diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp | 194 |
1 files changed, 97 insertions, 97 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp index 1d502f5354..69ddb21c31 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp @@ -52,33 +52,33 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx( __asm__ __volatile__( "ptrue p0.b\n" "1:" // Height loop - "ldr x26, [%x[args_ptr], %[offsetof_Bpanel]]\n" - "ldr x25, [%x[args_ptr], %[offsetof_N]]\n" - "str x26, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x24, %x[Apanel]\n" + "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n" + "ldr x26, [%x[args_ptr], %[offsetof_N]]\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "mov x25, %x[Apanel]\n" "2:" // Width loop - "ldr x26, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x24, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" "cnth x23, ALL, MUL #2\n" - "add x22, x26, x20, LSL #1\n" + "add x22, x24, x20, LSL #1\n" "add x21, x22, x20, LSL #1\n" "add x20, x21, x20, LSL #1\n" - "cmp x25, x23\n" + "cmp x26, x23\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov %x[Apanel], x24\n" + "mov %x[Apanel], x25\n" "bgt 3f\n" "dech x23\n" - "cmp x25, x23\n" - "mov x21, x26\n" + "cmp x26, x23\n" + "mov x21, x24\n" "bgt 3f\n" - "mov x22, x26\n" + "mov x22, x24\n" "3:" // B setup done "ldr x20, [%x[args_ptr], %[offsetof_K]]\n" "cmp x20, #0x2\n" "mov z8.b, #0x0\n" "mov z9.b, #0x0\n" "mov z10.b, #0x0\n" - "ld1h { z0.h }, p0/Z, [x26]\n" + "ld1h { z0.h }, p0/Z, [x24]\n" "mov z11.b, #0x0\n" "mov z12.b, #0x0\n" "ld1h { z1.h }, p0/Z, [x22]\n" @@ -116,12 +116,12 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx( "fmla z11.h, p0/M, z0.h, z4.h\n" "fmla z12.h, p0/M, z1.h, z4.h\n" "fmla z13.h, p0/M, z2.h, z4.h\n" - "ld1rh { z4.h }, p0/Z, [%x[Apanel], #10]\n" + "ld1rh { z7.h }, p0/Z, [%x[Apanel], #10]\n" "fmla z14.h, p0/M, z0.h, z5.h\n" "fmla z15.h, p0/M, z1.h, z5.h\n" "cmp x20, #0x2\n" "fmla z16.h, p0/M, z2.h, z5.h\n" - "ld1rh { z5.h }, p0/Z, [%x[Apanel], #12]\n" + "ld1rh { z4.h }, p0/Z, [%x[Apanel], #12]\n" "fmla z17.h, p0/M, z0.h, z6.h\n" "fmla z18.h, p0/M, z1.h, z6.h\n" "fmla z19.h, p0/M, z2.h, z6.h\n" @@ -130,57 +130,57 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx( "fmla z21.h, p0/M, z1.h, z3.h\n" "fmla z22.h, p0/M, z2.h, z3.h\n" "ld1rh { z3.h }, p0/Z, [%x[Apanel], #16]\n" - "fmla z23.h, p0/M, z0.h, z4.h\n" - "fmla z24.h, p0/M, z1.h, z4.h\n" - "fmla z25.h, p0/M, z2.h, z4.h\n" - "ld1rh { z4.h }, p0/Z, [%x[Apanel], #18]\n" - "fmla z26.h, p0/M, z0.h, z5.h\n" - "fmla z27.h, p0/M, z1.h, z5.h\n" - "fmla z28.h, p0/M, z2.h, z5.h\n" - "ld1rh { z5.h }, p0/Z, [%x[Apanel], #20]\n" + "fmla z23.h, p0/M, z0.h, z7.h\n" + "fmla z24.h, p0/M, z1.h, z7.h\n" + "fmla z25.h, p0/M, z2.h, z7.h\n" + "ld1rh { z5.h }, p0/Z, [%x[Apanel], #18]\n" + "fmla z26.h, p0/M, z0.h, z4.h\n" + "fmla z27.h, p0/M, z1.h, z4.h\n" + "fmla z28.h, p0/M, z2.h, z4.h\n" + "ld1rh { z4.h }, p0/Z, [%x[Apanel], #20]\n" "fmla z29.h, p0/M, z0.h, z6.h\n" - "ld1h { z0.h }, p0/Z, [x26, #1, MUL VL]\n" + "ld1h { z7.h }, p0/Z, [x24, #1, MUL VL]\n" "fmla z30.h, p0/M, z1.h, z6.h\n" "fmla z31.h, p0/M, z2.h, z6.h\n" - "ld1h { z1.h }, p0/Z, [x22, #1, MUL VL]\n" + "ld1h { z6.h }, p0/Z, [x22, #1, MUL VL]\n" "ld1h { z2.h }, p0/Z, [x21, #1, MUL VL]\n" - "fmla z8.h, p0/M, z0.h, z3.h\n" - "ld1rh { z6.h }, p0/Z, [%x[Apanel], #22]\n" - "fmla z9.h, p0/M, z1.h, z3.h\n" + "fmla z8.h, p0/M, z7.h, z3.h\n" + "ld1rh { z1.h }, p0/Z, [%x[Apanel], #22]\n" + "fmla z9.h, p0/M, z6.h, z3.h\n" "fmla z10.h, p0/M, z2.h, z3.h\n" - "fmla z11.h, p0/M, z0.h, z4.h\n" + "fmla z11.h, p0/M, z7.h, z5.h\n" "ld1rh { z3.h }, p0/Z, [%x[Apanel], #24]\n" - "fmla z12.h, p0/M, z1.h, z4.h\n" - "fmla z13.h, p0/M, z2.h, z4.h\n" - "ld1rh { z4.h }, p0/Z, [%x[Apanel], #26]\n" - "fmla z14.h, p0/M, z0.h, z5.h\n" - "fmla z15.h, p0/M, z1.h, z5.h\n" - "addvl x26, x26, #2\n" - "fmla z16.h, p0/M, z2.h, z5.h\n" - "ld1rh { z5.h }, p0/Z, [%x[Apanel], #28]\n" - "fmla z17.h, p0/M, z0.h, z6.h\n" - "fmla z18.h, p0/M, z1.h, z6.h\n" - "fmla z19.h, p0/M, z2.h, z6.h\n" - "ld1rh { z6.h }, p0/Z, [%x[Apanel], #30]\n" + "fmla z12.h, p0/M, z6.h, z5.h\n" + "fmla z13.h, p0/M, z2.h, z5.h\n" + "ld1rh { z5.h }, p0/Z, [%x[Apanel], #26]\n" + "fmla z14.h, p0/M, z7.h, z4.h\n" + "fmla z15.h, p0/M, z6.h, z4.h\n" + "addvl x24, x24, #2\n" + "fmla z16.h, p0/M, z2.h, z4.h\n" + "ld1rh { z0.h }, p0/Z, [%x[Apanel], #28]\n" + "fmla z17.h, p0/M, z7.h, z1.h\n" + "fmla z18.h, p0/M, z6.h, z1.h\n" + "fmla z19.h, p0/M, z2.h, z1.h\n" + "ld1rh { z1.h }, p0/Z, [%x[Apanel], #30]\n" "addvl x22, x22, #2\n" "addvl x21, x21, #2\n" "add %x[Apanel], %x[Apanel], #0x20\n" - "fmla z20.h, p0/M, z0.h, z3.h\n" - "fmla z21.h, p0/M, z1.h, z3.h\n" + "fmla z20.h, p0/M, z7.h, z3.h\n" + "fmla z21.h, p0/M, z6.h, z3.h\n" "fmla z22.h, p0/M, z2.h, z3.h\n" "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n" - "fmla z23.h, p0/M, z0.h, z4.h\n" - "fmla z24.h, p0/M, z1.h, z4.h\n" - "fmla z25.h, p0/M, z2.h, z4.h\n" - "fmla z26.h, p0/M, z0.h, z5.h\n" + "fmla z23.h, p0/M, z7.h, z5.h\n" + "fmla z24.h, p0/M, z6.h, z5.h\n" + "fmla z25.h, p0/M, z2.h, z5.h\n" + "fmla z26.h, p0/M, z7.h, z0.h\n" "ld1rh { z4.h }, p0/Z, [%x[Apanel], #2]\n" - "fmla z27.h, p0/M, z1.h, z5.h\n" - "fmla z28.h, p0/M, z2.h, z5.h\n" + "fmla z27.h, p0/M, z6.h, z0.h\n" + "fmla z28.h, p0/M, z2.h, z0.h\n" "ld1rh { z5.h }, p0/Z, [%x[Apanel], #4]\n" - "fmla z29.h, p0/M, z0.h, z6.h\n" - "ld1h { z0.h }, p0/Z, [x26]\n" - "fmla z30.h, p0/M, z1.h, z6.h\n" - "fmla z31.h, p0/M, z2.h, z6.h\n" + "fmla z29.h, p0/M, z7.h, z1.h\n" + "ld1h { z0.h }, p0/Z, [x24]\n" + "fmla z30.h, p0/M, z6.h, z1.h\n" + "fmla z31.h, p0/M, z2.h, z1.h\n" "ld1h { z1.h }, p0/Z, [x22]\n" "ld1h { z2.h }, p0/Z, [x21]\n" "ld1rh { z6.h }, p0/Z, [%x[Apanel], #6]\n" @@ -188,9 +188,9 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx( "5:" // main loop skip "fmla z8.h, p0/M, z0.h, z3.h\n" "fmla z9.h, p0/M, z1.h, z3.h\n" - "addvl x26, x26, #1\n" + "addvl x24, x24, #1\n" "fmla z10.h, p0/M, z2.h, z3.h\n" - "ld1rh { z3.h }, p0/Z, [%x[Apanel], #8]\n" + "ld1rh { z7.h }, p0/Z, [%x[Apanel], #8]\n" "fmla z11.h, p0/M, z0.h, z4.h\n" "fmla z12.h, p0/M, z1.h, z4.h\n" "fmla z13.h, p0/M, z2.h, z4.h\n" @@ -203,11 +203,11 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx( "fmla z17.h, p0/M, z0.h, z6.h\n" "fmla z18.h, p0/M, z1.h, z6.h\n" "fmla z19.h, p0/M, z2.h, z6.h\n" - "ld1rh { z6.h }, p0/Z, [%x[Apanel], #14]\n" - "fmla z20.h, p0/M, z0.h, z3.h\n" - "fmla z21.h, p0/M, z1.h, z3.h\n" + "ld1rh { z3.h }, p0/Z, [%x[Apanel], #14]\n" + "fmla z20.h, p0/M, z0.h, z7.h\n" + "fmla z21.h, p0/M, z1.h, z7.h\n" "addvl x21, x21, #1\n" - "fmla z22.h, p0/M, z2.h, z3.h\n" + "fmla z22.h, p0/M, z2.h, z7.h\n" "fmla z23.h, p0/M, z0.h, z4.h\n" "add %x[Apanel], %x[Apanel], #0x10\n" "fmla z24.h, p0/M, z1.h, z4.h\n" @@ -215,50 +215,50 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx( "fmla z26.h, p0/M, z0.h, z5.h\n" "fmla z27.h, p0/M, z1.h, z5.h\n" "fmla z28.h, p0/M, z2.h, z5.h\n" - "fmla z29.h, p0/M, z0.h, z6.h\n" - "fmla z30.h, p0/M, z1.h, z6.h\n" - "fmla z31.h, p0/M, z2.h, z6.h\n" + "fmla z29.h, p0/M, z0.h, z3.h\n" + "fmla z30.h, p0/M, z1.h, z3.h\n" + "fmla z31.h, p0/M, z2.h, z3.h\n" "cbz x20, 6f\n" - "ld1h { z0.h }, p0/Z, [x26]\n" - "ld1h { z1.h }, p0/Z, [x22]\n" - "ld1h { z2.h }, p0/Z, [x21]\n" + "ld1h { z6.h }, p0/Z, [x24]\n" + "ld1h { z5.h }, p0/Z, [x22]\n" + "ld1h { z4.h }, p0/Z, [x21]\n" "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n" - "fmla z8.h, p0/M, z0.h, z3.h\n" - "ld1rh { z4.h }, p0/Z, [%x[Apanel], #2]\n" - "ld1rh { z5.h }, p0/Z, [%x[Apanel], #4]\n" - "fmla z9.h, p0/M, z1.h, z3.h\n" - "ld1rh { z6.h }, p0/Z, [%x[Apanel], #6]\n" - "fmla z10.h, p0/M, z2.h, z3.h\n" - "fmla z11.h, p0/M, z0.h, z4.h\n" - "fmla z12.h, p0/M, z1.h, z4.h\n" - "fmla z13.h, p0/M, z2.h, z4.h\n" + "fmla z8.h, p0/M, z6.h, z3.h\n" + "ld1rh { z2.h }, p0/Z, [%x[Apanel], #2]\n" + "ld1rh { z1.h }, p0/Z, [%x[Apanel], #4]\n" + "fmla z9.h, p0/M, z5.h, z3.h\n" + "ld1rh { z0.h }, p0/Z, [%x[Apanel], #6]\n" + "fmla z10.h, p0/M, z4.h, z3.h\n" + "fmla z11.h, p0/M, z6.h, z2.h\n" + "fmla z12.h, p0/M, z5.h, z2.h\n" + "fmla z13.h, p0/M, z4.h, z2.h\n" "ld1rh { z3.h }, p0/Z, [%x[Apanel], #8]\n" - "fmla z14.h, p0/M, z0.h, z5.h\n" - "fmla z15.h, p0/M, z1.h, z5.h\n" - "ld1rh { z4.h }, p0/Z, [%x[Apanel], #10]\n" - "fmla z16.h, p0/M, z2.h, z5.h\n" - "fmla z17.h, p0/M, z0.h, z6.h\n" - "ld1rh { z5.h }, p0/Z, [%x[Apanel], #12]\n" - "fmla z18.h, p0/M, z1.h, z6.h\n" - "fmla z19.h, p0/M, z2.h, z6.h\n" - "ld1rh { z6.h }, p0/Z, [%x[Apanel], #14]\n" - "fmla z20.h, p0/M, z0.h, z3.h\n" - "fmla z21.h, p0/M, z1.h, z3.h\n" + "fmla z14.h, p0/M, z6.h, z1.h\n" + "fmla z15.h, p0/M, z5.h, z1.h\n" + "ld1rh { z2.h }, p0/Z, [%x[Apanel], #10]\n" + "fmla z16.h, p0/M, z4.h, z1.h\n" + "fmla z17.h, p0/M, z6.h, z0.h\n" + "ld1rh { z1.h }, p0/Z, [%x[Apanel], #12]\n" + "fmla z18.h, p0/M, z5.h, z0.h\n" + "fmla z19.h, p0/M, z4.h, z0.h\n" + "ld1rh { z0.h }, p0/Z, [%x[Apanel], #14]\n" + "fmla z20.h, p0/M, z6.h, z3.h\n" + "fmla z21.h, p0/M, z5.h, z3.h\n" "add %x[Apanel], %x[Apanel], #0x10\n" - "fmla z22.h, p0/M, z2.h, z3.h\n" - "fmla z23.h, p0/M, z0.h, z4.h\n" - "fmla z24.h, p0/M, z1.h, z4.h\n" - "fmla z25.h, p0/M, z2.h, z4.h\n" - "fmla z26.h, p0/M, z0.h, z5.h\n" - "fmla z27.h, p0/M, z1.h, z5.h\n" - "fmla z28.h, p0/M, z2.h, z5.h\n" - "fmla z29.h, p0/M, z0.h, z6.h\n" - "fmla z30.h, p0/M, z1.h, z6.h\n" - "fmla z31.h, p0/M, z2.h, z6.h\n" + "fmla z22.h, p0/M, z4.h, z3.h\n" + "fmla z23.h, p0/M, z6.h, z2.h\n" + "fmla z24.h, p0/M, z5.h, z2.h\n" + "fmla z25.h, p0/M, z4.h, z2.h\n" + "fmla z26.h, p0/M, z6.h, z1.h\n" + "fmla z27.h, p0/M, z5.h, z1.h\n" + "fmla z28.h, p0/M, z4.h, z1.h\n" + "fmla z29.h, p0/M, z6.h, z0.h\n" + "fmla z30.h, p0/M, z5.h, z0.h\n" + "fmla z31.h, p0/M, z4.h, z0.h\n" "6:" // multiply loop done - "dech x25, ALL, MUL #3\n" + "dech x26, ALL, MUL #3\n" "st1h { z8.h }, p0, [%x[Cpanel]]\n" - "cmp x25, XZR\n" + "cmp x26, XZR\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" @@ -289,7 +289,7 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx( "bne 1b\n" : [Apanel] "+&r" (Apanel), [Cpanel] "+&r" (Cpanel), [ablocks] "+&r" (ablocks) : [args_ptr] "r" (&ka), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_Bpanel] "I" (offsetof(KernelArgs, Bpanel)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)) - : "cc", "memory", "p0", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "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", "p0", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "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" ); } |