diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp | 679 |
1 files changed, 339 insertions, 340 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp index 9ae51af59b..66481f04f9 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp @@ -139,11 +139,11 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "6:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 7f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" "cbnz x28, 8f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -159,12 +159,12 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "9:" // Height 1: Multiply loop: Main loop "fmla z8.s, p4/M, z6.s, z0.s\n" "fmla z9.s, p4/M, z7.s, z0.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" "add x26, x26, #0x4\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" + "fmla z10.s, p4/M, z17.s, z0.s\n" + "fmla z11.s, p4/M, z16.s, z0.s\n" "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1w { z6.s }, p4/Z, [x10]\n" @@ -174,27 +174,27 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "fmla z8.s, p4/M, z6.s, z0.s\n" "fmla z9.s, p4/M, z7.s, z0.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x10, #3, MUL VL]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" + "fmla z10.s, p4/M, z17.s, z0.s\n" + "fmla z11.s, p4/M, z16.s, z0.s\n" "addvl x10, x10, #4\n" "bne 6b\n" "tbz %x[flags], #1, 11f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p4/Z, [x20]\n" + "ld1rw { z17.s }, p4/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p4/Z, [x20]\n" - "fmin z8.s, p4/M, z8.s, z1.s\n" - "fmin z9.s, p4/M, z9.s, z1.s\n" - "fmin z10.s, p4/M, z10.s, z1.s\n" - "fmin z11.s, p4/M, z11.s, z1.s\n" - "fmax z8.s, p4/M, z8.s, z0.s\n" - "fmax z9.s, p4/M, z9.s, z0.s\n" - "fmax z10.s, p4/M, z10.s, z0.s\n" - "fmax z11.s, p4/M, z11.s, z0.s\n" + "ld1rw { z16.s }, p4/Z, [x20]\n" + "fmin z8.s, p4/M, z8.s, z17.s\n" + "fmin z9.s, p4/M, z9.s, z17.s\n" + "fmin z10.s, p4/M, z10.s, z17.s\n" + "fmin z11.s, p4/M, z11.s, z17.s\n" + "fmax z8.s, p4/M, z8.s, z16.s\n" + "fmax z9.s, p4/M, z9.s, z16.s\n" + "fmax z10.s, p4/M, z10.s, z16.s\n" + "fmax z11.s, p4/M, z11.s, z16.s\n" "11:" // Height 1: No activation "st1w { z8.s }, p3, [x9]\n" "st1w { z9.s }, p2, [x9, #1, MUL VL]\n" @@ -234,15 +234,15 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "15:" // Height 2: no bias "tbz %x[flags], #0, 16f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" + "add x20, x9, x20, LSL #2\n" "ld1w { z8.s }, p3/Z, [x9]\n" "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n" "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p3/Z, [x25]\n" - "ld1w { z13.s }, p2/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p1/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p0/Z, [x25, #3, MUL VL]\n" + "ld1w { z12.s }, p3/Z, [x20]\n" + "ld1w { z13.s }, p2/Z, [x20, #1, MUL VL]\n" + "ld1w { z14.s }, p1/Z, [x20, #2, MUL VL]\n" + "ld1w { z15.s }, p0/Z, [x20, #3, MUL VL]\n" "b 17f\n" "16:" // Height 2: no accumulate "mov z8.b, #0x0\n" @@ -258,12 +258,12 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "18:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 19f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" "cbnz x28, 20f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -271,7 +271,7 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "b 20f\n" "19:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" "20:" // Height 2: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -282,18 +282,18 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "21:" // Height 2: Multiply loop: Main loop "fmla z8.s, p4/M, z6.s, z0.s\n" "fmla z12.s, p4/M, z6.s, z1.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x10, #2, MUL VL]\n" "add x26, x26, #0x4\n" "fmla z9.s, p4/M, z7.s, z0.s\n" "fmla z13.s, p4/M, z7.s, z1.s\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" "subs x27, x27, #0x1\n" "add x25, x25, #0x4\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z14.s, p4/M, z6.s, z1.s\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" - "fmla z15.s, p4/M, z7.s, z1.s\n" + "fmla z10.s, p4/M, z17.s, z0.s\n" + "fmla z14.s, p4/M, z17.s, z1.s\n" + "fmla z11.s, p4/M, z16.s, z0.s\n" + "fmla z15.s, p4/M, z16.s, z1.s\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1rw { z1.s }, p4/Z, [x25]\n" "ld1w { z6.s }, p4/Z, [x10]\n" @@ -303,41 +303,41 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "fmla z8.s, p4/M, z6.s, z0.s\n" "fmla z12.s, p4/M, z6.s, z1.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x10, #2, MUL VL]\n" "fmla z9.s, p4/M, z7.s, z0.s\n" "fmla z13.s, p4/M, z7.s, z1.s\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x10, #3, MUL VL]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z14.s, p4/M, z6.s, z1.s\n" + "fmla z10.s, p4/M, z17.s, z0.s\n" + "fmla z14.s, p4/M, z17.s, z1.s\n" "addvl x10, x10, #4\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" - "fmla z15.s, p4/M, z7.s, z1.s\n" + "fmla z11.s, p4/M, z16.s, z0.s\n" + "fmla z15.s, p4/M, z16.s, z1.s\n" "bne 18b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "add x25, x9, x20, LSL #2\n" "tbz %x[flags], #1, 23f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p4/Z, [x20]\n" + "ld1rw { z17.s }, p4/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p4/Z, [x20]\n" - "fmin z8.s, p4/M, z8.s, z1.s\n" - "fmin z9.s, p4/M, z9.s, z1.s\n" - "fmin z10.s, p4/M, z10.s, z1.s\n" - "fmin z11.s, p4/M, z11.s, z1.s\n" - "fmin z12.s, p4/M, z12.s, z1.s\n" - "fmin z13.s, p4/M, z13.s, z1.s\n" - "fmin z14.s, p4/M, z14.s, z1.s\n" - "fmin z15.s, p4/M, z15.s, z1.s\n" - "fmax z8.s, p4/M, z8.s, z0.s\n" - "fmax z9.s, p4/M, z9.s, z0.s\n" - "fmax z10.s, p4/M, z10.s, z0.s\n" - "fmax z11.s, p4/M, z11.s, z0.s\n" - "fmax z12.s, p4/M, z12.s, z0.s\n" - "fmax z13.s, p4/M, z13.s, z0.s\n" - "fmax z14.s, p4/M, z14.s, z0.s\n" - "fmax z15.s, p4/M, z15.s, z0.s\n" + "ld1rw { z16.s }, p4/Z, [x20]\n" + "fmin z8.s, p4/M, z8.s, z17.s\n" + "fmin z9.s, p4/M, z9.s, z17.s\n" + "fmin z10.s, p4/M, z10.s, z17.s\n" + "fmin z11.s, p4/M, z11.s, z17.s\n" + "fmin z12.s, p4/M, z12.s, z17.s\n" + "fmin z13.s, p4/M, z13.s, z17.s\n" + "fmin z14.s, p4/M, z14.s, z17.s\n" + "fmin z15.s, p4/M, z15.s, z17.s\n" + "fmax z8.s, p4/M, z8.s, z16.s\n" + "fmax z9.s, p4/M, z9.s, z16.s\n" + "fmax z10.s, p4/M, z10.s, z16.s\n" + "fmax z11.s, p4/M, z11.s, z16.s\n" + "fmax z12.s, p4/M, z12.s, z16.s\n" + "fmax z13.s, p4/M, z13.s, z16.s\n" + "fmax z14.s, p4/M, z14.s, z16.s\n" + "fmax z15.s, p4/M, z15.s, z16.s\n" "23:" // Height 2: No activation "st1w { z8.s }, p3, [x9]\n" "st1w { z9.s }, p2, [x9, #1, MUL VL]\n" @@ -385,20 +385,20 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "27:" // Height 3: no bias "tbz %x[flags], #0, 28f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x21, x9, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z8.s }, p3/Z, [x9]\n" "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n" "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p3/Z, [x25]\n" - "ld1w { z13.s }, p2/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p1/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p0/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p3/Z, [x24]\n" - "ld1w { z17.s }, p2/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p1/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p0/Z, [x24, #3, MUL VL]\n" + "ld1w { z12.s }, p3/Z, [x21]\n" + "ld1w { z13.s }, p2/Z, [x21, #1, MUL VL]\n" + "ld1w { z14.s }, p1/Z, [x21, #2, MUL VL]\n" + "ld1w { z15.s }, p0/Z, [x21, #3, MUL VL]\n" + "ld1w { z16.s }, p3/Z, [x20]\n" + "ld1w { z17.s }, p2/Z, [x20, #1, MUL VL]\n" + "ld1w { z18.s }, p1/Z, [x20, #2, MUL VL]\n" + "ld1w { z19.s }, p0/Z, [x20, #3, MUL VL]\n" "b 29f\n" "28:" // Height 3: no accumulate "mov z8.b, #0x0\n" @@ -418,13 +418,13 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "30:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 31f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" "cbnz x28, 32f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -433,8 +433,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "b 32f\n" "31:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" "32:" // Height 3: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -450,21 +450,21 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "subs x27, x27, #0x1\n" "fmla z16.s, p4/M, z6.s, z2.s\n" "fmla z9.s, p4/M, z7.s, z0.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z21.s }, p4/Z, [x10, #2, MUL VL]\n" "add x25, x25, #0x4\n" "fmla z13.s, p4/M, z7.s, z1.s\n" "fmla z17.s, p4/M, z7.s, z2.s\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" "add x24, x24, #0x4\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z14.s, p4/M, z6.s, z1.s\n" - "fmla z18.s, p4/M, z6.s, z2.s\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" + "fmla z10.s, p4/M, z21.s, z0.s\n" + "fmla z14.s, p4/M, z21.s, z1.s\n" + "fmla z18.s, p4/M, z21.s, z2.s\n" + "fmla z11.s, p4/M, z20.s, z0.s\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1w { z6.s }, p4/Z, [x10]\n" - "fmla z15.s, p4/M, z7.s, z1.s\n" - "fmla z19.s, p4/M, z7.s, z2.s\n" + "fmla z15.s, p4/M, z20.s, z1.s\n" + "fmla z19.s, p4/M, z20.s, z2.s\n" "ld1rw { z1.s }, p4/Z, [x25]\n" "ld1rw { z2.s }, p4/Z, [x24]\n" "ld1w { z7.s }, p4/Z, [x10, #1, MUL VL]\n" @@ -476,51 +476,51 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "add x28, x28, #0x1\n" "fmla z16.s, p4/M, z6.s, z2.s\n" "fmla z9.s, p4/M, z7.s, z0.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z21.s }, p4/Z, [x10, #2, MUL VL]\n" "cmp x28, x20\n" "fmla z13.s, p4/M, z7.s, z1.s\n" "fmla z17.s, p4/M, z7.s, z2.s\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z14.s, p4/M, z6.s, z1.s\n" - "fmla z18.s, p4/M, z6.s, z2.s\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" - "fmla z15.s, p4/M, z7.s, z1.s\n" - "fmla z19.s, p4/M, z7.s, z2.s\n" + "fmla z10.s, p4/M, z21.s, z0.s\n" + "fmla z14.s, p4/M, z21.s, z1.s\n" + "fmla z18.s, p4/M, z21.s, z2.s\n" + "fmla z11.s, p4/M, z20.s, z0.s\n" + "fmla z15.s, p4/M, z20.s, z1.s\n" + "fmla z19.s, p4/M, z20.s, z2.s\n" "bne 30b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "add x25, x9, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" "tbz %x[flags], #1, 35f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p4/Z, [x20]\n" + "ld1rw { z21.s }, p4/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p4/Z, [x20]\n" - "fmin z8.s, p4/M, z8.s, z1.s\n" - "fmin z9.s, p4/M, z9.s, z1.s\n" - "fmin z10.s, p4/M, z10.s, z1.s\n" - "fmin z11.s, p4/M, z11.s, z1.s\n" - "fmin z12.s, p4/M, z12.s, z1.s\n" - "fmin z13.s, p4/M, z13.s, z1.s\n" - "fmin z14.s, p4/M, z14.s, z1.s\n" - "fmin z15.s, p4/M, z15.s, z1.s\n" - "fmin z16.s, p4/M, z16.s, z1.s\n" - "fmin z17.s, p4/M, z17.s, z1.s\n" - "fmin z18.s, p4/M, z18.s, z1.s\n" - "fmin z19.s, p4/M, z19.s, z1.s\n" - "fmax z8.s, p4/M, z8.s, z0.s\n" - "fmax z9.s, p4/M, z9.s, z0.s\n" - "fmax z10.s, p4/M, z10.s, z0.s\n" - "fmax z11.s, p4/M, z11.s, z0.s\n" - "fmax z12.s, p4/M, z12.s, z0.s\n" - "fmax z13.s, p4/M, z13.s, z0.s\n" - "fmax z14.s, p4/M, z14.s, z0.s\n" - "fmax z15.s, p4/M, z15.s, z0.s\n" - "fmax z16.s, p4/M, z16.s, z0.s\n" - "fmax z17.s, p4/M, z17.s, z0.s\n" - "fmax z18.s, p4/M, z18.s, z0.s\n" - "fmax z19.s, p4/M, z19.s, z0.s\n" + "ld1rw { z20.s }, p4/Z, [x20]\n" + "fmin z8.s, p4/M, z8.s, z21.s\n" + "fmin z9.s, p4/M, z9.s, z21.s\n" + "fmin z10.s, p4/M, z10.s, z21.s\n" + "fmin z11.s, p4/M, z11.s, z21.s\n" + "fmin z12.s, p4/M, z12.s, z21.s\n" + "fmin z13.s, p4/M, z13.s, z21.s\n" + "fmin z14.s, p4/M, z14.s, z21.s\n" + "fmin z15.s, p4/M, z15.s, z21.s\n" + "fmin z16.s, p4/M, z16.s, z21.s\n" + "fmin z17.s, p4/M, z17.s, z21.s\n" + "fmin z18.s, p4/M, z18.s, z21.s\n" + "fmin z19.s, p4/M, z19.s, z21.s\n" + "fmax z8.s, p4/M, z8.s, z20.s\n" + "fmax z9.s, p4/M, z9.s, z20.s\n" + "fmax z10.s, p4/M, z10.s, z20.s\n" + "fmax z11.s, p4/M, z11.s, z20.s\n" + "fmax z12.s, p4/M, z12.s, z20.s\n" + "fmax z13.s, p4/M, z13.s, z20.s\n" + "fmax z14.s, p4/M, z14.s, z20.s\n" + "fmax z15.s, p4/M, z15.s, z20.s\n" + "fmax z16.s, p4/M, z16.s, z20.s\n" + "fmax z17.s, p4/M, z17.s, z20.s\n" + "fmax z18.s, p4/M, z18.s, z20.s\n" + "fmax z19.s, p4/M, z19.s, z20.s\n" "35:" // Height 3: No activation "st1w { z8.s }, p3, [x9]\n" "st1w { z9.s }, p2, [x9, #1, MUL VL]\n" @@ -576,25 +576,25 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "39:" // Height 4: no bias "tbz %x[flags], #0, 40f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x22, x9, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "ld1w { z8.s }, p3/Z, [x9]\n" - "add x23, x24, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n" "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p3/Z, [x25]\n" - "ld1w { z13.s }, p2/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p1/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p0/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p3/Z, [x24]\n" - "ld1w { z17.s }, p2/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p1/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p0/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p3/Z, [x23]\n" - "ld1w { z21.s }, p2/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p1/Z, [x23, #2, MUL VL]\n" - "ld1w { z23.s }, p0/Z, [x23, #3, MUL VL]\n" + "ld1w { z12.s }, p3/Z, [x22]\n" + "ld1w { z13.s }, p2/Z, [x22, #1, MUL VL]\n" + "ld1w { z14.s }, p1/Z, [x22, #2, MUL VL]\n" + "ld1w { z15.s }, p0/Z, [x22, #3, MUL VL]\n" + "ld1w { z16.s }, p3/Z, [x21]\n" + "ld1w { z17.s }, p2/Z, [x21, #1, MUL VL]\n" + "ld1w { z18.s }, p1/Z, [x21, #2, MUL VL]\n" + "ld1w { z19.s }, p0/Z, [x21, #3, MUL VL]\n" + "ld1w { z20.s }, p3/Z, [x20]\n" + "ld1w { z21.s }, p2/Z, [x20, #1, MUL VL]\n" + "ld1w { z22.s }, p1/Z, [x20, #2, MUL VL]\n" + "ld1w { z23.s }, p0/Z, [x20, #3, MUL VL]\n" "b 41f\n" "40:" // Height 4: no accumulate "mov z8.b, #0x0\n" @@ -618,14 +618,14 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "42:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 43f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" "cbnz x28, 44f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -635,9 +635,9 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "b 44f\n" "43:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "44:" // Height 4: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -654,7 +654,7 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "subs x27, x27, #0x1\n" "fmla z16.s, p4/M, z6.s, z2.s\n" "fmla z20.s, p4/M, z6.s, z3.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z25.s }, p4/Z, [x10, #2, MUL VL]\n" "add x25, x25, #0x4\n" "fmla z9.s, p4/M, z7.s, z0.s\n" "fmla z13.s, p4/M, z7.s, z1.s\n" @@ -662,19 +662,19 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "add x23, x23, #0x4\n" "fmla z17.s, p4/M, z7.s, z2.s\n" "fmla z21.s, p4/M, z7.s, z3.s\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z24.s }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z14.s, p4/M, z6.s, z1.s\n" - "fmla z18.s, p4/M, z6.s, z2.s\n" - "fmla z22.s, p4/M, z6.s, z3.s\n" + "fmla z10.s, p4/M, z25.s, z0.s\n" + "fmla z14.s, p4/M, z25.s, z1.s\n" + "fmla z18.s, p4/M, z25.s, z2.s\n" + "fmla z22.s, p4/M, z25.s, z3.s\n" "ld1w { z6.s }, p4/Z, [x10]\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" - "fmla z15.s, p4/M, z7.s, z1.s\n" + "fmla z11.s, p4/M, z24.s, z0.s\n" + "fmla z15.s, p4/M, z24.s, z1.s\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1rw { z1.s }, p4/Z, [x25]\n" - "fmla z19.s, p4/M, z7.s, z2.s\n" - "fmla z23.s, p4/M, z7.s, z3.s\n" + "fmla z19.s, p4/M, z24.s, z2.s\n" + "fmla z23.s, p4/M, z24.s, z3.s\n" "ld1rw { z2.s }, p4/Z, [x24]\n" "ld1rw { z3.s }, p4/Z, [x23]\n" "ld1w { z7.s }, p4/Z, [x10, #1, MUL VL]\n" @@ -686,22 +686,22 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "add x28, x28, #0x1\n" "fmla z16.s, p4/M, z6.s, z2.s\n" "fmla z20.s, p4/M, z6.s, z3.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z25.s }, p4/Z, [x10, #2, MUL VL]\n" "cmp x28, x20\n" "fmla z9.s, p4/M, z7.s, z0.s\n" "fmla z13.s, p4/M, z7.s, z1.s\n" "fmla z17.s, p4/M, z7.s, z2.s\n" "fmla z21.s, p4/M, z7.s, z3.s\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z24.s }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z14.s, p4/M, z6.s, z1.s\n" - "fmla z18.s, p4/M, z6.s, z2.s\n" - "fmla z22.s, p4/M, z6.s, z3.s\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" - "fmla z15.s, p4/M, z7.s, z1.s\n" - "fmla z19.s, p4/M, z7.s, z2.s\n" - "fmla z23.s, p4/M, z7.s, z3.s\n" + "fmla z10.s, p4/M, z25.s, z0.s\n" + "fmla z14.s, p4/M, z25.s, z1.s\n" + "fmla z18.s, p4/M, z25.s, z2.s\n" + "fmla z22.s, p4/M, z25.s, z3.s\n" + "fmla z11.s, p4/M, z24.s, z0.s\n" + "fmla z15.s, p4/M, z24.s, z1.s\n" + "fmla z19.s, p4/M, z24.s, z2.s\n" + "fmla z23.s, p4/M, z24.s, z3.s\n" "bne 42b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "add x25, x9, x20, LSL #2\n" @@ -709,41 +709,41 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "add x23, x24, x20, LSL #2\n" "tbz %x[flags], #1, 47f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p4/Z, [x20]\n" + "ld1rw { z25.s }, p4/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p4/Z, [x20]\n" - "fmin z8.s, p4/M, z8.s, z1.s\n" - "fmin z9.s, p4/M, z9.s, z1.s\n" - "fmin z10.s, p4/M, z10.s, z1.s\n" - "fmin z11.s, p4/M, z11.s, z1.s\n" - "fmin z12.s, p4/M, z12.s, z1.s\n" - "fmin z13.s, p4/M, z13.s, z1.s\n" - "fmin z14.s, p4/M, z14.s, z1.s\n" - "fmin z15.s, p4/M, z15.s, z1.s\n" - "fmin z16.s, p4/M, z16.s, z1.s\n" - "fmin z17.s, p4/M, z17.s, z1.s\n" - "fmin z18.s, p4/M, z18.s, z1.s\n" - "fmin z19.s, p4/M, z19.s, z1.s\n" - "fmin z20.s, p4/M, z20.s, z1.s\n" - "fmin z21.s, p4/M, z21.s, z1.s\n" - "fmin z22.s, p4/M, z22.s, z1.s\n" - "fmin z23.s, p4/M, z23.s, z1.s\n" - "fmax z8.s, p4/M, z8.s, z0.s\n" - "fmax z9.s, p4/M, z9.s, z0.s\n" - "fmax z10.s, p4/M, z10.s, z0.s\n" - "fmax z11.s, p4/M, z11.s, z0.s\n" - "fmax z12.s, p4/M, z12.s, z0.s\n" - "fmax z13.s, p4/M, z13.s, z0.s\n" - "fmax z14.s, p4/M, z14.s, z0.s\n" - "fmax z15.s, p4/M, z15.s, z0.s\n" - "fmax z16.s, p4/M, z16.s, z0.s\n" - "fmax z17.s, p4/M, z17.s, z0.s\n" - "fmax z18.s, p4/M, z18.s, z0.s\n" - "fmax z19.s, p4/M, z19.s, z0.s\n" - "fmax z20.s, p4/M, z20.s, z0.s\n" - "fmax z21.s, p4/M, z21.s, z0.s\n" - "fmax z22.s, p4/M, z22.s, z0.s\n" - "fmax z23.s, p4/M, z23.s, z0.s\n" + "ld1rw { z24.s }, p4/Z, [x20]\n" + "fmin z8.s, p4/M, z8.s, z25.s\n" + "fmin z9.s, p4/M, z9.s, z25.s\n" + "fmin z10.s, p4/M, z10.s, z25.s\n" + "fmin z11.s, p4/M, z11.s, z25.s\n" + "fmin z12.s, p4/M, z12.s, z25.s\n" + "fmin z13.s, p4/M, z13.s, z25.s\n" + "fmin z14.s, p4/M, z14.s, z25.s\n" + "fmin z15.s, p4/M, z15.s, z25.s\n" + "fmin z16.s, p4/M, z16.s, z25.s\n" + "fmin z17.s, p4/M, z17.s, z25.s\n" + "fmin z18.s, p4/M, z18.s, z25.s\n" + "fmin z19.s, p4/M, z19.s, z25.s\n" + "fmin z20.s, p4/M, z20.s, z25.s\n" + "fmin z21.s, p4/M, z21.s, z25.s\n" + "fmin z22.s, p4/M, z22.s, z25.s\n" + "fmin z23.s, p4/M, z23.s, z25.s\n" + "fmax z8.s, p4/M, z8.s, z24.s\n" + "fmax z9.s, p4/M, z9.s, z24.s\n" + "fmax z10.s, p4/M, z10.s, z24.s\n" + "fmax z11.s, p4/M, z11.s, z24.s\n" + "fmax z12.s, p4/M, z12.s, z24.s\n" + "fmax z13.s, p4/M, z13.s, z24.s\n" + "fmax z14.s, p4/M, z14.s, z24.s\n" + "fmax z15.s, p4/M, z15.s, z24.s\n" + "fmax z16.s, p4/M, z16.s, z24.s\n" + "fmax z17.s, p4/M, z17.s, z24.s\n" + "fmax z18.s, p4/M, z18.s, z24.s\n" + "fmax z19.s, p4/M, z19.s, z24.s\n" + "fmax z20.s, p4/M, z20.s, z24.s\n" + "fmax z21.s, p4/M, z21.s, z24.s\n" + "fmax z22.s, p4/M, z22.s, z24.s\n" + "fmax z23.s, p4/M, z23.s, z24.s\n" "47:" // Height 4: No activation "st1w { z8.s }, p3, [x9]\n" "st1w { z9.s }, p2, [x9, #1, MUL VL]\n" @@ -807,30 +807,30 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "51:" // Height 5: no bias "tbz %x[flags], #0, 52f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z8.s }, p3/Z, [x9]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x9, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" + "ld1w { z8.s }, p3/Z, [x9]\n" + "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n" "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p3/Z, [x25]\n" - "ld1w { z13.s }, p2/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p1/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p0/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p3/Z, [x24]\n" - "ld1w { z17.s }, p2/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p1/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p0/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p3/Z, [x23]\n" - "ld1w { z21.s }, p2/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p1/Z, [x23, #2, MUL VL]\n" - "ld1w { z23.s }, p0/Z, [x23, #3, MUL VL]\n" - "ld1w { z24.s }, p3/Z, [x22]\n" - "ld1w { z25.s }, p2/Z, [x22, #1, MUL VL]\n" - "ld1w { z26.s }, p1/Z, [x22, #2, MUL VL]\n" - "ld1w { z27.s }, p0/Z, [x22, #3, MUL VL]\n" + "ld1w { z12.s }, p3/Z, [x23]\n" + "ld1w { z13.s }, p2/Z, [x23, #1, MUL VL]\n" + "ld1w { z14.s }, p1/Z, [x23, #2, MUL VL]\n" + "ld1w { z15.s }, p0/Z, [x23, #3, MUL VL]\n" + "ld1w { z16.s }, p3/Z, [x22]\n" + "ld1w { z17.s }, p2/Z, [x22, #1, MUL VL]\n" + "ld1w { z18.s }, p1/Z, [x22, #2, MUL VL]\n" + "ld1w { z19.s }, p0/Z, [x22, #3, MUL VL]\n" + "ld1w { z20.s }, p3/Z, [x21]\n" + "ld1w { z21.s }, p2/Z, [x21, #1, MUL VL]\n" + "ld1w { z22.s }, p1/Z, [x21, #2, MUL VL]\n" + "ld1w { z23.s }, p0/Z, [x21, #3, MUL VL]\n" + "ld1w { z24.s }, p3/Z, [x20]\n" + "ld1w { z25.s }, p2/Z, [x20, #1, MUL VL]\n" + "ld1w { z26.s }, p1/Z, [x20, #2, MUL VL]\n" + "ld1w { z27.s }, p0/Z, [x20, #3, MUL VL]\n" "b 53f\n" "52:" // Height 5: no accumulate "mov z8.b, #0x0\n" @@ -858,15 +858,15 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "54:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 55f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" "cbnz x28, 56f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -877,10 +877,10 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "b 56f\n" "55:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "56:" // Height 5: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -902,29 +902,29 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "add x24, x24, #0x4\n" "fmla z24.s, p4/M, z6.s, z4.s\n" "fmla z9.s, p4/M, z7.s, z0.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z29.s }, p4/Z, [x10, #2, MUL VL]\n" "add x23, x23, #0x4\n" "fmla z13.s, p4/M, z7.s, z1.s\n" "fmla z17.s, p4/M, z7.s, z2.s\n" "add x22, x22, #0x4\n" "fmla z21.s, p4/M, z7.s, z3.s\n" "fmla z25.s, p4/M, z7.s, z4.s\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z28.s }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z14.s, p4/M, z6.s, z1.s\n" - "fmla z18.s, p4/M, z6.s, z2.s\n" - "fmla z22.s, p4/M, z6.s, z3.s\n" - "fmla z26.s, p4/M, z6.s, z4.s\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" + "fmla z10.s, p4/M, z29.s, z0.s\n" + "fmla z14.s, p4/M, z29.s, z1.s\n" + "fmla z18.s, p4/M, z29.s, z2.s\n" + "fmla z22.s, p4/M, z29.s, z3.s\n" + "fmla z26.s, p4/M, z29.s, z4.s\n" + "fmla z11.s, p4/M, z28.s, z0.s\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1w { z6.s }, p4/Z, [x10]\n" - "fmla z15.s, p4/M, z7.s, z1.s\n" - "fmla z19.s, p4/M, z7.s, z2.s\n" + "fmla z15.s, p4/M, z28.s, z1.s\n" + "fmla z19.s, p4/M, z28.s, z2.s\n" "ld1rw { z1.s }, p4/Z, [x25]\n" "ld1rw { z2.s }, p4/Z, [x24]\n" - "fmla z23.s, p4/M, z7.s, z3.s\n" - "fmla z27.s, p4/M, z7.s, z4.s\n" + "fmla z23.s, p4/M, z28.s, z3.s\n" + "fmla z27.s, p4/M, z28.s, z4.s\n" "ld1rw { z3.s }, p4/Z, [x23]\n" "ld1rw { z4.s }, p4/Z, [x22]\n" "ld1w { z7.s }, p4/Z, [x10, #1, MUL VL]\n" @@ -939,23 +939,23 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "cmp x28, x20\n" "fmla z24.s, p4/M, z6.s, z4.s\n" "fmla z9.s, p4/M, z7.s, z0.s\n" - "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n" + "ld1w { z29.s }, p4/Z, [x10, #2, MUL VL]\n" "fmla z13.s, p4/M, z7.s, z1.s\n" "fmla z17.s, p4/M, z7.s, z2.s\n" "fmla z21.s, p4/M, z7.s, z3.s\n" "fmla z25.s, p4/M, z7.s, z4.s\n" - "ld1w { z7.s }, p4/Z, [x10, #3, MUL VL]\n" + "ld1w { z28.s }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "fmla z10.s, p4/M, z6.s, z0.s\n" - "fmla z14.s, p4/M, z6.s, z1.s\n" - "fmla z18.s, p4/M, z6.s, z2.s\n" - "fmla z22.s, p4/M, z6.s, z3.s\n" - "fmla z26.s, p4/M, z6.s, z4.s\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" - "fmla z15.s, p4/M, z7.s, z1.s\n" - "fmla z19.s, p4/M, z7.s, z2.s\n" - "fmla z23.s, p4/M, z7.s, z3.s\n" - "fmla z27.s, p4/M, z7.s, z4.s\n" + "fmla z10.s, p4/M, z29.s, z0.s\n" + "fmla z14.s, p4/M, z29.s, z1.s\n" + "fmla z18.s, p4/M, z29.s, z2.s\n" + "fmla z22.s, p4/M, z29.s, z3.s\n" + "fmla z26.s, p4/M, z29.s, z4.s\n" + "fmla z11.s, p4/M, z28.s, z0.s\n" + "fmla z15.s, p4/M, z28.s, z1.s\n" + "fmla z19.s, p4/M, z28.s, z2.s\n" + "fmla z23.s, p4/M, z28.s, z3.s\n" + "fmla z27.s, p4/M, z28.s, z4.s\n" "bne 54b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "add x25, x9, x20, LSL #2\n" @@ -964,49 +964,49 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "add x22, x23, x20, LSL #2\n" "tbz %x[flags], #1, 59f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p4/Z, [x20]\n" + "ld1rw { z29.s }, p4/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p4/Z, [x20]\n" - "fmin z8.s, p4/M, z8.s, z1.s\n" - "fmin z9.s, p4/M, z9.s, z1.s\n" - "fmin z10.s, p4/M, z10.s, z1.s\n" - "fmin z11.s, p4/M, z11.s, z1.s\n" - "fmin z12.s, p4/M, z12.s, z1.s\n" - "fmin z13.s, p4/M, z13.s, z1.s\n" - "fmin z14.s, p4/M, z14.s, z1.s\n" - "fmin z15.s, p4/M, z15.s, z1.s\n" - "fmin z16.s, p4/M, z16.s, z1.s\n" - "fmin z17.s, p4/M, z17.s, z1.s\n" - "fmin z18.s, p4/M, z18.s, z1.s\n" - "fmin z19.s, p4/M, z19.s, z1.s\n" - "fmin z20.s, p4/M, z20.s, z1.s\n" - "fmin z21.s, p4/M, z21.s, z1.s\n" - "fmin z22.s, p4/M, z22.s, z1.s\n" - "fmin z23.s, p4/M, z23.s, z1.s\n" - "fmin z24.s, p4/M, z24.s, z1.s\n" - "fmin z25.s, p4/M, z25.s, z1.s\n" - "fmin z26.s, p4/M, z26.s, z1.s\n" - "fmin z27.s, p4/M, z27.s, z1.s\n" - "fmax z8.s, p4/M, z8.s, z0.s\n" - "fmax z9.s, p4/M, z9.s, z0.s\n" - "fmax z10.s, p4/M, z10.s, z0.s\n" - "fmax z11.s, p4/M, z11.s, z0.s\n" - "fmax z12.s, p4/M, z12.s, z0.s\n" - "fmax z13.s, p4/M, z13.s, z0.s\n" - "fmax z14.s, p4/M, z14.s, z0.s\n" - "fmax z15.s, p4/M, z15.s, z0.s\n" - "fmax z16.s, p4/M, z16.s, z0.s\n" - "fmax z17.s, p4/M, z17.s, z0.s\n" - "fmax z18.s, p4/M, z18.s, z0.s\n" - "fmax z19.s, p4/M, z19.s, z0.s\n" - "fmax z20.s, p4/M, z20.s, z0.s\n" - "fmax z21.s, p4/M, z21.s, z0.s\n" - "fmax z22.s, p4/M, z22.s, z0.s\n" - "fmax z23.s, p4/M, z23.s, z0.s\n" - "fmax z24.s, p4/M, z24.s, z0.s\n" - "fmax z25.s, p4/M, z25.s, z0.s\n" - "fmax z26.s, p4/M, z26.s, z0.s\n" - "fmax z27.s, p4/M, z27.s, z0.s\n" + "ld1rw { z28.s }, p4/Z, [x20]\n" + "fmin z8.s, p4/M, z8.s, z29.s\n" + "fmin z9.s, p4/M, z9.s, z29.s\n" + "fmin z10.s, p4/M, z10.s, z29.s\n" + "fmin z11.s, p4/M, z11.s, z29.s\n" + "fmin z12.s, p4/M, z12.s, z29.s\n" + "fmin z13.s, p4/M, z13.s, z29.s\n" + "fmin z14.s, p4/M, z14.s, z29.s\n" + "fmin z15.s, p4/M, z15.s, z29.s\n" + "fmin z16.s, p4/M, z16.s, z29.s\n" + "fmin z17.s, p4/M, z17.s, z29.s\n" + "fmin z18.s, p4/M, z18.s, z29.s\n" + "fmin z19.s, p4/M, z19.s, z29.s\n" + "fmin z20.s, p4/M, z20.s, z29.s\n" + "fmin z21.s, p4/M, z21.s, z29.s\n" + "fmin z22.s, p4/M, z22.s, z29.s\n" + "fmin z23.s, p4/M, z23.s, z29.s\n" + "fmin z24.s, p4/M, z24.s, z29.s\n" + "fmin z25.s, p4/M, z25.s, z29.s\n" + "fmin z26.s, p4/M, z26.s, z29.s\n" + "fmin z27.s, p4/M, z27.s, z29.s\n" + "fmax z8.s, p4/M, z8.s, z28.s\n" + "fmax z9.s, p4/M, z9.s, z28.s\n" + "fmax z10.s, p4/M, z10.s, z28.s\n" + "fmax z11.s, p4/M, z11.s, z28.s\n" + "fmax z12.s, p4/M, z12.s, z28.s\n" + "fmax z13.s, p4/M, z13.s, z28.s\n" + "fmax z14.s, p4/M, z14.s, z28.s\n" + "fmax z15.s, p4/M, z15.s, z28.s\n" + "fmax z16.s, p4/M, z16.s, z28.s\n" + "fmax z17.s, p4/M, z17.s, z28.s\n" + "fmax z18.s, p4/M, z18.s, z28.s\n" + "fmax z19.s, p4/M, z19.s, z28.s\n" + "fmax z20.s, p4/M, z20.s, z28.s\n" + "fmax z21.s, p4/M, z21.s, z28.s\n" + "fmax z22.s, p4/M, z22.s, z28.s\n" + "fmax z23.s, p4/M, z23.s, z28.s\n" + "fmax z24.s, p4/M, z24.s, z28.s\n" + "fmax z25.s, p4/M, z25.s, z28.s\n" + "fmax z26.s, p4/M, z26.s, z28.s\n" + "fmax z27.s, p4/M, z27.s, z28.s\n" "59:" // Height 5: No activation "st1w { z8.s }, p3, [x9]\n" "st1w { z9.s }, p2, [x9, #1, MUL VL]\n" @@ -1081,35 +1081,35 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "63:" // Height 6: no bias "tbz %x[flags], #0, 64f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z8.s }, p3/Z, [x9]\n" + "add x24, x9, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" + "ld1w { z8.s }, p3/Z, [x9]\n" "add x22, x23, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n" - "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p3/Z, [x25]\n" - "ld1w { z13.s }, p2/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p1/Z, [x25, #2, MUL VL]\n" - "ld1w { z15.s }, p0/Z, [x25, #3, MUL VL]\n" - "ld1w { z16.s }, p3/Z, [x24]\n" - "ld1w { z17.s }, p2/Z, [x24, #1, MUL VL]\n" - "ld1w { z18.s }, p1/Z, [x24, #2, MUL VL]\n" - "ld1w { z19.s }, p0/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p3/Z, [x23]\n" - "ld1w { z21.s }, p2/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p1/Z, [x23, #2, MUL VL]\n" - "ld1w { z23.s }, p0/Z, [x23, #3, MUL VL]\n" - "ld1w { z24.s }, p3/Z, [x22]\n" - "ld1w { z25.s }, p2/Z, [x22, #1, MUL VL]\n" - "ld1w { z26.s }, p1/Z, [x22, #2, MUL VL]\n" - "ld1w { z27.s }, p0/Z, [x22, #3, MUL VL]\n" - "ld1w { z28.s }, p3/Z, [x21]\n" - "ld1w { z29.s }, p2/Z, [x21, #1, MUL VL]\n" - "ld1w { z30.s }, p1/Z, [x21, #2, MUL VL]\n" - "ld1w { z31.s }, p0/Z, [x21, #3, MUL VL]\n" + "ld1w { z12.s }, p3/Z, [x24]\n" + "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n" + "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n" + "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n" + "ld1w { z16.s }, p3/Z, [x23]\n" + "ld1w { z17.s }, p2/Z, [x23, #1, MUL VL]\n" + "ld1w { z18.s }, p1/Z, [x23, #2, MUL VL]\n" + "ld1w { z19.s }, p0/Z, [x23, #3, MUL VL]\n" + "ld1w { z20.s }, p3/Z, [x22]\n" + "ld1w { z21.s }, p2/Z, [x22, #1, MUL VL]\n" + "ld1w { z22.s }, p1/Z, [x22, #2, MUL VL]\n" + "ld1w { z23.s }, p0/Z, [x22, #3, MUL VL]\n" + "ld1w { z24.s }, p3/Z, [x21]\n" + "ld1w { z25.s }, p2/Z, [x21, #1, MUL VL]\n" + "ld1w { z26.s }, p1/Z, [x21, #2, MUL VL]\n" + "ld1w { z27.s }, p0/Z, [x21, #3, MUL VL]\n" + "ld1w { z28.s }, p3/Z, [x20]\n" + "ld1w { z29.s }, p2/Z, [x20, #1, MUL VL]\n" + "ld1w { z30.s }, p1/Z, [x20, #2, MUL VL]\n" + "ld1w { z31.s }, p0/Z, [x20, #3, MUL VL]\n" "b 65f\n" "64:" // Height 6: no accumulate "mov z8.b, #0x0\n" @@ -1141,16 +1141,16 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "66:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 67f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "ldr x21, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "ldr x21, [x20, #0x28]\n" "cbnz x28, 68f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1162,11 +1162,11 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "b 68f\n" "67:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "68:" // Height 6: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -1355,7 +1355,6 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "74:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "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" @@ -1363,4 +1362,4 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx ( } } // namespace arm_gemm -#endif // ARM_COMPUTE_ENABLE_SVE +#endif // ARM_COMPUTE_ENABLE_SVE |