aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp513
1 files changed, 217 insertions, 296 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp
index adee900337..e507bc5551 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021, 2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -21,309 +21,230 @@
* 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_dot_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_dot_8x3VL(
+ const bfloat16 *Apanel,
+ const bfloat16 *Bpanel,
+ float *Cpanel,
+ int ablocks,
+ int bblocks,
+ int K) {
- K /= 2;
- 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/2) - 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"
- "add %[a_ptr], %[a_ptr], #0x40\n"
- "mov z17.s, #0\n"
- "addvl %[b_ptr], %[b_ptr], #3\n"
- "mov z18.s, #0\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 0x64604088 // bfdot z8.s, z4.h, z0.h[0]\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x64684089 // bfdot z9.s, z4.h, z0.h[1]\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #-0x10]\n"
- ".inst 0x6470408a // bfdot z10.s, z4.h, z0.h[2]\n"
- "subs %[loops], %[loops], #0x1\n"
- ".inst 0x6478408b // bfdot z11.s, z4.h, z0.h[3]\n"
- ".inst 0x64614094 // bfdot z20.s, z4.h, z1.h[0]\n"
- ".inst 0x64694095 // bfdot z21.s, z4.h, z1.h[1]\n"
- ".inst 0x64714096 // bfdot z22.s, z4.h, z1.h[2]\n"
- ".inst 0x64794097 // bfdot z23.s, z4.h, z1.h[3]\n"
- "ld1h z4.h, p0/z, [%[b_ptr]]\n"
- ".inst 0x646040ac // bfdot z12.s, z5.h, z0.h[0]\n"
- ".inst 0x646840ad // bfdot z13.s, z5.h, z0.h[1]\n"
- ".inst 0x647040ae // bfdot z14.s, z5.h, z0.h[2]\n"
- ".inst 0x647840af // bfdot z15.s, z5.h, z0.h[3]\n"
- ".inst 0x646140b8 // bfdot z24.s, z5.h, z1.h[0]\n"
- ".inst 0x646940b9 // bfdot z25.s, z5.h, z1.h[1]\n"
- ".inst 0x647140ba // bfdot z26.s, z5.h, z1.h[2]\n"
- ".inst 0x647940bb // bfdot z27.s, z5.h, z1.h[3]\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #1, MUL VL]\n"
- ".inst 0x646040d0 // bfdot z16.s, z6.h, z0.h[0]\n"
- ".inst 0x646840d1 // bfdot z17.s, z6.h, z0.h[1]\n"
- ".inst 0x647040d2 // bfdot z18.s, z6.h, z0.h[2]\n"
- ".inst 0x647840d3 // bfdot z19.s, z6.h, z0.h[3]\n"
- "ld1rqh z0.h, p0/z, [%[a_ptr]]\n"
- ".inst 0x646140dc // bfdot z28.s, z6.h, z1.h[0]\n"
- ".inst 0x646940dd // bfdot z29.s, z6.h, z1.h[1]\n"
- ".inst 0x647140de // bfdot z30.s, z6.h, z1.h[2]\n"
- ".inst 0x647940df // bfdot z31.s, z6.h, z1.h[3]\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #2, MUL VL]\n"
- ".inst 0x64624088 // bfdot z8.s, z4.h, z2.h[0]\n"
- "ld1rqh z1.h, p0/z, [%[a_ptr], #0x10]\n"
- ".inst 0x646a4089 // bfdot z9.s, z4.h, z2.h[1]\n"
- "add %[a_ptr], %[a_ptr], #0x40\n"
- ".inst 0x6472408a // bfdot z10.s, z4.h, z2.h[2]\n"
- "addvl %[b_ptr], %[b_ptr], #6\n"
- ".inst 0x647a408b // bfdot z11.s, z4.h, z2.h[3]\n"
- ".inst 0x64634094 // bfdot z20.s, z4.h, z3.h[0]\n"
- ".inst 0x646b4095 // bfdot z21.s, z4.h, z3.h[1]\n"
- ".inst 0x64734096 // bfdot z22.s, z4.h, z3.h[2]\n"
- ".inst 0x647b4097 // bfdot z23.s, z4.h, z3.h[3]\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #-3, MUL VL]\n"
- ".inst 0x646240ac // bfdot z12.s, z5.h, z2.h[0]\n"
- ".inst 0x646a40ad // bfdot z13.s, z5.h, z2.h[1]\n"
- ".inst 0x647240ae // bfdot z14.s, z5.h, z2.h[2]\n"
- ".inst 0x647a40af // bfdot z15.s, z5.h, z2.h[3]\n"
- ".inst 0x646340b8 // bfdot z24.s, z5.h, z3.h[0]\n"
- ".inst 0x646b40b9 // bfdot z25.s, z5.h, z3.h[1]\n"
- ".inst 0x647340ba // bfdot z26.s, z5.h, z3.h[2]\n"
- ".inst 0x647b40bb // bfdot z27.s, z5.h, z3.h[3]\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #-2, MUL VL]\n"
- ".inst 0x646240d0 // bfdot z16.s, z6.h, z2.h[0]\n"
- ".inst 0x646a40d1 // bfdot z17.s, z6.h, z2.h[1]\n"
- ".inst 0x647240d2 // bfdot z18.s, z6.h, z2.h[2]\n"
- ".inst 0x647a40d3 // bfdot z19.s, z6.h, z2.h[3]\n"
- "ld1rqh z2.h, p0/z, [%[a_ptr], #-0x20]\n"
- ".inst 0x646340dc // bfdot z28.s, z6.h, z3.h[0]\n"
- ".inst 0x646b40dd // bfdot z29.s, z6.h, z3.h[1]\n"
- ".inst 0x647340de // bfdot z30.s, z6.h, z3.h[2]\n"
- ".inst 0x647b40df // bfdot z31.s, z6.h, z3.h[3]\n"
- "b.ne 2b\n"
- "1:\n"
- "cbz %[tails], 3f\n"
- ".inst 0x64604088 // bfdot z8.s, z4.h, z0.h[0]\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x64684089 // bfdot z9.s, z4.h, z0.h[1]\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #-0x10]\n"
- ".inst 0x6470408a // bfdot z10.s, z4.h, z0.h[2]\n"
- ".inst 0x6478408b // bfdot z11.s, z4.h, z0.h[3]\n"
- ".inst 0x64614094 // bfdot z20.s, z4.h, z1.h[0]\n"
- ".inst 0x64694095 // bfdot z21.s, z4.h, z1.h[1]\n"
- ".inst 0x64714096 // bfdot z22.s, z4.h, z1.h[2]\n"
- ".inst 0x64794097 // bfdot z23.s, z4.h, z1.h[3]\n"
- "ld1h z4.h, p0/z, [%[b_ptr]]\n"
- ".inst 0x646040ac // bfdot z12.s, z5.h, z0.h[0]\n"
- ".inst 0x646840ad // bfdot z13.s, z5.h, z0.h[1]\n"
- ".inst 0x647040ae // bfdot z14.s, z5.h, z0.h[2]\n"
- ".inst 0x647840af // bfdot z15.s, z5.h, z0.h[3]\n"
- ".inst 0x646140b8 // bfdot z24.s, z5.h, z1.h[0]\n"
- ".inst 0x646940b9 // bfdot z25.s, z5.h, z1.h[1]\n"
- ".inst 0x647140ba // bfdot z26.s, z5.h, z1.h[2]\n"
- ".inst 0x647940bb // bfdot z27.s, z5.h, z1.h[3]\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #1, MUL VL]\n"
- ".inst 0x646040d0 // bfdot z16.s, z6.h, z0.h[0]\n"
- ".inst 0x646840d1 // bfdot z17.s, z6.h, z0.h[1]\n"
- ".inst 0x647040d2 // bfdot z18.s, z6.h, z0.h[2]\n"
- ".inst 0x647840d3 // bfdot z19.s, z6.h, z0.h[3]\n"
- "ld1rqh z0.h, p0/z, [%[a_ptr]]\n"
- ".inst 0x646140dc // bfdot z28.s, z6.h, z1.h[0]\n"
- ".inst 0x646940dd // bfdot z29.s, z6.h, z1.h[1]\n"
- ".inst 0x647140de // bfdot z30.s, z6.h, z1.h[2]\n"
- ".inst 0x647940df // bfdot z31.s, z6.h, z1.h[3]\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #2, MUL VL]\n"
- ".inst 0x64624088 // bfdot z8.s, z4.h, z2.h[0]\n"
- "ld1rqh z1.h, p0/z, [%[a_ptr], #0x10]\n"
- ".inst 0x646a4089 // bfdot z9.s, z4.h, z2.h[1]\n"
- "add %[a_ptr], %[a_ptr], #0x20\n"
- ".inst 0x6472408a // bfdot z10.s, z4.h, z2.h[2]\n"
- "addvl %[b_ptr], %[b_ptr], #6\n"
- ".inst 0x647a408b // bfdot z11.s, z4.h, z2.h[3]\n"
- ".inst 0x64634094 // bfdot z20.s, z4.h, z3.h[0]\n"
- ".inst 0x646b4095 // bfdot z21.s, z4.h, z3.h[1]\n"
- ".inst 0x64734096 // bfdot z22.s, z4.h, z3.h[2]\n"
- ".inst 0x647b4097 // bfdot z23.s, z4.h, z3.h[3]\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #-3, MUL VL]\n"
- ".inst 0x646240ac // bfdot z12.s, z5.h, z2.h[0]\n"
- ".inst 0x646a40ad // bfdot z13.s, z5.h, z2.h[1]\n"
- ".inst 0x647240ae // bfdot z14.s, z5.h, z2.h[2]\n"
- ".inst 0x647a40af // bfdot z15.s, z5.h, z2.h[3]\n"
- ".inst 0x646340b8 // bfdot z24.s, z5.h, z3.h[0]\n"
- ".inst 0x646b40b9 // bfdot z25.s, z5.h, z3.h[1]\n"
- ".inst 0x647340ba // bfdot z26.s, z5.h, z3.h[2]\n"
- ".inst 0x647b40bb // bfdot z27.s, z5.h, z3.h[3]\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #-2, MUL VL]\n"
- ".inst 0x646240d0 // bfdot z16.s, z6.h, z2.h[0]\n"
- ".inst 0x646a40d1 // bfdot z17.s, z6.h, z2.h[1]\n"
- ".inst 0x647240d2 // bfdot z18.s, z6.h, z2.h[2]\n"
- ".inst 0x647a40d3 // bfdot z19.s, z6.h, z2.h[3]\n"
- ".inst 0x646340dc // bfdot z28.s, z6.h, z3.h[0]\n"
- ".inst 0x646b40dd // bfdot z29.s, z6.h, z3.h[1]\n"
- ".inst 0x647340de // bfdot z30.s, z6.h, z3.h[2]\n"
- ".inst 0x647b40df // bfdot z31.s, z6.h, z3.h[3]\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x64604088 // bfdot z8.s, z4.h, z0.h[0]\n"
- ".inst 0x64684089 // bfdot z9.s, z4.h, z0.h[1]\n"
- ".inst 0x6470408a // bfdot z10.s, z4.h, z0.h[2]\n"
- ".inst 0x6478408b // bfdot z11.s, z4.h, z0.h[3]\n"
- "st1w z8.s, p0, [%[c_ptr]]\n"
- ".inst 0x64614094 // bfdot z20.s, z4.h, z1.h[0]\n"
- ".inst 0x64694095 // bfdot z21.s, z4.h, z1.h[1]\n"
- ".inst 0x64714096 // bfdot z22.s, z4.h, z1.h[2]\n"
- ".inst 0x64794097 // bfdot z23.s, z4.h, z1.h[3]\n"
- ".inst 0x646040ac // bfdot z12.s, z5.h, z0.h[0]\n"
- ".inst 0x646840ad // bfdot z13.s, z5.h, z0.h[1]\n"
- ".inst 0x647040ae // bfdot z14.s, z5.h, z0.h[2]\n"
- ".inst 0x647840af // bfdot z15.s, z5.h, z0.h[3]\n"
- "st1w z12.s, p0, [%[c_ptr], #1, MUL VL]\n"
- ".inst 0x646140b8 // bfdot z24.s, z5.h, z1.h[0]\n"
- ".inst 0x646940b9 // bfdot z25.s, z5.h, z1.h[1]\n"
- ".inst 0x647140ba // bfdot z26.s, z5.h, z1.h[2]\n"
- ".inst 0x647940bb // bfdot z27.s, z5.h, z1.h[3]\n"
- ".inst 0x646040d0 // bfdot z16.s, z6.h, z0.h[0]\n"
- ".inst 0x646840d1 // bfdot z17.s, z6.h, z0.h[1]\n"
- ".inst 0x647040d2 // bfdot z18.s, z6.h, z0.h[2]\n"
- ".inst 0x647840d3 // bfdot z19.s, z6.h, z0.h[3]\n"
- "st1w z16.s, p0, [%[c_ptr], #2, MUL VL]\n"
- ".inst 0x646140dc // bfdot z28.s, z6.h, z1.h[0]\n"
- ".inst 0x646940dd // bfdot z29.s, z6.h, z1.h[1]\n"
- ".inst 0x647140de // bfdot z30.s, z6.h, z1.h[2]\n"
- "st1w z9.s, p0, [%[c_ptr], #3, MUL VL]\n"
- ".inst 0x647940df // bfdot z31.s, z6.h, z1.h[3]\n"
- "b 4f\n"
- "3:\n"
- ".inst 0x64604088 // bfdot z8.s, z4.h, z0.h[0]\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x64684089 // bfdot z9.s, z4.h, z0.h[1]\n"
- "ld1rqh z3.h, p0/z, [%[a_ptr], #-0x10]\n"
- ".inst 0x6470408a // bfdot z10.s, z4.h, z0.h[2]\n"
- "addvl %[b_ptr], %[b_ptr], #3\n"
- ".inst 0x6478408b // bfdot z11.s, z4.h, z0.h[3]\n"
- ".inst 0x64614094 // bfdot z20.s, z4.h, z1.h[0]\n"
- ".inst 0x64694095 // bfdot z21.s, z4.h, z1.h[1]\n"
- ".inst 0x64714096 // bfdot z22.s, z4.h, z1.h[2]\n"
- ".inst 0x64794097 // bfdot z23.s, z4.h, z1.h[3]\n"
- "ld1h z4.h, p0/z, [%[b_ptr], #-3, MUL VL]\n"
- ".inst 0x646040ac // bfdot z12.s, z5.h, z0.h[0]\n"
- ".inst 0x646840ad // bfdot z13.s, z5.h, z0.h[1]\n"
- ".inst 0x647040ae // bfdot z14.s, z5.h, z0.h[2]\n"
- ".inst 0x647840af // bfdot z15.s, z5.h, z0.h[3]\n"
- ".inst 0x646140b8 // bfdot z24.s, z5.h, z1.h[0]\n"
- ".inst 0x646940b9 // bfdot z25.s, z5.h, z1.h[1]\n"
- ".inst 0x647140ba // bfdot z26.s, z5.h, z1.h[2]\n"
- ".inst 0x647940bb // bfdot z27.s, z5.h, z1.h[3]\n"
- "ld1h z5.h, p0/z, [%[b_ptr], #-2, MUL VL]\n"
- ".inst 0x646040d0 // bfdot z16.s, z6.h, z0.h[0]\n"
- ".inst 0x646840d1 // bfdot z17.s, z6.h, z0.h[1]\n"
- ".inst 0x647040d2 // bfdot z18.s, z6.h, z0.h[2]\n"
- ".inst 0x647840d3 // bfdot z19.s, z6.h, z0.h[3]\n"
- ".inst 0x646140dc // bfdot z28.s, z6.h, z1.h[0]\n"
- ".inst 0x646940dd // bfdot z29.s, z6.h, z1.h[1]\n"
- ".inst 0x647140de // bfdot z30.s, z6.h, z1.h[2]\n"
- ".inst 0x647940df // bfdot z31.s, z6.h, z1.h[3]\n"
- "ld1h z6.h, p0/z, [%[b_ptr], #-1, MUL VL]\n"
- ".inst 0x64624088 // bfdot z8.s, z4.h, z2.h[0]\n"
- ".inst 0x646a4089 // bfdot z9.s, z4.h, z2.h[1]\n"
- ".inst 0x6472408a // bfdot z10.s, z4.h, z2.h[2]\n"
- ".inst 0x647a408b // bfdot z11.s, z4.h, z2.h[3]\n"
- "st1w z8.s, p0, [%[c_ptr]]\n"
- ".inst 0x64634094 // bfdot z20.s, z4.h, z3.h[0]\n"
- ".inst 0x646b4095 // bfdot z21.s, z4.h, z3.h[1]\n"
- ".inst 0x64734096 // bfdot z22.s, z4.h, z3.h[2]\n"
- ".inst 0x647b4097 // bfdot z23.s, z4.h, z3.h[3]\n"
- ".inst 0x646240ac // bfdot z12.s, z5.h, z2.h[0]\n"
- ".inst 0x646a40ad // bfdot z13.s, z5.h, z2.h[1]\n"
- ".inst 0x647240ae // bfdot z14.s, z5.h, z2.h[2]\n"
- ".inst 0x647a40af // bfdot z15.s, z5.h, z2.h[3]\n"
- "st1w z12.s, p0, [%[c_ptr], #1, MUL VL]\n"
- ".inst 0x646340b8 // bfdot z24.s, z5.h, z3.h[0]\n"
- ".inst 0x646b40b9 // bfdot z25.s, z5.h, z3.h[1]\n"
- ".inst 0x647340ba // bfdot z26.s, z5.h, z3.h[2]\n"
- ".inst 0x647b40bb // bfdot z27.s, z5.h, z3.h[3]\n"
- ".inst 0x646240d0 // bfdot z16.s, z6.h, z2.h[0]\n"
- ".inst 0x646a40d1 // bfdot z17.s, z6.h, z2.h[1]\n"
- ".inst 0x647240d2 // bfdot z18.s, z6.h, z2.h[2]\n"
- ".inst 0x647a40d3 // bfdot z19.s, z6.h, z2.h[3]\n"
- "st1w z16.s, p0, [%[c_ptr], #2, MUL VL]\n"
- ".inst 0x646340dc // bfdot z28.s, z6.h, z3.h[0]\n"
- ".inst 0x646b40dd // bfdot z29.s, z6.h, z3.h[1]\n"
- ".inst 0x647340de // bfdot z30.s, z6.h, z3.h[2]\n"
- "st1w z9.s, p0, [%[c_ptr], #3, MUL VL]\n"
- ".inst 0x647b40df // bfdot z31.s, z6.h, z3.h[3]\n"
- "4:\n"
- "st1w z13.s, p0, [%[c_ptr], #4, MUL VL]\n"
- "st1w z17.s, p0, [%[c_ptr], #5, MUL VL]\n"
- "st1w z10.s, p0, [%[c_ptr], #6, MUL VL]\n"
- "st1w z14.s, p0, [%[c_ptr], #7, MUL VL]\n"
- "addvl %[c_ptr], %[c_ptr], #16\n"
- "st1w z18.s, p0, [%[c_ptr], #-8, MUL VL]\n"
- "st1w z11.s, p0, [%[c_ptr], #-7, MUL VL]\n"
- "st1w z15.s, p0, [%[c_ptr], #-6, MUL VL]\n"
- "st1w z19.s, p0, [%[c_ptr], #-5, MUL VL]\n"
- "st1w z20.s, p0, [%[c_ptr], #-4, MUL VL]\n"
- "st1w z24.s, p0, [%[c_ptr], #-3, MUL VL]\n"
- "st1w z28.s, p0, [%[c_ptr], #-2, MUL VL]\n"
- "st1w z21.s, p0, [%[c_ptr], #-1, MUL VL]\n"
- "st1w z25.s, p0, [%[c_ptr]]\n"
- "st1w z29.s, p0, [%[c_ptr], #1, MUL VL]\n"
- "st1w z22.s, p0, [%[c_ptr], #2, MUL VL]\n"
- "st1w z26.s, p0, [%[c_ptr], #3, MUL VL]\n"
- "st1w z30.s, p0, [%[c_ptr], #4, MUL VL]\n"
- "st1w z23.s, p0, [%[c_ptr], #5, MUL VL]\n"
- "st1w z27.s, p0, [%[c_ptr], #6, MUL VL]\n"
- "st1w z31.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", "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"
+ "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
+ "mov z10.b, #0x0\n"
+ "mov z11.b, #0x0\n"
+ "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
+ "mov z12.b, #0x0\n"
+ "mov z13.b, #0x0\n"
+ "ld1h { z4.h }, p0/Z, [x22]\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"
+ "ld1h { z6.h }, p0/Z, [x22, #2, MUL VL]\n"
+ "mov z18.b, #0x0\n"
+ "mov z19.b, #0x0\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
+ ".inst 0x64604088 // bfdot z8.s, z4.h, z0.h[0]\n"
+ ".inst 0x6468408b // bfdot z11.s, z4.h, z0.h[1]\n"
+ "ld1rqh { z3.h }, p0/Z, [%x[Apanel], #32]\n"
+ ".inst 0x6470408e // bfdot z14.s, z4.h, z0.h[2]\n"
+ ".inst 0x64784091 // bfdot z17.s, z4.h, z0.h[3]\n"
+ "ld1rqh { z7.h }, p0/Z, [%x[Apanel], #48]\n"
+ ".inst 0x64614094 // bfdot z20.s, z4.h, z1.h[0]\n"
+ ".inst 0x64694097 // bfdot z23.s, z4.h, z1.h[1]\n"
+ "sub x20, x20, #0x2\n"
+ ".inst 0x6471409a // bfdot z26.s, z4.h, z1.h[2]\n"
+ ".inst 0x6479409d // bfdot z29.s, z4.h, z1.h[3]\n"
+ "ld1h { z4.h }, p0/Z, [x22, #3, MUL VL]\n"
+ ".inst 0x646040a9 // bfdot z9.s, z5.h, z0.h[0]\n"
+ ".inst 0x646840ac // bfdot z12.s, z5.h, z0.h[1]\n"
+ "cmp x20, #0x2\n"
+ ".inst 0x647040af // bfdot z15.s, z5.h, z0.h[2]\n"
+ ".inst 0x647840b2 // bfdot z18.s, z5.h, z0.h[3]\n"
+ "add %x[Apanel], %x[Apanel], #0x40\n"
+ ".inst 0x646140b5 // bfdot z21.s, z5.h, z1.h[0]\n"
+ ".inst 0x646940b8 // bfdot z24.s, z5.h, z1.h[1]\n"
+ ".inst 0x647140bb // bfdot z27.s, z5.h, z1.h[2]\n"
+ ".inst 0x647940be // bfdot z30.s, z5.h, z1.h[3]\n"
+ "ld1h { z5.h }, p0/Z, [x22, #4, MUL VL]\n"
+ ".inst 0x646040ca // bfdot z10.s, z6.h, z0.h[0]\n"
+ ".inst 0x646840cd // bfdot z13.s, z6.h, z0.h[1]\n"
+ ".inst 0x647040d0 // bfdot z16.s, z6.h, z0.h[2]\n"
+ ".inst 0x647840d3 // bfdot z19.s, z6.h, z0.h[3]\n"
+ "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
+ ".inst 0x646140d6 // bfdot z22.s, z6.h, z1.h[0]\n"
+ ".inst 0x646940d9 // bfdot z25.s, z6.h, z1.h[1]\n"
+ ".inst 0x647140dc // bfdot z28.s, z6.h, z1.h[2]\n"
+ ".inst 0x647940df // bfdot z31.s, z6.h, z1.h[3]\n"
+ "ld1h { z2.h }, p0/Z, [x22, #5, MUL VL]\n"
+ "addvl x22, x22, #6\n"
+ ".inst 0x64634088 // bfdot z8.s, z4.h, z3.h[0]\n"
+ ".inst 0x646b408b // bfdot z11.s, z4.h, z3.h[1]\n"
+ "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
+ ".inst 0x6473408e // bfdot z14.s, z4.h, z3.h[2]\n"
+ ".inst 0x647b4091 // bfdot z17.s, z4.h, z3.h[3]\n"
+ ".inst 0x64674094 // bfdot z20.s, z4.h, z7.h[0]\n"
+ ".inst 0x646f4097 // bfdot z23.s, z4.h, z7.h[1]\n"
+ ".inst 0x6477409a // bfdot z26.s, z4.h, z7.h[2]\n"
+ ".inst 0x647f409d // bfdot z29.s, z4.h, z7.h[3]\n"
+ "ld1h { z4.h }, p0/Z, [x22]\n"
+ ".inst 0x646340a9 // bfdot z9.s, z5.h, z3.h[0]\n"
+ ".inst 0x646b40ac // bfdot z12.s, z5.h, z3.h[1]\n"
+ ".inst 0x647340af // bfdot z15.s, z5.h, z3.h[2]\n"
+ ".inst 0x647b40b2 // bfdot z18.s, z5.h, z3.h[3]\n"
+ ".inst 0x646740b5 // bfdot z21.s, z5.h, z7.h[0]\n"
+ ".inst 0x646f40b8 // bfdot z24.s, z5.h, z7.h[1]\n"
+ ".inst 0x647740bb // bfdot z27.s, z5.h, z7.h[2]\n"
+ ".inst 0x647f40be // bfdot z30.s, z5.h, z7.h[3]\n"
+ "ld1h { z5.h }, p0/Z, [x22, #1, MUL VL]\n"
+ ".inst 0x6463404a // bfdot z10.s, z2.h, z3.h[0]\n"
+ ".inst 0x646b404d // bfdot z13.s, z2.h, z3.h[1]\n"
+ ".inst 0x64734050 // bfdot z16.s, z2.h, z3.h[2]\n"
+ ".inst 0x647b4053 // bfdot z19.s, z2.h, z3.h[3]\n"
+ ".inst 0x64674056 // bfdot z22.s, z2.h, z7.h[0]\n"
+ ".inst 0x646f4059 // bfdot z25.s, z2.h, z7.h[1]\n"
+ ".inst 0x6477405c // bfdot z28.s, z2.h, z7.h[2]\n"
+ ".inst 0x647f405f // bfdot z31.s, z2.h, z7.h[3]\n"
+ "ld1h { z6.h }, p0/Z, [x22, #2, MUL VL]\n"
+ "bge 3b\n"
+ "4:" // main loop skip
+ ".inst 0x64604088 // bfdot z8.s, z4.h, z0.h[0]\n"
+ ".inst 0x6468408b // bfdot z11.s, z4.h, z0.h[1]\n"
+ "add %x[Apanel], %x[Apanel], #0x20\n"
+ ".inst 0x6470408e // bfdot z14.s, z4.h, z0.h[2]\n"
+ ".inst 0x64784091 // bfdot z17.s, z4.h, z0.h[3]\n"
+ "addvl x22, x22, #3\n"
+ ".inst 0x64614094 // bfdot z20.s, z4.h, z1.h[0]\n"
+ ".inst 0x64694097 // bfdot z23.s, z4.h, z1.h[1]\n"
+ ".inst 0x6471409a // bfdot z26.s, z4.h, z1.h[2]\n"
+ ".inst 0x6479409d // bfdot z29.s, z4.h, z1.h[3]\n"
+ ".inst 0x646040a9 // bfdot z9.s, z5.h, z0.h[0]\n"
+ ".inst 0x646840ac // bfdot z12.s, z5.h, z0.h[1]\n"
+ ".inst 0x647040af // bfdot z15.s, z5.h, z0.h[2]\n"
+ ".inst 0x647840b2 // bfdot z18.s, z5.h, z0.h[3]\n"
+ ".inst 0x646140b5 // bfdot z21.s, z5.h, z1.h[0]\n"
+ ".inst 0x646940b8 // bfdot z24.s, z5.h, z1.h[1]\n"
+ ".inst 0x647140bb // bfdot z27.s, z5.h, z1.h[2]\n"
+ ".inst 0x647940be // bfdot z30.s, z5.h, z1.h[3]\n"
+ ".inst 0x646040ca // bfdot z10.s, z6.h, z0.h[0]\n"
+ ".inst 0x646840cd // bfdot z13.s, z6.h, z0.h[1]\n"
+ ".inst 0x647040d0 // bfdot z16.s, z6.h, z0.h[2]\n"
+ ".inst 0x647840d3 // bfdot z19.s, z6.h, z0.h[3]\n"
+ ".inst 0x646140d6 // bfdot z22.s, z6.h, z1.h[0]\n"
+ ".inst 0x646940d9 // bfdot z25.s, z6.h, z1.h[1]\n"
+ ".inst 0x647140dc // bfdot z28.s, z6.h, z1.h[2]\n"
+ ".inst 0x647940df // bfdot z31.s, z6.h, z1.h[3]\n"
+ "cbz x20, 5f\n"
+ "ld1rqh { z4.h }, p0/Z, [%x[Apanel]]\n"
+ "ld1rqh { z3.h }, p0/Z, [%x[Apanel], #16]\n"
+ "add %x[Apanel], %x[Apanel], #0x20\n"
+ "ld1h { z2.h }, p0/Z, [x22]\n"
+ "ld1h { z1.h }, p0/Z, [x22, #1, MUL VL]\n"
+ ".inst 0x64644048 // bfdot z8.s, z2.h, z4.h[0]\n"
+ "ld1h { z0.h }, p0/Z, [x22, #2, MUL VL]\n"
+ ".inst 0x646c404b // bfdot z11.s, z2.h, z4.h[1]\n"
+ ".inst 0x6474404e // bfdot z14.s, z2.h, z4.h[2]\n"
+ ".inst 0x647c4051 // bfdot z17.s, z2.h, z4.h[3]\n"
+ ".inst 0x64634054 // bfdot z20.s, z2.h, z3.h[0]\n"
+ "addvl x22, x22, #3\n"
+ ".inst 0x646b4057 // bfdot z23.s, z2.h, z3.h[1]\n"
+ ".inst 0x6473405a // bfdot z26.s, z2.h, z3.h[2]\n"
+ ".inst 0x647b405d // bfdot z29.s, z2.h, z3.h[3]\n"
+ ".inst 0x64644029 // bfdot z9.s, z1.h, z4.h[0]\n"
+ ".inst 0x646c402c // bfdot z12.s, z1.h, z4.h[1]\n"
+ ".inst 0x6474402f // bfdot z15.s, z1.h, z4.h[2]\n"
+ ".inst 0x647c4032 // bfdot z18.s, z1.h, z4.h[3]\n"
+ ".inst 0x64634035 // bfdot z21.s, z1.h, z3.h[0]\n"
+ ".inst 0x646b4038 // bfdot z24.s, z1.h, z3.h[1]\n"
+ ".inst 0x6473403b // bfdot z27.s, z1.h, z3.h[2]\n"
+ ".inst 0x647b403e // bfdot z30.s, z1.h, z3.h[3]\n"
+ ".inst 0x6464400a // bfdot z10.s, z0.h, z4.h[0]\n"
+ ".inst 0x646c400d // bfdot z13.s, z0.h, z4.h[1]\n"
+ ".inst 0x64744010 // bfdot z16.s, z0.h, z4.h[2]\n"
+ ".inst 0x647c4013 // bfdot z19.s, z0.h, z4.h[3]\n"
+ ".inst 0x64634016 // bfdot z22.s, z0.h, z3.h[0]\n"
+ ".inst 0x646b4019 // bfdot z25.s, z0.h, z3.h[1]\n"
+ ".inst 0x6473401c // bfdot z28.s, z0.h, z3.h[2]\n"
+ ".inst 0x647b401f // bfdot z31.s, z0.h, z3.h[3]\n"
+ "5:" // multiply loop done
+ "st1w { z8.s }, p0, [%x[Cpanel]]\n"
+ "subs x23, x23, #0x1\n"
+ "st1w { z9.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
+ "st1w { z10.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+ "st1w { z11.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+ "st1w { z12.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+ "st1w { z13.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
+ "st1w { z14.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
+ "st1w { z15.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
+ "addvl %x[Cpanel], %x[Cpanel], #16\n"
+ "st1w { z16.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
+ "st1w { z17.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
+ "st1w { z18.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
+ "st1w { z19.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
+ "st1w { z20.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
+ "st1w { z21.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
+ "st1w { z22.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
+ "st1w { z23.s }, p0, [%x[Cpanel], #-1, MUL VL]\n"
+ "st1w { z24.s }, p0, [%x[Cpanel]]\n"
+ "st1w { z25.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
+ "st1w { z26.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+ "st1w { z27.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+ "st1w { z28.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+ "st1w { z29.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
+ "st1w { z30.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
+ "st1w { z31.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 23:27:10 +0000