diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp | 1033 |
1 files changed, 516 insertions, 517 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp index 8d05c1ffb3..2b2a0684f9 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp @@ -140,22 +140,22 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "b 5f\n" "3:" // Height 1: no bias "tbz %x[flags], #0, 4f\n" - "ld1w { z9.s }, p6/Z, [x27]\n" - "ld1w { z10.s }, p5/Z, [x27, #1, MUL VL]\n" - "zip1 z8.d, z9.d, z14.d\n" - "zip2 z14.d, z9.d, z14.d\n" - "ld1w { z11.s }, p4/Z, [x27, #2, MUL VL]\n" - "ld1w { z12.s }, p3/Z, [x27, #3, MUL VL]\n" - "zip1 z9.d, z10.d, z15.d\n" - "zip2 z15.d, z10.d, z15.d\n" - "ld1w { z13.s }, p2/Z, [x27, #4, MUL VL]\n" + "ld1w { z21.s }, p6/Z, [x27]\n" + "ld1w { z20.s }, p5/Z, [x27, #1, MUL VL]\n" + "zip1 z8.d, z21.d, z14.d\n" + "zip2 z14.d, z21.d, z14.d\n" + "ld1w { z23.s }, p4/Z, [x27, #2, MUL VL]\n" + "ld1w { z22.s }, p3/Z, [x27, #3, MUL VL]\n" + "zip1 z9.d, z20.d, z15.d\n" + "zip2 z15.d, z20.d, z15.d\n" + "ld1w { z21.s }, p2/Z, [x27, #4, MUL VL]\n" "ld1w { z20.s }, p1/Z, [x27, #5, MUL VL]\n" - "zip1 z10.d, z11.d, z16.d\n" - "zip2 z16.d, z11.d, z16.d\n" - "zip1 z11.d, z12.d, z17.d\n" - "zip2 z17.d, z12.d, z17.d\n" - "zip1 z12.d, z13.d, z18.d\n" - "zip2 z18.d, z13.d, z18.d\n" + "zip1 z10.d, z23.d, z16.d\n" + "zip2 z16.d, z23.d, z16.d\n" + "zip1 z11.d, z22.d, z17.d\n" + "zip2 z17.d, z22.d, z17.d\n" + "zip1 z12.d, z21.d, z18.d\n" + "zip2 z18.d, z21.d, z18.d\n" "zip1 z13.d, z20.d, z19.d\n" "zip2 z19.d, z20.d, z19.d\n" "b 5f\n" @@ -177,11 +177,11 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "6:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 7f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" "cbnz x26, 8f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -193,69 +193,69 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "ble 10f\n" "9:" // Height 1: Multiply loop: Main loop head "whilelt p0.s, XZR, x25\n" - "ld1rqw { z0.s }, p0/Z, [x24]\n" - ".inst 0x658abc00 // bfcvt z0.h, p7/M, z0.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z4.h }, p7/Z, [x28]\n" - "ld1h { z5.h }, p7/Z, [x28, #1, MUL VL]\n" - ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n" - ".inst 0x6465e40e // bfmmla z14.s, z0.h, z5.h\n" - "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - "ld1h { z4.h }, p7/Z, [x28, #4, MUL VL]\n" - "ld1h { z5.h }, p7/Z, [x28, #5, MUL VL]\n" - ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n" - ".inst 0x6465e410 // bfmmla z16.s, z0.h, z5.h\n" - "ld1h { z6.h }, p7/Z, [x28, #6, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #7, MUL VL]\n" + "ld1rqw { z24.s }, p0/Z, [x24]\n" + ".inst 0x658abf18 // bfcvt z24.h, p7/M, z24.s\n" + "uzp1 z24.h, z24.h, z24.h\n" + "ld1h { z21.h }, p7/Z, [x28]\n" + "ld1h { z20.h }, p7/Z, [x28, #1, MUL VL]\n" + ".inst 0x6475e708 // bfmmla z8.s, z24.h, z21.h\n" + ".inst 0x6474e70e // bfmmla z14.s, z24.h, z20.h\n" + "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #3, MUL VL]\n" + ".inst 0x6475e709 // bfmmla z9.s, z24.h, z21.h\n" + ".inst 0x6474e70f // bfmmla z15.s, z24.h, z20.h\n" + "ld1h { z21.h }, p7/Z, [x28, #4, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #5, MUL VL]\n" + ".inst 0x6475e70a // bfmmla z10.s, z24.h, z21.h\n" + ".inst 0x6474e710 // bfmmla z16.s, z24.h, z20.h\n" + "ld1h { z21.h }, p7/Z, [x28, #6, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - ".inst 0x6467e411 // bfmmla z17.s, z0.h, z7.h\n" - "ld1h { z4.h }, p7/Z, [x28, #-8, MUL VL]\n" - "ld1h { z5.h }, p7/Z, [x28, #-7, MUL VL]\n" + ".inst 0x6475e70b // bfmmla z11.s, z24.h, z21.h\n" + ".inst 0x6474e711 // bfmmla z17.s, z24.h, z20.h\n" + "ld1h { z23.h }, p7/Z, [x28, #-8, MUL VL]\n" + "ld1h { z22.h }, p7/Z, [x28, #-7, MUL VL]\n" "sub x25, x25, #0x4\n" - "ld1h { z6.h }, p7/Z, [x28, #-6, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #-5, MUL VL]\n" + "ld1h { z21.h }, p7/Z, [x28, #-6, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #-5, MUL VL]\n" "cmp x25, #0x4\n" - ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n" - ".inst 0x6465e412 // bfmmla z18.s, z0.h, z5.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" + ".inst 0x6477e70c // bfmmla z12.s, z24.h, z23.h\n" + ".inst 0x6476e712 // bfmmla z18.s, z24.h, z22.h\n" + ".inst 0x6475e70d // bfmmla z13.s, z24.h, z21.h\n" "add x24, x24, #0x10\n" "addvl x28, x28, #-4\n" - ".inst 0x6467e413 // bfmmla z19.s, z0.h, z7.h\n" + ".inst 0x6474e713 // bfmmla z19.s, z24.h, z20.h\n" "bgt 9b\n" "10:" // Height 1: Multiply loop: Single iteration only "whilelt p0.s, XZR, x25\n" - "ld1rqw { z0.s }, p0/Z, [x24]\n" - ".inst 0x658abc00 // bfcvt z0.h, p7/M, z0.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z4.h }, p7/Z, [x28]\n" - "ld1h { z5.h }, p7/Z, [x28, #1, MUL VL]\n" - ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n" - ".inst 0x6465e40e // bfmmla z14.s, z0.h, z5.h\n" - "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - "ld1h { z4.h }, p7/Z, [x28, #4, MUL VL]\n" - "ld1h { z5.h }, p7/Z, [x28, #5, MUL VL]\n" - ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n" - ".inst 0x6465e410 // bfmmla z16.s, z0.h, z5.h\n" - "ld1h { z6.h }, p7/Z, [x28, #6, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #7, MUL VL]\n" + "ld1rqw { z23.s }, p0/Z, [x24]\n" + ".inst 0x658abef7 // bfcvt z23.h, p7/M, z23.s\n" + "uzp1 z23.h, z23.h, z23.h\n" + "ld1h { z21.h }, p7/Z, [x28]\n" + "ld1h { z20.h }, p7/Z, [x28, #1, MUL VL]\n" + ".inst 0x6475e6e8 // bfmmla z8.s, z23.h, z21.h\n" + ".inst 0x6474e6ee // bfmmla z14.s, z23.h, z20.h\n" + "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #3, MUL VL]\n" + ".inst 0x6475e6e9 // bfmmla z9.s, z23.h, z21.h\n" + ".inst 0x6474e6ef // bfmmla z15.s, z23.h, z20.h\n" + "ld1h { z21.h }, p7/Z, [x28, #4, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #5, MUL VL]\n" + ".inst 0x6475e6ea // bfmmla z10.s, z23.h, z21.h\n" + ".inst 0x6474e6f0 // bfmmla z16.s, z23.h, z20.h\n" + "ld1h { z21.h }, p7/Z, [x28, #6, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - ".inst 0x6467e411 // bfmmla z17.s, z0.h, z7.h\n" - "ld1h { z4.h }, p7/Z, [x28, #-8, MUL VL]\n" - "ld1h { z5.h }, p7/Z, [x28, #-7, MUL VL]\n" - ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n" - "ld1h { z6.h }, p7/Z, [x28, #-6, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #-5, MUL VL]\n" - ".inst 0x6465e412 // bfmmla z18.s, z0.h, z5.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6467e413 // bfmmla z19.s, z0.h, z7.h\n" + ".inst 0x6475e6eb // bfmmla z11.s, z23.h, z21.h\n" + ".inst 0x6474e6f1 // bfmmla z17.s, z23.h, z20.h\n" + "ld1h { z20.h }, p7/Z, [x28, #-8, MUL VL]\n" + "ld1h { z22.h }, p7/Z, [x28, #-7, MUL VL]\n" + ".inst 0x6474e6ec // bfmmla z12.s, z23.h, z20.h\n" + "ld1h { z21.h }, p7/Z, [x28, #-6, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #-5, MUL VL]\n" + ".inst 0x6476e6f2 // bfmmla z18.s, z23.h, z22.h\n" + ".inst 0x6475e6ed // bfmmla z13.s, z23.h, z21.h\n" + ".inst 0x6474e6f3 // bfmmla z19.s, z23.h, z20.h\n" "addvl x28, x28, #-4\n" "11:" // Height 1: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -270,21 +270,21 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "uzp1 z13.d, z13.d, z19.d\n" "tbz %x[flags], #1, 12f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p7/Z, [x20]\n" + "ld1rw { z21.s }, p7/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p7/Z, [x20]\n" - "fmin z8.s, p7/M, z8.s, z1.s\n" - "fmin z9.s, p7/M, z9.s, z1.s\n" - "fmin z10.s, p7/M, z10.s, z1.s\n" - "fmin z11.s, p7/M, z11.s, z1.s\n" - "fmin z12.s, p7/M, z12.s, z1.s\n" - "fmin z13.s, p7/M, z13.s, z1.s\n" - "fmax z8.s, p7/M, z8.s, z0.s\n" - "fmax z9.s, p7/M, z9.s, z0.s\n" - "fmax z10.s, p7/M, z10.s, z0.s\n" - "fmax z11.s, p7/M, z11.s, z0.s\n" - "fmax z12.s, p7/M, z12.s, z0.s\n" - "fmax z13.s, p7/M, z13.s, z0.s\n" + "ld1rw { z20.s }, p7/Z, [x20]\n" + "fmin z8.s, p7/M, z8.s, z21.s\n" + "fmin z9.s, p7/M, z9.s, z21.s\n" + "fmin z10.s, p7/M, z10.s, z21.s\n" + "fmin z11.s, p7/M, z11.s, z21.s\n" + "fmin z12.s, p7/M, z12.s, z21.s\n" + "fmin z13.s, p7/M, z13.s, z21.s\n" + "fmax z8.s, p7/M, z8.s, z20.s\n" + "fmax z9.s, p7/M, z9.s, z20.s\n" + "fmax z10.s, p7/M, z10.s, z20.s\n" + "fmax z11.s, p7/M, z11.s, z20.s\n" + "fmax z12.s, p7/M, z12.s, z20.s\n" + "fmax z13.s, p7/M, z13.s, z20.s\n" "12:" // Height 1: No activation "st1w { z8.s }, p6, [x27]\n" "st1w { z9.s }, p5, [x27, #1, MUL VL]\n" @@ -340,29 +340,29 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "16:" // Height 2: no bias "tbz %x[flags], #0, 17f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x27, x20, LSL #2\n" - "ld1w { z9.s }, p6/Z, [x27]\n" - "ld1w { z10.s }, p5/Z, [x27, #1, MUL VL]\n" - "ld1w { z11.s }, p4/Z, [x27, #2, MUL VL]\n" - "ld1w { z12.s }, p3/Z, [x27, #3, MUL VL]\n" - "ld1w { z13.s }, p2/Z, [x27, #4, MUL VL]\n" + "add x20, x27, x20, LSL #2\n" + "ld1w { z16.s }, p6/Z, [x27]\n" + "ld1w { z17.s }, p5/Z, [x27, #1, MUL VL]\n" + "ld1w { z19.s }, p4/Z, [x27, #2, MUL VL]\n" + "ld1w { z22.s }, p3/Z, [x27, #3, MUL VL]\n" + "ld1w { z21.s }, p2/Z, [x27, #4, MUL VL]\n" "ld1w { z20.s }, p1/Z, [x27, #5, MUL VL]\n" - "ld1w { z14.s }, p6/Z, [x23]\n" - "zip1 z8.d, z9.d, z14.d\n" - "zip2 z14.d, z9.d, z14.d\n" - "ld1w { z15.s }, p5/Z, [x23, #1, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x23, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z15.d\n" - "zip2 z15.d, z10.d, z15.d\n" - "ld1w { z17.s }, p3/Z, [x23, #3, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x23, #4, MUL VL]\n" - "zip1 z10.d, z11.d, z16.d\n" - "zip2 z16.d, z11.d, z16.d\n" - "ld1w { z19.s }, p1/Z, [x23, #5, MUL VL]\n" - "zip1 z11.d, z12.d, z17.d\n" - "zip2 z17.d, z12.d, z17.d\n" - "zip1 z12.d, z13.d, z18.d\n" - "zip2 z18.d, z13.d, z18.d\n" + "ld1w { z14.s }, p6/Z, [x20]\n" + "zip1 z8.d, z16.d, z14.d\n" + "zip2 z14.d, z16.d, z14.d\n" + "ld1w { z15.s }, p5/Z, [x20, #1, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x20, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z15.d\n" + "zip2 z15.d, z17.d, z15.d\n" + "ld1w { z17.s }, p3/Z, [x20, #3, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x20, #4, MUL VL]\n" + "zip1 z10.d, z19.d, z16.d\n" + "zip2 z16.d, z19.d, z16.d\n" + "ld1w { z19.s }, p1/Z, [x20, #5, MUL VL]\n" + "zip1 z11.d, z22.d, z17.d\n" + "zip2 z17.d, z22.d, z17.d\n" + "zip1 z12.d, z21.d, z18.d\n" + "zip2 z18.d, z21.d, z18.d\n" "zip1 z13.d, z20.d, z19.d\n" "zip2 z19.d, z20.d, z19.d\n" "b 18f\n" @@ -384,12 +384,12 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "19:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 20f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" "cbnz x26, 21f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -397,85 +397,85 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "b 21f\n" "20:" // Height 2: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "21:" // Height 2: input setup done "cmp x25, #0x4\n" "ble 23f\n" "22:" // Height 2: Multiply loop: Main loop head "whilelt p0.s, XZR, x25\n" - "ld1rqw { z0.s }, p0/Z, [x24]\n" - "ld1rqw { z1.s }, p0/Z, [x23]\n" - ".inst 0x658abc00 // bfcvt z0.h, p7/M, z0.s\n" - ".inst 0x658abc21 // bfcvt z1.h, p7/M, z1.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z4.h }, p7/Z, [x28]\n" - "ld1h { z5.h }, p7/Z, [x28, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "trn1 z0.d, z0.d, z1.d\n" - "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #3, MUL VL]\n" - ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n" - ".inst 0x6465e40e // bfmmla z14.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #4, MUL VL]\n" - "ld1h { z5.h }, p7/Z, [x28, #5, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #6, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #7, MUL VL]\n" + "ld1rqw { z24.s }, p0/Z, [x24]\n" + "ld1rqw { z20.s }, p0/Z, [x23]\n" + ".inst 0x658abf18 // bfcvt z24.h, p7/M, z24.s\n" + ".inst 0x658abe94 // bfcvt z20.h, p7/M, z20.s\n" + "uzp1 z24.h, z24.h, z24.h\n" + "ld1h { z23.h }, p7/Z, [x28]\n" + "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n" + "uzp1 z20.h, z20.h, z20.h\n" + "trn1 z24.d, z24.d, z20.d\n" + "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #3, MUL VL]\n" + ".inst 0x6477e708 // bfmmla z8.s, z24.h, z23.h\n" + ".inst 0x6476e70e // bfmmla z14.s, z24.h, z22.h\n" + "ld1h { z23.h }, p7/Z, [x28, #4, MUL VL]\n" + "ld1h { z22.h }, p7/Z, [x28, #5, MUL VL]\n" + ".inst 0x6475e709 // bfmmla z9.s, z24.h, z21.h\n" + ".inst 0x6474e70f // bfmmla z15.s, z24.h, z20.h\n" + "ld1h { z21.h }, p7/Z, [x28, #6, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n" - ".inst 0x6465e410 // bfmmla z16.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #-8, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - ".inst 0x6467e411 // bfmmla z17.s, z0.h, z7.h\n" - "ld1h { z5.h }, p7/Z, [x28, #-7, MUL VL]\n" - "ld1h { z6.h }, p7/Z, [x28, #-6, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #-5, MUL VL]\n" + ".inst 0x6477e70a // bfmmla z10.s, z24.h, z23.h\n" + ".inst 0x6476e710 // bfmmla z16.s, z24.h, z22.h\n" + "ld1h { z23.h }, p7/Z, [x28, #-8, MUL VL]\n" + ".inst 0x6475e70b // bfmmla z11.s, z24.h, z21.h\n" + ".inst 0x6474e711 // bfmmla z17.s, z24.h, z20.h\n" + "ld1h { z22.h }, p7/Z, [x28, #-7, MUL VL]\n" + "ld1h { z21.h }, p7/Z, [x28, #-6, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #-5, MUL VL]\n" "sub x25, x25, #0x4\n" "cmp x25, #0x4\n" - ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n" - ".inst 0x6465e412 // bfmmla z18.s, z0.h, z5.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" + ".inst 0x6477e70c // bfmmla z12.s, z24.h, z23.h\n" + ".inst 0x6476e712 // bfmmla z18.s, z24.h, z22.h\n" + ".inst 0x6475e70d // bfmmla z13.s, z24.h, z21.h\n" "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" - ".inst 0x6467e413 // bfmmla z19.s, z0.h, z7.h\n" + ".inst 0x6474e713 // bfmmla z19.s, z24.h, z20.h\n" "addvl x28, x28, #-4\n" "bgt 22b\n" "23:" // Height 2: Multiply loop: Single iteration only "whilelt p0.s, XZR, x25\n" - "ld1rqw { z0.s }, p0/Z, [x24]\n" - "ld1rqw { z1.s }, p0/Z, [x23]\n" - ".inst 0x658abc00 // bfcvt z0.h, p7/M, z0.s\n" - ".inst 0x658abc21 // bfcvt z1.h, p7/M, z1.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z4.h }, p7/Z, [x28]\n" - "ld1h { z5.h }, p7/Z, [x28, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "trn1 z0.d, z0.d, z1.d\n" - "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #3, MUL VL]\n" - ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n" - ".inst 0x6465e40e // bfmmla z14.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #4, MUL VL]\n" - "ld1h { z5.h }, p7/Z, [x28, #5, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #6, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #7, MUL VL]\n" + "ld1rqw { z24.s }, p0/Z, [x24]\n" + "ld1rqw { z20.s }, p0/Z, [x23]\n" + ".inst 0x658abf18 // bfcvt z24.h, p7/M, z24.s\n" + ".inst 0x658abe94 // bfcvt z20.h, p7/M, z20.s\n" + "uzp1 z24.h, z24.h, z24.h\n" + "ld1h { z23.h }, p7/Z, [x28]\n" + "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n" + "uzp1 z20.h, z20.h, z20.h\n" + "trn1 z24.d, z24.d, z20.d\n" + "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #3, MUL VL]\n" + ".inst 0x6477e708 // bfmmla z8.s, z24.h, z23.h\n" + ".inst 0x6476e70e // bfmmla z14.s, z24.h, z22.h\n" + "ld1h { z23.h }, p7/Z, [x28, #4, MUL VL]\n" + "ld1h { z22.h }, p7/Z, [x28, #5, MUL VL]\n" + ".inst 0x6475e709 // bfmmla z9.s, z24.h, z21.h\n" + ".inst 0x6474e70f // bfmmla z15.s, z24.h, z20.h\n" + "ld1h { z21.h }, p7/Z, [x28, #6, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n" - ".inst 0x6465e410 // bfmmla z16.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #-8, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - ".inst 0x6467e411 // bfmmla z17.s, z0.h, z7.h\n" - "ld1h { z5.h }, p7/Z, [x28, #-7, MUL VL]\n" - "ld1h { z6.h }, p7/Z, [x28, #-6, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #-5, MUL VL]\n" - ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n" - ".inst 0x6465e412 // bfmmla z18.s, z0.h, z5.h\n" + ".inst 0x6477e70a // bfmmla z10.s, z24.h, z23.h\n" + ".inst 0x6476e710 // bfmmla z16.s, z24.h, z22.h\n" + "ld1h { z23.h }, p7/Z, [x28, #-8, MUL VL]\n" + ".inst 0x6475e70b // bfmmla z11.s, z24.h, z21.h\n" + ".inst 0x6474e711 // bfmmla z17.s, z24.h, z20.h\n" + "ld1h { z22.h }, p7/Z, [x28, #-7, MUL VL]\n" + "ld1h { z21.h }, p7/Z, [x28, #-6, MUL VL]\n" + "ld1h { z20.h }, p7/Z, [x28, #-5, MUL VL]\n" + ".inst 0x6477e70c // bfmmla z12.s, z24.h, z23.h\n" + ".inst 0x6476e712 // bfmmla z18.s, z24.h, z22.h\n" "addvl x28, x28, #-4\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6467e413 // bfmmla z19.s, z0.h, z7.h\n" + ".inst 0x6475e70d // bfmmla z13.s, z24.h, z21.h\n" + ".inst 0x6474e713 // bfmmla z19.s, z24.h, z20.h\n" "24:" // Height 2: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" @@ -497,33 +497,33 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "uzp2 z13.d, z13.d, z19.d\n" "tbz %x[flags], #1, 25f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p7/Z, [x20]\n" + "ld1rw { z20.s }, p7/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p7/Z, [x20]\n" - "fmin z4.s, p7/M, z4.s, z1.s\n" - "fmin z14.s, p7/M, z14.s, z1.s\n" - "fmin z15.s, p7/M, z15.s, z1.s\n" - "fmin z16.s, p7/M, z16.s, z1.s\n" - "fmin z17.s, p7/M, z17.s, z1.s\n" - "fmin z18.s, p7/M, z18.s, z1.s\n" - "fmin z8.s, p7/M, z8.s, z1.s\n" - "fmin z9.s, p7/M, z9.s, z1.s\n" - "fmin z10.s, p7/M, z10.s, z1.s\n" - "fmin z11.s, p7/M, z11.s, z1.s\n" - "fmin z12.s, p7/M, z12.s, z1.s\n" - "fmin z13.s, p7/M, z13.s, z1.s\n" - "fmax z4.s, p7/M, z4.s, z0.s\n" - "fmax z14.s, p7/M, z14.s, z0.s\n" - "fmax z15.s, p7/M, z15.s, z0.s\n" - "fmax z16.s, p7/M, z16.s, z0.s\n" - "fmax z17.s, p7/M, z17.s, z0.s\n" - "fmax z18.s, p7/M, z18.s, z0.s\n" - "fmax z8.s, p7/M, z8.s, z0.s\n" - "fmax z9.s, p7/M, z9.s, z0.s\n" - "fmax z10.s, p7/M, z10.s, z0.s\n" - "fmax z11.s, p7/M, z11.s, z0.s\n" - "fmax z12.s, p7/M, z12.s, z0.s\n" - "fmax z13.s, p7/M, z13.s, z0.s\n" + "ld1rw { z19.s }, p7/Z, [x20]\n" + "fmin z4.s, p7/M, z4.s, z20.s\n" + "fmin z14.s, p7/M, z14.s, z20.s\n" + "fmin z15.s, p7/M, z15.s, z20.s\n" + "fmin z16.s, p7/M, z16.s, z20.s\n" + "fmin z17.s, p7/M, z17.s, z20.s\n" + "fmin z18.s, p7/M, z18.s, z20.s\n" + "fmin z8.s, p7/M, z8.s, z20.s\n" + "fmin z9.s, p7/M, z9.s, z20.s\n" + "fmin z10.s, p7/M, z10.s, z20.s\n" + "fmin z11.s, p7/M, z11.s, z20.s\n" + "fmin z12.s, p7/M, z12.s, z20.s\n" + "fmin z13.s, p7/M, z13.s, z20.s\n" + "fmax z4.s, p7/M, z4.s, z19.s\n" + "fmax z14.s, p7/M, z14.s, z19.s\n" + "fmax z15.s, p7/M, z15.s, z19.s\n" + "fmax z16.s, p7/M, z16.s, z19.s\n" + "fmax z17.s, p7/M, z17.s, z19.s\n" + "fmax z18.s, p7/M, z18.s, z19.s\n" + "fmax z8.s, p7/M, z8.s, z19.s\n" + "fmax z9.s, p7/M, z9.s, z19.s\n" + "fmax z10.s, p7/M, z10.s, z19.s\n" + "fmax z11.s, p7/M, z11.s, z19.s\n" + "fmax z12.s, p7/M, z12.s, z19.s\n" + "fmax z13.s, p7/M, z13.s, z19.s\n" "25:" // Height 2: No activation "st1w { z4.s }, p6, [x27]\n" "st1w { z14.s }, p5, [x27, #1, MUL VL]\n" @@ -597,38 +597,38 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "29:" // Height 3: no bias "tbz %x[flags], #0, 30f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x27, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "ld1w { z9.s }, p6/Z, [x27]\n" - "ld1w { z10.s }, p5/Z, [x27, #1, MUL VL]\n" - "ld1w { z11.s }, p4/Z, [x27, #2, MUL VL]\n" - "ld1w { z12.s }, p3/Z, [x27, #3, MUL VL]\n" - "ld1w { z13.s }, p2/Z, [x27, #4, MUL VL]\n" + "add x21, x27, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z16.s }, p6/Z, [x27]\n" + "ld1w { z17.s }, p5/Z, [x27, #1, MUL VL]\n" + "ld1w { z19.s }, p4/Z, [x27, #2, MUL VL]\n" + "ld1w { z22.s }, p3/Z, [x27, #3, MUL VL]\n" + "ld1w { z24.s }, p2/Z, [x27, #4, MUL VL]\n" "ld1w { z20.s }, p1/Z, [x27, #5, MUL VL]\n" - "ld1w { z14.s }, p6/Z, [x23]\n" - "zip1 z8.d, z9.d, z14.d\n" - "zip2 z14.d, z9.d, z14.d\n" - "ld1w { z15.s }, p5/Z, [x23, #1, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x23, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z15.d\n" - "zip2 z15.d, z10.d, z15.d\n" - "ld1w { z17.s }, p3/Z, [x23, #3, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x23, #4, MUL VL]\n" - "zip1 z10.d, z11.d, z16.d\n" - "zip2 z16.d, z11.d, z16.d\n" - "ld1w { z19.s }, p1/Z, [x23, #5, MUL VL]\n" - "ld1w { z21.s }, p6/Z, [x22]\n" - "zip1 z11.d, z12.d, z17.d\n" - "zip2 z17.d, z12.d, z17.d\n" - "ld1w { z22.s }, p5/Z, [x22, #1, MUL VL]\n" - "ld1w { z23.s }, p4/Z, [x22, #2, MUL VL]\n" - "zip1 z12.d, z13.d, z18.d\n" - "zip2 z18.d, z13.d, z18.d\n" - "ld1w { z24.s }, p3/Z, [x22, #3, MUL VL]\n" - "ld1w { z25.s }, p2/Z, [x22, #4, MUL VL]\n" + "ld1w { z14.s }, p6/Z, [x21]\n" + "zip1 z8.d, z16.d, z14.d\n" + "zip2 z14.d, z16.d, z14.d\n" + "ld1w { z15.s }, p5/Z, [x21, #1, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x21, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z15.d\n" + "zip2 z15.d, z17.d, z15.d\n" + "ld1w { z17.s }, p3/Z, [x21, #3, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x21, #4, MUL VL]\n" + "zip1 z10.d, z19.d, z16.d\n" + "zip2 z16.d, z19.d, z16.d\n" + "ld1w { z19.s }, p1/Z, [x21, #5, MUL VL]\n" + "ld1w { z21.s }, p6/Z, [x20]\n" + "zip1 z11.d, z22.d, z17.d\n" + "zip2 z17.d, z22.d, z17.d\n" + "ld1w { z22.s }, p5/Z, [x20, #1, MUL VL]\n" + "ld1w { z23.s }, p4/Z, [x20, #2, MUL VL]\n" + "zip1 z12.d, z24.d, z18.d\n" + "zip2 z18.d, z24.d, z18.d\n" + "ld1w { z24.s }, p3/Z, [x20, #3, MUL VL]\n" + "ld1w { z25.s }, p2/Z, [x20, #4, MUL VL]\n" "zip1 z13.d, z20.d, z19.d\n" "zip2 z19.d, z20.d, z19.d\n" - "ld1w { z4.s }, p1/Z, [x22, #5, MUL VL]\n" + "ld1w { z0.s }, p1/Z, [x20, #5, MUL VL]\n" "zip1 z20.d, z21.d, z26.d\n" "zip2 z26.d, z21.d, z26.d\n" "zip1 z21.d, z22.d, z27.d\n" @@ -639,8 +639,8 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "zip2 z29.d, z24.d, z29.d\n" "zip1 z24.d, z25.d, z30.d\n" "zip2 z30.d, z25.d, z30.d\n" - "zip1 z25.d, z4.d, z31.d\n" - "zip2 z31.d, z4.d, z31.d\n" + "zip1 z25.d, z0.d, z31.d\n" + "zip2 z31.d, z0.d, z31.d\n" "b 31f\n" "30:" // Height 3: no accumulate "mov z8.b, #0x0\n" @@ -672,13 +672,13 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "32:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 33f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" - "ldr x22, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" + "ldr x22, [x20, #0x10]\n" "cbnz x26, 34f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -687,117 +687,117 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "b 34f\n" "33:" // Height 3: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "34:" // Height 3: input setup done "cmp x25, #0x4\n" "ble 36f\n" "35:" // Height 3: Multiply loop: Main loop head "whilelt p0.s, XZR, x25\n" - "ld1rqw { z0.s }, p0/Z, [x24]\n" - "ld1rqw { z1.s }, p0/Z, [x23]\n" + "ld1rqw { z5.s }, p0/Z, [x24]\n" + "ld1rqw { z0.s }, p0/Z, [x23]\n" + ".inst 0x658abca5 // bfcvt z5.h, p7/M, z5.s\n" + "ld1rqw { z4.s }, p0/Z, [x22]\n" ".inst 0x658abc00 // bfcvt z0.h, p7/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x22]\n" - ".inst 0x658abc21 // bfcvt z1.h, p7/M, z1.s\n" + "uzp1 z5.h, z5.h, z5.h\n" + "ld1h { z3.h }, p7/Z, [x28]\n" "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z4.h }, p7/Z, [x28]\n" - "uzp1 z1.h, z1.h, z1.h\n" - ".inst 0x658abc42 // bfcvt z2.h, p7/M, z2.s\n" - "ld1h { z5.h }, p7/Z, [x28, #1, MUL VL]\n" - "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n" - "trn1 z0.d, z0.d, z1.d\n" - "uzp1 z2.h, z2.h, z2.h\n" - "ld1h { z7.h }, p7/Z, [x28, #3, MUL VL]\n" - ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n" - ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n" - ".inst 0x6465e40e // bfmmla z14.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #4, MUL VL]\n" + ".inst 0x658abc84 // bfcvt z4.h, p7/M, z4.s\n" + "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n" + "ld1h { z1.h }, p7/Z, [x28, #2, MUL VL]\n" + "trn1 z5.d, z5.d, z0.d\n" + "uzp1 z4.h, z4.h, z4.h\n" + "ld1h { z0.h }, p7/Z, [x28, #3, MUL VL]\n" + ".inst 0x6463e4a8 // bfmmla z8.s, z5.h, z3.h\n" + ".inst 0x6463e494 // bfmmla z20.s, z4.h, z3.h\n" + ".inst 0x6462e4ae // bfmmla z14.s, z5.h, z2.h\n" + "ld1h { z3.h }, p7/Z, [x28, #4, MUL VL]\n" "sub x25, x25, #0x4\n" - ".inst 0x6465e45a // bfmmla z26.s, z2.h, z5.h\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - "ld1h { z5.h }, p7/Z, [x28, #5, MUL VL]\n" + ".inst 0x6462e49a // bfmmla z26.s, z4.h, z2.h\n" + ".inst 0x6461e4a9 // bfmmla z9.s, z5.h, z1.h\n" + "ld1h { z2.h }, p7/Z, [x28, #5, MUL VL]\n" "cmp x25, #0x4\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #6, MUL VL]\n" + ".inst 0x6461e495 // bfmmla z21.s, z4.h, z1.h\n" + ".inst 0x6460e4af // bfmmla z15.s, z5.h, z0.h\n" + "ld1h { z1.h }, p7/Z, [x28, #6, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x6467e45b // bfmmla z27.s, z2.h, z7.h\n" - "ld1h { z7.h }, p7/Z, [x28, #7, MUL VL]\n" + ".inst 0x6460e49b // bfmmla z27.s, z4.h, z0.h\n" + "ld1h { z0.h }, p7/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n" - ".inst 0x6464e456 // bfmmla z22.s, z2.h, z4.h\n" - ".inst 0x6465e410 // bfmmla z16.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #-8, MUL VL]\n" + ".inst 0x6463e4aa // bfmmla z10.s, z5.h, z3.h\n" + ".inst 0x6463e496 // bfmmla z22.s, z4.h, z3.h\n" + ".inst 0x6462e4b0 // bfmmla z16.s, z5.h, z2.h\n" + "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n" "add x23, x23, #0x10\n" - ".inst 0x6465e45c // bfmmla z28.s, z2.h, z5.h\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - "ld1h { z5.h }, p7/Z, [x28, #-7, MUL VL]\n" + ".inst 0x6462e49c // bfmmla z28.s, z4.h, z2.h\n" + ".inst 0x6461e4ab // bfmmla z11.s, z5.h, z1.h\n" + "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n" "add x22, x22, #0x10\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - ".inst 0x6467e411 // bfmmla z17.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #-6, MUL VL]\n" - ".inst 0x6467e45d // bfmmla z29.s, z2.h, z7.h\n" - "ld1h { z7.h }, p7/Z, [x28, #-5, MUL VL]\n" - ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n" + ".inst 0x6461e497 // bfmmla z23.s, z4.h, z1.h\n" + ".inst 0x6460e4b1 // bfmmla z17.s, z5.h, z0.h\n" + "ld1h { z1.h }, p7/Z, [x28, #-6, MUL VL]\n" + ".inst 0x6460e49d // bfmmla z29.s, z4.h, z0.h\n" + "ld1h { z0.h }, p7/Z, [x28, #-5, MUL VL]\n" + ".inst 0x6463e4ac // bfmmla z12.s, z5.h, z3.h\n" "addvl x28, x28, #-4\n" - ".inst 0x6464e458 // bfmmla z24.s, z2.h, z4.h\n" - ".inst 0x6465e412 // bfmmla z18.s, z0.h, z5.h\n" - ".inst 0x6465e45e // bfmmla z30.s, z2.h, z5.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e459 // bfmmla z25.s, z2.h, z6.h\n" - ".inst 0x6467e413 // bfmmla z19.s, z0.h, z7.h\n" - ".inst 0x6467e45f // bfmmla z31.s, z2.h, z7.h\n" + ".inst 0x6463e498 // bfmmla z24.s, z4.h, z3.h\n" + ".inst 0x6462e4b2 // bfmmla z18.s, z5.h, z2.h\n" + ".inst 0x6462e49e // bfmmla z30.s, z4.h, z2.h\n" + ".inst 0x6461e4ad // bfmmla z13.s, z5.h, z1.h\n" + ".inst 0x6461e499 // bfmmla z25.s, z4.h, z1.h\n" + ".inst 0x6460e4b3 // bfmmla z19.s, z5.h, z0.h\n" + ".inst 0x6460e49f // bfmmla z31.s, z4.h, z0.h\n" "bgt 35b\n" "36:" // Height 3: Multiply loop: Single iteration only "whilelt p0.s, XZR, x25\n" - "ld1rqw { z0.s }, p0/Z, [x24]\n" - "ld1rqw { z1.s }, p0/Z, [x23]\n" + "ld1rqw { z5.s }, p0/Z, [x24]\n" + "ld1rqw { z0.s }, p0/Z, [x23]\n" + ".inst 0x658abca5 // bfcvt z5.h, p7/M, z5.s\n" + "ld1rqw { z4.s }, p0/Z, [x22]\n" ".inst 0x658abc00 // bfcvt z0.h, p7/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x22]\n" - ".inst 0x658abc21 // bfcvt z1.h, p7/M, z1.s\n" + "uzp1 z5.h, z5.h, z5.h\n" + "ld1h { z3.h }, p7/Z, [x28]\n" "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z4.h }, p7/Z, [x28]\n" - "uzp1 z1.h, z1.h, z1.h\n" - ".inst 0x658abc42 // bfcvt z2.h, p7/M, z2.s\n" - "ld1h { z5.h }, p7/Z, [x28, #1, MUL VL]\n" - "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n" - "trn1 z0.d, z0.d, z1.d\n" - "uzp1 z2.h, z2.h, z2.h\n" - "ld1h { z7.h }, p7/Z, [x28, #3, MUL VL]\n" - ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n" - ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n" - ".inst 0x6465e40e // bfmmla z14.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #4, MUL VL]\n" - ".inst 0x6465e45a // bfmmla z26.s, z2.h, z5.h\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - "ld1h { z5.h }, p7/Z, [x28, #5, MUL VL]\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #6, MUL VL]\n" - ".inst 0x6467e45b // bfmmla z27.s, z2.h, z7.h\n" - "ld1h { z7.h }, p7/Z, [x28, #7, MUL VL]\n" + ".inst 0x658abc84 // bfcvt z4.h, p7/M, z4.s\n" + "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n" + "ld1h { z1.h }, p7/Z, [x28, #2, MUL VL]\n" + "trn1 z5.d, z5.d, z0.d\n" + "uzp1 z4.h, z4.h, z4.h\n" + "ld1h { z0.h }, p7/Z, [x28, #3, MUL VL]\n" + ".inst 0x6463e4a8 // bfmmla z8.s, z5.h, z3.h\n" + ".inst 0x6463e494 // bfmmla z20.s, z4.h, z3.h\n" + ".inst 0x6462e4ae // bfmmla z14.s, z5.h, z2.h\n" + "ld1h { z3.h }, p7/Z, [x28, #4, MUL VL]\n" + ".inst 0x6462e49a // bfmmla z26.s, z4.h, z2.h\n" + ".inst 0x6461e4a9 // bfmmla z9.s, z5.h, z1.h\n" + "ld1h { z2.h }, p7/Z, [x28, #5, MUL VL]\n" + ".inst 0x6461e495 // bfmmla z21.s, z4.h, z1.h\n" + ".inst 0x6460e4af // bfmmla z15.s, z5.h, z0.h\n" + "ld1h { z1.h }, p7/Z, [x28, #6, MUL VL]\n" + ".inst 0x6460e49b // bfmmla z27.s, z4.h, z0.h\n" + "ld1h { z0.h }, p7/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n" - ".inst 0x6464e456 // bfmmla z22.s, z2.h, z4.h\n" - ".inst 0x6465e410 // bfmmla z16.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #-8, MUL VL]\n" - ".inst 0x6465e45c // bfmmla z28.s, z2.h, z5.h\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - "ld1h { z5.h }, p7/Z, [x28, #-7, MUL VL]\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - ".inst 0x6467e411 // bfmmla z17.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #-6, MUL VL]\n" - ".inst 0x6467e45d // bfmmla z29.s, z2.h, z7.h\n" - "ld1h { z7.h }, p7/Z, [x28, #-5, MUL VL]\n" - ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n" + ".inst 0x6463e4aa // bfmmla z10.s, z5.h, z3.h\n" + ".inst 0x6463e496 // bfmmla z22.s, z4.h, z3.h\n" + ".inst 0x6462e4b0 // bfmmla z16.s, z5.h, z2.h\n" + "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n" + ".inst 0x6462e49c // bfmmla z28.s, z4.h, z2.h\n" + ".inst 0x6461e4ab // bfmmla z11.s, z5.h, z1.h\n" + "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n" + ".inst 0x6461e497 // bfmmla z23.s, z4.h, z1.h\n" + ".inst 0x6460e4b1 // bfmmla z17.s, z5.h, z0.h\n" + "ld1h { z1.h }, p7/Z, [x28, #-6, MUL VL]\n" + ".inst 0x6460e49d // bfmmla z29.s, z4.h, z0.h\n" + "ld1h { z0.h }, p7/Z, [x28, #-5, MUL VL]\n" + ".inst 0x6463e4ac // bfmmla z12.s, z5.h, z3.h\n" "addvl x28, x28, #-4\n" - ".inst 0x6464e458 // bfmmla z24.s, z2.h, z4.h\n" - ".inst 0x6465e412 // bfmmla z18.s, z0.h, z5.h\n" - ".inst 0x6465e45e // bfmmla z30.s, z2.h, z5.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e459 // bfmmla z25.s, z2.h, z6.h\n" - ".inst 0x6467e413 // bfmmla z19.s, z0.h, z7.h\n" - ".inst 0x6467e45f // bfmmla z31.s, z2.h, z7.h\n" + ".inst 0x6463e498 // bfmmla z24.s, z4.h, z3.h\n" + ".inst 0x6462e4b2 // bfmmla z18.s, z5.h, z2.h\n" + ".inst 0x6462e49e // bfmmla z30.s, z4.h, z2.h\n" + ".inst 0x6461e4ad // bfmmla z13.s, z5.h, z1.h\n" + ".inst 0x6461e499 // bfmmla z25.s, z4.h, z1.h\n" + ".inst 0x6460e4b3 // bfmmla z19.s, z5.h, z0.h\n" + ".inst 0x6460e49f // bfmmla z31.s, z4.h, z0.h\n" "37:" // Height 3: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" @@ -826,45 +826,45 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "uzp1 z25.d, z25.d, z31.d\n" "tbz %x[flags], #1, 38f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p7/Z, [x20]\n" - "add x20, %x[args_ptr], %[offset_min]\n" "ld1rw { z0.s }, p7/Z, [x20]\n" - "fmin z4.s, p7/M, z4.s, z1.s\n" - "fmin z14.s, p7/M, z14.s, z1.s\n" - "fmin z15.s, p7/M, z15.s, z1.s\n" - "fmin z16.s, p7/M, z16.s, z1.s\n" - "fmin z17.s, p7/M, z17.s, z1.s\n" - "fmin z18.s, p7/M, z18.s, z1.s\n" - "fmin z8.s, p7/M, z8.s, z1.s\n" - "fmin z9.s, p7/M, z9.s, z1.s\n" - "fmin z10.s, p7/M, z10.s, z1.s\n" - "fmin z11.s, p7/M, z11.s, z1.s\n" - "fmin z12.s, p7/M, z12.s, z1.s\n" - "fmin z13.s, p7/M, z13.s, z1.s\n" - "fmin z20.s, p7/M, z20.s, z1.s\n" - "fmin z21.s, p7/M, z21.s, z1.s\n" - "fmin z22.s, p7/M, z22.s, z1.s\n" - "fmin z23.s, p7/M, z23.s, z1.s\n" - "fmin z24.s, p7/M, z24.s, z1.s\n" - "fmin z25.s, p7/M, z25.s, z1.s\n" - "fmax z4.s, p7/M, z4.s, z0.s\n" - "fmax z14.s, p7/M, z14.s, z0.s\n" - "fmax z15.s, p7/M, z15.s, z0.s\n" - "fmax z16.s, p7/M, z16.s, z0.s\n" - "fmax z17.s, p7/M, z17.s, z0.s\n" - "fmax z18.s, p7/M, z18.s, z0.s\n" - "fmax z8.s, p7/M, z8.s, z0.s\n" - "fmax z9.s, p7/M, z9.s, z0.s\n" - "fmax z10.s, p7/M, z10.s, z0.s\n" - "fmax z11.s, p7/M, z11.s, z0.s\n" - "fmax z12.s, p7/M, z12.s, z0.s\n" - "fmax z13.s, p7/M, z13.s, z0.s\n" - "fmax z20.s, p7/M, z20.s, z0.s\n" - "fmax z21.s, p7/M, z21.s, z0.s\n" - "fmax z22.s, p7/M, z22.s, z0.s\n" - "fmax z23.s, p7/M, z23.s, z0.s\n" - "fmax z24.s, p7/M, z24.s, z0.s\n" - "fmax z25.s, p7/M, z25.s, z0.s\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z19.s }, p7/Z, [x20]\n" + "fmin z4.s, p7/M, z4.s, z0.s\n" + "fmin z14.s, p7/M, z14.s, z0.s\n" + "fmin z15.s, p7/M, z15.s, z0.s\n" + "fmin z16.s, p7/M, z16.s, z0.s\n" + "fmin z17.s, p7/M, z17.s, z0.s\n" + "fmin z18.s, p7/M, z18.s, z0.s\n" + "fmin z8.s, p7/M, z8.s, z0.s\n" + "fmin z9.s, p7/M, z9.s, z0.s\n" + "fmin z10.s, p7/M, z10.s, z0.s\n" + "fmin z11.s, p7/M, z11.s, z0.s\n" + "fmin z12.s, p7/M, z12.s, z0.s\n" + "fmin z13.s, p7/M, z13.s, z0.s\n" + "fmin z20.s, p7/M, z20.s, z0.s\n" + "fmin z21.s, p7/M, z21.s, z0.s\n" + "fmin z22.s, p7/M, z22.s, z0.s\n" + "fmin z23.s, p7/M, z23.s, z0.s\n" + "fmin z24.s, p7/M, z24.s, z0.s\n" + "fmin z25.s, p7/M, z25.s, z0.s\n" + "fmax z4.s, p7/M, z4.s, z19.s\n" + "fmax z14.s, p7/M, z14.s, z19.s\n" + "fmax z15.s, p7/M, z15.s, z19.s\n" + "fmax z16.s, p7/M, z16.s, z19.s\n" + "fmax z17.s, p7/M, z17.s, z19.s\n" + "fmax z18.s, p7/M, z18.s, z19.s\n" + "fmax z8.s, p7/M, z8.s, z19.s\n" + "fmax z9.s, p7/M, z9.s, z19.s\n" + "fmax z10.s, p7/M, z10.s, z19.s\n" + "fmax z11.s, p7/M, z11.s, z19.s\n" + "fmax z12.s, p7/M, z12.s, z19.s\n" + "fmax z13.s, p7/M, z13.s, z19.s\n" + "fmax z20.s, p7/M, z20.s, z19.s\n" + "fmax z21.s, p7/M, z21.s, z19.s\n" + "fmax z22.s, p7/M, z22.s, z19.s\n" + "fmax z23.s, p7/M, z23.s, z19.s\n" + "fmax z24.s, p7/M, z24.s, z19.s\n" + "fmax z25.s, p7/M, z25.s, z19.s\n" "38:" // Height 3: No activation "st1w { z4.s }, p6, [x27]\n" "st1w { z14.s }, p5, [x27, #1, MUL VL]\n" @@ -947,57 +947,57 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "42:" // Height 4: no bias "tbz %x[flags], #0, 43f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x27, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "ld1w { z9.s }, p6/Z, [x27]\n" + "add x22, x27, x20, LSL #2\n" "add x21, x22, x20, LSL #2\n" - "ld1w { z10.s }, p5/Z, [x27, #1, MUL VL]\n" - "ld1w { z11.s }, p4/Z, [x27, #2, MUL VL]\n" - "ld1w { z12.s }, p3/Z, [x27, #3, MUL VL]\n" - "ld1w { z13.s }, p2/Z, [x27, #4, MUL VL]\n" + "ld1w { z16.s }, p6/Z, [x27]\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z17.s }, p5/Z, [x27, #1, MUL VL]\n" + "ld1w { z19.s }, p4/Z, [x27, #2, MUL VL]\n" + "ld1w { z22.s }, p3/Z, [x27, #3, MUL VL]\n" + "ld1w { z24.s }, p2/Z, [x27, #4, MUL VL]\n" "ld1w { z20.s }, p1/Z, [x27, #5, MUL VL]\n" - "ld1w { z14.s }, p6/Z, [x23]\n" - "zip1 z8.d, z9.d, z14.d\n" - "zip2 z14.d, z9.d, z14.d\n" - "ld1w { z15.s }, p5/Z, [x23, #1, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x23, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z15.d\n" - "zip2 z15.d, z10.d, z15.d\n" - "ld1w { z17.s }, p3/Z, [x23, #3, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x23, #4, MUL VL]\n" - "zip1 z10.d, z11.d, z16.d\n" - "zip2 z16.d, z11.d, z16.d\n" - "ld1w { z19.s }, p1/Z, [x23, #5, MUL VL]\n" - "ld1w { z21.s }, p6/Z, [x22]\n" - "zip1 z11.d, z12.d, z17.d\n" - "zip2 z17.d, z12.d, z17.d\n" - "ld1w { z22.s }, p5/Z, [x22, #1, MUL VL]\n" - "ld1w { z23.s }, p4/Z, [x22, #2, MUL VL]\n" - "zip1 z12.d, z13.d, z18.d\n" - "zip2 z18.d, z13.d, z18.d\n" - "ld1w { z24.s }, p3/Z, [x22, #3, MUL VL]\n" - "ld1w { z25.s }, p2/Z, [x22, #4, MUL VL]\n" + "ld1w { z14.s }, p6/Z, [x22]\n" + "zip1 z8.d, z16.d, z14.d\n" + "zip2 z14.d, z16.d, z14.d\n" + "ld1w { z15.s }, p5/Z, [x22, #1, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x22, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z15.d\n" + "zip2 z15.d, z17.d, z15.d\n" + "ld1w { z17.s }, p3/Z, [x22, #3, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x22, #4, MUL VL]\n" + "zip1 z10.d, z19.d, z16.d\n" + "zip2 z16.d, z19.d, z16.d\n" + "ld1w { z19.s }, p1/Z, [x22, #5, MUL VL]\n" + "ld1w { z21.s }, p6/Z, [x21]\n" + "zip1 z11.d, z22.d, z17.d\n" + "zip2 z17.d, z22.d, z17.d\n" + "ld1w { z22.s }, p5/Z, [x21, #1, MUL VL]\n" + "ld1w { z23.s }, p4/Z, [x21, #2, MUL VL]\n" + "zip1 z12.d, z24.d, z18.d\n" + "zip2 z18.d, z24.d, z18.d\n" + "ld1w { z24.s }, p3/Z, [x21, #3, MUL VL]\n" + "ld1w { z25.s }, p2/Z, [x21, #4, MUL VL]\n" "zip1 z13.d, z20.d, z19.d\n" "zip2 z19.d, z20.d, z19.d\n" - "ld1w { z4.s }, p1/Z, [x22, #5, MUL VL]\n" - "ld1w { z26.s }, p6/Z, [x21]\n" + "ld1w { z0.s }, p1/Z, [x21, #5, MUL VL]\n" + "ld1w { z26.s }, p6/Z, [x20]\n" "zip1 z20.d, z21.d, z26.d\n" "zip2 z26.d, z21.d, z26.d\n" - "ld1w { z27.s }, p5/Z, [x21, #1, MUL VL]\n" - "ld1w { z28.s }, p4/Z, [x21, #2, MUL VL]\n" + "ld1w { z27.s }, p5/Z, [x20, #1, MUL VL]\n" + "ld1w { z28.s }, p4/Z, [x20, #2, MUL VL]\n" "zip1 z21.d, z22.d, z27.d\n" "zip2 z27.d, z22.d, z27.d\n" - "ld1w { z29.s }, p3/Z, [x21, #3, MUL VL]\n" - "ld1w { z30.s }, p2/Z, [x21, #4, MUL VL]\n" + "ld1w { z29.s }, p3/Z, [x20, #3, MUL VL]\n" + "ld1w { z30.s }, p2/Z, [x20, #4, MUL VL]\n" "zip1 z22.d, z23.d, z28.d\n" "zip2 z28.d, z23.d, z28.d\n" - "ld1w { z31.s }, p1/Z, [x21, #5, MUL VL]\n" + "ld1w { z31.s }, p1/Z, [x20, #5, MUL VL]\n" "zip1 z23.d, z24.d, z29.d\n" "zip2 z29.d, z24.d, z29.d\n" "zip1 z24.d, z25.d, z30.d\n" "zip2 z30.d, z25.d, z30.d\n" - "zip1 z25.d, z4.d, z31.d\n" - "zip2 z31.d, z4.d, z31.d\n" + "zip1 z25.d, z0.d, z31.d\n" + "zip2 z31.d, z0.d, z31.d\n" "b 44f\n" "43:" // Height 4: no accumulate "mov z8.b, #0x0\n" @@ -1029,14 +1029,14 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "45:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 46f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" - "ldr x22, [x21, #0x10]\n" - "ldr x21, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" + "ldr x22, [x20, #0x10]\n" + "ldr x21, [x20, #0x18]\n" "cbnz x26, 47f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -1046,127 +1046,127 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "b 47f\n" "46:" // Height 4: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "47:" // Height 4: input setup done "cmp x25, #0x4\n" "ble 49f\n" "48:" // Height 4: Multiply loop: Main loop head "whilelt p0.s, XZR, x25\n" - "ld1rqw { z0.s }, p0/Z, [x24]\n" - "ld1rqw { z1.s }, p0/Z, [x23]\n" - ".inst 0x658abc00 // bfcvt z0.h, p7/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x22]\n" - "ld1rqw { z3.s }, p0/Z, [x21]\n" - ".inst 0x658abc21 // bfcvt z1.h, p7/M, z1.s\n" - ".inst 0x658abc42 // bfcvt z2.h, p7/M, z2.s\n" - ".inst 0x658abc63 // bfcvt z3.h, p7/M, z3.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z4.h }, p7/Z, [x28]\n" - "ld1h { z5.h }, p7/Z, [x28, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "uzp1 z2.h, z2.h, z2.h\n" - "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #3, MUL VL]\n" - "uzp1 z3.h, z3.h, z3.h\n" - "trn1 z0.d, z0.d, z1.d\n" - ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n" + "ld1rqw { z7.s }, p0/Z, [x24]\n" + "ld1rqw { z6.s }, p0/Z, [x23]\n" + ".inst 0x658abce7 // bfcvt z7.h, p7/M, z7.s\n" + "ld1rqw { z5.s }, p0/Z, [x22]\n" + "ld1rqw { z4.s }, p0/Z, [x21]\n" + ".inst 0x658abcc6 // bfcvt z6.h, p7/M, z6.s\n" + ".inst 0x658abca5 // bfcvt z5.h, p7/M, z5.s\n" + ".inst 0x658abc84 // bfcvt z4.h, p7/M, z4.s\n" + "uzp1 z7.h, z7.h, z7.h\n" + "ld1h { z3.h }, p7/Z, [x28]\n" + "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n" + "uzp1 z6.h, z6.h, z6.h\n" + "uzp1 z5.h, z5.h, z5.h\n" + "ld1h { z1.h }, p7/Z, [x28, #2, MUL VL]\n" + "ld1h { z0.h }, p7/Z, [x28, #3, MUL VL]\n" + "uzp1 z4.h, z4.h, z4.h\n" + "trn1 z7.d, z7.d, z6.d\n" + ".inst 0x6463e4e8 // bfmmla z8.s, z7.h, z3.h\n" "sub x25, x25, #0x4\n" - "trn1 z2.d, z2.d, z3.d\n" - ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n" - ".inst 0x6465e40e // bfmmla z14.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #4, MUL VL]\n" - ".inst 0x6465e45a // bfmmla z26.s, z2.h, z5.h\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - "ld1h { z5.h }, p7/Z, [x28, #5, MUL VL]\n" + "trn1 z5.d, z5.d, z4.d\n" + ".inst 0x6463e4b4 // bfmmla z20.s, z5.h, z3.h\n" + ".inst 0x6462e4ee // bfmmla z14.s, z7.h, z2.h\n" + "ld1h { z3.h }, p7/Z, [x28, #4, MUL VL]\n" + ".inst 0x6462e4ba // bfmmla z26.s, z5.h, z2.h\n" + ".inst 0x6461e4e9 // bfmmla z9.s, z7.h, z1.h\n" + "ld1h { z2.h }, p7/Z, [x28, #5, MUL VL]\n" "cmp x25, #0x4\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #6, MUL VL]\n" + ".inst 0x6461e4b5 // bfmmla z21.s, z5.h, z1.h\n" + ".inst 0x6460e4ef // bfmmla z15.s, z7.h, z0.h\n" + "ld1h { z1.h }, p7/Z, [x28, #6, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x6467e45b // bfmmla z27.s, z2.h, z7.h\n" - "ld1h { z7.h }, p7/Z, [x28, #7, MUL VL]\n" + ".inst 0x6460e4bb // bfmmla z27.s, z5.h, z0.h\n" + "ld1h { z0.h }, p7/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n" - ".inst 0x6464e456 // bfmmla z22.s, z2.h, z4.h\n" - ".inst 0x6465e410 // bfmmla z16.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #-8, MUL VL]\n" + ".inst 0x6463e4ea // bfmmla z10.s, z7.h, z3.h\n" + ".inst 0x6463e4b6 // bfmmla z22.s, z5.h, z3.h\n" + ".inst 0x6462e4f0 // bfmmla z16.s, z7.h, z2.h\n" + "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n" "add x23, x23, #0x10\n" - ".inst 0x6465e45c // bfmmla z28.s, z2.h, z5.h\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - "ld1h { z5.h }, p7/Z, [x28, #-7, MUL VL]\n" + ".inst 0x6462e4bc // bfmmla z28.s, z5.h, z2.h\n" + ".inst 0x6461e4eb // bfmmla z11.s, z7.h, z1.h\n" + "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n" "add x22, x22, #0x10\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - ".inst 0x6467e411 // bfmmla z17.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #-6, MUL VL]\n" + ".inst 0x6461e4b7 // bfmmla z23.s, z5.h, z1.h\n" + ".inst 0x6460e4f1 // bfmmla z17.s, z7.h, z0.h\n" + "ld1h { z1.h }, p7/Z, [x28, #-6, MUL VL]\n" "add x21, x21, #0x10\n" - ".inst 0x6467e45d // bfmmla z29.s, z2.h, z7.h\n" - "ld1h { z7.h }, p7/Z, [x28, #-5, MUL VL]\n" - ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n" + ".inst 0x6460e4bd // bfmmla z29.s, z5.h, z0.h\n" + "ld1h { z0.h }, p7/Z, [x28, #-5, MUL VL]\n" + ".inst 0x6463e4ec // bfmmla z12.s, z7.h, z3.h\n" "addvl x28, x28, #-4\n" - ".inst 0x6464e458 // bfmmla z24.s, z2.h, z4.h\n" - ".inst 0x6465e412 // bfmmla z18.s, z0.h, z5.h\n" - ".inst 0x6465e45e // bfmmla z30.s, z2.h, z5.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e459 // bfmmla z25.s, z2.h, z6.h\n" - ".inst 0x6467e413 // bfmmla z19.s, z0.h, z7.h\n" - ".inst 0x6467e45f // bfmmla z31.s, z2.h, z7.h\n" + ".inst 0x6463e4b8 // bfmmla z24.s, z5.h, z3.h\n" + ".inst 0x6462e4f2 // bfmmla z18.s, z7.h, z2.h\n" + ".inst 0x6462e4be // bfmmla z30.s, z5.h, z2.h\n" + ".inst 0x6461e4ed // bfmmla z13.s, z7.h, z1.h\n" + ".inst 0x6461e4b9 // bfmmla z25.s, z5.h, z1.h\n" + ".inst 0x6460e4f3 // bfmmla z19.s, z7.h, z0.h\n" + ".inst 0x6460e4bf // bfmmla z31.s, z5.h, z0.h\n" "bgt 48b\n" "49:" // Height 4: Multiply loop: Single iteration only "whilelt p0.s, XZR, x25\n" - "ld1rqw { z0.s }, p0/Z, [x24]\n" - "ld1rqw { z1.s }, p0/Z, [x23]\n" - ".inst 0x658abc00 // bfcvt z0.h, p7/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x22]\n" - "ld1rqw { z3.s }, p0/Z, [x21]\n" - ".inst 0x658abc21 // bfcvt z1.h, p7/M, z1.s\n" - ".inst 0x658abc42 // bfcvt z2.h, p7/M, z2.s\n" - ".inst 0x658abc63 // bfcvt z3.h, p7/M, z3.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z4.h }, p7/Z, [x28]\n" - "ld1h { z5.h }, p7/Z, [x28, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "uzp1 z2.h, z2.h, z2.h\n" - "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n" - "ld1h { z7.h }, p7/Z, [x28, #3, MUL VL]\n" - "uzp1 z3.h, z3.h, z3.h\n" - "trn1 z0.d, z0.d, z1.d\n" - ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n" - "trn1 z2.d, z2.d, z3.d\n" - ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n" - ".inst 0x6465e40e // bfmmla z14.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #4, MUL VL]\n" - ".inst 0x6465e45a // bfmmla z26.s, z2.h, z5.h\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - "ld1h { z5.h }, p7/Z, [x28, #5, MUL VL]\n" - ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #6, MUL VL]\n" - ".inst 0x6467e45b // bfmmla z27.s, z2.h, z7.h\n" - "ld1h { z7.h }, p7/Z, [x28, #7, MUL VL]\n" + "ld1rqw { z7.s }, p0/Z, [x24]\n" + "ld1rqw { z6.s }, p0/Z, [x23]\n" + ".inst 0x658abce7 // bfcvt z7.h, p7/M, z7.s\n" + "ld1rqw { z5.s }, p0/Z, [x22]\n" + "ld1rqw { z4.s }, p0/Z, [x21]\n" + ".inst 0x658abcc6 // bfcvt z6.h, p7/M, z6.s\n" + ".inst 0x658abca5 // bfcvt z5.h, p7/M, z5.s\n" + ".inst 0x658abc84 // bfcvt z4.h, p7/M, z4.s\n" + "uzp1 z7.h, z7.h, z7.h\n" + "ld1h { z3.h }, p7/Z, [x28]\n" + "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n" + "uzp1 z6.h, z6.h, z6.h\n" + "uzp1 z5.h, z5.h, z5.h\n" + "ld1h { z1.h }, p7/Z, [x28, #2, MUL VL]\n" + "ld1h { z0.h }, p7/Z, [x28, #3, MUL VL]\n" + "uzp1 z4.h, z4.h, z4.h\n" + "trn1 z7.d, z7.d, z6.d\n" + ".inst 0x6463e4e8 // bfmmla z8.s, z7.h, z3.h\n" + "trn1 z5.d, z5.d, z4.d\n" + ".inst 0x6463e4b4 // bfmmla z20.s, z5.h, z3.h\n" + ".inst 0x6462e4ee // bfmmla z14.s, z7.h, z2.h\n" + "ld1h { z3.h }, p7/Z, [x28, #4, MUL VL]\n" + ".inst 0x6462e4ba // bfmmla z26.s, z5.h, z2.h\n" + ".inst 0x6461e4e9 // bfmmla z9.s, z7.h, z1.h\n" + "ld1h { z2.h }, p7/Z, [x28, #5, MUL VL]\n" + ".inst 0x6461e4b5 // bfmmla z21.s, z5.h, z1.h\n" + ".inst 0x6460e4ef // bfmmla z15.s, z7.h, z0.h\n" + "ld1h { z1.h }, p7/Z, [x28, #6, MUL VL]\n" + ".inst 0x6460e4bb // bfmmla z27.s, z5.h, z0.h\n" + "ld1h { z0.h }, p7/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n" - ".inst 0x6464e456 // bfmmla z22.s, z2.h, z4.h\n" - ".inst 0x6465e410 // bfmmla z16.s, z0.h, z5.h\n" - "ld1h { z4.h }, p7/Z, [x28, #-8, MUL VL]\n" - ".inst 0x6465e45c // bfmmla z28.s, z2.h, z5.h\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - "ld1h { z5.h }, p7/Z, [x28, #-7, MUL VL]\n" - ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" - ".inst 0x6467e411 // bfmmla z17.s, z0.h, z7.h\n" - "ld1h { z6.h }, p7/Z, [x28, #-6, MUL VL]\n" - ".inst 0x6467e45d // bfmmla z29.s, z2.h, z7.h\n" - "ld1h { z7.h }, p7/Z, [x28, #-5, MUL VL]\n" - ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n" + ".inst 0x6463e4ea // bfmmla z10.s, z7.h, z3.h\n" + ".inst 0x6463e4b6 // bfmmla z22.s, z5.h, z3.h\n" + ".inst 0x6462e4f0 // bfmmla z16.s, z7.h, z2.h\n" + "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n" + ".inst 0x6462e4bc // bfmmla z28.s, z5.h, z2.h\n" + ".inst 0x6461e4eb // bfmmla z11.s, z7.h, z1.h\n" + "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n" + ".inst 0x6461e4b7 // bfmmla z23.s, z5.h, z1.h\n" + ".inst 0x6460e4f1 // bfmmla z17.s, z7.h, z0.h\n" + "ld1h { z1.h }, p7/Z, [x28, #-6, MUL VL]\n" + ".inst 0x6460e4bd // bfmmla z29.s, z5.h, z0.h\n" + "ld1h { z0.h }, p7/Z, [x28, #-5, MUL VL]\n" + ".inst 0x6463e4ec // bfmmla z12.s, z7.h, z3.h\n" "addvl x28, x28, #-4\n" - ".inst 0x6464e458 // bfmmla z24.s, z2.h, z4.h\n" - ".inst 0x6465e412 // bfmmla z18.s, z0.h, z5.h\n" - ".inst 0x6465e45e // bfmmla z30.s, z2.h, z5.h\n" - ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" - ".inst 0x6466e459 // bfmmla z25.s, z2.h, z6.h\n" - ".inst 0x6467e413 // bfmmla z19.s, z0.h, z7.h\n" - ".inst 0x6467e45f // bfmmla z31.s, z2.h, z7.h\n" + ".inst 0x6463e4b8 // bfmmla z24.s, z5.h, z3.h\n" + ".inst 0x6462e4f2 // bfmmla z18.s, z7.h, z2.h\n" + ".inst 0x6462e4be // bfmmla z30.s, z5.h, z2.h\n" + ".inst 0x6461e4ed // bfmmla z13.s, z7.h, z1.h\n" + ".inst 0x6461e4b9 // bfmmla z25.s, z5.h, z1.h\n" + ".inst 0x6460e4f3 // bfmmla z19.s, z7.h, z0.h\n" + ".inst 0x6460e4bf // bfmmla z31.s, z5.h, z0.h\n" "50:" // Height 4: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" @@ -1295,7 +1295,6 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "54:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" @@ -1303,4 +1302,4 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL ( } } // namespace arm_gemm -#endif // __ARM_FEATURE_SVE +#endif // ARM_COMPUTE_ENABLE_SVE |