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 | 1974 |
1 files changed, 987 insertions, 987 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 3c7e562c89..eb057e7734 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 @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022 Arm Limited. + * Copyright (c) 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -10,16 +10,16 @@ * sell copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * - * The above copyright notice and this permission notice shall be included in - * all copies or substantial portions of the Software. + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING - * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS - * IN THE SOFTWARE. + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. */ #ifdef ARM_COMPUTE_ENABLE_SVE @@ -105,53 +105,53 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cmp %x[M], #0x2\n" "bgt 29f\n" "beq 15f\n" - "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[bias]\n" - "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x12, %x[output_ptr]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x15, %x[bias]\n" + "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "mov x13, %x[output_ptr]\n" "2:" // Height 1: Column loop - "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x10, x11, x19, LSL #2\n" - "cntw x20, ALL, MUL #3\n" - "add x9, x10, x19, LSL #2\n" - "add x28, x9, x19, LSL #2\n" - "add x19, x28, x19, LSL #2\n" - "cmp x13, x20\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" + "add x11, x12, x20, LSL #2\n" + "cntw x21, ALL, MUL #3\n" + "add x10, x11, x20, LSL #2\n" + "add x9, x10, x20, LSL #2\n" + "add x20, x9, x20, LSL #2\n" + "cmp x14, x21\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 3f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x28, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x9, x12\n" "bgt 3f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x9, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x10, x12\n" "bgt 3f\n" - "mov x10, x11\n" + "mov x11, x12\n" "3:" // Height 1: B setup done - "mov x19, #0x0\n" - "whilelt p4.s, x19, x13\n" - "incw x19\n" - "whilelt p3.s, x19, x13\n" - "incw x19\n" - "whilelt p2.s, x19, x13\n" - "incw x19\n" - "whilelt p1.s, x19, x13\n" - "cbz x14, 4f\n" - "ld1w { z8.s }, p5/Z, [x14]\n" - "ld1w { z9.s }, p5/Z, [x14, #1, MUL VL]\n" - "ld1w { z10.s }, p5/Z, [x14, #2, MUL VL]\n" - "ld1w { z11.s }, p5/Z, [x14, #3, MUL VL]\n" - "addvl x14, x14, #4\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x14\n" + "incw x20\n" + "whilelt p3.s, x20, x14\n" + "incw x20\n" + "whilelt p2.s, x20, x14\n" + "incw x20\n" + "whilelt p1.s, x20, x14\n" + "cbz x15, 4f\n" + "ld1w { z8.s }, p5/Z, [x15]\n" + "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" + "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" + "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n" + "addvl x15, x15, #4\n" "b 6f\n" "4:" // Height 1: no bias "tbz %x[flags], #0, 5f\n" - "ld1w { z8.s }, p4/Z, [x12]\n" - "ld1w { z9.s }, p3/Z, [x12, #1, MUL VL]\n" - "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" - "ld1w { z11.s }, p1/Z, [x12, #3, MUL VL]\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" "b 6f\n" "5:" // Height 1: no accumulate "mov z8.b, #0x0\n" @@ -159,134 +159,134 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov z10.b, #0x0\n" "mov z11.b, #0x0\n" "6:" // Height 1: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "7:" // Height 1: String loop - "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w26, [x19, x27, LSL #0x2]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "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" "tbz %x[flags], #3, 8f\n" - "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" - "add x20, x20, x19, LSL #3\n" - "ldr x25, [x20, #0x0]\n" - "cbnz x27, 9f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #2\n" + "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" + "add x21, x21, x20, LSL #3\n" + "ldr x26, [x21, #0x0]\n" + "cbnz x28, 9f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #2\n" "b 9f\n" "8:" // Height 1: setup direct input - "mov x25, %x[input_ptr]\n" + "mov x26, %x[input_ptr]\n" "9:" // Height 1: input setup done - "cmp x26, #0x4\n" + "cmp x27, #0x4\n" "ble 11f\n" "10:" // Height 1: Multiply loop: Main loop head - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1w { z6.s }, p5/Z, [x11]\n" + "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, [x10]\n" + "ld1w { z7.s }, p5/Z, [x11]\n" "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x9]\n" + "ld1w { z6.s }, p5/Z, [x10]\n" "fmla z10.s, z6.s, z0.s[0]\n" - "ld1w { z7.s }, p5/Z, [x28]\n" + "ld1w { z7.s }, p5/Z, [x9]\n" "fmla z11.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x11, #1, MUL VL]\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, [x10, #1, MUL VL]\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, [x9, #1, MUL VL]\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, [x28, #1, MUL VL]\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, [x11, #2, MUL VL]\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, [x10, #2, MUL VL]\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, [x9, #2, MUL VL]\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, [x28, #2, MUL VL]\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, [x11, #3, MUL VL]\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, [x10, #3, MUL VL]\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, [x9, #3, MUL VL]\n" - "sub x26, x26, #0x4\n" - "ld1w { z7.s }, p5/Z, [x28, #3, MUL VL]\n" - "cmp x26, #0x4\n" + "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n" + "sub x27, x27, #0x4\n" + "ld1w { z7.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" - "add x25, x25, #0x10\n" + "add x26, x26, #0x10\n" + "addvl x12, x12, #4\n" "addvl x11, x11, #4\n" "addvl x10, x10, #4\n" "addvl x9, x9, #4\n" - "addvl x28, x28, #4\n" "bgt 10b\n" "11:" // Height 1: Multiply loop: Single iteration only - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1w { z6.s }, p5/Z, [x11]\n" + "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, [x10]\n" + "ld1w { z7.s }, p5/Z, [x11]\n" "fmla z9.s, z7.s, z0.s[0]\n" - "ld1w { z6.s }, p5/Z, [x9]\n" - "subs x26, x26, #0x1\n" - "ld1w { z7.s }, p5/Z, [x28]\n" + "ld1w { z6.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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "addvl x28, x28, #1\n" "ble 12f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "ld1w { z7.s }, p5/Z, [x28]\n" - "subs x26, x26, #0x1\n" + "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z7.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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "addvl x28, x28, #1\n" "ble 12f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "ld1w { z7.s }, p5/Z, [x28]\n" - "subs x26, x26, #0x1\n" + "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z7.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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "addvl x28, x28, #1\n" "ble 12f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "ld1w { z7.s }, p5/Z, [x28]\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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "addvl x28, x28, #1\n" "12:" // Height 1: Multiply loop: multiply skip - "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x27, x27, #0x1\n" - "cmp x27, x19\n" + "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 x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" "fmin 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" @@ -296,74 +296,74 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmax z10.s, p5/M, z10.s, z0.s\n" "fmax z11.s, p5/M, z11.s, z0.s\n" "13:" // Height 1: No activation - "st1w { z8.s }, p4, [x12]\n" - "st1w { z9.s }, p3, [x12, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x12, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x12, #3, MUL VL]\n" - "addvl x12, x12, #4\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" "14:" // Height 1: Writeback done - "decw x13, ALL, MUL #4\n" - "cmp x13, XZR\n" + "decw x14, ALL, MUL #4\n" + "cmp x14, XZR\n" "bgt 2b\n" "b 86f\n" "15:" // Height 2 - "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[bias]\n" - "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x12, %x[output_ptr]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x15, %x[bias]\n" + "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "mov x13, %x[output_ptr]\n" "16:" // Height 2: Column loop - "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x10, x11, x19, LSL #2\n" - "cntw x20, ALL, MUL #3\n" - "add x9, x10, x19, LSL #2\n" - "add x28, x9, x19, LSL #2\n" - "add x19, x28, x19, LSL #2\n" - "cmp x13, x20\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" + "add x11, x12, x20, LSL #2\n" + "cntw x21, ALL, MUL #3\n" + "add x10, x11, x20, LSL #2\n" + "add x9, x10, x20, LSL #2\n" + "add x20, x9, x20, LSL #2\n" + "cmp x14, x21\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 17f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x28, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x9, x12\n" "bgt 17f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x9, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x10, x12\n" "bgt 17f\n" - "mov x10, x11\n" + "mov x11, x12\n" "17:" // Height 2: B setup done - "mov x19, #0x0\n" - "whilelt p4.s, x19, x13\n" - "incw x19\n" - "whilelt p3.s, x19, x13\n" - "incw x19\n" - "whilelt p2.s, x19, x13\n" - "incw x19\n" - "whilelt p1.s, x19, x13\n" - "cbz x14, 18f\n" - "ld1w { z8.s }, p5/Z, [x14]\n" - "ld1w { z9.s }, p5/Z, [x14, #1, MUL VL]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x14\n" + "incw x20\n" + "whilelt p3.s, x20, x14\n" + "incw x20\n" + "whilelt p2.s, x20, x14\n" + "incw x20\n" + "whilelt p1.s, x20, x14\n" + "cbz x15, 18f\n" + "ld1w { z8.s }, p5/Z, [x15]\n" + "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" - "ld1w { z10.s }, p5/Z, [x14, #2, MUL VL]\n" - "ld1w { z11.s }, p5/Z, [x14, #3, MUL VL]\n" + "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" + "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n" "mov z14.d, z10.d\n" "mov z15.d, z11.d\n" - "addvl x14, x14, #4\n" + "addvl x15, x15, #4\n" "b 20f\n" "18:" // Height 2: no bias "tbz %x[flags], #0, 19f\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" - "ld1w { z8.s }, p4/Z, [x12]\n" - "ld1w { z9.s }, p3/Z, [x12, #1, MUL VL]\n" - "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" - "ld1w { z11.s }, p1/Z, [x12, #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" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, 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" "b 20f\n" "19:" // Height 2: no accumulate "mov z8.b, #0x0\n" @@ -375,174 +375,174 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov z14.b, #0x0\n" "mov z15.b, #0x0\n" "20:" // Height 2: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "21:" // Height 2: String loop - "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w26, [x19, x27, LSL #0x2]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "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" "tbz %x[flags], #3, 22f\n" - "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" - "add x20, x20, x19, LSL #3\n" - "ldr x25, [x20, #0x0]\n" - "ldr x24, [x20, #0x8]\n" - "cbnz x27, 23f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #2\n" - "add x24, x24, x19, LSL #2\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" + "cbnz x28, 23f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #2\n" + "add x25, x25, x20, LSL #2\n" "b 23f\n" "22:" // Height 2: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #2\n" + "mov x26, %x[input_ptr]\n" + "add x25, x26, x20, LSL #2\n" "23:" // Height 2: input setup done - "cmp x26, #0x4\n" + "cmp x27, #0x4\n" "ble 25f\n" "24:" // Height 2: Multiply loop: Main loop head - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "sub x26, x26, #0x4\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\n" + "whilelt p0.s, XZR, x27\n" + "ld1rqw { z0.s }, p0/Z, [x26]\n" + "ld1rqw { z1.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, [x9]\n" - "ld1w { z7.s }, p5/Z, [x28]\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, [x11, #1, MUL VL]\n" - "cmp x26, #0x4\n" + "ld1w { z6.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, [x10, #1, MUL VL]\n" - "add x25, x25, #0x10\n" + "ld1w { z7.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, [x9, #1, MUL VL]\n" - "add x24, x24, #0x10\n" + "ld1w { z6.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, [x28, #1, MUL VL]\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, [x11, #2, MUL VL]\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, [x10, #2, MUL VL]\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, [x9, #2, MUL VL]\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, [x28, #2, MUL VL]\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, [x11, #3, MUL VL]\n" - "addvl x11, x11, #4\n" + "ld1w { z6.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, [x10, #3, MUL VL]\n" - "addvl x10, x10, #4\n" + "ld1w { z7.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, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" + "ld1w { z6.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, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" + "ld1w { z7.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" "bgt 24b\n" "25:" // Height 2: Multiply loop: Single iteration only - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "subs x26, x26, #0x1\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\n" + "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, [x9]\n" - "ld1w { z7.s }, p5/Z, [x28]\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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "addvl x10, x10, #1\n" "fmla z11.s, z7.s, z0.s[0]\n" "fmla z15.s, z7.s, z1.s[0]\n" + "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "addvl x28, x28, #1\n" "ble 26f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "ld1w { z7.s }, p5/Z, [x28]\n" - "subs x26, x26, #0x1\n" + "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z7.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" - "addvl x11, x11, #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" + "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "addvl x28, x28, #1\n" "ble 26f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "ld1w { z7.s }, p5/Z, [x28]\n" - "subs x26, x26, #0x1\n" + "ld1w { z6.s }, p5/Z, [x10]\n" + "ld1w { z7.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" - "addvl x11, x11, #1\n" + "addvl x12, x12, #1\n" "fmla z11.s, z7.s, z0.s[2]\n" "fmla z15.s, z7.s, z1.s[2]\n" + "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "addvl x28, x28, #1\n" "ble 26f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "ld1w { z7.s }, p5/Z, [x28]\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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "addvl x10, x10, #1\n" "fmla z11.s, z7.s, z0.s[3]\n" "fmla z15.s, z7.s, z1.s[3]\n" + "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "addvl x28, x28, #1\n" "26:" // Height 2: Multiply loop: multiply skip - "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x27, x27, #0x1\n" - "cmp x27, x19\n" + "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x28, x28, #0x1\n" + "cmp x28, x20\n" "bne 21b\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x13, x20, LSL #2\n" "tbz %x[flags], #1, 27f\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" "fmin 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" @@ -560,87 +560,87 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmax z14.s, p5/M, z14.s, z0.s\n" "fmax z15.s, p5/M, z15.s, z0.s\n" "27:" // Height 2: No activation - "st1w { z8.s }, p4, [x12]\n" - "st1w { z9.s }, p3, [x12, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x12, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x12, #3, MUL VL]\n" - "addvl x12, x12, #4\n" - "st1w { z12.s }, p4, [x24]\n" - "st1w { z13.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z15.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x25]\n" + "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x25, #3, MUL VL]\n" "28:" // Height 2: Writeback done - "decw x13, ALL, MUL #4\n" - "cmp x13, XZR\n" + "decw x14, ALL, MUL #4\n" + "cmp x14, XZR\n" "bgt 16b\n" "b 86f\n" "29:" // Height 3 - "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[bias]\n" - "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x12, %x[output_ptr]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x15, %x[bias]\n" + "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "mov x13, %x[output_ptr]\n" "30:" // Height 3: Column loop - "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x10, x11, x19, LSL #2\n" - "cntw x20, ALL, MUL #3\n" - "add x9, x10, x19, LSL #2\n" - "add x28, x9, x19, LSL #2\n" - "add x19, x28, x19, LSL #2\n" - "cmp x13, x20\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" + "add x11, x12, x20, LSL #2\n" + "cntw x21, ALL, MUL #3\n" + "add x10, x11, x20, LSL #2\n" + "add x9, x10, x20, LSL #2\n" + "add x20, x9, x20, LSL #2\n" + "cmp x14, x21\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 31f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x28, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x9, x12\n" "bgt 31f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x9, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x10, x12\n" "bgt 31f\n" - "mov x10, x11\n" + "mov x11, x12\n" "31:" // Height 3: B setup done - "mov x19, #0x0\n" - "whilelt p4.s, x19, x13\n" - "incw x19\n" - "whilelt p3.s, x19, x13\n" - "incw x19\n" - "whilelt p2.s, x19, x13\n" - "incw x19\n" - "whilelt p1.s, x19, x13\n" - "cbz x14, 32f\n" - "ld1w { z8.s }, p5/Z, [x14]\n" - "ld1w { z9.s }, p5/Z, [x14, #1, MUL VL]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x14\n" + "incw x20\n" + "whilelt p3.s, x20, x14\n" + "incw x20\n" + "whilelt p2.s, x20, x14\n" + "incw x20\n" + "whilelt p1.s, x20, x14\n" + "cbz x15, 32f\n" + "ld1w { z8.s }, p5/Z, [x15]\n" + "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" - "ld1w { z10.s }, p5/Z, [x14, #2, MUL VL]\n" - "ld1w { z11.s }, p5/Z, [x14, #3, MUL VL]\n" + "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" + "ld1w { z11.s }, 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 x14, x14, #4\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 x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "ld1w { z8.s }, p4/Z, [x12]\n" - "ld1w { z9.s }, p3/Z, [x12, #1, MUL VL]\n" - "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" - "ld1w { z11.s }, p1/Z, [x12, #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" + "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" + "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" "b 34f\n" "33:" // Height 3: no accumulate "mov z8.b, #0x0\n" @@ -656,102 +656,102 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov z18.b, #0x0\n" "mov z19.b, #0x0\n" "34:" // Height 3: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "35:" // Height 3: String loop - "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w26, [x19, x27, LSL #0x2]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "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" "tbz %x[flags], #3, 36f\n" - "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" - "add x20, x20, x19, LSL #3\n" - "ldr x25, [x20, #0x0]\n" - "ldr x24, [x20, #0x8]\n" - "ldr x23, [x20, #0x10]\n" - "cbnz x27, 37f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #2\n" - "add x24, x24, x19, LSL #2\n" - "add x23, x23, x19, LSL #2\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" + "cbnz x28, 37f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #2\n" + "add x25, x25, x20, LSL #2\n" + "add x24, x24, x20, LSL #2\n" "b 37f\n" "36:" // Height 3: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" + "mov x26, %x[input_ptr]\n" + "add x25, x26, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "37:" // Height 3: input setup done - "cmp x26, #0x4\n" + "cmp x27, #0x4\n" "ble 39f\n" "38:" // Height 3: Multiply loop: Main loop head - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "sub x26, x26, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x23]\n" - "ld1w { z6.s }, p5/Z, [x11]\n" + "whilelt p0.s, XZR, x27\n" + "ld1rqw { z0.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, [x10]\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, [x9]\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, [x28]\n" - "cmp x26, #0x4\n" + "ld1w { z7.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" + "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" - "add x24, x24, #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, [x11, #1, MUL VL]\n" - "add x23, x23, #0x10\n" + "ld1w { z6.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, [x10, #1, MUL VL]\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, [x9, #1, MUL VL]\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, [x28, #1, MUL VL]\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, [x11, #2, MUL VL]\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, [x10, #2, MUL VL]\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, [x9, #2, MUL VL]\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, [x28, #2, MUL VL]\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, [x11, #3, MUL VL]\n" - "addvl x11, x11, #4\n" + "ld1w { z6.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, [x10, #3, MUL VL]\n" - "addvl x10, x10, #4\n" + "ld1w { z7.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, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" + "ld1w { z6.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, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" + "ld1w { z7.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" @@ -760,109 +760,109 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z19.s, z7.s, z2.s[3]\n" "bgt 38b\n" "39:" // Height 3: Multiply loop: Single iteration only - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "subs x26, x26, #0x1\n" - "ld1rqw { z2.s }, p0/Z, [x23]\n" - "ld1w { z6.s }, p5/Z, [x11]\n" + "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" + "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, [x10]\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, [x9]\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, [x28]\n" - "addvl x11, x11, #1\n" + "ld1w { z7.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" + "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "addvl x9, x9, #1\n" "fmla z18.s, z6.s, z2.s[0]\n" "fmla z11.s, z7.s, z0.s[0]\n" - "addvl x28, x28, #1\n" + "addvl x9, x9, #1\n" "fmla z15.s, z7.s, z1.s[0]\n" "fmla z19.s, z7.s, z2.s[0]\n" "ble 40f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "subs x26, x26, #0x1\n" + "ld1w { z6.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, [x28]\n" - "addvl x11, x11, #1\n" + "ld1w { z7.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" + "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "addvl x9, x9, #1\n" "fmla z18.s, z6.s, z2.s[1]\n" "fmla z11.s, z7.s, z0.s[1]\n" - "addvl x28, x28, #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" "ble 40f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "subs x26, x26, #0x1\n" + "ld1w { z6.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, [x28]\n" - "addvl x11, x11, #1\n" + "ld1w { z7.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" + "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "addvl x9, x9, #1\n" "fmla z18.s, z6.s, z2.s[2]\n" "fmla z11.s, z7.s, z0.s[2]\n" - "addvl x28, x28, #1\n" + "addvl x9, x9, #1\n" "fmla z15.s, z7.s, z1.s[2]\n" "fmla z19.s, z7.s, z2.s[2]\n" "ble 40f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "addvl x11, x11, #1\n" + "ld1w { z6.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, [x28]\n" - "addvl x10, x10, #1\n" + "ld1w { z7.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" + "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" - "addvl x28, x28, #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" "40:" // Height 3: Multiply loop: multiply skip - "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x27, x27, #0x1\n" - "cmp x27, x19\n" + "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x28, x28, #0x1\n" + "cmp x28, x20\n" "bne 35b\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x13, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "tbz %x[flags], #1, 41f\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" "fmin 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" @@ -888,71 +888,71 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmax z18.s, p5/M, z18.s, z0.s\n" "fmax z19.s, p5/M, z19.s, z0.s\n" "41:" // Height 3: No activation - "st1w { z8.s }, p4, [x12]\n" - "st1w { z9.s }, p3, [x12, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x12, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x12, #3, MUL VL]\n" - "addvl x12, x12, #4\n" - "st1w { z12.s }, p4, [x24]\n" - "st1w { z13.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z15.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x23]\n" - "st1w { z17.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x23, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x25]\n" + "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x24]\n" + "st1w { z17.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x24, #3, MUL VL]\n" "42:" // Height 3: Writeback done - "decw x13, ALL, MUL #4\n" - "cmp x13, XZR\n" + "decw x14, ALL, MUL #4\n" + "cmp x14, XZR\n" "bgt 30b\n" "b 86f\n" "43:" // Height 4 - "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[bias]\n" - "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x12, %x[output_ptr]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x15, %x[bias]\n" + "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "mov x13, %x[output_ptr]\n" "44:" // Height 4: Column loop - "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x10, x11, x19, LSL #2\n" - "cntw x20, ALL, MUL #3\n" - "add x9, x10, x19, LSL #2\n" - "add x28, x9, x19, LSL #2\n" - "add x19, x28, x19, LSL #2\n" - "cmp x13, x20\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" + "add x11, x12, x20, LSL #2\n" + "cntw x21, ALL, MUL #3\n" + "add x10, x11, x20, LSL #2\n" + "add x9, x10, x20, LSL #2\n" + "add x20, x9, x20, LSL #2\n" + "cmp x14, x21\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 45f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x28, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x9, x12\n" "bgt 45f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x9, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x10, x12\n" "bgt 45f\n" - "mov x10, x11\n" + "mov x11, x12\n" "45:" // Height 4: B setup done - "mov x19, #0x0\n" - "whilelt p4.s, x19, x13\n" - "incw x19\n" - "whilelt p3.s, x19, x13\n" - "incw x19\n" - "whilelt p2.s, x19, x13\n" - "incw x19\n" - "whilelt p1.s, x19, x13\n" - "cbz x14, 46f\n" - "ld1w { z8.s }, p5/Z, [x14]\n" - "ld1w { z9.s }, p5/Z, [x14, #1, MUL VL]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x14\n" + "incw x20\n" + "whilelt p3.s, x20, x14\n" + "incw x20\n" + "whilelt p2.s, x20, x14\n" + "incw x20\n" + "whilelt p1.s, x20, x14\n" + "cbz x15, 46f\n" + "ld1w { z8.s }, p5/Z, [x15]\n" + "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" - "ld1w { z10.s }, p5/Z, [x14, #2, MUL VL]\n" - "ld1w { z11.s }, p5/Z, [x14, #3, MUL VL]\n" + "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" + "ld1w { z11.s }, 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 x14, x14, #4\n" + "addvl x15, x15, #4\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" "mov z20.d, z8.d\n" @@ -962,26 +962,26 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 48f\n" "46:" // Height 4: no bias "tbz %x[flags], #0, 47f\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "ld1w { z8.s }, p4/Z, [x12]\n" - "add x22, x23, x19, LSL #2\n" - "ld1w { z9.s }, p3/Z, [x12, #1, MUL VL]\n" - "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" - "ld1w { z11.s }, p1/Z, [x12, #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" + "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" + "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" "b 48f\n" "47:" // Height 4: no accumulate "mov z8.b, #0x0\n" @@ -1001,121 +1001,121 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov z22.b, #0x0\n" "mov z23.b, #0x0\n" "48:" // Height 4: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "49:" // Height 4: String loop - "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w26, [x19, x27, LSL #0x2]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "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" "tbz %x[flags], #3, 50f\n" - "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" - "add x20, x20, x19, LSL #3\n" - "ldr x25, [x20, #0x0]\n" - "ldr x24, [x20, #0x8]\n" - "ldr x23, [x20, #0x10]\n" - "ldr x22, [x20, #0x18]\n" - "cbnz x27, 51f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #2\n" - "add x24, x24, x19, LSL #2\n" - "add x23, x23, x19, LSL #2\n" - "add x22, x22, x19, LSL #2\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" + "cbnz x28, 51f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #2\n" + "add x25, x25, x20, LSL #2\n" + "add x24, x24, x20, LSL #2\n" + "add x23, x23, x20, LSL #2\n" "b 51f\n" "50:" // Height 4: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "add x22, x23, x19, LSL #2\n" + "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" "51:" // Height 4: input setup done - "cmp x26, #0x4\n" + "cmp x27, #0x4\n" "ble 53f\n" "52:" // Height 4: Multiply loop: Main loop head - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "sub x26, x26, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x23]\n" - "ld1rqw { z3.s }, p0/Z, [x22]\n" - "cmp x26, #0x4\n" - "add x25, x25, #0x10\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\n" + "whilelt p0.s, XZR, x27\n" + "ld1rqw { z0.s }, p0/Z, [x26]\n" + "ld1rqw { z1.s }, p0/Z, [x25]\n" + "sub x27, x27, #0x4\n" + "ld1rqw { z2.s }, p0/Z, [x24]\n" + "ld1rqw { z3.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, [x9]\n" - "add x24, x24, #0x10\n" + "ld1w { z6.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" + "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" - "add x22, x22, #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, [x28]\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, [x11, #1, MUL VL]\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, [x10, #1, MUL VL]\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, [x9, #1, MUL VL]\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, [x28, #1, MUL VL]\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, [x11, #2, MUL VL]\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, [x10, #2, MUL VL]\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, [x9, #2, MUL VL]\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, [x28, #2, MUL VL]\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, [x11, #3, MUL VL]\n" - "addvl x11, x11, #4\n" + "ld1w { z6.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, [x10, #3, MUL VL]\n" - "addvl x10, x10, #4\n" + "ld1w { z7.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, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" + "ld1w { z6.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, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" + "ld1w { z7.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" @@ -1126,28 +1126,28 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z23.s, z7.s, z3.s[3]\n" "bgt 52b\n" "53:" // Height 4: Multiply loop: Single iteration only - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "subs x26, x26, #0x1\n" - "ld1rqw { z2.s }, p0/Z, [x23]\n" - "ld1rqw { z3.s }, p0/Z, [x22]\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\n" + "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" + "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, [x9]\n" - "addvl x11, x11, #1\n" + "ld1w { z6.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" + "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "addvl x9, x9, #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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -1157,25 +1157,25 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z19.s, z7.s, z2.s[0]\n" "fmla z23.s, z7.s, z3.s[0]\n" "ble 54f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "subs x26, x26, #0x1\n" + "ld1w { z6.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" + "addvl x12, x12, #1\n" "addvl x11, x11, #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" - "ld1w { z7.s }, p5/Z, [x28]\n" - "addvl x9, x9, #1\n" + "ld1w { z7.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" - "addvl x28, x28, #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" @@ -1183,25 +1183,25 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z19.s, z7.s, z2.s[1]\n" "fmla z23.s, z7.s, z3.s[1]\n" "ble 54f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "subs x26, x26, #0x1\n" + "ld1w { z6.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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "addvl x10, x10, #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, [x28]\n" - "addvl x9, x9, #1\n" + "ld1w { z7.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" - "addvl x28, x28, #1\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" @@ -1209,22 +1209,22 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z19.s, z7.s, z2.s[2]\n" "fmla z23.s, z7.s, z3.s[2]\n" "ble 54f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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, [x9]\n" - "addvl x11, x11, #1\n" + "ld1w { z6.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" + "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "addvl x9, x9, #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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -1234,19 +1234,19 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z19.s, z7.s, z2.s[3]\n" "fmla z23.s, z7.s, z3.s[3]\n" "54:" // Height 4: Multiply loop: multiply skip - "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x27, x27, #0x1\n" - "cmp x27, x19\n" + "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x28, x28, #0x1\n" + "cmp x28, x20\n" "bne 49b\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "add x22, x23, x19, LSL #2\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 x23, x24, x20, LSL #2\n" "tbz %x[flags], #1, 55f\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" "fmin 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" @@ -1280,75 +1280,75 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmax z22.s, p5/M, z22.s, z0.s\n" "fmax z23.s, p5/M, z23.s, z0.s\n" "55:" // Height 4: No activation - "st1w { z8.s }, p4, [x12]\n" - "st1w { z9.s }, p3, [x12, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x12, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x12, #3, MUL VL]\n" - "addvl x12, x12, #4\n" - "st1w { z12.s }, p4, [x24]\n" - "st1w { z13.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z15.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x23]\n" - "st1w { z17.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z20.s }, p4, [x22]\n" - "st1w { z21.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z22.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z23.s }, p1, [x22, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x25]\n" + "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x24]\n" + "st1w { z17.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z20.s }, p4, [x23]\n" + "st1w { z21.s }, p3, [x23, #1, MUL VL]\n" + "st1w { z22.s }, p2, [x23, #2, MUL VL]\n" + "st1w { z23.s }, p1, [x23, #3, MUL VL]\n" "56:" // Height 4: Writeback done - "decw x13, ALL, MUL #4\n" - "cmp x13, XZR\n" + "decw x14, ALL, MUL #4\n" + "cmp x14, XZR\n" "bgt 44b\n" "b 86f\n" "57:" // Height 5 - "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[bias]\n" - "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x12, %x[output_ptr]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x15, %x[bias]\n" + "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "mov x13, %x[output_ptr]\n" "58:" // Height 5: Column loop - "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x10, x11, x19, LSL #2\n" - "cntw x20, ALL, MUL #3\n" - "add x9, x10, x19, LSL #2\n" - "add x28, x9, x19, LSL #2\n" - "add x19, x28, x19, LSL #2\n" - "cmp x13, x20\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" + "add x11, x12, x20, LSL #2\n" + "cntw x21, ALL, MUL #3\n" + "add x10, x11, x20, LSL #2\n" + "add x9, x10, x20, LSL #2\n" + "add x20, x9, x20, LSL #2\n" + "cmp x14, x21\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 59f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x28, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x9, x12\n" "bgt 59f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x9, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x10, x12\n" "bgt 59f\n" - "mov x10, x11\n" + "mov x11, x12\n" "59:" // Height 5: B setup done - "mov x19, #0x0\n" - "whilelt p4.s, x19, x13\n" - "incw x19\n" - "whilelt p3.s, x19, x13\n" - "incw x19\n" - "whilelt p2.s, x19, x13\n" - "incw x19\n" - "whilelt p1.s, x19, x13\n" - "cbz x14, 60f\n" - "ld1w { z8.s }, p5/Z, [x14]\n" - "ld1w { z9.s }, p5/Z, [x14, #1, MUL VL]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x14\n" + "incw x20\n" + "whilelt p3.s, x20, x14\n" + "incw x20\n" + "whilelt p2.s, x20, x14\n" + "incw x20\n" + "whilelt p1.s, x20, x14\n" + "cbz x15, 60f\n" + "ld1w { z8.s }, p5/Z, [x15]\n" + "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" - "ld1w { z10.s }, p5/Z, [x14, #2, MUL VL]\n" - "ld1w { z11.s }, p5/Z, [x14, #3, MUL VL]\n" + "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" + "ld1w { z11.s }, 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 x14, x14, #4\n" + "addvl x15, x15, #4\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" "mov z20.d, z8.d\n" @@ -1362,31 +1362,31 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 62f\n" "60:" // Height 5: no bias "tbz %x[flags], #0, 61f\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "ld1w { z8.s }, p4/Z, [x12]\n" - "add x22, x23, x19, LSL #2\n" - "add x21, x22, x19, LSL #2\n" - "ld1w { z9.s }, p3/Z, [x12, #1, MUL VL]\n" - "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" - "ld1w { z11.s }, p1/Z, [x12, #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" + "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 x22, x23, 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" "b 62f\n" "61:" // Height 5: no accumulate "mov z8.b, #0x0\n" @@ -1410,140 +1410,140 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov z26.b, #0x0\n" "mov z27.b, #0x0\n" "62:" // Height 5: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "63:" // Height 5: String loop - "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w26, [x19, x27, LSL #0x2]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "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" "tbz %x[flags], #3, 64f\n" - "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" - "add x20, x20, x19, LSL #3\n" - "ldr x25, [x20, #0x0]\n" - "ldr x24, [x20, #0x8]\n" - "ldr x23, [x20, #0x10]\n" - "ldr x22, [x20, #0x18]\n" - "ldr x21, [x20, #0x20]\n" - "cbnz x27, 65f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #2\n" - "add x24, x24, x19, LSL #2\n" - "add x23, x23, x19, LSL #2\n" - "add x22, x22, x19, LSL #2\n" - "add x21, x21, x19, LSL #2\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" + "cbnz x28, 65f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #2\n" + "add x25, x25, x20, LSL #2\n" + "add x24, x24, x20, LSL #2\n" + "add x23, x23, x20, LSL #2\n" + "add x22, x22, x20, LSL #2\n" "b 65f\n" "64:" // Height 5: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "add x22, x23, x19, LSL #2\n" - "add x21, x22, x19, LSL #2\n" + "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" "65:" // Height 5: input setup done - "cmp x26, #0x4\n" + "cmp x27, #0x4\n" "ble 67f\n" "66:" // Height 5: Multiply loop: Main loop head - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "sub x26, x26, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x23]\n" - "ld1rqw { z3.s }, p0/Z, [x22]\n" - "cmp x26, #0x4\n" - "add x25, x25, #0x10\n" - "ld1rqw { z4.s }, p0/Z, [x21]\n" - "ld1w { z6.s }, p5/Z, [x11]\n" + "whilelt p0.s, XZR, x27\n" + "ld1rqw { z0.s }, p0/Z, [x26]\n" + "ld1rqw { z1.s }, p0/Z, [x25]\n" + "sub x27, x27, #0x4\n" + "ld1rqw { z2.s }, p0/Z, [x24]\n" + "ld1rqw { z3.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, [x10]\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" - "add x24, x24, #0x10\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, [x9]\n" - "add x23, x23, #0x10\n" + "ld1w { z6.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" + "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - "add x21, x21, #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, [x28]\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, [x11, #1, MUL VL]\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, [x10, #1, MUL VL]\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, [x9, #1, MUL VL]\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, [x28, #1, MUL VL]\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, [x11, #2, MUL VL]\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, [x10, #2, MUL VL]\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, [x9, #2, MUL VL]\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, [x28, #2, MUL VL]\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, [x11, #3, MUL VL]\n" - "addvl x11, x11, #4\n" + "ld1w { z6.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, [x10, #3, MUL VL]\n" - "addvl x10, x10, #4\n" + "ld1w { z7.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, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" + "ld1w { z6.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, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" + "ld1w { z7.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" @@ -1556,31 +1556,31 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z27.s, z7.s, z4.s[3]\n" "bgt 66b\n" "67:" // Height 5: Multiply loop: Single iteration only - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "subs x26, x26, #0x1\n" - "ld1rqw { z2.s }, p0/Z, [x23]\n" - "ld1rqw { z3.s }, p0/Z, [x22]\n" - "ld1rqw { z4.s }, p0/Z, [x21]\n" - "ld1w { z6.s }, p5/Z, [x11]\n" + "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" + "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, [x10]\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" - "addvl x11, x11, #1\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, [x9]\n" - "addvl x10, x10, #1\n" + "ld1w { z6.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" - "addvl x9, x9, #1\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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -1592,25 +1592,25 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z23.s, z7.s, z3.s[0]\n" "fmla z27.s, z7.s, z4.s[0]\n" "ble 68f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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" - "subs x26, x26, #0x1\n" - "addvl x11, x11, #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, [x9]\n" - "addvl x10, x10, #1\n" + "ld1w { z6.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" - "addvl x9, x9, #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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -1622,25 +1622,25 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z23.s, z7.s, z3.s[1]\n" "fmla z27.s, z7.s, z4.s[1]\n" "ble 68f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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" - "subs x26, x26, #0x1\n" - "addvl x11, x11, #1\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, [x9]\n" - "addvl x10, x10, #1\n" + "ld1w { z6.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" - "addvl x9, x9, #1\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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -1652,24 +1652,24 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z23.s, z7.s, z3.s[2]\n" "fmla z27.s, z7.s, z4.s[2]\n" "ble 68f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "addvl x10, x10, #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, [x9]\n" - "addvl x9, x9, #1\n" + "ld1w { z6.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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -1681,20 +1681,20 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z23.s, z7.s, z3.s[3]\n" "fmla z27.s, z7.s, z4.s[3]\n" "68:" // Height 5: Multiply loop: multiply skip - "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x27, x27, #0x1\n" - "cmp x27, x19\n" + "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x28, x28, #0x1\n" + "cmp x28, x20\n" "bne 63b\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "add x22, x23, x19, LSL #2\n" - "add x21, x22, x19, LSL #2\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 x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" "tbz %x[flags], #1, 69f\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" "fmin 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" @@ -1736,82 +1736,82 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmax z26.s, p5/M, z26.s, z0.s\n" "fmax z27.s, p5/M, z27.s, z0.s\n" "69:" // Height 5: No activation - "st1w { z8.s }, p4, [x12]\n" - "st1w { z9.s }, p3, [x12, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x12, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x12, #3, MUL VL]\n" - "addvl x12, x12, #4\n" - "st1w { z12.s }, p4, [x24]\n" - "st1w { z13.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z15.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x23]\n" - "st1w { z17.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z20.s }, p4, [x22]\n" - "st1w { z21.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z22.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z23.s }, p1, [x22, #3, MUL VL]\n" - "st1w { z24.s }, p4, [x21]\n" - "st1w { z25.s }, p3, [x21, #1, MUL VL]\n" - "st1w { z26.s }, p2, [x21, #2, MUL VL]\n" - "st1w { z27.s }, p1, [x21, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x25]\n" + "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x24]\n" + "st1w { z17.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z20.s }, p4, [x23]\n" + "st1w { z21.s }, p3, [x23, #1, MUL VL]\n" + "st1w { z22.s }, p2, [x23, #2, MUL VL]\n" + "st1w { z23.s }, p1, [x23, #3, MUL VL]\n" + "st1w { z24.s }, p4, [x22]\n" + "st1w { z25.s }, p3, [x22, #1, MUL VL]\n" + "st1w { z26.s }, p2, [x22, #2, MUL VL]\n" + "st1w { z27.s }, p1, [x22, #3, MUL VL]\n" "70:" // Height 5: Writeback done - "decw x13, ALL, MUL #4\n" - "cmp x13, XZR\n" + "decw x14, ALL, MUL #4\n" + "cmp x14, XZR\n" "bgt 58b\n" "b 86f\n" "71:" // Height 6 - "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x20, #0x18\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "mov x14, %x[bias]\n" - "mov x12, %x[output_ptr]\n" - "madd %x[output_ptr], x19, x20, %x[output_ptr]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "mov x21, #0x18\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\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 x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x10, x11, x19, LSL #2\n" - "cntw x20, ALL, MUL #3\n" - "add x9, x10, x19, LSL #2\n" - "add x28, x9, x19, LSL #2\n" - "add x19, x28, x19, LSL #2\n" - "cmp x13, x20\n" - "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" + "add x11, x12, x20, LSL #2\n" + "cntw x21, ALL, MUL #3\n" + "add x10, x11, x20, LSL #2\n" + "add x9, x10, x20, LSL #2\n" + "add x20, x9, x20, LSL #2\n" + "cmp x14, x21\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 73f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x28, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x9, x12\n" "bgt 73f\n" - "decw x20\n" - "cmp x13, x20\n" - "mov x9, x11\n" + "decw x21\n" + "cmp x14, x21\n" + "mov x10, x12\n" "bgt 73f\n" - "mov x10, x11\n" + "mov x11, x12\n" "73:" // Height 6: B setup done - "mov x19, #0x0\n" - "whilelt p4.s, x19, x13\n" - "incw x19\n" - "whilelt p3.s, x19, x13\n" - "incw x19\n" - "whilelt p2.s, x19, x13\n" - "incw x19\n" - "whilelt p1.s, x19, x13\n" - "cbz x14, 74f\n" - "ld1w { z8.s }, p5/Z, [x14]\n" - "ld1w { z9.s }, p5/Z, [x14, #1, MUL VL]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x14\n" + "incw x20\n" + "whilelt p3.s, x20, x14\n" + "incw x20\n" + "whilelt p2.s, x20, x14\n" + "incw x20\n" + "whilelt p1.s, x20, x14\n" + "cbz x15, 74f\n" + "ld1w { z8.s }, p5/Z, [x15]\n" + "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" "mov z12.d, z8.d\n" "mov z13.d, z9.d\n" - "ld1w { z10.s }, p5/Z, [x14, #2, MUL VL]\n" - "ld1w { z11.s }, p5/Z, [x14, #3, MUL VL]\n" + "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" + "ld1w { z11.s }, 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 x14, x14, #4\n" + "addvl x15, x15, #4\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" "mov z20.d, z8.d\n" @@ -1829,36 +1829,36 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 76f\n" "74:" // Height 6: no bias "tbz %x[flags], #0, 75f\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "ld1w { z8.s }, p4/Z, [x12]\n" - "add x22, x23, x19, LSL #2\n" - "add x21, x22, x19, LSL #2\n" - "ld1w { z9.s }, p3/Z, [x12, #1, MUL VL]\n" - "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" - "add x20, x21, x19, LSL #2\n" - "ld1w { z11.s }, p1/Z, [x12, #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" + "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 x22, x23, 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" + "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" "b 76f\n" "75:" // Height 6: no accumulate "mov z8.b, #0x0\n" @@ -1886,159 +1886,159 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov z30.b, #0x0\n" "mov z31.b, #0x0\n" "76:" // Height 6: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "77:" // Height 6: String loop - "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w26, [x19, x27, LSL #0x2]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "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" "tbz %x[flags], #3, 78f\n" - "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" - "add x20, x20, x19, LSL #3\n" - "ldr x25, [x20, #0x0]\n" - "ldr x24, [x20, #0x8]\n" - "ldr x23, [x20, #0x10]\n" - "ldr x22, [x20, #0x18]\n" - "ldr x21, [x20, #0x20]\n" - "ldr x20, [x20, #0x28]\n" - "cbnz x27, 79f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #2\n" - "add x24, x24, x19, LSL #2\n" - "add x23, x23, x19, LSL #2\n" - "add x22, x22, x19, LSL #2\n" - "add x21, x21, x19, LSL #2\n" - "add x20, x20, x19, LSL #2\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" + "cbnz x28, 79f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #2\n" + "add x25, x25, x20, LSL #2\n" + "add x24, x24, x20, LSL #2\n" + "add x23, x23, x20, LSL #2\n" + "add x22, x22, x20, LSL #2\n" + "add x21, x21, x20, LSL #2\n" "b 79f\n" "78:" // Height 6: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "add x22, x23, x19, LSL #2\n" - "add x21, x22, x19, LSL #2\n" - "add x20, x21, x19, LSL #2\n" + "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" "79:" // Height 6: input setup done - "cmp x26, #0x4\n" + "cmp x27, #0x4\n" "ble 81f\n" "80:" // Height 6: Multiply loop: Main loop head - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "sub x26, x26, #0x4\n" - "ld1rqw { z2.s }, p0/Z, [x23]\n" - "ld1rqw { z3.s }, p0/Z, [x22]\n" - "cmp x26, #0x4\n" + "whilelt p0.s, XZR, x27\n" + "ld1rqw { z0.s }, p0/Z, [x26]\n" + "ld1rqw { z1.s }, p0/Z, [x25]\n" + "sub x27, x27, #0x4\n" + "ld1rqw { z2.s }, p0/Z, [x24]\n" + "ld1rqw { z3.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" "add x25, x25, #0x10\n" - "ld1rqw { z4.s }, p0/Z, [x21]\n" - "ld1rqw { z5.s }, p0/Z, [x20]\n" "add x24, x24, #0x10\n" - "add x23, x23, #0x10\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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" + "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - "add x21, x21, #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, [x9]\n" - "add x20, x20, #0x10\n" + "ld1w { z6.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, [x28]\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, [x11, #1, MUL VL]\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, [x10, #1, MUL VL]\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, [x9, #1, MUL VL]\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, [x28, #1, MUL VL]\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, [x11, #2, MUL VL]\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, [x10, #2, MUL VL]\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, [x9, #2, MUL VL]\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, [x28, #2, MUL VL]\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, [x11, #3, MUL VL]\n" - "addvl x11, x11, #4\n" + "ld1w { z6.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, [x10, #3, MUL VL]\n" - "addvl x10, x10, #4\n" + "ld1w { z7.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, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" + "ld1w { z6.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, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" + "ld1w { z7.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" @@ -2053,34 +2053,34 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z31.s, z7.s, z5.s[3]\n" "bgt 80b\n" "81:" // Height 6: Multiply loop: Single iteration only - "whilelt p0.s, XZR, x26\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "ld1rqw { z1.s }, p0/Z, [x24]\n" - "subs x26, x26, #0x1\n" - "ld1rqw { z2.s }, p0/Z, [x23]\n" - "ld1rqw { z3.s }, p0/Z, [x22]\n" - "ld1rqw { z4.s }, p0/Z, [x21]\n" - "ld1rqw { z5.s }, p0/Z, [x20]\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\n" + "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" + "ld1rqw { z2.s }, p0/Z, [x24]\n" + "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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "addvl x10, x10, #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, [x9]\n" - "addvl x9, x9, #1\n" + "ld1w { z6.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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -2094,27 +2094,27 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z27.s, z7.s, z4.s[0]\n" "fmla z31.s, z7.s, z5.s[0]\n" "ble 82f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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" - "subs x26, x26, #0x1\n" - "addvl x11, x11, #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, [x9]\n" - "addvl x10, x10, #1\n" + "ld1w { z6.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" - "addvl x9, x9, #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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -2128,27 +2128,27 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z27.s, z7.s, z4.s[1]\n" "fmla z31.s, z7.s, z5.s[1]\n" "ble 82f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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" - "subs x26, x26, #0x1\n" - "addvl x11, x11, #1\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, [x9]\n" - "addvl x10, x10, #1\n" + "ld1w { z6.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" - "addvl x9, x9, #1\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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -2162,26 +2162,26 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z27.s, z7.s, z4.s[2]\n" "fmla z31.s, z7.s, z5.s[2]\n" "ble 82f\n" - "ld1w { z6.s }, p5/Z, [x11]\n" - "ld1w { z7.s }, p5/Z, [x10]\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" + "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" - "addvl x10, x10, #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, [x9]\n" - "addvl x9, x9, #1\n" + "ld1w { z6.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, [x28]\n" - "addvl x28, x28, #1\n" + "ld1w { z7.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" @@ -2195,21 +2195,21 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z27.s, z7.s, z4.s[3]\n" "fmla z31.s, z7.s, z5.s[3]\n" "82:" // Height 6: Multiply loop: multiply skip - "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x27, x27, #0x1\n" - "cmp x27, x19\n" + "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x28, x28, #0x1\n" + "cmp x28, x20\n" "bne 77b\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x12, x19, LSL #2\n" - "add x23, x24, x19, LSL #2\n" - "add x22, x23, x19, LSL #2\n" - "add x21, x22, x19, LSL #2\n" - "add x20, x21, x19, LSL #2\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 x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "tbz %x[flags], #1, 83f\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" "fmin 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" @@ -2259,50 +2259,50 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmax z30.s, p5/M, z30.s, z0.s\n" "fmax z31.s, p5/M, z31.s, z0.s\n" "83:" // Height 6: No activation - "st1w { z8.s }, p4, [x12]\n" - "st1w { z9.s }, p3, [x12, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x12, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x12, #3, MUL VL]\n" - "addvl x12, x12, #4\n" - "st1w { z12.s }, p4, [x24]\n" - "st1w { z13.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z15.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x23]\n" - "st1w { z17.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z20.s }, p4, [x22]\n" - "st1w { z21.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z22.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z23.s }, p1, [x22, #3, MUL VL]\n" - "st1w { z24.s }, p4, [x21]\n" - "st1w { z25.s }, p3, [x21, #1, MUL VL]\n" - "st1w { z26.s }, p2, [x21, #2, MUL VL]\n" - "st1w { z27.s }, p1, [x21, #3, MUL VL]\n" - "st1w { z28.s }, p4, [x20]\n" - "st1w { z29.s }, p3, [x20, #1, MUL VL]\n" - "st1w { z30.s }, p2, [x20, #2, MUL VL]\n" - "st1w { z31.s }, p1, [x20, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x25]\n" + "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x24]\n" + "st1w { z17.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z20.s }, p4, [x23]\n" + "st1w { z21.s }, p3, [x23, #1, MUL VL]\n" + "st1w { z22.s }, p2, [x23, #2, MUL VL]\n" + "st1w { z23.s }, p1, [x23, #3, MUL VL]\n" + "st1w { z24.s }, p4, [x22]\n" + "st1w { z25.s }, p3, [x22, #1, MUL VL]\n" + "st1w { z26.s }, p2, [x22, #2, MUL VL]\n" + "st1w { z27.s }, p1, [x22, #3, MUL VL]\n" + "st1w { z28.s }, p4, [x21]\n" + "st1w { z29.s }, p3, [x21, #1, MUL VL]\n" + "st1w { z30.s }, p2, [x21, #2, MUL VL]\n" + "st1w { z31.s }, p1, [x21, #3, MUL VL]\n" "84:" // Height 6: Writeback done - "decw x13, ALL, MUL #4\n" - "cmp x13, XZR\n" + "decw x14, ALL, MUL #4\n" + "cmp x14, XZR\n" "bgt 72b\n" "subs %x[M], %x[M], #0x6\n" "beq 86f\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 85f\n" - "add x20, x20, #0x6\n" - "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "add x21, x21, #0x6\n" + "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "b 1b\n" "85:" // Update direct input - "mov x19, #0x18\n" - "madd %x[input_ptr], x19, x20, %x[input_ptr]\n" + "mov x20, #0x18\n" + "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "86:" // Exit : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_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", "x19", "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" + : "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" ); } |