diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp | 678 |
1 files changed, 339 insertions, 339 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp index 32fcac3a45..8e4fd4388e 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp @@ -163,11 +163,11 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "7:" // 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, 8f\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, 9f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -183,12 +183,12 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "10:" // 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]\n" - "ld1w { z7.s }, p4/Z, [x9]\n" + "ld1w { z17.s }, p4/Z, [x10]\n" + "ld1w { z16.s }, p4/Z, [x9]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\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" "add x26, x26, #0x4\n" "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -201,12 +201,12 @@ void sve_ffhybrid_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]\n" - "ld1w { z7.s }, p4/Z, [x9]\n" + "ld1w { z17.s }, p4/Z, [x10]\n" + "ld1w { z16.s }, p4/Z, [x9]\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 x12, x12, #1\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" @@ -214,17 +214,17 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "bne 7b\n" "tbz %x[flags], #1, 12f\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" "12:" // Height 1: No activation "st1w { z8.s }, p3, [x13]\n" "st1w { z9.s }, p2, [x13, #1, MUL VL]\n" @@ -285,15 +285,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "17:" // Height 2: no bias "tbz %x[flags], #0, 18f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" + "add x20, x13, x20, LSL #2\n" "ld1w { z8.s }, p3/Z, [x13]\n" "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p0/Z, [x13, #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 19f\n" "18:" // Height 2: no accumulate "mov z8.b, #0x0\n" @@ -309,12 +309,12 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "20:" // 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, 21f\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, 22f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -322,7 +322,7 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "b 22f\n" "21:" // 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" "22:" // Height 2: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -333,19 +333,19 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "23:" // 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]\n" + "ld1w { z17.s }, p4/Z, [x10]\n" "addvl x12, x12, #1\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, [x9]\n" + "ld1w { z16.s }, p4/Z, [x9]\n" "addvl x11, x11, #1\n" "add x26, x26, #0x4\n" "subs x27, x27, #0x1\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" "add x25, x25, #0x4\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" "addvl x10, x10, #1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1rw { z1.s }, p4/Z, [x25]\n" @@ -357,18 +357,18 @@ void sve_ffhybrid_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]\n" + "ld1w { z17.s }, p4/Z, [x10]\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, [x9]\n" + "ld1w { z16.s }, p4/Z, [x9]\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 x12, x12, #1\n" "addvl x11, x11, #1\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" "addvl x10, x10, #1\n" "addvl x9, x9, #1\n" "bne 20b\n" @@ -376,25 +376,25 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "add x25, x13, x20, LSL #2\n" "tbz %x[flags], #1, 25f\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" "25:" // Height 2: No activation "st1w { z8.s }, p3, [x13]\n" "st1w { z9.s }, p2, [x13, #1, MUL VL]\n" @@ -463,20 +463,20 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "30:" // Height 3: no bias "tbz %x[flags], #0, 31f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x21, x13, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z8.s }, p3/Z, [x13]\n" "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p0/Z, [x13, #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 32f\n" "31:" // Height 3: no accumulate "mov z8.b, #0x0\n" @@ -496,13 +496,13 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "33:" // 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, 34f\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, 35f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -511,8 +511,8 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "b 35f\n" "34:" // 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" "35:" // Height 3: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -528,22 +528,22 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "addvl x11, x11, #1\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]\n" + "ld1w { z21.s }, p4/Z, [x10]\n" "add x26, x26, #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, [x9]\n" + "ld1w { z20.s }, p4/Z, [x9]\n" "subs x27, x27, #0x1\n" "add x25, x25, #0x4\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" "addvl x10, x10, #1\n" "addvl x9, x9, #1\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 { z0.s }, p4/Z, [x26]\n" "ld1rw { z1.s }, p4/Z, [x25]\n" "ld1rw { z2.s }, p4/Z, [x24]\n" @@ -557,54 +557,54 @@ void sve_ffhybrid_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]\n" + "ld1w { z21.s }, p4/Z, [x10]\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, [x9]\n" + "ld1w { z20.s }, p4/Z, [x9]\n" "addvl x12, x12, #1\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, z21.s, z0.s\n" + "fmla z14.s, p4/M, z21.s, z1.s\n" "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" - "fmla z18.s, p4/M, z6.s, z2.s\n" - "fmla z11.s, p4/M, z7.s, z0.s\n" + "fmla z18.s, p4/M, z21.s, z2.s\n" + "fmla z11.s, p4/M, z20.s, z0.s\n" "addvl x9, x9, #1\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" "bne 33b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "add x25, x13, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" "tbz %x[flags], #1, 38f\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" "38:" // Height 3: No activation "st1w { z8.s }, p3, [x13]\n" "st1w { z9.s }, p2, [x13, #1, MUL VL]\n" @@ -681,25 +681,25 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "43:" // Height 4: no bias "tbz %x[flags], #0, 44f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x22, x13, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "ld1w { z8.s }, p3/Z, [x13]\n" - "add x23, x24, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p0/Z, [x13, #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 45f\n" "44:" // Height 4: no accumulate "mov z8.b, #0x0\n" @@ -723,14 +723,14 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "46:" // 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, 47f\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, 48f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -740,9 +740,9 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "b 48f\n" "47:" // 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" "48:" // Height 4: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -759,7 +759,7 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "addvl x11, x11, #1\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]\n" + "ld1w { z25.s }, p4/Z, [x10]\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" @@ -767,22 +767,22 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "add x25, x25, #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, [x9]\n" + "ld1w { z24.s }, p4/Z, [x9]\n" "add x24, x24, #0x4\n" "add x23, x23, #0x4\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, z25.s, z0.s\n" + "fmla z14.s, p4/M, z25.s, z1.s\n" "addvl x10, x10, #1\n" - "fmla z18.s, p4/M, z6.s, z2.s\n" - "fmla z22.s, p4/M, z6.s, z3.s\n" + "fmla z18.s, p4/M, z25.s, z2.s\n" + "fmla z22.s, p4/M, z25.s, z3.s\n" "addvl x9, x9, #1\n" "ld1w { z6.s }, p4/Z, [x12]\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, [x11]\n" @@ -794,7 +794,7 @@ void sve_ffhybrid_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]\n" + "ld1w { z25.s }, p4/Z, [x10]\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" @@ -802,17 +802,17 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "addvl x11, x11, #1\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, [x9]\n" + "ld1w { z24.s }, p4/Z, [x9]\n" "addvl x10, x10, #1\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, z25.s, z0.s\n" + "fmla z14.s, p4/M, z25.s, z1.s\n" "addvl x9, x9, #1\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 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 46b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "add x25, x13, x20, LSL #2\n" @@ -820,41 +820,41 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "add x23, x24, x20, LSL #2\n" "tbz %x[flags], #1, 51f\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" "51:" // Height 4: No activation "st1w { z8.s }, p3, [x13]\n" "st1w { z9.s }, p2, [x13, #1, MUL VL]\n" @@ -939,30 +939,30 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "56:" // Height 5: no bias "tbz %x[flags], #0, 57f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z8.s }, p3/Z, [x13]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x13, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" + "ld1w { z8.s }, p3/Z, [x13]\n" + "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p0/Z, [x13, #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 58f\n" "57:" // Height 5: no accumulate "mov z8.b, #0x0\n" @@ -990,15 +990,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "59:" // 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, 60f\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, 61f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1009,10 +1009,10 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "b 61f\n" "60:" // 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" "61:" // Height 5: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -1034,7 +1034,7 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "subs x27, x27, #0x1\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]\n" + "ld1w { z29.s }, p4/Z, [x10]\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" @@ -1042,24 +1042,24 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "add x23, x23, #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, [x9]\n" + "ld1w { z28.s }, p4/Z, [x9]\n" "add x22, x22, #0x4\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, z29.s, z0.s\n" + "fmla z14.s, p4/M, z29.s, z1.s\n" "addvl x10, x10, #1\n" "addvl x9, x9, #1\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 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, [x12]\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, [x11]\n" @@ -1075,25 +1075,25 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "addvl x12, x12, #1\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]\n" + "ld1w { z29.s }, p4/Z, [x10]\n" "addvl x11, x11, #1\n" "fmla z13.s, p4/M, z7.s, z1.s\n" "fmla z17.s, p4/M, z7.s, z2.s\n" "addvl x10, x10, #1\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, [x9]\n" + "ld1w { z28.s }, p4/Z, [x9]\n" "addvl x9, x9, #1\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 59b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "add x25, x13, x20, LSL #2\n" @@ -1102,49 +1102,49 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "add x22, x23, x20, LSL #2\n" "tbz %x[flags], #1, 64f\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" "64:" // Height 5: No activation "st1w { z8.s }, p3, [x13]\n" "st1w { z9.s }, p2, [x13, #1, MUL VL]\n" @@ -1240,35 +1240,35 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "69:" // Height 6: no bias "tbz %x[flags], #0, 70f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z8.s }, p3/Z, [x13]\n" + "add x24, x13, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" + "ld1w { z8.s }, p3/Z, [x13]\n" "add x22, x23, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n" - "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z11.s }, p0/Z, [x13, #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 71f\n" "70:" // Height 6: no accumulate "mov z8.b, #0x0\n" @@ -1300,16 +1300,16 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "72:" // 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, 73f\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, 74f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1321,11 +1321,11 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( "b 74f\n" "73:" // 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" "74:" // Height 6: input setup done "subs x27, x27, #0x1\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -1527,4 +1527,4 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx ( } } // namespace arm_gemm -#endif // ARM_COMPUTE_ENABLE_SVE +#endif // ARM_COMPUTE_ENABLE_SVE |