diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp | 1075 |
1 files changed, 536 insertions, 539 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp index 78a84fd89b..d5ccf3476c 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp @@ -49,19 +49,18 @@ void sve_ffhybrid_fp16_mla_6x4VL ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; - void *output_ptr = {}; - const __fp16 *bias = {}; } ka; unsigned long flags=0; + void *output_ptr; void *input_ptr; if (output_arg.is_indirect) { - ka.output_ptr=(void *)(output_arg.indirect.ptr); + output_ptr=(void *)(output_arg.indirect.ptr); ka.output_offset=output_arg.indirect.offset; flags |= 0x4; } else { - ka.output_ptr=(void *)(output_arg.direct.base); + output_ptr=(void *)(output_arg.direct.base); ka.output_offset=output_arg.direct.stride; } @@ -82,7 +81,6 @@ void sve_ffhybrid_fp16_mla_6x4VL ( ka.string_lengths = string_lengths; ka.N = N; ka.B_ptr = B_ptr; - ka.bias = bias; ka.B_stride = B_stride; switch(act.type) { default: @@ -108,15 +106,15 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "bgt 29f\n" "beq 15f\n" "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x15, %x[bias]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x13, %x[output_ptr]\n" "2:" // Height 1: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "cnth x21, ALL, MUL #3\n" "add x11, x12, x20, LSL #1\n" + "cnth x21, ALL, MUL #3\n" "add x10, x11, x20, LSL #1\n" "add x9, x10, x20, LSL #1\n" "add x20, x9, x20, LSL #1\n" @@ -124,12 +122,12 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 3f\n" "dech x21\n" - "mov x9, x12\n" "cmp x14, x21\n" + "mov x9, x12\n" "bgt 3f\n" "dech x21\n" - "mov x10, x12\n" "cmp x14, x21\n" + "mov x10, x12\n" "bgt 3f\n" "mov x11, x12\n" "3:" // Height 1: B setup done @@ -164,8 +162,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "mov x28, #0x0\n" "7:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 8f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -181,201 +179,201 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 11f\n" "10:" // Height 1: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1h { z17.h }, p5/Z, [x12]\n" - "ld1h { z16.h }, p5/Z, [x11]\n" - "sub x27, x27, #0x8\n" - "cmp x27, #0x8\n" "ld1rqh { z0.h }, p0/Z, [x26]\n" - "add x26, x26, #0x10\n" - "fmla z8.h, z17.h, z0.h[0]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x12]\n" + "fmla z8.h, z16.h, z0.h[0]\n" + "ld1h { z16.h }, p5/Z, [x11]\n" "fmla z9.h, z16.h, z0.h[0]\n" + "ld1h { z16.h }, p5/Z, [x10]\n" + "fmla z10.h, z16.h, z0.h[0]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "fmla z10.h, z17.h, z0.h[0]\n" - "ld1h { z17.h }, p5/Z, [x12, #1, MUL VL]\n" "fmla z11.h, z16.h, z0.h[0]\n" + "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n" + "fmla z8.h, z16.h, z0.h[1]\n" "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n" - "fmla z8.h, z17.h, z0.h[1]\n" - "ld1h { z17.h }, p5/Z, [x10, #1, MUL VL]\n" "fmla z9.h, z16.h, z0.h[1]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z10.h, z16.h, z0.h[1]\n" "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n" - "fmla z10.h, z17.h, z0.h[1]\n" - "ld1h { z17.h }, p5/Z, [x12, #2, MUL VL]\n" "fmla z11.h, z16.h, z0.h[1]\n" + "ld1h { z16.h }, p5/Z, [x12, #2, MUL VL]\n" + "fmla z8.h, z16.h, z0.h[2]\n" "ld1h { z16.h }, p5/Z, [x11, #2, MUL VL]\n" - "fmla z8.h, z17.h, z0.h[2]\n" - "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n" "fmla z9.h, z16.h, z0.h[2]\n" + "ld1h { z16.h }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z10.h, z16.h, z0.h[2]\n" "ld1h { z16.h }, p5/Z, [x9, #2, MUL VL]\n" - "fmla z10.h, z17.h, z0.h[2]\n" - "ld1h { z17.h }, p5/Z, [x12, #3, MUL VL]\n" "fmla z11.h, z16.h, z0.h[2]\n" + "ld1h { z16.h }, p5/Z, [x12, #3, MUL VL]\n" + "fmla z8.h, z16.h, z0.h[3]\n" "ld1h { z16.h }, p5/Z, [x11, #3, MUL VL]\n" - "fmla z8.h, z17.h, z0.h[3]\n" - "ld1h { z17.h }, p5/Z, [x10, #3, MUL VL]\n" "fmla z9.h, z16.h, z0.h[3]\n" + "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z10.h, z16.h, z0.h[3]\n" "ld1h { z16.h }, p5/Z, [x9, #3, MUL VL]\n" - "fmla z10.h, z17.h, z0.h[3]\n" - "ld1h { z17.h }, p5/Z, [x12, #4, MUL VL]\n" "fmla z11.h, z16.h, z0.h[3]\n" + "ld1h { z16.h }, p5/Z, [x12, #4, MUL VL]\n" + "fmla z8.h, z16.h, z0.h[4]\n" "ld1h { z16.h }, p5/Z, [x11, #4, MUL VL]\n" - "fmla z8.h, z17.h, z0.h[4]\n" - "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n" "fmla z9.h, z16.h, z0.h[4]\n" + "ld1h { z16.h }, p5/Z, [x10, #4, MUL VL]\n" + "fmla z10.h, z16.h, z0.h[4]\n" "ld1h { z16.h }, p5/Z, [x9, #4, MUL VL]\n" - "fmla z10.h, z17.h, z0.h[4]\n" - "ld1h { z17.h }, p5/Z, [x12, #5, MUL VL]\n" "fmla z11.h, z16.h, z0.h[4]\n" + "ld1h { z16.h }, p5/Z, [x12, #5, MUL VL]\n" + "fmla z8.h, z16.h, z0.h[5]\n" "ld1h { z16.h }, p5/Z, [x11, #5, MUL VL]\n" - "fmla z8.h, z17.h, z0.h[5]\n" - "ld1h { z17.h }, p5/Z, [x10, #5, MUL VL]\n" "fmla z9.h, z16.h, z0.h[5]\n" + "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n" + "fmla z10.h, z16.h, z0.h[5]\n" "ld1h { z16.h }, p5/Z, [x9, #5, MUL VL]\n" - "fmla z10.h, z17.h, z0.h[5]\n" - "ld1h { z17.h }, p5/Z, [x12, #6, MUL VL]\n" "fmla z11.h, z16.h, z0.h[5]\n" + "ld1h { z16.h }, p5/Z, [x12, #6, MUL VL]\n" + "fmla z8.h, z16.h, z0.h[6]\n" "ld1h { z16.h }, p5/Z, [x11, #6, MUL VL]\n" - "fmla z8.h, z17.h, z0.h[6]\n" - "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n" "fmla z9.h, z16.h, z0.h[6]\n" + "ld1h { z16.h }, p5/Z, [x10, #6, MUL VL]\n" + "fmla z10.h, z16.h, z0.h[6]\n" "ld1h { z16.h }, p5/Z, [x9, #6, MUL VL]\n" - "fmla z10.h, z17.h, z0.h[6]\n" - "ld1h { z17.h }, p5/Z, [x12, #7, MUL VL]\n" - "addvl x12, x12, #8\n" "fmla z11.h, z16.h, z0.h[6]\n" + "ld1h { z16.h }, p5/Z, [x12, #7, MUL VL]\n" + "fmla z8.h, z16.h, z0.h[7]\n" "ld1h { z16.h }, p5/Z, [x11, #7, MUL VL]\n" - "addvl x11, x11, #8\n" - "fmla z8.h, z17.h, z0.h[7]\n" - "ld1h { z17.h }, p5/Z, [x10, #7, MUL VL]\n" - "addvl x10, x10, #8\n" "fmla z9.h, z16.h, z0.h[7]\n" + "ld1h { z17.h }, p5/Z, [x10, #7, MUL VL]\n" + "sub x27, x27, #0x8\n" "ld1h { z16.h }, p5/Z, [x9, #7, MUL VL]\n" - "addvl x9, x9, #8\n" + "cmp x27, #0x8\n" "fmla z10.h, z17.h, z0.h[7]\n" "fmla z11.h, z16.h, z0.h[7]\n" + "add x26, x26, #0x10\n" + "addvl x12, x12, #8\n" + "addvl x11, x11, #8\n" + "addvl x10, x10, #8\n" + "addvl x9, x9, #8\n" "bgt 10b\n" "11:" // Height 1: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" - "ld1h { z17.h }, p5/Z, [x12]\n" + "ld1rqh { z0.h }, p0/Z, [x26]\n" + "ld1h { z16.h }, p5/Z, [x12]\n" + "fmla z8.h, z16.h, z0.h[0]\n" "ld1h { z16.h }, p5/Z, [x11]\n" + "fmla z9.h, z16.h, z0.h[0]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "subs x27, x27, #0x1\n" + "ld1h { z16.h }, p5/Z, [x9]\n" + "fmla z10.h, z17.h, z0.h[0]\n" + "fmla z11.h, z16.h, z0.h[0]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "ld1rqh { z0.h }, p0/Z, [x26]\n" - "fmla z8.h, z17.h, z0.h[0]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" - "fmla z9.h, z16.h, z0.h[0]\n" - "ld1h { z16.h }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.h, z17.h, z0.h[0]\n" - "fmla z11.h, z16.h, z0.h[0]\n" "ble 12f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[1]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[1]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[1]\n" "fmla z11.h, z16.h, z0.h[1]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 12f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[2]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[2]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[2]\n" "fmla z11.h, z16.h, z0.h[2]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 12f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[3]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[3]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[3]\n" "fmla z11.h, z16.h, z0.h[3]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 12f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[4]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[4]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[4]\n" "fmla z11.h, z16.h, z0.h[4]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 12f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[5]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[5]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[5]\n" "fmla z11.h, z16.h, z0.h[5]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 12f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[6]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[6]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[6]\n" "fmla z11.h, z16.h, z0.h[6]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 12f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[7]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[7]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" "fmla z10.h, z17.h, z0.h[7]\n" "fmla z11.h, z16.h, z0.h[7]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "12:" // Height 1: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" "bne 7b\n" "tbz %x[flags], #1, 13f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rh { z17.h }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rh { z17.h }, p5/Z, [x21]\n" "ld1rh { z16.h }, p5/Z, [x20]\n" "fmin z8.h, p5/M, z8.h, z17.h\n" "fmin z9.h, p5/M, z9.h, z17.h\n" @@ -398,15 +396,15 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "b 86f\n" "15:" // Height 2 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x15, %x[bias]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x13, %x[output_ptr]\n" "16:" // Height 2: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "cnth x21, ALL, MUL #3\n" "add x11, x12, x20, LSL #1\n" + "cnth x21, ALL, MUL #3\n" "add x10, x11, x20, LSL #1\n" "add x9, x10, x20, LSL #1\n" "add x20, x9, x20, LSL #1\n" @@ -414,12 +412,12 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 17f\n" "dech x21\n" - "mov x9, x12\n" "cmp x14, x21\n" + "mov x9, x12\n" "bgt 17f\n" "dech x21\n" - "mov x10, x12\n" "cmp x14, x21\n" + "mov x10, x12\n" "bgt 17f\n" "mov x11, x12\n" "17:" // Height 2: B setup done @@ -434,22 +432,22 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "cbz x15, 18f\n" "ld1h { z8.h }, p5/Z, [x15]\n" "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n" - "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" - "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" - "addvl x15, x15, #4\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" + "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" + "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" "mov z14.d, z10.d\n" "mov z15.d, z11.d\n" + "addvl x15, x15, #4\n" "b 20f\n" "18:" // Height 2: no bias "tbz %x[flags], #0, 19f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x20, x13, x20, LSL #1\n" "ld1h { z8.h }, p4/Z, [x13]\n" "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n" "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n" "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n" - "add x20, x13, x20, LSL #1\n" "ld1h { z12.h }, p4/Z, [x20]\n" "ld1h { z13.h }, p3/Z, [x20, #1, MUL VL]\n" "ld1h { z14.h }, p2/Z, [x20, #2, MUL VL]\n" @@ -468,8 +466,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "mov x28, #0x0\n" "21:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 22f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -488,29 +486,29 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 25f\n" "24:" // Height 2: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1h { z17.h }, p5/Z, [x12]\n" - "ld1h { z16.h }, p5/Z, [x11]\n" - "sub x27, x27, #0x8\n" - "cmp x27, #0x8\n" "ld1rqh { z1.h }, p0/Z, [x26]\n" - "add x26, x26, #0x10\n" "ld1rqh { z0.h }, p0/Z, [x25]\n" - "add x25, x25, #0x10\n" + "sub x27, x27, #0x8\n" + "ld1h { z17.h }, p5/Z, [x12]\n" + "ld1h { z16.h }, p5/Z, [x11]\n" "fmla z8.h, z17.h, z1.h[0]\n" "fmla z12.h, z17.h, z0.h[0]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" "fmla z9.h, z16.h, z1.h[0]\n" "fmla z13.h, z16.h, z0.h[0]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" "fmla z10.h, z17.h, z1.h[0]\n" "fmla z14.h, z17.h, z0.h[0]\n" "ld1h { z17.h }, p5/Z, [x12, #1, MUL VL]\n" + "cmp x27, #0x8\n" "fmla z11.h, z16.h, z1.h[0]\n" "fmla z15.h, z16.h, z0.h[0]\n" "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n" + "add x26, x26, #0x10\n" "fmla z8.h, z17.h, z1.h[1]\n" "fmla z12.h, z17.h, z0.h[1]\n" "ld1h { z17.h }, p5/Z, [x10, #1, MUL VL]\n" + "add x25, x25, #0x10\n" "fmla z9.h, z16.h, z1.h[1]\n" "fmla z13.h, z16.h, z0.h[1]\n" "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n" @@ -597,161 +595,161 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "bgt 24b\n" "25:" // Height 2: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" - "ld1h { z17.h }, p5/Z, [x12]\n" - "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "ld1rqh { z0.h }, p0/Z, [x26]\n" "ld1rqh { z1.h }, p0/Z, [x25]\n" + "subs x27, x27, #0x1\n" + "ld1h { z17.h }, p5/Z, [x12]\n" + "ld1h { z16.h }, p5/Z, [x11]\n" "fmla z8.h, z17.h, z0.h[0]\n" "fmla z12.h, z17.h, z1.h[0]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[0]\n" "fmla z13.h, z16.h, z1.h[0]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" "fmla z10.h, z17.h, z0.h[0]\n" "fmla z14.h, z17.h, z1.h[0]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z11.h, z16.h, z0.h[0]\n" "fmla z15.h, z16.h, z1.h[0]\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 26f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[1]\n" "fmla z12.h, z17.h, z1.h[1]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[1]\n" "fmla z13.h, z16.h, z1.h[1]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[1]\n" "fmla z14.h, z17.h, z1.h[1]\n" + "addvl x12, x12, #1\n" "fmla z11.h, z16.h, z0.h[1]\n" "fmla z15.h, z16.h, z1.h[1]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 26f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[2]\n" "fmla z12.h, z17.h, z1.h[2]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[2]\n" "fmla z13.h, z16.h, z1.h[2]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[2]\n" "fmla z14.h, z17.h, z1.h[2]\n" + "addvl x12, x12, #1\n" "fmla z11.h, z16.h, z0.h[2]\n" "fmla z15.h, z16.h, z1.h[2]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 26f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[3]\n" "fmla z12.h, z17.h, z1.h[3]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[3]\n" "fmla z13.h, z16.h, z1.h[3]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[3]\n" "fmla z14.h, z17.h, z1.h[3]\n" + "addvl x12, x12, #1\n" "fmla z11.h, z16.h, z0.h[3]\n" "fmla z15.h, z16.h, z1.h[3]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 26f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[4]\n" "fmla z12.h, z17.h, z1.h[4]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[4]\n" "fmla z13.h, z16.h, z1.h[4]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[4]\n" "fmla z14.h, z17.h, z1.h[4]\n" + "addvl x12, x12, #1\n" "fmla z11.h, z16.h, z0.h[4]\n" "fmla z15.h, z16.h, z1.h[4]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 26f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[5]\n" "fmla z12.h, z17.h, z1.h[5]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[5]\n" "fmla z13.h, z16.h, z1.h[5]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[5]\n" "fmla z14.h, z17.h, z1.h[5]\n" + "addvl x12, x12, #1\n" "fmla z11.h, z16.h, z0.h[5]\n" "fmla z15.h, z16.h, z1.h[5]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 26f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[6]\n" "fmla z12.h, z17.h, z1.h[6]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[6]\n" "fmla z13.h, z16.h, z1.h[6]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "subs x27, x27, #0x1\n" "fmla z10.h, z17.h, z0.h[6]\n" "fmla z14.h, z17.h, z1.h[6]\n" + "addvl x12, x12, #1\n" "fmla z11.h, z16.h, z0.h[6]\n" "fmla z15.h, z16.h, z1.h[6]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "ble 26f\n" "ld1h { z17.h }, p5/Z, [x12]\n" "ld1h { z16.h }, p5/Z, [x11]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z17.h, z0.h[7]\n" "fmla z12.h, z17.h, z1.h[7]\n" - "ld1h { z17.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z16.h, z0.h[7]\n" "fmla z13.h, z16.h, z1.h[7]\n" + "ld1h { z17.h }, p5/Z, [x10]\n" "ld1h { z16.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" "fmla z10.h, z17.h, z0.h[7]\n" "fmla z14.h, z17.h, z1.h[7]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z11.h, z16.h, z0.h[7]\n" "fmla z15.h, z16.h, z1.h[7]\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "26:" // Height 2: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" "bne 21b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x26, x13, x20, LSL #1\n" + "add x25, x13, x20, LSL #1\n" "tbz %x[flags], #1, 27f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rh { z17.h }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rh { z17.h }, p5/Z, [x21]\n" "ld1rh { z16.h }, p5/Z, [x20]\n" "fmin z8.h, p5/M, z8.h, z17.h\n" "fmin z9.h, p5/M, z9.h, z17.h\n" @@ -775,10 +773,10 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "st1h { z10.h }, p2, [x13, #2, MUL VL]\n" "st1h { z11.h }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1h { z12.h }, p4, [x26]\n" - "st1h { z13.h }, p3, [x26, #1, MUL VL]\n" - "st1h { z14.h }, p2, [x26, #2, MUL VL]\n" - "st1h { z15.h }, p1, [x26, #3, MUL VL]\n" + "st1h { z12.h }, p4, [x25]\n" + "st1h { z13.h }, p3, [x25, #1, MUL VL]\n" + "st1h { z14.h }, p2, [x25, #2, MUL VL]\n" + "st1h { z15.h }, p1, [x25, #3, MUL VL]\n" "28:" // Height 2: Writeback done "dech x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -786,15 +784,15 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "b 86f\n" "29:" // Height 3 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x15, %x[bias]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x13, %x[output_ptr]\n" "30:" // Height 3: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "cnth x21, ALL, MUL #3\n" "add x11, x12, x20, LSL #1\n" + "cnth x21, ALL, MUL #3\n" "add x10, x11, x20, LSL #1\n" "add x9, x10, x20, LSL #1\n" "add x20, x9, x20, LSL #1\n" @@ -802,12 +800,12 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 31f\n" "dech x21\n" - "mov x9, x12\n" "cmp x14, x21\n" + "mov x9, x12\n" "bgt 31f\n" "dech x21\n" - "mov x10, x12\n" "cmp x14, x21\n" + "mov x10, x12\n" "bgt 31f\n" "mov x11, x12\n" "31:" // Height 3: B setup done @@ -822,27 +820,27 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "cbz x15, 32f\n" "ld1h { z8.h }, p5/Z, [x15]\n" "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n" - "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" - "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" - "addvl x15, x15, #4\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" - "mov z16.d, z8.d\n" - "mov z17.d, z9.d\n" + "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" + "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" "mov z14.d, z10.d\n" "mov z15.d, z11.d\n" + "mov z16.d, z8.d\n" + "mov z17.d, z9.d\n" + "addvl x15, x15, #4\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" "b 34f\n" "32:" // Height 3: no bias "tbz %x[flags], #0, 33f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x21, x13, x20, LSL #1\n" + "add x20, x21, x20, LSL #1\n" "ld1h { z8.h }, p4/Z, [x13]\n" "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n" "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n" "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n" - "add x21, x13, x20, LSL #1\n" - "add x20, x21, x20, LSL #1\n" "ld1h { z12.h }, p4/Z, [x21]\n" "ld1h { z13.h }, p3/Z, [x21, #1, MUL VL]\n" "ld1h { z14.h }, p2/Z, [x21, #2, MUL VL]\n" @@ -869,8 +867,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "mov x28, #0x0\n" "35:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 36f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -892,126 +890,126 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 39f\n" "38:" // Height 3: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1h { z21.h }, p5/Z, [x12]\n" - "ld1h { z20.h }, p5/Z, [x11]\n" - "sub x27, x27, #0x8\n" - "cmp x27, #0x8\n" "ld1rqh { z2.h }, p0/Z, [x26]\n" - "add x26, x26, #0x10\n" "ld1rqh { z1.h }, p0/Z, [x25]\n" - "add x25, x25, #0x10\n" + "sub x27, x27, #0x8\n" "ld1rqh { z0.h }, p0/Z, [x24]\n" - "add x24, x24, #0x10\n" + "ld1h { z21.h }, p5/Z, [x12]\n" "fmla z8.h, z21.h, z2.h[0]\n" "fmla z12.h, z21.h, z1.h[0]\n" - "fmla z9.h, z20.h, z2.h[0]\n" - "fmla z13.h, z20.h, z1.h[0]\n" + "ld1h { z20.h }, p5/Z, [x11]\n" "fmla z16.h, z21.h, z0.h[0]\n" + "fmla z9.h, z20.h, z2.h[0]\n" "ld1h { z21.h }, p5/Z, [x10]\n" + "fmla z13.h, z20.h, z1.h[0]\n" "fmla z17.h, z20.h, z0.h[0]\n" "ld1h { z20.h }, p5/Z, [x9]\n" + "cmp x27, #0x8\n" "fmla z10.h, z21.h, z2.h[0]\n" "fmla z14.h, z21.h, z1.h[0]\n" + "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" "fmla z18.h, z21.h, z0.h[0]\n" - "ld1h { z21.h }, p5/Z, [x12, #1, MUL VL]\n" "fmla z11.h, z20.h, z2.h[0]\n" + "ld1h { z21.h }, p5/Z, [x12, #1, MUL VL]\n" + "add x24, x24, #0x10\n" "fmla z15.h, z20.h, z1.h[0]\n" "fmla z19.h, z20.h, z0.h[0]\n" "ld1h { z20.h }, p5/Z, [x11, #1, MUL VL]\n" "fmla z8.h, z21.h, z2.h[1]\n" "fmla z12.h, z21.h, z1.h[1]\n" "fmla z16.h, z21.h, z0.h[1]\n" - "ld1h { z21.h }, p5/Z, [x10, #1, MUL VL]\n" "fmla z9.h, z20.h, z2.h[1]\n" + "ld1h { z21.h }, p5/Z, [x10, #1, MUL VL]\n" "fmla z13.h, z20.h, z1.h[1]\n" "fmla z17.h, z20.h, z0.h[1]\n" "ld1h { z20.h }, p5/Z, [x9, #1, MUL VL]\n" "fmla z10.h, z21.h, z2.h[1]\n" "fmla z14.h, z21.h, z1.h[1]\n" "fmla z18.h, z21.h, z0.h[1]\n" - "ld1h { z21.h }, p5/Z, [x12, #2, MUL VL]\n" "fmla z11.h, z20.h, z2.h[1]\n" + "ld1h { z21.h }, p5/Z, [x12, #2, MUL VL]\n" "fmla z15.h, z20.h, z1.h[1]\n" "fmla z19.h, z20.h, z0.h[1]\n" "ld1h { z20.h }, p5/Z, [x11, #2, MUL VL]\n" "fmla z8.h, z21.h, z2.h[2]\n" "fmla z12.h, z21.h, z1.h[2]\n" "fmla z16.h, z21.h, z0.h[2]\n" - "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n" "fmla z9.h, z20.h, z2.h[2]\n" + "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n" "fmla z13.h, z20.h, z1.h[2]\n" "fmla z17.h, z20.h, z0.h[2]\n" "ld1h { z20.h }, p5/Z, [x9, #2, MUL VL]\n" "fmla z10.h, z21.h, z2.h[2]\n" "fmla z14.h, z21.h, z1.h[2]\n" "fmla z18.h, z21.h, z0.h[2]\n" - "ld1h { z21.h }, p5/Z, [x12, #3, MUL VL]\n" "fmla z11.h, z20.h, z2.h[2]\n" + "ld1h { z21.h }, p5/Z, [x12, #3, MUL VL]\n" "fmla z15.h, z20.h, z1.h[2]\n" "fmla z19.h, z20.h, z0.h[2]\n" "ld1h { z20.h }, p5/Z, [x11, #3, MUL VL]\n" "fmla z8.h, z21.h, z2.h[3]\n" "fmla z12.h, z21.h, z1.h[3]\n" "fmla z16.h, z21.h, z0.h[3]\n" - "ld1h { z21.h }, p5/Z, [x10, #3, MUL VL]\n" "fmla z9.h, z20.h, z2.h[3]\n" + "ld1h { z21.h }, p5/Z, [x10, #3, MUL VL]\n" "fmla z13.h, z20.h, z1.h[3]\n" "fmla z17.h, z20.h, z0.h[3]\n" "ld1h { z20.h }, p5/Z, [x9, #3, MUL VL]\n" "fmla z10.h, z21.h, z2.h[3]\n" "fmla z14.h, z21.h, z1.h[3]\n" "fmla z18.h, z21.h, z0.h[3]\n" - "ld1h { z21.h }, p5/Z, [x12, #4, MUL VL]\n" "fmla z11.h, z20.h, z2.h[3]\n" + "ld1h { z21.h }, p5/Z, [x12, #4, MUL VL]\n" "fmla z15.h, z20.h, z1.h[3]\n" "fmla z19.h, z20.h, z0.h[3]\n" "ld1h { z20.h }, p5/Z, [x11, #4, MUL VL]\n" "fmla z8.h, z21.h, z2.h[4]\n" "fmla z12.h, z21.h, z1.h[4]\n" "fmla z16.h, z21.h, z0.h[4]\n" - "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n" "fmla z9.h, z20.h, z2.h[4]\n" + "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n" "fmla z13.h, z20.h, z1.h[4]\n" "fmla z17.h, z20.h, z0.h[4]\n" "ld1h { z20.h }, p5/Z, [x9, #4, MUL VL]\n" "fmla z10.h, z21.h, z2.h[4]\n" "fmla z14.h, z21.h, z1.h[4]\n" "fmla z18.h, z21.h, z0.h[4]\n" - "ld1h { z21.h }, p5/Z, [x12, #5, MUL VL]\n" "fmla z11.h, z20.h, z2.h[4]\n" + "ld1h { z21.h }, p5/Z, [x12, #5, MUL VL]\n" "fmla z15.h, z20.h, z1.h[4]\n" "fmla z19.h, z20.h, z0.h[4]\n" "ld1h { z20.h }, p5/Z, [x11, #5, MUL VL]\n" "fmla z8.h, z21.h, z2.h[5]\n" "fmla z12.h, z21.h, z1.h[5]\n" "fmla z16.h, z21.h, z0.h[5]\n" - "ld1h { z21.h }, p5/Z, [x10, #5, MUL VL]\n" "fmla z9.h, z20.h, z2.h[5]\n" + "ld1h { z21.h }, p5/Z, [x10, #5, MUL VL]\n" "fmla z13.h, z20.h, z1.h[5]\n" "fmla z17.h, z20.h, z0.h[5]\n" "ld1h { z20.h }, p5/Z, [x9, #5, MUL VL]\n" "fmla z10.h, z21.h, z2.h[5]\n" "fmla z14.h, z21.h, z1.h[5]\n" "fmla z18.h, z21.h, z0.h[5]\n" - "ld1h { z21.h }, p5/Z, [x12, #6, MUL VL]\n" "fmla z11.h, z20.h, z2.h[5]\n" + "ld1h { z21.h }, p5/Z, [x12, #6, MUL VL]\n" "fmla z15.h, z20.h, z1.h[5]\n" "fmla z19.h, z20.h, z0.h[5]\n" "ld1h { z20.h }, p5/Z, [x11, #6, MUL VL]\n" "fmla z8.h, z21.h, z2.h[6]\n" "fmla z12.h, z21.h, z1.h[6]\n" "fmla z16.h, z21.h, z0.h[6]\n" - "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n" "fmla z9.h, z20.h, z2.h[6]\n" + "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n" "fmla z13.h, z20.h, z1.h[6]\n" "fmla z17.h, z20.h, z0.h[6]\n" "ld1h { z20.h }, p5/Z, [x9, #6, MUL VL]\n" "fmla z10.h, z21.h, z2.h[6]\n" "fmla z14.h, z21.h, z1.h[6]\n" "fmla z18.h, z21.h, z0.h[6]\n" + "fmla z11.h, z20.h, z2.h[6]\n" "ld1h { z21.h }, p5/Z, [x12, #7, MUL VL]\n" "addvl x12, x12, #8\n" - "fmla z11.h, z20.h, z2.h[6]\n" "fmla z15.h, z20.h, z1.h[6]\n" "fmla z19.h, z20.h, z0.h[6]\n" "ld1h { z20.h }, p5/Z, [x11, #7, MUL VL]\n" @@ -1019,9 +1017,9 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z8.h, z21.h, z2.h[7]\n" "fmla z12.h, z21.h, z1.h[7]\n" "fmla z16.h, z21.h, z0.h[7]\n" + "fmla z9.h, z20.h, z2.h[7]\n" "ld1h { z21.h }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - "fmla z9.h, z20.h, z2.h[7]\n" "fmla z13.h, z20.h, z1.h[7]\n" "fmla z17.h, z20.h, z0.h[7]\n" "ld1h { z20.h }, p5/Z, [x9, #7, MUL VL]\n" @@ -1035,179 +1033,179 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "bgt 38b\n" "39:" // Height 3: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" - "ld1h { z21.h }, p5/Z, [x12]\n" - "ld1h { z20.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "ld1rqh { z0.h }, p0/Z, [x26]\n" "ld1rqh { z1.h }, p0/Z, [x25]\n" + "subs x27, x27, #0x1\n" "ld1rqh { z2.h }, p0/Z, [x24]\n" + "ld1h { z21.h }, p5/Z, [x12]\n" "fmla z8.h, z21.h, z0.h[0]\n" "fmla z12.h, z21.h, z1.h[0]\n" - "fmla z9.h, z20.h, z0.h[0]\n" - "fmla z13.h, z20.h, z1.h[0]\n" + "ld1h { z20.h }, p5/Z, [x11]\n" "fmla z16.h, z21.h, z2.h[0]\n" + "fmla z9.h, z20.h, z0.h[0]\n" "ld1h { z21.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "fmla z13.h, z20.h, z1.h[0]\n" "fmla z17.h, z20.h, z2.h[0]\n" "ld1h { z20.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x12, x12, #1\n" "fmla z10.h, z21.h, z0.h[0]\n" "fmla z14.h, z21.h, z1.h[0]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z18.h, z21.h, z2.h[0]\n" "fmla z11.h, z20.h, z0.h[0]\n" + "addvl x9, x9, #1\n" "fmla z15.h, z20.h, z1.h[0]\n" "fmla z19.h, z20.h, z2.h[0]\n" "ble 40f\n" "ld1h { z21.h }, p5/Z, [x12]\n" "ld1h { z20.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z21.h, z0.h[1]\n" "fmla z12.h, z21.h, z1.h[1]\n" "fmla z16.h, z21.h, z2.h[1]\n" - "ld1h { z21.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z20.h, z0.h[1]\n" + "ld1h { z21.h }, p5/Z, [x10]\n" + "subs x27, x27, #0x1\n" "fmla z13.h, z20.h, z1.h[1]\n" "fmla z17.h, z20.h, z2.h[1]\n" "ld1h { z20.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x12, x12, #1\n" "fmla z10.h, z21.h, z0.h[1]\n" "fmla z14.h, z21.h, z1.h[1]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z18.h, z21.h, z2.h[1]\n" "fmla z11.h, z20.h, z0.h[1]\n" + "addvl x9, x9, #1\n" "fmla z15.h, z20.h, z1.h[1]\n" "fmla z19.h, z20.h, z2.h[1]\n" "ble 40f\n" "ld1h { z21.h }, p5/Z, [x12]\n" "ld1h { z20.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z21.h, z0.h[2]\n" "fmla z12.h, z21.h, z1.h[2]\n" "fmla z16.h, z21.h, z2.h[2]\n" - "ld1h { z21.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z20.h, z0.h[2]\n" + "ld1h { z21.h }, p5/Z, [x10]\n" + "subs x27, x27, #0x1\n" "fmla z13.h, z20.h, z1.h[2]\n" "fmla z17.h, z20.h, z2.h[2]\n" "ld1h { z20.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x12, x12, #1\n" "fmla z10.h, z21.h, z0.h[2]\n" "fmla z14.h, z21.h, z1.h[2]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z18.h, z21.h, z2.h[2]\n" "fmla z11.h, z20.h, z0.h[2]\n" + "addvl x9, x9, #1\n" "fmla z15.h, z20.h, z1.h[2]\n" "fmla z19.h, z20.h, z2.h[2]\n" "ble 40f\n" "ld1h { z21.h }, p5/Z, [x12]\n" "ld1h { z20.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z21.h, z0.h[3]\n" "fmla z12.h, z21.h, z1.h[3]\n" "fmla z16.h, z21.h, z2.h[3]\n" - "ld1h { z21.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z20.h, z0.h[3]\n" + "ld1h { z21.h }, p5/Z, [x10]\n" + "subs x27, x27, #0x1\n" "fmla z13.h, z20.h, z1.h[3]\n" "fmla z17.h, z20.h, z2.h[3]\n" "ld1h { z20.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x12, x12, #1\n" "fmla z10.h, z21.h, z0.h[3]\n" "fmla z14.h, z21.h, z1.h[3]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z18.h, z21.h, z2.h[3]\n" "fmla z11.h, z20.h, z0.h[3]\n" + "addvl x9, x9, #1\n" "fmla z15.h, z20.h, z1.h[3]\n" "fmla z19.h, z20.h, z2.h[3]\n" "ble 40f\n" "ld1h { z21.h }, p5/Z, [x12]\n" "ld1h { z20.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z21.h, z0.h[4]\n" "fmla z12.h, z21.h, z1.h[4]\n" "fmla z16.h, z21.h, z2.h[4]\n" - "ld1h { z21.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z20.h, z0.h[4]\n" + "ld1h { z21.h }, p5/Z, [x10]\n" + "subs x27, x27, #0x1\n" "fmla z13.h, z20.h, z1.h[4]\n" "fmla z17.h, z20.h, z2.h[4]\n" "ld1h { z20.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x12, x12, #1\n" "fmla z10.h, z21.h, z0.h[4]\n" "fmla z14.h, z21.h, z1.h[4]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z18.h, z21.h, z2.h[4]\n" "fmla z11.h, z20.h, z0.h[4]\n" + "addvl x9, x9, #1\n" "fmla z15.h, z20.h, z1.h[4]\n" "fmla z19.h, z20.h, z2.h[4]\n" "ble 40f\n" "ld1h { z21.h }, p5/Z, [x12]\n" "ld1h { z20.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z21.h, z0.h[5]\n" "fmla z12.h, z21.h, z1.h[5]\n" "fmla z16.h, z21.h, z2.h[5]\n" - "ld1h { z21.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z20.h, z0.h[5]\n" + "ld1h { z21.h }, p5/Z, [x10]\n" + "subs x27, x27, #0x1\n" "fmla z13.h, z20.h, z1.h[5]\n" "fmla z17.h, z20.h, z2.h[5]\n" "ld1h { z20.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x12, x12, #1\n" "fmla z10.h, z21.h, z0.h[5]\n" "fmla z14.h, z21.h, z1.h[5]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z18.h, z21.h, z2.h[5]\n" "fmla z11.h, z20.h, z0.h[5]\n" + "addvl x9, x9, #1\n" "fmla z15.h, z20.h, z1.h[5]\n" "fmla z19.h, z20.h, z2.h[5]\n" "ble 40f\n" "ld1h { z21.h }, p5/Z, [x12]\n" "ld1h { z20.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z21.h, z0.h[6]\n" "fmla z12.h, z21.h, z1.h[6]\n" "fmla z16.h, z21.h, z2.h[6]\n" - "ld1h { z21.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z20.h, z0.h[6]\n" + "ld1h { z21.h }, p5/Z, [x10]\n" + "subs x27, x27, #0x1\n" "fmla z13.h, z20.h, z1.h[6]\n" "fmla z17.h, z20.h, z2.h[6]\n" "ld1h { z20.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x12, x12, #1\n" "fmla z10.h, z21.h, z0.h[6]\n" "fmla z14.h, z21.h, z1.h[6]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z18.h, z21.h, z2.h[6]\n" "fmla z11.h, z20.h, z0.h[6]\n" + "addvl x9, x9, #1\n" "fmla z15.h, z20.h, z1.h[6]\n" "fmla z19.h, z20.h, z2.h[6]\n" "ble 40f\n" "ld1h { z21.h }, p5/Z, [x12]\n" "ld1h { z20.h }, p5/Z, [x11]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z21.h, z0.h[7]\n" "fmla z12.h, z21.h, z1.h[7]\n" "fmla z16.h, z21.h, z2.h[7]\n" - "ld1h { z21.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" "fmla z9.h, z20.h, z0.h[7]\n" + "ld1h { z21.h }, p5/Z, [x10]\n" + "addvl x12, x12, #1\n" "fmla z13.h, z20.h, z1.h[7]\n" "fmla z17.h, z20.h, z2.h[7]\n" "ld1h { z20.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x11, x11, #1\n" "fmla z10.h, z21.h, z0.h[7]\n" "fmla z14.h, z21.h, z1.h[7]\n" + "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "fmla z18.h, z21.h, z2.h[7]\n" "fmla z11.h, z20.h, z0.h[7]\n" "fmla z15.h, z20.h, z1.h[7]\n" @@ -1218,12 +1216,12 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "cmp x28, x20\n" "bne 35b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x26, x13, x20, LSL #1\n" - "add x25, x26, x20, LSL #1\n" + "add x25, x13, x20, LSL #1\n" + "add x24, x25, x20, LSL #1\n" "tbz %x[flags], #1, 41f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rh { z21.h }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rh { z21.h }, p5/Z, [x21]\n" "ld1rh { z20.h }, p5/Z, [x20]\n" "fmin z8.h, p5/M, z8.h, z21.h\n" "fmin z9.h, p5/M, z9.h, z21.h\n" @@ -1255,14 +1253,14 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "st1h { z10.h }, p2, [x13, #2, MUL VL]\n" "st1h { z11.h }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1h { z12.h }, p4, [x26]\n" - "st1h { z13.h }, p3, [x26, #1, MUL VL]\n" - "st1h { z14.h }, p2, [x26, #2, MUL VL]\n" - "st1h { z15.h }, p1, [x26, #3, MUL VL]\n" - "st1h { z16.h }, p4, [x25]\n" - "st1h { z17.h }, p3, [x25, #1, MUL VL]\n" - "st1h { z18.h }, p2, [x25, #2, MUL VL]\n" - "st1h { z19.h }, p1, [x25, #3, MUL VL]\n" + "st1h { z12.h }, p4, [x25]\n" + "st1h { z13.h }, p3, [x25, #1, MUL VL]\n" + "st1h { z14.h }, p2, [x25, #2, MUL VL]\n" + "st1h { z15.h }, p1, [x25, #3, MUL VL]\n" + "st1h { z16.h }, p4, [x24]\n" + "st1h { z17.h }, p3, [x24, #1, MUL VL]\n" + "st1h { z18.h }, p2, [x24, #2, MUL VL]\n" + "st1h { z19.h }, p1, [x24, #3, MUL VL]\n" "42:" // Height 3: Writeback done "dech x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -1270,15 +1268,15 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "b 86f\n" "43:" // Height 4 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x15, %x[bias]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x13, %x[output_ptr]\n" "44:" // Height 4: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "cnth x21, ALL, MUL #3\n" "add x11, x12, x20, LSL #1\n" + "cnth x21, ALL, MUL #3\n" "add x10, x11, x20, LSL #1\n" "add x9, x10, x20, LSL #1\n" "add x20, x9, x20, LSL #1\n" @@ -1286,12 +1284,12 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 45f\n" "dech x21\n" - "mov x9, x12\n" "cmp x14, x21\n" + "mov x9, x12\n" "bgt 45f\n" "dech x21\n" - "mov x10, x12\n" "cmp x14, x21\n" + "mov x10, x12\n" "bgt 45f\n" "mov x11, x12\n" "45:" // Height 4: B setup done @@ -1306,18 +1304,18 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "cbz x15, 46f\n" "ld1h { z8.h }, p5/Z, [x15]\n" "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n" - "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" - "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" - "addvl x15, x15, #4\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" - "mov z16.d, z8.d\n" - "mov z17.d, z9.d\n" - "mov z20.d, z8.d\n" + "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" + "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" "mov z14.d, z10.d\n" "mov z15.d, z11.d\n" + "mov z16.d, z8.d\n" + "mov z17.d, z9.d\n" + "addvl x15, x15, #4\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" + "mov z20.d, z8.d\n" "mov z21.d, z9.d\n" "mov z22.d, z10.d\n" "mov z23.d, z11.d\n" @@ -1325,13 +1323,13 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "46:" // Height 4: no bias "tbz %x[flags], #0, 47f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x22, x13, x20, LSL #1\n" + "add x21, x22, x20, LSL #1\n" "ld1h { z8.h }, p4/Z, [x13]\n" + "add x20, x21, x20, LSL #1\n" "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n" "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n" "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n" - "add x22, x13, x20, LSL #1\n" - "add x21, x22, x20, LSL #1\n" - "add x20, x21, x20, LSL #1\n" "ld1h { z12.h }, p4/Z, [x22]\n" "ld1h { z13.h }, p3/Z, [x22, #1, MUL VL]\n" "ld1h { z14.h }, p2/Z, [x22, #2, MUL VL]\n" @@ -1366,8 +1364,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "mov x28, #0x0\n" "49:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 50f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1392,25 +1390,25 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 53f\n" "52:" // Height 4: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1h { z25.h }, p5/Z, [x12]\n" - "ld1h { z24.h }, p5/Z, [x11]\n" - "sub x27, x27, #0x8\n" - "cmp x27, #0x8\n" "ld1rqh { z3.h }, p0/Z, [x26]\n" - "add x26, x26, #0x10\n" "ld1rqh { z2.h }, p0/Z, [x25]\n" - "add x25, x25, #0x10\n" + "sub x27, x27, #0x8\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x24, x24, #0x10\n" "ld1rqh { z0.h }, p0/Z, [x23]\n" - "add x23, x23, #0x10\n" + "cmp x27, #0x8\n" + "add x26, x26, #0x10\n" + "ld1h { z25.h }, p5/Z, [x12]\n" + "ld1h { z24.h }, p5/Z, [x11]\n" "fmla z8.h, z25.h, z3.h[0]\n" "fmla z12.h, z25.h, z2.h[0]\n" - "fmla z9.h, z24.h, z3.h[0]\n" - "fmla z13.h, z24.h, z2.h[0]\n" "fmla z16.h, z25.h, z1.h[0]\n" "fmla z20.h, z25.h, z0.h[0]\n" "ld1h { z25.h }, p5/Z, [x10]\n" + "add x25, x25, #0x10\n" + "fmla z9.h, z24.h, z3.h[0]\n" + "fmla z13.h, z24.h, z2.h[0]\n" + "add x24, x24, #0x10\n" + "add x23, x23, #0x10\n" "fmla z17.h, z24.h, z1.h[0]\n" "fmla z21.h, z24.h, z0.h[0]\n" "ld1h { z24.h }, p5/Z, [x9]\n" @@ -1569,22 +1567,22 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "bgt 52b\n" "53:" // Height 4: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" - "ld1h { z25.h }, p5/Z, [x12]\n" - "ld1h { z24.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "ld1rqh { z0.h }, p0/Z, [x26]\n" "ld1rqh { z1.h }, p0/Z, [x25]\n" + "subs x27, x27, #0x1\n" "ld1rqh { z2.h }, p0/Z, [x24]\n" "ld1rqh { z3.h }, p0/Z, [x23]\n" + "ld1h { z25.h }, p5/Z, [x12]\n" + "ld1h { z24.h }, p5/Z, [x11]\n" "fmla z8.h, z25.h, z0.h[0]\n" "fmla z12.h, z25.h, z1.h[0]\n" - "fmla z9.h, z24.h, z0.h[0]\n" - "fmla z13.h, z24.h, z1.h[0]\n" "fmla z16.h, z25.h, z2.h[0]\n" "fmla z20.h, z25.h, z3.h[0]\n" "ld1h { z25.h }, p5/Z, [x10]\n" + "addvl x12, x12, #1\n" + "fmla z9.h, z24.h, z0.h[0]\n" + "fmla z13.h, z24.h, z1.h[0]\n" + "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "fmla z17.h, z24.h, z2.h[0]\n" "fmla z21.h, z24.h, z3.h[0]\n" @@ -1601,23 +1599,23 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 54f\n" "ld1h { z25.h }, p5/Z, [x12]\n" "ld1h { z24.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z25.h, z0.h[1]\n" "fmla z12.h, z25.h, z1.h[1]\n" "fmla z16.h, z25.h, z2.h[1]\n" "fmla z20.h, z25.h, z3.h[1]\n" "ld1h { z25.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "subs x27, x27, #0x1\n" "fmla z9.h, z24.h, z0.h[1]\n" "fmla z13.h, z24.h, z1.h[1]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z17.h, z24.h, z2.h[1]\n" "fmla z21.h, z24.h, z3.h[1]\n" "ld1h { z24.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x10, x10, #1\n" "fmla z10.h, z25.h, z0.h[1]\n" "fmla z14.h, z25.h, z1.h[1]\n" + "addvl x9, x9, #1\n" "fmla z18.h, z25.h, z2.h[1]\n" "fmla z22.h, z25.h, z3.h[1]\n" "fmla z11.h, z24.h, z0.h[1]\n" @@ -1627,23 +1625,23 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 54f\n" "ld1h { z25.h }, p5/Z, [x12]\n" "ld1h { z24.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z25.h, z0.h[2]\n" "fmla z12.h, z25.h, z1.h[2]\n" "fmla z16.h, z25.h, z2.h[2]\n" "fmla z20.h, z25.h, z3.h[2]\n" "ld1h { z25.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "subs x27, x27, #0x1\n" "fmla z9.h, z24.h, z0.h[2]\n" "fmla z13.h, z24.h, z1.h[2]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z17.h, z24.h, z2.h[2]\n" "fmla z21.h, z24.h, z3.h[2]\n" "ld1h { z24.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x10, x10, #1\n" "fmla z10.h, z25.h, z0.h[2]\n" "fmla z14.h, z25.h, z1.h[2]\n" + "addvl x9, x9, #1\n" "fmla z18.h, z25.h, z2.h[2]\n" "fmla z22.h, z25.h, z3.h[2]\n" "fmla z11.h, z24.h, z0.h[2]\n" @@ -1653,23 +1651,23 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 54f\n" "ld1h { z25.h }, p5/Z, [x12]\n" "ld1h { z24.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z25.h, z0.h[3]\n" "fmla z12.h, z25.h, z1.h[3]\n" "fmla z16.h, z25.h, z2.h[3]\n" "fmla z20.h, z25.h, z3.h[3]\n" "ld1h { z25.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "subs x27, x27, #0x1\n" "fmla z9.h, z24.h, z0.h[3]\n" "fmla z13.h, z24.h, z1.h[3]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z17.h, z24.h, z2.h[3]\n" "fmla z21.h, z24.h, z3.h[3]\n" "ld1h { z24.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x10, x10, #1\n" "fmla z10.h, z25.h, z0.h[3]\n" "fmla z14.h, z25.h, z1.h[3]\n" + "addvl x9, x9, #1\n" "fmla z18.h, z25.h, z2.h[3]\n" "fmla z22.h, z25.h, z3.h[3]\n" "fmla z11.h, z24.h, z0.h[3]\n" @@ -1679,23 +1677,23 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 54f\n" "ld1h { z25.h }, p5/Z, [x12]\n" "ld1h { z24.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z25.h, z0.h[4]\n" "fmla z12.h, z25.h, z1.h[4]\n" "fmla z16.h, z25.h, z2.h[4]\n" "fmla z20.h, z25.h, z3.h[4]\n" "ld1h { z25.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "subs x27, x27, #0x1\n" "fmla z9.h, z24.h, z0.h[4]\n" "fmla z13.h, z24.h, z1.h[4]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z17.h, z24.h, z2.h[4]\n" "fmla z21.h, z24.h, z3.h[4]\n" "ld1h { z24.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x10, x10, #1\n" "fmla z10.h, z25.h, z0.h[4]\n" "fmla z14.h, z25.h, z1.h[4]\n" + "addvl x9, x9, #1\n" "fmla z18.h, z25.h, z2.h[4]\n" "fmla z22.h, z25.h, z3.h[4]\n" "fmla z11.h, z24.h, z0.h[4]\n" @@ -1705,23 +1703,23 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 54f\n" "ld1h { z25.h }, p5/Z, [x12]\n" "ld1h { z24.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z25.h, z0.h[5]\n" "fmla z12.h, z25.h, z1.h[5]\n" "fmla z16.h, z25.h, z2.h[5]\n" "fmla z20.h, z25.h, z3.h[5]\n" "ld1h { z25.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "subs x27, x27, #0x1\n" "fmla z9.h, z24.h, z0.h[5]\n" "fmla z13.h, z24.h, z1.h[5]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z17.h, z24.h, z2.h[5]\n" "fmla z21.h, z24.h, z3.h[5]\n" "ld1h { z24.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x10, x10, #1\n" "fmla z10.h, z25.h, z0.h[5]\n" "fmla z14.h, z25.h, z1.h[5]\n" + "addvl x9, x9, #1\n" "fmla z18.h, z25.h, z2.h[5]\n" "fmla z22.h, z25.h, z3.h[5]\n" "fmla z11.h, z24.h, z0.h[5]\n" @@ -1731,23 +1729,23 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 54f\n" "ld1h { z25.h }, p5/Z, [x12]\n" "ld1h { z24.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z25.h, z0.h[6]\n" "fmla z12.h, z25.h, z1.h[6]\n" "fmla z16.h, z25.h, z2.h[6]\n" "fmla z20.h, z25.h, z3.h[6]\n" "ld1h { z25.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "subs x27, x27, #0x1\n" "fmla z9.h, z24.h, z0.h[6]\n" "fmla z13.h, z24.h, z1.h[6]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z17.h, z24.h, z2.h[6]\n" "fmla z21.h, z24.h, z3.h[6]\n" "ld1h { z24.h }, p5/Z, [x9]\n" - "addvl x9, x9, #1\n" + "addvl x10, x10, #1\n" "fmla z10.h, z25.h, z0.h[6]\n" "fmla z14.h, z25.h, z1.h[6]\n" + "addvl x9, x9, #1\n" "fmla z18.h, z25.h, z2.h[6]\n" "fmla z22.h, z25.h, z3.h[6]\n" "fmla z11.h, z24.h, z0.h[6]\n" @@ -1757,16 +1755,16 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 54f\n" "ld1h { z25.h }, p5/Z, [x12]\n" "ld1h { z24.h }, p5/Z, [x11]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z25.h, z0.h[7]\n" "fmla z12.h, z25.h, z1.h[7]\n" "fmla z16.h, z25.h, z2.h[7]\n" "fmla z20.h, z25.h, z3.h[7]\n" "ld1h { z25.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x12, x12, #1\n" "fmla z9.h, z24.h, z0.h[7]\n" "fmla z13.h, z24.h, z1.h[7]\n" + "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z17.h, z24.h, z2.h[7]\n" "fmla z21.h, z24.h, z3.h[7]\n" "ld1h { z24.h }, p5/Z, [x9]\n" @@ -1785,13 +1783,13 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "cmp x28, x20\n" "bne 49b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x26, x13, x20, LSL #1\n" - "add x25, x26, x20, LSL #1\n" + "add x25, x13, x20, LSL #1\n" "add x24, x25, x20, LSL #1\n" + "add x23, x24, x20, LSL #1\n" "tbz %x[flags], #1, 55f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rh { z25.h }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rh { z25.h }, p5/Z, [x21]\n" "ld1rh { z24.h }, p5/Z, [x20]\n" "fmin z8.h, p5/M, z8.h, z25.h\n" "fmin z9.h, p5/M, z9.h, z25.h\n" @@ -1831,18 +1829,18 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "st1h { z10.h }, p2, [x13, #2, MUL VL]\n" "st1h { z11.h }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1h { z12.h }, p4, [x26]\n" - "st1h { z13.h }, p3, [x26, #1, MUL VL]\n" - "st1h { z14.h }, p2, [x26, #2, MUL VL]\n" - "st1h { z15.h }, p1, [x26, #3, MUL VL]\n" - "st1h { z16.h }, p4, [x25]\n" - "st1h { z17.h }, p3, [x25, #1, MUL VL]\n" - "st1h { z18.h }, p2, [x25, #2, MUL VL]\n" - "st1h { z19.h }, p1, [x25, #3, MUL VL]\n" - "st1h { z20.h }, p4, [x24]\n" - "st1h { z21.h }, p3, [x24, #1, MUL VL]\n" - "st1h { z22.h }, p2, [x24, #2, MUL VL]\n" - "st1h { z23.h }, p1, [x24, #3, MUL VL]\n" + "st1h { z12.h }, p4, [x25]\n" + "st1h { z13.h }, p3, [x25, #1, MUL VL]\n" + "st1h { z14.h }, p2, [x25, #2, MUL VL]\n" + "st1h { z15.h }, p1, [x25, #3, MUL VL]\n" + "st1h { z16.h }, p4, [x24]\n" + "st1h { z17.h }, p3, [x24, #1, MUL VL]\n" + "st1h { z18.h }, p2, [x24, #2, MUL VL]\n" + "st1h { z19.h }, p1, [x24, #3, MUL VL]\n" + "st1h { z20.h }, p4, [x23]\n" + "st1h { z21.h }, p3, [x23, #1, MUL VL]\n" + "st1h { z22.h }, p2, [x23, #2, MUL VL]\n" + "st1h { z23.h }, p1, [x23, #3, MUL VL]\n" "56:" // Height 4: Writeback done "dech x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -1850,15 +1848,15 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "b 86f\n" "57:" // Height 5 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x15, %x[bias]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x13, %x[output_ptr]\n" "58:" // Height 5: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "cnth x21, ALL, MUL #3\n" "add x11, x12, x20, LSL #1\n" + "cnth x21, ALL, MUL #3\n" "add x10, x11, x20, LSL #1\n" "add x9, x10, x20, LSL #1\n" "add x20, x9, x20, LSL #1\n" @@ -1866,12 +1864,12 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 59f\n" "dech x21\n" - "mov x9, x12\n" "cmp x14, x21\n" + "mov x9, x12\n" "bgt 59f\n" "dech x21\n" - "mov x10, x12\n" "cmp x14, x21\n" + "mov x10, x12\n" "bgt 59f\n" "mov x11, x12\n" "59:" // Height 5: B setup done @@ -1886,18 +1884,18 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "cbz x15, 60f\n" "ld1h { z8.h }, p5/Z, [x15]\n" "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n" - "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" - "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" - "addvl x15, x15, #4\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" - "mov z16.d, z8.d\n" - "mov z17.d, z9.d\n" - "mov z20.d, z8.d\n" + "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" + "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" "mov z14.d, z10.d\n" "mov z15.d, z11.d\n" + "mov z16.d, z8.d\n" + "mov z17.d, z9.d\n" + "addvl x15, x15, #4\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" + "mov z20.d, z8.d\n" "mov z21.d, z9.d\n" "mov z22.d, z10.d\n" "mov z23.d, z11.d\n" @@ -1909,16 +1907,16 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "60:" // Height 5: no bias "tbz %x[flags], #0, 61f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x23, x13, x20, LSL #1\n" + "add x22, x23, x20, LSL #1\n" "ld1h { z8.h }, p4/Z, [x13]\n" + "add x21, x22, x20, LSL #1\n" + "add x20, x21, x20, LSL #1\n" "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n" "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n" "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n" - "add x23, x13, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" - "add x21, x22, x20, LSL #1\n" "ld1h { z12.h }, p4/Z, [x23]\n" "ld1h { z13.h }, p3/Z, [x23, #1, MUL VL]\n" - "add x20, x21, x20, LSL #1\n" "ld1h { z14.h }, p2/Z, [x23, #2, MUL VL]\n" "ld1h { z15.h }, p1/Z, [x23, #3, MUL VL]\n" "ld1h { z16.h }, p4/Z, [x22]\n" @@ -1959,8 +1957,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "mov x28, #0x0\n" "63:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 64f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1988,29 +1986,29 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 67f\n" "66:" // Height 5: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1h { z29.h }, p5/Z, [x12]\n" - "ld1h { z28.h }, p5/Z, [x11]\n" - "sub x27, x27, #0x8\n" - "cmp x27, #0x8\n" "ld1rqh { z4.h }, p0/Z, [x26]\n" - "add x26, x26, #0x10\n" "ld1rqh { z3.h }, p0/Z, [x25]\n" - "add x25, x25, #0x10\n" + "sub x27, x27, #0x8\n" "ld1rqh { z2.h }, p0/Z, [x24]\n" - "add x24, x24, #0x10\n" "ld1rqh { z1.h }, p0/Z, [x23]\n" + "cmp x27, #0x8\n" + "add x26, x26, #0x10\n" "ld1rqh { z0.h }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "add x22, x22, #0x10\n" + "ld1h { z29.h }, p5/Z, [x12]\n" "fmla z8.h, z29.h, z4.h[0]\n" "fmla z12.h, z29.h, z3.h[0]\n" - "fmla z9.h, z28.h, z4.h[0]\n" + "ld1h { z28.h }, p5/Z, [x11]\n" "fmla z16.h, z29.h, z2.h[0]\n" "fmla z20.h, z29.h, z1.h[0]\n" + "add x25, x25, #0x10\n" "fmla z24.h, z29.h, z0.h[0]\n" - "fmla z13.h, z28.h, z3.h[0]\n" + "fmla z9.h, z28.h, z4.h[0]\n" "ld1h { z29.h }, p5/Z, [x10]\n" + "add x24, x24, #0x10\n" + "fmla z13.h, z28.h, z3.h[0]\n" "fmla z17.h, z28.h, z2.h[0]\n" + "add x23, x23, #0x10\n" + "add x22, x22, #0x10\n" "fmla z21.h, z28.h, z1.h[0]\n" "fmla z25.h, z28.h, z0.h[0]\n" "ld1h { z28.h }, p5/Z, [x9]\n" @@ -2019,8 +2017,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z18.h, z29.h, z2.h[0]\n" "fmla z22.h, z29.h, z1.h[0]\n" "fmla z26.h, z29.h, z0.h[0]\n" - "ld1h { z29.h }, p5/Z, [x12, #1, MUL VL]\n" "fmla z11.h, z28.h, z4.h[0]\n" + "ld1h { z29.h }, p5/Z, [x12, #1, MUL VL]\n" "fmla z15.h, z28.h, z3.h[0]\n" "fmla z19.h, z28.h, z2.h[0]\n" "fmla z23.h, z28.h, z1.h[0]\n" @@ -2031,8 +2029,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z16.h, z29.h, z2.h[1]\n" "fmla z20.h, z29.h, z1.h[1]\n" "fmla z24.h, z29.h, z0.h[1]\n" - "ld1h { z29.h }, p5/Z, [x10, #1, MUL VL]\n" "fmla z9.h, z28.h, z4.h[1]\n" + "ld1h { z29.h }, p5/Z, [x10, #1, MUL VL]\n" "fmla z13.h, z28.h, z3.h[1]\n" "fmla z17.h, z28.h, z2.h[1]\n" "fmla z21.h, z28.h, z1.h[1]\n" @@ -2043,8 +2041,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z18.h, z29.h, z2.h[1]\n" "fmla z22.h, z29.h, z1.h[1]\n" "fmla z26.h, z29.h, z0.h[1]\n" - "ld1h { z29.h }, p5/Z, [x12, #2, MUL VL]\n" "fmla z11.h, z28.h, z4.h[1]\n" + "ld1h { z29.h }, p5/Z, [x12, #2, MUL VL]\n" "fmla z15.h, z28.h, z3.h[1]\n" "fmla z19.h, z28.h, z2.h[1]\n" "fmla z23.h, z28.h, z1.h[1]\n" @@ -2055,8 +2053,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z16.h, z29.h, z2.h[2]\n" "fmla z20.h, z29.h, z1.h[2]\n" "fmla z24.h, z29.h, z0.h[2]\n" - "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n" "fmla z9.h, z28.h, z4.h[2]\n" + "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n" "fmla z13.h, z28.h, z3.h[2]\n" "fmla z17.h, z28.h, z2.h[2]\n" "fmla z21.h, z28.h, z1.h[2]\n" @@ -2067,8 +2065,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z18.h, z29.h, z2.h[2]\n" "fmla z22.h, z29.h, z1.h[2]\n" "fmla z26.h, z29.h, z0.h[2]\n" - "ld1h { z29.h }, p5/Z, [x12, #3, MUL VL]\n" "fmla z11.h, z28.h, z4.h[2]\n" + "ld1h { z29.h }, p5/Z, [x12, #3, MUL VL]\n" "fmla z15.h, z28.h, z3.h[2]\n" "fmla z19.h, z28.h, z2.h[2]\n" "fmla z23.h, z28.h, z1.h[2]\n" @@ -2079,8 +2077,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z16.h, z29.h, z2.h[3]\n" "fmla z20.h, z29.h, z1.h[3]\n" "fmla z24.h, z29.h, z0.h[3]\n" - "ld1h { z29.h }, p5/Z, [x10, #3, MUL VL]\n" "fmla z9.h, z28.h, z4.h[3]\n" + "ld1h { z29.h }, p5/Z, [x10, #3, MUL VL]\n" "fmla z13.h, z28.h, z3.h[3]\n" "fmla z17.h, z28.h, z2.h[3]\n" "fmla z21.h, z28.h, z1.h[3]\n" @@ -2091,8 +2089,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z18.h, z29.h, z2.h[3]\n" "fmla z22.h, z29.h, z1.h[3]\n" "fmla z26.h, z29.h, z0.h[3]\n" - "ld1h { z29.h }, p5/Z, [x12, #4, MUL VL]\n" "fmla z11.h, z28.h, z4.h[3]\n" + "ld1h { z29.h }, p5/Z, [x12, #4, MUL VL]\n" "fmla z15.h, z28.h, z3.h[3]\n" "fmla z19.h, z28.h, z2.h[3]\n" "fmla z23.h, z28.h, z1.h[3]\n" @@ -2103,8 +2101,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z16.h, z29.h, z2.h[4]\n" "fmla z20.h, z29.h, z1.h[4]\n" "fmla z24.h, z29.h, z0.h[4]\n" - "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n" "fmla z9.h, z28.h, z4.h[4]\n" + "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n" "fmla z13.h, z28.h, z3.h[4]\n" "fmla z17.h, z28.h, z2.h[4]\n" "fmla z21.h, z28.h, z1.h[4]\n" @@ -2115,8 +2113,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z18.h, z29.h, z2.h[4]\n" "fmla z22.h, z29.h, z1.h[4]\n" "fmla z26.h, z29.h, z0.h[4]\n" - "ld1h { z29.h }, p5/Z, [x12, #5, MUL VL]\n" "fmla z11.h, z28.h, z4.h[4]\n" + "ld1h { z29.h }, p5/Z, [x12, #5, MUL VL]\n" "fmla z15.h, z28.h, z3.h[4]\n" "fmla z19.h, z28.h, z2.h[4]\n" "fmla z23.h, z28.h, z1.h[4]\n" @@ -2127,8 +2125,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z16.h, z29.h, z2.h[5]\n" "fmla z20.h, z29.h, z1.h[5]\n" "fmla z24.h, z29.h, z0.h[5]\n" - "ld1h { z29.h }, p5/Z, [x10, #5, MUL VL]\n" "fmla z9.h, z28.h, z4.h[5]\n" + "ld1h { z29.h }, p5/Z, [x10, #5, MUL VL]\n" "fmla z13.h, z28.h, z3.h[5]\n" "fmla z17.h, z28.h, z2.h[5]\n" "fmla z21.h, z28.h, z1.h[5]\n" @@ -2139,8 +2137,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z18.h, z29.h, z2.h[5]\n" "fmla z22.h, z29.h, z1.h[5]\n" "fmla z26.h, z29.h, z0.h[5]\n" - "ld1h { z29.h }, p5/Z, [x12, #6, MUL VL]\n" "fmla z11.h, z28.h, z4.h[5]\n" + "ld1h { z29.h }, p5/Z, [x12, #6, MUL VL]\n" "fmla z15.h, z28.h, z3.h[5]\n" "fmla z19.h, z28.h, z2.h[5]\n" "fmla z23.h, z28.h, z1.h[5]\n" @@ -2151,8 +2149,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z16.h, z29.h, z2.h[6]\n" "fmla z20.h, z29.h, z1.h[6]\n" "fmla z24.h, z29.h, z0.h[6]\n" - "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n" "fmla z9.h, z28.h, z4.h[6]\n" + "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n" "fmla z13.h, z28.h, z3.h[6]\n" "fmla z17.h, z28.h, z2.h[6]\n" "fmla z21.h, z28.h, z1.h[6]\n" @@ -2163,30 +2161,30 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "fmla z18.h, z29.h, z2.h[6]\n" "fmla z22.h, z29.h, z1.h[6]\n" "fmla z26.h, z29.h, z0.h[6]\n" + "fmla z11.h, z28.h, z4.h[6]\n" "ld1h { z29.h }, p5/Z, [x12, #7, MUL VL]\n" "addvl x12, x12, #8\n" - "fmla z11.h, z28.h, z4.h[6]\n" "fmla z15.h, z28.h, z3.h[6]\n" "fmla z19.h, z28.h, z2.h[6]\n" "fmla z23.h, z28.h, z1.h[6]\n" "fmla z27.h, z28.h, z0.h[6]\n" "ld1h { z28.h }, p5/Z, [x11, #7, MUL VL]\n" - "fmla z8.h, z29.h, z4.h[7]\n" "addvl x11, x11, #8\n" + "fmla z8.h, z29.h, z4.h[7]\n" "fmla z12.h, z29.h, z3.h[7]\n" "fmla z16.h, z29.h, z2.h[7]\n" "fmla z20.h, z29.h, z1.h[7]\n" "fmla z24.h, z29.h, z0.h[7]\n" + "fmla z9.h, z28.h, z4.h[7]\n" "ld1h { z29.h }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - "fmla z9.h, z28.h, z4.h[7]\n" "fmla z13.h, z28.h, z3.h[7]\n" "fmla z17.h, z28.h, z2.h[7]\n" "fmla z21.h, z28.h, z1.h[7]\n" "fmla z25.h, z28.h, z0.h[7]\n" "ld1h { z28.h }, p5/Z, [x9, #7, MUL VL]\n" - "fmla z10.h, z29.h, z4.h[7]\n" "addvl x9, x9, #8\n" + "fmla z10.h, z29.h, z4.h[7]\n" "fmla z14.h, z29.h, z3.h[7]\n" "fmla z18.h, z29.h, z2.h[7]\n" "fmla z22.h, z29.h, z1.h[7]\n" @@ -2199,25 +2197,25 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "bgt 66b\n" "67:" // Height 5: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" - "ld1h { z29.h }, p5/Z, [x12]\n" - "ld1h { z28.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "ld1rqh { z0.h }, p0/Z, [x26]\n" "ld1rqh { z1.h }, p0/Z, [x25]\n" + "subs x27, x27, #0x1\n" "ld1rqh { z2.h }, p0/Z, [x24]\n" "ld1rqh { z3.h }, p0/Z, [x23]\n" "ld1rqh { z4.h }, p0/Z, [x22]\n" + "ld1h { z29.h }, p5/Z, [x12]\n" "fmla z8.h, z29.h, z0.h[0]\n" "fmla z12.h, z29.h, z1.h[0]\n" - "fmla z9.h, z28.h, z0.h[0]\n" - "fmla z13.h, z28.h, z1.h[0]\n" + "ld1h { z28.h }, p5/Z, [x11]\n" "fmla z16.h, z29.h, z2.h[0]\n" "fmla z20.h, z29.h, z3.h[0]\n" + "addvl x12, x12, #1\n" "fmla z24.h, z29.h, z4.h[0]\n" - "fmla z17.h, z28.h, z2.h[0]\n" + "fmla z9.h, z28.h, z0.h[0]\n" "ld1h { z29.h }, p5/Z, [x10]\n" + "addvl x11, x11, #1\n" + "fmla z13.h, z28.h, z1.h[0]\n" + "fmla z17.h, z28.h, z2.h[0]\n" "addvl x10, x10, #1\n" "fmla z21.h, z28.h, z3.h[0]\n" "fmla z25.h, z28.h, z4.h[0]\n" @@ -2236,19 +2234,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 68f\n" "ld1h { z29.h }, p5/Z, [x12]\n" "ld1h { z28.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z29.h, z0.h[1]\n" "fmla z12.h, z29.h, z1.h[1]\n" "fmla z16.h, z29.h, z2.h[1]\n" "fmla z20.h, z29.h, z3.h[1]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z29.h, z4.h[1]\n" "fmla z9.h, z28.h, z0.h[1]\n" "ld1h { z29.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z13.h, z28.h, z1.h[1]\n" "fmla z17.h, z28.h, z2.h[1]\n" + "addvl x10, x10, #1\n" "fmla z21.h, z28.h, z3.h[1]\n" "fmla z25.h, z28.h, z4.h[1]\n" "ld1h { z28.h }, p5/Z, [x9]\n" @@ -2266,19 +2264,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 68f\n" "ld1h { z29.h }, p5/Z, [x12]\n" "ld1h { z28.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z29.h, z0.h[2]\n" "fmla z12.h, z29.h, z1.h[2]\n" "fmla z16.h, z29.h, z2.h[2]\n" "fmla z20.h, z29.h, z3.h[2]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z29.h, z4.h[2]\n" "fmla z9.h, z28.h, z0.h[2]\n" "ld1h { z29.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z13.h, z28.h, z1.h[2]\n" "fmla z17.h, z28.h, z2.h[2]\n" + "addvl x10, x10, #1\n" "fmla z21.h, z28.h, z3.h[2]\n" "fmla z25.h, z28.h, z4.h[2]\n" "ld1h { z28.h }, p5/Z, [x9]\n" @@ -2296,19 +2294,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 68f\n" "ld1h { z29.h }, p5/Z, [x12]\n" "ld1h { z28.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z29.h, z0.h[3]\n" "fmla z12.h, z29.h, z1.h[3]\n" "fmla z16.h, z29.h, z2.h[3]\n" "fmla z20.h, z29.h, z3.h[3]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z29.h, z4.h[3]\n" "fmla z9.h, z28.h, z0.h[3]\n" "ld1h { z29.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z13.h, z28.h, z1.h[3]\n" "fmla z17.h, z28.h, z2.h[3]\n" + "addvl x10, x10, #1\n" "fmla z21.h, z28.h, z3.h[3]\n" "fmla z25.h, z28.h, z4.h[3]\n" "ld1h { z28.h }, p5/Z, [x9]\n" @@ -2326,19 +2324,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 68f\n" "ld1h { z29.h }, p5/Z, [x12]\n" "ld1h { z28.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z29.h, z0.h[4]\n" "fmla z12.h, z29.h, z1.h[4]\n" "fmla z16.h, z29.h, z2.h[4]\n" "fmla z20.h, z29.h, z3.h[4]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z29.h, z4.h[4]\n" "fmla z9.h, z28.h, z0.h[4]\n" "ld1h { z29.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z13.h, z28.h, z1.h[4]\n" "fmla z17.h, z28.h, z2.h[4]\n" + "addvl x10, x10, #1\n" "fmla z21.h, z28.h, z3.h[4]\n" "fmla z25.h, z28.h, z4.h[4]\n" "ld1h { z28.h }, p5/Z, [x9]\n" @@ -2356,19 +2354,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 68f\n" "ld1h { z29.h }, p5/Z, [x12]\n" "ld1h { z28.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z29.h, z0.h[5]\n" "fmla z12.h, z29.h, z1.h[5]\n" "fmla z16.h, z29.h, z2.h[5]\n" "fmla z20.h, z29.h, z3.h[5]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z29.h, z4.h[5]\n" "fmla z9.h, z28.h, z0.h[5]\n" "ld1h { z29.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z13.h, z28.h, z1.h[5]\n" "fmla z17.h, z28.h, z2.h[5]\n" + "addvl x10, x10, #1\n" "fmla z21.h, z28.h, z3.h[5]\n" "fmla z25.h, z28.h, z4.h[5]\n" "ld1h { z28.h }, p5/Z, [x9]\n" @@ -2386,19 +2384,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 68f\n" "ld1h { z29.h }, p5/Z, [x12]\n" "ld1h { z28.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z29.h, z0.h[6]\n" "fmla z12.h, z29.h, z1.h[6]\n" "fmla z16.h, z29.h, z2.h[6]\n" "fmla z20.h, z29.h, z3.h[6]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z29.h, z4.h[6]\n" "fmla z9.h, z28.h, z0.h[6]\n" "ld1h { z29.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z13.h, z28.h, z1.h[6]\n" "fmla z17.h, z28.h, z2.h[6]\n" + "addvl x10, x10, #1\n" "fmla z21.h, z28.h, z3.h[6]\n" "fmla z25.h, z28.h, z4.h[6]\n" "ld1h { z28.h }, p5/Z, [x9]\n" @@ -2416,12 +2414,12 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 68f\n" "ld1h { z29.h }, p5/Z, [x12]\n" "ld1h { z28.h }, p5/Z, [x11]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z29.h, z0.h[7]\n" "fmla z12.h, z29.h, z1.h[7]\n" "fmla z16.h, z29.h, z2.h[7]\n" "fmla z20.h, z29.h, z3.h[7]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z24.h, z29.h, z4.h[7]\n" "fmla z9.h, z28.h, z0.h[7]\n" "ld1h { z29.h }, p5/Z, [x10]\n" @@ -2448,14 +2446,14 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "cmp x28, x20\n" "bne 63b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x26, x13, x20, LSL #1\n" - "add x25, x26, x20, LSL #1\n" + "add x25, x13, 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" "tbz %x[flags], #1, 69f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rh { z29.h }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rh { z29.h }, p5/Z, [x21]\n" "ld1rh { z28.h }, p5/Z, [x20]\n" "fmin z8.h, p5/M, z8.h, z29.h\n" "fmin z9.h, p5/M, z9.h, z29.h\n" @@ -2503,22 +2501,22 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "st1h { z10.h }, p2, [x13, #2, MUL VL]\n" "st1h { z11.h }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1h { z12.h }, p4, [x26]\n" - "st1h { z13.h }, p3, [x26, #1, MUL VL]\n" - "st1h { z14.h }, p2, [x26, #2, MUL VL]\n" - "st1h { z15.h }, p1, [x26, #3, MUL VL]\n" - "st1h { z16.h }, p4, [x25]\n" - "st1h { z17.h }, p3, [x25, #1, MUL VL]\n" - "st1h { z18.h }, p2, [x25, #2, MUL VL]\n" - "st1h { z19.h }, p1, [x25, #3, MUL VL]\n" - "st1h { z20.h }, p4, [x24]\n" - "st1h { z21.h }, p3, [x24, #1, MUL VL]\n" - "st1h { z22.h }, p2, [x24, #2, MUL VL]\n" - "st1h { z23.h }, p1, [x24, #3, MUL VL]\n" - "st1h { z24.h }, p4, [x23]\n" - "st1h { z25.h }, p3, [x23, #1, MUL VL]\n" - "st1h { z26.h }, p2, [x23, #2, MUL VL]\n" - "st1h { z27.h }, p1, [x23, #3, MUL VL]\n" + "st1h { z12.h }, p4, [x25]\n" + "st1h { z13.h }, p3, [x25, #1, MUL VL]\n" + "st1h { z14.h }, p2, [x25, #2, MUL VL]\n" + "st1h { z15.h }, p1, [x25, #3, MUL VL]\n" + "st1h { z16.h }, p4, [x24]\n" + "st1h { z17.h }, p3, [x24, #1, MUL VL]\n" + "st1h { z18.h }, p2, [x24, #2, MUL VL]\n" + "st1h { z19.h }, p1, [x24, #3, MUL VL]\n" + "st1h { z20.h }, p4, [x23]\n" + "st1h { z21.h }, p3, [x23, #1, MUL VL]\n" + "st1h { z22.h }, p2, [x23, #2, MUL VL]\n" + "st1h { z23.h }, p1, [x23, #3, MUL VL]\n" + "st1h { z24.h }, p4, [x22]\n" + "st1h { z25.h }, p3, [x22, #1, MUL VL]\n" + "st1h { z26.h }, p2, [x22, #2, MUL VL]\n" + "st1h { z27.h }, p1, [x22, #3, MUL VL]\n" "70:" // Height 5: Writeback done "dech x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -2526,19 +2524,18 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "b 86f\n" "71:" // Height 6 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" - "mov x21, #0xc\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "mov x21, #0xc\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" - "madd x21, x20, x21, x13\n" - "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x15, %x[bias]\n" + "mov x13, %x[output_ptr]\n" + "madd %x[output_ptr], x20, x21, %x[output_ptr]\n" "72:" // Height 6: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "cnth x21, ALL, MUL #3\n" "add x11, x12, x20, LSL #1\n" + "cnth x21, ALL, MUL #3\n" "add x10, x11, x20, LSL #1\n" "add x9, x10, x20, LSL #1\n" "add x20, x9, x20, LSL #1\n" @@ -2546,12 +2543,12 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 73f\n" "dech x21\n" - "mov x9, x12\n" "cmp x14, x21\n" + "mov x9, x12\n" "bgt 73f\n" "dech x21\n" - "mov x10, x12\n" "cmp x14, x21\n" + "mov x10, x12\n" "bgt 73f\n" "mov x11, x12\n" "73:" // Height 6: B setup done @@ -2566,18 +2563,18 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "cbz x15, 74f\n" "ld1h { z8.h }, p5/Z, [x15]\n" "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n" - "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" - "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" - "addvl x15, x15, #4\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" - "mov z16.d, z8.d\n" - "mov z17.d, z9.d\n" - "mov z20.d, z8.d\n" + "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n" + "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n" "mov z14.d, z10.d\n" "mov z15.d, z11.d\n" + "mov z16.d, z8.d\n" + "mov z17.d, z9.d\n" + "addvl x15, x15, #4\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" + "mov z20.d, z8.d\n" "mov z21.d, z9.d\n" "mov z22.d, z10.d\n" "mov z23.d, z11.d\n" @@ -2593,17 +2590,17 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "74:" // Height 6: no bias "tbz %x[flags], #0, 75f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x24, x13, x20, LSL #1\n" + "add x23, x24, x20, LSL #1\n" "ld1h { z8.h }, p4/Z, [x13]\n" + "add x22, x23, x20, LSL #1\n" + "add x21, x22, x20, LSL #1\n" "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n" "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n" + "add x20, x21, x20, LSL #1\n" "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n" - "add x24, x13, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" "ld1h { z12.h }, p4/Z, [x24]\n" "ld1h { z13.h }, p3/Z, [x24, #1, MUL VL]\n" - "add x21, x22, x20, LSL #1\n" - "add x20, x21, x20, LSL #1\n" "ld1h { z14.h }, p2/Z, [x24, #2, MUL VL]\n" "ld1h { z15.h }, p1/Z, [x24, #3, MUL VL]\n" "ld1h { z16.h }, p4/Z, [x23]\n" @@ -2652,8 +2649,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "mov x28, #0x0\n" "77:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 78f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -2684,29 +2681,29 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 81f\n" "80:" // Height 6: Multiply loop: Main loop head "whilelt p0.h, XZR, x27\n" - "ld1h { z1.h }, p5/Z, [x12]\n" - "ld1h { z0.h }, p5/Z, [x11]\n" - "sub x27, x27, #0x8\n" - "cmp x27, #0x8\n" "ld1rqh { z7.h }, p0/Z, [x26]\n" - "add x26, x26, #0x10\n" "ld1rqh { z6.h }, p0/Z, [x25]\n" - "add x25, x25, #0x10\n" + "sub x27, x27, #0x8\n" "ld1rqh { z5.h }, p0/Z, [x24]\n" - "add x24, x24, #0x10\n" "ld1rqh { z4.h }, p0/Z, [x23]\n" + "cmp x27, #0x8\n" + "add x26, x26, #0x10\n" "ld1rqh { z3.h }, p0/Z, [x22]\n" "ld1rqh { z2.h }, p0/Z, [x21]\n" - "add x23, x23, #0x10\n" - "add x22, x22, #0x10\n" + "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" + "ld1h { z1.h }, p5/Z, [x12]\n" + "ld1h { z0.h }, p5/Z, [x11]\n" "fmla z8.h, z1.h, z7.h[0]\n" "fmla z12.h, z1.h, z6.h[0]\n" - "add x21, x21, #0x10\n" "fmla z16.h, z1.h, z5.h[0]\n" "fmla z20.h, z1.h, z4.h[0]\n" + "add x23, x23, #0x10\n" + "add x22, x22, #0x10\n" "fmla z24.h, z1.h, z3.h[0]\n" "fmla z28.h, z1.h, z2.h[0]\n" "ld1h { z1.h }, p5/Z, [x10]\n" + "add x21, x21, #0x10\n" "fmla z9.h, z0.h, z7.h[0]\n" "fmla z13.h, z0.h, z6.h[0]\n" "fmla z17.h, z0.h, z5.h[0]\n" @@ -2929,27 +2926,27 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "bgt 80b\n" "81:" // Height 6: Multiply loop: Single iteration only "whilelt p0.h, XZR, x27\n" - "ld1h { z7.h }, p5/Z, [x12]\n" - "ld1h { z6.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "ld1rqh { z0.h }, p0/Z, [x26]\n" "ld1rqh { z1.h }, p0/Z, [x25]\n" + "subs x27, x27, #0x1\n" "ld1rqh { z2.h }, p0/Z, [x24]\n" "ld1rqh { z3.h }, p0/Z, [x23]\n" "ld1rqh { z4.h }, p0/Z, [x22]\n" "ld1rqh { z5.h }, p0/Z, [x21]\n" + "ld1h { z7.h }, p5/Z, [x12]\n" + "ld1h { z6.h }, p5/Z, [x11]\n" "fmla z8.h, z7.h, z0.h[0]\n" "fmla z12.h, z7.h, z1.h[0]\n" - "fmla z9.h, z6.h, z0.h[0]\n" - "fmla z13.h, z6.h, z1.h[0]\n" "fmla z16.h, z7.h, z2.h[0]\n" "fmla z20.h, z7.h, z3.h[0]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z24.h, z7.h, z4.h[0]\n" "fmla z28.h, z7.h, z5.h[0]\n" "ld1h { z7.h }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" + "fmla z9.h, z6.h, z0.h[0]\n" + "fmla z13.h, z6.h, z1.h[0]\n" "fmla z17.h, z6.h, z2.h[0]\n" "fmla z21.h, z6.h, z3.h[0]\n" "fmla z25.h, z6.h, z4.h[0]\n" @@ -2971,19 +2968,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 82f\n" "ld1h { z7.h }, p5/Z, [x12]\n" "ld1h { z6.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z7.h, z0.h[1]\n" "fmla z12.h, z7.h, z1.h[1]\n" "fmla z16.h, z7.h, z2.h[1]\n" "fmla z20.h, z7.h, z3.h[1]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z7.h, z4.h[1]\n" "fmla z28.h, z7.h, z5.h[1]\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z9.h, z6.h, z0.h[1]\n" "fmla z13.h, z6.h, z1.h[1]\n" + "addvl x10, x10, #1\n" "fmla z17.h, z6.h, z2.h[1]\n" "fmla z21.h, z6.h, z3.h[1]\n" "fmla z25.h, z6.h, z4.h[1]\n" @@ -3005,19 +3002,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 82f\n" "ld1h { z7.h }, p5/Z, [x12]\n" "ld1h { z6.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z7.h, z0.h[2]\n" "fmla z12.h, z7.h, z1.h[2]\n" "fmla z16.h, z7.h, z2.h[2]\n" "fmla z20.h, z7.h, z3.h[2]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z7.h, z4.h[2]\n" "fmla z28.h, z7.h, z5.h[2]\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z9.h, z6.h, z0.h[2]\n" "fmla z13.h, z6.h, z1.h[2]\n" + "addvl x10, x10, #1\n" "fmla z17.h, z6.h, z2.h[2]\n" "fmla z21.h, z6.h, z3.h[2]\n" "fmla z25.h, z6.h, z4.h[2]\n" @@ -3039,19 +3036,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 82f\n" "ld1h { z7.h }, p5/Z, [x12]\n" "ld1h { z6.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z7.h, z0.h[3]\n" "fmla z12.h, z7.h, z1.h[3]\n" "fmla z16.h, z7.h, z2.h[3]\n" "fmla z20.h, z7.h, z3.h[3]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z7.h, z4.h[3]\n" "fmla z28.h, z7.h, z5.h[3]\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z9.h, z6.h, z0.h[3]\n" "fmla z13.h, z6.h, z1.h[3]\n" + "addvl x10, x10, #1\n" "fmla z17.h, z6.h, z2.h[3]\n" "fmla z21.h, z6.h, z3.h[3]\n" "fmla z25.h, z6.h, z4.h[3]\n" @@ -3073,19 +3070,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 82f\n" "ld1h { z7.h }, p5/Z, [x12]\n" "ld1h { z6.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z7.h, z0.h[4]\n" "fmla z12.h, z7.h, z1.h[4]\n" "fmla z16.h, z7.h, z2.h[4]\n" "fmla z20.h, z7.h, z3.h[4]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z7.h, z4.h[4]\n" "fmla z28.h, z7.h, z5.h[4]\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z9.h, z6.h, z0.h[4]\n" "fmla z13.h, z6.h, z1.h[4]\n" + "addvl x10, x10, #1\n" "fmla z17.h, z6.h, z2.h[4]\n" "fmla z21.h, z6.h, z3.h[4]\n" "fmla z25.h, z6.h, z4.h[4]\n" @@ -3107,19 +3104,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 82f\n" "ld1h { z7.h }, p5/Z, [x12]\n" "ld1h { z6.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z7.h, z0.h[5]\n" "fmla z12.h, z7.h, z1.h[5]\n" "fmla z16.h, z7.h, z2.h[5]\n" "fmla z20.h, z7.h, z3.h[5]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z7.h, z4.h[5]\n" "fmla z28.h, z7.h, z5.h[5]\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z9.h, z6.h, z0.h[5]\n" "fmla z13.h, z6.h, z1.h[5]\n" + "addvl x10, x10, #1\n" "fmla z17.h, z6.h, z2.h[5]\n" "fmla z21.h, z6.h, z3.h[5]\n" "fmla z25.h, z6.h, z4.h[5]\n" @@ -3141,19 +3138,19 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 82f\n" "ld1h { z7.h }, p5/Z, [x12]\n" "ld1h { z6.h }, p5/Z, [x11]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z7.h, z0.h[6]\n" "fmla z12.h, z7.h, z1.h[6]\n" "fmla z16.h, z7.h, z2.h[6]\n" "fmla z20.h, z7.h, z3.h[6]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" "fmla z24.h, z7.h, z4.h[6]\n" "fmla z28.h, z7.h, z5.h[6]\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "addvl x10, x10, #1\n" + "addvl x11, x11, #1\n" "fmla z9.h, z6.h, z0.h[6]\n" "fmla z13.h, z6.h, z1.h[6]\n" + "addvl x10, x10, #1\n" "fmla z17.h, z6.h, z2.h[6]\n" "fmla z21.h, z6.h, z3.h[6]\n" "fmla z25.h, z6.h, z4.h[6]\n" @@ -3175,12 +3172,12 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "ble 82f\n" "ld1h { z7.h }, p5/Z, [x12]\n" "ld1h { z6.h }, p5/Z, [x11]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z8.h, z7.h, z0.h[7]\n" "fmla z12.h, z7.h, z1.h[7]\n" "fmla z16.h, z7.h, z2.h[7]\n" "fmla z20.h, z7.h, z3.h[7]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z24.h, z7.h, z4.h[7]\n" "fmla z28.h, z7.h, z5.h[7]\n" "ld1h { z7.h }, p5/Z, [x10]\n" @@ -3211,15 +3208,15 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "cmp x28, x20\n" "bne 77b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x26, x13, x20, LSL #1\n" - "add x25, x26, x20, LSL #1\n" + "add x25, x13, 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" "tbz %x[flags], #1, 83f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rh { z1.h }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rh { z1.h }, p5/Z, [x21]\n" "ld1rh { z0.h }, p5/Z, [x20]\n" "fmin z8.h, p5/M, z8.h, z1.h\n" "fmin z9.h, p5/M, z9.h, z1.h\n" @@ -3275,26 +3272,26 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "st1h { z10.h }, p2, [x13, #2, MUL VL]\n" "st1h { z11.h }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1h { z12.h }, p4, [x26]\n" - "st1h { z13.h }, p3, [x26, #1, MUL VL]\n" - "st1h { z14.h }, p2, [x26, #2, MUL VL]\n" - "st1h { z15.h }, p1, [x26, #3, MUL VL]\n" - "st1h { z16.h }, p4, [x25]\n" - "st1h { z17.h }, p3, [x25, #1, MUL VL]\n" - "st1h { z18.h }, p2, [x25, #2, MUL VL]\n" - "st1h { z19.h }, p1, [x25, #3, MUL VL]\n" - "st1h { z20.h }, p4, [x24]\n" - "st1h { z21.h }, p3, [x24, #1, MUL VL]\n" - "st1h { z22.h }, p2, [x24, #2, MUL VL]\n" - "st1h { z23.h }, p1, [x24, #3, MUL VL]\n" - "st1h { z24.h }, p4, [x23]\n" - "st1h { z25.h }, p3, [x23, #1, MUL VL]\n" - "st1h { z26.h }, p2, [x23, #2, MUL VL]\n" - "st1h { z27.h }, p1, [x23, #3, MUL VL]\n" - "st1h { z28.h }, p4, [x22]\n" - "st1h { z29.h }, p3, [x22, #1, MUL VL]\n" - "st1h { z30.h }, p2, [x22, #2, MUL VL]\n" - "st1h { z31.h }, p1, [x22, #3, MUL VL]\n" + "st1h { z12.h }, p4, [x25]\n" + "st1h { z13.h }, p3, [x25, #1, MUL VL]\n" + "st1h { z14.h }, p2, [x25, #2, MUL VL]\n" + "st1h { z15.h }, p1, [x25, #3, MUL VL]\n" + "st1h { z16.h }, p4, [x24]\n" + "st1h { z17.h }, p3, [x24, #1, MUL VL]\n" + "st1h { z18.h }, p2, [x24, #2, MUL VL]\n" + "st1h { z19.h }, p1, [x24, #3, MUL VL]\n" + "st1h { z20.h }, p4, [x23]\n" + "st1h { z21.h }, p3, [x23, #1, MUL VL]\n" + "st1h { z22.h }, p2, [x23, #2, MUL VL]\n" + "st1h { z23.h }, p1, [x23, #3, MUL VL]\n" + "st1h { z24.h }, p4, [x22]\n" + "st1h { z25.h }, p3, [x22, #1, MUL VL]\n" + "st1h { z26.h }, p2, [x22, #2, MUL VL]\n" + "st1h { z27.h }, p1, [x22, #3, MUL VL]\n" + "st1h { z28.h }, p4, [x21]\n" + "st1h { z29.h }, p3, [x21, #1, MUL VL]\n" + "st1h { z30.h }, p2, [x21, #2, MUL VL]\n" + "st1h { z31.h }, p1, [x21, #3, MUL VL]\n" "84:" // Height 6: Writeback done "dech x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -3311,8 +3308,8 @@ void sve_ffhybrid_fp16_mla_6x4VL ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "86:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr) - : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [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_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) + : [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_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "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" ); } |