diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp | 513 |
1 files changed, 217 insertions, 296 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp index 1e05a308b5..c7f32ff7a9 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,308 +21,229 @@ * 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 "../../asmlib.hpp" +#include <cstddef> namespace arm_gemm { -void sve_interleaved_fp32_mla_8x3VL(const float *Apanel, const float *Bpanel, float *Cpanel, int ablocks, int bblocks, int K) { - const float *a_ptr = Apanel; - float *c_ptr = Cpanel; - - const long loops_count = (K / 2) - 1; - const long tails_count = K % 2; +void sve_interleaved_fp32_mla_8x3VL( + const float *Apanel, + const float *Bpanel, + float *Cpanel, + int ablocks, + int bblocks, + int K) { - for (int yb=0; yb<ablocks; yb++) { - const float *a_ptr0 = a_ptr; - const float *b_ptr = Bpanel; + struct KernelArgs { + size_t K = {}; + const float *Bpanel = {}; + size_t bblocks = {}; + } ka; - for (int xb=0; xb<bblocks; xb++) { - a_ptr = a_ptr0; - long loops = loops_count; - long tails = tails_count; + ka.K = (K/1) - 1; + ka.Bpanel = Bpanel; + ka.bblocks = bblocks; - __asm __volatile ( - "mov z8.s, #0\n" - "ptrue p0.s\n" - "mov z9.s, #0\n" - "mov z10.s, #0\n" - "mov z11.s, #0\n" - "ld1rqw z0.s, p0/z, [%[a_ptr]]\n" - "mov z12.s, #0\n" - "ld1w z4.s, p0/z, [%[b_ptr]]\n" - "mov z13.s, #0\n" - "ld1rqw z1.s, p0/z, [%[a_ptr], #0x10]\n" - "mov z14.s, #0\n" - "ld1w z5.s, p0/z, [%[b_ptr], #1, MUL VL]\n" - "mov z15.s, #0\n" - "ld1rqw z2.s, 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" - "fmla z8.s, z4.s, z0.s[0]\n" - "ld1w z6.s, p0/z, [%[b_ptr], #-1, MUL VL]\n" - "fmla z9.s, z4.s, z0.s[1]\n" - "ld1rqw z3.s, p0/z, [%[a_ptr], #-0x10]\n" - "fmla z10.s, z4.s, z0.s[2]\n" - "subs %[loops], %[loops], #0x1\n" - "fmla z11.s, z4.s, z0.s[3]\n" - "fmla z20.s, z4.s, z1.s[0]\n" - "fmla z21.s, z4.s, z1.s[1]\n" - "fmla z22.s, z4.s, z1.s[2]\n" - "fmla z23.s, z4.s, z1.s[3]\n" - "ld1w z4.s, p0/z, [%[b_ptr]]\n" - "fmla z12.s, z5.s, z0.s[0]\n" - "fmla z13.s, z5.s, z0.s[1]\n" - "fmla z14.s, z5.s, z0.s[2]\n" - "fmla z15.s, z5.s, z0.s[3]\n" - "fmla z24.s, z5.s, z1.s[0]\n" - "fmla z25.s, z5.s, z1.s[1]\n" - "fmla z26.s, z5.s, z1.s[2]\n" - "fmla z27.s, z5.s, z1.s[3]\n" - "ld1w z5.s, p0/z, [%[b_ptr], #1, MUL VL]\n" - "fmla z16.s, z6.s, z0.s[0]\n" - "fmla z17.s, z6.s, z0.s[1]\n" - "fmla z18.s, z6.s, z0.s[2]\n" - "fmla z19.s, z6.s, z0.s[3]\n" - "ld1rqw z0.s, p0/z, [%[a_ptr]]\n" - "fmla z28.s, z6.s, z1.s[0]\n" - "fmla z29.s, z6.s, z1.s[1]\n" - "fmla z30.s, z6.s, z1.s[2]\n" - "fmla z31.s, z6.s, z1.s[3]\n" - "ld1w z6.s, p0/z, [%[b_ptr], #2, MUL VL]\n" - "fmla z8.s, z4.s, z2.s[0]\n" - "ld1rqw z1.s, p0/z, [%[a_ptr], #0x10]\n" - "fmla z9.s, z4.s, z2.s[1]\n" - "add %[a_ptr], %[a_ptr], #0x40\n" - "fmla z10.s, z4.s, z2.s[2]\n" - "addvl %[b_ptr], %[b_ptr], #6\n" - "fmla z11.s, z4.s, z2.s[3]\n" - "fmla z20.s, z4.s, z3.s[0]\n" - "fmla z21.s, z4.s, z3.s[1]\n" - "fmla z22.s, z4.s, z3.s[2]\n" - "fmla z23.s, z4.s, z3.s[3]\n" - "ld1w z4.s, p0/z, [%[b_ptr], #-3, MUL VL]\n" - "fmla z12.s, z5.s, z2.s[0]\n" - "fmla z13.s, z5.s, z2.s[1]\n" - "fmla z14.s, z5.s, z2.s[2]\n" - "fmla z15.s, z5.s, z2.s[3]\n" - "fmla z24.s, z5.s, z3.s[0]\n" - "fmla z25.s, z5.s, z3.s[1]\n" - "fmla z26.s, z5.s, z3.s[2]\n" - "fmla z27.s, z5.s, z3.s[3]\n" - "ld1w z5.s, p0/z, [%[b_ptr], #-2, MUL VL]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z17.s, z6.s, z2.s[1]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z19.s, z6.s, z2.s[3]\n" - "ld1rqw z2.s, p0/z, [%[a_ptr], #-0x20]\n" - "fmla z28.s, z6.s, z3.s[0]\n" - "fmla z29.s, z6.s, z3.s[1]\n" - "fmla z30.s, z6.s, z3.s[2]\n" - "fmla z31.s, z6.s, z3.s[3]\n" - "b.ne 2b\n" - "1:\n" - "cbz %[tails], 3f\n" - "fmla z8.s, z4.s, z0.s[0]\n" - "ld1w z6.s, p0/z, [%[b_ptr], #-1, MUL VL]\n" - "fmla z9.s, z4.s, z0.s[1]\n" - "ld1rqw z3.s, p0/z, [%[a_ptr], #-0x10]\n" - "fmla z10.s, z4.s, z0.s[2]\n" - "fmla z11.s, z4.s, z0.s[3]\n" - "fmla z20.s, z4.s, z1.s[0]\n" - "fmla z21.s, z4.s, z1.s[1]\n" - "fmla z22.s, z4.s, z1.s[2]\n" - "fmla z23.s, z4.s, z1.s[3]\n" - "ld1w z4.s, p0/z, [%[b_ptr]]\n" - "fmla z12.s, z5.s, z0.s[0]\n" - "fmla z13.s, z5.s, z0.s[1]\n" - "fmla z14.s, z5.s, z0.s[2]\n" - "fmla z15.s, z5.s, z0.s[3]\n" - "fmla z24.s, z5.s, z1.s[0]\n" - "fmla z25.s, z5.s, z1.s[1]\n" - "fmla z26.s, z5.s, z1.s[2]\n" - "fmla z27.s, z5.s, z1.s[3]\n" - "ld1w z5.s, p0/z, [%[b_ptr], #1, MUL VL]\n" - "fmla z16.s, z6.s, z0.s[0]\n" - "fmla z17.s, z6.s, z0.s[1]\n" - "fmla z18.s, z6.s, z0.s[2]\n" - "fmla z19.s, z6.s, z0.s[3]\n" - "ld1rqw z0.s, p0/z, [%[a_ptr]]\n" - "fmla z28.s, z6.s, z1.s[0]\n" - "fmla z29.s, z6.s, z1.s[1]\n" - "fmla z30.s, z6.s, z1.s[2]\n" - "fmla z31.s, z6.s, z1.s[3]\n" - "ld1w z6.s, p0/z, [%[b_ptr], #2, MUL VL]\n" - "fmla z8.s, z4.s, z2.s[0]\n" - "ld1rqw z1.s, p0/z, [%[a_ptr], #0x10]\n" - "fmla z9.s, z4.s, z2.s[1]\n" - "add %[a_ptr], %[a_ptr], #0x20\n" - "fmla z10.s, z4.s, z2.s[2]\n" - "addvl %[b_ptr], %[b_ptr], #6\n" - "fmla z11.s, z4.s, z2.s[3]\n" - "fmla z20.s, z4.s, z3.s[0]\n" - "fmla z21.s, z4.s, z3.s[1]\n" - "fmla z22.s, z4.s, z3.s[2]\n" - "fmla z23.s, z4.s, z3.s[3]\n" - "ld1w z4.s, p0/z, [%[b_ptr], #-3, MUL VL]\n" - "fmla z12.s, z5.s, z2.s[0]\n" - "fmla z13.s, z5.s, z2.s[1]\n" - "fmla z14.s, z5.s, z2.s[2]\n" - "fmla z15.s, z5.s, z2.s[3]\n" - "fmla z24.s, z5.s, z3.s[0]\n" - "fmla z25.s, z5.s, z3.s[1]\n" - "fmla z26.s, z5.s, z3.s[2]\n" - "fmla z27.s, z5.s, z3.s[3]\n" - "ld1w z5.s, p0/z, [%[b_ptr], #-2, MUL VL]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z17.s, z6.s, z2.s[1]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z19.s, z6.s, z2.s[3]\n" - "fmla z28.s, z6.s, z3.s[0]\n" - "fmla z29.s, z6.s, z3.s[1]\n" - "fmla z30.s, z6.s, z3.s[2]\n" - "fmla z31.s, z6.s, z3.s[3]\n" - "ld1w z6.s, p0/z, [%[b_ptr], #-1, MUL VL]\n" - "fmla z8.s, z4.s, z0.s[0]\n" - "fmla z9.s, z4.s, z0.s[1]\n" - "fmla z10.s, z4.s, z0.s[2]\n" - "fmla z11.s, z4.s, z0.s[3]\n" - "st1w z8.s, p0, [%[c_ptr]]\n" - "fmla z20.s, z4.s, z1.s[0]\n" - "fmla z21.s, z4.s, z1.s[1]\n" - "fmla z22.s, z4.s, z1.s[2]\n" - "fmla z23.s, z4.s, z1.s[3]\n" - "fmla z12.s, z5.s, z0.s[0]\n" - "fmla z13.s, z5.s, z0.s[1]\n" - "fmla z14.s, z5.s, z0.s[2]\n" - "fmla z15.s, z5.s, z0.s[3]\n" - "st1w z12.s, p0, [%[c_ptr], #1, MUL VL]\n" - "fmla z24.s, z5.s, z1.s[0]\n" - "fmla z25.s, z5.s, z1.s[1]\n" - "fmla z26.s, z5.s, z1.s[2]\n" - "fmla z27.s, z5.s, z1.s[3]\n" - "fmla z16.s, z6.s, z0.s[0]\n" - "fmla z17.s, z6.s, z0.s[1]\n" - "fmla z18.s, z6.s, z0.s[2]\n" - "fmla z19.s, z6.s, z0.s[3]\n" - "st1w z16.s, p0, [%[c_ptr], #2, MUL VL]\n" - "fmla z28.s, z6.s, z1.s[0]\n" - "fmla z29.s, z6.s, z1.s[1]\n" - "fmla z30.s, z6.s, z1.s[2]\n" - "st1w z9.s, p0, [%[c_ptr], #3, MUL VL]\n" - "fmla z31.s, z6.s, z1.s[3]\n" - "b 4f\n" - "3:\n" - "fmla z8.s, z4.s, z0.s[0]\n" - "ld1w z6.s, p0/z, [%[b_ptr], #-1, MUL VL]\n" - "fmla z9.s, z4.s, z0.s[1]\n" - "ld1rqw z3.s, p0/z, [%[a_ptr], #-0x10]\n" - "fmla z10.s, z4.s, z0.s[2]\n" - "addvl %[b_ptr], %[b_ptr], #3\n" - "fmla z11.s, z4.s, z0.s[3]\n" - "fmla z20.s, z4.s, z1.s[0]\n" - "fmla z21.s, z4.s, z1.s[1]\n" - "fmla z22.s, z4.s, z1.s[2]\n" - "fmla z23.s, z4.s, z1.s[3]\n" - "ld1w z4.s, p0/z, [%[b_ptr], #-3, MUL VL]\n" - "fmla z12.s, z5.s, z0.s[0]\n" - "fmla z13.s, z5.s, z0.s[1]\n" - "fmla z14.s, z5.s, z0.s[2]\n" - "fmla z15.s, z5.s, z0.s[3]\n" - "fmla z24.s, z5.s, z1.s[0]\n" - "fmla z25.s, z5.s, z1.s[1]\n" - "fmla z26.s, z5.s, z1.s[2]\n" - "fmla z27.s, z5.s, z1.s[3]\n" - "ld1w z5.s, p0/z, [%[b_ptr], #-2, MUL VL]\n" - "fmla z16.s, z6.s, z0.s[0]\n" - "fmla z17.s, z6.s, z0.s[1]\n" - "fmla z18.s, z6.s, z0.s[2]\n" - "fmla z19.s, z6.s, z0.s[3]\n" - "fmla z28.s, z6.s, z1.s[0]\n" - "fmla z29.s, z6.s, z1.s[1]\n" - "fmla z30.s, z6.s, z1.s[2]\n" - "fmla z31.s, z6.s, z1.s[3]\n" - "ld1w z6.s, p0/z, [%[b_ptr], #-1, MUL VL]\n" - "fmla z8.s, z4.s, z2.s[0]\n" - "fmla z9.s, z4.s, z2.s[1]\n" - "fmla z10.s, z4.s, z2.s[2]\n" - "fmla z11.s, z4.s, z2.s[3]\n" - "st1w z8.s, p0, [%[c_ptr]]\n" - "fmla z20.s, z4.s, z3.s[0]\n" - "fmla z21.s, z4.s, z3.s[1]\n" - "fmla z22.s, z4.s, z3.s[2]\n" - "fmla z23.s, z4.s, z3.s[3]\n" - "fmla z12.s, z5.s, z2.s[0]\n" - "fmla z13.s, z5.s, z2.s[1]\n" - "fmla z14.s, z5.s, z2.s[2]\n" - "fmla z15.s, z5.s, z2.s[3]\n" - "st1w z12.s, p0, [%[c_ptr], #1, MUL VL]\n" - "fmla z24.s, z5.s, z3.s[0]\n" - "fmla z25.s, z5.s, z3.s[1]\n" - "fmla z26.s, z5.s, z3.s[2]\n" - "fmla z27.s, z5.s, z3.s[3]\n" - "fmla z16.s, z6.s, z2.s[0]\n" - "fmla z17.s, z6.s, z2.s[1]\n" - "fmla z18.s, z6.s, z2.s[2]\n" - "fmla z19.s, z6.s, z2.s[3]\n" - "st1w z16.s, p0, [%[c_ptr], #2, MUL VL]\n" - "fmla z28.s, z6.s, z3.s[0]\n" - "fmla z29.s, z6.s, z3.s[1]\n" - "fmla z30.s, z6.s, z3.s[2]\n" - "st1w z9.s, p0, [%[c_ptr], #3, MUL VL]\n" - "fmla z31.s, z6.s, z3.s[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" + "ld1rqw { z0.s }, p0/Z, [%x[Apanel]]\n" + "mov z10.b, #0x0\n" + "mov z11.b, #0x0\n" + "ld1rqw { z1.s }, p0/Z, [%x[Apanel], #16]\n" + "mov z12.b, #0x0\n" + "mov z13.b, #0x0\n" + "ld1w { z4.s }, p0/Z, [x22]\n" + "mov z14.b, #0x0\n" + "mov z15.b, #0x0\n" + "ld1w { z5.s }, p0/Z, [x22, #1, MUL VL]\n" + "mov z16.b, #0x0\n" + "mov z17.b, #0x0\n" + "ld1w { z6.s }, 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 + "fmla z8.s, z4.s, z0.s[0]\n" + "fmla z11.s, z4.s, z0.s[1]\n" + "ld1rqw { z3.s }, p0/Z, [%x[Apanel], #32]\n" + "fmla z14.s, z4.s, z0.s[2]\n" + "fmla z17.s, z4.s, z0.s[3]\n" + "ld1rqw { z7.s }, p0/Z, [%x[Apanel], #48]\n" + "fmla z20.s, z4.s, z1.s[0]\n" + "fmla z23.s, z4.s, z1.s[1]\n" + "sub x20, x20, #0x2\n" + "fmla z26.s, z4.s, z1.s[2]\n" + "fmla z29.s, z4.s, z1.s[3]\n" + "ld1w { z4.s }, p0/Z, [x22, #3, MUL VL]\n" + "fmla z9.s, z5.s, z0.s[0]\n" + "fmla z12.s, z5.s, z0.s[1]\n" + "cmp x20, #0x2\n" + "fmla z15.s, z5.s, z0.s[2]\n" + "fmla z18.s, z5.s, z0.s[3]\n" + "add %x[Apanel], %x[Apanel], #0x40\n" + "fmla z21.s, z5.s, z1.s[0]\n" + "fmla z24.s, z5.s, z1.s[1]\n" + "fmla z27.s, z5.s, z1.s[2]\n" + "fmla z30.s, z5.s, z1.s[3]\n" + "ld1w { z5.s }, p0/Z, [x22, #4, MUL VL]\n" + "fmla z10.s, z6.s, z0.s[0]\n" + "fmla z13.s, z6.s, z0.s[1]\n" + "fmla z16.s, z6.s, z0.s[2]\n" + "fmla z19.s, z6.s, z0.s[3]\n" + "ld1rqw { z0.s }, p0/Z, [%x[Apanel]]\n" + "fmla z22.s, z6.s, z1.s[0]\n" + "fmla z25.s, z6.s, z1.s[1]\n" + "fmla z28.s, z6.s, z1.s[2]\n" + "fmla z31.s, z6.s, z1.s[3]\n" + "ld1w { z2.s }, p0/Z, [x22, #5, MUL VL]\n" + "addvl x22, x22, #6\n" + "fmla z8.s, z4.s, z3.s[0]\n" + "fmla z11.s, z4.s, z3.s[1]\n" + "ld1rqw { z1.s }, p0/Z, [%x[Apanel], #16]\n" + "fmla z14.s, z4.s, z3.s[2]\n" + "fmla z17.s, z4.s, z3.s[3]\n" + "fmla z20.s, z4.s, z7.s[0]\n" + "fmla z23.s, z4.s, z7.s[1]\n" + "fmla z26.s, z4.s, z7.s[2]\n" + "fmla z29.s, z4.s, z7.s[3]\n" + "ld1w { z4.s }, p0/Z, [x22]\n" + "fmla z9.s, z5.s, z3.s[0]\n" + "fmla z12.s, z5.s, z3.s[1]\n" + "fmla z15.s, z5.s, z3.s[2]\n" + "fmla z18.s, z5.s, z3.s[3]\n" + "fmla z21.s, z5.s, z7.s[0]\n" + "fmla z24.s, z5.s, z7.s[1]\n" + "fmla z27.s, z5.s, z7.s[2]\n" + "fmla z30.s, z5.s, z7.s[3]\n" + "ld1w { z5.s }, p0/Z, [x22, #1, MUL VL]\n" + "fmla z10.s, z2.s, z3.s[0]\n" + "fmla z13.s, z2.s, z3.s[1]\n" + "fmla z16.s, z2.s, z3.s[2]\n" + "fmla z19.s, z2.s, z3.s[3]\n" + "fmla z22.s, z2.s, z7.s[0]\n" + "fmla z25.s, z2.s, z7.s[1]\n" + "fmla z28.s, z2.s, z7.s[2]\n" + "fmla z31.s, z2.s, z7.s[3]\n" + "ld1w { z6.s }, p0/Z, [x22, #2, MUL VL]\n" + "bge 3b\n" + "4:" // main loop skip + "fmla z8.s, z4.s, z0.s[0]\n" + "fmla z11.s, z4.s, z0.s[1]\n" + "add %x[Apanel], %x[Apanel], #0x20\n" + "fmla z14.s, z4.s, z0.s[2]\n" + "fmla z17.s, z4.s, z0.s[3]\n" + "addvl x22, x22, #3\n" + "fmla z20.s, z4.s, z1.s[0]\n" + "fmla z23.s, z4.s, z1.s[1]\n" + "fmla z26.s, z4.s, z1.s[2]\n" + "fmla z29.s, z4.s, z1.s[3]\n" + "fmla z9.s, z5.s, z0.s[0]\n" + "fmla z12.s, z5.s, z0.s[1]\n" + "fmla z15.s, z5.s, z0.s[2]\n" + "fmla z18.s, z5.s, z0.s[3]\n" + "fmla z21.s, z5.s, z1.s[0]\n" + "fmla z24.s, z5.s, z1.s[1]\n" + "fmla z27.s, z5.s, z1.s[2]\n" + "fmla z30.s, z5.s, z1.s[3]\n" + "fmla z10.s, z6.s, z0.s[0]\n" + "fmla z13.s, z6.s, z0.s[1]\n" + "fmla z16.s, z6.s, z0.s[2]\n" + "fmla z19.s, z6.s, z0.s[3]\n" + "fmla z22.s, z6.s, z1.s[0]\n" + "fmla z25.s, z6.s, z1.s[1]\n" + "fmla z28.s, z6.s, z1.s[2]\n" + "fmla z31.s, z6.s, z1.s[3]\n" + "cbz x20, 5f\n" + "ld1rqw { z4.s }, p0/Z, [%x[Apanel]]\n" + "ld1rqw { z3.s }, p0/Z, [%x[Apanel], #16]\n" + "add %x[Apanel], %x[Apanel], #0x20\n" + "ld1w { z2.s }, p0/Z, [x22]\n" + "ld1w { z1.s }, p0/Z, [x22, #1, MUL VL]\n" + "fmla z8.s, z2.s, z4.s[0]\n" + "ld1w { z0.s }, p0/Z, [x22, #2, MUL VL]\n" + "fmla z11.s, z2.s, z4.s[1]\n" + "fmla z14.s, z2.s, z4.s[2]\n" + "fmla z17.s, z2.s, z4.s[3]\n" + "fmla z20.s, z2.s, z3.s[0]\n" + "addvl x22, x22, #3\n" + "fmla z23.s, z2.s, z3.s[1]\n" + "fmla z26.s, z2.s, z3.s[2]\n" + "fmla z29.s, z2.s, z3.s[3]\n" + "fmla z9.s, z1.s, z4.s[0]\n" + "fmla z12.s, z1.s, z4.s[1]\n" + "fmla z15.s, z1.s, z4.s[2]\n" + "fmla z18.s, z1.s, z4.s[3]\n" + "fmla z21.s, z1.s, z3.s[0]\n" + "fmla z24.s, z1.s, z3.s[1]\n" + "fmla z27.s, z1.s, z3.s[2]\n" + "fmla z30.s, z1.s, z3.s[3]\n" + "fmla z10.s, z0.s, z4.s[0]\n" + "fmla z13.s, z0.s, z4.s[1]\n" + "fmla z16.s, z0.s, z4.s[2]\n" + "fmla z19.s, z0.s, z4.s[3]\n" + "fmla z22.s, z0.s, z3.s[0]\n" + "fmla z25.s, z0.s, z3.s[1]\n" + "fmla z28.s, z0.s, z3.s[2]\n" + "fmla z31.s, z0.s, z3.s[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 |