diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp | 1661 |
1 files changed, 830 insertions, 831 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp index e69293e3f1..9bb67f18d2 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021, 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 @@ -103,42 +103,42 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "cmp %x[M], #0x2\n" "bgt 27f\n" "beq 14f\n" + "mov x12, %x[bias]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[bias]\n" - "mov x28, %x[output_ptr]\n" + "mov x9, %x[output_ptr]\n" "2:" // Height 1: Column loop - "mov x19, #0x0\n" - "whilelt p4.s, x19, x11\n" - "incw x19\n" - "whilelt p3.s, x19, x11\n" - "incw x19\n" - "whilelt p2.s, x19, x11\n" - "incw x19\n" - "whilelt p1.s, x19, x11\n" - "cbz x9, 3f\n" - "ld1w { z8.s }, p5/Z, [x9]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x11\n" + "incw x20\n" + "whilelt p3.s, x20, x11\n" + "incw x20\n" + "whilelt p2.s, x20, x11\n" + "incw x20\n" + "whilelt p1.s, x20, x11\n" + "cbz x12, 3f\n" + "ld1w { z8.s }, p5/Z, [x12]\n" + "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" "zip2 z12.d, z8.d, z8.d\n" - "ld1w { z9.s }, p5/Z, [x9, #1, MUL VL]\n" "zip1 z8.d, z8.d, z8.d\n" - "ld1w { z10.s }, p5/Z, [x9, #2, MUL VL]\n" - "ld1w { z11.s }, p5/Z, [x9, #3, MUL VL]\n" + "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" + "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" "zip2 z13.d, z9.d, z9.d\n" - "addvl x9, x9, #4\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" + "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" "b 5f\n" "3:" // Height 1: no bias "tbz %x[flags], #0, 4f\n" - "ld1w { z9.s }, p4/Z, [x28]\n" + "ld1w { z9.s }, p4/Z, [x9]\n" + "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" "zip1 z8.d, z9.d, z12.d\n" - "ld1w { z10.s }, p3/Z, [x28, #1, MUL VL]\n" "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z11.s }, p2/Z, [x28, #2, MUL VL]\n" - "ld1w { z16.s }, p1/Z, [x28, #3, MUL VL]\n" + "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" "zip1 z9.d, z10.d, z13.d\n" "zip2 z13.d, z10.d, z13.d\n" "zip1 z10.d, z11.d, z14.d\n" @@ -156,185 +156,185 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "mov z14.b, #0x0\n" "mov z15.b, #0x0\n" "5:" // Height 1: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "6:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w26, [x20, x27, LSL #0x2]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 7f\n" - "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" - "add x20, x20, x19, LSL #3\n" - "ldr x25, [x20, #0x0]\n" - "cbnz x27, 8f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #1\n" + "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" + "add x21, x21, x20, LSL #3\n" + "ldr x26, [x21, #0x0]\n" + "cbnz x28, 8f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #1\n" "b 8f\n" "7:" // Height 1: setup direct input - "mov x25, %x[input_ptr]\n" + "mov x26, %x[input_ptr]\n" "8:" // Height 1: input setup done - "cmp x26, #0x8\n" + "cmp x27, #0x8\n" "ble 10f\n" "9:" // Height 1: Multiply loop: Main loop head + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "trn1 z0.d, z1.d, z2.d\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "sub x26, x26, #0x8\n" - "trn2 z1.d, z1.d, z2.d\n" - "cmp x26, #0x8\n" - "add x25, x25, #0x10\n" ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" + "trn2 z1.d, z1.d, z2.d\n" ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #-8, MUL VL]\n" ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #-8, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #-7, MUL VL]\n" ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #-6, MUL VL]\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #-6, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #-5, MUL VL]\n" ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #-4, MUL VL]\n" ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #-4, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #-3, MUL VL]\n" ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #-2, MUL VL]\n" ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #-2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #-1, MUL VL]\n" + "sub x27, x27, #0x8\n" + "cmp x27, #0x8\n" ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" + "add x26, x26, #0x10\n" "bgt 9b\n" "10:" // Height 1: Multiply loop: Single iteration only + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "trn1 z0.d, z1.d, z2.d\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "subs x26, x26, #0x4\n" - "trn2 z1.d, z1.d, z2.d\n" ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" - "addvl x10, x10, #8\n" + "subs x27, x27, #0x4\n" + "trn2 z1.d, z1.d, z2.d\n" ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" + "addvl x10, x10, #8\n" "ble 11f\n" "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" - "addvl x10, x10, #8\n" ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" + "addvl x10, x10, #8\n" "11:" // 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 6b\n" "uzp1 z8.d, z8.d, z12.d\n" "uzp1 z9.d, z9.d, z13.d\n" "uzp1 z10.d, z10.d, z14.d\n" "uzp1 z11.d, z11.d, z15.d\n" "tbz %x[flags], #1, 12f\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" - "fmin z8.s, p5/M, z8.s, z0.s\n" - "fmin z9.s, p5/M, z9.s, z0.s\n" - "fmin z10.s, p5/M, z10.s, z0.s\n" - "fmin z11.s, p5/M, z11.s, z0.s\n" - "fmax z8.s, p5/M, z8.s, z1.s\n" - "fmax z9.s, p5/M, z9.s, z1.s\n" - "fmax z10.s, p5/M, z10.s, z1.s\n" - "fmax z11.s, p5/M, z11.s, z1.s\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" + "fmin z11.s, p5/M, z11.s, z1.s\n" + "fmax z8.s, p5/M, z8.s, z0.s\n" + "fmax z9.s, p5/M, z9.s, z0.s\n" + "fmax z10.s, p5/M, z10.s, z0.s\n" + "fmax z11.s, p5/M, z11.s, z0.s\n" "12:" // Height 1: No activation - "st1w { z8.s }, p4, [x28]\n" - "st1w { z9.s }, p3, [x28, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x28, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" + "st1w { z8.s }, p4, [x9]\n" + "st1w { z9.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" "13:" // Height 1: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" "bgt 2b\n" "b 80f\n" "14:" // Height 2 + "mov x12, %x[bias]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" - "mov x9, %x[bias]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x28, %x[output_ptr]\n" + "mov x9, %x[output_ptr]\n" "15:" // Height 2: Column loop - "mov x19, #0x0\n" - "whilelt p4.s, x19, x11\n" - "incw x19\n" - "whilelt p3.s, x19, x11\n" - "incw x19\n" - "whilelt p2.s, x19, x11\n" - "incw x19\n" - "whilelt p1.s, x19, x11\n" - "cbz x9, 16f\n" - "ld1w { z8.s }, p5/Z, [x9]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x11\n" + "incw x20\n" + "whilelt p3.s, x20, x11\n" + "incw x20\n" + "whilelt p2.s, x20, x11\n" + "incw x20\n" + "whilelt p1.s, x20, x11\n" + "cbz x12, 16f\n" + "ld1w { z8.s }, p5/Z, [x12]\n" + "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" "zip2 z12.d, z8.d, z8.d\n" - "ld1w { z9.s }, p5/Z, [x9, #1, MUL VL]\n" "zip1 z8.d, z8.d, z8.d\n" - "ld1w { z10.s }, p5/Z, [x9, #2, MUL VL]\n" - "ld1w { z11.s }, p5/Z, [x9, #3, MUL VL]\n" + "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" + "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" "zip2 z13.d, z9.d, z9.d\n" - "addvl x9, x9, #4\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" + "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" "b 18f\n" "16:" // Height 2: no bias "tbz %x[flags], #0, 17f\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "ld1w { z9.s }, p4/Z, [x28]\n" - "add x24, x28, x19, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x28, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x28, #2, MUL VL]\n" - "ld1w { z16.s }, p1/Z, [x28, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x24]\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x9, x20, LSL #2\n" + "ld1w { z9.s }, p4/Z, [x9]\n" + "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x25]\n" "zip1 z8.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" "zip1 z9.d, z10.d, z13.d\n" "zip2 z13.d, z10.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" "zip1 z10.d, z11.d, z14.d\n" "zip2 z14.d, z11.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" @@ -350,122 +350,122 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "mov z14.b, #0x0\n" "mov z15.b, #0x0\n" "18:" // Height 2: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "19:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w26, [x20, x27, LSL #0x2]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 20f\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, 21f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #1\n" - "add x24, x24, x19, LSL #1\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, 21f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #1\n" + "add x25, x25, x20, LSL #1\n" "b 21f\n" "20:" // Height 2: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #1\n" + "mov x26, %x[input_ptr]\n" + "add x25, x26, x20, LSL #1\n" "21:" // Height 2: input setup done - "cmp x26, #0x8\n" + "cmp x27, #0x8\n" "ble 23f\n" "22:" // Height 2: Multiply loop: Main loop head + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" + "trn1 z0.d, z1.d, z2.d\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "sub x26, x26, #0x8\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "cmp x26, #0x8\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" - "trn1 z0.d, z1.d, z2.d\n" - "add x25, x25, #0x10\n" - "trn2 z1.d, z1.d, z2.d\n" - "add x24, x24, #0x10\n" ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" + "trn2 z1.d, z1.d, z2.d\n" ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #-8, MUL VL]\n" ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #-8, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #-7, MUL VL]\n" ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #-6, MUL VL]\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #-6, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #-5, MUL VL]\n" ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #-4, MUL VL]\n" ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #-4, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #-3, MUL VL]\n" ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #-2, MUL VL]\n" ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #-2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #-1, MUL VL]\n" + "sub x27, x27, #0x8\n" + "cmp x27, #0x8\n" ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" + "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" "bgt 22b\n" "23:" // Height 2: Multiply loop: Single iteration only + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" + "trn1 z0.d, z1.d, z2.d\n" "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "subs x26, x26, #0x4\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" - "trn1 z0.d, z1.d, z2.d\n" - "trn2 z1.d, z1.d, z2.d\n" ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" - "addvl x10, x10, #8\n" + "subs x27, x27, #0x4\n" + "trn2 z1.d, z1.d, z2.d\n" ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" + "addvl x10, x10, #8\n" "ble 24f\n" "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" - "addvl x10, x10, #8\n" ".inst 0x6467e42b // bfmmla z11.s, z1.h, z7.h\n" ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" + "addvl x10, x10, #8\n" "24:" // 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 19b\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp1 z7.d, z8.d, z12.d\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp2 z8.d, z8.d, z12.d\n" - "add x24, x28, x19, LSL #2\n" + "add x25, x9, x20, LSL #2\n" "uzp1 z12.d, z9.d, z13.d\n" "uzp2 z9.d, z9.d, z13.d\n" "uzp1 z13.d, z10.d, z14.d\n" @@ -473,70 +473,70 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "uzp1 z14.d, z11.d, z15.d\n" "uzp2 z11.d, z11.d, z15.d\n" "tbz %x[flags], #1, 25f\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" - "fmin z7.s, p5/M, z7.s, z0.s\n" - "fmin z12.s, p5/M, z12.s, z0.s\n" - "fmin z13.s, p5/M, z13.s, z0.s\n" - "fmin z14.s, p5/M, z14.s, z0.s\n" - "fmin z8.s, p5/M, z8.s, z0.s\n" - "fmax z7.s, p5/M, z7.s, z1.s\n" - "fmax z12.s, p5/M, z12.s, z1.s\n" - "fmax z13.s, p5/M, z13.s, z1.s\n" - "fmax z14.s, p5/M, z14.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z0.s\n" - "fmin z10.s, p5/M, z10.s, z0.s\n" - "fmin z11.s, p5/M, z11.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z1.s\n" - "fmax z10.s, p5/M, z10.s, z1.s\n" - "fmax z11.s, p5/M, z11.s, z1.s\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" + "fmin z7.s, p5/M, z7.s, z1.s\n" + "fmin z12.s, p5/M, z12.s, z1.s\n" + "fmin z13.s, p5/M, z13.s, z1.s\n" + "fmin z14.s, p5/M, z14.s, z1.s\n" + "fmin z8.s, p5/M, z8.s, z1.s\n" + "fmin z9.s, p5/M, z9.s, z1.s\n" + "fmin z10.s, p5/M, z10.s, z1.s\n" + "fmin z11.s, p5/M, z11.s, z1.s\n" + "fmax z7.s, p5/M, z7.s, z0.s\n" + "fmax z12.s, p5/M, z12.s, z0.s\n" + "fmax z13.s, p5/M, z13.s, z0.s\n" + "fmax z14.s, p5/M, z14.s, z0.s\n" + "fmax z8.s, p5/M, z8.s, z0.s\n" + "fmax z9.s, p5/M, z9.s, z0.s\n" + "fmax z10.s, p5/M, z10.s, z0.s\n" + "fmax z11.s, p5/M, z11.s, z0.s\n" "25:" // Height 2: No activation - "st1w { z7.s }, p4, [x28]\n" - "st1w { z12.s }, p3, [x28, #1, MUL VL]\n" - "st1w { z13.s }, p2, [x28, #2, MUL VL]\n" - "st1w { z14.s }, p1, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" - "st1w { z8.s }, p4, [x24]\n" - "st1w { z9.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z7.s }, p4, [x9]\n" + "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "st1w { z8.s }, p4, [x25]\n" + "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x25, #3, MUL VL]\n" "26:" // Height 2: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" "bgt 15b\n" "b 80f\n" "27:" // Height 3 + "mov x12, %x[bias]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" - "mov x9, %x[bias]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x28, %x[output_ptr]\n" + "mov x9, %x[output_ptr]\n" "28:" // Height 3: Column loop - "mov x19, #0x0\n" - "whilelt p4.s, x19, x11\n" - "incw x19\n" - "whilelt p3.s, x19, x11\n" - "incw x19\n" - "whilelt p2.s, x19, x11\n" - "incw x19\n" - "whilelt p1.s, x19, x11\n" - "cbz x9, 29f\n" - "ld1w { z8.s }, p5/Z, [x9]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x11\n" + "incw x20\n" + "whilelt p3.s, x20, x11\n" + "incw x20\n" + "whilelt p2.s, x20, x11\n" + "incw x20\n" + "whilelt p1.s, x20, x11\n" + "cbz x12, 29f\n" + "ld1w { z8.s }, p5/Z, [x12]\n" + "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" "zip2 z12.d, z8.d, z8.d\n" - "ld1w { z9.s }, p5/Z, [x9, #1, MUL VL]\n" "zip1 z8.d, z8.d, z8.d\n" - "ld1w { z10.s }, p5/Z, [x9, #2, MUL VL]\n" - "mov z16.d, z8.d\n" - "ld1w { z11.s }, p5/Z, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" + "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" + "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" + "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" + "mov z16.d, z8.d\n" "mov z20.d, z12.d\n" "mov z17.d, z9.d\n" "mov z21.d, z13.d\n" @@ -547,29 +547,29 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "b 31f\n" "29:" // Height 3: no bias "tbz %x[flags], #0, 30f\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "ld1w { z9.s }, p4/Z, [x28]\n" - "add x24, x28, x19, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x28, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x28, #2, MUL VL]\n" - "add x23, x24, x19, LSL #2\n" - "ld1w { z16.s }, p1/Z, [x28, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x24]\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x9, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" + "ld1w { z9.s }, p4/Z, [x9]\n" + "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x25]\n" "zip1 z8.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" "zip1 z9.d, z10.d, z13.d\n" - "ld1w { z17.s }, p4/Z, [x23]\n" "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x24]\n" "zip1 z10.d, z11.d, z14.d\n" - "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" + "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" "zip1 z17.d, z18.d, z21.d\n" @@ -597,61 +597,58 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "mov z22.b, #0x0\n" "mov z23.b, #0x0\n" "31:" // Height 3: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "32:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w26, [x20, x27, LSL #0x2]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 33f\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, 34f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #1\n" - "add x24, x24, x19, LSL #1\n" - "add x23, x23, x19, LSL #1\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, 34f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #1\n" + "add x25, x25, x20, LSL #1\n" + "add x24, x24, x20, LSL #1\n" "b 34f\n" "33:" // Height 3: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #1\n" - "add x23, x24, x19, LSL #1\n" + "mov x26, %x[input_ptr]\n" + "add x25, x26, x20, LSL #1\n" + "add x24, x25, x20, LSL #1\n" "34:" // Height 3: input setup done - "cmp x26, #0x8\n" + "cmp x27, #0x8\n" "ble 36f\n" "35:" // Height 3: Multiply loop: Main loop head - "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "sub x26, x26, #0x8\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" + "ld1rqh { z3.h }, p0/Z, [x24]\n" "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x23]\n" - "cmp x26, #0x8\n" "trn2 z1.d, z1.d, z2.d\n" - "add x25, x25, #0x10\n" - "add x24, x24, #0x10\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - "add x23, x23, #0x10\n" + "ld1h { z7.h }, p5/Z, [x10]\n" "trn1 z2.d, z3.d, z4.d\n" - "trn2 z3.d, z3.d, z4.d\n" + "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" + "trn2 z3.d, z3.d, z4.d\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" + "sub x27, x27, #0x8\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" + "cmp x27, #0x8\n" ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" @@ -659,12 +656,15 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #-8, MUL VL]\n" + "add x26, x26, #0x10\n" ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #-7, MUL VL]\n" + "add x25, x25, #0x10\n" ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #-6, MUL VL]\n" + "add x24, x24, #0x10\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #-5, MUL VL]\n" @@ -686,29 +686,29 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" "bgt 35b\n" "36:" // Height 3: Multiply loop: Single iteration only - "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "subs x26, x26, #0x4\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" + "ld1rqh { z3.h }, p0/Z, [x24]\n" "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x23]\n" "trn2 z1.d, z1.d, z2.d\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10]\n" "trn1 z2.d, z3.d, z4.d\n" - "trn2 z3.d, z3.d, z4.d\n" + "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" + "subs x27, x27, #0x4\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" + "trn2 z3.d, z3.d, z4.d\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" @@ -722,12 +722,12 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" "ble 37f\n" "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" @@ -747,17 +747,17 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" "37:" // 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 32b\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x9, x20, LSL #2\n" "uzp1 z7.d, z8.d, z12.d\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp2 z8.d, z8.d, z12.d\n" - "add x24, x28, x19, LSL #2\n" "uzp1 z12.d, z9.d, z13.d\n" "uzp2 z9.d, z9.d, z13.d\n" - "add x23, x24, x19, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "uzp1 z13.d, z10.d, z14.d\n" "uzp2 z10.d, z10.d, z14.d\n" "uzp1 z14.d, z11.d, z15.d\n" @@ -767,82 +767,82 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "uzp1 z18.d, z18.d, z22.d\n" "uzp1 z19.d, z19.d, z23.d\n" "tbz %x[flags], #1, 38f\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" - "fmin z7.s, p5/M, z7.s, z0.s\n" - "fmin z12.s, p5/M, z12.s, z0.s\n" - "fmin z13.s, p5/M, z13.s, z0.s\n" - "fmin z14.s, p5/M, z14.s, z0.s\n" - "fmin z8.s, p5/M, z8.s, z0.s\n" - "fmax z7.s, p5/M, z7.s, z1.s\n" - "fmax z12.s, p5/M, z12.s, z1.s\n" - "fmax z13.s, p5/M, z13.s, z1.s\n" - "fmax z14.s, p5/M, z14.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z0.s\n" - "fmin z10.s, p5/M, z10.s, z0.s\n" - "fmin z11.s, p5/M, z11.s, z0.s\n" - "fmin z16.s, p5/M, z16.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z1.s\n" - "fmax z10.s, p5/M, z10.s, z1.s\n" - "fmax z11.s, p5/M, z11.s, z1.s\n" - "fmax z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z0.s\n" - "fmin z18.s, p5/M, z18.s, z0.s\n" - "fmin z19.s, p5/M, z19.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z1.s\n" - "fmax z18.s, p5/M, z18.s, z1.s\n" - "fmax z19.s, p5/M, z19.s, z1.s\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" + "fmin z7.s, p5/M, z7.s, z1.s\n" + "fmin z12.s, p5/M, z12.s, z1.s\n" + "fmin z13.s, p5/M, z13.s, z1.s\n" + "fmin z14.s, p5/M, z14.s, z1.s\n" + "fmin z8.s, p5/M, z8.s, z1.s\n" + "fmin z9.s, p5/M, z9.s, z1.s\n" + "fmin z10.s, p5/M, z10.s, z1.s\n" + "fmin z11.s, p5/M, z11.s, z1.s\n" + "fmin z16.s, p5/M, z16.s, z1.s\n" + "fmin z17.s, p5/M, z17.s, z1.s\n" + "fmin z18.s, p5/M, z18.s, z1.s\n" + "fmin z19.s, p5/M, z19.s, z1.s\n" + "fmax z7.s, p5/M, z7.s, z0.s\n" + "fmax z12.s, p5/M, z12.s, z0.s\n" + "fmax z13.s, p5/M, z13.s, z0.s\n" + "fmax z14.s, p5/M, z14.s, z0.s\n" + "fmax z8.s, p5/M, z8.s, z0.s\n" + "fmax z9.s, p5/M, z9.s, z0.s\n" + "fmax z10.s, p5/M, z10.s, z0.s\n" + "fmax z11.s, p5/M, z11.s, z0.s\n" + "fmax z16.s, p5/M, z16.s, z0.s\n" + "fmax z17.s, p5/M, z17.s, z0.s\n" + "fmax z18.s, p5/M, z18.s, z0.s\n" + "fmax z19.s, p5/M, z19.s, z0.s\n" "38:" // Height 3: No activation - "st1w { z7.s }, p4, [x28]\n" - "st1w { z12.s }, p3, [x28, #1, MUL VL]\n" - "st1w { z13.s }, p2, [x28, #2, MUL VL]\n" - "st1w { z14.s }, p1, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" - "st1w { z8.s }, p4, [x24]\n" - "st1w { z9.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z11.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 { z7.s }, p4, [x9]\n" + "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "st1w { z8.s }, p4, [x25]\n" + "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z11.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" "39:" // Height 3: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" "bgt 28b\n" "b 80f\n" "40:" // Height 4 + "mov x12, %x[bias]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" - "mov x9, %x[bias]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x28, %x[output_ptr]\n" + "mov x9, %x[output_ptr]\n" "41:" // Height 4: Column loop - "mov x19, #0x0\n" - "whilelt p4.s, x19, x11\n" - "incw x19\n" - "whilelt p3.s, x19, x11\n" - "incw x19\n" - "whilelt p2.s, x19, x11\n" - "incw x19\n" - "whilelt p1.s, x19, x11\n" - "cbz x9, 42f\n" - "ld1w { z8.s }, p5/Z, [x9]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x11\n" + "incw x20\n" + "whilelt p3.s, x20, x11\n" + "incw x20\n" + "whilelt p2.s, x20, x11\n" + "incw x20\n" + "whilelt p1.s, x20, x11\n" + "cbz x12, 42f\n" + "ld1w { z8.s }, p5/Z, [x12]\n" + "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" "zip2 z12.d, z8.d, z8.d\n" - "ld1w { z9.s }, p5/Z, [x9, #1, MUL VL]\n" "zip1 z8.d, z8.d, z8.d\n" - "ld1w { z10.s }, p5/Z, [x9, #2, MUL VL]\n" - "mov z16.d, z8.d\n" - "ld1w { z11.s }, p5/Z, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" + "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" + "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" + "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" + "mov z16.d, z8.d\n" "mov z20.d, z12.d\n" "mov z17.d, z9.d\n" "mov z21.d, z13.d\n" @@ -853,38 +853,38 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "b 44f\n" "42:" // Height 4: no bias "tbz %x[flags], #0, 43f\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "ld1w { z9.s }, p4/Z, [x28]\n" - "add x24, x28, x19, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x28, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x28, #2, MUL VL]\n" - "add x23, x24, x19, LSL #2\n" - "ld1w { z16.s }, p1/Z, [x28, #3, MUL VL]\n" - "add x22, x23, x19, LSL #2\n" - "ld1w { z12.s }, p4/Z, [x24]\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x9, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" + "ld1w { z9.s }, p4/Z, [x9]\n" + "add x23, x24, x20, LSL #2\n" + "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x25]\n" "zip1 z8.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" "zip1 z9.d, z10.d, z13.d\n" - "ld1w { z17.s }, p4/Z, [x23]\n" "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x24]\n" "zip1 z10.d, z11.d, z14.d\n" - "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" - "ld1w { z20.s }, p4/Z, [x22]\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x23]\n" "zip1 z16.d, z17.d, z20.d\n" - "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n" "zip2 z20.d, z17.d, z20.d\n" + "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" "zip1 z17.d, z18.d, z21.d\n" "zip2 z21.d, z18.d, z21.d\n" + "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" "zip1 z18.d, z19.d, z22.d\n" "zip2 z22.d, z19.d, z22.d\n" "zip1 z19.d, z24.d, z23.d\n" @@ -908,66 +908,62 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "mov z22.b, #0x0\n" "mov z23.b, #0x0\n" "44:" // Height 4: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "45:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w26, [x20, x27, LSL #0x2]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 46f\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, 47f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #1\n" - "add x24, x24, x19, LSL #1\n" - "add x23, x23, x19, LSL #1\n" - "add x22, x22, x19, LSL #1\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, 47f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #1\n" + "add x25, x25, x20, LSL #1\n" + "add x24, x24, x20, LSL #1\n" + "add x23, x23, x20, LSL #1\n" "b 47f\n" "46:" // Height 4: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #1\n" - "add x23, x24, x19, LSL #1\n" - "add x22, x23, x19, LSL #1\n" + "mov x26, %x[input_ptr]\n" + "add x25, x26, x20, LSL #1\n" + "add x24, x25, x20, LSL #1\n" + "add x23, x24, x20, LSL #1\n" "47:" // Height 4: input setup done - "cmp x26, #0x8\n" + "cmp x27, #0x8\n" "ble 49f\n" "48:" // Height 4: Multiply loop: Main loop head - "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "sub x26, x26, #0x8\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "cmp x26, #0x8\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x23]\n" - "add x25, x25, #0x10\n" + "ld1rqh { z3.h }, p0/Z, [x24]\n" + "ld1rqh { z4.h }, p0/Z, [x23]\n" "trn2 z1.d, z1.d, z2.d\n" - "ld1rqh { z4.h }, p0/Z, [x22]\n" - "add x24, x24, #0x10\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - "add x23, x23, #0x10\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" - "add x22, x22, #0x10\n" "trn1 z2.d, z3.d, z4.d\n" - "trn2 z3.d, z3.d, z4.d\n" + "ld1h { z7.h }, p5/Z, [x10]\n" + "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" + "trn2 z3.d, z3.d, z4.d\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" + "sub x27, x27, #0x8\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" + "cmp x27, #0x8\n" ".inst 0x6466e40e // bfmmla z14.s, z0.h, z6.h\n" ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" @@ -975,15 +971,19 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #-8, MUL VL]\n" + "add x26, x26, #0x10\n" ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #-7, MUL VL]\n" + "add x25, x25, #0x10\n" ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #-6, MUL VL]\n" + "add x24, x24, #0x10\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #-5, MUL VL]\n" + "add x23, x23, #0x10\n" ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #-4, MUL VL]\n" @@ -1002,30 +1002,30 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" "bgt 48b\n" "49:" // Height 4: Multiply loop: Single iteration only - "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" - "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "subs x26, x26, #0x4\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x23]\n" + "ld1rqh { z3.h }, p0/Z, [x24]\n" + "ld1rqh { z4.h }, p0/Z, [x23]\n" "trn2 z1.d, z1.d, z2.d\n" - "ld1rqh { z4.h }, p0/Z, [x22]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" "trn1 z2.d, z3.d, z4.d\n" - "trn2 z3.d, z3.d, z4.d\n" + "ld1h { z7.h }, p5/Z, [x10]\n" + "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" + "subs x27, x27, #0x4\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" + "trn2 z3.d, z3.d, z4.d\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" @@ -1039,12 +1039,12 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e457 // bfmmla z23.s, z2.h, z6.h\n" "ble 50f\n" "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" @@ -1064,19 +1064,19 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n" ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" "50:" // 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 45b\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x9, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "uzp1 z7.d, z8.d, z12.d\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp2 z8.d, z8.d, z12.d\n" - "add x24, x28, x19, LSL #2\n" "uzp1 z12.d, z9.d, z13.d\n" + "add x23, x24, x20, LSL #2\n" "uzp2 z9.d, z9.d, z13.d\n" - "add x23, x24, x19, LSL #2\n" "uzp1 z13.d, z10.d, z14.d\n" - "add x22, x23, x19, LSL #2\n" "uzp2 z10.d, z10.d, z14.d\n" "uzp1 z14.d, z11.d, z15.d\n" "uzp2 z11.d, z11.d, z15.d\n" @@ -1089,94 +1089,94 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "uzp1 z22.d, z19.d, z23.d\n" "uzp2 z19.d, z19.d, z23.d\n" "tbz %x[flags], #1, 51f\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" - "fmin z7.s, p5/M, z7.s, z0.s\n" - "fmin z12.s, p5/M, z12.s, z0.s\n" - "fmin z13.s, p5/M, z13.s, z0.s\n" - "fmin z14.s, p5/M, z14.s, z0.s\n" - "fmin z8.s, p5/M, z8.s, z0.s\n" - "fmax z7.s, p5/M, z7.s, z1.s\n" - "fmax z12.s, p5/M, z12.s, z1.s\n" - "fmax z13.s, p5/M, z13.s, z1.s\n" - "fmax z14.s, p5/M, z14.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z0.s\n" - "fmin z10.s, p5/M, z10.s, z0.s\n" - "fmin z11.s, p5/M, z11.s, z0.s\n" - "fmin z15.s, p5/M, z15.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z1.s\n" - "fmax z10.s, p5/M, z10.s, z1.s\n" - "fmax z11.s, p5/M, z11.s, z1.s\n" - "fmax z15.s, p5/M, z15.s, z1.s\n" - "fmin z20.s, p5/M, z20.s, z0.s\n" - "fmin z21.s, p5/M, z21.s, z0.s\n" - "fmin z22.s, p5/M, z22.s, z0.s\n" - "fmin z16.s, p5/M, z16.s, z0.s\n" - "fmax z20.s, p5/M, z20.s, z1.s\n" - "fmax z21.s, p5/M, z21.s, z1.s\n" - "fmax z22.s, p5/M, z22.s, z1.s\n" - "fmax z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z0.s\n" - "fmin z18.s, p5/M, z18.s, z0.s\n" - "fmin z19.s, p5/M, z19.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z1.s\n" - "fmax z18.s, p5/M, z18.s, z1.s\n" - "fmax z19.s, p5/M, z19.s, z1.s\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" + "fmin z7.s, p5/M, z7.s, z1.s\n" + "fmin z12.s, p5/M, z12.s, z1.s\n" + "fmin z13.s, p5/M, z13.s, z1.s\n" + "fmin z14.s, p5/M, z14.s, z1.s\n" + "fmin z8.s, p5/M, z8.s, z1.s\n" + "fmin z9.s, p5/M, z9.s, z1.s\n" + "fmin z10.s, p5/M, z10.s, z1.s\n" + "fmin z11.s, p5/M, z11.s, z1.s\n" + "fmin z15.s, p5/M, z15.s, z1.s\n" + "fmin z20.s, p5/M, z20.s, z1.s\n" + "fmin z21.s, p5/M, z21.s, z1.s\n" + "fmin z22.s, p5/M, z22.s, z1.s\n" + "fmin z16.s, p5/M, z16.s, z1.s\n" + "fmin z17.s, p5/M, z17.s, z1.s\n" + "fmin z18.s, p5/M, z18.s, z1.s\n" + "fmin z19.s, p5/M, z19.s, z1.s\n" + "fmax z7.s, p5/M, z7.s, z0.s\n" + "fmax z12.s, p5/M, z12.s, z0.s\n" + "fmax z13.s, p5/M, z13.s, z0.s\n" + "fmax z14.s, p5/M, z14.s, z0.s\n" + "fmax z8.s, p5/M, z8.s, z0.s\n" + "fmax z9.s, p5/M, z9.s, z0.s\n" + "fmax z10.s, p5/M, z10.s, z0.s\n" + "fmax z11.s, p5/M, z11.s, z0.s\n" + "fmax z15.s, p5/M, z15.s, z0.s\n" + "fmax z20.s, p5/M, z20.s, z0.s\n" + "fmax z21.s, p5/M, z21.s, z0.s\n" + "fmax z22.s, p5/M, z22.s, z0.s\n" + "fmax z16.s, p5/M, z16.s, z0.s\n" + "fmax z17.s, p5/M, z17.s, z0.s\n" + "fmax z18.s, p5/M, z18.s, z0.s\n" + "fmax z19.s, p5/M, z19.s, z0.s\n" "51:" // Height 4: No activation - "st1w { z7.s }, p4, [x28]\n" - "st1w { z12.s }, p3, [x28, #1, MUL VL]\n" - "st1w { z13.s }, p2, [x28, #2, MUL VL]\n" - "st1w { z14.s }, p1, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" - "st1w { z8.s }, p4, [x24]\n" - "st1w { z9.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z15.s }, p4, [x23]\n" - "st1w { z20.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z21.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z22.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x22]\n" - "st1w { z17.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x22, #3, MUL VL]\n" + "st1w { z7.s }, p4, [x9]\n" + "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "st1w { z8.s }, p4, [x25]\n" + "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z15.s }, p4, [x24]\n" + "st1w { z20.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z21.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z22.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" "52:" // Height 4: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" "bgt 41b\n" "b 80f\n" "53:" // Height 5 + "mov x12, %x[bias]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" - "mov x9, %x[bias]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x28, %x[output_ptr]\n" + "mov x9, %x[output_ptr]\n" "54:" // Height 5: Column loop - "mov x19, #0x0\n" - "whilelt p4.s, x19, x11\n" - "incw x19\n" - "whilelt p3.s, x19, x11\n" - "incw x19\n" - "whilelt p2.s, x19, x11\n" - "incw x19\n" - "whilelt p1.s, x19, x11\n" - "cbz x9, 55f\n" - "ld1w { z8.s }, p5/Z, [x9]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x11\n" + "incw x20\n" + "whilelt p3.s, x20, x11\n" + "incw x20\n" + "whilelt p2.s, x20, x11\n" + "incw x20\n" + "whilelt p1.s, x20, x11\n" + "cbz x12, 55f\n" + "ld1w { z8.s }, p5/Z, [x12]\n" + "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" "zip2 z12.d, z8.d, z8.d\n" - "ld1w { z9.s }, p5/Z, [x9, #1, MUL VL]\n" "zip1 z8.d, z8.d, z8.d\n" - "ld1w { z10.s }, p5/Z, [x9, #2, MUL VL]\n" - "mov z16.d, z8.d\n" - "ld1w { z11.s }, p5/Z, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" + "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" + "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" + "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" + "mov z16.d, z8.d\n" "mov z20.d, z12.d\n" "mov z17.d, z9.d\n" "mov z21.d, z13.d\n" @@ -1195,47 +1195,47 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "b 57f\n" "55:" // Height 5: no bias "tbz %x[flags], #0, 56f\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "ld1w { z9.s }, p4/Z, [x28]\n" - "add x24, x28, x19, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x28, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x28, #2, MUL VL]\n" - "add x23, x24, x19, LSL #2\n" - "ld1w { z16.s }, p1/Z, [x28, #3, MUL VL]\n" - "add x22, x23, x19, LSL #2\n" - "ld1w { z12.s }, p4/Z, [x24]\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x9, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" + "ld1w { z9.s }, p4/Z, [x9]\n" + "add x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" + "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x25]\n" "zip1 z8.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" - "add x21, x22, x19, LSL #2\n" "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" "zip1 z9.d, z10.d, z13.d\n" - "ld1w { z17.s }, p4/Z, [x23]\n" "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x24]\n" "zip1 z10.d, z11.d, z14.d\n" - "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" - "ld1w { z20.s }, p4/Z, [x22]\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x23]\n" "zip1 z16.d, z17.d, z20.d\n" - "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n" "zip2 z20.d, z17.d, z20.d\n" - "ld1w { z25.s }, p4/Z, [x21]\n" + "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" "zip1 z17.d, z18.d, z21.d\n" - "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n" "zip2 z21.d, z18.d, z21.d\n" - "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z25.s }, p4/Z, [x22]\n" "zip1 z18.d, z19.d, z22.d\n" - "ld1w { z6.s }, p1/Z, [x21, #3, MUL VL]\n" "zip2 z22.d, z19.d, z22.d\n" + "ld1w { z26.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z27.s }, p2/Z, [x22, #2, MUL VL]\n" "zip1 z19.d, z24.d, z23.d\n" "zip2 z23.d, z24.d, z23.d\n" + "ld1w { z6.s }, p1/Z, [x22, #3, MUL VL]\n" "zip1 z24.d, z25.d, z28.d\n" "zip2 z28.d, z25.d, z28.d\n" "zip1 z25.d, z26.d, z29.d\n" @@ -1271,72 +1271,72 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "mov z30.b, #0x0\n" "mov z31.b, #0x0\n" "57:" // Height 5: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "58:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w26, [x20, x27, LSL #0x2]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 59f\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, 60f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #1\n" - "add x24, x24, x19, LSL #1\n" - "add x23, x23, x19, LSL #1\n" - "add x22, x22, x19, LSL #1\n" - "add x21, x21, x19, LSL #1\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, 60f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #1\n" + "add x25, x25, x20, LSL #1\n" + "add x24, x24, x20, LSL #1\n" + "add x23, x23, x20, LSL #1\n" + "add x22, x22, x20, LSL #1\n" "b 60f\n" "59:" // Height 5: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #1\n" - "add x23, x24, x19, LSL #1\n" - "add x22, x23, x19, LSL #1\n" - "add x21, x22, x19, LSL #1\n" + "mov x26, %x[input_ptr]\n" + "add x25, x26, x20, LSL #1\n" + "add x24, x25, x20, LSL #1\n" + "add x23, x24, x20, LSL #1\n" + "add x22, x23, x20, LSL #1\n" "60:" // Height 5: input setup done - "cmp x26, #0x8\n" + "cmp x27, #0x8\n" "ble 62f\n" "61:" // Height 5: Multiply loop: Main loop head - "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" + "ld1rqh { z3.h }, p0/Z, [x24]\n" + "ld1rqh { z4.h }, p0/Z, [x23]\n" "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x23]\n" - "sub x26, x26, #0x8\n" "trn2 z1.d, z1.d, z2.d\n" - "ld1rqh { z4.h }, p0/Z, [x22]\n" - "cmp x26, #0x8\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - "ld1rqh { z5.h }, p0/Z, [x21]\n" - "add x25, x25, #0x10\n" + "ld1rqh { z5.h }, p0/Z, [x22]\n" "trn1 z2.d, z3.d, z4.d\n" - "add x24, x24, #0x10\n" "trn2 z3.d, z3.d, z4.d\n" - "add x23, x23, #0x10\n" + "ld1h { z7.h }, p5/Z, [x10]\n" "trn1 z4.d, z5.d, z6.d\n" - "add x22, x22, #0x10\n" "trn2 z5.d, z5.d, z6.d\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" - "add x21, x21, #0x10\n" + ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" ".inst 0x6467e498 // bfmmla z24.s, z4.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + "sub x27, x27, #0x8\n" ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + "cmp x27, #0x8\n" + "add x26, x26, #0x10\n" ".inst 0x6466e49c // bfmmla z28.s, z4.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" + "add x25, x25, #0x10\n" ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" ".inst 0x6467e499 // bfmmla z25.s, z4.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" + "add x24, x24, #0x10\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" + "add x23, x23, #0x10\n" + "add x22, x22, #0x10\n" ".inst 0x6466e49d // bfmmla z29.s, z4.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" @@ -1388,25 +1388,25 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e4bf // bfmmla z31.s, z5.h, z6.h\n" "bgt 61b\n" "62:" // Height 5: Multiply loop: Single iteration only - "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" + "ld1rqh { z3.h }, p0/Z, [x24]\n" + "ld1rqh { z4.h }, p0/Z, [x23]\n" "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x23]\n" - "subs x26, x26, #0x4\n" "trn2 z1.d, z1.d, z2.d\n" - "ld1rqh { z4.h }, p0/Z, [x22]\n" - "ld1rqh { z5.h }, p0/Z, [x21]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" + "ld1rqh { z5.h }, p0/Z, [x22]\n" "trn1 z2.d, z3.d, z4.d\n" "trn2 z3.d, z3.d, z4.d\n" + "ld1h { z7.h }, p5/Z, [x10]\n" "trn1 z4.d, z5.d, z6.d\n" "trn2 z5.d, z5.d, z6.d\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" ".inst 0x6467e498 // bfmmla z24.s, z4.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + "subs x27, x27, #0x4\n" ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" ".inst 0x6466e49c // bfmmla z28.s, z4.h, z6.h\n" @@ -1427,8 +1427,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" ".inst 0x6466e49e // bfmmla z30.s, z4.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" - "addvl x10, x10, #8\n" ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" + "addvl x10, x10, #8\n" ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" ".inst 0x6467e49b // bfmmla z27.s, z4.h, z7.h\n" ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" @@ -1436,28 +1436,28 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e49f // bfmmla z31.s, z4.h, z6.h\n" "ble 63f\n" "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" ".inst 0x6467e4b8 // bfmmla z24.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" ".inst 0x6466e4bc // bfmmla z28.s, z5.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" ".inst 0x6467e4b9 // bfmmla z25.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" ".inst 0x6466e4bd // bfmmla z29.s, z5.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" ".inst 0x6467e4ba // bfmmla z26.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" ".inst 0x6466e4be // bfmmla z30.s, z5.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" @@ -1469,21 +1469,21 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" ".inst 0x6466e4bf // bfmmla z31.s, z5.h, z6.h\n" "63:" // 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 58b\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x9, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "uzp1 z7.d, z8.d, z12.d\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x23, x24, x20, LSL #2\n" "uzp2 z8.d, z8.d, z12.d\n" - "add x24, x28, x19, LSL #2\n" "uzp1 z12.d, z9.d, z13.d\n" + "add x22, x23, x20, LSL #2\n" "uzp2 z9.d, z9.d, z13.d\n" - "add x23, x24, x19, LSL #2\n" "uzp1 z13.d, z10.d, z14.d\n" - "add x22, x23, x19, LSL #2\n" "uzp2 z10.d, z10.d, z14.d\n" - "add x21, x22, x19, LSL #2\n" "uzp1 z14.d, z11.d, z15.d\n" "uzp2 z11.d, z11.d, z15.d\n" "uzp1 z15.d, z16.d, z20.d\n" @@ -1499,109 +1499,109 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "uzp1 z26.d, z26.d, z30.d\n" "uzp1 z27.d, z27.d, z31.d\n" "tbz %x[flags], #1, 64f\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" - "fmin z7.s, p5/M, z7.s, z0.s\n" - "fmin z12.s, p5/M, z12.s, z0.s\n" - "fmin z13.s, p5/M, z13.s, z0.s\n" - "fmin z14.s, p5/M, z14.s, z0.s\n" - "fmin z8.s, p5/M, z8.s, z0.s\n" - "fmax z7.s, p5/M, z7.s, z1.s\n" - "fmax z12.s, p5/M, z12.s, z1.s\n" - "fmax z13.s, p5/M, z13.s, z1.s\n" - "fmax z14.s, p5/M, z14.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z0.s\n" - "fmin z10.s, p5/M, z10.s, z0.s\n" - "fmin z11.s, p5/M, z11.s, z0.s\n" - "fmin z15.s, p5/M, z15.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z1.s\n" - "fmax z10.s, p5/M, z10.s, z1.s\n" - "fmax z11.s, p5/M, z11.s, z1.s\n" - "fmax z15.s, p5/M, z15.s, z1.s\n" - "fmin z20.s, p5/M, z20.s, z0.s\n" - "fmin z21.s, p5/M, z21.s, z0.s\n" - "fmin z22.s, p5/M, z22.s, z0.s\n" - "fmin z16.s, p5/M, z16.s, z0.s\n" - "fmax z20.s, p5/M, z20.s, z1.s\n" - "fmax z21.s, p5/M, z21.s, z1.s\n" - "fmax z22.s, p5/M, z22.s, z1.s\n" - "fmax z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z0.s\n" - "fmin z18.s, p5/M, z18.s, z0.s\n" - "fmin z19.s, p5/M, z19.s, z0.s\n" - "fmin z24.s, p5/M, z24.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z1.s\n" - "fmax z18.s, p5/M, z18.s, z1.s\n" - "fmax z19.s, p5/M, z19.s, z1.s\n" - "fmax z24.s, p5/M, z24.s, z1.s\n" - "fmin z25.s, p5/M, z25.s, z0.s\n" - "fmin z26.s, p5/M, z26.s, z0.s\n" - "fmin z27.s, p5/M, z27.s, z0.s\n" - "fmax z25.s, p5/M, z25.s, z1.s\n" - "fmax z26.s, p5/M, z26.s, z1.s\n" - "fmax z27.s, p5/M, z27.s, z1.s\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" + "fmin z7.s, p5/M, z7.s, z1.s\n" + "fmin z12.s, p5/M, z12.s, z1.s\n" + "fmin z13.s, p5/M, z13.s, z1.s\n" + "fmin z14.s, p5/M, z14.s, z1.s\n" + "fmin z8.s, p5/M, z8.s, z1.s\n" + "fmin z9.s, p5/M, z9.s, z1.s\n" + "fmin z10.s, p5/M, z10.s, z1.s\n" + "fmin z11.s, p5/M, z11.s, z1.s\n" + "fmin z15.s, p5/M, z15.s, z1.s\n" + "fmin z20.s, p5/M, z20.s, z1.s\n" + "fmin z21.s, p5/M, z21.s, z1.s\n" + "fmin z22.s, p5/M, z22.s, z1.s\n" + "fmin z16.s, p5/M, z16.s, z1.s\n" + "fmin z17.s, p5/M, z17.s, z1.s\n" + "fmin z18.s, p5/M, z18.s, z1.s\n" + "fmin z19.s, p5/M, z19.s, z1.s\n" + "fmin z24.s, p5/M, z24.s, z1.s\n" + "fmin z25.s, p5/M, z25.s, z1.s\n" + "fmin z26.s, p5/M, z26.s, z1.s\n" + "fmin z27.s, p5/M, z27.s, z1.s\n" + "fmax z7.s, p5/M, z7.s, z0.s\n" + "fmax z12.s, p5/M, z12.s, z0.s\n" + "fmax z13.s, p5/M, z13.s, z0.s\n" + "fmax z14.s, p5/M, z14.s, z0.s\n" + "fmax z8.s, p5/M, z8.s, z0.s\n" + "fmax z9.s, p5/M, z9.s, z0.s\n" + "fmax z10.s, p5/M, z10.s, z0.s\n" + "fmax z11.s, p5/M, z11.s, z0.s\n" + "fmax z15.s, p5/M, z15.s, z0.s\n" + "fmax z20.s, p5/M, z20.s, z0.s\n" + "fmax z21.s, p5/M, z21.s, z0.s\n" + "fmax z22.s, p5/M, z22.s, z0.s\n" + "fmax z16.s, p5/M, z16.s, z0.s\n" + "fmax z17.s, p5/M, z17.s, z0.s\n" + "fmax z18.s, p5/M, z18.s, z0.s\n" + "fmax z19.s, p5/M, z19.s, z0.s\n" + "fmax z24.s, p5/M, z24.s, z0.s\n" + "fmax z25.s, p5/M, z25.s, z0.s\n" + "fmax z26.s, p5/M, z26.s, z0.s\n" + "fmax z27.s, p5/M, z27.s, z0.s\n" "64:" // Height 5: No activation - "st1w { z7.s }, p4, [x28]\n" - "st1w { z12.s }, p3, [x28, #1, MUL VL]\n" - "st1w { z13.s }, p2, [x28, #2, MUL VL]\n" - "st1w { z14.s }, p1, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" - "st1w { z8.s }, p4, [x24]\n" - "st1w { z9.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z15.s }, p4, [x23]\n" - "st1w { z20.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z21.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z22.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x22]\n" - "st1w { z17.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z19.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 { z7.s }, p4, [x9]\n" + "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "st1w { z8.s }, p4, [x25]\n" + "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z15.s }, p4, [x24]\n" + "st1w { z20.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z21.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z22.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 { 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" "65:" // Height 5: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" "bgt 54b\n" "b 80f\n" "66:" // Height 6 + "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" + "mov x20, #0x18\n" + "mov x12, %x[bias]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" - "mov x9, %x[bias]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x28, %x[output_ptr]\n" - "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "mov x19, #0x18\n" - "madd %x[output_ptr], x20, x19, %x[output_ptr]\n" + "mov x9, %x[output_ptr]\n" + "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" "67:" // Height 6: Column loop - "mov x19, #0x0\n" - "whilelt p4.s, x19, x11\n" - "incw x19\n" - "whilelt p3.s, x19, x11\n" - "incw x19\n" - "whilelt p2.s, x19, x11\n" - "incw x19\n" - "whilelt p1.s, x19, x11\n" - "cbz x9, 68f\n" - "ld1w { z8.s }, p5/Z, [x9]\n" + "mov x20, #0x0\n" + "whilelt p4.s, x20, x11\n" + "incw x20\n" + "whilelt p3.s, x20, x11\n" + "incw x20\n" + "whilelt p2.s, x20, x11\n" + "incw x20\n" + "whilelt p1.s, x20, x11\n" + "cbz x12, 68f\n" + "ld1w { z8.s }, p5/Z, [x12]\n" + "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" "zip2 z12.d, z8.d, z8.d\n" - "ld1w { z9.s }, p5/Z, [x9, #1, MUL VL]\n" "zip1 z8.d, z8.d, z8.d\n" - "ld1w { z10.s }, p5/Z, [x9, #2, MUL VL]\n" - "mov z16.d, z8.d\n" - "ld1w { z11.s }, p5/Z, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" + "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" + "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" + "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" + "mov z16.d, z8.d\n" "mov z20.d, z12.d\n" "mov z17.d, z9.d\n" "mov z21.d, z13.d\n" @@ -1620,55 +1620,55 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "b 70f\n" "68:" // Height 6: no bias "tbz %x[flags], #0, 69f\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "ld1w { z9.s }, p4/Z, [x28]\n" - "add x24, x28, x19, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x28, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x28, #2, MUL VL]\n" - "add x23, x24, x19, LSL #2\n" - "ld1w { z16.s }, p1/Z, [x28, #3, MUL VL]\n" - "add x22, x23, x19, LSL #2\n" - "ld1w { z12.s }, p4/Z, [x24]\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x9, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" + "ld1w { z9.s }, p4/Z, [x9]\n" + "add x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" + "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "add x21, x22, x20, LSL #2\n" + "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x25]\n" "zip1 z8.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" - "add x21, x22, x19, LSL #2\n" + "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "add x20, x21, x19, LSL #2\n" "zip1 z9.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x24]\n" "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z17.s }, p4/Z, [x23]\n" "zip1 z10.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x23]\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z20.s }, p4/Z, [x22]\n" - "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n" "zip1 z16.d, z17.d, z20.d\n" - "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" "zip2 z20.d, z17.d, z20.d\n" - "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n" "zip1 z17.d, z18.d, z21.d\n" - "ld1w { z25.s }, p4/Z, [x21]\n" + "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z25.s }, p4/Z, [x22]\n" "zip2 z21.d, z18.d, z21.d\n" - "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n" "zip1 z18.d, z19.d, z22.d\n" - "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z26.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z27.s }, p2/Z, [x22, #2, MUL VL]\n" "zip2 z22.d, z19.d, z22.d\n" - "ld1w { z6.s }, p1/Z, [x21, #3, MUL VL]\n" "zip1 z19.d, z24.d, z23.d\n" - "ld1w { z28.s }, p4/Z, [x20]\n" + "ld1w { z6.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z28.s }, p4/Z, [x21]\n" "zip2 z23.d, z24.d, z23.d\n" - "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n" - "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z24.d, z25.d, z28.d\n" - "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n" + "ld1w { z29.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z30.s }, p2/Z, [x21, #2, MUL VL]\n" "zip2 z28.d, z25.d, z28.d\n" "zip1 z25.d, z26.d, z29.d\n" + "ld1w { z31.s }, p1/Z, [x21, #3, MUL VL]\n" "zip2 z29.d, z26.d, z29.d\n" "zip1 z26.d, z27.d, z30.d\n" "zip2 z30.d, z27.d, z30.d\n" @@ -1701,80 +1701,80 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "mov z30.b, #0x0\n" "mov z31.b, #0x0\n" "70:" // Height 6: setup done - "mov x27, #0x0\n" + "mov x28, #0x0\n" "71:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w26, [x20, x27, LSL #0x2]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 72f\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, 73f\n" - "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x25, x25, x19, LSL #1\n" - "add x24, x24, x19, LSL #1\n" - "add x23, x23, x19, LSL #1\n" - "add x22, x22, x19, LSL #1\n" - "add x21, x21, x19, LSL #1\n" - "add x20, x20, x19, LSL #1\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, 73f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #1\n" + "add x25, x25, x20, LSL #1\n" + "add x24, x24, x20, LSL #1\n" + "add x23, x23, x20, LSL #1\n" + "add x22, x22, x20, LSL #1\n" + "add x21, x21, x20, LSL #1\n" "b 73f\n" "72:" // Height 6: setup direct input - "mov x25, %x[input_ptr]\n" - "add x24, x25, x19, LSL #1\n" - "add x23, x24, x19, LSL #1\n" - "add x22, x23, x19, LSL #1\n" - "add x21, x22, x19, LSL #1\n" - "add x20, x21, x19, LSL #1\n" + "mov x26, %x[input_ptr]\n" + "add x25, x26, x20, LSL #1\n" + "add x24, x25, x20, LSL #1\n" + "add x23, x24, x20, LSL #1\n" + "add x22, x23, x20, LSL #1\n" + "add x21, x22, x20, LSL #1\n" "73:" // Height 6: input setup done - "cmp x26, #0x8\n" + "cmp x27, #0x8\n" "ble 75f\n" "74:" // Height 6: Multiply loop: Main loop head - "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" - "sub x26, x26, #0x8\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x23]\n" - "cmp x26, #0x8\n" + "ld1rqh { z3.h }, p0/Z, [x24]\n" + "ld1rqh { z4.h }, p0/Z, [x23]\n" "trn2 z1.d, z1.d, z2.d\n" - "ld1rqh { z4.h }, p0/Z, [x22]\n" - "add x25, x25, #0x10\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - "ld1rqh { z5.h }, p0/Z, [x21]\n" - "add x24, x24, #0x10\n" "trn1 z2.d, z3.d, z4.d\n" - "ld1rqh { z6.h }, p0/Z, [x20]\n" - "add x23, x23, #0x10\n" + "ld1rqh { z5.h }, p0/Z, [x22]\n" + "ld1rqh { z6.h }, p0/Z, [x21]\n" "trn2 z3.d, z3.d, z4.d\n" - "add x22, x22, #0x10\n" - "add x21, x21, #0x10\n" - ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" - "add x20, x20, #0x10\n" "trn1 z4.d, z5.d, z6.d\n" "trn2 z5.d, z5.d, z6.d\n" + "ld1h { z7.h }, p5/Z, [x10]\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" + ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" ".inst 0x6467e498 // bfmmla z24.s, z4.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + "sub x27, x27, #0x8\n" ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" + "cmp x27, #0x8\n" + "add x26, x26, #0x10\n" ".inst 0x6466e49c // bfmmla z28.s, z4.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n" + "add x25, x25, #0x10\n" ".inst 0x6467e451 // bfmmla z17.s, z2.h, z7.h\n" ".inst 0x6467e499 // bfmmla z25.s, z4.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" + "add x24, x24, #0x10\n" ".inst 0x6466e40d // bfmmla z13.s, z0.h, z6.h\n" ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n" + "add x23, x23, #0x10\n" + "add x22, x22, #0x10\n" ".inst 0x6466e49d // bfmmla z29.s, z4.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e40a // bfmmla z10.s, z0.h, z7.h\n" + "add x21, x21, #0x10\n" ".inst 0x6467e452 // bfmmla z18.s, z2.h, z7.h\n" ".inst 0x6467e49a // bfmmla z26.s, z4.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" @@ -1823,26 +1823,26 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e4bf // bfmmla z31.s, z5.h, z6.h\n" "bgt 74b\n" "75:" // Height 6: Multiply loop: Single iteration only - "ld1h { z7.h }, p5/Z, [x10]\n" - "whilelt p0.h, XZR, x26\n" - "subs x26, x26, #0x4\n" - "ld1rqh { z1.h }, p0/Z, [x25]\n" - "ld1rqh { z2.h }, p0/Z, [x24]\n" + "whilelt p0.h, XZR, x27\n" + "ld1rqh { z1.h }, p0/Z, [x26]\n" + "ld1rqh { z2.h }, p0/Z, [x25]\n" "trn1 z0.d, z1.d, z2.d\n" - "ld1rqh { z3.h }, p0/Z, [x23]\n" + "ld1rqh { z3.h }, p0/Z, [x24]\n" + "ld1rqh { z4.h }, p0/Z, [x23]\n" "trn2 z1.d, z1.d, z2.d\n" - "ld1rqh { z4.h }, p0/Z, [x22]\n" - "ld1rqh { z5.h }, p0/Z, [x21]\n" - ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" - "ld1rqh { z6.h }, p0/Z, [x20]\n" "trn1 z2.d, z3.d, z4.d\n" + "ld1rqh { z5.h }, p0/Z, [x22]\n" + "ld1rqh { z6.h }, p0/Z, [x21]\n" "trn2 z3.d, z3.d, z4.d\n" "trn1 z4.d, z5.d, z6.d\n" "trn2 z5.d, z5.d, z6.d\n" + "ld1h { z7.h }, p5/Z, [x10]\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e408 // bfmmla z8.s, z0.h, z7.h\n" ".inst 0x6467e450 // bfmmla z16.s, z2.h, z7.h\n" ".inst 0x6467e498 // bfmmla z24.s, z4.h, z7.h\n" "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" + "subs x27, x27, #0x4\n" ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n" ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n" ".inst 0x6466e49c // bfmmla z28.s, z4.h, z6.h\n" @@ -1863,8 +1863,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n" ".inst 0x6466e49e // bfmmla z30.s, z4.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" - "addvl x10, x10, #8\n" ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n" + "addvl x10, x10, #8\n" ".inst 0x6467e453 // bfmmla z19.s, z2.h, z7.h\n" ".inst 0x6467e49b // bfmmla z27.s, z4.h, z7.h\n" ".inst 0x6466e40f // bfmmla z15.s, z0.h, z6.h\n" @@ -1872,28 +1872,28 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e49f // bfmmla z31.s, z4.h, z6.h\n" "ble 76f\n" "ld1h { z7.h }, p5/Z, [x10]\n" - ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6467e428 // bfmmla z8.s, z1.h, z7.h\n" ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" ".inst 0x6467e4b8 // bfmmla z24.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e42c // bfmmla z12.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n" ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" ".inst 0x6466e4bc // bfmmla z28.s, z5.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n" ".inst 0x6467e429 // bfmmla z9.s, z1.h, z7.h\n" ".inst 0x6467e471 // bfmmla z17.s, z3.h, z7.h\n" ".inst 0x6467e4b9 // bfmmla z25.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e42d // bfmmla z13.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #4, MUL VL]\n" ".inst 0x6466e475 // bfmmla z21.s, z3.h, z6.h\n" ".inst 0x6466e4bd // bfmmla z29.s, z5.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #5, MUL VL]\n" ".inst 0x6467e42a // bfmmla z10.s, z1.h, z7.h\n" ".inst 0x6467e472 // bfmmla z18.s, z3.h, z7.h\n" ".inst 0x6467e4ba // bfmmla z26.s, z5.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n" + "ld1h { z7.h }, p5/Z, [x10, #6, MUL VL]\n" ".inst 0x6466e476 // bfmmla z22.s, z3.h, z6.h\n" ".inst 0x6466e4be // bfmmla z30.s, z5.h, z6.h\n" "ld1h { z6.h }, p5/Z, [x10, #7, MUL VL]\n" @@ -1905,23 +1905,23 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( ".inst 0x6466e477 // bfmmla z23.s, z3.h, z6.h\n" ".inst 0x6466e4bf // bfmmla z31.s, z5.h, z6.h\n" "76:" // 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 71b\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x25, x9, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "uzp1 z7.d, z8.d, z12.d\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" "uzp2 z8.d, z8.d, z12.d\n" - "add x24, x28, x19, LSL #2\n" "uzp1 z12.d, z9.d, z13.d\n" "uzp2 z9.d, z9.d, z13.d\n" - "add x23, x24, x19, LSL #2\n" "uzp1 z13.d, z10.d, z14.d\n" - "add x22, x23, x19, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "uzp2 z10.d, z10.d, z14.d\n" - "add x21, x22, x19, LSL #2\n" "uzp1 z14.d, z11.d, z15.d\n" - "add x20, x21, x19, LSL #2\n" "uzp2 z11.d, z11.d, z15.d\n" "uzp1 z15.d, z16.d, z20.d\n" "uzp2 z16.d, z16.d, z20.d\n" @@ -1940,104 +1940,103 @@ void sve_hybrid_bf16fp32_mmla_6x4VL ( "uzp1 z30.d, z27.d, z31.d\n" "uzp2 z27.d, z27.d, z31.d\n" "tbz %x[flags], #1, 77f\n" - "add x19, %x[args_ptr], %[offset_min]\n" - "ld1rw { z1.s }, p5/Z, [x19]\n" - "add x19, %x[args_ptr], %[offset_max]\n" - "ld1rw { z0.s }, p5/Z, [x19]\n" - "fmin z7.s, p5/M, z7.s, z0.s\n" - "fmin z12.s, p5/M, z12.s, z0.s\n" - "fmin z13.s, p5/M, z13.s, z0.s\n" - "fmin z14.s, p5/M, z14.s, z0.s\n" - "fmin z8.s, p5/M, z8.s, z0.s\n" - "fmax z7.s, p5/M, z7.s, z1.s\n" - "fmax z12.s, p5/M, z12.s, z1.s\n" - "fmax z13.s, p5/M, z13.s, z1.s\n" - "fmax z14.s, p5/M, z14.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z0.s\n" - "fmin z10.s, p5/M, z10.s, z0.s\n" - "fmin z11.s, p5/M, z11.s, z0.s\n" - "fmin z15.s, p5/M, z15.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z1.s\n" - "fmax z10.s, p5/M, z10.s, z1.s\n" - "fmax z11.s, p5/M, z11.s, z1.s\n" - "fmax z15.s, p5/M, z15.s, z1.s\n" - "fmin z20.s, p5/M, z20.s, z0.s\n" - "fmin z21.s, p5/M, z21.s, z0.s\n" - "fmin z22.s, p5/M, z22.s, z0.s\n" - "fmin z16.s, p5/M, z16.s, z0.s\n" - "fmax z20.s, p5/M, z20.s, z1.s\n" - "fmax z21.s, p5/M, z21.s, z1.s\n" - "fmax z22.s, p5/M, z22.s, z1.s\n" - "fmax z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z0.s\n" - "fmin z18.s, p5/M, z18.s, z0.s\n" - "fmin z19.s, p5/M, z19.s, z0.s\n" - "fmin z23.s, p5/M, z23.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z1.s\n" - "fmax z18.s, p5/M, z18.s, z1.s\n" - "fmax z19.s, p5/M, z19.s, z1.s\n" - "fmax z23.s, p5/M, z23.s, z1.s\n" - "fmin z28.s, p5/M, z28.s, z0.s\n" - "fmin z29.s, p5/M, z29.s, z0.s\n" - "fmin z30.s, p5/M, z30.s, z0.s\n" - "fmin z24.s, p5/M, z24.s, z0.s\n" - "fmax z28.s, p5/M, z28.s, z1.s\n" - "fmax z29.s, p5/M, z29.s, z1.s\n" - "fmax z30.s, p5/M, z30.s, z1.s\n" - "fmax z24.s, p5/M, z24.s, z1.s\n" - "fmin z25.s, p5/M, z25.s, z0.s\n" - "fmin z26.s, p5/M, z26.s, z0.s\n" - "fmin z27.s, p5/M, z27.s, z0.s\n" - "fmax z25.s, p5/M, z25.s, z1.s\n" - "fmax z26.s, p5/M, z26.s, z1.s\n" - "fmax z27.s, p5/M, z27.s, z1.s\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x20]\n" + "fmin z7.s, p5/M, z7.s, z1.s\n" + "fmin z12.s, p5/M, z12.s, z1.s\n" + "fmin z13.s, p5/M, z13.s, z1.s\n" + "fmin z14.s, p5/M, z14.s, z1.s\n" + "fmin z8.s, p5/M, z8.s, z1.s\n" + "fmin z9.s, p5/M, z9.s, z1.s\n" + "fmin z10.s, p5/M, z10.s, z1.s\n" + "fmin z11.s, p5/M, z11.s, z1.s\n" + "fmin z15.s, p5/M, z15.s, z1.s\n" + "fmin z20.s, p5/M, z20.s, z1.s\n" + "fmin z21.s, p5/M, z21.s, z1.s\n" + "fmin z22.s, p5/M, z22.s, z1.s\n" + "fmin z16.s, p5/M, z16.s, z1.s\n" + "fmin z17.s, p5/M, z17.s, z1.s\n" + "fmin z18.s, p5/M, z18.s, z1.s\n" + "fmin z19.s, p5/M, z19.s, z1.s\n" + "fmin z23.s, p5/M, z23.s, z1.s\n" + "fmin z28.s, p5/M, z28.s, z1.s\n" + "fmin z29.s, p5/M, z29.s, z1.s\n" + "fmin z30.s, p5/M, z30.s, z1.s\n" + "fmin z24.s, p5/M, z24.s, z1.s\n" + "fmin z25.s, p5/M, z25.s, z1.s\n" + "fmin z26.s, p5/M, z26.s, z1.s\n" + "fmin z27.s, p5/M, z27.s, z1.s\n" + "fmax z7.s, p5/M, z7.s, z0.s\n" + "fmax z12.s, p5/M, z12.s, z0.s\n" + "fmax z13.s, p5/M, z13.s, z0.s\n" + "fmax z14.s, p5/M, z14.s, z0.s\n" + "fmax z8.s, p5/M, z8.s, z0.s\n" + "fmax z9.s, p5/M, z9.s, z0.s\n" + "fmax z10.s, p5/M, z10.s, z0.s\n" + "fmax z11.s, p5/M, z11.s, z0.s\n" + "fmax z15.s, p5/M, z15.s, z0.s\n" + "fmax z20.s, p5/M, z20.s, z0.s\n" + "fmax z21.s, p5/M, z21.s, z0.s\n" + "fmax z22.s, p5/M, z22.s, z0.s\n" + "fmax z16.s, p5/M, z16.s, z0.s\n" + "fmax z17.s, p5/M, z17.s, z0.s\n" + "fmax z18.s, p5/M, z18.s, z0.s\n" + "fmax z19.s, p5/M, z19.s, z0.s\n" + "fmax z23.s, p5/M, z23.s, z0.s\n" + "fmax z28.s, p5/M, z28.s, z0.s\n" + "fmax z29.s, p5/M, z29.s, z0.s\n" + "fmax z30.s, p5/M, z30.s, z0.s\n" + "fmax z24.s, p5/M, z24.s, z0.s\n" + "fmax z25.s, p5/M, z25.s, z0.s\n" + "fmax z26.s, p5/M, z26.s, z0.s\n" + "fmax z27.s, p5/M, z27.s, z0.s\n" "77:" // Height 6: No activation - "st1w { z7.s }, p4, [x28]\n" - "st1w { z12.s }, p3, [x28, #1, MUL VL]\n" - "st1w { z13.s }, p2, [x28, #2, MUL VL]\n" - "st1w { z14.s }, p1, [x28, #3, MUL VL]\n" - "addvl x28, x28, #4\n" - "st1w { z8.s }, p4, [x24]\n" - "st1w { z9.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z15.s }, p4, [x23]\n" - "st1w { z20.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z21.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z22.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x22]\n" - "st1w { z17.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x22, #3, MUL VL]\n" - "st1w { z23.s }, p4, [x21]\n" - "st1w { z28.s }, p3, [x21, #1, MUL VL]\n" - "st1w { z29.s }, p2, [x21, #2, MUL VL]\n" - "st1w { z30.s }, p1, [x21, #3, MUL VL]\n" - "st1w { z24.s }, p4, [x20]\n" - "st1w { z25.s }, p3, [x20, #1, MUL VL]\n" - "st1w { z26.s }, p2, [x20, #2, MUL VL]\n" - "st1w { z27.s }, p1, [x20, #3, MUL VL]\n" + "st1w { z7.s }, p4, [x9]\n" + "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "st1w { z8.s }, p4, [x25]\n" + "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z15.s }, p4, [x24]\n" + "st1w { z20.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z21.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z22.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 { z23.s }, p4, [x22]\n" + "st1w { z28.s }, p3, [x22, #1, MUL VL]\n" + "st1w { z29.s }, p2, [x22, #2, MUL VL]\n" + "st1w { z30.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" "78:" // Height 6: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" "bgt 67b\n" "subs %x[M], %x[M], #0x6\n" "beq 80f\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 79f\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" "79:" // Update direct input - "mov x19, #0xc\n" - "madd %x[input_ptr], x19, x20, %x[input_ptr]\n" + "mov x20, #0xc\n" + "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "80:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "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", "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" ); } |