aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp628
1 files changed, 264 insertions, 364 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp
index e43404e608..ba7185752a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021, 2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -21,377 +21,277 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#ifdef __ARM_FEATURE_SVE
+#ifdef ARM_COMPUTE_ENABLE_SVE
+#include <cstddef>
#include "../../bfloat.hpp"
-#include "../../asmlib.hpp"
namespace arm_gemm {
-void sve_interleaved_bf16fp32_mmla_8x3VL(const bfloat16 *Apanel, const bfloat16 *Bpanel, float *Cpanel, int ablocks, int bblocks, int K) {
- const bfloat16 *a_ptr = Apanel;
- float *c_ptr = Cpanel;
+void sve_interleaved_bf16fp32_mmla_8x3VL(
+ const bfloat16 *Apanel,
+ const bfloat16 *Bpanel,
+ float *Cpanel,
+ int ablocks,
+ int bblocks,
+ int K) {
- K /= 4;
- const long loops_count = (K / 2) - 1;
- const long tails_count = K % 2;
+ struct KernelArgs {
+ size_t K = {};
+ const bfloat16 *Bpanel = {};
+ size_t bblocks = {};
+ } ka;
- for (int yb=0; yb<ablocks; yb++) {
- const bfloat16 *a_ptr0 = a_ptr;
- const bfloat16 *b_ptr = Bpanel;
+ ka.K = (K/4) - 1;
+ ka.Bpanel = Bpanel;
+ ka.bblocks = bblocks;
- for (int xb=0; xb<bblocks; xb++) {
- a_ptr = a_ptr0;
- long loops = loops_count;
- long tails = tails_count;
-
- __asm __volatile (
- "mov z8.s, #0\n"
- "ptrue p0.h\n"
- "mov z9.s, #0\n"
- "mov z10.s, #0\n"
- "mov z11.s, #0\n"
- "ld1rqh z0.h, p0/z, [%[a_ptr]]\n"
- "mov z12.s, #0\n"
- "ld1h z4.h, p0/z, [%[b_ptr]]\n"
- "mov z13.s, #0\n"
- "ld1rqh z1.h, p0/z, [%[a_ptr], #0x10]\n"
- "mov z14.s, #0\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #1, MUL VL]\n"
- "mov z15.s, #0\n"
- "ld1rqh z2.h, p0/z, [%[a_ptr], #0x20]\n"
- "mov z16.s, #0\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #2, MUL VL]\n"
- "mov z17.s, #0\n"
- "add %[a_ptr], %[a_ptr], #0x40\n"
- "mov z18.s, #0\n"
- "addvl %[b_ptr], %[b_ptr], #4\n"
- "mov z19.s, #0\n"
- "mov z20.s, #0\n"
- "mov z21.s, #0\n"
- "mov z22.s, #0\n"
- "mov z23.s, #0\n"
- "mov z24.s, #0\n"
- "mov z25.s, #0\n"
- "mov z26.s, #0\n"
- "mov z27.s, #0\n"
- "mov z28.s, #0\n"
- "mov z29.s, #0\n"
- "mov z30.s, #0\n"
- "mov z31.s, #0\n"
- "cbz %[loops], 1f\n"
- "2:\n"
- ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x6464e42e // bfmmla z14.s, z1.h, z4.h\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #-0x10]\n"
- ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n"
- "subs %[loops], %[loops], #0x1\n"
- ".inst 0x6465e409 // bfmmla z9.s, z0.h, z5.h\n"
- ".inst 0x6464e47a // bfmmla z26.s, z3.h, z4.h\n"
- "ld1h z4.h, p0/z, [%[b_ptr]]\n"
- ".inst 0x6465e42f // bfmmla z15.s, z1.h, z5.h\n"
- ".inst 0x6465e455 // bfmmla z21.s, z2.h, z5.h\n"
- ".inst 0x6465e47b // bfmmla z27.s, z3.h, z5.h\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #1, MUL VL]\n"
- ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n"
- ".inst 0x6466e430 // bfmmla z16.s, z1.h, z6.h\n"
- ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n"
- ".inst 0x6466e47c // bfmmla z28.s, z3.h, z6.h\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #2, MUL VL]\n"
- ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n"
- ".inst 0x6467e431 // bfmmla z17.s, z1.h, z7.h\n"
- ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n"
- ".inst 0x6467e47d // bfmmla z29.s, z3.h, z7.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #3, MUL VL]\n"
- ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n"
- ".inst 0x6464e432 // bfmmla z18.s, z1.h, z4.h\n"
- ".inst 0x6464e458 // bfmmla z24.s, z2.h, z4.h\n"
- ".inst 0x6464e47e // bfmmla z30.s, z3.h, z4.h\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #4, MUL VL]\n"
- ".inst 0x6465e40d // bfmmla z13.s, z0.h, z5.h\n"
- "ld1rqh z0.h, p0/z, [%[a_ptr]]\n"
- ".inst 0x6465e433 // bfmmla z19.s, z1.h, z5.h\n"
- "ld1rqh z1.h, p0/z, [%[a_ptr], #0x10]\n"
- ".inst 0x6465e459 // bfmmla z25.s, z2.h, z5.h\n"
- "ld1rqh z2.h, p0/z, [%[a_ptr], #0x20]\n"
- ".inst 0x6465e47f // bfmmla z31.s, z3.h, z5.h\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #5, MUL VL]\n"
- ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #0x30]\n"
- ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n"
- "add %[a_ptr], %[a_ptr], #0x80\n"
- ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n"
- "addvl %[b_ptr], %[b_ptr], #12\n"
- ".inst 0x6466e47a // bfmmla z26.s, z3.h, z6.h\n"
- ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n"
- ".inst 0x6467e42f // bfmmla z15.s, z1.h, z7.h\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-6, MUL VL]\n"
- ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n"
- ".inst 0x6467e47b // bfmmla z27.s, z3.h, z7.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #-5, MUL VL]\n"
- ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n"
- ".inst 0x6464e430 // bfmmla z16.s, z1.h, z4.h\n"
- ".inst 0x6464e456 // bfmmla z22.s, z2.h, z4.h\n"
- ".inst 0x6464e47c // bfmmla z28.s, z3.h, z4.h\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #-4, MUL VL]\n"
- ".inst 0x6465e40b // bfmmla z11.s, z0.h, z5.h\n"
- ".inst 0x6465e431 // bfmmla z17.s, z1.h, z5.h\n"
- ".inst 0x6465e457 // bfmmla z23.s, z2.h, z5.h\n"
- ".inst 0x6465e47d // bfmmla z29.s, z3.h, z5.h\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #-3, MUL VL]\n"
- ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n"
- ".inst 0x6466e432 // bfmmla z18.s, z1.h, z6.h\n"
- ".inst 0x6466e458 // bfmmla z24.s, z2.h, z6.h\n"
- ".inst 0x6466e47e // bfmmla z30.s, z3.h, z6.h\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-2, MUL VL]\n"
- ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n"
- "ld1rqh z0.h, p0/z, [%[a_ptr], #-0x40]\n"
- ".inst 0x6467e433 // bfmmla z19.s, z1.h, z7.h\n"
- "ld1rqh z1.h, p0/z, [%[a_ptr], #-0x30]\n"
- ".inst 0x6467e459 // bfmmla z25.s, z2.h, z7.h\n"
- "ld1rqh z2.h, p0/z, [%[a_ptr], #-0x20]\n"
- ".inst 0x6467e47f // bfmmla z31.s, z3.h, z7.h\n"
- "b.ne 2b\n"
- "1:\n"
- "cbz %[tails], 3f\n"
- ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x6464e42e // bfmmla z14.s, z1.h, z4.h\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #-0x10]\n"
- ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n"
- ".inst 0x6465e409 // bfmmla z9.s, z0.h, z5.h\n"
- ".inst 0x6465e42f // bfmmla z15.s, z1.h, z5.h\n"
- ".inst 0x6464e47a // bfmmla z26.s, z3.h, z4.h\n"
- "ld1h z4.h, p0/z, [%[b_ptr]]\n"
- ".inst 0x6465e455 // bfmmla z21.s, z2.h, z5.h\n"
- ".inst 0x6465e47b // bfmmla z27.s, z3.h, z5.h\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #1, MUL VL]\n"
- ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n"
- ".inst 0x6466e430 // bfmmla z16.s, z1.h, z6.h\n"
- ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n"
- ".inst 0x6466e47c // bfmmla z28.s, z3.h, z6.h\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #2, MUL VL]\n"
- ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n"
- ".inst 0x6467e431 // bfmmla z17.s, z1.h, z7.h\n"
- ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n"
- ".inst 0x6467e47d // bfmmla z29.s, z3.h, z7.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #3, MUL VL]\n"
- ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n"
- ".inst 0x6464e432 // bfmmla z18.s, z1.h, z4.h\n"
- ".inst 0x6464e458 // bfmmla z24.s, z2.h, z4.h\n"
- ".inst 0x6464e47e // bfmmla z30.s, z3.h, z4.h\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #4, MUL VL]\n"
- ".inst 0x6465e40d // bfmmla z13.s, z0.h, z5.h\n"
- "ld1rqh z0.h, p0/z, [%[a_ptr]]\n"
- ".inst 0x6465e433 // bfmmla z19.s, z1.h, z5.h\n"
- "ld1rqh z1.h, p0/z, [%[a_ptr], #0x10]\n"
- ".inst 0x6465e459 // bfmmla z25.s, z2.h, z5.h\n"
- "ld1rqh z2.h, p0/z, [%[a_ptr], #0x20]\n"
- ".inst 0x6465e47f // bfmmla z31.s, z3.h, z5.h\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #5, MUL VL]\n"
- ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #0x30]\n"
- ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n"
- "add %[a_ptr], %[a_ptr], #0x80\n"
- ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n"
- "addvl %[b_ptr], %[b_ptr], #14\n"
- ".inst 0x6466e47a // bfmmla z26.s, z3.h, z6.h\n"
- ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n"
- ".inst 0x6467e42f // bfmmla z15.s, z1.h, z7.h\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-8, MUL VL]\n"
- ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n"
- ".inst 0x6467e47b // bfmmla z27.s, z3.h, z7.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #-7, MUL VL]\n"
- ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n"
- ".inst 0x6464e430 // bfmmla z16.s, z1.h, z4.h\n"
- ".inst 0x6464e456 // bfmmla z22.s, z2.h, z4.h\n"
- ".inst 0x6464e47c // bfmmla z28.s, z3.h, z4.h\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #-6, MUL VL]\n"
- ".inst 0x6465e40b // bfmmla z11.s, z0.h, z5.h\n"
- ".inst 0x6465e431 // bfmmla z17.s, z1.h, z5.h\n"
- ".inst 0x6465e457 // bfmmla z23.s, z2.h, z5.h\n"
- ".inst 0x6465e47d // bfmmla z29.s, z3.h, z5.h\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #-5, MUL VL]\n"
- ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n"
- ".inst 0x6466e432 // bfmmla z18.s, z1.h, z6.h\n"
- ".inst 0x6466e458 // bfmmla z24.s, z2.h, z6.h\n"
- ".inst 0x6466e47e // bfmmla z30.s, z3.h, z6.h\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-4, MUL VL]\n"
- ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n"
- "ld1rqh z0.h, p0/z, [%[a_ptr], #-0x40]\n"
- ".inst 0x6467e433 // bfmmla z19.s, z1.h, z7.h\n"
- "ld1rqh z1.h, p0/z, [%[a_ptr], #-0x30]\n"
- ".inst 0x6467e459 // bfmmla z25.s, z2.h, z7.h\n"
- "ld1rqh z2.h, p0/z, [%[a_ptr], #-0x20]\n"
- ".inst 0x6467e47f // bfmmla z31.s, z3.h, z7.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #-3, MUL VL]\n"
- ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #-0x10]\n"
- ".inst 0x6464e42e // bfmmla z14.s, z1.h, z4.h\n"
- ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n"
- ".inst 0x6465e409 // bfmmla z9.s, z0.h, z5.h\n"
- ".inst 0x6464e47a // bfmmla z26.s, z3.h, z4.h\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #-2, MUL VL]\n"
- ".inst 0x6465e42f // bfmmla z15.s, z1.h, z5.h\n"
- ".inst 0x6465e455 // bfmmla z21.s, z2.h, z5.h\n"
- ".inst 0x6465e47b // bfmmla z27.s, z3.h, z5.h\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n"
- ".inst 0x6466e430 // bfmmla z16.s, z1.h, z6.h\n"
- ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n"
- ".inst 0x6466e47c // bfmmla z28.s, z3.h, z6.h\n"
- "uzp1 z6.d, z14.d, z15.d\n"
- ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n"
- ".inst 0x6467e431 // bfmmla z17.s, z1.h, z7.h\n"
- ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n"
- ".inst 0x6467e47d // bfmmla z29.s, z3.h, z7.h\n"
- ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n"
- "uzp1 z7.d, z16.d, z17.d\n"
- ".inst 0x6464e432 // bfmmla z18.s, z1.h, z4.h\n"
- ".inst 0x6464e458 // bfmmla z24.s, z2.h, z4.h\n"
- ".inst 0x6464e47e // bfmmla z30.s, z3.h, z4.h\n"
- "uzp2 z4.d, z10.d, z11.d\n"
- ".inst 0x6465e40d // bfmmla z13.s, z0.h, z5.h\n"
- "uzp1 z0.d, z8.d, z9.d\n"
- ".inst 0x6465e433 // bfmmla z19.s, z1.h, z5.h\n"
- "uzp1 z1.d, z10.d, z11.d\n"
- ".inst 0x6465e459 // bfmmla z25.s, z2.h, z5.h\n"
- "st1w z0.s, p0, [%[c_ptr]]\n"
- "uzp1 z2.d, z12.d, z13.d\n"
- "uzp1 z0.d, z18.d, z19.d\n"
- ".inst 0x6465e47f // bfmmla z31.s, z3.h, z5.h\n"
- "st1w z1.s, p0, [%[c_ptr], #1, MUL VL]\n"
- "uzp2 z3.d, z8.d, z9.d\n"
- "uzp2 z5.d, z12.d, z13.d\n"
- "uzp2 z1.d, z14.d, z15.d\n"
- "st1w z2.s, p0, [%[c_ptr], #2, MUL VL]\n"
- "b 4f\n"
- "3:\n"
- ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x6464e42e // bfmmla z14.s, z1.h, z4.h\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #-0x10]\n"
- ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n"
- "add %[a_ptr], %[a_ptr], #0x40\n"
- ".inst 0x6465e409 // bfmmla z9.s, z0.h, z5.h\n"
- "addvl %[b_ptr], %[b_ptr], #8\n"
- ".inst 0x6464e47a // bfmmla z26.s, z3.h, z4.h\n"
- ".inst 0x6465e42f // bfmmla z15.s, z1.h, z5.h\n"
- ".inst 0x6465e455 // bfmmla z21.s, z2.h, z5.h\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #-8, MUL VL]\n"
- ".inst 0x6465e47b // bfmmla z27.s, z3.h, z5.h\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #-7, MUL VL]\n"
- ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n"
- ".inst 0x6466e430 // bfmmla z16.s, z1.h, z6.h\n"
- ".inst 0x6466e456 // bfmmla z22.s, z2.h, z6.h\n"
- ".inst 0x6466e47c // bfmmla z28.s, z3.h, z6.h\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-6, MUL VL]\n"
- ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n"
- ".inst 0x6467e431 // bfmmla z17.s, z1.h, z7.h\n"
- ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n"
- ".inst 0x6467e47d // bfmmla z29.s, z3.h, z7.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #-5, MUL VL]\n"
- ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n"
- ".inst 0x6464e432 // bfmmla z18.s, z1.h, z4.h\n"
- ".inst 0x6464e458 // bfmmla z24.s, z2.h, z4.h\n"
- ".inst 0x6464e47e // bfmmla z30.s, z3.h, z4.h\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #-4, MUL VL]\n"
- ".inst 0x6465e40d // bfmmla z13.s, z0.h, z5.h\n"
- "ld1rqh z0.h, p0/z, [%[a_ptr], #-0x40]\n"
- ".inst 0x6465e433 // bfmmla z19.s, z1.h, z5.h\n"
- "ld1rqh z1.h, p0/z, [%[a_ptr], #-0x30]\n"
- ".inst 0x6465e459 // bfmmla z25.s, z2.h, z5.h\n"
- "ld1rqh z2.h, p0/z, [%[a_ptr], #-0x20]\n"
- ".inst 0x6465e47f // bfmmla z31.s, z3.h, z5.h\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #-3, MUL VL]\n"
- ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #-0x10]\n"
- ".inst 0x6466e42e // bfmmla z14.s, z1.h, z6.h\n"
- ".inst 0x6466e454 // bfmmla z20.s, z2.h, z6.h\n"
- ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n"
- ".inst 0x6466e47a // bfmmla z26.s, z3.h, z6.h\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-2, MUL VL]\n"
- ".inst 0x6467e42f // bfmmla z15.s, z1.h, z7.h\n"
- ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n"
- ".inst 0x6467e47b // bfmmla z27.s, z3.h, z7.h\n"
- "ld1h z7.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x6464e40a // bfmmla z10.s, z0.h, z4.h\n"
- ".inst 0x6464e430 // bfmmla z16.s, z1.h, z4.h\n"
- ".inst 0x6464e456 // bfmmla z22.s, z2.h, z4.h\n"
- ".inst 0x6464e47c // bfmmla z28.s, z3.h, z4.h\n"
- ".inst 0x6465e40b // bfmmla z11.s, z0.h, z5.h\n"
- ".inst 0x6465e431 // bfmmla z17.s, z1.h, z5.h\n"
- ".inst 0x6465e457 // bfmmla z23.s, z2.h, z5.h\n"
- ".inst 0x6465e47d // bfmmla z29.s, z3.h, z5.h\n"
- "uzp2 z4.d, z10.d, z11.d\n"
- ".inst 0x6466e40c // bfmmla z12.s, z0.h, z6.h\n"
- ".inst 0x6466e432 // bfmmla z18.s, z1.h, z6.h\n"
- ".inst 0x6466e458 // bfmmla z24.s, z2.h, z6.h\n"
- ".inst 0x6466e47e // bfmmla z30.s, z3.h, z6.h\n"
- "uzp1 z6.d, z14.d, z15.d\n"
- ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n"
- "uzp1 z0.d, z8.d, z9.d\n"
- ".inst 0x6467e433 // bfmmla z19.s, z1.h, z7.h\n"
- "uzp1 z1.d, z10.d, z11.d\n"
- "uzp2 z5.d, z12.d, z13.d\n"
- "st1w z0.s, p0, [%[c_ptr]]\n"
- ".inst 0x6467e459 // bfmmla z25.s, z2.h, z7.h\n"
- "uzp1 z2.d, z12.d, z13.d\n"
- "uzp1 z0.d, z18.d, z19.d\n"
- "st1w z1.s, p0, [%[c_ptr], #1, MUL VL]\n"
- "uzp2 z1.d, z14.d, z15.d\n"
- ".inst 0x6467e47f // bfmmla z31.s, z3.h, z7.h\n"
- "uzp2 z3.d, z8.d, z9.d\n"
- "st1w z2.s, p0, [%[c_ptr], #2, MUL VL]\n"
- "uzp1 z7.d, z16.d, z17.d\n"
- "4:\n"
- "uzp2 z2.d, z16.d, z17.d\n"
- "st1w z3.s, p0, [%[c_ptr], #3, MUL VL]\n"
- "uzp2 z3.d, z18.d, z19.d\n"
- "st1w z4.s, p0, [%[c_ptr], #4, MUL VL]\n"
- "uzp1 z4.d, z20.d, z21.d\n"
- "st1w z5.s, p0, [%[c_ptr], #5, MUL VL]\n"
- "uzp1 z5.d, z22.d, z23.d\n"
- "st1w z6.s, p0, [%[c_ptr], #6, MUL VL]\n"
- "uzp1 z6.d, z24.d, z25.d\n"
- "st1w z7.s, p0, [%[c_ptr], #7, MUL VL]\n"
- "addvl %[c_ptr], %[c_ptr], #16\n"
- "uzp2 z7.d, z20.d, z21.d\n"
- "st1w z0.s, p0, [%[c_ptr], #-8, MUL VL]\n"
- "uzp2 z0.d, z22.d, z23.d\n"
- "st1w z1.s, p0, [%[c_ptr], #-7, MUL VL]\n"
- "uzp2 z1.d, z24.d, z25.d\n"
- "st1w z2.s, p0, [%[c_ptr], #-6, MUL VL]\n"
- "uzp1 z2.d, z26.d, z27.d\n"
- "st1w z3.s, p0, [%[c_ptr], #-5, MUL VL]\n"
- "uzp1 z3.d, z28.d, z29.d\n"
- "st1w z4.s, p0, [%[c_ptr], #-4, MUL VL]\n"
- "uzp1 z4.d, z30.d, z31.d\n"
- "st1w z5.s, p0, [%[c_ptr], #-3, MUL VL]\n"
- "uzp2 z5.d, z26.d, z27.d\n"
- "st1w z6.s, p0, [%[c_ptr], #-2, MUL VL]\n"
- "uzp2 z6.d, z28.d, z29.d\n"
- "st1w z7.s, p0, [%[c_ptr], #-1, MUL VL]\n"
- "uzp2 z7.d, z30.d, z31.d\n"
- "st1w z0.s, p0, [%[c_ptr]]\n"
- "st1w z1.s, p0, [%[c_ptr], #1, MUL VL]\n"
- "st1w z2.s, p0, [%[c_ptr], #2, MUL VL]\n"
- "st1w z3.s, p0, [%[c_ptr], #3, MUL VL]\n"
- "st1w z4.s, p0, [%[c_ptr], #4, MUL VL]\n"
- "st1w z5.s, p0, [%[c_ptr], #5, MUL VL]\n"
- "st1w z6.s, p0, [%[c_ptr], #6, MUL VL]\n"
- "st1w z7.s, 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 x23, [%x[args_ptr], %[offsetof_bblocks]]\n"
+ "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n"
+ "mov x21, %x[Apanel]\n"
+ "2:" // Width loop
+ "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+ "mov %x[Apanel], x21\n"
+ "cmp x20, #0x2\n"
+ "mov z8.b, #0x0\n"
+ "mov z9.b, #0x0\n"
+ "ld1h { z4.h }, p0/Z, [x22]\n"
+ "mov z10.b, #0x0\n"
+ "mov z11.b, #0x0\n"
+ "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
+ "mov z12.b, #0x0\n"
+ "mov z13.b, #0x0\n"
+ "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
+ "mov z14.b, #0x0\n"
+ "mov z15.b, #0x0\n"
+ "ld1h { z5.h }, p0/Z, [x22, #1, MUL VL]\n"
+ "mov z16.b, #0x0\n"
+ "mov z17.b, #0x0\n"
+ "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #32]\n"
+ "mov z18.b, #0x0\n"
+ "mov z19.b, #0x0\n"
+ "addvl x22, x22, #2\n"
+ "mov z20.b, #0x0\n"
+ "mov z21.b, #0x0\n"
+ "add %x[Apanel], %x[Apanel], #0x30\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
+ "ld1rqh { z6.h }, p0/Z, [%x[Apanel]]\n"
+ ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n"
+ ".inst 0x6465e40b // bfmmla z11.s, z0.h, z5.h\n"
+ ".inst 0x6464e42e // bfmmla z14.s, z1.h, z4.h\n"
+ ".inst 0x6465e431 // bfmmla z17.s, z1.h, z5.h\n"
+ "ld1h { z7.h }, p0/Z, [x22]\n"
+ ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n"
+ ".inst 0x6465e457 // bfmmla z23.s, z2.h, z5.h\n"
+ "ld1h { z3.h }, p0/Z, [x22, #1, MUL VL]\n"
+ ".inst 0x6464e4da // bfmmla z26.s, z6.h, z4.h\n"
+ ".inst 0x6465e4dd // bfmmla z29.s, z6.h, z5.h\n"
+ "ld1h { z5.h }, p0/Z, [x22, #2, MUL VL]\n"
+ "ld1h { z4.h }, p0/Z, [x22, #3, MUL VL]\n"
+ ".inst 0x6467e409 // bfmmla z9.s, z0.h, z7.h\n"
+ ".inst 0x6463e40c // bfmmla z12.s, z0.h, z3.h\n"
+ ".inst 0x6467e42f // bfmmla z15.s, z1.h, z7.h\n"
+ ".inst 0x6463e432 // bfmmla z18.s, z1.h, z3.h\n"
+ "sub x20, x20, #0x2\n"
+ ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n"
+ ".inst 0x6463e458 // bfmmla z24.s, z2.h, z3.h\n"
+ "cmp x20, #0x2\n"
+ ".inst 0x6467e4db // bfmmla z27.s, z6.h, z7.h\n"
+ ".inst 0x6463e4de // bfmmla z30.s, z6.h, z3.h\n"
+ "ld1h { z3.h }, p0/Z, [x22, #4, MUL VL]\n"
+ ".inst 0x6465e40a // bfmmla z10.s, z0.h, z5.h\n"
+ ".inst 0x6464e40d // bfmmla z13.s, z0.h, z4.h\n"
+ "ld1rqh { z0.h }, p0/Z, [%x[Apanel], #16]\n"
+ ".inst 0x6465e430 // bfmmla z16.s, z1.h, z5.h\n"
+ ".inst 0x6464e433 // bfmmla z19.s, z1.h, z4.h\n"
+ "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #32]\n"
+ ".inst 0x6465e456 // bfmmla z22.s, z2.h, z5.h\n"
+ ".inst 0x6464e459 // bfmmla z25.s, z2.h, z4.h\n"
+ "ld1h { z7.h }, p0/Z, [x22, #5, MUL VL]\n"
+ ".inst 0x6465e4dc // bfmmla z28.s, z6.h, z5.h\n"
+ ".inst 0x6464e4df // bfmmla z31.s, z6.h, z4.h\n"
+ "ld1rqh { z5.h }, p0/Z, [%x[Apanel], #48]\n"
+ "ld1rqh { z6.h }, p0/Z, [%x[Apanel], #64]\n"
+ "ld1h { z2.h }, p0/Z, [x22, #6, MUL VL]\n"
+ ".inst 0x6463e408 // bfmmla z8.s, z0.h, z3.h\n"
+ "ld1h { z4.h }, p0/Z, [x22, #7, MUL VL]\n"
+ "addvl x22, x22, #16\n"
+ ".inst 0x6467e40b // bfmmla z11.s, z0.h, z7.h\n"
+ ".inst 0x6463e42e // bfmmla z14.s, z1.h, z3.h\n"
+ ".inst 0x6467e431 // bfmmla z17.s, z1.h, z7.h\n"
+ ".inst 0x6463e4b4 // bfmmla z20.s, z5.h, z3.h\n"
+ ".inst 0x6467e4b7 // bfmmla z23.s, z5.h, z7.h\n"
+ ".inst 0x6463e4da // bfmmla z26.s, z6.h, z3.h\n"
+ ".inst 0x6467e4dd // bfmmla z29.s, z6.h, z7.h\n"
+ "ld1h { z3.h }, p0/Z, [x22, #-8, MUL VL]\n"
+ "ld1h { z7.h }, p0/Z, [x22, #-7, MUL VL]\n"
+ ".inst 0x6462e409 // bfmmla z9.s, z0.h, z2.h\n"
+ ".inst 0x6464e40c // bfmmla z12.s, z0.h, z4.h\n"
+ ".inst 0x6462e42f // bfmmla z15.s, z1.h, z2.h\n"
+ ".inst 0x6464e432 // bfmmla z18.s, z1.h, z4.h\n"
+ ".inst 0x6462e4b5 // bfmmla z21.s, z5.h, z2.h\n"
+ ".inst 0x6464e4b8 // bfmmla z24.s, z5.h, z4.h\n"
+ ".inst 0x6462e4db // bfmmla z27.s, z6.h, z2.h\n"
+ ".inst 0x6464e4de // bfmmla z30.s, z6.h, z4.h\n"
+ "ld1h { z4.h }, p0/Z, [x22, #-6, MUL VL]\n"
+ ".inst 0x6463e40a // bfmmla z10.s, z0.h, z3.h\n"
+ ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n"
+ "ld1rqh { z0.h }, p0/Z, [%x[Apanel], #80]\n"
+ ".inst 0x6463e430 // bfmmla z16.s, z1.h, z3.h\n"
+ ".inst 0x6467e433 // bfmmla z19.s, z1.h, z7.h\n"
+ "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #96]\n"
+ ".inst 0x6463e4b6 // bfmmla z22.s, z5.h, z3.h\n"
+ ".inst 0x6467e4b9 // bfmmla z25.s, z5.h, z7.h\n"
+ "ld1h { z5.h }, p0/Z, [x22, #-5, MUL VL]\n"
+ ".inst 0x6463e4dc // bfmmla z28.s, z6.h, z3.h\n"
+ ".inst 0x6467e4df // bfmmla z31.s, z6.h, z7.h\n"
+ "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #112]\n"
+ "add %x[Apanel], %x[Apanel], #0x80\n"
+ "addvl x22, x22, #-4\n"
+ "bge 3b\n"
+ "4:" // main loop skip
+ "ld1rqh { z7.h }, p0/Z, [%x[Apanel]]\n"
+ ".inst 0x6464e408 // bfmmla z8.s, z0.h, z4.h\n"
+ ".inst 0x6465e40b // bfmmla z11.s, z0.h, z5.h\n"
+ ".inst 0x6464e42e // bfmmla z14.s, z1.h, z4.h\n"
+ ".inst 0x6465e431 // bfmmla z17.s, z1.h, z5.h\n"
+ "ld1h { z6.h }, p0/Z, [x22]\n"
+ ".inst 0x6464e454 // bfmmla z20.s, z2.h, z4.h\n"
+ ".inst 0x6465e457 // bfmmla z23.s, z2.h, z5.h\n"
+ "ld1h { z3.h }, p0/Z, [x22, #1, MUL VL]\n"
+ ".inst 0x6464e4fa // bfmmla z26.s, z7.h, z4.h\n"
+ ".inst 0x6465e4fd // bfmmla z29.s, z7.h, z5.h\n"
+ "ld1h { z5.h }, p0/Z, [x22, #2, MUL VL]\n"
+ "ld1h { z4.h }, p0/Z, [x22, #3, MUL VL]\n"
+ ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n"
+ ".inst 0x6463e40c // bfmmla z12.s, z0.h, z3.h\n"
+ ".inst 0x6466e42f // bfmmla z15.s, z1.h, z6.h\n"
+ ".inst 0x6463e432 // bfmmla z18.s, z1.h, z3.h\n"
+ "add %x[Apanel], %x[Apanel], #0x10\n"
+ ".inst 0x6466e455 // bfmmla z21.s, z2.h, z6.h\n"
+ ".inst 0x6463e458 // bfmmla z24.s, z2.h, z3.h\n"
+ "addvl x22, x22, #4\n"
+ ".inst 0x6466e4fb // bfmmla z27.s, z7.h, z6.h\n"
+ ".inst 0x6463e4fe // bfmmla z30.s, z7.h, z3.h\n"
+ ".inst 0x6465e40a // bfmmla z10.s, z0.h, z5.h\n"
+ ".inst 0x6464e40d // bfmmla z13.s, z0.h, z4.h\n"
+ ".inst 0x6465e430 // bfmmla z16.s, z1.h, z5.h\n"
+ ".inst 0x6464e433 // bfmmla z19.s, z1.h, z4.h\n"
+ ".inst 0x6465e456 // bfmmla z22.s, z2.h, z5.h\n"
+ ".inst 0x6464e459 // bfmmla z25.s, z2.h, z4.h\n"
+ ".inst 0x6465e4fc // bfmmla z28.s, z7.h, z5.h\n"
+ ".inst 0x6464e4ff // bfmmla z31.s, z7.h, z4.h\n"
+ "cbz x20, 5f\n"
+ "ld1h { z1.h }, p0/Z, [x22]\n"
+ "ld1rqh { z7.h }, p0/Z, [%x[Apanel]]\n"
+ ".inst 0x6461e4e8 // bfmmla z8.s, z7.h, z1.h\n"
+ "ld1rqh { z6.h }, p0/Z, [%x[Apanel], #16]\n"
+ "ld1h { z0.h }, p0/Z, [x22, #1, MUL VL]\n"
+ ".inst 0x6460e4eb // bfmmla z11.s, z7.h, z0.h\n"
+ "ld1rqh { z5.h }, p0/Z, [%x[Apanel], #32]\n"
+ "ld1rqh { z4.h }, p0/Z, [%x[Apanel], #48]\n"
+ ".inst 0x6461e4ce // bfmmla z14.s, z6.h, z1.h\n"
+ ".inst 0x6460e4d1 // bfmmla z17.s, z6.h, z0.h\n"
+ ".inst 0x6461e4b4 // bfmmla z20.s, z5.h, z1.h\n"
+ "ld1h { z3.h }, p0/Z, [x22, #2, MUL VL]\n"
+ ".inst 0x6460e4b7 // bfmmla z23.s, z5.h, z0.h\n"
+ ".inst 0x6461e49a // bfmmla z26.s, z4.h, z1.h\n"
+ "ld1h { z2.h }, p0/Z, [x22, #3, MUL VL]\n"
+ ".inst 0x6460e49d // bfmmla z29.s, z4.h, z0.h\n"
+ "ld1h { z1.h }, p0/Z, [x22, #4, MUL VL]\n"
+ "ld1h { z0.h }, p0/Z, [x22, #5, MUL VL]\n"
+ ".inst 0x6463e4e9 // bfmmla z9.s, z7.h, z3.h\n"
+ ".inst 0x6462e4ec // bfmmla z12.s, z7.h, z2.h\n"
+ "addvl x22, x22, #6\n"
+ ".inst 0x6463e4cf // bfmmla z15.s, z6.h, z3.h\n"
+ ".inst 0x6462e4d2 // bfmmla z18.s, z6.h, z2.h\n"
+ "add %x[Apanel], %x[Apanel], #0x40\n"
+ ".inst 0x6463e4b5 // bfmmla z21.s, z5.h, z3.h\n"
+ ".inst 0x6462e4b8 // bfmmla z24.s, z5.h, z2.h\n"
+ ".inst 0x6463e49b // bfmmla z27.s, z4.h, z3.h\n"
+ ".inst 0x6462e49e // bfmmla z30.s, z4.h, z2.h\n"
+ ".inst 0x6461e4ea // bfmmla z10.s, z7.h, z1.h\n"
+ ".inst 0x6460e4ed // bfmmla z13.s, z7.h, z0.h\n"
+ ".inst 0x6461e4d0 // bfmmla z16.s, z6.h, z1.h\n"
+ ".inst 0x6460e4d3 // bfmmla z19.s, z6.h, z0.h\n"
+ ".inst 0x6461e4b6 // bfmmla z22.s, z5.h, z1.h\n"
+ ".inst 0x6460e4b9 // bfmmla z25.s, z5.h, z0.h\n"
+ ".inst 0x6461e49c // bfmmla z28.s, z4.h, z1.h\n"
+ ".inst 0x6460e49f // bfmmla z31.s, z4.h, z0.h\n"
+ "5:" // multiply loop done
+ "uzp1 z0.d, z8.d, z11.d\n"
+ "uzp2 z8.d, z8.d, z11.d\n"
+ "st1w { z0.s }, p0, [%x[Cpanel]]\n"
+ "uzp1 z0.d, z9.d, z12.d\n"
+ "uzp2 z9.d, z9.d, z12.d\n"
+ "st1w { z0.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
+ "uzp1 z0.d, z10.d, z13.d\n"
+ "uzp2 z10.d, z10.d, z13.d\n"
+ "st1w { z0.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+ "st1w { z8.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+ "uzp1 z0.d, z14.d, z17.d\n"
+ "uzp2 z14.d, z14.d, z17.d\n"
+ "st1w { z9.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+ "uzp1 z1.d, z15.d, z18.d\n"
+ "subs x23, x23, #0x1\n"
+ "st1w { z10.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
+ "uzp2 z15.d, z15.d, z18.d\n"
+ "uzp1 z17.d, z16.d, z19.d\n"
+ "st1w { z0.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
+ "uzp2 z16.d, z16.d, z19.d\n"
+ "uzp1 z0.d, z20.d, z23.d\n"
+ "st1w { z1.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
+ "addvl %x[Cpanel], %x[Cpanel], #16\n"
+ "uzp2 z20.d, z20.d, z23.d\n"
+ "st1w { z17.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
+ "uzp1 z23.d, z21.d, z24.d\n"
+ "uzp2 z21.d, z21.d, z24.d\n"
+ "st1w { z14.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
+ "uzp1 z19.d, z22.d, z25.d\n"
+ "uzp2 z22.d, z22.d, z25.d\n"
+ "st1w { z15.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
+ "uzp1 z18.d, z26.d, z29.d\n"
+ "uzp2 z26.d, z26.d, z29.d\n"
+ "st1w { z16.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
+ "uzp1 z17.d, z27.d, z30.d\n"
+ "uzp2 z27.d, z27.d, z30.d\n"
+ "st1w { z0.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
+ "uzp1 z16.d, z28.d, z31.d\n"
+ "uzp2 z28.d, z28.d, z31.d\n"
+ "st1w { z23.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
+ "st1w { z19.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
+ "st1w { z20.s }, p0, [%x[Cpanel], #-1, MUL VL]\n"
+ "st1w { z21.s }, p0, [%x[Cpanel]]\n"
+ "st1w { z22.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
+ "st1w { z18.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+ "st1w { z17.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+ "st1w { z16.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+ "st1w { z26.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
+ "st1w { z27.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
+ "st1w { z28.s }, 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", "x20", "x21", "x22", "x23", "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_FEATURE_SVE
+#endif // ARM_COMPUTE_ENABLE_SVE
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-03 15:30:01 +0000