diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp | 147 |
1 files changed, 42 insertions, 105 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp index c151179a1f..09d5d8d96d 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp @@ -161,13 +161,12 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x8\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - "cmp x26, #0x8\n" + "add x25, x25, #0x10\n" "fmla z10.h, z6.h, z0.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - "prfm pldl1keep, [x25, #0x80]\n" "fmla z11.h, z7.h, z0.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" "fmla z8.h, z6.h, z0.h[1]\n" @@ -235,7 +234,6 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - "add x25, x25, #0x10\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" @@ -318,9 +316,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "fmla z10.h, z6.h, z0.h[7]\n" "fmla z11.h, z7.h, z0.h[7]\n" "11:" // Height 1: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 6b\n" "tbz %x[flags], #1, 12f\n" @@ -424,16 +421,14 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x8\n" "fmla z9.h, z7.h, z0.h[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" "fmla z12.h, z6.h, z1.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - "cmp x26, #0x8\n" "fmla z13.h, z7.h, z1.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - "prfm pldl1keep, [x24, #0x80]\n" "fmla z10.h, z6.h, z0.h[0]\n" "fmla z14.h, z6.h, z1.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" @@ -533,9 +528,7 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "fmla z9.h, z7.h, z0.h[0]\n" - "add x24, x24, #0x10\n" "fmla z12.h, z6.h, z1.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" "fmla z13.h, z7.h, z1.h[0]\n" @@ -650,10 +643,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "fmla z11.h, z7.h, z0.h[7]\n" "fmla z15.h, z7.h, z1.h[7]\n" "24:" // Height 2: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 19b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -787,21 +778,18 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x8\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1rqh { z2.h }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" + "add x25, x25, #0x10\n" "fmla z12.h, z6.h, z1.h[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" + "fmla z13.h, z7.h, z1.h[0]\n" "add x23, x23, #0x10\n" "fmla z16.h, z6.h, z2.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - "cmp x26, #0x8\n" - "fmla z13.h, z7.h, z1.h[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" "fmla z17.h, z7.h, z2.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - "prfm pldl1keep, [x23, #0x80]\n" "fmla z10.h, z6.h, z0.h[0]\n" "fmla z14.h, z6.h, z1.h[0]\n" "fmla z18.h, z6.h, z2.h[0]\n" @@ -931,12 +919,9 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1rqh { z2.h }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" "fmla z12.h, z6.h, z1.h[0]\n" - "add x23, x23, #0x10\n" "fmla z13.h, z7.h, z1.h[0]\n" "fmla z16.h, z6.h, z2.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" @@ -1082,11 +1067,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "fmla z15.h, z7.h, z1.h[7]\n" "fmla z19.h, z7.h, z2.h[7]\n" "37:" // Height 3: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 32b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -1249,26 +1231,22 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x8\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1rqh { z2.h }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" + "add x25, x25, #0x10\n" "fmla z12.h, z6.h, z1.h[0]\n" "ld1rqh { z3.h }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" + "add x24, x24, #0x10\n" "fmla z16.h, z6.h, z2.h[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" - "add x22, x22, #0x10\n" + "add x23, x23, #0x10\n" "fmla z13.h, z7.h, z1.h[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "cmp x26, #0x8\n" + "add x22, x22, #0x10\n" + "fmla z17.h, z7.h, z2.h[0]\n" "fmla z20.h, z6.h, z3.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z17.h, z7.h, z2.h[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" "fmla z21.h, z7.h, z3.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - "prfm pldl1keep, [x22, #0x80]\n" "fmla z10.h, z6.h, z0.h[0]\n" "fmla z14.h, z6.h, z1.h[0]\n" "fmla z18.h, z6.h, z2.h[0]\n" @@ -1428,19 +1406,15 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1rqh { z2.h }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "fmla z12.h, z6.h, z1.h[0]\n" "ld1rqh { z3.h }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "fmla z16.h, z6.h, z2.h[0]\n" - "add x22, x22, #0x10\n" + "fmla z12.h, z6.h, z1.h[0]\n" "fmla z13.h, z7.h, z1.h[0]\n" - "fmla z17.h, z7.h, z2.h[0]\n" + "fmla z16.h, z6.h, z2.h[0]\n" "fmla z20.h, z6.h, z3.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z17.h, z7.h, z2.h[0]\n" "fmla z21.h, z7.h, z3.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" @@ -1613,12 +1587,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "fmla z19.h, z7.h, z2.h[7]\n" "fmla z23.h, z7.h, z3.h[7]\n" "50:" // Height 4: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 45b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -1810,32 +1780,27 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x8\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1rqh { z2.h }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" + "add x25, x25, #0x10\n" "fmla z12.h, z6.h, z1.h[0]\n" "ld1rqh { z3.h }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" + "add x24, x24, #0x10\n" "fmla z16.h, z6.h, z2.h[0]\n" "ld1rqh { z4.h }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" + "add x23, x23, #0x10\n" "fmla z13.h, z7.h, z1.h[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x22, x22, #0x10\n" + "fmla z17.h, z7.h, z2.h[0]\n" "add x21, x21, #0x10\n" "fmla z20.h, z6.h, z3.h[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "cmp x26, #0x8\n" "fmla z24.h, z6.h, z4.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z17.h, z7.h, z2.h[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" "fmla z21.h, z7.h, z3.h[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" "fmla z25.h, z7.h, z4.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" "fmla z10.h, z6.h, z0.h[0]\n" - "prfm pldl1keep, [x21, #0x80]\n" "fmla z14.h, z6.h, z1.h[0]\n" "fmla z18.h, z6.h, z2.h[0]\n" "fmla z22.h, z6.h, z3.h[0]\n" @@ -2024,22 +1989,17 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1rqh { z2.h }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "fmla z12.h, z6.h, z1.h[0]\n" "ld1rqh { z3.h }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "fmla z16.h, z6.h, z2.h[0]\n" + "fmla z12.h, z6.h, z1.h[0]\n" "ld1rqh { z4.h }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" "fmla z13.h, z7.h, z1.h[0]\n" - "add x21, x21, #0x10\n" - "fmla z17.h, z7.h, z2.h[0]\n" + "fmla z16.h, z6.h, z2.h[0]\n" "fmla z20.h, z6.h, z3.h[0]\n" "fmla z24.h, z6.h, z4.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z17.h, z7.h, z2.h[0]\n" "fmla z21.h, z7.h, z3.h[0]\n" "fmla z25.h, z7.h, z4.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" @@ -2243,13 +2203,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "fmla z23.h, z7.h, z3.h[7]\n" "fmla z27.h, z7.h, z4.h[7]\n" "63:" // Height 5: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "prfm pldl1keep, [x21, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 58b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -2473,37 +2428,31 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x8\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1rqh { z2.h }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" + "add x25, x25, #0x10\n" "fmla z12.h, z6.h, z1.h[0]\n" "ld1rqh { z3.h }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" + "add x24, x24, #0x10\n" "fmla z16.h, z6.h, z2.h[0]\n" "ld1rqh { z4.h }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" + "add x23, x23, #0x10\n" "fmla z13.h, z7.h, z1.h[0]\n" "ld1rqh { z5.h }, p0/Z, [x20]\n" - "add x21, x21, #0x10\n" + "add x22, x22, #0x10\n" "fmla z20.h, z6.h, z3.h[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x21, x21, #0x10\n" + "fmla z17.h, z7.h, z2.h[0]\n" "add x20, x20, #0x10\n" "fmla z24.h, z6.h, z4.h[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "cmp x26, #0x8\n" "fmla z28.h, z6.h, z5.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z17.h, z7.h, z2.h[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" "fmla z21.h, z7.h, z3.h[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" "fmla z25.h, z7.h, z4.h[0]\n" - "prfm pldl1keep, [x21, #0x80]\n" "fmla z29.h, z7.h, z5.h[0]\n" "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" "fmla z10.h, z6.h, z0.h[0]\n" - "prfm pldl1keep, [x20, #0x80]\n" "fmla z14.h, z6.h, z1.h[0]\n" "fmla z18.h, z6.h, z2.h[0]\n" "fmla z22.h, z6.h, z3.h[0]\n" @@ -2722,25 +2671,19 @@ void sve_hybrid_fp16_mla_6x4VL ( "ld1rqh { z0.h }, p0/Z, [x25]\n" "fmla z8.h, z6.h, z0.h[0]\n" "ld1rqh { z1.h }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "fmla z9.h, z7.h, z0.h[0]\n" "ld1rqh { z2.h }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "fmla z12.h, z6.h, z1.h[0]\n" "ld1rqh { z3.h }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "fmla z16.h, z6.h, z2.h[0]\n" + "fmla z12.h, z6.h, z1.h[0]\n" "ld1rqh { z4.h }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" "fmla z13.h, z7.h, z1.h[0]\n" "ld1rqh { z5.h }, p0/Z, [x20]\n" - "add x21, x21, #0x10\n" + "fmla z16.h, z6.h, z2.h[0]\n" "fmla z20.h, z6.h, z3.h[0]\n" - "add x20, x20, #0x10\n" - "fmla z17.h, z7.h, z2.h[0]\n" "fmla z24.h, z6.h, z4.h[0]\n" "fmla z28.h, z6.h, z5.h[0]\n" "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z17.h, z7.h, z2.h[0]\n" "fmla z21.h, z7.h, z3.h[0]\n" "fmla z25.h, z7.h, z4.h[0]\n" "fmla z29.h, z7.h, z5.h[0]\n" @@ -2975,14 +2918,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "fmla z27.h, z7.h, z4.h[7]\n" "fmla z31.h, z7.h, z5.h[7]\n" "76:" // Height 6: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "prfm pldl1keep, [x21, #0x80]\n" - "prfm pldl1keep, [x20, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 71b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" |