diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp | 467 |
1 files changed, 233 insertions, 234 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp index 51e9aa1b40..1a483210f3 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp @@ -115,11 +115,11 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "5:" // 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, 6f\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, 7f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -135,12 +135,12 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "8:" // Height 1: Multiply loop: Main loop "sdot z8.s, z6.b, z0.b\n" "sdot z9.s, z7.b, z0.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" "add x26, x26, #0x4\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z11.s, z7.b, z0.b\n" + "sdot z10.s, z17.b, z0.b\n" + "sdot z11.s, z16.b, z0.b\n" "subs x27, x27, #0x4\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1b { z6.b }, p4/Z, [x10]\n" @@ -150,12 +150,12 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "sdot z8.s, z6.b, z0.b\n" "sdot z9.s, z7.b, z0.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z11.s, z7.b, z0.b\n" + "sdot z10.s, z17.b, z0.b\n" + "sdot z11.s, z16.b, z0.b\n" "addvl x10, x10, #4\n" "bne 5b\n" "st1w { z8.s }, p3, [x9]\n" @@ -183,15 +183,15 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "whilelt p0.s, x20, x11\n" "tbz %x[flags], #0, 13f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, 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, [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 { 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 14f\n" "13:" // Height 2: no accumulate "mov z8.s, #0x0\n" @@ -207,12 +207,12 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "15:" // 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, 16f\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, 17f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -220,7 +220,7 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "b 17f\n" "16:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" + "add x25, x26, x21\n" "17:" // Height 2: input setup done "subs x27, x27, #0x4\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -231,18 +231,18 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "18:" // Height 2: Multiply loop: Main loop "sdot z8.s, z6.b, z0.b\n" "sdot z12.s, z6.b, z1.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n" "add x26, x26, #0x4\n" "sdot z9.s, z7.b, z0.b\n" "sdot z13.s, z7.b, z1.b\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" "subs x27, x27, #0x4\n" "add x25, x25, #0x4\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z14.s, z6.b, z1.b\n" - "sdot z11.s, z7.b, z0.b\n" - "sdot z15.s, z7.b, z1.b\n" + "sdot z10.s, z17.b, z0.b\n" + "sdot z14.s, z17.b, z1.b\n" + "sdot z11.s, z16.b, z0.b\n" + "sdot z15.s, z16.b, z1.b\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1rw { z1.s }, p4/Z, [x25]\n" "ld1b { z6.b }, p4/Z, [x10]\n" @@ -252,29 +252,29 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "sdot z8.s, z6.b, z0.b\n" "sdot z12.s, z6.b, z1.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n" "sdot z9.s, z7.b, z0.b\n" "sdot z13.s, z7.b, z1.b\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z14.s, z6.b, z1.b\n" + "sdot z10.s, z17.b, z0.b\n" + "sdot z14.s, z17.b, z1.b\n" "addvl x10, x10, #4\n" - "sdot z11.s, z7.b, z0.b\n" - "sdot z15.s, z7.b, z1.b\n" + "sdot z11.s, z16.b, z0.b\n" + "sdot z15.s, z16.b, z1.b\n" "bne 15b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" + "add x20, x9, x20, LSL #2\n" "st1w { z8.s }, p3, [x9]\n" "st1w { z9.s }, p2, [x9, #1, MUL VL]\n" "st1w { z10.s }, p1, [x9, #2, MUL VL]\n" "st1w { z11.s }, p0, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z12.s }, p3, [x24]\n" - "st1w { z13.s }, p2, [x24, #1, MUL VL]\n" - "st1w { z14.s }, p1, [x24, #2, MUL VL]\n" - "st1w { z15.s }, p0, [x24, #3, MUL VL]\n" + "st1w { z12.s }, p3, [x20]\n" + "st1w { z13.s }, p2, [x20, #1, MUL VL]\n" + "st1w { z14.s }, p1, [x20, #2, MUL VL]\n" + "st1w { z15.s }, p0, [x20, #3, MUL VL]\n" "20:" // Height 2: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -295,20 +295,20 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "whilelt p0.s, x20, x11\n" "tbz %x[flags], #0, 23f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, 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, [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 { 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 24f\n" "23:" // Height 3: no accumulate "mov z8.s, #0x0\n" @@ -328,13 +328,13 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "25:" // 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, 26f\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, 27f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -343,8 +343,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "b 27f\n" "26:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" "27:" // Height 3: input setup done "subs x27, x27, #0x4\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -360,21 +360,21 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "subs x27, x27, #0x4\n" "sdot z16.s, z6.b, z2.b\n" "sdot z9.s, z7.b, z0.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n" "add x25, x25, #0x4\n" "sdot z13.s, z7.b, z1.b\n" "sdot z17.s, z7.b, z2.b\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z20.b }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" "add x24, x24, #0x4\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z14.s, z6.b, z1.b\n" - "sdot z18.s, z6.b, z2.b\n" - "sdot z11.s, z7.b, z0.b\n" + "sdot z10.s, z21.b, z0.b\n" + "sdot z14.s, z21.b, z1.b\n" + "sdot z18.s, z21.b, z2.b\n" + "sdot z11.s, z20.b, z0.b\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1b { z6.b }, p4/Z, [x10]\n" - "sdot z15.s, z7.b, z1.b\n" - "sdot z19.s, z7.b, z2.b\n" + "sdot z15.s, z20.b, z1.b\n" + "sdot z19.s, z20.b, z2.b\n" "ld1rw { z1.s }, p4/Z, [x25]\n" "ld1rw { z2.s }, p4/Z, [x24]\n" "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n" @@ -386,35 +386,35 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "add x28, x28, #0x1\n" "sdot z16.s, z6.b, z2.b\n" "sdot z9.s, z7.b, z0.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n" "cmp x28, x20\n" "sdot z13.s, z7.b, z1.b\n" "sdot z17.s, z7.b, z2.b\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z20.b }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z14.s, z6.b, z1.b\n" - "sdot z18.s, z6.b, z2.b\n" - "sdot z11.s, z7.b, z0.b\n" - "sdot z15.s, z7.b, z1.b\n" - "sdot z19.s, z7.b, z2.b\n" + "sdot z10.s, z21.b, z0.b\n" + "sdot z14.s, z21.b, z1.b\n" + "sdot z18.s, z21.b, z2.b\n" + "sdot z11.s, z20.b, z0.b\n" + "sdot z15.s, z20.b, z1.b\n" + "sdot z19.s, z20.b, z2.b\n" "bne 25b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" + "add x21, x9, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "st1w { z8.s }, p3, [x9]\n" "st1w { z9.s }, p2, [x9, #1, MUL VL]\n" "st1w { z10.s }, p1, [x9, #2, MUL VL]\n" "st1w { z11.s }, p0, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z12.s }, p3, [x24]\n" - "st1w { z13.s }, p2, [x24, #1, MUL VL]\n" - "st1w { z14.s }, p1, [x24, #2, MUL VL]\n" - "st1w { z15.s }, p0, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p3, [x23]\n" - "st1w { z17.s }, p2, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p1, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p0, [x23, #3, MUL VL]\n" + "st1w { z12.s }, p3, [x21]\n" + "st1w { z13.s }, p2, [x21, #1, MUL VL]\n" + "st1w { z14.s }, p1, [x21, #2, MUL VL]\n" + "st1w { z15.s }, p0, [x21, #3, MUL VL]\n" + "st1w { z16.s }, p3, [x20]\n" + "st1w { z17.s }, p2, [x20, #1, MUL VL]\n" + "st1w { z18.s }, p1, [x20, #2, MUL VL]\n" + "st1w { z19.s }, p0, [x20, #3, MUL VL]\n" "30:" // Height 3: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -435,25 +435,25 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "whilelt p0.s, x20, x11\n" "tbz %x[flags], #0, 33f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, 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 x22, x23, 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, [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 { 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 34f\n" "33:" // Height 4: no accumulate "mov z8.s, #0x0\n" @@ -477,14 +477,14 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "35:" // 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, 36f\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, 37f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -494,9 +494,9 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "b 37f\n" "36:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" + "add x23, x24, x21\n" "37:" // Height 4: input setup done "subs x27, x27, #0x4\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -513,7 +513,7 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "subs x27, x27, #0x4\n" "sdot z16.s, z6.b, z2.b\n" "sdot z20.s, z6.b, z3.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z25.b }, p4/Z, [x10, #2, MUL VL]\n" "add x25, x25, #0x4\n" "sdot z9.s, z7.b, z0.b\n" "sdot z13.s, z7.b, z1.b\n" @@ -521,19 +521,19 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "add x23, x23, #0x4\n" "sdot z17.s, z7.b, z2.b\n" "sdot z21.s, z7.b, z3.b\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z24.b }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z14.s, z6.b, z1.b\n" - "sdot z18.s, z6.b, z2.b\n" - "sdot z22.s, z6.b, z3.b\n" + "sdot z10.s, z25.b, z0.b\n" + "sdot z14.s, z25.b, z1.b\n" + "sdot z18.s, z25.b, z2.b\n" + "sdot z22.s, z25.b, z3.b\n" "ld1b { z6.b }, p4/Z, [x10]\n" - "sdot z11.s, z7.b, z0.b\n" - "sdot z15.s, z7.b, z1.b\n" + "sdot z11.s, z24.b, z0.b\n" + "sdot z15.s, z24.b, z1.b\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1rw { z1.s }, p4/Z, [x25]\n" - "sdot z19.s, z7.b, z2.b\n" - "sdot z23.s, z7.b, z3.b\n" + "sdot z19.s, z24.b, z2.b\n" + "sdot z23.s, z24.b, z3.b\n" "ld1rw { z2.s }, p4/Z, [x24]\n" "ld1rw { z3.s }, p4/Z, [x23]\n" "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n" @@ -545,44 +545,44 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "add x28, x28, #0x1\n" "sdot z16.s, z6.b, z2.b\n" "sdot z20.s, z6.b, z3.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z25.b }, p4/Z, [x10, #2, MUL VL]\n" "cmp x28, x20\n" "sdot z9.s, z7.b, z0.b\n" "sdot z13.s, z7.b, z1.b\n" "sdot z17.s, z7.b, z2.b\n" "sdot z21.s, z7.b, z3.b\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z24.b }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z14.s, z6.b, z1.b\n" - "sdot z18.s, z6.b, z2.b\n" - "sdot z22.s, z6.b, z3.b\n" - "sdot z11.s, z7.b, z0.b\n" - "sdot z15.s, z7.b, z1.b\n" - "sdot z19.s, z7.b, z2.b\n" - "sdot z23.s, z7.b, z3.b\n" + "sdot z10.s, z25.b, z0.b\n" + "sdot z14.s, z25.b, z1.b\n" + "sdot z18.s, z25.b, z2.b\n" + "sdot z22.s, z25.b, z3.b\n" + "sdot z11.s, z24.b, z0.b\n" + "sdot z15.s, z24.b, z1.b\n" + "sdot z19.s, z24.b, z2.b\n" + "sdot z23.s, z24.b, z3.b\n" "bne 35b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" + "add x22, x9, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "st1w { z8.s }, p3, [x9]\n" - "add x22, x23, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "st1w { z9.s }, p2, [x9, #1, MUL VL]\n" "st1w { z10.s }, p1, [x9, #2, MUL VL]\n" "st1w { z11.s }, p0, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z12.s }, p3, [x24]\n" - "st1w { z13.s }, p2, [x24, #1, MUL VL]\n" - "st1w { z14.s }, p1, [x24, #2, MUL VL]\n" - "st1w { z15.s }, p0, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p3, [x23]\n" - "st1w { z17.s }, p2, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p1, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p0, [x23, #3, MUL VL]\n" - "st1w { z20.s }, p3, [x22]\n" - "st1w { z21.s }, p2, [x22, #1, MUL VL]\n" - "st1w { z22.s }, p1, [x22, #2, MUL VL]\n" - "st1w { z23.s }, p0, [x22, #3, MUL VL]\n" + "st1w { z12.s }, p3, [x22]\n" + "st1w { z13.s }, p2, [x22, #1, MUL VL]\n" + "st1w { z14.s }, p1, [x22, #2, MUL VL]\n" + "st1w { z15.s }, p0, [x22, #3, MUL VL]\n" + "st1w { z16.s }, p3, [x21]\n" + "st1w { z17.s }, p2, [x21, #1, MUL VL]\n" + "st1w { z18.s }, p1, [x21, #2, MUL VL]\n" + "st1w { z19.s }, p0, [x21, #3, MUL VL]\n" + "st1w { z20.s }, p3, [x20]\n" + "st1w { z21.s }, p2, [x20, #1, MUL VL]\n" + "st1w { z22.s }, p1, [x20, #2, MUL VL]\n" + "st1w { z23.s }, p0, [x20, #3, MUL VL]\n" "40:" // Height 4: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -603,30 +603,30 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "whilelt p0.s, x20, x11\n" "tbz %x[flags], #0, 43f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "ld1w { z8.s }, p3/Z, [x9]\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, [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 { 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 44f\n" "43:" // Height 5: no accumulate "mov z8.s, #0x0\n" @@ -654,15 +654,15 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "45:" // 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, 46f\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, 47f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -673,10 +673,10 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "b 47f\n" "46:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" - "add x22, x23, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" + "add x23, x24, x21\n" + "add x22, x23, x21\n" "47:" // Height 5: input setup done "subs x27, x27, #0x4\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -698,29 +698,29 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "add x24, x24, #0x4\n" "sdot z24.s, z6.b, z4.b\n" "sdot z9.s, z7.b, z0.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n" "add x23, x23, #0x4\n" "sdot z13.s, z7.b, z1.b\n" "sdot z17.s, z7.b, z2.b\n" "add x22, x22, #0x4\n" "sdot z21.s, z7.b, z3.b\n" "sdot z25.s, z7.b, z4.b\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z28.b }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z14.s, z6.b, z1.b\n" - "sdot z18.s, z6.b, z2.b\n" - "sdot z22.s, z6.b, z3.b\n" - "sdot z26.s, z6.b, z4.b\n" - "sdot z11.s, z7.b, z0.b\n" + "sdot z10.s, z29.b, z0.b\n" + "sdot z14.s, z29.b, z1.b\n" + "sdot z18.s, z29.b, z2.b\n" + "sdot z22.s, z29.b, z3.b\n" + "sdot z26.s, z29.b, z4.b\n" + "sdot z11.s, z28.b, z0.b\n" "ld1rw { z0.s }, p4/Z, [x26]\n" "ld1b { z6.b }, p4/Z, [x10]\n" - "sdot z15.s, z7.b, z1.b\n" - "sdot z19.s, z7.b, z2.b\n" + "sdot z15.s, z28.b, z1.b\n" + "sdot z19.s, z28.b, z2.b\n" "ld1rw { z1.s }, p4/Z, [x25]\n" "ld1rw { z2.s }, p4/Z, [x24]\n" - "sdot z23.s, z7.b, z3.b\n" - "sdot z27.s, z7.b, z4.b\n" + "sdot z23.s, z28.b, z3.b\n" + "sdot z27.s, z28.b, z4.b\n" "ld1rw { z3.s }, p4/Z, [x23]\n" "ld1rw { z4.s }, p4/Z, [x22]\n" "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n" @@ -735,50 +735,50 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "cmp x28, x20\n" "sdot z24.s, z6.b, z4.b\n" "sdot z9.s, z7.b, z0.b\n" - "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n" + "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n" "sdot z13.s, z7.b, z1.b\n" "sdot z17.s, z7.b, z2.b\n" "sdot z21.s, z7.b, z3.b\n" "sdot z25.s, z7.b, z4.b\n" - "ld1b { z7.b }, p4/Z, [x10, #3, MUL VL]\n" + "ld1b { z28.b }, p4/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" - "sdot z10.s, z6.b, z0.b\n" - "sdot z14.s, z6.b, z1.b\n" - "sdot z18.s, z6.b, z2.b\n" - "sdot z22.s, z6.b, z3.b\n" - "sdot z26.s, z6.b, z4.b\n" - "sdot z11.s, z7.b, z0.b\n" - "sdot z15.s, z7.b, z1.b\n" - "sdot z19.s, z7.b, z2.b\n" - "sdot z23.s, z7.b, z3.b\n" - "sdot z27.s, z7.b, z4.b\n" + "sdot z10.s, z29.b, z0.b\n" + "sdot z14.s, z29.b, z1.b\n" + "sdot z18.s, z29.b, z2.b\n" + "sdot z22.s, z29.b, z3.b\n" + "sdot z26.s, z29.b, z4.b\n" + "sdot z11.s, z28.b, z0.b\n" + "sdot z15.s, z28.b, z1.b\n" + "sdot z19.s, z28.b, z2.b\n" + "sdot z23.s, z28.b, z3.b\n" + "sdot z27.s, z28.b, z4.b\n" "bne 45b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "st1w { z8.s }, p3, [x9]\n" + "add x23, x9, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" + "st1w { z8.s }, p3, [x9]\n" "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "st1w { z9.s }, p2, [x9, #1, MUL VL]\n" "st1w { z10.s }, p1, [x9, #2, MUL VL]\n" "st1w { z11.s }, p0, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z12.s }, p3, [x24]\n" - "st1w { z13.s }, p2, [x24, #1, MUL VL]\n" - "st1w { z14.s }, p1, [x24, #2, MUL VL]\n" - "st1w { z15.s }, p0, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p3, [x23]\n" - "st1w { z17.s }, p2, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p1, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p0, [x23, #3, MUL VL]\n" - "st1w { z20.s }, p3, [x22]\n" - "st1w { z21.s }, p2, [x22, #1, MUL VL]\n" - "st1w { z22.s }, p1, [x22, #2, MUL VL]\n" - "st1w { z23.s }, p0, [x22, #3, MUL VL]\n" - "st1w { z24.s }, p3, [x21]\n" - "st1w { z25.s }, p2, [x21, #1, MUL VL]\n" - "st1w { z26.s }, p1, [x21, #2, MUL VL]\n" - "st1w { z27.s }, p0, [x21, #3, MUL VL]\n" + "st1w { z12.s }, p3, [x23]\n" + "st1w { z13.s }, p2, [x23, #1, MUL VL]\n" + "st1w { z14.s }, p1, [x23, #2, MUL VL]\n" + "st1w { z15.s }, p0, [x23, #3, MUL VL]\n" + "st1w { z16.s }, p3, [x22]\n" + "st1w { z17.s }, p2, [x22, #1, MUL VL]\n" + "st1w { z18.s }, p1, [x22, #2, MUL VL]\n" + "st1w { z19.s }, p0, [x22, #3, MUL VL]\n" + "st1w { z20.s }, p3, [x21]\n" + "st1w { z21.s }, p2, [x21, #1, MUL VL]\n" + "st1w { z22.s }, p1, [x21, #2, MUL VL]\n" + "st1w { z23.s }, p0, [x21, #3, MUL VL]\n" + "st1w { z24.s }, p3, [x20]\n" + "st1w { z25.s }, p2, [x20, #1, MUL VL]\n" + "st1w { z26.s }, p1, [x20, #2, MUL VL]\n" + "st1w { z27.s }, p0, [x20, #3, MUL VL]\n" "50:" // Height 5: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -862,16 +862,16 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "55:" // 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, 56f\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, 57f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -883,11 +883,11 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "b 57f\n" "56:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" - "add x22, x23, x20\n" - "add x21, x22, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" + "add x23, x24, x21\n" + "add x22, x23, x21\n" + "add x21, x22, x21\n" "57:" // Height 6: input setup done "subs x27, x27, #0x4\n" "ld1rw { z0.s }, p4/Z, [x26]\n" @@ -1022,7 +1022,6 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "62:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [flags] "r" (flags), [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", "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" @@ -1030,4 +1029,4 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx ( } } // namespace arm_gemm -#endif // ARM_COMPUTE_ENABLE_SVE +#endif // ARM_COMPUTE_ENABLE_SVE |