".inst 0x6467e498 // bfmmla z24.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" + "ld1rqh { z6.h }, p0/Z, [x26]\n" + "ld1rqh { z1.h }, p0/Z, [x25]\n" + "ld1rqh { z7.h }, p0/Z, [x24]\n" + "ld1rqh { z2.h }, p0/Z, [x23]\n" + "trn1 z5.d, z6.d, z1.d\n" + "trn2 z6.d, z6.d, z1.d\n" + "ld1rqh { z4.h }, p0/Z, [x22]\n" + "trn1 z3.d, z7.d, z2.d\n" + "trn2 z7.d, z7.d, z2.d\n" + "ld1h { z1.h }, p5/Z, [x12]\n" + "trn1 z2.d, z4.d, z0.d\n" + "trn2 z4.d, z4.d, z0.d\n" + "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6461e4a8 // bfmmla z8.s, z5.h, z1.h\n" + ".inst 0x6461e470 // bfmmla z16.s, z3.h, z1.h\n" + ".inst 0x6461e458 // bfmmla z24.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x11]\n" "sub x27, x27, #0x8\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + ".inst 0x6460e4ac // bfmmla z12.s, z5.h, z0.h\n" + ".inst 0x6460e474 // bfmmla z20.s, z3.h, z0.h\n" "cmp x27, #0x8\n" "add x26, x26, #0x10\n" - ".inst 0x6466e49c // bfmmla z28.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" + ".inst 0x6460e45c // bfmmla z28.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6461e4a9 // bfmmla z9.s, z5.h, z1.h\n" "add x25, x25, #0x10\n" - ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" - ".inst 0x6467e499 // bfmmla z25.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" + ".inst 0x6461e471 // bfmmla z17.s, z3.h, z1.h\n" + ".inst 0x6461e459 // bfmmla z25.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10]\n" "add x24, x24, #0x10\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" + ".inst 0x6460e4ad // bfmmla z13.s, z5.h, z0.h\n" + ".inst 0x6460e475 // bfmmla z21.s, z3.h, z0.h\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - ".inst 0x6466e49d // bfmmla z29.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" - ".inst 0x6467e49a // bfmmla z26.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" - ".inst 0x6466e49e // bfmmla z30.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" - ".inst 0x6467e49b // bfmmla z27.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x12, #2, MUL VL]\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - ".inst 0x6466e49f // bfmmla z31.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x12, #3, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" + ".inst 0x6460e45d // bfmmla z29.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6461e4aa // bfmmla z10.s, z5.h, z1.h\n" + ".inst 0x6461e472 // bfmmla z18.s, z3.h, z1.h\n" + ".inst 0x6461e45a // bfmmla z26.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x9]\n" + ".inst 0x6460e4ae // bfmmla z14.s, z5.h, z0.h\n" + ".inst 0x6460e476 // bfmmla z22.s, z3.h, z0.h\n" + ".inst 0x6460e45e // bfmmla z30.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6461e4ab // bfmmla z11.s, z5.h, z1.h\n" + ".inst 0x6461e473 // bfmmla z19.s, z3.h, z1.h\n" + ".inst 0x6461e45b // bfmmla z27.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x12, #2, MUL VL]\n" + ".inst 0x6460e4af // bfmmla z15.s, z5.h, z0.h\n" + ".inst 0x6460e477 // bfmmla z23.s, z3.h, z0.h\n" + ".inst 0x6460e45f // bfmmla z31.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x12, #3, MUL VL]\n" + ".inst 0x6461e4c8 // bfmmla z8.s, z6.h, z1.h\n" "addvl x12, x12, #4\n" - ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - ".inst 0x6467e4b8 // bfmmla z24.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x11, #2, MUL VL]\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" - ".inst 0x6466e4bc // bfmmla z28.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #3, MUL VL]\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" + ".inst 0x6461e4f0 // bfmmla z16.s, z7.h, z1.h\n" + ".inst 0x6461e498 // bfmmla z24.s, z4.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x11, #2, MUL VL]\n" + ".inst 0x6460e4cc // bfmmla z12.s, z6.h, z0.h\n" + ".inst 0x6460e4f4 // bfmmla z20.s, z7.h, z0.h\n" + ".inst 0x6460e49c // bfmmla z28.s, z4.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x11, #3, MUL VL]\n" + ".inst 0x6461e4c9 // bfmmla z9.s, z6.h, z1.h\n" "addvl x11, x11, #4\n" - ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" - ".inst 0x6467e4b9 // bfmmla z25.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" - ".inst 0x6466e4bd // bfmmla z29.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" + ".inst 0x6461e4f1 // bfmmla z17.s, z7.h, z1.h\n" + ".inst 0x6461e499 // bfmmla z25.s, z4.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6460e4cd // bfmmla z13.s, z6.h, z0.h\n" + ".inst 0x6460e4f5 // bfmmla z21.s, z7.h, z0.h\n" + ".inst 0x6460e49d // bfmmla z29.s, z4.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6461e4ca // bfmmla z10.s, z6.h, z1.h\n" "addvl x10, x10, #4\n" - ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" - ".inst 0x6467e4ba // bfmmla z26.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9, #2, MUL VL]\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" - ".inst 0x6466e4be // bfmmla z30.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #3, MUL VL]\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" + ".inst 0x6461e4f2 // bfmmla z18.s, z7.h, z1.h\n" + ".inst 0x6461e49a // bfmmla z26.s, z4.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x9, #2, MUL VL]\n" + ".inst 0x6460e4ce // bfmmla z14.s, z6.h, z0.h\n" + ".inst 0x6460e4f6 // bfmmla z22.s, z7.h, z0.h\n" + ".inst 0x6460e49e // bfmmla z30.s, z4.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x9, #3, MUL VL]\n" + ".inst 0x6461e4cb // bfmmla z11.s, z6.h, z1.h\n" "addvl x9, x9, #4\n" - ".inst 0x6467e473 // bfmmla z19.s, z3.h, z7.h\n" - ".inst 0x6467e4bb // bfmmla z27.s, z5.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" - ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" - ".inst 0x6466e4bf // bfmmla z31.s, z5.h, z6.h\n" + ".inst 0x6461e4f3 // bfmmla z19.s, z7.h, z1.h\n" + ".inst 0x6461e49b // bfmmla z27.s, z4.h, z1.h\n" + ".inst 0x6460e4cf // bfmmla z15.s, z6.h, z0.h\n" + ".inst 0x6460e4f7 // bfmmla z23.s, z7.h, z0.h\n" + ".inst 0x6460e49f // bfmmla z31.s, z4.h, z0.h\n" "bgt 66b\n" "67:" // Height 5: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" + "ld1rqh { z4.h }, p0/Z, [x25]\n" "ld1rqh { z3.h }, p0/Z, [x24]\n" - "ld1rqh { z4.h }, p0/Z, [x23]\n" - "trn1 z0.d, z1.d, z2.d\n" - "trn2 z1.d, z1.d, z2.d\n" + "ld1rqh { z2.h }, p0/Z, [x23]\n" + "trn1 z7.d, z1.d, z4.d\n" + "trn2 z1.d, z1.d, z4.d\n" "ld1rqh { z5.h }, p0/Z, [x22]\n" - "trn1 z2.d, z3.d, z4.d\n" - "trn2 z3.d, z3.d, z4.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "trn1 z4.d, z5.d, z6.d\n" - "trn2 z5.d, z5.d, z6.d\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - ".inst 0x6467e498 // bfmmla z24.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" + "trn1 z6.d, z3.d, z2.d\n" + "trn2 z3.d, z3.d, z2.d\n" + "ld1h { z2.h }, p5/Z, [x12]\n" + "trn1 z4.d, z5.d, z0.d\n" + "trn2 z5.d, z5.d, z0.d\n" + "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6462e4e8 // bfmmla z8.s, z7.h, z2.h\n" + ".inst 0x6462e4d0 // bfmmla z16.s, z6.h, z2.h\n" + ".inst 0x6462e498 // bfmmla z24.s, z4.h, z2.h\n" + "ld1h { z2.h }, p5/Z, [x11]\n" "subs x27, x27, #0x4\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + ".inst 0x6460e4ec // bfmmla z12.s, z7.h, z0.h\n" + ".inst 0x6460e4d4 // bfmmla z20.s, z6.h, z0.h\n" "addvl x12, x12, #2\n" - ".inst 0x6466e49c // bfmmla z28.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" + ".inst 0x6460e49c // bfmmla z28.s, z4.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6462e4e9 // bfmmla z9.s, z7.h, z2.h\n" "addvl x11, x11, #2\n" - ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" - ".inst 0x6467e499 // bfmmla z25.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - ".inst 0x6466e49d // bfmmla z29.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" + ".inst 0x6462e4d1 // bfmmla z17.s, z6.h, z2.h\n" + ".inst 0x6462e499 // bfmmla z25.s, z4.h, z2.h\n" + "ld1h { z2.h }, p5/Z, [x10]\n" + ".inst 0x6460e4ed // bfmmla z13.s, z7.h, z0.h\n" + ".inst 0x6460e4d5 // bfmmla z21.s, z6.h, z0.h\n" + ".inst 0x6460e49d // bfmmla z29.s, z4.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6462e4ea // bfmmla z10.s, z7.h, z2.h\n" "addvl x10, x10, #2\n" - ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" - ".inst 0x6467e49a // bfmmla z26.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" - ".inst 0x6466e49e // bfmmla z30.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" + ".inst 0x6462e4d2 // bfmmla z18.s, z6.h, z2.h\n" + ".inst 0x6462e49a // bfmmla z26.s, z4.h, z2.h\n" + "ld1h { z2.h }, p5/Z, [x9]\n" + ".inst 0x6460e4ee // bfmmla z14.s, z7.h, z0.h\n" + ".inst 0x6460e4d6 // bfmmla z22.s, z6.h, z0.h\n" + ".inst 0x6460e49e // bfmmla z30.s, z4.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6462e4eb // bfmmla z11.s, z7.h, z2.h\n" "addvl x9, x9, #2\n" - ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" - ".inst 0x6467e49b // bfmmla z27.s, z4.h, z7.h\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - ".inst 0x6466e49f // bfmmla z31.s, z4.h, z6.h\n" + ".inst 0x6462e4d3 // bfmmla z19.s, z6.h, z2.h\n" + ".inst 0x6462e49b // bfmmla z27.s, z4.h, z2.h\n" + ".inst 0x6460e4ef // bfmmla z15.s, z7.h, z0.h\n" + ".inst 0x6460e4d7 // bfmmla z23.s, z6.h, z0.h\n" + ".inst 0x6460e49f // bfmmla z31.s, z4.h, z0.h\n" "ble 68f\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - ".inst 0x6467e4b8 // bfmmla z24.s, z5.h, z7.h\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" + "ld1h { z2.h }, p5/Z, [x12]\n" + "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6462e428 // bfmmla z8.s, z1.h, z2.h\n" + ".inst 0x6462e470 // bfmmla z16.s, z3.h, z2.h\n" + ".inst 0x6462e4b8 // bfmmla z24.s, z5.h, z2.h\n" + ".inst 0x6460e42c // bfmmla z12.s, z1.h, z0.h\n" + "ld1h { z2.h }, p5/Z, [x11]\n" "addvl x12, x12, #2\n" - ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" - ".inst 0x6466e4bc // bfmmla z28.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6460e474 // bfmmla z20.s, z3.h, z0.h\n" + ".inst 0x6460e4bc // bfmmla z28.s, z5.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n" "addvl x11, x11, #2\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" - ".inst 0x6467e4b9 // bfmmla z25.s, z5.h, z7.h\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" - ".inst 0x6466e4bd // bfmmla z29.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6462e429 // bfmmla z9.s, z1.h, z2.h\n" + ".inst 0x6462e471 // bfmmla z17.s, z3.h, z2.h\n" + ".inst 0x6462e4b9 // bfmmla z25.s, z5.h, z2.h\n" + ".inst 0x6460e42d // bfmmla z13.s, z1.h, z0.h\n" + "ld1h { z2.h }, p5/Z, [x10]\n" + ".inst 0x6460e475 // bfmmla z21.s, z3.h, z0.h\n" + ".inst 0x6460e4bd // bfmmla z29.s, z5.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" "addvl x10, x10, #2\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" - ".inst 0x6467e4ba // bfmmla z26.s, z5.h, z7.h\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" - ".inst 0x6466e4be // bfmmla z30.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6462e42a // bfmmla z10.s, z1.h, z2.h\n" + ".inst 0x6462e472 // bfmmla z18.s, z3.h, z2.h\n" + ".inst 0x6462e4ba // bfmmla z26.s, z5.h, z2.h\n" + ".inst 0x6460e42e // bfmmla z14.s, z1.h, z0.h\n" + "ld1h { z2.h }, p5/Z, [x9]\n" + ".inst 0x6460e476 // bfmmla z22.s, z3.h, z0.h\n" + ".inst 0x6460e4be // bfmmla z30.s, z5.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n" "addvl x9, x9, #2\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6467e473 // bfmmla z19.s, z3.h, z7.h\n" - ".inst 0x6467e4bb // bfmmla z27.s, z5.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" - ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" - ".inst 0x6466e4bf // bfmmla z31.s, z5.h, z6.h\n" + ".inst 0x6462e42b // bfmmla z11.s, z1.h, z2.h\n" + ".inst 0x6462e473 // bfmmla z19.s, z3.h, z2.h\n" + ".inst 0x6462e4bb // bfmmla z27.s, z5.h, z2.h\n" + ".inst 0x6460e42f // bfmmla z15.s, z1.h, z0.h\n" + ".inst 0x6460e477 // bfmmla z23.s, z3.h, z0.h\n" + ".inst 0x6460e4bf // bfmmla z31.s, z5.h, z0.h\n" "68:" // Height 5: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1653,49 +1653,49 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "uzp1 z27.d, z27.d, z31.d\n" "tbz %x[flags], #1, 69f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" - "add x20, %x[args_ptr], %[offset_min]\n" "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z7.s, p5/M, z7.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 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 z15.s, p5/M, z15.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 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 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 z7.s, p5/M, z7.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 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 z15.s, p5/M, z15.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 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 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" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z23.s }, p5/Z, [x20]\n" + "fmin z7.s, p5/M, z7.s, z0.s\n" + "fmin z12.s, p5/M, z12.s, z0.s\n" + "fmin z13.s, p5/M, z13.s, z0.s\n" + "fmin z14.s, p5/M, z14.s, z0.s\n" + "fmin z8.s, p5/M, z8.s, z0.s\n" + "fmin z9.s, p5/M, z9.s, z0.s\n" + "fmin z10.s, p5/M, z10.s, z0.s\n" + "fmin z11.s, p5/M, z11.s, z0.s\n" + "fmin z15.s, p5/M, z15.s, z0.s\n" + "fmin z20.s, p5/M, z20.s, z0.s\n" + "fmin z21.s, p5/M, z21.s, z0.s\n" + "fmin z22.s, p5/M, z22.s, z0.s\n" + "fmin z16.s, p5/M, z16.s, z0.s\n" + "fmin z17.s, p5/M, z17.s, z0.s\n" + "fmin z18.s, p5/M, z18.s, z0.s\n" + "fmin z19.s, p5/M, z19.s, z0.s\n" + "fmin z24.s, p5/M, z24.s, z0.s\n" + "fmin z25.s, p5/M, z25.s, z0.s\n" + "fmin z26.s, p5/M, z26.s, z0.s\n" + "fmin z27.s, p5/M, z27.s, z0.s\n" + "fmax z7.s, p5/M, z7.s, z23.s\n" + "fmax z12.s, p5/M, z12.s, z23.s\n" + "fmax z13.s, p5/M, z13.s, z23.s\n" + "fmax z14.s, p5/M, z14.s, z23.s\n" + "fmax z8.s, p5/M, z8.s, z23.s\n" + "fmax z9.s, p5/M, z9.s, z23.s\n" + "fmax z10.s, p5/M, z10.s, z23.s\n" + "fmax z11.s, p5/M, z11.s, z23.s\n" + "fmax z15.s, p5/M, z15.s, z23.s\n" + "fmax z20.s, p5/M, z20.s, z23.s\n" + "fmax z21.s, p5/M, z21.s, z23.s\n" + "fmax z22.s, p5/M, z22.s, z23.s\n" + "fmax z16.s, p5/M, z16.s, z23.s\n" + "fmax z17.s, p5/M, z17.s, z23.s\n" + "fmax z18.s, p5/M, z18.s, z23.s\n" + "fmax z19.s, p5/M, z19.s, z23.s\n" + "fmax z24.s, p5/M, z24.s, z23.s\n" + "fmax z25.s, p5/M, z25.s, z23.s\n" + "fmax z26.s, p5/M, z26.s, z23.s\n" + "fmax z27.s, p5/M, z27.s, z23.s\n" "69:" // Height 5: No activation "st1w { z7.s }, p4, [x13]\n" "st1w { z12.s }, p3, [x13, #1, MUL VL]\n" @@ -1795,59 +1795,59 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "74:" // Height 6: no bias "tbz %x[flags], #0, 75f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x13]\n" + "add x24, x13, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" + "ld1w { z17.s }, p4/Z, [x13]\n" "add x22, x23, x20, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x13, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x13, #2, MUL VL]\n" "add x21, x22, x20, LSL #2\n" + "ld1w { z18.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z20.s }, p2/Z, [x13, #2, MUL VL]\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip2 z12.d, z9.d, z12.d\n" - "zip1 z9.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x24]\n" - "zip2 z13.d, z10.d, z13.d\n" - "zip1 z10.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" - "zip2 z14.d, z11.d, z14.d\n" + "ld1w { z12.s }, p4/Z, [x24]\n" + "zip1 z8.d, z17.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" + "zip2 z12.d, z17.d, z12.d\n" + "zip1 z9.d, z18.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x23]\n" + "zip2 z13.d, z18.d, z13.d\n" + "zip1 z10.d, z20.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" + "zip2 z14.d, z20.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" + "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x22]\n" "zip2 z15.d, z16.d, z15.d\n" "zip1 z16.d, z17.d, z20.d\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n" "zip2 z20.d, z17.d, z20.d\n" "zip1 z17.d, z18.d, z21.d\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z25.s }, p4/Z, [x22]\n" + "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z25.s }, p4/Z, [x21]\n" "zip2 z21.d, z18.d, z21.d\n" "zip1 z18.d, z19.d, z22.d\n" - "ld1w { z26.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z27.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n" "zip2 z22.d, z19.d, z22.d\n" "zip1 z19.d, z24.d, z23.d\n" - "ld1w { z6.s }, p1/Z, [x22, #3, MUL VL]\n" - "ld1w { z28.s }, p4/Z, [x21]\n" + "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z28.s }, p4/Z, [x20]\n" "zip2 z23.d, z24.d, z23.d\n" "zip1 z24.d, z25.d, z28.d\n" - "ld1w { z29.s }, p3/Z, [x21, #1, MUL VL]\n" - "ld1w { z30.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n" "zip2 z28.d, z25.d, z28.d\n" "zip1 z25.d, z26.d, z29.d\n" - "ld1w { z31.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n" "zip2 z29.d, z26.d, z29.d\n" "zip1 z26.d, z27.d, z30.d\n" "zip2 z30.d, z27.d, z30.d\n" - "zip1 z27.d, z6.d, z31.d\n" - "zip2 z31.d, z6.d, z31.d\n" + "zip1 z27.d, z0.d, z31.d\n" + "zip2 z31.d, z0.d, z31.d\n" "b 76f\n" "75:" // Height 6: no accumulate "mov z8.b, #0x0\n" @@ -1879,16 +1879,16 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "77:" // 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, 78f\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, 79f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -1900,193 +1900,193 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL ( "b 79f\n" "78:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" - "add x21, x22, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" + "add x22, x23, x21, LSL #1\n" + "add x21, x22, x21, LSL #1\n" "79:" // Height 6: input setup done "cmp x27, #0x8\n" "ble 81f\n" "80:" // Height 6: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x24]\n" - "ld1rqh { z4.h }, p0/Z, [x23]\n" - "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1rqh { z5.h }, p0/Z, [x22]\n" - "ld1rqh { z6.h }, p0/Z, [x21]\n" - "trn2 z3.d, z3.d, z4.d\n" - "trn1 z4.d, z5.d, z6.d\n" - "trn2 z5.d, z5.d, z6.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - ".inst 0x6467e498 // bfmmla z24.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" + "ld1rqh { z7.h }, p0/Z, [x26]\n" + "ld1rqh { z0.h }, p0/Z, [x25]\n" + "trn1 z6.d, z7.d, z0.d\n" + "ld1rqh { z5.h }, p0/Z, [x24]\n" + "ld1rqh { z1.h }, p0/Z, [x23]\n" + "trn2 z7.d, z7.d, z0.d\n" + "trn1 z4.d, z5.d, z1.d\n" + "ld1rqh { z3.h }, p0/Z, [x22]\n" + "ld1rqh { z0.h }, p0/Z, [x21]\n" + "trn2 z5.d, z5.d, z1.d\n" + "trn1 z2.d, z3.d, z0.d\n" + "trn2 z3.d, z3.d, z0.d\n" + "ld1h { z1.h }, p5/Z, [x12]\n" + "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6461e4c8 // bfmmla z8.s, z6.h, z1.h\n" + ".inst 0x6461e490 // bfmmla z16.s, z4.h, z1.h\n" + ".inst 0x6461e458 // bfmmla z24.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x11]\n" "sub x27, x27, #0x8\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + ".inst 0x6460e4cc // bfmmla z12.s, z6.h, z0.h\n" + ".inst 0x6460e494 // bfmmla z20.s, z4.h, z0.h\n" "cmp x27, #0x8\n" "add x26, x26, #0x10\n" - ".inst 0x6466e49c // bfmmla z28.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" + ".inst 0x6460e45c // bfmmla z28.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6461e4c9 // bfmmla z9.s, z6.h, z1.h\n" "add x25, x25, #0x10\n" - ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" - ".inst 0x6467e499 // bfmmla z25.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" + ".inst 0x6461e491 // bfmmla z17.s, z4.h, z1.h\n" + ".inst 0x6461e459 // bfmmla z25.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10]\n" "add x24, x24, #0x10\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" + ".inst 0x6460e4cd // bfmmla z13.s, z6.h, z0.h\n" + ".inst 0x6460e495 // bfmmla z21.s, z4.h, z0.h\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - ".inst 0x6466e49d // bfmmla z29.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" + ".inst 0x6460e45d // bfmmla z29.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6461e4ca // bfmmla z10.s, z6.h, z1.h\n" "add x21, x21, #0x10\n" - ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" - ".inst 0x6467e49a // bfmmla z26.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" - ".inst 0x6466e49e // bfmmla z30.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" - ".inst 0x6467e49b // bfmmla z27.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x12, #2, MUL VL]\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - ".inst 0x6466e49f // bfmmla z31.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x12, #3, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" + ".inst 0x6461e492 // bfmmla z18.s, z4.h, z1.h\n" + ".inst 0x6461e45a // bfmmla z26.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x9]\n" + ".inst 0x6460e4ce // bfmmla z14.s, z6.h, z0.h\n" + ".inst 0x6460e496 // bfmmla z22.s, z4.h, z0.h\n" + ".inst 0x6460e45e // bfmmla z30.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6461e4cb // bfmmla z11.s, z6.h, z1.h\n" + ".inst 0x6461e493 // bfmmla z19.s, z4.h, z1.h\n" + ".inst 0x6461e45b // bfmmla z27.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x12, #2, MUL VL]\n" + ".inst 0x6460e4cf // bfmmla z15.s, z6.h, z0.h\n" + ".inst 0x6460e497 // bfmmla z23.s, z4.h, z0.h\n" + ".inst 0x6460e45f // bfmmla z31.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x12, #3, MUL VL]\n" + ".inst 0x6461e4e8 // bfmmla z8.s, z7.h, z1.h\n" "addvl x12, x12, #4\n" - ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - ".inst 0x6467e4b8 // bfmmla z24.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x11, #2, MUL VL]\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" - ".inst 0x6466e4bc // bfmmla z28.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #3, MUL VL]\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" + ".inst 0x6461e4b0 // bfmmla z16.s, z5.h, z1.h\n" + ".inst 0x6461e478 // bfmmla z24.s, z3.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x11, #2, MUL VL]\n" + ".inst 0x6460e4ec // bfmmla z12.s, z7.h, z0.h\n" + ".inst 0x6460e4b4 // bfmmla z20.s, z5.h, z0.h\n" + ".inst 0x6460e47c // bfmmla z28.s, z3.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x11, #3, MUL VL]\n" + ".inst 0x6461e4e9 // bfmmla z9.s, z7.h, z1.h\n" "addvl x11, x11, #4\n" - ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" - ".inst 0x6467e4b9 // bfmmla z25.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" - ".inst 0x6466e4bd // bfmmla z29.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" + ".inst 0x6461e4b1 // bfmmla z17.s, z5.h, z1.h\n" + ".inst 0x6461e479 // bfmmla z25.s, z3.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6460e4ed // bfmmla z13.s, z7.h, z0.h\n" + ".inst 0x6460e4b5 // bfmmla z21.s, z5.h, z0.h\n" + ".inst 0x6460e47d // bfmmla z29.s, z3.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6461e4ea // bfmmla z10.s, z7.h, z1.h\n" "addvl x10, x10, #4\n" - ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" - ".inst 0x6467e4ba // bfmmla z26.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9, #2, MUL VL]\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" - ".inst 0x6466e4be // bfmmla z30.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #3, MUL VL]\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" + ".inst 0x6461e4b2 // bfmmla z18.s, z5.h, z1.h\n" + ".inst 0x6461e47a // bfmmla z26.s, z3.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x9, #2, MUL VL]\n" + ".inst 0x6460e4ee // bfmmla z14.s, z7.h, z0.h\n" + ".inst 0x6460e4b6 // bfmmla z22.s, z5.h, z0.h\n" + ".inst 0x6460e47e // bfmmla z30.s, z3.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x9, #3, MUL VL]\n" + ".inst 0x6461e4eb // bfmmla z11.s, z7.h, z1.h\n" "addvl x9, x9, #4\n" - ".inst 0x6467e473 // bfmmla z19.s, z3.h, z7.h\n" - ".inst 0x6467e4bb // bfmmla z27.s, z5.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" - ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" - ".inst 0x6466e4bf // bfmmla z31.s, z5.h, z6.h\n" + ".inst 0x6461e4b3 // bfmmla z19.s, z5.h, z1.h\n" + ".inst 0x6461e47b // bfmmla z27.s, z3.h, z1.h\n" + ".inst 0x6460e4ef // bfmmla z15.s, z7.h, z0.h\n" + ".inst 0x6460e4b7 // bfmmla z23.s, z5.h, z0.h\n" + ".inst 0x6460e47f // bfmmla z31.s, z3.h, z0.h\n" "bgt 80b\n" "81:" // Height 6: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" "ld1rqh { z1.h }, p0/Z, [x26]\n" - "ld1rqh { z2.h }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" + "ld1rqh { z0.h }, p0/Z, [x25]\n" + "trn1 z7.d, z1.d, z0.d\n" "ld1rqh { z3.h }, p0/Z, [x24]\n" - "ld1rqh { z4.h }, p0/Z, [x23]\n" - "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z4.d\n" + "ld1rqh { z2.h }, p0/Z, [x23]\n" + "trn2 z1.d, z1.d, z0.d\n" + "trn1 z6.d, z3.d, z2.d\n" "ld1rqh { z5.h }, p0/Z, [x22]\n" - "ld1rqh { z6.h }, p0/Z, [x21]\n" - "trn2 z3.d, z3.d, z4.d\n" - "trn1 z4.d, z5.d, z6.d\n" - "trn2 z5.d, z5.d, z6.d\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - ".inst 0x6467e498 // bfmmla z24.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" + "ld1rqh { z0.h }, p0/Z, [x21]\n" + "trn2 z3.d, z3.d, z2.d\n" + "trn1 z4.d, z5.d, z0.d\n" + "trn2 z5.d, z5.d, z0.d\n" + "ld1h { z2.h }, p5/Z, [x12]\n" + "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6462e4e8 // bfmmla z8.s, z7.h, z2.h\n" + ".inst 0x6462e4d0 // bfmmla z16.s, z6.h, z2.h\n" + ".inst 0x6462e498 // bfmmla z24.s, z4.h, z2.h\n" + "ld1h { z2.h }, p5/Z, [x11]\n" "subs x27, x27, #0x4\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + ".inst 0x6460e4ec // bfmmla z12.s, z7.h, z0.h\n" + ".inst 0x6460e4d4 // bfmmla z20.s, z6.h, z0.h\n" "addvl x12, x12, #2\n" - ".inst 0x6466e49c // bfmmla z28.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" - ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" + ".inst 0x6460e49c // bfmmla z28.s, z4.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6462e4e9 // bfmmla z9.s, z7.h, z2.h\n" "addvl x11, x11, #2\n" - ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" - ".inst 0x6467e499 // bfmmla z25.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - ".inst 0x6466e49d // bfmmla z29.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" + ".inst 0x6462e4d1 // bfmmla z17.s, z6.h, z2.h\n" + ".inst 0x6462e499 // bfmmla z25.s, z4.h, z2.h\n" + "ld1h { z2.h }, p5/Z, [x10]\n" + ".inst 0x6460e4ed // bfmmla z13.s, z7.h, z0.h\n" + ".inst 0x6460e4d5 // bfmmla z21.s, z6.h, z0.h\n" + ".inst 0x6460e49d // bfmmla z29.s, z4.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6462e4ea // bfmmla z10.s, z7.h, z2.h\n" "addvl x10, x10, #2\n" - ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" - ".inst 0x6467e49a // bfmmla z26.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" - ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" - ".inst 0x6466e49e // bfmmla z30.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" - ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" + ".inst 0x6462e4d2 // bfmmla z18.s, z6.h, z2.h\n" + ".inst 0x6462e49a // bfmmla z26.s, z4.h, z2.h\n" + "ld1h { z2.h }, p5/Z, [x9]\n" + ".inst 0x6460e4ee // bfmmla z14.s, z7.h, z0.h\n" + ".inst 0x6460e4d6 // bfmmla z22.s, z6.h, z0.h\n" + ".inst 0x6460e49e // bfmmla z30.s, z4.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6462e4eb // bfmmla z11.s, z7.h, z2.h\n" "addvl x9, x9, #2\n" - ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" - ".inst 0x6467e49b // bfmmla z27.s, z4.h, z7.h\n" - ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - ".inst 0x6466e49f // bfmmla z31.s, z4.h, z6.h\n" + ".inst 0x6462e4d3 // bfmmla z19.s, z6.h, z2.h\n" + ".inst 0x6462e49b // bfmmla z27.s, z4.h, z2.h\n" + ".inst 0x6460e4ef // bfmmla z15.s, z7.h, z0.h\n" + ".inst 0x6460e4d7 // bfmmla z23.s, z6.h, z0.h\n" + ".inst 0x6460e49f // bfmmla z31.s, z4.h, z0.h\n" "ble 82f\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x12, #1, MUL VL]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - ".inst 0x6467e4b8 // bfmmla z24.s, z5.h, z7.h\n" - ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x11]\n" + "ld1h { z2.h }, p5/Z, [x12]\n" + "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n" + ".inst 0x6462e428 // bfmmla z8.s, z1.h, z2.h\n" + ".inst 0x6462e470 // bfmmla z16.s, z3.h, z2.h\n" + ".inst 0x6462e4b8 // bfmmla z24.s, z5.h, z2.h\n" + ".inst 0x6460e42c // bfmmla z12.s, z1.h, z0.h\n" + "ld1h { z2.h }, p5/Z, [x11]\n" "addvl x12, x12, #2\n" - ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" - ".inst 0x6466e4bc // bfmmla z28.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x11, #1, MUL VL]\n" + ".inst 0x6460e474 // bfmmla z20.s, z3.h, z0.h\n" + ".inst 0x6460e4bc // bfmmla z28.s, z5.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n" "addvl x11, x11, #2\n" - ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" - ".inst 0x6467e4b9 // bfmmla z25.s, z5.h, z7.h\n" - ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" - ".inst 0x6466e4bd // bfmmla z29.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6462e429 // bfmmla z9.s, z1.h, z2.h\n" + ".inst 0x6462e471 // bfmmla z17.s, z3.h, z2.h\n" + ".inst 0x6462e4b9 // bfmmla z25.s, z5.h, z2.h\n" + ".inst 0x6460e42d // bfmmla z13.s, z1.h, z0.h\n" + "ld1h { z2.h }, p5/Z, [x10]\n" + ".inst 0x6460e475 // bfmmla z21.s, z3.h, z0.h\n" + ".inst 0x6460e4bd // bfmmla z29.s, z5.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" "addvl x10, x10, #2\n" - ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" - ".inst 0x6467e4ba // bfmmla z26.s, z5.h, z7.h\n" - ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" - "ld1h { z7.h }, p5/Z, [x9]\n" - ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" - ".inst 0x6466e4be // bfmmla z30.s, z5.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x9, #1, MUL VL]\n" + ".inst 0x6462e42a // bfmmla z10.s, z1.h, z2.h\n" + ".inst 0x6462e472 // bfmmla z18.s, z3.h, z2.h\n" + ".inst 0x6462e4ba // bfmmla z26.s, z5.h, z2.h\n" + ".inst 0x6460e42e // bfmmla z14.s, z1.h, z0.h\n" + "ld1h { z2.h }, p5/Z, [x9]\n" + ".inst 0x6460e476 // bfmmla z22.s, z3.h, z0.h\n" + ".inst 0x6460e4be // bfmmla z30.s, z5.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n" "addvl x9, x9, #2\n" - ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" - ".inst 0x6467e473 // bfmmla z19.s, z3.h, z7.h\n" - ".inst 0x6467e4bb // bfmmla z27.s, z5.h, z7.h\n" - ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" - ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" - ".inst 0x6466e4bf // bfmmla z31.s, z5.h, z6.h\n" + ".inst 0x6462e42b // bfmmla z11.s, z1.h, z2.h\n" + ".inst 0x6462e473 // bfmmla z19.s, z3.h, z2.h\n" + ".inst 0x6462e4bb // bfmmla z27.s, z5.h, z2.h\n" + ".inst 0x6460e42f // bfmmla z15.s, z1.h, z0.h\n" + ".inst 0x6460e477 // bfmmla z23.s, z3.h, z0.h\n" + ".inst 0x6460e4bf // bfmmla z31.s, z5.h, z0.h\n" "82:" // Height 6: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" |