diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp | 2282 |
1 files changed, 1141 insertions, 1141 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp index eb057e7734..b1ab31e618 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp @@ -163,11 +163,11 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "7:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 8f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" "cbnz x28, 9f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -180,40 +180,40 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "10:" // Height 1: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x12, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "ld1w { z7.s }, p5/Z, [x11, #1, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9, #1, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x12, #2, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "ld1w { z7.s }, p5/Z, [x11, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9, #2, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x12, #3, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "ld1w { z7.s }, p5/Z, [x11, #3, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x12]\n" + "fmla z8.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z9.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x10]\n" + "fmla z10.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" + "fmla z11.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x12, #1, MUL VL]\n" + "fmla z8.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x11, #1, MUL VL]\n" + "fmla z9.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z10.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x9, #1, MUL VL]\n" + "fmla z11.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x12, #2, MUL VL]\n" + "fmla z8.s, z16.s, z0.s[2]\n" + "ld1w { z16.s }, p5/Z, [x11, #2, MUL VL]\n" + "fmla z9.s, z16.s, z0.s[2]\n" + "ld1w { z16.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z10.s, z16.s, z0.s[2]\n" + "ld1w { z16.s }, p5/Z, [x9, #2, MUL VL]\n" + "fmla z11.s, z16.s, z0.s[2]\n" + "ld1w { z16.s }, p5/Z, [x12, #3, MUL VL]\n" + "fmla z8.s, z16.s, z0.s[3]\n" + "ld1w { z16.s }, p5/Z, [x11, #3, MUL VL]\n" + "fmla z9.s, z16.s, z0.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10, #3, MUL VL]\n" "sub x27, x27, #0x4\n" - "ld1w { z7.s }, p5/Z, [x9, #3, MUL VL]\n" + "ld1w { z16.s }, p5/Z, [x9, #3, MUL VL]\n" "cmp x27, #0x4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" + "fmla z10.s, z17.s, z0.s[3]\n" + "fmla z11.s, z16.s, z0.s[3]\n" "add x26, x26, #0x10\n" "addvl x12, x12, #4\n" "addvl x11, x11, #4\n" @@ -223,56 +223,56 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "11:" // Height 1: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x12]\n" + "fmla z8.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z9.s, z16.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" "subs x27, x27, #0x1\n" - "ld1w { z7.s }, p5/Z, [x9]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" + "fmla z10.s, z17.s, z0.s[0]\n" + "fmla z11.s, z16.s, z0.s[0]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" "ble 12f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z8.s, z17.s, z0.s[1]\n" + "fmla z9.s, z16.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" + "fmla z10.s, z17.s, z0.s[1]\n" + "fmla z11.s, z16.s, z0.s[1]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" "ble 12f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z8.s, z17.s, z0.s[2]\n" + "fmla z9.s, z16.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" + "fmla z10.s, z17.s, z0.s[2]\n" + "fmla z11.s, z16.s, z0.s[2]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" "ble 12f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z8.s, z17.s, z0.s[3]\n" + "fmla z9.s, z16.s, z0.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" + "fmla z10.s, z17.s, z0.s[3]\n" + "fmla z11.s, z16.s, z0.s[3]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" @@ -284,17 +284,17 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "bne 7b\n" "tbz %x[flags], #1, 13f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z17.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" + "ld1rw { z16.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z17.s\n" + "fmin z9.s, p5/M, z9.s, z17.s\n" + "fmin z10.s, p5/M, z10.s, z17.s\n" + "fmin z11.s, p5/M, z11.s, z17.s\n" + "fmax z8.s, p5/M, z8.s, z16.s\n" + "fmax z9.s, p5/M, z9.s, z16.s\n" + "fmax z10.s, p5/M, z10.s, z16.s\n" + "fmax z11.s, p5/M, z11.s, z16.s\n" "13:" // Height 1: No activation "st1w { z8.s }, p4, [x13]\n" "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" @@ -355,15 +355,15 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "18:" // Height 2: no bias "tbz %x[flags], #0, 19f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" + "add x20, x13, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x13]\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x20]\n" + "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n" "b 20f\n" "19:" // Height 2: no accumulate "mov z8.b, #0x0\n" @@ -379,12 +379,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "21:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 22f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" "cbnz x28, 23f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -392,143 +392,143 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 23f\n" "22:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" "23:" // Height 2: input setup done "cmp x27, #0x4\n" "ble 25f\n" "24:" // Height 2: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "ld1rqw { z0.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "ld1w { z6.s }, p5/Z, [x12, #1, MUL VL]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z8.s, z17.s, z1.s[0]\n" + "fmla z12.s, z17.s, z0.s[0]\n" + "fmla z9.s, z16.s, z1.s[0]\n" + "fmla z13.s, z16.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" + "fmla z10.s, z17.s, z1.s[0]\n" + "fmla z14.s, z17.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x12, #1, MUL VL]\n" "cmp x27, #0x4\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11, #1, MUL VL]\n" + "fmla z11.s, z16.s, z1.s[0]\n" + "fmla z15.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x11, #1, MUL VL]\n" "add x26, x26, #0x10\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z8.s, z17.s, z1.s[1]\n" + "fmla z12.s, z17.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10, #1, MUL VL]\n" "add x25, x25, #0x10\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9, #1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "ld1w { z6.s }, p5/Z, [x12, #2, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "ld1w { z7.s }, p5/Z, [x11, #2, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9, #2, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "ld1w { z6.s }, p5/Z, [x12, #3, MUL VL]\n" + "fmla z9.s, z16.s, z1.s[1]\n" + "fmla z13.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x9, #1, MUL VL]\n" + "fmla z10.s, z17.s, z1.s[1]\n" + "fmla z14.s, z17.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x12, #2, MUL VL]\n" + "fmla z11.s, z16.s, z1.s[1]\n" + "fmla z15.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x11, #2, MUL VL]\n" + "fmla z8.s, z17.s, z1.s[2]\n" + "fmla z12.s, z17.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z16.s, z1.s[2]\n" + "fmla z13.s, z16.s, z0.s[2]\n" + "ld1w { z16.s }, p5/Z, [x9, #2, MUL VL]\n" + "fmla z10.s, z17.s, z1.s[2]\n" + "fmla z14.s, z17.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x12, #3, MUL VL]\n" "addvl x12, x12, #4\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "ld1w { z7.s }, p5/Z, [x11, #3, MUL VL]\n" + "fmla z11.s, z16.s, z1.s[2]\n" + "fmla z15.s, z16.s, z0.s[2]\n" + "ld1w { z16.s }, p5/Z, [x11, #3, MUL VL]\n" "addvl x11, x11, #4\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z8.s, z17.s, z1.s[3]\n" + "fmla z12.s, z17.s, z0.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "ld1w { z7.s }, p5/Z, [x9, #3, MUL VL]\n" + "fmla z9.s, z16.s, z1.s[3]\n" + "fmla z13.s, z16.s, z0.s[3]\n" + "ld1w { z16.s }, p5/Z, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" + "fmla z10.s, z17.s, z1.s[3]\n" + "fmla z14.s, z17.s, z0.s[3]\n" + "fmla z11.s, z16.s, z1.s[3]\n" + "fmla z15.s, z16.s, z0.s[3]\n" "bgt 24b\n" "25:" // Height 2: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" "ld1rqw { z1.s }, p0/Z, [x25]\n" "subs x27, x27, #0x1\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z8.s, z17.s, z0.s[0]\n" + "fmla z12.s, z17.s, z1.s[0]\n" + "fmla z9.s, z16.s, z0.s[0]\n" + "fmla z13.s, z16.s, z1.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" + "fmla z10.s, z17.s, z0.s[0]\n" + "fmla z14.s, z17.s, z1.s[0]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" + "fmla z11.s, z16.s, z0.s[0]\n" + "fmla z15.s, z16.s, z1.s[0]\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" "ble 26f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z8.s, z17.s, z0.s[1]\n" + "fmla z12.s, z17.s, z1.s[1]\n" + "fmla z9.s, z16.s, z0.s[1]\n" + "fmla z13.s, z16.s, z1.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" + "fmla z10.s, z17.s, z0.s[1]\n" + "fmla z14.s, z17.s, z1.s[1]\n" "addvl x12, x12, #1\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" + "fmla z11.s, z16.s, z0.s[1]\n" + "fmla z15.s, z16.s, z1.s[1]\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" "ble 26f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z8.s, z17.s, z0.s[2]\n" + "fmla z12.s, z17.s, z1.s[2]\n" + "fmla z9.s, z16.s, z0.s[2]\n" + "fmla z13.s, z16.s, z1.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" + "fmla z10.s, z17.s, z0.s[2]\n" + "fmla z14.s, z17.s, z1.s[2]\n" "addvl x12, x12, #1\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" + "fmla z11.s, z16.s, z0.s[2]\n" + "fmla z15.s, z16.s, z1.s[2]\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" "ble 26f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" + "ld1w { z16.s }, p5/Z, [x11]\n" + "fmla z8.s, z17.s, z0.s[3]\n" + "fmla z12.s, z17.s, z1.s[3]\n" + "fmla z9.s, z16.s, z0.s[3]\n" + "fmla z13.s, z16.s, z1.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" + "fmla z10.s, z17.s, z0.s[3]\n" + "fmla z14.s, z17.s, z1.s[3]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" + "fmla z11.s, z16.s, z0.s[3]\n" + "fmla z15.s, z16.s, z1.s[3]\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" "26:" // Height 2: Multiply loop: multiply skip @@ -540,25 +540,25 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "add x25, x13, x20, LSL #2\n" "tbz %x[flags], #1, 27f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z17.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" + "ld1rw { z16.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z17.s\n" + "fmin z9.s, p5/M, z9.s, z17.s\n" + "fmin z10.s, p5/M, z10.s, z17.s\n" + "fmin z11.s, p5/M, z11.s, z17.s\n" + "fmin z12.s, p5/M, z12.s, z17.s\n" + "fmin z13.s, p5/M, z13.s, z17.s\n" + "fmin z14.s, p5/M, z14.s, z17.s\n" + "fmin z15.s, p5/M, z15.s, z17.s\n" + "fmax z8.s, p5/M, z8.s, z16.s\n" + "fmax z9.s, p5/M, z9.s, z16.s\n" + "fmax z10.s, p5/M, z10.s, z16.s\n" + "fmax z11.s, p5/M, z11.s, z16.s\n" + "fmax z12.s, p5/M, z12.s, z16.s\n" + "fmax z13.s, p5/M, z13.s, z16.s\n" + "fmax z14.s, p5/M, z14.s, z16.s\n" + "fmax z15.s, p5/M, z15.s, z16.s\n" "27:" // Height 2: No activation "st1w { z8.s }, p4, [x13]\n" "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" @@ -627,20 +627,20 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "32:" // Height 3: no bias "tbz %x[flags], #0, 33f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x21, x13, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x13]\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x24]\n" - "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x21]\n" + "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x20]\n" + "ld1w { z17.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x20, #3, MUL VL]\n" "b 34f\n" "33:" // Height 3: no accumulate "mov z8.b, #0x0\n" @@ -660,13 +660,13 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "35:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 36f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" "cbnz x28, 37f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -675,89 +675,89 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 37f\n" "36:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" "37:" // Height 3: input setup done "cmp x27, #0x4\n" "ble 39f\n" "38:" // Height 3: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" + "ld1rqw { z2.s }, p0/Z, [x26]\n" "ld1rqw { z1.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "ld1rqw { z0.s }, p0/Z, [x24]\n" + "ld1w { z21.s }, p5/Z, [x12]\n" + "fmla z8.s, z21.s, z2.s[0]\n" + "fmla z12.s, z21.s, z1.s[0]\n" + "ld1w { z20.s }, p5/Z, [x11]\n" + "fmla z16.s, z21.s, z0.s[0]\n" + "fmla z9.s, z20.s, z2.s[0]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" + "fmla z13.s, z20.s, z1.s[0]\n" + "fmla z17.s, z20.s, z0.s[0]\n" + "ld1w { z20.s }, p5/Z, [x9]\n" "cmp x27, #0x4\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" + "fmla z10.s, z21.s, z2.s[0]\n" + "fmla z14.s, z21.s, z1.s[0]\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x12, #1, MUL VL]\n" + "fmla z18.s, z21.s, z0.s[0]\n" + "fmla z11.s, z20.s, z2.s[0]\n" + "ld1w { z21.s }, p5/Z, [x12, #1, MUL VL]\n" "add x24, x24, #0x10\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9, #1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x12, #2, MUL VL]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "ld1w { z7.s }, p5/Z, [x11, #2, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9, #2, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x12, #3, MUL VL]\n" + "fmla z15.s, z20.s, z1.s[0]\n" + "fmla z19.s, z20.s, z0.s[0]\n" + "ld1w { z20.s }, p5/Z, [x11, #1, MUL VL]\n" + "fmla z8.s, z21.s, z2.s[1]\n" + "fmla z12.s, z21.s, z1.s[1]\n" + "fmla z16.s, z21.s, z0.s[1]\n" + "fmla z9.s, z20.s, z2.s[1]\n" + "ld1w { z21.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[1]\n" + "fmla z17.s, z20.s, z0.s[1]\n" + "ld1w { z20.s }, p5/Z, [x9, #1, MUL VL]\n" + "fmla z10.s, z21.s, z2.s[1]\n" + "fmla z14.s, z21.s, z1.s[1]\n" + "fmla z18.s, z21.s, z0.s[1]\n" + "fmla z11.s, z20.s, z2.s[1]\n" + "ld1w { z21.s }, p5/Z, [x12, #2, MUL VL]\n" + "fmla z15.s, z20.s, z1.s[1]\n" + "fmla z19.s, z20.s, z0.s[1]\n" + "ld1w { z20.s }, p5/Z, [x11, #2, MUL VL]\n" + "fmla z8.s, z21.s, z2.s[2]\n" + "fmla z12.s, z21.s, z1.s[2]\n" + "fmla z16.s, z21.s, z0.s[2]\n" + "fmla z9.s, z20.s, z2.s[2]\n" + "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[2]\n" + "fmla z17.s, z20.s, z0.s[2]\n" + "ld1w { z20.s }, p5/Z, [x9, #2, MUL VL]\n" + "fmla z10.s, z21.s, z2.s[2]\n" + "fmla z14.s, z21.s, z1.s[2]\n" + "fmla z18.s, z21.s, z0.s[2]\n" + "fmla z11.s, z20.s, z2.s[2]\n" + "ld1w { z21.s }, p5/Z, [x12, #3, MUL VL]\n" "addvl x12, x12, #4\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "ld1w { z7.s }, p5/Z, [x11, #3, MUL VL]\n" + "fmla z15.s, z20.s, z1.s[2]\n" + "fmla z19.s, z20.s, z0.s[2]\n" + "ld1w { z20.s }, p5/Z, [x11, #3, MUL VL]\n" "addvl x11, x11, #4\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z8.s, z21.s, z2.s[3]\n" + "fmla z12.s, z21.s, z1.s[3]\n" + "fmla z16.s, z21.s, z0.s[3]\n" + "fmla z9.s, z20.s, z2.s[3]\n" + "ld1w { z21.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "ld1w { z7.s }, p5/Z, [x9, #3, MUL VL]\n" + "fmla z13.s, z20.s, z1.s[3]\n" + "fmla z17.s, z20.s, z0.s[3]\n" + "ld1w { z20.s }, p5/Z, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" + "fmla z10.s, z21.s, z2.s[3]\n" + "fmla z14.s, z21.s, z1.s[3]\n" + "fmla z18.s, z21.s, z0.s[3]\n" + "fmla z11.s, z20.s, z2.s[3]\n" + "fmla z15.s, z20.s, z1.s[3]\n" + "fmla z19.s, z20.s, z0.s[3]\n" "bgt 38b\n" "39:" // Height 3: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" @@ -765,91 +765,91 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ld1rqw { z1.s }, p0/Z, [x25]\n" "subs x27, x27, #0x1\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "ld1w { z21.s }, p5/Z, [x12]\n" + "fmla z8.s, z21.s, z0.s[0]\n" + "fmla z12.s, z21.s, z1.s[0]\n" + "ld1w { z20.s }, p5/Z, [x11]\n" + "fmla z16.s, z21.s, z2.s[0]\n" + "fmla z9.s, z20.s, z0.s[0]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" + "fmla z13.s, z20.s, z1.s[0]\n" + "fmla z17.s, z20.s, z2.s[0]\n" + "ld1w { z20.s }, p5/Z, [x9]\n" "addvl x12, x12, #1\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" + "fmla z10.s, z21.s, z0.s[0]\n" + "fmla z14.s, z21.s, z1.s[0]\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" + "fmla z18.s, z21.s, z2.s[0]\n" + "fmla z11.s, z20.s, z0.s[0]\n" "addvl x9, x9, #1\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" + "fmla z15.s, z20.s, z1.s[0]\n" + "fmla z19.s, z20.s, z2.s[0]\n" "ble 40f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z21.s }, p5/Z, [x12]\n" + "ld1w { z20.s }, p5/Z, [x11]\n" + "fmla z8.s, z21.s, z0.s[1]\n" + "fmla z12.s, z21.s, z1.s[1]\n" + "fmla z16.s, z21.s, z2.s[1]\n" + "fmla z9.s, z20.s, z0.s[1]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" "subs x27, x27, #0x1\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z13.s, z20.s, z1.s[1]\n" + "fmla z17.s, z20.s, z2.s[1]\n" + "ld1w { z20.s }, p5/Z, [x9]\n" "addvl x12, x12, #1\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" + "fmla z10.s, z21.s, z0.s[1]\n" + "fmla z14.s, z21.s, z1.s[1]\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" + "fmla z18.s, z21.s, z2.s[1]\n" + "fmla z11.s, z20.s, z0.s[1]\n" "addvl x9, x9, #1\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" + "fmla z15.s, z20.s, z1.s[1]\n" + "fmla z19.s, z20.s, z2.s[1]\n" "ble 40f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z21.s }, p5/Z, [x12]\n" + "ld1w { z20.s }, p5/Z, [x11]\n" + "fmla z8.s, z21.s, z0.s[2]\n" + "fmla z12.s, z21.s, z1.s[2]\n" + "fmla z16.s, z21.s, z2.s[2]\n" + "fmla z9.s, z20.s, z0.s[2]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" "subs x27, x27, #0x1\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z13.s, z20.s, z1.s[2]\n" + "fmla z17.s, z20.s, z2.s[2]\n" + "ld1w { z20.s }, p5/Z, [x9]\n" "addvl x12, x12, #1\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" + "fmla z10.s, z21.s, z0.s[2]\n" + "fmla z14.s, z21.s, z1.s[2]\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" + "fmla z18.s, z21.s, z2.s[2]\n" + "fmla z11.s, z20.s, z0.s[2]\n" "addvl x9, x9, #1\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" + "fmla z15.s, z20.s, z1.s[2]\n" + "fmla z19.s, z20.s, z2.s[2]\n" "ble 40f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z21.s }, p5/Z, [x12]\n" + "ld1w { z20.s }, p5/Z, [x11]\n" + "fmla z8.s, z21.s, z0.s[3]\n" + "fmla z12.s, z21.s, z1.s[3]\n" + "fmla z16.s, z21.s, z2.s[3]\n" + "fmla z9.s, z20.s, z0.s[3]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" "addvl x12, x12, #1\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z13.s, z20.s, z1.s[3]\n" + "fmla z17.s, z20.s, z2.s[3]\n" + "ld1w { z20.s }, p5/Z, [x9]\n" "addvl x11, x11, #1\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" + "fmla z10.s, z21.s, z0.s[3]\n" + "fmla z14.s, z21.s, z1.s[3]\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" + "fmla z18.s, z21.s, z2.s[3]\n" + "fmla z11.s, z20.s, z0.s[3]\n" + "fmla z15.s, z20.s, z1.s[3]\n" + "fmla z19.s, z20.s, z2.s[3]\n" "40:" // Height 3: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -860,33 +860,33 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "add x24, x25, x20, LSL #2\n" "tbz %x[flags], #1, 41f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z21.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmin z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z1.s\n" - "fmin z18.s, p5/M, z18.s, z1.s\n" - "fmin z19.s, p5/M, z19.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" - "fmax z16.s, p5/M, z16.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z0.s\n" - "fmax z18.s, p5/M, z18.s, z0.s\n" - "fmax z19.s, p5/M, z19.s, z0.s\n" + "ld1rw { z20.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z21.s\n" + "fmin z9.s, p5/M, z9.s, z21.s\n" + "fmin z10.s, p5/M, z10.s, z21.s\n" + "fmin z11.s, p5/M, z11.s, z21.s\n" + "fmin z12.s, p5/M, z12.s, z21.s\n" + "fmin z13.s, p5/M, z13.s, z21.s\n" + "fmin z14.s, p5/M, z14.s, z21.s\n" + "fmin z15.s, p5/M, z15.s, z21.s\n" + "fmin z16.s, p5/M, z16.s, z21.s\n" + "fmin z17.s, p5/M, z17.s, z21.s\n" + "fmin z18.s, p5/M, z18.s, z21.s\n" + "fmin z19.s, p5/M, z19.s, z21.s\n" + "fmax z8.s, p5/M, z8.s, z20.s\n" + "fmax z9.s, p5/M, z9.s, z20.s\n" + "fmax z10.s, p5/M, z10.s, z20.s\n" + "fmax z11.s, p5/M, z11.s, z20.s\n" + "fmax z12.s, p5/M, z12.s, z20.s\n" + "fmax z13.s, p5/M, z13.s, z20.s\n" + "fmax z14.s, p5/M, z14.s, z20.s\n" + "fmax z15.s, p5/M, z15.s, z20.s\n" + "fmax z16.s, p5/M, z16.s, z20.s\n" + "fmax z17.s, p5/M, z17.s, z20.s\n" + "fmax z18.s, p5/M, z18.s, z20.s\n" + "fmax z19.s, p5/M, z19.s, z20.s\n" "41:" // Height 3: No activation "st1w { z8.s }, p4, [x13]\n" "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" @@ -963,25 +963,25 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "46:" // Height 4: no bias "tbz %x[flags], #0, 47f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x22, x13, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x13]\n" - "add x23, x24, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x24]\n" - "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x22]\n" + "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x21]\n" + "ld1w { z17.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x20]\n" + "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n" "b 48f\n" "47:" // Height 4: no accumulate "mov z8.b, #0x0\n" @@ -1005,14 +1005,14 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "49:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 50f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" "cbnz x28, 51f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1022,108 +1022,108 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 51f\n" "50:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "51:" // Height 4: input setup done "cmp x27, #0x4\n" "ble 53f\n" "52:" // Height 4: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" + "ld1rqw { z3.s }, p0/Z, [x26]\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" + "ld1rqw { z1.s }, p0/Z, [x24]\n" + "ld1rqw { z0.s }, p0/Z, [x23]\n" "cmp x27, #0x4\n" "add x26, x26, #0x10\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z25.s }, p5/Z, [x12]\n" + "ld1w { z24.s }, p5/Z, [x11]\n" + "fmla z8.s, z25.s, z3.s[0]\n" + "fmla z12.s, z25.s, z2.s[0]\n" + "fmla z16.s, z25.s, z1.s[0]\n" + "fmla z20.s, z25.s, z0.s[0]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" "add x25, x25, #0x10\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" + "fmla z9.s, z24.s, z3.s[0]\n" + "fmla z13.s, z24.s, z2.s[0]\n" "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "ld1w { z6.s }, p5/Z, [x12, #1, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9, #1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "ld1w { z6.s }, p5/Z, [x12, #2, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "ld1w { z7.s }, p5/Z, [x11, #2, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9, #2, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "ld1w { z6.s }, p5/Z, [x12, #3, MUL VL]\n" + "fmla z17.s, z24.s, z1.s[0]\n" + "fmla z21.s, z24.s, z0.s[0]\n" + "ld1w { z24.s }, p5/Z, [x9]\n" + "fmla z10.s, z25.s, z3.s[0]\n" + "fmla z14.s, z25.s, z2.s[0]\n" + "fmla z18.s, z25.s, z1.s[0]\n" + "fmla z22.s, z25.s, z0.s[0]\n" + "ld1w { z25.s }, p5/Z, [x12, #1, MUL VL]\n" + "fmla z11.s, z24.s, z3.s[0]\n" + "fmla z15.s, z24.s, z2.s[0]\n" + "fmla z19.s, z24.s, z1.s[0]\n" + "fmla z23.s, z24.s, z0.s[0]\n" + "ld1w { z24.s }, p5/Z, [x11, #1, MUL VL]\n" + "fmla z8.s, z25.s, z3.s[1]\n" + "fmla z12.s, z25.s, z2.s[1]\n" + "fmla z16.s, z25.s, z1.s[1]\n" + "fmla z20.s, z25.s, z0.s[1]\n" + "ld1w { z25.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z9.s, z24.s, z3.s[1]\n" + "fmla z13.s, z24.s, z2.s[1]\n" + "fmla z17.s, z24.s, z1.s[1]\n" + "fmla z21.s, z24.s, z0.s[1]\n" + "ld1w { z24.s }, p5/Z, [x9, #1, MUL VL]\n" + "fmla z10.s, z25.s, z3.s[1]\n" + "fmla z14.s, z25.s, z2.s[1]\n" + "fmla z18.s, z25.s, z1.s[1]\n" + "fmla z22.s, z25.s, z0.s[1]\n" + "ld1w { z25.s }, p5/Z, [x12, #2, MUL VL]\n" + "fmla z11.s, z24.s, z3.s[1]\n" + "fmla z15.s, z24.s, z2.s[1]\n" + "fmla z19.s, z24.s, z1.s[1]\n" + "fmla z23.s, z24.s, z0.s[1]\n" + "ld1w { z24.s }, p5/Z, [x11, #2, MUL VL]\n" + "fmla z8.s, z25.s, z3.s[2]\n" + "fmla z12.s, z25.s, z2.s[2]\n" + "fmla z16.s, z25.s, z1.s[2]\n" + "fmla z20.s, z25.s, z0.s[2]\n" + "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z24.s, z3.s[2]\n" + "fmla z13.s, z24.s, z2.s[2]\n" + "fmla z17.s, z24.s, z1.s[2]\n" + "fmla z21.s, z24.s, z0.s[2]\n" + "ld1w { z24.s }, p5/Z, [x9, #2, MUL VL]\n" + "fmla z10.s, z25.s, z3.s[2]\n" + "fmla z14.s, z25.s, z2.s[2]\n" + "fmla z18.s, z25.s, z1.s[2]\n" + "fmla z22.s, z25.s, z0.s[2]\n" + "ld1w { z25.s }, p5/Z, [x12, #3, MUL VL]\n" "addvl x12, x12, #4\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "ld1w { z7.s }, p5/Z, [x11, #3, MUL VL]\n" + "fmla z11.s, z24.s, z3.s[2]\n" + "fmla z15.s, z24.s, z2.s[2]\n" + "fmla z19.s, z24.s, z1.s[2]\n" + "fmla z23.s, z24.s, z0.s[2]\n" + "ld1w { z24.s }, p5/Z, [x11, #3, MUL VL]\n" "addvl x11, x11, #4\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z8.s, z25.s, z3.s[3]\n" + "fmla z12.s, z25.s, z2.s[3]\n" + "fmla z16.s, z25.s, z1.s[3]\n" + "fmla z20.s, z25.s, z0.s[3]\n" + "ld1w { z25.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "ld1w { z7.s }, p5/Z, [x9, #3, MUL VL]\n" + "fmla z9.s, z24.s, z3.s[3]\n" + "fmla z13.s, z24.s, z2.s[3]\n" + "fmla z17.s, z24.s, z1.s[3]\n" + "fmla z21.s, z24.s, z0.s[3]\n" + "ld1w { z24.s }, p5/Z, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" + "fmla z10.s, z25.s, z3.s[3]\n" + "fmla z14.s, z25.s, z2.s[3]\n" + "fmla z18.s, z25.s, z1.s[3]\n" + "fmla z22.s, z25.s, z0.s[3]\n" + "fmla z11.s, z24.s, z3.s[3]\n" + "fmla z15.s, z24.s, z2.s[3]\n" + "fmla z19.s, z24.s, z1.s[3]\n" + "fmla z23.s, z24.s, z0.s[3]\n" "bgt 52b\n" "53:" // Height 4: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" @@ -1132,107 +1132,107 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "subs x27, x27, #0x1\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" "ld1rqw { z3.s }, p0/Z, [x23]\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z25.s }, p5/Z, [x12]\n" + "ld1w { z24.s }, p5/Z, [x11]\n" + "fmla z8.s, z25.s, z0.s[0]\n" + "fmla z12.s, z25.s, z1.s[0]\n" + "fmla z16.s, z25.s, z2.s[0]\n" + "fmla z20.s, z25.s, z3.s[0]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" "addvl x12, x12, #1\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" + "fmla z9.s, z24.s, z0.s[0]\n" + "fmla z13.s, z24.s, z1.s[0]\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z17.s, z24.s, z2.s[0]\n" + "fmla z21.s, z24.s, z3.s[0]\n" + "ld1w { z24.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" + "fmla z10.s, z25.s, z0.s[0]\n" + "fmla z14.s, z25.s, z1.s[0]\n" + "fmla z18.s, z25.s, z2.s[0]\n" + "fmla z22.s, z25.s, z3.s[0]\n" + "fmla z11.s, z24.s, z0.s[0]\n" + "fmla z15.s, z24.s, z1.s[0]\n" + "fmla z19.s, z24.s, z2.s[0]\n" + "fmla z23.s, z24.s, z3.s[0]\n" "ble 54f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z25.s }, p5/Z, [x12]\n" + "ld1w { z24.s }, p5/Z, [x11]\n" + "fmla z8.s, z25.s, z0.s[1]\n" + "fmla z12.s, z25.s, z1.s[1]\n" + "fmla z16.s, z25.s, z2.s[1]\n" + "fmla z20.s, z25.s, z3.s[1]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" "subs x27, x27, #0x1\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" + "fmla z9.s, z24.s, z0.s[1]\n" + "fmla z13.s, z24.s, z1.s[1]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z17.s, z24.s, z2.s[1]\n" + "fmla z21.s, z24.s, z3.s[1]\n" + "ld1w { z24.s }, p5/Z, [x9]\n" "addvl x10, x10, #1\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" + "fmla z10.s, z25.s, z0.s[1]\n" + "fmla z14.s, z25.s, z1.s[1]\n" "addvl x9, x9, #1\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" + "fmla z18.s, z25.s, z2.s[1]\n" + "fmla z22.s, z25.s, z3.s[1]\n" + "fmla z11.s, z24.s, z0.s[1]\n" + "fmla z15.s, z24.s, z1.s[1]\n" + "fmla z19.s, z24.s, z2.s[1]\n" + "fmla z23.s, z24.s, z3.s[1]\n" "ble 54f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z25.s }, p5/Z, [x12]\n" + "ld1w { z24.s }, p5/Z, [x11]\n" + "fmla z8.s, z25.s, z0.s[2]\n" + "fmla z12.s, z25.s, z1.s[2]\n" + "fmla z16.s, z25.s, z2.s[2]\n" + "fmla z20.s, z25.s, z3.s[2]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" "subs x27, x27, #0x1\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" + "fmla z9.s, z24.s, z0.s[2]\n" + "fmla z13.s, z24.s, z1.s[2]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z17.s, z24.s, z2.s[2]\n" + "fmla z21.s, z24.s, z3.s[2]\n" + "ld1w { z24.s }, p5/Z, [x9]\n" "addvl x10, x10, #1\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" + "fmla z10.s, z25.s, z0.s[2]\n" + "fmla z14.s, z25.s, z1.s[2]\n" "addvl x9, x9, #1\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" + "fmla z18.s, z25.s, z2.s[2]\n" + "fmla z22.s, z25.s, z3.s[2]\n" + "fmla z11.s, z24.s, z0.s[2]\n" + "fmla z15.s, z24.s, z1.s[2]\n" + "fmla z19.s, z24.s, z2.s[2]\n" + "fmla z23.s, z24.s, z3.s[2]\n" "ble 54f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z25.s }, p5/Z, [x12]\n" + "ld1w { z24.s }, p5/Z, [x11]\n" + "fmla z8.s, z25.s, z0.s[3]\n" + "fmla z12.s, z25.s, z1.s[3]\n" + "fmla z16.s, z25.s, z2.s[3]\n" + "fmla z20.s, z25.s, z3.s[3]\n" + "ld1w { z25.s }, p5/Z, [x10]\n" "addvl x12, x12, #1\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" + "fmla z9.s, z24.s, z0.s[3]\n" + "fmla z13.s, z24.s, z1.s[3]\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z17.s, z24.s, z2.s[3]\n" + "fmla z21.s, z24.s, z3.s[3]\n" + "ld1w { z24.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" + "fmla z10.s, z25.s, z0.s[3]\n" + "fmla z14.s, z25.s, z1.s[3]\n" + "fmla z18.s, z25.s, z2.s[3]\n" + "fmla z22.s, z25.s, z3.s[3]\n" + "fmla z11.s, z24.s, z0.s[3]\n" + "fmla z15.s, z24.s, z1.s[3]\n" + "fmla z19.s, z24.s, z2.s[3]\n" + "fmla z23.s, z24.s, z3.s[3]\n" "54:" // Height 4: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1244,41 +1244,41 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "add x23, x24, x20, LSL #2\n" "tbz %x[flags], #1, 55f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z25.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmin z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z1.s\n" - "fmin z18.s, p5/M, z18.s, z1.s\n" - "fmin z19.s, p5/M, z19.s, z1.s\n" - "fmin z20.s, p5/M, z20.s, z1.s\n" - "fmin z21.s, p5/M, z21.s, z1.s\n" - "fmin z22.s, p5/M, z22.s, z1.s\n" - "fmin z23.s, p5/M, z23.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" - "fmax z16.s, p5/M, z16.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z0.s\n" - "fmax z18.s, p5/M, z18.s, z0.s\n" - "fmax z19.s, p5/M, z19.s, z0.s\n" - "fmax z20.s, p5/M, z20.s, z0.s\n" - "fmax z21.s, p5/M, z21.s, z0.s\n" - "fmax z22.s, p5/M, z22.s, z0.s\n" - "fmax z23.s, p5/M, z23.s, z0.s\n" + "ld1rw { z24.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z25.s\n" + "fmin z9.s, p5/M, z9.s, z25.s\n" + "fmin z10.s, p5/M, z10.s, z25.s\n" + "fmin z11.s, p5/M, z11.s, z25.s\n" + "fmin z12.s, p5/M, z12.s, z25.s\n" + "fmin z13.s, p5/M, z13.s, z25.s\n" + "fmin z14.s, p5/M, z14.s, z25.s\n" + "fmin z15.s, p5/M, z15.s, z25.s\n" + "fmin z16.s, p5/M, z16.s, z25.s\n" + "fmin z17.s, p5/M, z17.s, z25.s\n" + "fmin z18.s, p5/M, z18.s, z25.s\n" + "fmin z19.s, p5/M, z19.s, z25.s\n" + "fmin z20.s, p5/M, z20.s, z25.s\n" + "fmin z21.s, p5/M, z21.s, z25.s\n" + "fmin z22.s, p5/M, z22.s, z25.s\n" + "fmin z23.s, p5/M, z23.s, z25.s\n" + "fmax z8.s, p5/M, z8.s, z24.s\n" + "fmax z9.s, p5/M, z9.s, z24.s\n" + "fmax z10.s, p5/M, z10.s, z24.s\n" + "fmax z11.s, p5/M, z11.s, z24.s\n" + "fmax z12.s, p5/M, z12.s, z24.s\n" + "fmax z13.s, p5/M, z13.s, z24.s\n" + "fmax z14.s, p5/M, z14.s, z24.s\n" + "fmax z15.s, p5/M, z15.s, z24.s\n" + "fmax z16.s, p5/M, z16.s, z24.s\n" + "fmax z17.s, p5/M, z17.s, z24.s\n" + "fmax z18.s, p5/M, z18.s, z24.s\n" + "fmax z19.s, p5/M, z19.s, z24.s\n" + "fmax z20.s, p5/M, z20.s, z24.s\n" + "fmax z21.s, p5/M, z21.s, z24.s\n" + "fmax z22.s, p5/M, z22.s, z24.s\n" + "fmax z23.s, p5/M, z23.s, z24.s\n" "55:" // Height 4: No activation "st1w { z8.s }, p4, [x13]\n" "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" @@ -1363,30 +1363,30 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "60:" // Height 5: no bias "tbz %x[flags], #0, 61f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z8.s }, p4/Z, [x13]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x13, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" + "ld1w { z8.s }, p4/Z, [x13]\n" + "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x24]\n" - "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z24.s }, p4/Z, [x22]\n" - "ld1w { z25.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z26.s }, p2/Z, [x22, #2, MUL VL]\n" - "ld1w { z27.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x23]\n" + "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x22]\n" + "ld1w { z17.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x21]\n" + "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z24.s }, p4/Z, [x20]\n" + "ld1w { z25.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z26.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z27.s }, p1/Z, [x20, #3, MUL VL]\n" "b 62f\n" "61:" // Height 5: no accumulate "mov z8.b, #0x0\n" @@ -1414,15 +1414,15 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "63:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 64f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" "cbnz x28, 65f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1433,127 +1433,127 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 65f\n" "64:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "65:" // Height 5: input setup done "cmp x27, #0x4\n" "ble 67f\n" "66:" // Height 5: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" + "ld1rqw { z4.s }, p0/Z, [x26]\n" + "ld1rqw { z3.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" + "ld1rqw { z1.s }, p0/Z, [x23]\n" "cmp x27, #0x4\n" "add x26, x26, #0x10\n" - "ld1rqw { z4.s }, p0/Z, [x22]\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" + "ld1rqw { z0.s }, p0/Z, [x22]\n" + "ld1w { z29.s }, p5/Z, [x12]\n" + "fmla z8.s, z29.s, z4.s[0]\n" + "fmla z12.s, z29.s, z3.s[0]\n" + "ld1w { z28.s }, p5/Z, [x11]\n" + "fmla z16.s, z29.s, z2.s[0]\n" + "fmla z20.s, z29.s, z1.s[0]\n" "add x25, x25, #0x10\n" - "fmla z24.s, z6.s, z4.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z29.s, z0.s[0]\n" + "fmla z9.s, z28.s, z4.s[0]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" "add x24, x24, #0x10\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" + "fmla z13.s, z28.s, z3.s[0]\n" + "fmla z17.s, z28.s, z2.s[0]\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "fmla z25.s, z7.s, z4.s[0]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z26.s, z6.s, z4.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x12, #1, MUL VL]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "fmla z27.s, z7.s, z4.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" - "fmla z24.s, z6.s, z4.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "fmla z25.s, z7.s, z4.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9, #1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z26.s, z6.s, z4.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x12, #2, MUL VL]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "fmla z27.s, z7.s, z4.s[1]\n" - "ld1w { z7.s }, p5/Z, [x11, #2, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" - "fmla z24.s, z6.s, z4.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "fmla z25.s, z7.s, z4.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9, #2, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z26.s, z6.s, z4.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x12, #3, MUL VL]\n" + "fmla z21.s, z28.s, z1.s[0]\n" + "fmla z25.s, z28.s, z0.s[0]\n" + "ld1w { z28.s }, p5/Z, [x9]\n" + "fmla z10.s, z29.s, z4.s[0]\n" + "fmla z14.s, z29.s, z3.s[0]\n" + "fmla z18.s, z29.s, z2.s[0]\n" + "fmla z22.s, z29.s, z1.s[0]\n" + "fmla z26.s, z29.s, z0.s[0]\n" + "fmla z11.s, z28.s, z4.s[0]\n" + "ld1w { z29.s }, p5/Z, [x12, #1, MUL VL]\n" + "fmla z15.s, z28.s, z3.s[0]\n" + "fmla z19.s, z28.s, z2.s[0]\n" + "fmla z23.s, z28.s, z1.s[0]\n" + "fmla z27.s, z28.s, z0.s[0]\n" + "ld1w { z28.s }, p5/Z, [x11, #1, MUL VL]\n" + "fmla z8.s, z29.s, z4.s[1]\n" + "fmla z12.s, z29.s, z3.s[1]\n" + "fmla z16.s, z29.s, z2.s[1]\n" + "fmla z20.s, z29.s, z1.s[1]\n" + "fmla z24.s, z29.s, z0.s[1]\n" + "fmla z9.s, z28.s, z4.s[1]\n" + "ld1w { z29.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z13.s, z28.s, z3.s[1]\n" + "fmla z17.s, z28.s, z2.s[1]\n" + "fmla z21.s, z28.s, z1.s[1]\n" + "fmla z25.s, z28.s, z0.s[1]\n" + "ld1w { z28.s }, p5/Z, [x9, #1, MUL VL]\n" + "fmla z10.s, z29.s, z4.s[1]\n" + "fmla z14.s, z29.s, z3.s[1]\n" + "fmla z18.s, z29.s, z2.s[1]\n" + "fmla z22.s, z29.s, z1.s[1]\n" + "fmla z26.s, z29.s, z0.s[1]\n" + "fmla z11.s, z28.s, z4.s[1]\n" + "ld1w { z29.s }, p5/Z, [x12, #2, MUL VL]\n" + "fmla z15.s, z28.s, z3.s[1]\n" + "fmla z19.s, z28.s, z2.s[1]\n" + "fmla z23.s, z28.s, z1.s[1]\n" + "fmla z27.s, z28.s, z0.s[1]\n" + "ld1w { z28.s }, p5/Z, [x11, #2, MUL VL]\n" + "fmla z8.s, z29.s, z4.s[2]\n" + "fmla z12.s, z29.s, z3.s[2]\n" + "fmla z16.s, z29.s, z2.s[2]\n" + "fmla z20.s, z29.s, z1.s[2]\n" + "fmla z24.s, z29.s, z0.s[2]\n" + "fmla z9.s, z28.s, z4.s[2]\n" + "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z13.s, z28.s, z3.s[2]\n" + "fmla z17.s, z28.s, z2.s[2]\n" + "fmla z21.s, z28.s, z1.s[2]\n" + "fmla z25.s, z28.s, z0.s[2]\n" + "ld1w { z28.s }, p5/Z, [x9, #2, MUL VL]\n" + "fmla z10.s, z29.s, z4.s[2]\n" + "fmla z14.s, z29.s, z3.s[2]\n" + "fmla z18.s, z29.s, z2.s[2]\n" + "fmla z22.s, z29.s, z1.s[2]\n" + "fmla z26.s, z29.s, z0.s[2]\n" + "fmla z11.s, z28.s, z4.s[2]\n" + "ld1w { z29.s }, p5/Z, [x12, #3, MUL VL]\n" "addvl x12, x12, #4\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "fmla z27.s, z7.s, z4.s[2]\n" - "ld1w { z7.s }, p5/Z, [x11, #3, MUL VL]\n" + "fmla z15.s, z28.s, z3.s[2]\n" + "fmla z19.s, z28.s, z2.s[2]\n" + "fmla z23.s, z28.s, z1.s[2]\n" + "fmla z27.s, z28.s, z0.s[2]\n" + "ld1w { z28.s }, p5/Z, [x11, #3, MUL VL]\n" "addvl x11, x11, #4\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "fmla z24.s, z6.s, z4.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z8.s, z29.s, z4.s[3]\n" + "fmla z12.s, z29.s, z3.s[3]\n" + "fmla z16.s, z29.s, z2.s[3]\n" + "fmla z20.s, z29.s, z1.s[3]\n" + "fmla z24.s, z29.s, z0.s[3]\n" + "fmla z9.s, z28.s, z4.s[3]\n" + "ld1w { z29.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "fmla z25.s, z7.s, z4.s[3]\n" - "ld1w { z7.s }, p5/Z, [x9, #3, MUL VL]\n" + "fmla z13.s, z28.s, z3.s[3]\n" + "fmla z17.s, z28.s, z2.s[3]\n" + "fmla z21.s, z28.s, z1.s[3]\n" + "fmla z25.s, z28.s, z0.s[3]\n" + "ld1w { z28.s }, p5/Z, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z26.s, z6.s, z4.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" - "fmla z27.s, z7.s, z4.s[3]\n" + "fmla z10.s, z29.s, z4.s[3]\n" + "fmla z14.s, z29.s, z3.s[3]\n" + "fmla z18.s, z29.s, z2.s[3]\n" + "fmla z22.s, z29.s, z1.s[3]\n" + "fmla z26.s, z29.s, z0.s[3]\n" + "fmla z11.s, z28.s, z4.s[3]\n" + "fmla z15.s, z28.s, z3.s[3]\n" + "fmla z19.s, z28.s, z2.s[3]\n" + "fmla z23.s, z28.s, z1.s[3]\n" + "fmla z27.s, z28.s, z0.s[3]\n" "bgt 66b\n" "67:" // Height 5: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" @@ -1563,123 +1563,123 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ld1rqw { z2.s }, p0/Z, [x24]\n" "ld1rqw { z3.s }, p0/Z, [x23]\n" "ld1rqw { z4.s }, p0/Z, [x22]\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" + "ld1w { z29.s }, p5/Z, [x12]\n" + "fmla z8.s, z29.s, z0.s[0]\n" + "fmla z12.s, z29.s, z1.s[0]\n" + "ld1w { z28.s }, p5/Z, [x11]\n" + "fmla z16.s, z29.s, z2.s[0]\n" + "fmla z20.s, z29.s, z3.s[0]\n" "addvl x12, x12, #1\n" - "fmla z24.s, z6.s, z4.s[0]\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z29.s, z4.s[0]\n" + "fmla z9.s, z28.s, z0.s[0]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" "addvl x11, x11, #1\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" + "fmla z13.s, z28.s, z1.s[0]\n" + "fmla z17.s, z28.s, z2.s[0]\n" "addvl x10, x10, #1\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "fmla z25.s, z7.s, z4.s[0]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z21.s, z28.s, z3.s[0]\n" + "fmla z25.s, z28.s, z4.s[0]\n" + "ld1w { z28.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z26.s, z6.s, z4.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "fmla z27.s, z7.s, z4.s[0]\n" + "fmla z10.s, z29.s, z0.s[0]\n" + "fmla z14.s, z29.s, z1.s[0]\n" + "fmla z18.s, z29.s, z2.s[0]\n" + "fmla z22.s, z29.s, z3.s[0]\n" + "fmla z26.s, z29.s, z4.s[0]\n" + "fmla z11.s, z28.s, z0.s[0]\n" + "fmla z15.s, z28.s, z1.s[0]\n" + "fmla z19.s, z28.s, z2.s[0]\n" + "fmla z23.s, z28.s, z3.s[0]\n" + "fmla z27.s, z28.s, z4.s[0]\n" "ble 68f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" + "ld1w { z29.s }, p5/Z, [x12]\n" + "ld1w { z28.s }, p5/Z, [x11]\n" + "fmla z8.s, z29.s, z0.s[1]\n" + "fmla z12.s, z29.s, z1.s[1]\n" + "fmla z16.s, z29.s, z2.s[1]\n" + "fmla z20.s, z29.s, z3.s[1]\n" "subs x27, x27, #0x1\n" "addvl x12, x12, #1\n" - "fmla z24.s, z6.s, z4.s[1]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z29.s, z4.s[1]\n" + "fmla z9.s, z28.s, z0.s[1]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" "addvl x11, x11, #1\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" + "fmla z13.s, z28.s, z1.s[1]\n" + "fmla z17.s, z28.s, z2.s[1]\n" "addvl x10, x10, #1\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "fmla z25.s, z7.s, z4.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z21.s, z28.s, z3.s[1]\n" + "fmla z25.s, z28.s, z4.s[1]\n" + "ld1w { z28.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z26.s, z6.s, z4.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "fmla z27.s, z7.s, z4.s[1]\n" + "fmla z10.s, z29.s, z0.s[1]\n" + "fmla z14.s, z29.s, z1.s[1]\n" + "fmla z18.s, z29.s, z2.s[1]\n" + "fmla z22.s, z29.s, z3.s[1]\n" + "fmla z26.s, z29.s, z4.s[1]\n" + "fmla z11.s, z28.s, z0.s[1]\n" + "fmla z15.s, z28.s, z1.s[1]\n" + "fmla z19.s, z28.s, z2.s[1]\n" + "fmla z23.s, z28.s, z3.s[1]\n" + "fmla z27.s, z28.s, z4.s[1]\n" "ble 68f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" + "ld1w { z29.s }, p5/Z, [x12]\n" + "ld1w { z28.s }, p5/Z, [x11]\n" + "fmla z8.s, z29.s, z0.s[2]\n" + "fmla z12.s, z29.s, z1.s[2]\n" + "fmla z16.s, z29.s, z2.s[2]\n" + "fmla z20.s, z29.s, z3.s[2]\n" "subs x27, x27, #0x1\n" "addvl x12, x12, #1\n" - "fmla z24.s, z6.s, z4.s[2]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z29.s, z4.s[2]\n" + "fmla z9.s, z28.s, z0.s[2]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" "addvl x11, x11, #1\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" + "fmla z13.s, z28.s, z1.s[2]\n" + "fmla z17.s, z28.s, z2.s[2]\n" "addvl x10, x10, #1\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "fmla z25.s, z7.s, z4.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z21.s, z28.s, z3.s[2]\n" + "fmla z25.s, z28.s, z4.s[2]\n" + "ld1w { z28.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z26.s, z6.s, z4.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "fmla z27.s, z7.s, z4.s[2]\n" + "fmla z10.s, z29.s, z0.s[2]\n" + "fmla z14.s, z29.s, z1.s[2]\n" + "fmla z18.s, z29.s, z2.s[2]\n" + "fmla z22.s, z29.s, z3.s[2]\n" + "fmla z26.s, z29.s, z4.s[2]\n" + "fmla z11.s, z28.s, z0.s[2]\n" + "fmla z15.s, z28.s, z1.s[2]\n" + "fmla z19.s, z28.s, z2.s[2]\n" + "fmla z23.s, z28.s, z3.s[2]\n" + "fmla z27.s, z28.s, z4.s[2]\n" "ble 68f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" + "ld1w { z29.s }, p5/Z, [x12]\n" + "ld1w { z28.s }, p5/Z, [x11]\n" + "fmla z8.s, z29.s, z0.s[3]\n" + "fmla z12.s, z29.s, z1.s[3]\n" + "fmla z16.s, z29.s, z2.s[3]\n" + "fmla z20.s, z29.s, z3.s[3]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "fmla z24.s, z6.s, z4.s[3]\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z29.s, z4.s[3]\n" + "fmla z9.s, z28.s, z0.s[3]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "fmla z25.s, z7.s, z4.s[3]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z13.s, z28.s, z1.s[3]\n" + "fmla z17.s, z28.s, z2.s[3]\n" + "fmla z21.s, z28.s, z3.s[3]\n" + "fmla z25.s, z28.s, z4.s[3]\n" + "ld1w { z28.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z26.s, z6.s, z4.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" - "fmla z27.s, z7.s, z4.s[3]\n" + "fmla z10.s, z29.s, z0.s[3]\n" + "fmla z14.s, z29.s, z1.s[3]\n" + "fmla z18.s, z29.s, z2.s[3]\n" + "fmla z22.s, z29.s, z3.s[3]\n" + "fmla z26.s, z29.s, z4.s[3]\n" + "fmla z11.s, z28.s, z0.s[3]\n" + "fmla z15.s, z28.s, z1.s[3]\n" + "fmla z19.s, z28.s, z2.s[3]\n" + "fmla z23.s, z28.s, z3.s[3]\n" + "fmla z27.s, z28.s, z4.s[3]\n" "68:" // Height 5: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1692,49 +1692,49 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "add x22, x23, x20, LSL #2\n" "tbz %x[flags], #1, 69f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z29.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmin z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z1.s\n" - "fmin z18.s, p5/M, z18.s, z1.s\n" - "fmin z19.s, p5/M, z19.s, z1.s\n" - "fmin z20.s, p5/M, z20.s, z1.s\n" - "fmin z21.s, p5/M, z21.s, z1.s\n" - "fmin z22.s, p5/M, z22.s, z1.s\n" - "fmin z23.s, p5/M, z23.s, z1.s\n" - "fmin z24.s, p5/M, z24.s, z1.s\n" - "fmin z25.s, p5/M, z25.s, z1.s\n" - "fmin z26.s, p5/M, z26.s, z1.s\n" - "fmin z27.s, p5/M, z27.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" - "fmax z16.s, p5/M, z16.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z0.s\n" - "fmax z18.s, p5/M, z18.s, z0.s\n" - "fmax z19.s, p5/M, z19.s, z0.s\n" - "fmax z20.s, p5/M, z20.s, z0.s\n" - "fmax z21.s, p5/M, z21.s, z0.s\n" - "fmax z22.s, p5/M, z22.s, z0.s\n" - "fmax z23.s, p5/M, z23.s, z0.s\n" - "fmax z24.s, p5/M, z24.s, z0.s\n" - "fmax z25.s, p5/M, z25.s, z0.s\n" - "fmax z26.s, p5/M, z26.s, z0.s\n" - "fmax z27.s, p5/M, z27.s, z0.s\n" + "ld1rw { z28.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z29.s\n" + "fmin z9.s, p5/M, z9.s, z29.s\n" + "fmin z10.s, p5/M, z10.s, z29.s\n" + "fmin z11.s, p5/M, z11.s, z29.s\n" + "fmin z12.s, p5/M, z12.s, z29.s\n" + "fmin z13.s, p5/M, z13.s, z29.s\n" + "fmin z14.s, p5/M, z14.s, z29.s\n" + "fmin z15.s, p5/M, z15.s, z29.s\n" + "fmin z16.s, p5/M, z16.s, z29.s\n" + "fmin z17.s, p5/M, z17.s, z29.s\n" + "fmin z18.s, p5/M, z18.s, z29.s\n" + "fmin z19.s, p5/M, z19.s, z29.s\n" + "fmin z20.s, p5/M, z20.s, z29.s\n" + "fmin z21.s, p5/M, z21.s, z29.s\n" + "fmin z22.s, p5/M, z22.s, z29.s\n" + "fmin z23.s, p5/M, z23.s, z29.s\n" + "fmin z24.s, p5/M, z24.s, z29.s\n" + "fmin z25.s, p5/M, z25.s, z29.s\n" + "fmin z26.s, p5/M, z26.s, z29.s\n" + "fmin z27.s, p5/M, z27.s, z29.s\n" + "fmax z8.s, p5/M, z8.s, z28.s\n" + "fmax z9.s, p5/M, z9.s, z28.s\n" + "fmax z10.s, p5/M, z10.s, z28.s\n" + "fmax z11.s, p5/M, z11.s, z28.s\n" + "fmax z12.s, p5/M, z12.s, z28.s\n" + "fmax z13.s, p5/M, z13.s, z28.s\n" + "fmax z14.s, p5/M, z14.s, z28.s\n" + "fmax z15.s, p5/M, z15.s, z28.s\n" + "fmax z16.s, p5/M, z16.s, z28.s\n" + "fmax z17.s, p5/M, z17.s, z28.s\n" + "fmax z18.s, p5/M, z18.s, z28.s\n" + "fmax z19.s, p5/M, z19.s, z28.s\n" + "fmax z20.s, p5/M, z20.s, z28.s\n" + "fmax z21.s, p5/M, z21.s, z28.s\n" + "fmax z22.s, p5/M, z22.s, z28.s\n" + "fmax z23.s, p5/M, z23.s, z28.s\n" + "fmax z24.s, p5/M, z24.s, z28.s\n" + "fmax z25.s, p5/M, z25.s, z28.s\n" + "fmax z26.s, p5/M, z26.s, z28.s\n" + "fmax z27.s, p5/M, z27.s, z28.s\n" "69:" // Height 5: No activation "st1w { z8.s }, p4, [x13]\n" "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" @@ -1830,35 +1830,35 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "74:" // Height 6: no bias "tbz %x[flags], #0, 75f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z8.s }, p4/Z, [x13]\n" + "add x24, x13, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" + "ld1w { z8.s }, p4/Z, [x13]\n" "add x22, x23, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" - "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p4/Z, [x24]\n" - "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z24.s }, p4/Z, [x22]\n" - "ld1w { z25.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z26.s }, p2/Z, [x22, #2, MUL VL]\n" - "ld1w { z27.s }, p1/Z, [x22, #3, MUL VL]\n" - "ld1w { z28.s }, p4/Z, [x21]\n" - "ld1w { z29.s }, p3/Z, [x21, #1, MUL VL]\n" - "ld1w { z30.s }, p2/Z, [x21, #2, MUL VL]\n" - "ld1w { z31.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x24]\n" + "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x23]\n" + "ld1w { z17.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x22]\n" + "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z24.s }, p4/Z, [x21]\n" + "ld1w { z25.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z26.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z27.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z28.s }, p4/Z, [x20]\n" + "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n" "b 76f\n" "75:" // Height 6: no accumulate "mov z8.b, #0x0\n" @@ -1890,16 +1890,16 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "77:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 78f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "ldr x21, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "ldr x21, [x20, #0x28]\n" "cbnz x28, 79f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1911,146 +1911,146 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 79f\n" "78:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "79:" // Height 6: input setup done "cmp x27, #0x4\n" "ble 81f\n" "80:" // Height 6: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" + "ld1rqw { z7.s }, p0/Z, [x26]\n" + "ld1rqw { z6.s }, p0/Z, [x25]\n" "sub x27, x27, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" + "ld1rqw { z5.s }, p0/Z, [x24]\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" "cmp x27, #0x4\n" "add x26, x26, #0x10\n" - "ld1rqw { z4.s }, p0/Z, [x22]\n" - "ld1rqw { z5.s }, p0/Z, [x21]\n" + "ld1rqw { z3.s }, p0/Z, [x22]\n" + "ld1rqw { z2.s }, p0/Z, [x21]\n" "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" + "ld1w { z1.s }, p5/Z, [x12]\n" + "ld1w { z0.s }, p5/Z, [x11]\n" + "fmla z8.s, z1.s, z7.s[0]\n" + "fmla z12.s, z1.s, z6.s[0]\n" + "fmla z16.s, z1.s, z5.s[0]\n" + "fmla z20.s, z1.s, z4.s[0]\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - "fmla z24.s, z6.s, z4.s[0]\n" - "fmla z28.s, z6.s, z5.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z1.s, z3.s[0]\n" + "fmla z28.s, z1.s, z2.s[0]\n" + "ld1w { z1.s }, p5/Z, [x10]\n" "add x21, x21, #0x10\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "fmla z25.s, z7.s, z4.s[0]\n" - "fmla z29.s, z7.s, z5.s[0]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z26.s, z6.s, z4.s[0]\n" - "fmla z30.s, z6.s, z5.s[0]\n" - "ld1w { z6.s }, p5/Z, [x12, #1, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "fmla z27.s, z7.s, z4.s[0]\n" - "fmla z31.s, z7.s, z5.s[0]\n" - "ld1w { z7.s }, p5/Z, [x11, #1, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" - "fmla z24.s, z6.s, z4.s[1]\n" - "fmla z28.s, z6.s, z5.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "fmla z25.s, z7.s, z4.s[1]\n" - "fmla z29.s, z7.s, z5.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9, #1, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z26.s, z6.s, z4.s[1]\n" - "fmla z30.s, z6.s, z5.s[1]\n" - "ld1w { z6.s }, p5/Z, [x12, #2, MUL VL]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "fmla z27.s, z7.s, z4.s[1]\n" - "fmla z31.s, z7.s, z5.s[1]\n" - "ld1w { z7.s }, p5/Z, [x11, #2, MUL VL]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" - "fmla z24.s, z6.s, z4.s[2]\n" - "fmla z28.s, z6.s, z5.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "fmla z25.s, z7.s, z4.s[2]\n" - "fmla z29.s, z7.s, z5.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9, #2, MUL VL]\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z26.s, z6.s, z4.s[2]\n" - "fmla z30.s, z6.s, z5.s[2]\n" - "ld1w { z6.s }, p5/Z, [x12, #3, MUL VL]\n" + "fmla z9.s, z0.s, z7.s[0]\n" + "fmla z13.s, z0.s, z6.s[0]\n" + "fmla z17.s, z0.s, z5.s[0]\n" + "fmla z21.s, z0.s, z4.s[0]\n" + "fmla z25.s, z0.s, z3.s[0]\n" + "fmla z29.s, z0.s, z2.s[0]\n" + "ld1w { z0.s }, p5/Z, [x9]\n" + "fmla z10.s, z1.s, z7.s[0]\n" + "fmla z14.s, z1.s, z6.s[0]\n" + "fmla z18.s, z1.s, z5.s[0]\n" + "fmla z22.s, z1.s, z4.s[0]\n" + "fmla z26.s, z1.s, z3.s[0]\n" + "fmla z30.s, z1.s, z2.s[0]\n" + "ld1w { z1.s }, p5/Z, [x12, #1, MUL VL]\n" + "fmla z11.s, z0.s, z7.s[0]\n" + "fmla z15.s, z0.s, z6.s[0]\n" + "fmla z19.s, z0.s, z5.s[0]\n" + "fmla z23.s, z0.s, z4.s[0]\n" + "fmla z27.s, z0.s, z3.s[0]\n" + "fmla z31.s, z0.s, z2.s[0]\n" + "ld1w { z0.s }, p5/Z, [x11, #1, MUL VL]\n" + "fmla z8.s, z1.s, z7.s[1]\n" + "fmla z12.s, z1.s, z6.s[1]\n" + "fmla z16.s, z1.s, z5.s[1]\n" + "fmla z20.s, z1.s, z4.s[1]\n" + "fmla z24.s, z1.s, z3.s[1]\n" + "fmla z28.s, z1.s, z2.s[1]\n" + "ld1w { z1.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z9.s, z0.s, z7.s[1]\n" + "fmla z13.s, z0.s, z6.s[1]\n" + "fmla z17.s, z0.s, z5.s[1]\n" + "fmla z21.s, z0.s, z4.s[1]\n" + "fmla z25.s, z0.s, z3.s[1]\n" + "fmla z29.s, z0.s, z2.s[1]\n" + "ld1w { z0.s }, p5/Z, [x9, #1, MUL VL]\n" + "fmla z10.s, z1.s, z7.s[1]\n" + "fmla z14.s, z1.s, z6.s[1]\n" + "fmla z18.s, z1.s, z5.s[1]\n" + "fmla z22.s, z1.s, z4.s[1]\n" + "fmla z26.s, z1.s, z3.s[1]\n" + "fmla z30.s, z1.s, z2.s[1]\n" + "ld1w { z1.s }, p5/Z, [x12, #2, MUL VL]\n" + "fmla z11.s, z0.s, z7.s[1]\n" + "fmla z15.s, z0.s, z6.s[1]\n" + "fmla z19.s, z0.s, z5.s[1]\n" + "fmla z23.s, z0.s, z4.s[1]\n" + "fmla z27.s, z0.s, z3.s[1]\n" + "fmla z31.s, z0.s, z2.s[1]\n" + "ld1w { z0.s }, p5/Z, [x11, #2, MUL VL]\n" + "fmla z8.s, z1.s, z7.s[2]\n" + "fmla z12.s, z1.s, z6.s[2]\n" + "fmla z16.s, z1.s, z5.s[2]\n" + "fmla z20.s, z1.s, z4.s[2]\n" + "fmla z24.s, z1.s, z3.s[2]\n" + "fmla z28.s, z1.s, z2.s[2]\n" + "ld1w { z1.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z0.s, z7.s[2]\n" + "fmla z13.s, z0.s, z6.s[2]\n" + "fmla z17.s, z0.s, z5.s[2]\n" + "fmla z21.s, z0.s, z4.s[2]\n" + "fmla z25.s, z0.s, z3.s[2]\n" + "fmla z29.s, z0.s, z2.s[2]\n" + "ld1w { z0.s }, p5/Z, [x9, #2, MUL VL]\n" + "fmla z10.s, z1.s, z7.s[2]\n" + "fmla z14.s, z1.s, z6.s[2]\n" + "fmla z18.s, z1.s, z5.s[2]\n" + "fmla z22.s, z1.s, z4.s[2]\n" + "fmla z26.s, z1.s, z3.s[2]\n" + "fmla z30.s, z1.s, z2.s[2]\n" + "ld1w { z1.s }, p5/Z, [x12, #3, MUL VL]\n" "addvl x12, x12, #4\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "fmla z27.s, z7.s, z4.s[2]\n" - "fmla z31.s, z7.s, z5.s[2]\n" - "ld1w { z7.s }, p5/Z, [x11, #3, MUL VL]\n" + "fmla z11.s, z0.s, z7.s[2]\n" + "fmla z15.s, z0.s, z6.s[2]\n" + "fmla z19.s, z0.s, z5.s[2]\n" + "fmla z23.s, z0.s, z4.s[2]\n" + "fmla z27.s, z0.s, z3.s[2]\n" + "fmla z31.s, z0.s, z2.s[2]\n" + "ld1w { z0.s }, p5/Z, [x11, #3, MUL VL]\n" "addvl x11, x11, #4\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" - "fmla z24.s, z6.s, z4.s[3]\n" - "fmla z28.s, z6.s, z5.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" + "fmla z8.s, z1.s, z7.s[3]\n" + "fmla z12.s, z1.s, z6.s[3]\n" + "fmla z16.s, z1.s, z5.s[3]\n" + "fmla z20.s, z1.s, z4.s[3]\n" + "fmla z24.s, z1.s, z3.s[3]\n" + "fmla z28.s, z1.s, z2.s[3]\n" + "ld1w { z1.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "fmla z25.s, z7.s, z4.s[3]\n" - "fmla z29.s, z7.s, z5.s[3]\n" - "ld1w { z7.s }, p5/Z, [x9, #3, MUL VL]\n" + "fmla z9.s, z0.s, z7.s[3]\n" + "fmla z13.s, z0.s, z6.s[3]\n" + "fmla z17.s, z0.s, z5.s[3]\n" + "fmla z21.s, z0.s, z4.s[3]\n" + "fmla z25.s, z0.s, z3.s[3]\n" + "fmla z29.s, z0.s, z2.s[3]\n" + "ld1w { z0.s }, p5/Z, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z26.s, z6.s, z4.s[3]\n" - "fmla z30.s, z6.s, z5.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" - "fmla z27.s, z7.s, z4.s[3]\n" - "fmla z31.s, z7.s, z5.s[3]\n" + "fmla z10.s, z1.s, z7.s[3]\n" + "fmla z14.s, z1.s, z6.s[3]\n" + "fmla z18.s, z1.s, z5.s[3]\n" + "fmla z22.s, z1.s, z4.s[3]\n" + "fmla z26.s, z1.s, z3.s[3]\n" + "fmla z30.s, z1.s, z2.s[3]\n" + "fmla z11.s, z0.s, z7.s[3]\n" + "fmla z15.s, z0.s, z6.s[3]\n" + "fmla z19.s, z0.s, z5.s[3]\n" + "fmla z23.s, z0.s, z4.s[3]\n" + "fmla z27.s, z0.s, z3.s[3]\n" + "fmla z31.s, z0.s, z2.s[3]\n" "bgt 80b\n" "81:" // Height 6: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" @@ -2061,139 +2061,139 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ld1rqw { z3.s }, p0/Z, [x23]\n" "ld1rqw { z4.s }, p0/Z, [x22]\n" "ld1rqw { z5.s }, p0/Z, [x21]\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[0]\n" - "fmla z12.s, z6.s, z1.s[0]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z20.s, z6.s, z3.s[0]\n" + "ld1w { z7.s }, p5/Z, [x12]\n" + "ld1w { z6.s }, p5/Z, [x11]\n" + "fmla z8.s, z7.s, z0.s[0]\n" + "fmla z12.s, z7.s, z1.s[0]\n" + "fmla z16.s, z7.s, z2.s[0]\n" + "fmla z20.s, z7.s, z3.s[0]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "fmla z24.s, z6.s, z4.s[0]\n" - "fmla z28.s, z6.s, z5.s[0]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z7.s, z4.s[0]\n" + "fmla z28.s, z7.s, z5.s[0]\n" + "ld1w { z7.s }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" - "fmla z9.s, z7.s, z0.s[0]\n" - "fmla z13.s, z7.s, z1.s[0]\n" - "fmla z17.s, z7.s, z2.s[0]\n" - "fmla z21.s, z7.s, z3.s[0]\n" - "fmla z25.s, z7.s, z4.s[0]\n" - "fmla z29.s, z7.s, z5.s[0]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z9.s, z6.s, z0.s[0]\n" + "fmla z13.s, z6.s, z1.s[0]\n" + "fmla z17.s, z6.s, z2.s[0]\n" + "fmla z21.s, z6.s, z3.s[0]\n" + "fmla z25.s, z6.s, z4.s[0]\n" + "fmla z29.s, z6.s, z5.s[0]\n" + "ld1w { z6.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[0]\n" - "fmla z14.s, z6.s, z1.s[0]\n" - "fmla z18.s, z6.s, z2.s[0]\n" - "fmla z22.s, z6.s, z3.s[0]\n" - "fmla z26.s, z6.s, z4.s[0]\n" - "fmla z30.s, z6.s, z5.s[0]\n" - "fmla z11.s, z7.s, z0.s[0]\n" - "fmla z15.s, z7.s, z1.s[0]\n" - "fmla z19.s, z7.s, z2.s[0]\n" - "fmla z23.s, z7.s, z3.s[0]\n" - "fmla z27.s, z7.s, z4.s[0]\n" - "fmla z31.s, z7.s, z5.s[0]\n" + "fmla z10.s, z7.s, z0.s[0]\n" + "fmla z14.s, z7.s, z1.s[0]\n" + "fmla z18.s, z7.s, z2.s[0]\n" + "fmla z22.s, z7.s, z3.s[0]\n" + "fmla z26.s, z7.s, z4.s[0]\n" + "fmla z30.s, z7.s, z5.s[0]\n" + "fmla z11.s, z6.s, z0.s[0]\n" + "fmla z15.s, z6.s, z1.s[0]\n" + "fmla z19.s, z6.s, z2.s[0]\n" + "fmla z23.s, z6.s, z3.s[0]\n" + "fmla z27.s, z6.s, z4.s[0]\n" + "fmla z31.s, z6.s, z5.s[0]\n" "ble 82f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[1]\n" - "fmla z12.s, z6.s, z1.s[1]\n" - "fmla z16.s, z6.s, z2.s[1]\n" - "fmla z20.s, z6.s, z3.s[1]\n" + "ld1w { z7.s }, p5/Z, [x12]\n" + "ld1w { z6.s }, p5/Z, [x11]\n" + "fmla z8.s, z7.s, z0.s[1]\n" + "fmla z12.s, z7.s, z1.s[1]\n" + "fmla z16.s, z7.s, z2.s[1]\n" + "fmla z20.s, z7.s, z3.s[1]\n" "subs x27, x27, #0x1\n" "addvl x12, x12, #1\n" - "fmla z24.s, z6.s, z4.s[1]\n" - "fmla z28.s, z6.s, z5.s[1]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z7.s, z4.s[1]\n" + "fmla z28.s, z7.s, z5.s[1]\n" + "ld1w { z7.s }, p5/Z, [x10]\n" "addvl x11, x11, #1\n" - "fmla z9.s, z7.s, z0.s[1]\n" - "fmla z13.s, z7.s, z1.s[1]\n" + "fmla z9.s, z6.s, z0.s[1]\n" + "fmla z13.s, z6.s, z1.s[1]\n" "addvl x10, x10, #1\n" - "fmla z17.s, z7.s, z2.s[1]\n" - "fmla z21.s, z7.s, z3.s[1]\n" - "fmla z25.s, z7.s, z4.s[1]\n" - "fmla z29.s, z7.s, z5.s[1]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z17.s, z6.s, z2.s[1]\n" + "fmla z21.s, z6.s, z3.s[1]\n" + "fmla z25.s, z6.s, z4.s[1]\n" + "fmla z29.s, z6.s, z5.s[1]\n" + "ld1w { z6.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[1]\n" - "fmla z14.s, z6.s, z1.s[1]\n" - "fmla z18.s, z6.s, z2.s[1]\n" - "fmla z22.s, z6.s, z3.s[1]\n" - "fmla z26.s, z6.s, z4.s[1]\n" - "fmla z30.s, z6.s, z5.s[1]\n" - "fmla z11.s, z7.s, z0.s[1]\n" - "fmla z15.s, z7.s, z1.s[1]\n" - "fmla z19.s, z7.s, z2.s[1]\n" - "fmla z23.s, z7.s, z3.s[1]\n" - "fmla z27.s, z7.s, z4.s[1]\n" - "fmla z31.s, z7.s, z5.s[1]\n" + "fmla z10.s, z7.s, z0.s[1]\n" + "fmla z14.s, z7.s, z1.s[1]\n" + "fmla z18.s, z7.s, z2.s[1]\n" + "fmla z22.s, z7.s, z3.s[1]\n" + "fmla z26.s, z7.s, z4.s[1]\n" + "fmla z30.s, z7.s, z5.s[1]\n" + "fmla z11.s, z6.s, z0.s[1]\n" + "fmla z15.s, z6.s, z1.s[1]\n" + "fmla z19.s, z6.s, z2.s[1]\n" + "fmla z23.s, z6.s, z3.s[1]\n" + "fmla z27.s, z6.s, z4.s[1]\n" + "fmla z31.s, z6.s, z5.s[1]\n" "ble 82f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[2]\n" - "fmla z12.s, z6.s, z1.s[2]\n" - "fmla z16.s, z6.s, z2.s[2]\n" - "fmla z20.s, z6.s, z3.s[2]\n" + "ld1w { z7.s }, p5/Z, [x12]\n" + "ld1w { z6.s }, p5/Z, [x11]\n" + "fmla z8.s, z7.s, z0.s[2]\n" + "fmla z12.s, z7.s, z1.s[2]\n" + "fmla z16.s, z7.s, z2.s[2]\n" + "fmla z20.s, z7.s, z3.s[2]\n" "subs x27, x27, #0x1\n" "addvl x12, x12, #1\n" - "fmla z24.s, z6.s, z4.s[2]\n" - "fmla z28.s, z6.s, z5.s[2]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z7.s, z4.s[2]\n" + "fmla z28.s, z7.s, z5.s[2]\n" + "ld1w { z7.s }, p5/Z, [x10]\n" "addvl x11, x11, #1\n" - "fmla z9.s, z7.s, z0.s[2]\n" - "fmla z13.s, z7.s, z1.s[2]\n" + "fmla z9.s, z6.s, z0.s[2]\n" + "fmla z13.s, z6.s, z1.s[2]\n" "addvl x10, x10, #1\n" - "fmla z17.s, z7.s, z2.s[2]\n" - "fmla z21.s, z7.s, z3.s[2]\n" - "fmla z25.s, z7.s, z4.s[2]\n" - "fmla z29.s, z7.s, z5.s[2]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z17.s, z6.s, z2.s[2]\n" + "fmla z21.s, z6.s, z3.s[2]\n" + "fmla z25.s, z6.s, z4.s[2]\n" + "fmla z29.s, z6.s, z5.s[2]\n" + "ld1w { z6.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[2]\n" - "fmla z14.s, z6.s, z1.s[2]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z22.s, z6.s, z3.s[2]\n" - "fmla z26.s, z6.s, z4.s[2]\n" - "fmla z30.s, z6.s, z5.s[2]\n" - "fmla z11.s, z7.s, z0.s[2]\n" - "fmla z15.s, z7.s, z1.s[2]\n" - "fmla z19.s, z7.s, z2.s[2]\n" - "fmla z23.s, z7.s, z3.s[2]\n" - "fmla z27.s, z7.s, z4.s[2]\n" - "fmla z31.s, z7.s, z5.s[2]\n" + "fmla z10.s, z7.s, z0.s[2]\n" + "fmla z14.s, z7.s, z1.s[2]\n" + "fmla z18.s, z7.s, z2.s[2]\n" + "fmla z22.s, z7.s, z3.s[2]\n" + "fmla z26.s, z7.s, z4.s[2]\n" + "fmla z30.s, z7.s, z5.s[2]\n" + "fmla z11.s, z6.s, z0.s[2]\n" + "fmla z15.s, z6.s, z1.s[2]\n" + "fmla z19.s, z6.s, z2.s[2]\n" + "fmla z23.s, z6.s, z3.s[2]\n" + "fmla z27.s, z6.s, z4.s[2]\n" + "fmla z31.s, z6.s, z5.s[2]\n" "ble 82f\n" - "ld1w { z6.s }, p5/Z, [x12]\n" - "ld1w { z7.s }, p5/Z, [x11]\n" - "fmla z8.s, z6.s, z0.s[3]\n" - "fmla z12.s, z6.s, z1.s[3]\n" - "fmla z16.s, z6.s, z2.s[3]\n" - "fmla z20.s, z6.s, z3.s[3]\n" + "ld1w { z7.s }, p5/Z, [x12]\n" + "ld1w { z6.s }, p5/Z, [x11]\n" + "fmla z8.s, z7.s, z0.s[3]\n" + "fmla z12.s, z7.s, z1.s[3]\n" + "fmla z16.s, z7.s, z2.s[3]\n" + "fmla z20.s, z7.s, z3.s[3]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "fmla z24.s, z6.s, z4.s[3]\n" - "fmla z28.s, z6.s, z5.s[3]\n" - "ld1w { z6.s }, p5/Z, [x10]\n" + "fmla z24.s, z7.s, z4.s[3]\n" + "fmla z28.s, z7.s, z5.s[3]\n" + "ld1w { z7.s }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" - "fmla z9.s, z7.s, z0.s[3]\n" - "fmla z13.s, z7.s, z1.s[3]\n" - "fmla z17.s, z7.s, z2.s[3]\n" - "fmla z21.s, z7.s, z3.s[3]\n" - "fmla z25.s, z7.s, z4.s[3]\n" - "fmla z29.s, z7.s, z5.s[3]\n" - "ld1w { z7.s }, p5/Z, [x9]\n" + "fmla z9.s, z6.s, z0.s[3]\n" + "fmla z13.s, z6.s, z1.s[3]\n" + "fmla z17.s, z6.s, z2.s[3]\n" + "fmla z21.s, z6.s, z3.s[3]\n" + "fmla z25.s, z6.s, z4.s[3]\n" + "fmla z29.s, z6.s, z5.s[3]\n" + "ld1w { z6.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" - "fmla z10.s, z6.s, z0.s[3]\n" - "fmla z14.s, z6.s, z1.s[3]\n" - "fmla z18.s, z6.s, z2.s[3]\n" - "fmla z22.s, z6.s, z3.s[3]\n" - "fmla z26.s, z6.s, z4.s[3]\n" - "fmla z30.s, z6.s, z5.s[3]\n" - "fmla z11.s, z7.s, z0.s[3]\n" - "fmla z15.s, z7.s, z1.s[3]\n" - "fmla z19.s, z7.s, z2.s[3]\n" - "fmla z23.s, z7.s, z3.s[3]\n" - "fmla z27.s, z7.s, z4.s[3]\n" - "fmla z31.s, z7.s, z5.s[3]\n" + "fmla z10.s, z7.s, z0.s[3]\n" + "fmla z14.s, z7.s, z1.s[3]\n" + "fmla z18.s, z7.s, z2.s[3]\n" + "fmla z22.s, z7.s, z3.s[3]\n" + "fmla z26.s, z7.s, z4.s[3]\n" + "fmla z30.s, z7.s, z5.s[3]\n" + "fmla z11.s, z6.s, z0.s[3]\n" + "fmla z15.s, z6.s, z1.s[3]\n" + "fmla z19.s, z6.s, z2.s[3]\n" + "fmla z23.s, z6.s, z3.s[3]\n" + "fmla z27.s, z6.s, z4.s[3]\n" + "fmla z31.s, z6.s, z5.s[3]\n" "82:" // Height 6: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -2307,4 +2307,4 @@ void sve_ffhybrid_fp32_mla_6x4VL ( } } // namespace arm_gemm -#endif // ARM_COMPUTE_ENABLE_SVE +#endif // ARM_COMPUTE_ENABLE_SVE |