diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-07-16 16:16:43 +0100 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-07-22 02:25:50 +0000 |
commit | 4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2 (patch) | |
tree | 2f8362d33cdad4212f4b96995681c68184c759e1 /src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp | |
parent | 59fd7a722e5bc7e85309d6200bc37a772721a719 (diff) | |
download | ComputeLibrary-4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2.tar.gz |
Update GEMM assembly kernels
- Introduce Fp32 kernels with internal calculations in Bfloat16 when
fast_mode is enabled
- Improve kernel selection heuristics
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I68a9e7e862b6fd2721b46e0d7cc791091c4ab279
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5965
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp | 502 |
1 files changed, 212 insertions, 290 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp index 0f1937acc5..f8e4b89b95 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -10,310 +10,232 @@ * 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 - -#include "../../asmlib.hpp" +#include <cstddef> namespace arm_gemm { -void sve_interleaved_fp16_mla_8x3VL(const __fp16 *Apanel, const __fp16 *Bpanel, __fp16 *Cpanel, int ablocks, int bblocks, int K) { - const __fp16 *a_ptr = Apanel; - __fp16 *c_ptr = Cpanel; - - const long loops_count = (K / 2) - 1; - const long tails_count = K % 2; +void sve_interleaved_fp16_mla_8x3VL( + const __fp16 *Apanel, const __fp16 *Bpanel, + __fp16 *Cpanel, int ablocks, int bblocks, int K) { - for (int yb=0; yb<ablocks; yb++) { - const __fp16 *a_ptr0 = a_ptr; - const __fp16 *b_ptr = Bpanel; + struct KernelArgs { + size_t bblocks = {}; + size_t K = {}; + const __fp16 *Bpanel = {}; + } ka; - for (int xb=0; xb<bblocks; xb++) { - a_ptr = a_ptr0; - long loops = loops_count; - long tails = tails_count; + ka.bblocks = bblocks; + ka.K = (K/1) - 1; + ka.Bpanel = Bpanel; - __asm __volatile ( - "mov z8.h, #0\n" - "ptrue p0.h\n" - "mov z9.h, #0\n" - "mov z10.h, #0\n" - "mov z11.h, #0\n" - "ld1rqh z0.h, p0/z, [%[a_ptr]]\n" - "mov z12.h, #0\n" - "ld1h z2.h, p0/z, [%[b_ptr]]\n" - "mov z13.h, #0\n" - "ld1h z3.h, p0/z, [%[b_ptr], #1, MUL VL]\n" - "mov z14.h, #0\n" - "ld1h z4.h, p0/z, [%[b_ptr], #2, MUL VL]\n" - "mov z15.h, #0\n" - "ld1h z5.h, p0/z, [%[b_ptr], #3, MUL VL]\n" - "mov z16.h, #0\n" - "ld1h z6.h, p0/z, [%[b_ptr], #4, MUL VL]\n" - "mov z17.h, #0\n" - "add %[a_ptr], %[a_ptr], #0x20\n" - "mov z18.h, #0\n" - "addvl %[b_ptr], %[b_ptr], #6\n" - "mov z19.h, #0\n" - "mov z20.h, #0\n" - "mov z21.h, #0\n" - "mov z22.h, #0\n" - "mov z23.h, #0\n" - "mov z24.h, #0\n" - "mov z25.h, #0\n" - "mov z26.h, #0\n" - "mov z27.h, #0\n" - "mov z28.h, #0\n" - "mov z29.h, #0\n" - "mov z30.h, #0\n" - "mov z31.h, #0\n" - "cbz %[loops], 1f\n" - "2:\n" - "fmla z8.h, z2.h, z0.h[0]\n" - "ld1h z7.h, p0/z, [%[b_ptr], #-1, MUL VL]\n" - "fmla z9.h, z2.h, z0.h[1]\n" - "ld1rqh z1.h, p0/z, [%[a_ptr], #-0x10]\n" - "fmla z10.h, z2.h, z0.h[2]\n" - "subs %[loops], %[loops], #0x1\n" - "fmla z11.h, z2.h, z0.h[3]\n" - "fmla z12.h, z2.h, z0.h[4]\n" - "fmla z13.h, z2.h, z0.h[5]\n" - "fmla z14.h, z2.h, z0.h[6]\n" - "fmla z15.h, z2.h, z0.h[7]\n" - "ld1h z2.h, p0/z, [%[b_ptr]]\n" - "fmla z16.h, z3.h, z0.h[0]\n" - "fmla z17.h, z3.h, z0.h[1]\n" - "fmla z18.h, z3.h, z0.h[2]\n" - "fmla z19.h, z3.h, z0.h[3]\n" - "fmla z20.h, z3.h, z0.h[4]\n" - "fmla z21.h, z3.h, z0.h[5]\n" - "fmla z22.h, z3.h, z0.h[6]\n" - "fmla z23.h, z3.h, z0.h[7]\n" - "ld1h z3.h, p0/z, [%[b_ptr], #1, MUL VL]\n" - "fmla z24.h, z4.h, z0.h[0]\n" - "fmla z25.h, z4.h, z0.h[1]\n" - "fmla z26.h, z4.h, z0.h[2]\n" - "fmla z27.h, z4.h, z0.h[3]\n" - "fmla z28.h, z4.h, z0.h[4]\n" - "fmla z29.h, z4.h, z0.h[5]\n" - "fmla z30.h, z4.h, z0.h[6]\n" - "fmla z31.h, z4.h, z0.h[7]\n" - "ld1h z4.h, p0/z, [%[b_ptr], #2, MUL VL]\n" - "fmla z8.h, z5.h, z1.h[0]\n" - "ld1rqh z0.h, p0/z, [%[a_ptr]]\n" - "fmla z9.h, z5.h, z1.h[1]\n" - "add %[a_ptr], %[a_ptr], #0x20\n" - "fmla z10.h, z5.h, z1.h[2]\n" - "addvl %[b_ptr], %[b_ptr], #6\n" - "fmla z11.h, z5.h, z1.h[3]\n" - "fmla z12.h, z5.h, z1.h[4]\n" - "fmla z13.h, z5.h, z1.h[5]\n" - "fmla z14.h, z5.h, z1.h[6]\n" - "fmla z15.h, z5.h, z1.h[7]\n" - "ld1h z5.h, p0/z, [%[b_ptr], #-3, MUL VL]\n" - "fmla z16.h, z6.h, z1.h[0]\n" - "fmla z17.h, z6.h, z1.h[1]\n" - "fmla z18.h, z6.h, z1.h[2]\n" - "fmla z19.h, z6.h, z1.h[3]\n" - "fmla z20.h, z6.h, z1.h[4]\n" - "fmla z21.h, z6.h, z1.h[5]\n" - "fmla z22.h, z6.h, z1.h[6]\n" - "fmla z23.h, z6.h, z1.h[7]\n" - "ld1h z6.h, p0/z, [%[b_ptr], #-2, MUL VL]\n" - "fmla z24.h, z7.h, z1.h[0]\n" - "fmla z25.h, z7.h, z1.h[1]\n" - "fmla z26.h, z7.h, z1.h[2]\n" - "fmla z27.h, z7.h, z1.h[3]\n" - "fmla z28.h, z7.h, z1.h[4]\n" - "fmla z29.h, z7.h, z1.h[5]\n" - "fmla z30.h, z7.h, z1.h[6]\n" - "fmla z31.h, z7.h, z1.h[7]\n" - "b.ne 2b\n" - "1:\n" - "cbz %[tails], 3f\n" - "fmla z8.h, z2.h, z0.h[0]\n" - "ld1h z7.h, p0/z, [%[b_ptr], #-1, MUL VL]\n" - "fmla z9.h, z2.h, z0.h[1]\n" - "ld1rqh z1.h, p0/z, [%[a_ptr], #-0x10]\n" - "fmla z10.h, z2.h, z0.h[2]\n" - "fmla z11.h, z2.h, z0.h[3]\n" - "fmla z12.h, z2.h, z0.h[4]\n" - "fmla z13.h, z2.h, z0.h[5]\n" - "fmla z14.h, z2.h, z0.h[6]\n" - "fmla z15.h, z2.h, z0.h[7]\n" - "ld1h z2.h, p0/z, [%[b_ptr]]\n" - "fmla z16.h, z3.h, z0.h[0]\n" - "fmla z17.h, z3.h, z0.h[1]\n" - "fmla z18.h, z3.h, z0.h[2]\n" - "fmla z19.h, z3.h, z0.h[3]\n" - "fmla z20.h, z3.h, z0.h[4]\n" - "fmla z21.h, z3.h, z0.h[5]\n" - "fmla z22.h, z3.h, z0.h[6]\n" - "fmla z23.h, z3.h, z0.h[7]\n" - "ld1h z3.h, p0/z, [%[b_ptr], #1, MUL VL]\n" - "fmla z24.h, z4.h, z0.h[0]\n" - "fmla z25.h, z4.h, z0.h[1]\n" - "fmla z26.h, z4.h, z0.h[2]\n" - "fmla z27.h, z4.h, z0.h[3]\n" - "fmla z28.h, z4.h, z0.h[4]\n" - "fmla z29.h, z4.h, z0.h[5]\n" - "fmla z30.h, z4.h, z0.h[6]\n" - "fmla z31.h, z4.h, z0.h[7]\n" - "ld1h z4.h, p0/z, [%[b_ptr], #2, MUL VL]\n" - "fmla z8.h, z5.h, z1.h[0]\n" - "ld1rqh z0.h, p0/z, [%[a_ptr]]\n" - "fmla z9.h, z5.h, z1.h[1]\n" - "add %[a_ptr], %[a_ptr], #0x10\n" - "fmla z10.h, z5.h, z1.h[2]\n" - "addvl %[b_ptr], %[b_ptr], #3\n" - "fmla z11.h, z5.h, z1.h[3]\n" - "fmla z12.h, z5.h, z1.h[4]\n" - "fmla z13.h, z5.h, z1.h[5]\n" - "fmla z14.h, z5.h, z1.h[6]\n" - "fmla z15.h, z5.h, z1.h[7]\n" - "fmla z16.h, z6.h, z1.h[0]\n" - "fmla z17.h, z6.h, z1.h[1]\n" - "fmla z18.h, z6.h, z1.h[2]\n" - "fmla z19.h, z6.h, z1.h[3]\n" - "fmla z20.h, z6.h, z1.h[4]\n" - "fmla z21.h, z6.h, z1.h[5]\n" - "fmla z22.h, z6.h, z1.h[6]\n" - "fmla z23.h, z6.h, z1.h[7]\n" - "fmla z24.h, z7.h, z1.h[0]\n" - "fmla z25.h, z7.h, z1.h[1]\n" - "fmla z26.h, z7.h, z1.h[2]\n" - "fmla z27.h, z7.h, z1.h[3]\n" - "fmla z28.h, z7.h, z1.h[4]\n" - "fmla z29.h, z7.h, z1.h[5]\n" - "fmla z30.h, z7.h, z1.h[6]\n" - "fmla z31.h, z7.h, z1.h[7]\n" - "fmla z8.h, z2.h, z0.h[0]\n" - "fmla z9.h, z2.h, z0.h[1]\n" - "fmla z10.h, z2.h, z0.h[2]\n" - "fmla z11.h, z2.h, z0.h[3]\n" - "st1h z8.h, p0, [%[c_ptr]]\n" - "fmla z12.h, z2.h, z0.h[4]\n" - "fmla z13.h, z2.h, z0.h[5]\n" - "fmla z14.h, z2.h, z0.h[6]\n" - "fmla z15.h, z2.h, z0.h[7]\n" - "fmla z16.h, z3.h, z0.h[0]\n" - "fmla z17.h, z3.h, z0.h[1]\n" - "fmla z18.h, z3.h, z0.h[2]\n" - "fmla z19.h, z3.h, z0.h[3]\n" - "st1h z16.h, p0, [%[c_ptr], #1, MUL VL]\n" - "fmla z20.h, z3.h, z0.h[4]\n" - "fmla z21.h, z3.h, z0.h[5]\n" - "fmla z22.h, z3.h, z0.h[6]\n" - "fmla z23.h, z3.h, z0.h[7]\n" - "fmla z24.h, z4.h, z0.h[0]\n" - "fmla z25.h, z4.h, z0.h[1]\n" - "fmla z26.h, z4.h, z0.h[2]\n" - "fmla z27.h, z4.h, z0.h[3]\n" - "st1h z24.h, p0, [%[c_ptr], #2, MUL VL]\n" - "fmla z28.h, z4.h, z0.h[4]\n" - "fmla z29.h, z4.h, z0.h[5]\n" - "fmla z30.h, z4.h, z0.h[6]\n" - "st1h z9.h, p0, [%[c_ptr], #3, MUL VL]\n" - "fmla z31.h, z4.h, z0.h[7]\n" - "b 4f\n" - "3:\n" - "fmla z8.h, z2.h, z0.h[0]\n" - "ld1h z7.h, p0/z, [%[b_ptr], #-1, MUL VL]\n" - "fmla z9.h, z2.h, z0.h[1]\n" - "ld1rqh z1.h, p0/z, [%[a_ptr], #-0x10]\n" - "fmla z10.h, z2.h, z0.h[2]\n" - "fmla z11.h, z2.h, z0.h[3]\n" - "fmla z12.h, z2.h, z0.h[4]\n" - "fmla z13.h, z2.h, z0.h[5]\n" - "fmla z14.h, z2.h, z0.h[6]\n" - "fmla z15.h, z2.h, z0.h[7]\n" - "fmla z16.h, z3.h, z0.h[0]\n" - "fmla z17.h, z3.h, z0.h[1]\n" - "fmla z18.h, z3.h, z0.h[2]\n" - "fmla z19.h, z3.h, z0.h[3]\n" - "fmla z20.h, z3.h, z0.h[4]\n" - "fmla z21.h, z3.h, z0.h[5]\n" - "fmla z22.h, z3.h, z0.h[6]\n" - "fmla z23.h, z3.h, z0.h[7]\n" - "fmla z24.h, z4.h, z0.h[0]\n" - "fmla z25.h, z4.h, z0.h[1]\n" - "fmla z26.h, z4.h, z0.h[2]\n" - "fmla z27.h, z4.h, z0.h[3]\n" - "fmla z28.h, z4.h, z0.h[4]\n" - "fmla z29.h, z4.h, z0.h[5]\n" - "fmla z30.h, z4.h, z0.h[6]\n" - "fmla z31.h, z4.h, z0.h[7]\n" - "fmla z8.h, z5.h, z1.h[0]\n" - "fmla z9.h, z5.h, z1.h[1]\n" - "fmla z10.h, z5.h, z1.h[2]\n" - "fmla z11.h, z5.h, z1.h[3]\n" - "st1h z8.h, p0, [%[c_ptr]]\n" - "fmla z12.h, z5.h, z1.h[4]\n" - "fmla z13.h, z5.h, z1.h[5]\n" - "fmla z14.h, z5.h, z1.h[6]\n" - "fmla z15.h, z5.h, z1.h[7]\n" - "fmla z16.h, z6.h, z1.h[0]\n" - "fmla z17.h, z6.h, z1.h[1]\n" - "fmla z18.h, z6.h, z1.h[2]\n" - "fmla z19.h, z6.h, z1.h[3]\n" - "st1h z16.h, p0, [%[c_ptr], #1, MUL VL]\n" - "fmla z20.h, z6.h, z1.h[4]\n" - "fmla z21.h, z6.h, z1.h[5]\n" - "fmla z22.h, z6.h, z1.h[6]\n" - "fmla z23.h, z6.h, z1.h[7]\n" - "fmla z24.h, z7.h, z1.h[0]\n" - "fmla z25.h, z7.h, z1.h[1]\n" - "fmla z26.h, z7.h, z1.h[2]\n" - "fmla z27.h, z7.h, z1.h[3]\n" - "st1h z24.h, p0, [%[c_ptr], #2, MUL VL]\n" - "fmla z28.h, z7.h, z1.h[4]\n" - "fmla z29.h, z7.h, z1.h[5]\n" - "fmla z30.h, z7.h, z1.h[6]\n" - "st1h z9.h, p0, [%[c_ptr], #3, MUL VL]\n" - "fmla z31.h, z7.h, z1.h[7]\n" - "4:\n" - "st1h z17.h, p0, [%[c_ptr], #4, MUL VL]\n" - "st1h z25.h, p0, [%[c_ptr], #5, MUL VL]\n" - "st1h z10.h, p0, [%[c_ptr], #6, MUL VL]\n" - "st1h z18.h, p0, [%[c_ptr], #7, MUL VL]\n" - "addvl %[c_ptr], %[c_ptr], #16\n" - "st1h z26.h, p0, [%[c_ptr], #-8, MUL VL]\n" - "st1h z11.h, p0, [%[c_ptr], #-7, MUL VL]\n" - "st1h z19.h, p0, [%[c_ptr], #-6, MUL VL]\n" - "st1h z27.h, p0, [%[c_ptr], #-5, MUL VL]\n" - "st1h z12.h, p0, [%[c_ptr], #-4, MUL VL]\n" - "st1h z20.h, p0, [%[c_ptr], #-3, MUL VL]\n" - "st1h z28.h, p0, [%[c_ptr], #-2, MUL VL]\n" - "st1h z13.h, p0, [%[c_ptr], #-1, MUL VL]\n" - "st1h z21.h, p0, [%[c_ptr]]\n" - "st1h z29.h, p0, [%[c_ptr], #1, MUL VL]\n" - "st1h z14.h, p0, [%[c_ptr], #2, MUL VL]\n" - "st1h z22.h, p0, [%[c_ptr], #3, MUL VL]\n" - "st1h z30.h, p0, [%[c_ptr], #4, MUL VL]\n" - "st1h z15.h, p0, [%[c_ptr], #5, MUL VL]\n" - "st1h z23.h, p0, [%[c_ptr], #6, MUL VL]\n" - "st1h z31.h, p0, [%[c_ptr], #7, MUL VL]\n" - "addvl %[c_ptr], %[c_ptr], #8\n" - : [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr), - [loops] "+r" (loops), [tails] "+r" (tails) - : - : "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" - ); - } - } + __asm__ __volatile__( + "ptrue p0.b\n" + "1:" // Height loop + "ldr x22, [%x[args_ptr], %[offsetof_bblocks]]\n" + "mov x21, %x[Apanel]\n" + "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n" + "2:" // Width loop + "mov z8.b, #0x0\n" + "mov z9.b, #0x0\n" + "ldr x19, [%x[args_ptr], %[offsetof_K]]\n" + "mov z10.b, #0x0\n" + "mov z11.b, #0x0\n" + "ld1h { z2.h }, p0/Z, [x20]\n" + "mov z12.b, #0x0\n" + "mov z13.b, #0x0\n" + "mov %x[Apanel], x21\n" + "mov z14.b, #0x0\n" + "mov z15.b, #0x0\n" + "cmp x19, #0x2\n" + "mov z16.b, #0x0\n" + "mov z17.b, #0x0\n" + "mov z18.b, #0x0\n" + "mov z19.b, #0x0\n" + "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n" + "mov z20.b, #0x0\n" + "mov z21.b, #0x0\n" + "mov z22.b, #0x0\n" + "mov z23.b, #0x0\n" + "mov z24.b, #0x0\n" + "mov z25.b, #0x0\n" + "mov z26.b, #0x0\n" + "mov z27.b, #0x0\n" + "mov z28.b, #0x0\n" + "mov z29.b, #0x0\n" + "mov z30.b, #0x0\n" + "mov z31.b, #0x0\n" + "blt 4f\n" + "3:" // main loop head + "fmla z8.h, z2.h, z0.h[0]\n" + "fmla z11.h, z2.h, z0.h[1]\n" + "ld1h { z3.h }, p0/Z, [x20, #1, MUL VL]\n" + "fmla z14.h, z2.h, z0.h[2]\n" + "fmla z17.h, z2.h, z0.h[3]\n" + "ld1h { z4.h }, p0/Z, [x20, #2, MUL VL]\n" + "fmla z20.h, z2.h, z0.h[4]\n" + "fmla z23.h, z2.h, z0.h[5]\n" + "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n" + "fmla z26.h, z2.h, z0.h[6]\n" + "fmla z29.h, z2.h, z0.h[7]\n" + "ld1h { z5.h }, p0/Z, [x20, #3, MUL VL]\n" + "fmla z9.h, z3.h, z0.h[0]\n" + "fmla z12.h, z3.h, z0.h[1]\n" + "ld1h { z6.h }, p0/Z, [x20, #4, MUL VL]\n" + "fmla z15.h, z3.h, z0.h[2]\n" + "fmla z18.h, z3.h, z0.h[3]\n" + "ld1h { z7.h }, p0/Z, [x20, #5, MUL VL]\n" + "fmla z21.h, z3.h, z0.h[4]\n" + "fmla z24.h, z3.h, z0.h[5]\n" + "sub x19, x19, #0x2\n" + "fmla z27.h, z3.h, z0.h[6]\n" + "fmla z30.h, z3.h, z0.h[7]\n" + "cmp x19, #0x2\n" + "fmla z10.h, z4.h, z0.h[0]\n" + "fmla z13.h, z4.h, z0.h[1]\n" + "add %x[Apanel], %x[Apanel], #0x20\n" + "fmla z16.h, z4.h, z0.h[2]\n" + "fmla z19.h, z4.h, z0.h[3]\n" + "addvl x20, x20, #6\n" + "fmla z22.h, z4.h, z0.h[4]\n" + "fmla z25.h, z4.h, z0.h[5]\n" + "fmla z28.h, z4.h, z0.h[6]\n" + "fmla z31.h, z4.h, z0.h[7]\n" + "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n" + "fmla z8.h, z5.h, z1.h[0]\n" + "fmla z11.h, z5.h, z1.h[1]\n" + "ld1h { z2.h }, p0/Z, [x20]\n" + "fmla z14.h, z5.h, z1.h[2]\n" + "fmla z17.h, z5.h, z1.h[3]\n" + "fmla z20.h, z5.h, z1.h[4]\n" + "fmla z23.h, z5.h, z1.h[5]\n" + "fmla z26.h, z5.h, z1.h[6]\n" + "fmla z29.h, z5.h, z1.h[7]\n" + "fmla z9.h, z6.h, z1.h[0]\n" + "fmla z12.h, z6.h, z1.h[1]\n" + "fmla z15.h, z6.h, z1.h[2]\n" + "fmla z18.h, z6.h, z1.h[3]\n" + "fmla z21.h, z6.h, z1.h[4]\n" + "fmla z24.h, z6.h, z1.h[5]\n" + "fmla z27.h, z6.h, z1.h[6]\n" + "fmla z30.h, z6.h, z1.h[7]\n" + "fmla z10.h, z7.h, z1.h[0]\n" + "fmla z13.h, z7.h, z1.h[1]\n" + "fmla z16.h, z7.h, z1.h[2]\n" + "fmla z19.h, z7.h, z1.h[3]\n" + "fmla z22.h, z7.h, z1.h[4]\n" + "fmla z25.h, z7.h, z1.h[5]\n" + "fmla z28.h, z7.h, z1.h[6]\n" + "fmla z31.h, z7.h, z1.h[7]\n" + "bge 3b\n" + "4:" // main loop skip + "fmla z8.h, z2.h, z0.h[0]\n" + "fmla z11.h, z2.h, z0.h[1]\n" + "ld1h { z3.h }, p0/Z, [x20, #1, MUL VL]\n" + "fmla z14.h, z2.h, z0.h[2]\n" + "fmla z17.h, z2.h, z0.h[3]\n" + "ld1h { z4.h }, p0/Z, [x20, #2, MUL VL]\n" + "fmla z20.h, z2.h, z0.h[4]\n" + "fmla z23.h, z2.h, z0.h[5]\n" + "add %x[Apanel], %x[Apanel], #0x10\n" + "fmla z26.h, z2.h, z0.h[6]\n" + "fmla z29.h, z2.h, z0.h[7]\n" + "addvl x20, x20, #3\n" + "fmla z9.h, z3.h, z0.h[0]\n" + "fmla z12.h, z3.h, z0.h[1]\n" + "fmla z15.h, z3.h, z0.h[2]\n" + "fmla z18.h, z3.h, z0.h[3]\n" + "fmla z21.h, z3.h, z0.h[4]\n" + "fmla z24.h, z3.h, z0.h[5]\n" + "fmla z27.h, z3.h, z0.h[6]\n" + "fmla z30.h, z3.h, z0.h[7]\n" + "fmla z10.h, z4.h, z0.h[0]\n" + "fmla z13.h, z4.h, z0.h[1]\n" + "fmla z16.h, z4.h, z0.h[2]\n" + "fmla z19.h, z4.h, z0.h[3]\n" + "fmla z22.h, z4.h, z0.h[4]\n" + "fmla z25.h, z4.h, z0.h[5]\n" + "fmla z28.h, z4.h, z0.h[6]\n" + "fmla z31.h, z4.h, z0.h[7]\n" + "cbz x19, 5f\n" + "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n" + "add %x[Apanel], %x[Apanel], #0x10\n" + "ld1h { z5.h }, p0/Z, [x20]\n" + "ld1h { z6.h }, p0/Z, [x20, #1, MUL VL]\n" + "ld1h { z7.h }, p0/Z, [x20, #2, MUL VL]\n" + "addvl x20, x20, #3\n" + "fmla z8.h, z5.h, z0.h[0]\n" + "fmla z11.h, z5.h, z0.h[1]\n" + "fmla z14.h, z5.h, z0.h[2]\n" + "fmla z17.h, z5.h, z0.h[3]\n" + "fmla z20.h, z5.h, z0.h[4]\n" + "fmla z23.h, z5.h, z0.h[5]\n" + "fmla z26.h, z5.h, z0.h[6]\n" + "fmla z29.h, z5.h, z0.h[7]\n" + "fmla z9.h, z6.h, z0.h[0]\n" + "fmla z12.h, z6.h, z0.h[1]\n" + "fmla z15.h, z6.h, z0.h[2]\n" + "fmla z18.h, z6.h, z0.h[3]\n" + "fmla z21.h, z6.h, z0.h[4]\n" + "fmla z24.h, z6.h, z0.h[5]\n" + "fmla z27.h, z6.h, z0.h[6]\n" + "fmla z30.h, z6.h, z0.h[7]\n" + "fmla z10.h, z7.h, z0.h[0]\n" + "fmla z13.h, z7.h, z0.h[1]\n" + "fmla z16.h, z7.h, z0.h[2]\n" + "fmla z19.h, z7.h, z0.h[3]\n" + "fmla z22.h, z7.h, z0.h[4]\n" + "fmla z25.h, z7.h, z0.h[5]\n" + "fmla z28.h, z7.h, z0.h[6]\n" + "fmla z31.h, z7.h, z0.h[7]\n" + "5:" // multiply loop done + "st1h { z8.h }, p0, [%x[Cpanel]]\n" + "subs x22, x22, #0x1\n" + "st1h { z9.h }, p0, [%x[Cpanel], #1, MUL VL]\n" + "st1h { z10.h }, p0, [%x[Cpanel], #2, MUL VL]\n" + "st1h { z11.h }, p0, [%x[Cpanel], #3, MUL VL]\n" + "st1h { z12.h }, p0, [%x[Cpanel], #4, MUL VL]\n" + "st1h { z13.h }, p0, [%x[Cpanel], #5, MUL VL]\n" + "st1h { z14.h }, p0, [%x[Cpanel], #6, MUL VL]\n" + "st1h { z15.h }, p0, [%x[Cpanel], #7, MUL VL]\n" + "addvl %x[Cpanel], %x[Cpanel], #16\n" + "st1h { z16.h }, p0, [%x[Cpanel], #-8, MUL VL]\n" + "st1h { z17.h }, p0, [%x[Cpanel], #-7, MUL VL]\n" + "st1h { z18.h }, p0, [%x[Cpanel], #-6, MUL VL]\n" + "st1h { z19.h }, p0, [%x[Cpanel], #-5, MUL VL]\n" + "st1h { z20.h }, p0, [%x[Cpanel], #-4, MUL VL]\n" + "st1h { z21.h }, p0, [%x[Cpanel], #-3, MUL VL]\n" + "st1h { z22.h }, p0, [%x[Cpanel], #-2, MUL VL]\n" + "st1h { z23.h }, p0, [%x[Cpanel], #-1, MUL VL]\n" + "st1h { z24.h }, p0, [%x[Cpanel]]\n" + "st1h { z25.h }, p0, [%x[Cpanel], #1, MUL VL]\n" + "st1h { z26.h }, p0, [%x[Cpanel], #2, MUL VL]\n" + "st1h { z27.h }, p0, [%x[Cpanel], #3, MUL VL]\n" + "st1h { z28.h }, p0, [%x[Cpanel], #4, MUL VL]\n" + "st1h { z29.h }, p0, [%x[Cpanel], #5, MUL VL]\n" + "st1h { z30.h }, p0, [%x[Cpanel], #6, MUL VL]\n" + "st1h { z31.h }, p0, [%x[Cpanel], #7, MUL VL]\n" + "addvl %x[Cpanel], %x[Cpanel], #8\n" + "bgt 2b\n" + "subs %x[ablocks], %x[ablocks], #0x1\n" + "bne 1b\n" + : [Apanel] "+&r" (Apanel), [Cpanel] "+&r" (Cpanel), [ablocks] "+&r" (ablocks) + : [args_ptr] "r" (&ka), [offsetof_Bpanel] "I" (offsetof(KernelArgs, Bpanel)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_bblocks] "I" (offsetof(KernelArgs, bblocks)) + : "cc", "memory", "p0", "x19", "x20", "x21", "x22", "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" + ); } } // namespace arm_gemm - #endif // ARM_COMPUTE_ENABLE_SVE |