diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp | 2308 |
1 files changed, 1110 insertions, 1198 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp index 0e45b06765..863325f7f5 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp @@ -27,6 +27,7 @@ #include "../../utils.hpp" #include <cassert> +#include <limits> namespace arm_gemm { @@ -94,230 +95,219 @@ void sve_hybrid_fp32_mla_8x1VL ( "ptrue p2.b\n" "1:" // Row loop "cmp %x[M], #0x8\n" - "bge 99f\n" + "bge 92f\n" "cmp %x[M], #0x6\n" - "bgt 85f\n" - "beq 71f\n" + "bgt 79f\n" + "beq 66f\n" "cmp %x[M], #0x4\n" - "bgt 57f\n" - "beq 43f\n" + "bgt 53f\n" + "beq 40f\n" "cmp %x[M], #0x2\n" - "bgt 29f\n" - "beq 15f\n" - "ldr x6, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x7, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x8, %x[bias]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 2f\n" - "ldr x17, [%x[output_ptr], #0x0]\n" - "add x17, x17, x19, LSL #2\n" - "b 3f\n" - "2:" // Height 1: setup direct output - "mov x17, %x[output_ptr]\n" - "3:" // Height 1: Column loop + "bgt 27f\n" + "beq 14f\n" + "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x11, %x[bias]\n" + "mov x10, %x[output_ptr]\n" + "2:" // Height 1: Column loop "mov x19, #0x0\n" - "whilelt p1.s, x19, x6\n" - "cbz x8, 4f\n" - "ld1w { z24.s }, p2/Z, [x8]\n" - "addvl x8, x8, #1\n" - "b 6f\n" - "4:" // Height 1: no bias - "tbz %x[flags], #0, 5f\n" - "ld1w { z24.s }, p1/Z, [x17]\n" - "b 6f\n" - "5:" // Height 1: no accumulate + "whilelt p1.s, x19, x13\n" + "cbz x11, 3f\n" + "ld1w { z24.s }, p2/Z, [x11]\n" + "addvl x11, x11, #1\n" + "b 5f\n" + "3:" // Height 1: no bias + "tbz %x[flags], #0, 4f\n" + "ld1w { z24.s }, p1/Z, [x10]\n" + "b 5f\n" + "4:" // Height 1: no accumulate "mov z24.b, #0x0\n" - "6:" // Height 1: setup done - "mov x16, #0x0\n" - "7:" // Height 1: String loop + "5:" // Height 1: setup done + "mov x9, #0x0\n" + "6:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w15, [x20, x16, LSL #0x2]\n" - "tbz %x[flags], #3, 8f\n" - "ldr x20, [%x[input_ptr], x16, LSL #0x3]\n" + "ldr w28, [x20, x9, LSL #0x2]\n" + "tbz %x[flags], #3, 7f\n" + "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x14, [x20, #0x0]\n" - "cbnz x16, 9f\n" + "ldr x27, [x20, #0x0]\n" + "cbnz x9, 8f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x14, x14, x19, LSL #2\n" - "b 9f\n" - "8:" // Height 1: setup direct input - "mov x14, %x[input_ptr]\n" - "9:" // Height 1: input setup done - "cmp x15, #0x4\n" - "ble 11f\n" - "10:" // Height 1: Multiply loop: Main loop head - "ld1w { z8.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "ld1w { z9.s }, p2/Z, [x7, #1, MUL VL]\n" - "sub x15, x15, #0x4\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" + "add x27, x27, x19, LSL #2\n" + "b 8f\n" + "7:" // Height 1: setup direct input + "mov x27, %x[input_ptr]\n" + "8:" // Height 1: input setup done + "cmp x28, #0x4\n" + "ble 10f\n" + "9:" // Height 1: Multiply loop: Main loop head + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n" + "sub x28, x28, #0x4\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" "fmla z24.s, z8.s, z0.s[0]\n" - "ld1w { z10.s }, p2/Z, [x7, #2, MUL VL]\n" - "add x14, x14, #0x10\n" + "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" + "add x27, x27, #0x10\n" "fmla z24.s, z9.s, z0.s[1]\n" - "ld1w { z11.s }, p2/Z, [x7, #3, MUL VL]\n" - "cmp x15, #0x4\n" + "ld1w { z11.s }, p2/Z, [x12, #3, MUL VL]\n" + "cmp x28, #0x4\n" "fmla z24.s, z10.s, z0.s[2]\n" - "prfm pldl1keep, [x14, #0x80]\n" - "addvl x7, x7, #4\n" + "prfm pldl1keep, [x27, #0x80]\n" + "addvl x12, x12, #4\n" "fmla z24.s, z11.s, z0.s[3]\n" - "bgt 10b\n" - "11:" // Height 1: Multiply loop: Single iteration only - "ld1w { z12.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "subs x15, x15, #0x1\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" - "fmla z24.s, z12.s, z0.s[0]\n" - "add x14, x14, #0x10\n" - "addvl x7, x7, #1\n" - "ble 12f\n" - "ld1w { z13.s }, p2/Z, [x7]\n" - "fmla z24.s, z13.s, z0.s[1]\n" - "subs x15, x15, #0x1\n" - "addvl x7, x7, #1\n" - "ble 12f\n" - "ld1w { z14.s }, p2/Z, [x7]\n" - "fmla z24.s, z14.s, z0.s[2]\n" - "subs x15, x15, #0x1\n" - "addvl x7, x7, #1\n" - "ble 12f\n" - "ld1w { z15.s }, p2/Z, [x7]\n" - "fmla z24.s, z15.s, z0.s[3]\n" - "addvl x7, x7, #1\n" - "12:" // Height 1: Multiply loop: multiply skip - "prfm pldl1keep, [x14, #0x80]\n" - "add x16, x16, #0x1\n" + "bgt 9b\n" + "10:" // Height 1: Multiply loop: Single iteration only + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "subs x28, x28, #0x1\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" + "fmla z24.s, z8.s, z0.s[0]\n" + "add x27, x27, #0x10\n" + "addvl x12, x12, #1\n" + "ble 11f\n" + "ld1w { z9.s }, p2/Z, [x12]\n" + "fmla z24.s, z9.s, z0.s[1]\n" + "subs x28, x28, #0x1\n" + "addvl x12, x12, #1\n" + "ble 11f\n" + "ld1w { z10.s }, p2/Z, [x12]\n" + "fmla z24.s, z10.s, z0.s[2]\n" + "subs x28, x28, #0x1\n" + "addvl x12, x12, #1\n" + "ble 11f\n" + "ld1w { z11.s }, p2/Z, [x12]\n" + "fmla z24.s, z11.s, z0.s[3]\n" + "addvl x12, x12, #1\n" + "11:" // Height 1: Multiply loop: multiply skip + "prfm pldl1keep, [x27, #0x80]\n" + "add x9, x9, #0x1\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x16, x19\n" - "bne 7b\n" - "tbz %x[flags], #1, 13f\n" + "cmp x9, x19\n" + "bne 6b\n" + "tbz %x[flags], #1, 12f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rw { z17.s }, p2/Z, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" "ld1rw { z16.s }, p2/Z, [x19]\n" "fmin z24.s, p2/M, z24.s, z16.s\n" "fmax z24.s, p2/M, z24.s, z17.s\n" - "13:" // Height 1: No activation - "st1w { z24.s }, p1, [x17]\n" - "addvl x17, x17, #1\n" - "14:" // Height 1: Writeback done - "decw x6\n" - "cmp x6, XZR\n" - "bgt 3b\n" - "b 114f\n" - "15:" // Height 2 - "ldr x6, [%x[args_ptr], %[offsetof_N]]\n" - "mov x8, %x[bias]\n" - "ldr x7, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 16f\n" - "ldr x17, [%x[output_ptr], #0x0]\n" - "add x17, x17, x19, LSL #2\n" - "ldr x13, [%x[output_ptr], #0x8]\n" - "add x13, x13, x19, LSL #2\n" - "b 17f\n" - "16:" // Height 2: setup direct output - "mov x17, %x[output_ptr]\n" - "add x13, x17, x19, LSL #2\n" - "17:" // Height 2: Column loop + "12:" // Height 1: No activation + "st1w { z24.s }, p1, [x10]\n" + "addvl x10, x10, #1\n" + "13:" // Height 1: Writeback done + "decw x13\n" + "cmp x13, XZR\n" + "bgt 2b\n" + "b 106f\n" + "14:" // Height 2 + "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" + "mov x11, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x10, %x[output_ptr]\n" + "15:" // Height 2: Column loop "mov x19, #0x0\n" - "whilelt p1.s, x19, x6\n" - "cbz x8, 18f\n" - "ld1w { z24.s }, p2/Z, [x8]\n" + "whilelt p1.s, x19, x13\n" + "cbz x11, 16f\n" + "ld1w { z24.s }, p2/Z, [x11]\n" "mov z25.d, z24.d\n" - "addvl x8, x8, #1\n" - "b 20f\n" - "18:" // Height 2: no bias - "tbz %x[flags], #0, 19f\n" - "ld1w { z24.s }, p1/Z, [x17]\n" - "ld1w { z25.s }, p1/Z, [x13]\n" - "b 20f\n" - "19:" // Height 2: no accumulate + "addvl x11, x11, #1\n" + "b 18f\n" + "16:" // Height 2: no bias + "tbz %x[flags], #0, 17f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z24.s }, p1/Z, [x10]\n" + "add x26, x10, x19, LSL #2\n" + "ld1w { z25.s }, p1/Z, [x26]\n" + "b 18f\n" + "17:" // Height 2: no accumulate "mov z24.b, #0x0\n" "mov z25.b, #0x0\n" - "20:" // Height 2: setup done - "mov x16, #0x0\n" - "21:" // Height 2: String loop + "18:" // Height 2: setup done + "mov x9, #0x0\n" + "19:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w15, [x20, x16, LSL #0x2]\n" - "tbz %x[flags], #3, 22f\n" - "ldr x20, [%x[input_ptr], x16, LSL #0x3]\n" + "ldr w28, [x20, x9, LSL #0x2]\n" + "tbz %x[flags], #3, 20f\n" + "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x14, [x20, #0x0]\n" - "ldr x12, [x20, #0x8]\n" - "cbnz x16, 23f\n" + "ldr x27, [x20, #0x0]\n" + "ldr x26, [x20, #0x8]\n" + "cbnz x9, 21f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x14, x14, x19, LSL #2\n" - "add x12, x12, x19, LSL #2\n" - "b 23f\n" - "22:" // Height 2: setup direct input - "mov x14, %x[input_ptr]\n" - "add x12, x14, x19, LSL #2\n" - "23:" // Height 2: input setup done - "cmp x15, #0x4\n" - "ble 25f\n" - "24:" // Height 2: Multiply loop: Main loop head - "ld1w { z8.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "ld1w { z9.s }, p2/Z, [x7, #1, MUL VL]\n" - "sub x15, x15, #0x4\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" + "add x27, x27, x19, LSL #2\n" + "add x26, x26, x19, LSL #2\n" + "b 21f\n" + "20:" // Height 2: setup direct input + "mov x27, %x[input_ptr]\n" + "add x26, x27, x19, LSL #2\n" + "21:" // Height 2: input setup done + "cmp x28, #0x4\n" + "ble 23f\n" + "22:" // Height 2: Multiply loop: Main loop head + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n" + "sub x28, x28, #0x4\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" "fmla z24.s, z8.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" "fmla z25.s, z8.s, z1.s[0]\n" - "ld1w { z10.s }, p2/Z, [x7, #2, MUL VL]\n" - "add x12, x12, #0x10\n" + "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" + "add x26, x26, #0x10\n" "fmla z24.s, z9.s, z0.s[1]\n" - "ld1w { z11.s }, p2/Z, [x7, #3, MUL VL]\n" - "cmp x15, #0x4\n" + "ld1w { z11.s }, p2/Z, [x12, #3, MUL VL]\n" + "cmp x28, #0x4\n" "fmla z25.s, z9.s, z1.s[1]\n" - "prfm pldl1keep, [x14, #0x80]\n" - "addvl x7, x7, #4\n" + "prfm pldl1keep, [x27, #0x80]\n" + "addvl x12, x12, #4\n" "fmla z24.s, z10.s, z0.s[2]\n" - "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla z25.s, z10.s, z1.s[2]\n" "fmla z24.s, z11.s, z0.s[3]\n" "fmla z25.s, z11.s, z1.s[3]\n" - "bgt 24b\n" - "25:" // Height 2: Multiply loop: Single iteration only - "ld1w { z12.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "subs x15, x15, #0x1\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" - "fmla z24.s, z12.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" - "fmla z25.s, z12.s, z1.s[0]\n" - "add x12, x12, #0x10\n" - "addvl x7, x7, #1\n" - "ble 26f\n" - "ld1w { z13.s }, p2/Z, [x7]\n" - "fmla z24.s, z13.s, z0.s[1]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z13.s, z1.s[1]\n" - "addvl x7, x7, #1\n" - "ble 26f\n" - "ld1w { z14.s }, p2/Z, [x7]\n" - "fmla z24.s, z14.s, z0.s[2]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z14.s, z1.s[2]\n" - "addvl x7, x7, #1\n" - "ble 26f\n" - "ld1w { z15.s }, p2/Z, [x7]\n" - "fmla z24.s, z15.s, z0.s[3]\n" - "addvl x7, x7, #1\n" - "fmla z25.s, z15.s, z1.s[3]\n" - "26:" // Height 2: Multiply loop: multiply skip - "prfm pldl1keep, [x14, #0x80]\n" - "add x16, x16, #0x1\n" - "prfm pldl1keep, [x12, #0x80]\n" + "bgt 22b\n" + "23:" // Height 2: Multiply loop: Single iteration only + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "subs x28, x28, #0x1\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" + "fmla z24.s, z8.s, z0.s[0]\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" + "fmla z25.s, z8.s, z1.s[0]\n" + "add x26, x26, #0x10\n" + "addvl x12, x12, #1\n" + "ble 24f\n" + "ld1w { z9.s }, p2/Z, [x12]\n" + "fmla z24.s, z9.s, z0.s[1]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z9.s, z1.s[1]\n" + "addvl x12, x12, #1\n" + "ble 24f\n" + "ld1w { z10.s }, p2/Z, [x12]\n" + "fmla z24.s, z10.s, z0.s[2]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z10.s, z1.s[2]\n" + "addvl x12, x12, #1\n" + "ble 24f\n" + "ld1w { z11.s }, p2/Z, [x12]\n" + "fmla z24.s, z11.s, z0.s[3]\n" + "addvl x12, x12, #1\n" + "fmla z25.s, z11.s, z1.s[3]\n" + "24:" // Height 2: Multiply loop: multiply skip + "prfm pldl1keep, [x27, #0x80]\n" + "add x9, x9, #0x1\n" + "prfm pldl1keep, [x26, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x16, x19\n" - "bne 21b\n" - "tbz %x[flags], #1, 27f\n" + "cmp x9, x19\n" + "bne 19b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x26, x10, x19, LSL #2\n" + "tbz %x[flags], #1, 25f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rw { z17.s }, p2/Z, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -326,151 +316,144 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmin z25.s, p2/M, z25.s, z16.s\n" "fmax z24.s, p2/M, z24.s, z17.s\n" "fmax z25.s, p2/M, z25.s, z17.s\n" - "27:" // Height 2: No activation - "st1w { z24.s }, p1, [x17]\n" - "addvl x17, x17, #1\n" - "st1w { z25.s }, p1, [x13]\n" - "addvl x13, x13, #1\n" - "28:" // Height 2: Writeback done - "decw x6\n" - "cmp x6, XZR\n" - "bgt 17b\n" - "b 114f\n" - "29:" // Height 3 - "ldr x6, [%x[args_ptr], %[offsetof_N]]\n" - "mov x8, %x[bias]\n" - "ldr x7, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 30f\n" - "ldr x17, [%x[output_ptr], #0x0]\n" - "add x17, x17, x19, LSL #2\n" - "ldr x13, [%x[output_ptr], #0x8]\n" - "ldr x11, [%x[output_ptr], #0x10]\n" - "add x13, x13, x19, LSL #2\n" - "add x11, x11, x19, LSL #2\n" - "b 31f\n" - "30:" // Height 3: setup direct output - "mov x17, %x[output_ptr]\n" - "add x13, x17, x19, LSL #2\n" - "add x11, x13, x19, LSL #2\n" - "31:" // Height 3: Column loop + "25:" // Height 2: No activation + "st1w { z24.s }, p1, [x10]\n" + "addvl x10, x10, #1\n" + "st1w { z25.s }, p1, [x26]\n" + "26:" // Height 2: Writeback done + "decw x13\n" + "cmp x13, XZR\n" + "bgt 15b\n" + "b 106f\n" + "27:" // Height 3 + "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" + "mov x11, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x10, %x[output_ptr]\n" + "28:" // Height 3: Column loop "mov x19, #0x0\n" - "whilelt p1.s, x19, x6\n" - "cbz x8, 32f\n" - "ld1w { z24.s }, p2/Z, [x8]\n" + "whilelt p1.s, x19, x13\n" + "cbz x11, 29f\n" + "ld1w { z24.s }, p2/Z, [x11]\n" "mov z25.d, z24.d\n" - "addvl x8, x8, #1\n" + "addvl x11, x11, #1\n" "mov z26.d, z24.d\n" - "b 34f\n" - "32:" // Height 3: no bias - "tbz %x[flags], #0, 33f\n" - "ld1w { z24.s }, p1/Z, [x17]\n" - "ld1w { z25.s }, p1/Z, [x13]\n" - "ld1w { z26.s }, p1/Z, [x11]\n" - "b 34f\n" - "33:" // Height 3: no accumulate + "b 31f\n" + "29:" // Height 3: no bias + "tbz %x[flags], #0, 30f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z24.s }, p1/Z, [x10]\n" + "add x26, x10, x19, LSL #2\n" + "ld1w { z25.s }, p1/Z, [x26]\n" + "add x25, x26, x19, LSL #2\n" + "ld1w { z26.s }, p1/Z, [x25]\n" + "b 31f\n" + "30:" // Height 3: no accumulate "mov z24.b, #0x0\n" "mov z25.b, #0x0\n" "mov z26.b, #0x0\n" - "34:" // Height 3: setup done - "mov x16, #0x0\n" - "35:" // Height 3: String loop + "31:" // Height 3: setup done + "mov x9, #0x0\n" + "32:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w15, [x20, x16, LSL #0x2]\n" - "tbz %x[flags], #3, 36f\n" - "ldr x20, [%x[input_ptr], x16, LSL #0x3]\n" + "ldr w28, [x20, x9, LSL #0x2]\n" + "tbz %x[flags], #3, 33f\n" + "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x14, [x20, #0x0]\n" - "ldr x12, [x20, #0x8]\n" - "ldr x10, [x20, #0x10]\n" - "cbnz x16, 37f\n" + "ldr x27, [x20, #0x0]\n" + "ldr x26, [x20, #0x8]\n" + "ldr x25, [x20, #0x10]\n" + "cbnz x9, 34f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x14, x14, x19, LSL #2\n" - "add x12, x12, x19, LSL #2\n" - "add x10, x10, x19, LSL #2\n" - "b 37f\n" - "36:" // Height 3: setup direct input - "mov x14, %x[input_ptr]\n" - "add x12, x14, x19, LSL #2\n" - "add x10, x12, x19, LSL #2\n" - "37:" // Height 3: input setup done - "cmp x15, #0x4\n" - "ble 39f\n" - "38:" // Height 3: Multiply loop: Main loop head - "ld1w { z8.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "ld1w { z9.s }, p2/Z, [x7, #1, MUL VL]\n" - "sub x15, x15, #0x4\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" + "add x27, x27, x19, LSL #2\n" + "add x26, x26, x19, LSL #2\n" + "add x25, x25, x19, LSL #2\n" + "b 34f\n" + "33:" // Height 3: setup direct input + "mov x27, %x[input_ptr]\n" + "add x26, x27, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "34:" // Height 3: input setup done + "cmp x28, #0x4\n" + "ble 36f\n" + "35:" // Height 3: Multiply loop: Main loop head + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n" + "sub x28, x28, #0x4\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" "fmla z24.s, z8.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" "fmla z25.s, z8.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" + "add x26, x26, #0x10\n" "fmla z24.s, z9.s, z0.s[1]\n" - "ld1w { z10.s }, p2/Z, [x7, #2, MUL VL]\n" - "add x10, x10, #0x10\n" + "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" + "add x25, x25, #0x10\n" "fmla z26.s, z8.s, z2.s[0]\n" - "ld1w { z11.s }, p2/Z, [x7, #3, MUL VL]\n" - "cmp x15, #0x4\n" + "ld1w { z11.s }, p2/Z, [x12, #3, MUL VL]\n" + "cmp x28, #0x4\n" "fmla z25.s, z9.s, z1.s[1]\n" - "prfm pldl1keep, [x14, #0x80]\n" - "addvl x7, x7, #4\n" + "prfm pldl1keep, [x27, #0x80]\n" + "addvl x12, x12, #4\n" "fmla z24.s, z10.s, z0.s[2]\n" - "prfm pldl1keep, [x12, #0x80]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla z26.s, z9.s, z2.s[1]\n" "fmla z25.s, z10.s, z1.s[2]\n" "fmla z24.s, z11.s, z0.s[3]\n" "fmla z26.s, z10.s, z2.s[2]\n" "fmla z25.s, z11.s, z1.s[3]\n" "fmla z26.s, z11.s, z2.s[3]\n" - "bgt 38b\n" - "39:" // Height 3: Multiply loop: Single iteration only - "ld1w { z12.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "subs x15, x15, #0x1\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" - "fmla z24.s, z12.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" - "fmla z25.s, z12.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" - "fmla z26.s, z12.s, z2.s[0]\n" - "add x10, x10, #0x10\n" - "addvl x7, x7, #1\n" - "ble 40f\n" - "ld1w { z13.s }, p2/Z, [x7]\n" - "fmla z24.s, z13.s, z0.s[1]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z13.s, z1.s[1]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z13.s, z2.s[1]\n" - "ble 40f\n" - "ld1w { z14.s }, p2/Z, [x7]\n" - "fmla z24.s, z14.s, z0.s[2]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z14.s, z1.s[2]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z14.s, z2.s[2]\n" - "ble 40f\n" - "ld1w { z15.s }, p2/Z, [x7]\n" - "fmla z24.s, z15.s, z0.s[3]\n" - "addvl x7, x7, #1\n" - "fmla z25.s, z15.s, z1.s[3]\n" - "fmla z26.s, z15.s, z2.s[3]\n" - "40:" // Height 3: Multiply loop: multiply skip - "prfm pldl1keep, [x14, #0x80]\n" - "add x16, x16, #0x1\n" - "prfm pldl1keep, [x12, #0x80]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "bgt 35b\n" + "36:" // Height 3: Multiply loop: Single iteration only + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "subs x28, x28, #0x1\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" + "fmla z24.s, z8.s, z0.s[0]\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" + "fmla z25.s, z8.s, z1.s[0]\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" + "add x26, x26, #0x10\n" + "fmla z26.s, z8.s, z2.s[0]\n" + "add x25, x25, #0x10\n" + "addvl x12, x12, #1\n" + "ble 37f\n" + "ld1w { z9.s }, p2/Z, [x12]\n" + "fmla z24.s, z9.s, z0.s[1]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z9.s, z1.s[1]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z9.s, z2.s[1]\n" + "ble 37f\n" + "ld1w { z10.s }, p2/Z, [x12]\n" + "fmla z24.s, z10.s, z0.s[2]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z10.s, z1.s[2]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z10.s, z2.s[2]\n" + "ble 37f\n" + "ld1w { z11.s }, p2/Z, [x12]\n" + "fmla z24.s, z11.s, z0.s[3]\n" + "addvl x12, x12, #1\n" + "fmla z25.s, z11.s, z1.s[3]\n" + "fmla z26.s, z11.s, z2.s[3]\n" + "37:" // Height 3: Multiply loop: multiply skip + "prfm pldl1keep, [x27, #0x80]\n" + "add x9, x9, #0x1\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x16, x19\n" - "bne 35b\n" - "tbz %x[flags], #1, 41f\n" + "cmp x9, x19\n" + "bne 32b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x26, x10, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "tbz %x[flags], #1, 38f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rw { z17.s }, p2/Z, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -481,117 +464,104 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmax z24.s, p2/M, z24.s, z17.s\n" "fmax z25.s, p2/M, z25.s, z17.s\n" "fmax z26.s, p2/M, z26.s, z17.s\n" - "41:" // Height 3: No activation - "st1w { z24.s }, p1, [x17]\n" - "addvl x17, x17, #1\n" - "st1w { z25.s }, p1, [x13]\n" - "addvl x13, x13, #1\n" - "st1w { z26.s }, p1, [x11]\n" - "addvl x11, x11, #1\n" - "42:" // Height 3: Writeback done - "decw x6\n" - "cmp x6, XZR\n" - "bgt 31b\n" - "b 114f\n" - "43:" // Height 4 - "ldr x6, [%x[args_ptr], %[offsetof_N]]\n" - "mov x8, %x[bias]\n" - "ldr x7, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 44f\n" - "ldr x17, [%x[output_ptr], #0x0]\n" - "add x17, x17, x19, LSL #2\n" - "ldr x13, [%x[output_ptr], #0x8]\n" - "ldr x11, [%x[output_ptr], #0x10]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x18]\n" - "add x11, x11, x19, LSL #2\n" - "add x9, x9, x19, LSL #2\n" - "b 45f\n" - "44:" // Height 4: setup direct output - "mov x17, %x[output_ptr]\n" - "add x13, x17, x19, LSL #2\n" - "add x11, x13, x19, LSL #2\n" - "add x9, x11, x19, LSL #2\n" - "45:" // Height 4: Column loop + "38:" // Height 3: No activation + "st1w { z24.s }, p1, [x10]\n" + "addvl x10, x10, #1\n" + "st1w { z25.s }, p1, [x26]\n" + "st1w { z26.s }, p1, [x25]\n" + "39:" // Height 3: Writeback done + "decw x13\n" + "cmp x13, XZR\n" + "bgt 28b\n" + "b 106f\n" + "40:" // Height 4 + "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" + "mov x11, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x10, %x[output_ptr]\n" + "41:" // Height 4: Column loop "mov x19, #0x0\n" - "whilelt p1.s, x19, x6\n" - "cbz x8, 46f\n" - "ld1w { z24.s }, p2/Z, [x8]\n" + "whilelt p1.s, x19, x13\n" + "cbz x11, 42f\n" + "ld1w { z24.s }, p2/Z, [x11]\n" "mov z25.d, z24.d\n" - "addvl x8, x8, #1\n" + "addvl x11, x11, #1\n" "mov z26.d, z24.d\n" "mov z27.d, z24.d\n" - "b 48f\n" - "46:" // Height 4: no bias - "tbz %x[flags], #0, 47f\n" - "ld1w { z24.s }, p1/Z, [x17]\n" - "ld1w { z25.s }, p1/Z, [x13]\n" - "ld1w { z26.s }, p1/Z, [x11]\n" - "ld1w { z27.s }, p1/Z, [x9]\n" - "b 48f\n" - "47:" // Height 4: no accumulate + "b 44f\n" + "42:" // Height 4: no bias + "tbz %x[flags], #0, 43f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z24.s }, p1/Z, [x10]\n" + "add x26, x10, x19, LSL #2\n" + "ld1w { z25.s }, p1/Z, [x26]\n" + "add x25, x26, x19, LSL #2\n" + "ld1w { z26.s }, p1/Z, [x25]\n" + "add x24, x25, x19, LSL #2\n" + "ld1w { z27.s }, p1/Z, [x24]\n" + "b 44f\n" + "43:" // Height 4: no accumulate "mov z24.b, #0x0\n" "mov z25.b, #0x0\n" "mov z26.b, #0x0\n" "mov z27.b, #0x0\n" - "48:" // Height 4: setup done - "mov x16, #0x0\n" - "49:" // Height 4: String loop + "44:" // Height 4: setup done + "mov x9, #0x0\n" + "45:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w15, [x20, x16, LSL #0x2]\n" - "tbz %x[flags], #3, 50f\n" - "ldr x20, [%x[input_ptr], x16, LSL #0x3]\n" + "ldr w28, [x20, x9, LSL #0x2]\n" + "tbz %x[flags], #3, 46f\n" + "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x14, [x20, #0x0]\n" - "ldr x12, [x20, #0x8]\n" - "ldr x10, [x20, #0x10]\n" - "ldr x28, [x20, #0x18]\n" - "cbnz x16, 51f\n" + "ldr x27, [x20, #0x0]\n" + "ldr x26, [x20, #0x8]\n" + "ldr x25, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "cbnz x9, 47f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x14, x14, x19, LSL #2\n" - "add x12, x12, x19, LSL #2\n" - "add x10, x10, x19, LSL #2\n" - "add x28, x28, x19, LSL #2\n" - "b 51f\n" - "50:" // Height 4: setup direct input - "mov x14, %x[input_ptr]\n" - "add x12, x14, x19, LSL #2\n" - "add x10, x12, x19, LSL #2\n" - "add x28, x10, x19, LSL #2\n" - "51:" // Height 4: input setup done - "cmp x15, #0x4\n" - "ble 53f\n" - "52:" // Height 4: Multiply loop: Main loop head - "ld1w { z8.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "ld1w { z9.s }, p2/Z, [x7, #1, MUL VL]\n" - "sub x15, x15, #0x4\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" + "add x27, x27, x19, LSL #2\n" + "add x26, x26, x19, LSL #2\n" + "add x25, x25, x19, LSL #2\n" + "add x24, x24, x19, LSL #2\n" + "b 47f\n" + "46:" // Height 4: setup direct input + "mov x27, %x[input_ptr]\n" + "add x26, x27, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "47:" // Height 4: input setup done + "cmp x28, #0x4\n" + "ble 49f\n" + "48:" // Height 4: Multiply loop: Main loop head + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n" + "sub x28, x28, #0x4\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" "fmla z24.s, z8.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" "fmla z25.s, z8.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" + "add x26, x26, #0x10\n" "fmla z24.s, z9.s, z0.s[1]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "fmla z26.s, z8.s, z2.s[0]\n" - "ld1w { z10.s }, p2/Z, [x7, #2, MUL VL]\n" - "add x28, x28, #0x10\n" + "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" + "add x24, x24, #0x10\n" "fmla z27.s, z8.s, z3.s[0]\n" - "ld1w { z11.s }, p2/Z, [x7, #3, MUL VL]\n" - "cmp x15, #0x4\n" + "ld1w { z11.s }, p2/Z, [x12, #3, MUL VL]\n" + "cmp x28, #0x4\n" "fmla z25.s, z9.s, z1.s[1]\n" - "prfm pldl1keep, [x14, #0x80]\n" - "addvl x7, x7, #4\n" + "prfm pldl1keep, [x27, #0x80]\n" + "addvl x12, x12, #4\n" "fmla z24.s, z10.s, z0.s[2]\n" - "prfm pldl1keep, [x12, #0x80]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla z26.s, z9.s, z2.s[1]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla z27.s, z9.s, z3.s[1]\n" "fmla z25.s, z10.s, z1.s[2]\n" "fmla z24.s, z11.s, z0.s[3]\n" @@ -600,57 +570,61 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmla z25.s, z11.s, z1.s[3]\n" "fmla z26.s, z11.s, z2.s[3]\n" "fmla z27.s, z11.s, z3.s[3]\n" - "bgt 52b\n" - "53:" // Height 4: Multiply loop: Single iteration only - "ld1w { z12.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "subs x15, x15, #0x1\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" - "fmla z24.s, z12.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" - "fmla z25.s, z12.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" - "fmla z26.s, z12.s, z2.s[0]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" - "fmla z27.s, z12.s, z3.s[0]\n" - "add x28, x28, #0x10\n" - "addvl x7, x7, #1\n" - "ble 54f\n" - "ld1w { z13.s }, p2/Z, [x7]\n" - "fmla z24.s, z13.s, z0.s[1]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z13.s, z1.s[1]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z13.s, z2.s[1]\n" - "fmla z27.s, z13.s, z3.s[1]\n" - "ble 54f\n" - "ld1w { z14.s }, p2/Z, [x7]\n" - "fmla z24.s, z14.s, z0.s[2]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z14.s, z1.s[2]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z14.s, z2.s[2]\n" - "fmla z27.s, z14.s, z3.s[2]\n" - "ble 54f\n" - "ld1w { z15.s }, p2/Z, [x7]\n" - "fmla z24.s, z15.s, z0.s[3]\n" - "addvl x7, x7, #1\n" - "fmla z25.s, z15.s, z1.s[3]\n" - "fmla z26.s, z15.s, z2.s[3]\n" - "fmla z27.s, z15.s, z3.s[3]\n" - "54:" // Height 4: Multiply loop: multiply skip - "prfm pldl1keep, [x14, #0x80]\n" - "add x16, x16, #0x1\n" - "prfm pldl1keep, [x12, #0x80]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "bgt 48b\n" + "49:" // Height 4: Multiply loop: Single iteration only + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "subs x28, x28, #0x1\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" + "fmla z24.s, z8.s, z0.s[0]\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" + "fmla z25.s, z8.s, z1.s[0]\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" + "add x26, x26, #0x10\n" + "fmla z26.s, z8.s, z2.s[0]\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" + "fmla z27.s, z8.s, z3.s[0]\n" + "add x24, x24, #0x10\n" + "addvl x12, x12, #1\n" + "ble 50f\n" + "ld1w { z9.s }, p2/Z, [x12]\n" + "fmla z24.s, z9.s, z0.s[1]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z9.s, z1.s[1]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z9.s, z2.s[1]\n" + "fmla z27.s, z9.s, z3.s[1]\n" + "ble 50f\n" + "ld1w { z10.s }, p2/Z, [x12]\n" + "fmla z24.s, z10.s, z0.s[2]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z10.s, z1.s[2]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z10.s, z2.s[2]\n" + "fmla z27.s, z10.s, z3.s[2]\n" + "ble 50f\n" + "ld1w { z11.s }, p2/Z, [x12]\n" + "fmla z24.s, z11.s, z0.s[3]\n" + "addvl x12, x12, #1\n" + "fmla z25.s, z11.s, z1.s[3]\n" + "fmla z26.s, z11.s, z2.s[3]\n" + "fmla z27.s, z11.s, z3.s[3]\n" + "50:" // Height 4: Multiply loop: multiply skip + "prfm pldl1keep, [x27, #0x80]\n" + "add x9, x9, #0x1\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x16, x19\n" - "bne 49b\n" - "tbz %x[flags], #1, 55f\n" + "cmp x9, x19\n" + "bne 45b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x26, x10, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "tbz %x[flags], #1, 51f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rw { z17.s }, p2/Z, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -663,134 +637,118 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmax z25.s, p2/M, z25.s, z17.s\n" "fmax z26.s, p2/M, z26.s, z17.s\n" "fmax z27.s, p2/M, z27.s, z17.s\n" - "55:" // Height 4: No activation - "st1w { z24.s }, p1, [x17]\n" - "addvl x17, x17, #1\n" - "st1w { z25.s }, p1, [x13]\n" - "addvl x13, x13, #1\n" - "st1w { z26.s }, p1, [x11]\n" - "addvl x11, x11, #1\n" - "st1w { z27.s }, p1, [x9]\n" - "addvl x9, x9, #1\n" - "56:" // Height 4: Writeback done - "decw x6\n" - "cmp x6, XZR\n" - "bgt 45b\n" - "b 114f\n" - "57:" // Height 5 - "ldr x6, [%x[args_ptr], %[offsetof_N]]\n" - "mov x8, %x[bias]\n" - "ldr x7, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 58f\n" - "ldr x17, [%x[output_ptr], #0x0]\n" - "add x17, x17, x19, LSL #2\n" - "ldr x13, [%x[output_ptr], #0x8]\n" - "ldr x11, [%x[output_ptr], #0x10]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x18]\n" - "ldr x27, [%x[output_ptr], #0x20]\n" - "add x11, x11, x19, LSL #2\n" - "add x9, x9, x19, LSL #2\n" - "add x27, x27, x19, LSL #2\n" - "b 59f\n" - "58:" // Height 5: setup direct output - "mov x17, %x[output_ptr]\n" - "add x13, x17, x19, LSL #2\n" - "add x11, x13, x19, LSL #2\n" - "add x9, x11, x19, LSL #2\n" - "add x27, x9, x19, LSL #2\n" - "59:" // Height 5: Column loop + "51:" // Height 4: No activation + "st1w { z24.s }, p1, [x10]\n" + "addvl x10, x10, #1\n" + "st1w { z25.s }, p1, [x26]\n" + "st1w { z26.s }, p1, [x25]\n" + "st1w { z27.s }, p1, [x24]\n" + "52:" // Height 4: Writeback done + "decw x13\n" + "cmp x13, XZR\n" + "bgt 41b\n" + "b 106f\n" + "53:" // Height 5 + "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" + "mov x11, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x10, %x[output_ptr]\n" + "54:" // Height 5: Column loop "mov x19, #0x0\n" - "whilelt p1.s, x19, x6\n" - "cbz x8, 60f\n" - "ld1w { z24.s }, p2/Z, [x8]\n" + "whilelt p1.s, x19, x13\n" + "cbz x11, 55f\n" + "ld1w { z24.s }, p2/Z, [x11]\n" "mov z25.d, z24.d\n" - "addvl x8, x8, #1\n" + "addvl x11, x11, #1\n" "mov z26.d, z24.d\n" "mov z27.d, z24.d\n" "mov z28.d, z24.d\n" - "b 62f\n" - "60:" // Height 5: no bias - "tbz %x[flags], #0, 61f\n" - "ld1w { z24.s }, p1/Z, [x17]\n" - "ld1w { z25.s }, p1/Z, [x13]\n" - "ld1w { z26.s }, p1/Z, [x11]\n" - "ld1w { z27.s }, p1/Z, [x9]\n" - "ld1w { z28.s }, p1/Z, [x27]\n" - "b 62f\n" - "61:" // Height 5: no accumulate + "b 57f\n" + "55:" // Height 5: no bias + "tbz %x[flags], #0, 56f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z24.s }, p1/Z, [x10]\n" + "add x26, x10, x19, LSL #2\n" + "ld1w { z25.s }, p1/Z, [x26]\n" + "add x25, x26, x19, LSL #2\n" + "ld1w { z26.s }, p1/Z, [x25]\n" + "add x24, x25, x19, LSL #2\n" + "ld1w { z27.s }, p1/Z, [x24]\n" + "add x23, x24, x19, LSL #2\n" + "ld1w { z28.s }, p1/Z, [x23]\n" + "b 57f\n" + "56:" // Height 5: no accumulate "mov z24.b, #0x0\n" "mov z25.b, #0x0\n" "mov z26.b, #0x0\n" "mov z27.b, #0x0\n" "mov z28.b, #0x0\n" - "62:" // Height 5: setup done - "mov x16, #0x0\n" - "63:" // Height 5: String loop + "57:" // Height 5: setup done + "mov x9, #0x0\n" + "58:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w15, [x20, x16, LSL #0x2]\n" - "tbz %x[flags], #3, 64f\n" - "ldr x20, [%x[input_ptr], x16, LSL #0x3]\n" + "ldr w28, [x20, x9, LSL #0x2]\n" + "tbz %x[flags], #3, 59f\n" + "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x14, [x20, #0x0]\n" - "ldr x12, [x20, #0x8]\n" - "ldr x10, [x20, #0x10]\n" - "ldr x28, [x20, #0x18]\n" - "ldr x26, [x20, #0x20]\n" - "cbnz x16, 65f\n" + "ldr x27, [x20, #0x0]\n" + "ldr x26, [x20, #0x8]\n" + "ldr x25, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "ldr x23, [x20, #0x20]\n" + "cbnz x9, 60f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x14, x14, x19, LSL #2\n" - "add x12, x12, x19, LSL #2\n" - "add x10, x10, x19, LSL #2\n" - "add x28, x28, x19, LSL #2\n" + "add x27, x27, x19, LSL #2\n" "add x26, x26, x19, LSL #2\n" - "b 65f\n" - "64:" // Height 5: setup direct input - "mov x14, %x[input_ptr]\n" - "add x12, x14, x19, LSL #2\n" - "add x10, x12, x19, LSL #2\n" - "add x28, x10, x19, LSL #2\n" - "add x26, x28, x19, LSL #2\n" - "65:" // Height 5: input setup done - "cmp x15, #0x4\n" - "ble 67f\n" - "66:" // Height 5: Multiply loop: Main loop head - "ld1w { z8.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "ld1w { z9.s }, p2/Z, [x7, #1, MUL VL]\n" - "sub x15, x15, #0x4\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" + "add x25, x25, x19, LSL #2\n" + "add x24, x24, x19, LSL #2\n" + "add x23, x23, x19, LSL #2\n" + "b 60f\n" + "59:" // Height 5: setup direct input + "mov x27, %x[input_ptr]\n" + "add x26, x27, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "60:" // Height 5: input setup done + "cmp x28, #0x4\n" + "ble 62f\n" + "61:" // Height 5: Multiply loop: Main loop head + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n" + "sub x28, x28, #0x4\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" "fmla z24.s, z8.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" "fmla z25.s, z8.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" + "add x26, x26, #0x10\n" "fmla z24.s, z9.s, z0.s[1]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "fmla z26.s, z8.s, z2.s[0]\n" - "ld1rqw { z4.s }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "fmla z27.s, z8.s, z3.s[0]\n" - "ld1w { z10.s }, p2/Z, [x7, #2, MUL VL]\n" - "add x26, x26, #0x10\n" + "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" + "add x23, x23, #0x10\n" "fmla z25.s, z9.s, z1.s[1]\n" - "ld1w { z11.s }, p2/Z, [x7, #3, MUL VL]\n" - "cmp x15, #0x4\n" + "ld1w { z11.s }, p2/Z, [x12, #3, MUL VL]\n" + "cmp x28, #0x4\n" "fmla z28.s, z8.s, z4.s[0]\n" - "prfm pldl1keep, [x14, #0x80]\n" - "addvl x7, x7, #4\n" + "prfm pldl1keep, [x27, #0x80]\n" + "addvl x12, x12, #4\n" "fmla z26.s, z9.s, z2.s[1]\n" - "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla z24.s, z10.s, z0.s[2]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla z27.s, z9.s, z3.s[1]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla z25.s, z10.s, z1.s[2]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "fmla z28.s, z9.s, z4.s[1]\n" "fmla z26.s, z10.s, z2.s[2]\n" "fmla z27.s, z10.s, z3.s[2]\n" @@ -800,64 +758,69 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmla z26.s, z11.s, z2.s[3]\n" "fmla z27.s, z11.s, z3.s[3]\n" "fmla z28.s, z11.s, z4.s[3]\n" - "bgt 66b\n" - "67:" // Height 5: Multiply loop: Single iteration only - "ld1w { z12.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "subs x15, x15, #0x1\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" - "fmla z24.s, z12.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" - "fmla z25.s, z12.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" - "fmla z26.s, z12.s, z2.s[0]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" - "fmla z27.s, z12.s, z3.s[0]\n" - "ld1rqw { z4.s }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" - "fmla z28.s, z12.s, z4.s[0]\n" + "bgt 61b\n" + "62:" // Height 5: Multiply loop: Single iteration only + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "subs x28, x28, #0x1\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" + "fmla z24.s, z8.s, z0.s[0]\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" + "fmla z25.s, z8.s, z1.s[0]\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" "add x26, x26, #0x10\n" - "addvl x7, x7, #1\n" - "ble 68f\n" - "ld1w { z13.s }, p2/Z, [x7]\n" - "fmla z24.s, z13.s, z0.s[1]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z13.s, z1.s[1]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z13.s, z2.s[1]\n" - "fmla z27.s, z13.s, z3.s[1]\n" - "fmla z28.s, z13.s, z4.s[1]\n" - "ble 68f\n" - "ld1w { z14.s }, p2/Z, [x7]\n" - "fmla z24.s, z14.s, z0.s[2]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z14.s, z1.s[2]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z14.s, z2.s[2]\n" - "fmla z27.s, z14.s, z3.s[2]\n" - "fmla z28.s, z14.s, z4.s[2]\n" - "ble 68f\n" - "ld1w { z15.s }, p2/Z, [x7]\n" - "fmla z24.s, z15.s, z0.s[3]\n" - "addvl x7, x7, #1\n" - "fmla z25.s, z15.s, z1.s[3]\n" - "fmla z26.s, z15.s, z2.s[3]\n" - "fmla z27.s, z15.s, z3.s[3]\n" - "fmla z28.s, z15.s, z4.s[3]\n" - "68:" // Height 5: Multiply loop: multiply skip - "prfm pldl1keep, [x14, #0x80]\n" - "add x16, x16, #0x1\n" - "prfm pldl1keep, [x12, #0x80]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "fmla z26.s, z8.s, z2.s[0]\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" + "fmla z27.s, z8.s, z3.s[0]\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" + "fmla z28.s, z8.s, z4.s[0]\n" + "add x23, x23, #0x10\n" + "addvl x12, x12, #1\n" + "ble 63f\n" + "ld1w { z9.s }, p2/Z, [x12]\n" + "fmla z24.s, z9.s, z0.s[1]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z9.s, z1.s[1]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z9.s, z2.s[1]\n" + "fmla z27.s, z9.s, z3.s[1]\n" + "fmla z28.s, z9.s, z4.s[1]\n" + "ble 63f\n" + "ld1w { z10.s }, p2/Z, [x12]\n" + "fmla z24.s, z10.s, z0.s[2]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z10.s, z1.s[2]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z10.s, z2.s[2]\n" + "fmla z27.s, z10.s, z3.s[2]\n" + "fmla z28.s, z10.s, z4.s[2]\n" + "ble 63f\n" + "ld1w { z11.s }, p2/Z, [x12]\n" + "fmla z24.s, z11.s, z0.s[3]\n" + "addvl x12, x12, #1\n" + "fmla z25.s, z11.s, z1.s[3]\n" + "fmla z26.s, z11.s, z2.s[3]\n" + "fmla z27.s, z11.s, z3.s[3]\n" + "fmla z28.s, z11.s, z4.s[3]\n" + "63:" // Height 5: Multiply loop: multiply skip + "prfm pldl1keep, [x27, #0x80]\n" + "add x9, x9, #0x1\n" "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x16, x19\n" - "bne 63b\n" - "tbz %x[flags], #1, 69f\n" + "cmp x9, x19\n" + "bne 58b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x26, x10, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "tbz %x[flags], #1, 64f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rw { z17.s }, p2/Z, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -872,150 +835,131 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmax z26.s, p2/M, z26.s, z17.s\n" "fmax z27.s, p2/M, z27.s, z17.s\n" "fmax z28.s, p2/M, z28.s, z17.s\n" - "69:" // Height 5: No activation - "st1w { z24.s }, p1, [x17]\n" - "addvl x17, x17, #1\n" - "st1w { z25.s }, p1, [x13]\n" - "addvl x13, x13, #1\n" - "st1w { z26.s }, p1, [x11]\n" - "addvl x11, x11, #1\n" - "st1w { z27.s }, p1, [x9]\n" - "addvl x9, x9, #1\n" - "st1w { z28.s }, p1, [x27]\n" - "addvl x27, x27, #1\n" - "70:" // Height 5: Writeback done - "decw x6\n" - "cmp x6, XZR\n" - "bgt 59b\n" - "b 114f\n" - "71:" // Height 6 - "ldr x6, [%x[args_ptr], %[offsetof_N]]\n" - "mov x8, %x[bias]\n" - "ldr x7, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 72f\n" - "ldr x17, [%x[output_ptr], #0x0]\n" - "add x17, x17, x19, LSL #2\n" - "ldr x13, [%x[output_ptr], #0x8]\n" - "ldr x11, [%x[output_ptr], #0x10]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x18]\n" - "ldr x27, [%x[output_ptr], #0x20]\n" - "add x11, x11, x19, LSL #2\n" - "ldr x25, [%x[output_ptr], #0x28]\n" - "add x9, x9, x19, LSL #2\n" - "add x27, x27, x19, LSL #2\n" - "add x25, x25, x19, LSL #2\n" - "b 73f\n" - "72:" // Height 6: setup direct output - "mov x17, %x[output_ptr]\n" - "add x13, x17, x19, LSL #2\n" - "add x11, x13, x19, LSL #2\n" - "add x9, x11, x19, LSL #2\n" - "add x27, x9, x19, LSL #2\n" - "add x25, x27, x19, LSL #2\n" - "73:" // Height 6: Column loop + "64:" // Height 5: No activation + "st1w { z24.s }, p1, [x10]\n" + "addvl x10, x10, #1\n" + "st1w { z25.s }, p1, [x26]\n" + "st1w { z26.s }, p1, [x25]\n" + "st1w { z27.s }, p1, [x24]\n" + "st1w { z28.s }, p1, [x23]\n" + "65:" // Height 5: Writeback done + "decw x13\n" + "cmp x13, XZR\n" + "bgt 54b\n" + "b 106f\n" + "66:" // Height 6 + "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" + "mov x11, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x10, %x[output_ptr]\n" + "67:" // Height 6: Column loop "mov x19, #0x0\n" - "whilelt p1.s, x19, x6\n" - "cbz x8, 74f\n" - "ld1w { z24.s }, p2/Z, [x8]\n" + "whilelt p1.s, x19, x13\n" + "cbz x11, 68f\n" + "ld1w { z24.s }, p2/Z, [x11]\n" "mov z25.d, z24.d\n" - "addvl x8, x8, #1\n" + "addvl x11, x11, #1\n" "mov z26.d, z24.d\n" "mov z27.d, z24.d\n" "mov z28.d, z24.d\n" "mov z29.d, z24.d\n" - "b 76f\n" - "74:" // Height 6: no bias - "tbz %x[flags], #0, 75f\n" - "ld1w { z24.s }, p1/Z, [x17]\n" - "ld1w { z25.s }, p1/Z, [x13]\n" - "ld1w { z26.s }, p1/Z, [x11]\n" - "ld1w { z27.s }, p1/Z, [x9]\n" - "ld1w { z28.s }, p1/Z, [x27]\n" - "ld1w { z29.s }, p1/Z, [x25]\n" - "b 76f\n" - "75:" // Height 6: no accumulate + "b 70f\n" + "68:" // Height 6: no bias + "tbz %x[flags], #0, 69f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z24.s }, p1/Z, [x10]\n" + "add x26, x10, x19, LSL #2\n" + "ld1w { z25.s }, p1/Z, [x26]\n" + "add x25, x26, x19, LSL #2\n" + "ld1w { z26.s }, p1/Z, [x25]\n" + "add x24, x25, x19, LSL #2\n" + "ld1w { z27.s }, p1/Z, [x24]\n" + "add x23, x24, x19, LSL #2\n" + "ld1w { z28.s }, p1/Z, [x23]\n" + "add x22, x23, x19, LSL #2\n" + "ld1w { z29.s }, p1/Z, [x22]\n" + "b 70f\n" + "69:" // Height 6: no accumulate "mov z24.b, #0x0\n" "mov z25.b, #0x0\n" "mov z26.b, #0x0\n" "mov z27.b, #0x0\n" "mov z28.b, #0x0\n" "mov z29.b, #0x0\n" - "76:" // Height 6: setup done - "mov x16, #0x0\n" - "77:" // Height 6: String loop + "70:" // Height 6: setup done + "mov x9, #0x0\n" + "71:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w15, [x20, x16, LSL #0x2]\n" - "tbz %x[flags], #3, 78f\n" - "ldr x20, [%x[input_ptr], x16, LSL #0x3]\n" + "ldr w28, [x20, x9, LSL #0x2]\n" + "tbz %x[flags], #3, 72f\n" + "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x14, [x20, #0x0]\n" - "ldr x12, [x20, #0x8]\n" - "ldr x10, [x20, #0x10]\n" - "ldr x28, [x20, #0x18]\n" - "ldr x26, [x20, #0x20]\n" - "ldr x24, [x20, #0x28]\n" - "cbnz x16, 79f\n" + "ldr x27, [x20, #0x0]\n" + "ldr x26, [x20, #0x8]\n" + "ldr x25, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "ldr x23, [x20, #0x20]\n" + "ldr x22, [x20, #0x28]\n" + "cbnz x9, 73f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x14, x14, x19, LSL #2\n" - "add x12, x12, x19, LSL #2\n" - "add x10, x10, x19, LSL #2\n" - "add x28, x28, x19, LSL #2\n" + "add x27, x27, x19, LSL #2\n" "add x26, x26, x19, LSL #2\n" + "add x25, x25, x19, LSL #2\n" "add x24, x24, x19, LSL #2\n" - "b 79f\n" - "78:" // Height 6: setup direct input - "mov x14, %x[input_ptr]\n" - "add x12, x14, x19, LSL #2\n" - "add x10, x12, x19, LSL #2\n" - "add x28, x10, x19, LSL #2\n" - "add x26, x28, x19, LSL #2\n" - "add x24, x26, x19, LSL #2\n" - "79:" // Height 6: input setup done - "cmp x15, #0x4\n" - "ble 81f\n" - "80:" // Height 6: Multiply loop: Main loop head - "ld1w { z8.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "ld1w { z9.s }, p2/Z, [x7, #1, MUL VL]\n" - "sub x15, x15, #0x4\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" + "add x23, x23, x19, LSL #2\n" + "add x22, x22, x19, LSL #2\n" + "b 73f\n" + "72:" // Height 6: setup direct input + "mov x27, %x[input_ptr]\n" + "add x26, x27, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "add x22, x23, x19, LSL #2\n" + "73:" // Height 6: input setup done + "cmp x28, #0x4\n" + "ble 75f\n" + "74:" // Height 6: Multiply loop: Main loop head + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n" + "sub x28, x28, #0x4\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" "fmla z24.s, z8.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" "fmla z25.s, z8.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" + "add x26, x26, #0x10\n" "fmla z24.s, z9.s, z0.s[1]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "fmla z26.s, z8.s, z2.s[0]\n" - "ld1rqw { z4.s }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "fmla z27.s, z8.s, z3.s[0]\n" - "ld1rqw { z5.s }, p0/Z, [x24]\n" - "add x26, x26, #0x10\n" + "ld1rqw { z5.s }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" "fmla z25.s, z9.s, z1.s[1]\n" - "ld1w { z10.s }, p2/Z, [x7, #2, MUL VL]\n" - "add x24, x24, #0x10\n" + "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" + "add x22, x22, #0x10\n" "fmla z28.s, z8.s, z4.s[0]\n" - "ld1w { z11.s }, p2/Z, [x7, #3, MUL VL]\n" - "cmp x15, #0x4\n" + "ld1w { z11.s }, p2/Z, [x12, #3, MUL VL]\n" + "cmp x28, #0x4\n" "fmla z29.s, z8.s, z5.s[0]\n" - "prfm pldl1keep, [x14, #0x80]\n" - "addvl x7, x7, #4\n" + "prfm pldl1keep, [x27, #0x80]\n" + "addvl x12, x12, #4\n" "fmla z26.s, z9.s, z2.s[1]\n" - "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla z27.s, z9.s, z3.s[1]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla z24.s, z10.s, z0.s[2]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla z28.s, z9.s, z4.s[1]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "fmla z29.s, z9.s, z5.s[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" "fmla z25.s, z10.s, z1.s[2]\n" "fmla z26.s, z10.s, z2.s[2]\n" "fmla z27.s, z10.s, z3.s[2]\n" @@ -1027,71 +971,77 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmla z27.s, z11.s, z3.s[3]\n" "fmla z28.s, z11.s, z4.s[3]\n" "fmla z29.s, z11.s, z5.s[3]\n" - "bgt 80b\n" - "81:" // Height 6: Multiply loop: Single iteration only - "ld1w { z12.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "subs x15, x15, #0x1\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" - "fmla z24.s, z12.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" - "fmla z25.s, z12.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" - "fmla z26.s, z12.s, z2.s[0]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" - "fmla z27.s, z12.s, z3.s[0]\n" - "ld1rqw { z4.s }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" - "fmla z28.s, z12.s, z4.s[0]\n" - "ld1rqw { z5.s }, p0/Z, [x24]\n" + "bgt 74b\n" + "75:" // Height 6: Multiply loop: Single iteration only + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "subs x28, x28, #0x1\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" + "fmla z24.s, z8.s, z0.s[0]\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" + "fmla z25.s, z8.s, z1.s[0]\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" "add x26, x26, #0x10\n" - "fmla z29.s, z12.s, z5.s[0]\n" + "fmla z26.s, z8.s, z2.s[0]\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" + "fmla z27.s, z8.s, z3.s[0]\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" "add x24, x24, #0x10\n" - "addvl x7, x7, #1\n" - "ble 82f\n" - "ld1w { z13.s }, p2/Z, [x7]\n" - "fmla z24.s, z13.s, z0.s[1]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z13.s, z1.s[1]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z13.s, z2.s[1]\n" - "fmla z27.s, z13.s, z3.s[1]\n" - "fmla z28.s, z13.s, z4.s[1]\n" - "fmla z29.s, z13.s, z5.s[1]\n" - "ble 82f\n" - "ld1w { z14.s }, p2/Z, [x7]\n" - "fmla z24.s, z14.s, z0.s[2]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z14.s, z1.s[2]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z14.s, z2.s[2]\n" - "fmla z27.s, z14.s, z3.s[2]\n" - "fmla z28.s, z14.s, z4.s[2]\n" - "fmla z29.s, z14.s, z5.s[2]\n" - "ble 82f\n" - "ld1w { z15.s }, p2/Z, [x7]\n" - "fmla z24.s, z15.s, z0.s[3]\n" - "addvl x7, x7, #1\n" - "fmla z25.s, z15.s, z1.s[3]\n" - "fmla z26.s, z15.s, z2.s[3]\n" - "fmla z27.s, z15.s, z3.s[3]\n" - "fmla z28.s, z15.s, z4.s[3]\n" - "fmla z29.s, z15.s, z5.s[3]\n" - "82:" // Height 6: Multiply loop: multiply skip - "prfm pldl1keep, [x14, #0x80]\n" - "add x16, x16, #0x1\n" - "prfm pldl1keep, [x12, #0x80]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "fmla z28.s, z8.s, z4.s[0]\n" + "ld1rqw { z5.s }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" + "fmla z29.s, z8.s, z5.s[0]\n" + "add x22, x22, #0x10\n" + "addvl x12, x12, #1\n" + "ble 76f\n" + "ld1w { z9.s }, p2/Z, [x12]\n" + "fmla z24.s, z9.s, z0.s[1]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z9.s, z1.s[1]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z9.s, z2.s[1]\n" + "fmla z27.s, z9.s, z3.s[1]\n" + "fmla z28.s, z9.s, z4.s[1]\n" + "fmla z29.s, z9.s, z5.s[1]\n" + "ble 76f\n" + "ld1w { z10.s }, p2/Z, [x12]\n" + "fmla z24.s, z10.s, z0.s[2]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z10.s, z1.s[2]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z10.s, z2.s[2]\n" + "fmla z27.s, z10.s, z3.s[2]\n" + "fmla z28.s, z10.s, z4.s[2]\n" + "fmla z29.s, z10.s, z5.s[2]\n" + "ble 76f\n" + "ld1w { z11.s }, p2/Z, [x12]\n" + "fmla z24.s, z11.s, z0.s[3]\n" + "addvl x12, x12, #1\n" + "fmla z25.s, z11.s, z1.s[3]\n" + "fmla z26.s, z11.s, z2.s[3]\n" + "fmla z27.s, z11.s, z3.s[3]\n" + "fmla z28.s, z11.s, z4.s[3]\n" + "fmla z29.s, z11.s, z5.s[3]\n" + "76:" // Height 6: Multiply loop: multiply skip + "prfm pldl1keep, [x27, #0x80]\n" + "add x9, x9, #0x1\n" "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x16, x19\n" - "bne 77b\n" - "tbz %x[flags], #1, 83f\n" + "cmp x9, x19\n" + "bne 71b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x26, x10, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "add x22, x23, x19, LSL #2\n" + "tbz %x[flags], #1, 77f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rw { z17.s }, p2/Z, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -1108,77 +1058,55 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmax z28.s, p2/M, z28.s, z17.s\n" "fmin z29.s, p2/M, z29.s, z16.s\n" "fmax z29.s, p2/M, z29.s, z17.s\n" - "83:" // Height 6: No activation - "st1w { z24.s }, p1, [x17]\n" - "addvl x17, x17, #1\n" - "st1w { z25.s }, p1, [x13]\n" - "addvl x13, x13, #1\n" - "st1w { z26.s }, p1, [x11]\n" - "addvl x11, x11, #1\n" - "st1w { z27.s }, p1, [x9]\n" - "addvl x9, x9, #1\n" - "st1w { z28.s }, p1, [x27]\n" - "addvl x27, x27, #1\n" - "st1w { z29.s }, p1, [x25]\n" - "addvl x25, x25, #1\n" - "84:" // Height 6: Writeback done - "decw x6\n" - "cmp x6, XZR\n" - "bgt 73b\n" - "b 114f\n" - "85:" // Height 7 - "ldr x6, [%x[args_ptr], %[offsetof_N]]\n" - "mov x8, %x[bias]\n" - "ldr x7, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 86f\n" - "ldr x17, [%x[output_ptr], #0x0]\n" - "add x17, x17, x19, LSL #2\n" - "ldr x13, [%x[output_ptr], #0x8]\n" - "ldr x11, [%x[output_ptr], #0x10]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x18]\n" - "ldr x27, [%x[output_ptr], #0x20]\n" - "add x11, x11, x19, LSL #2\n" - "ldr x25, [%x[output_ptr], #0x28]\n" - "add x9, x9, x19, LSL #2\n" - "ldr x23, [%x[output_ptr], #0x30]\n" - "add x27, x27, x19, LSL #2\n" - "add x25, x25, x19, LSL #2\n" - "add x23, x23, x19, LSL #2\n" - "b 87f\n" - "86:" // Height 7: setup direct output - "mov x17, %x[output_ptr]\n" - "add x13, x17, x19, LSL #2\n" - "add x11, x13, x19, LSL #2\n" - "add x9, x11, x19, LSL #2\n" - "add x27, x9, x19, LSL #2\n" - "add x25, x27, x19, LSL #2\n" - "add x23, x25, x19, LSL #2\n" - "87:" // Height 7: Column loop + "77:" // Height 6: No activation + "st1w { z24.s }, p1, [x10]\n" + "addvl x10, x10, #1\n" + "st1w { z25.s }, p1, [x26]\n" + "st1w { z26.s }, p1, [x25]\n" + "st1w { z27.s }, p1, [x24]\n" + "st1w { z28.s }, p1, [x23]\n" + "st1w { z29.s }, p1, [x22]\n" + "78:" // Height 6: Writeback done + "decw x13\n" + "cmp x13, XZR\n" + "bgt 67b\n" + "b 106f\n" + "79:" // Height 7 + "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" + "mov x11, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x10, %x[output_ptr]\n" + "80:" // Height 7: Column loop "mov x19, #0x0\n" - "whilelt p1.s, x19, x6\n" - "cbz x8, 88f\n" - "ld1w { z24.s }, p2/Z, [x8]\n" + "whilelt p1.s, x19, x13\n" + "cbz x11, 81f\n" + "ld1w { z24.s }, p2/Z, [x11]\n" "mov z25.d, z24.d\n" - "addvl x8, x8, #1\n" + "addvl x11, x11, #1\n" "mov z26.d, z24.d\n" "mov z27.d, z24.d\n" "mov z28.d, z24.d\n" "mov z29.d, z24.d\n" "mov z30.d, z24.d\n" - "b 90f\n" - "88:" // Height 7: no bias - "tbz %x[flags], #0, 89f\n" - "ld1w { z24.s }, p1/Z, [x17]\n" - "ld1w { z25.s }, p1/Z, [x13]\n" - "ld1w { z26.s }, p1/Z, [x11]\n" - "ld1w { z27.s }, p1/Z, [x9]\n" - "ld1w { z28.s }, p1/Z, [x27]\n" - "ld1w { z29.s }, p1/Z, [x25]\n" - "ld1w { z30.s }, p1/Z, [x23]\n" - "b 90f\n" - "89:" // Height 7: no accumulate + "b 83f\n" + "81:" // Height 7: no bias + "tbz %x[flags], #0, 82f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z24.s }, p1/Z, [x10]\n" + "add x26, x10, x19, LSL #2\n" + "ld1w { z25.s }, p1/Z, [x26]\n" + "add x25, x26, x19, LSL #2\n" + "ld1w { z26.s }, p1/Z, [x25]\n" + "add x24, x25, x19, LSL #2\n" + "ld1w { z27.s }, p1/Z, [x24]\n" + "add x23, x24, x19, LSL #2\n" + "ld1w { z28.s }, p1/Z, [x23]\n" + "add x22, x23, x19, LSL #2\n" + "ld1w { z29.s }, p1/Z, [x22]\n" + "add x21, x22, x19, LSL #2\n" + "ld1w { z30.s }, p1/Z, [x21]\n" + "b 83f\n" + "82:" // Height 7: no accumulate "mov z24.b, #0x0\n" "mov z25.b, #0x0\n" "mov z26.b, #0x0\n" @@ -1186,88 +1114,88 @@ void sve_hybrid_fp32_mla_8x1VL ( "mov z28.b, #0x0\n" "mov z29.b, #0x0\n" "mov z30.b, #0x0\n" - "90:" // Height 7: setup done - "mov x16, #0x0\n" - "91:" // Height 7: String loop + "83:" // Height 7: setup done + "mov x9, #0x0\n" + "84:" // Height 7: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w15, [x20, x16, LSL #0x2]\n" - "tbz %x[flags], #3, 92f\n" - "ldr x20, [%x[input_ptr], x16, LSL #0x3]\n" + "ldr w28, [x20, x9, LSL #0x2]\n" + "tbz %x[flags], #3, 85f\n" + "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x14, [x20, #0x0]\n" - "ldr x12, [x20, #0x8]\n" - "ldr x10, [x20, #0x10]\n" - "ldr x28, [x20, #0x18]\n" - "ldr x26, [x20, #0x20]\n" - "ldr x24, [x20, #0x28]\n" - "ldr x22, [x20, #0x30]\n" - "cbnz x16, 93f\n" + "ldr x27, [x20, #0x0]\n" + "ldr x26, [x20, #0x8]\n" + "ldr x25, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "ldr x23, [x20, #0x20]\n" + "ldr x22, [x20, #0x28]\n" + "ldr x21, [x20, #0x30]\n" + "cbnz x9, 86f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x14, x14, x19, LSL #2\n" - "add x12, x12, x19, LSL #2\n" - "add x10, x10, x19, LSL #2\n" - "add x28, x28, x19, LSL #2\n" + "add x27, x27, x19, LSL #2\n" "add x26, x26, x19, LSL #2\n" + "add x25, x25, x19, LSL #2\n" "add x24, x24, x19, LSL #2\n" + "add x23, x23, x19, LSL #2\n" "add x22, x22, x19, LSL #2\n" - "b 93f\n" - "92:" // Height 7: setup direct input - "mov x14, %x[input_ptr]\n" - "add x12, x14, x19, LSL #2\n" - "add x10, x12, x19, LSL #2\n" - "add x28, x10, x19, LSL #2\n" - "add x26, x28, x19, LSL #2\n" - "add x24, x26, x19, LSL #2\n" - "add x22, x24, x19, LSL #2\n" - "93:" // Height 7: input setup done - "cmp x15, #0x4\n" - "ble 95f\n" - "94:" // Height 7: Multiply loop: Main loop head - "ld1w { z8.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "ld1w { z9.s }, p2/Z, [x7, #1, MUL VL]\n" - "sub x15, x15, #0x4\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" + "add x21, x21, x19, LSL #2\n" + "b 86f\n" + "85:" // Height 7: setup direct input + "mov x27, %x[input_ptr]\n" + "add x26, x27, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "add x22, x23, x19, LSL #2\n" + "add x21, x22, x19, LSL #2\n" + "86:" // Height 7: input setup done + "cmp x28, #0x4\n" + "ble 88f\n" + "87:" // Height 7: Multiply loop: Main loop head + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n" + "sub x28, x28, #0x4\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" "fmla z24.s, z8.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" "fmla z25.s, z8.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" + "add x26, x26, #0x10\n" "fmla z24.s, z9.s, z0.s[1]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "fmla z26.s, z8.s, z2.s[0]\n" - "ld1rqw { z4.s }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "fmla z27.s, z8.s, z3.s[0]\n" - "ld1rqw { z5.s }, p0/Z, [x24]\n" - "add x26, x26, #0x10\n" + "ld1rqw { z5.s }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" "fmla z25.s, z9.s, z1.s[1]\n" - "ld1rqw { z6.s }, p0/Z, [x22]\n" - "add x24, x24, #0x10\n" - "fmla z28.s, z8.s, z4.s[0]\n" - "ld1w { z10.s }, p2/Z, [x7, #2, MUL VL]\n" + "ld1rqw { z6.s }, p0/Z, [x21]\n" "add x22, x22, #0x10\n" + "fmla z28.s, z8.s, z4.s[0]\n" + "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" + "add x21, x21, #0x10\n" "fmla z29.s, z8.s, z5.s[0]\n" - "ld1w { z11.s }, p2/Z, [x7, #3, MUL VL]\n" - "cmp x15, #0x4\n" + "ld1w { z11.s }, p2/Z, [x12, #3, MUL VL]\n" + "cmp x28, #0x4\n" "fmla z30.s, z8.s, z6.s[0]\n" - "prfm pldl1keep, [x14, #0x80]\n" - "addvl x7, x7, #4\n" + "prfm pldl1keep, [x27, #0x80]\n" + "addvl x12, x12, #4\n" "fmla z26.s, z9.s, z2.s[1]\n" - "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla z27.s, z9.s, z3.s[1]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla z28.s, z9.s, z4.s[1]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla z29.s, z9.s, z5.s[1]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "fmla z30.s, z9.s, z6.s[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "fmla z24.s, z10.s, z0.s[2]\n" "prfm pldl1keep, [x22, #0x80]\n" + "fmla z24.s, z10.s, z0.s[2]\n" + "prfm pldl1keep, [x21, #0x80]\n" "fmla z25.s, z10.s, z1.s[2]\n" "fmla z26.s, z10.s, z2.s[2]\n" "fmla z27.s, z10.s, z3.s[2]\n" @@ -1281,78 +1209,85 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmla z28.s, z11.s, z4.s[3]\n" "fmla z29.s, z11.s, z5.s[3]\n" "fmla z30.s, z11.s, z6.s[3]\n" - "bgt 94b\n" - "95:" // Height 7: Multiply loop: Single iteration only - "ld1w { z12.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "subs x15, x15, #0x1\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" - "fmla z24.s, z12.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" - "fmla z25.s, z12.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" - "fmla z26.s, z12.s, z2.s[0]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" - "fmla z27.s, z12.s, z3.s[0]\n" - "ld1rqw { z4.s }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" - "fmla z28.s, z12.s, z4.s[0]\n" - "ld1rqw { z5.s }, p0/Z, [x24]\n" + "bgt 87b\n" + "88:" // Height 7: Multiply loop: Single iteration only + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "subs x28, x28, #0x1\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" + "fmla z24.s, z8.s, z0.s[0]\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" + "fmla z25.s, z8.s, z1.s[0]\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" "add x26, x26, #0x10\n" - "fmla z29.s, z12.s, z5.s[0]\n" - "ld1rqw { z6.s }, p0/Z, [x22]\n" + "fmla z26.s, z8.s, z2.s[0]\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" + "fmla z27.s, z8.s, z3.s[0]\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" "add x24, x24, #0x10\n" - "fmla z30.s, z12.s, z6.s[0]\n" + "fmla z28.s, z8.s, z4.s[0]\n" + "ld1rqw { z5.s }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" + "fmla z29.s, z8.s, z5.s[0]\n" + "ld1rqw { z6.s }, p0/Z, [x21]\n" "add x22, x22, #0x10\n" - "addvl x7, x7, #1\n" - "ble 96f\n" - "ld1w { z13.s }, p2/Z, [x7]\n" - "fmla z24.s, z13.s, z0.s[1]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z13.s, z1.s[1]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z13.s, z2.s[1]\n" - "fmla z27.s, z13.s, z3.s[1]\n" - "fmla z28.s, z13.s, z4.s[1]\n" - "fmla z29.s, z13.s, z5.s[1]\n" - "fmla z30.s, z13.s, z6.s[1]\n" - "ble 96f\n" - "ld1w { z14.s }, p2/Z, [x7]\n" - "fmla z24.s, z14.s, z0.s[2]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z14.s, z1.s[2]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z14.s, z2.s[2]\n" - "fmla z27.s, z14.s, z3.s[2]\n" - "fmla z28.s, z14.s, z4.s[2]\n" - "fmla z29.s, z14.s, z5.s[2]\n" - "fmla z30.s, z14.s, z6.s[2]\n" - "ble 96f\n" - "ld1w { z15.s }, p2/Z, [x7]\n" - "fmla z24.s, z15.s, z0.s[3]\n" - "addvl x7, x7, #1\n" - "fmla z25.s, z15.s, z1.s[3]\n" - "fmla z26.s, z15.s, z2.s[3]\n" - "fmla z27.s, z15.s, z3.s[3]\n" - "fmla z28.s, z15.s, z4.s[3]\n" - "fmla z29.s, z15.s, z5.s[3]\n" - "fmla z30.s, z15.s, z6.s[3]\n" - "96:" // Height 7: Multiply loop: multiply skip - "prfm pldl1keep, [x14, #0x80]\n" - "add x16, x16, #0x1\n" - "prfm pldl1keep, [x12, #0x80]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "fmla z30.s, z8.s, z6.s[0]\n" + "add x21, x21, #0x10\n" + "addvl x12, x12, #1\n" + "ble 89f\n" + "ld1w { z9.s }, p2/Z, [x12]\n" + "fmla z24.s, z9.s, z0.s[1]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z9.s, z1.s[1]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z9.s, z2.s[1]\n" + "fmla z27.s, z9.s, z3.s[1]\n" + "fmla z28.s, z9.s, z4.s[1]\n" + "fmla z29.s, z9.s, z5.s[1]\n" + "fmla z30.s, z9.s, z6.s[1]\n" + "ble 89f\n" + "ld1w { z10.s }, p2/Z, [x12]\n" + "fmla z24.s, z10.s, z0.s[2]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z10.s, z1.s[2]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z10.s, z2.s[2]\n" + "fmla z27.s, z10.s, z3.s[2]\n" + "fmla z28.s, z10.s, z4.s[2]\n" + "fmla z29.s, z10.s, z5.s[2]\n" + "fmla z30.s, z10.s, z6.s[2]\n" + "ble 89f\n" + "ld1w { z11.s }, p2/Z, [x12]\n" + "fmla z24.s, z11.s, z0.s[3]\n" + "addvl x12, x12, #1\n" + "fmla z25.s, z11.s, z1.s[3]\n" + "fmla z26.s, z11.s, z2.s[3]\n" + "fmla z27.s, z11.s, z3.s[3]\n" + "fmla z28.s, z11.s, z4.s[3]\n" + "fmla z29.s, z11.s, z5.s[3]\n" + "fmla z30.s, z11.s, z6.s[3]\n" + "89:" // Height 7: Multiply loop: multiply skip + "prfm pldl1keep, [x27, #0x80]\n" + "add x9, x9, #0x1\n" "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x21, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x16, x19\n" - "bne 91b\n" - "tbz %x[flags], #1, 97f\n" + "cmp x9, x19\n" + "bne 84b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x26, x10, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "add x22, x23, x19, LSL #2\n" + "add x21, x22, x19, LSL #2\n" + "tbz %x[flags], #1, 90f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rw { z17.s }, p2/Z, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -1371,86 +1306,62 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmin z30.s, p2/M, z30.s, z16.s\n" "fmax z29.s, p2/M, z29.s, z17.s\n" "fmax z30.s, p2/M, z30.s, z17.s\n" - "97:" // Height 7: No activation - "st1w { z24.s }, p1, [x17]\n" - "addvl x17, x17, #1\n" - "st1w { z25.s }, p1, [x13]\n" - "addvl x13, x13, #1\n" - "st1w { z26.s }, p1, [x11]\n" - "addvl x11, x11, #1\n" - "st1w { z27.s }, p1, [x9]\n" - "addvl x9, x9, #1\n" - "st1w { z28.s }, p1, [x27]\n" - "addvl x27, x27, #1\n" - "st1w { z29.s }, p1, [x25]\n" - "addvl x25, x25, #1\n" - "st1w { z30.s }, p1, [x23]\n" - "addvl x23, x23, #1\n" - "98:" // Height 7: Writeback done - "decw x6\n" - "cmp x6, XZR\n" - "bgt 87b\n" - "b 114f\n" - "99:" // Height 8 - "ldr x6, [%x[args_ptr], %[offsetof_N]]\n" - "mov x8, %x[bias]\n" - "ldr x7, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 100f\n" - "ldr x17, [%x[output_ptr], #0x0]\n" - "add x17, x17, x19, LSL #2\n" - "ldr x13, [%x[output_ptr], #0x8]\n" - "ldr x11, [%x[output_ptr], #0x10]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x18]\n" - "ldr x27, [%x[output_ptr], #0x20]\n" - "add x11, x11, x19, LSL #2\n" - "ldr x25, [%x[output_ptr], #0x28]\n" - "add x9, x9, x19, LSL #2\n" - "ldr x23, [%x[output_ptr], #0x30]\n" - "ldr x21, [%x[output_ptr], #0x38]\n" - "add x27, x27, x19, LSL #2\n" - "add x25, x25, x19, LSL #2\n" - "add %x[output_ptr], %x[output_ptr], #0x40\n" - "add x23, x23, x19, LSL #2\n" - "add x21, x21, x19, LSL #2\n" - "b 101f\n" - "100:" // Height 8: setup direct output - "mov x17, %x[output_ptr]\n" - "add x13, x17, x19, LSL #2\n" - "add x11, x13, x19, LSL #2\n" - "add x9, x11, x19, LSL #2\n" - "add x27, x9, x19, LSL #2\n" - "add x25, x27, x19, LSL #2\n" - "add x23, x25, x19, LSL #2\n" - "add x21, x23, x19, LSL #2\n" - "add %x[output_ptr], x21, x19, LSL #2\n" - "101:" // Height 8: Column loop + "90:" // Height 7: No activation + "st1w { z24.s }, p1, [x10]\n" + "addvl x10, x10, #1\n" + "st1w { z25.s }, p1, [x26]\n" + "st1w { z26.s }, p1, [x25]\n" + "st1w { z27.s }, p1, [x24]\n" + "st1w { z28.s }, p1, [x23]\n" + "st1w { z29.s }, p1, [x22]\n" + "st1w { z30.s }, p1, [x21]\n" + "91:" // Height 7: Writeback done + "decw x13\n" + "cmp x13, XZR\n" + "bgt 80b\n" + "b 106f\n" + "92:" // Height 8 + "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" + "mov x11, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x10, %x[output_ptr]\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "mov x19, #0x20\n" + "madd %x[output_ptr], x20, x19, %x[output_ptr]\n" + "93:" // Height 8: Column loop "mov x19, #0x0\n" - "whilelt p1.s, x19, x6\n" - "cbz x8, 102f\n" - "ld1w { z24.s }, p2/Z, [x8]\n" + "whilelt p1.s, x19, x13\n" + "cbz x11, 94f\n" + "ld1w { z24.s }, p2/Z, [x11]\n" "mov z25.d, z24.d\n" - "addvl x8, x8, #1\n" + "addvl x11, x11, #1\n" "mov z26.d, z24.d\n" "mov z27.d, z24.d\n" "mov z28.d, z24.d\n" "mov z29.d, z24.d\n" "mov z30.d, z24.d\n" "mov z31.d, z24.d\n" - "b 104f\n" - "102:" // Height 8: no bias - "tbz %x[flags], #0, 103f\n" - "ld1w { z24.s }, p1/Z, [x17]\n" - "ld1w { z25.s }, p1/Z, [x13]\n" - "ld1w { z26.s }, p1/Z, [x11]\n" - "ld1w { z27.s }, p1/Z, [x9]\n" - "ld1w { z28.s }, p1/Z, [x27]\n" - "ld1w { z29.s }, p1/Z, [x25]\n" - "ld1w { z30.s }, p1/Z, [x23]\n" - "ld1w { z31.s }, p1/Z, [x21]\n" - "b 104f\n" - "103:" // Height 8: no accumulate + "b 96f\n" + "94:" // Height 8: no bias + "tbz %x[flags], #0, 95f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ld1w { z24.s }, p1/Z, [x10]\n" + "add x26, x10, x19, LSL #2\n" + "ld1w { z25.s }, p1/Z, [x26]\n" + "add x25, x26, x19, LSL #2\n" + "ld1w { z26.s }, p1/Z, [x25]\n" + "add x24, x25, x19, LSL #2\n" + "ld1w { z27.s }, p1/Z, [x24]\n" + "add x23, x24, x19, LSL #2\n" + "ld1w { z28.s }, p1/Z, [x23]\n" + "add x22, x23, x19, LSL #2\n" + "ld1w { z29.s }, p1/Z, [x22]\n" + "add x21, x22, x19, LSL #2\n" + "ld1w { z30.s }, p1/Z, [x21]\n" + "add x20, x21, x19, LSL #2\n" + "ld1w { z31.s }, p1/Z, [x20]\n" + "b 96f\n" + "95:" // Height 8: no accumulate "mov z24.b, #0x0\n" "mov z25.b, #0x0\n" "mov z26.b, #0x0\n" @@ -1459,94 +1370,94 @@ void sve_hybrid_fp32_mla_8x1VL ( "mov z29.b, #0x0\n" "mov z30.b, #0x0\n" "mov z31.b, #0x0\n" - "104:" // Height 8: setup done - "mov x16, #0x0\n" - "105:" // Height 8: String loop + "96:" // Height 8: setup done + "mov x9, #0x0\n" + "97:" // Height 8: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w15, [x20, x16, LSL #0x2]\n" - "tbz %x[flags], #3, 106f\n" - "ldr x20, [%x[input_ptr], x16, LSL #0x3]\n" + "ldr w28, [x20, x9, LSL #0x2]\n" + "tbz %x[flags], #3, 98f\n" + "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x14, [x20, #0x0]\n" - "ldr x12, [x20, #0x8]\n" - "ldr x10, [x20, #0x10]\n" - "ldr x28, [x20, #0x18]\n" - "ldr x26, [x20, #0x20]\n" - "ldr x24, [x20, #0x28]\n" - "ldr x22, [x20, #0x30]\n" + "ldr x27, [x20, #0x0]\n" + "ldr x26, [x20, #0x8]\n" + "ldr x25, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "ldr x23, [x20, #0x20]\n" + "ldr x22, [x20, #0x28]\n" + "ldr x21, [x20, #0x30]\n" "ldr x20, [x20, #0x38]\n" - "cbnz x16, 107f\n" + "cbnz x9, 99f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x14, x14, x19, LSL #2\n" - "add x12, x12, x19, LSL #2\n" - "add x10, x10, x19, LSL #2\n" - "add x28, x28, x19, LSL #2\n" + "add x27, x27, x19, LSL #2\n" "add x26, x26, x19, LSL #2\n" + "add x25, x25, x19, LSL #2\n" "add x24, x24, x19, LSL #2\n" + "add x23, x23, x19, LSL #2\n" "add x22, x22, x19, LSL #2\n" + "add x21, x21, x19, LSL #2\n" "add x20, x20, x19, LSL #2\n" - "b 107f\n" - "106:" // Height 8: setup direct input - "mov x14, %x[input_ptr]\n" - "add x12, x14, x19, LSL #2\n" - "add x10, x12, x19, LSL #2\n" - "add x28, x10, x19, LSL #2\n" - "add x26, x28, x19, LSL #2\n" - "add x24, x26, x19, LSL #2\n" - "add x22, x24, x19, LSL #2\n" - "add x20, x22, x19, LSL #2\n" - "107:" // Height 8: input setup done - "cmp x15, #0x4\n" - "ble 109f\n" - "108:" // Height 8: Multiply loop: Main loop head - "ld1w { z8.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "ld1w { z9.s }, p2/Z, [x7, #1, MUL VL]\n" - "sub x15, x15, #0x4\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" + "b 99f\n" + "98:" // Height 8: setup direct input + "mov x27, %x[input_ptr]\n" + "add x26, x27, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "add x22, x23, x19, LSL #2\n" + "add x21, x22, x19, LSL #2\n" + "add x20, x21, x19, LSL #2\n" + "99:" // Height 8: input setup done + "cmp x28, #0x4\n" + "ble 101f\n" + "100:" // Height 8: Multiply loop: Main loop head + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n" + "sub x28, x28, #0x4\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" "fmla z24.s, z8.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" "fmla z25.s, z8.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" + "add x26, x26, #0x10\n" "fmla z24.s, z9.s, z0.s[1]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" "fmla z26.s, z8.s, z2.s[0]\n" - "ld1rqw { z4.s }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" + "add x24, x24, #0x10\n" "fmla z27.s, z8.s, z3.s[0]\n" - "ld1rqw { z5.s }, p0/Z, [x24]\n" - "add x26, x26, #0x10\n" + "ld1rqw { z5.s }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" "fmla z25.s, z9.s, z1.s[1]\n" - "ld1rqw { z6.s }, p0/Z, [x22]\n" - "add x24, x24, #0x10\n" + "ld1rqw { z6.s }, p0/Z, [x21]\n" + "add x22, x22, #0x10\n" "fmla z28.s, z8.s, z4.s[0]\n" "ld1rqw { z7.s }, p0/Z, [x20]\n" - "add x22, x22, #0x10\n" + "add x21, x21, #0x10\n" "fmla z29.s, z8.s, z5.s[0]\n" - "ld1w { z10.s }, p2/Z, [x7, #2, MUL VL]\n" + "ld1w { z10.s }, p2/Z, [x12, #2, MUL VL]\n" "add x20, x20, #0x10\n" "fmla z30.s, z8.s, z6.s[0]\n" - "ld1w { z11.s }, p2/Z, [x7, #3, MUL VL]\n" - "cmp x15, #0x4\n" + "ld1w { z11.s }, p2/Z, [x12, #3, MUL VL]\n" + "cmp x28, #0x4\n" "fmla z31.s, z8.s, z7.s[0]\n" - "prfm pldl1keep, [x14, #0x80]\n" - "addvl x7, x7, #4\n" + "prfm pldl1keep, [x27, #0x80]\n" + "addvl x12, x12, #4\n" "fmla z26.s, z9.s, z2.s[1]\n" - "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla z27.s, z9.s, z3.s[1]\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla z28.s, z9.s, z4.s[1]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla z29.s, z9.s, z5.s[1]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "fmla z30.s, z9.s, z6.s[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "fmla z31.s, z9.s, z7.s[1]\n" "prfm pldl1keep, [x22, #0x80]\n" + "fmla z31.s, z9.s, z7.s[1]\n" + "prfm pldl1keep, [x21, #0x80]\n" "fmla z24.s, z10.s, z0.s[2]\n" "prfm pldl1keep, [x20, #0x80]\n" "fmla z25.s, z10.s, z1.s[2]\n" @@ -1564,85 +1475,93 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmla z29.s, z11.s, z5.s[3]\n" "fmla z30.s, z11.s, z6.s[3]\n" "fmla z31.s, z11.s, z7.s[3]\n" - "bgt 108b\n" - "109:" // Height 8: Multiply loop: Single iteration only - "ld1w { z12.s }, p2/Z, [x7]\n" - "whilelt p0.s, XZR, x15\n" - "subs x15, x15, #0x1\n" - "ld1rqw { z0.s }, p0/Z, [x14]\n" - "fmla z24.s, z12.s, z0.s[0]\n" - "ld1rqw { z1.s }, p0/Z, [x12]\n" - "add x14, x14, #0x10\n" - "fmla z25.s, z12.s, z1.s[0]\n" - "ld1rqw { z2.s }, p0/Z, [x10]\n" - "add x12, x12, #0x10\n" - "fmla z26.s, z12.s, z2.s[0]\n" - "ld1rqw { z3.s }, p0/Z, [x28]\n" - "add x10, x10, #0x10\n" - "fmla z27.s, z12.s, z3.s[0]\n" - "ld1rqw { z4.s }, p0/Z, [x26]\n" - "add x28, x28, #0x10\n" - "fmla z28.s, z12.s, z4.s[0]\n" - "ld1rqw { z5.s }, p0/Z, [x24]\n" + "bgt 100b\n" + "101:" // Height 8: Multiply loop: Single iteration only + "ld1w { z8.s }, p2/Z, [x12]\n" + "whilelt p0.s, XZR, x28\n" + "subs x28, x28, #0x1\n" + "ld1rqw { z0.s }, p0/Z, [x27]\n" + "fmla z24.s, z8.s, z0.s[0]\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x27, x27, #0x10\n" + "fmla z25.s, z8.s, z1.s[0]\n" + "ld1rqw { z2.s }, p0/Z, [x25]\n" "add x26, x26, #0x10\n" - "fmla z29.s, z12.s, z5.s[0]\n" - "ld1rqw { z6.s }, p0/Z, [x22]\n" + "fmla z26.s, z8.s, z2.s[0]\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x25, x25, #0x10\n" + "fmla z27.s, z8.s, z3.s[0]\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" "add x24, x24, #0x10\n" - "fmla z30.s, z12.s, z6.s[0]\n" - "ld1rqw { z7.s }, p0/Z, [x20]\n" + "fmla z28.s, z8.s, z4.s[0]\n" + "ld1rqw { z5.s }, p0/Z, [x22]\n" + "add x23, x23, #0x10\n" + "fmla z29.s, z8.s, z5.s[0]\n" + "ld1rqw { z6.s }, p0/Z, [x21]\n" "add x22, x22, #0x10\n" - "fmla z31.s, z12.s, z7.s[0]\n" + "fmla z30.s, z8.s, z6.s[0]\n" + "ld1rqw { z7.s }, p0/Z, [x20]\n" + "add x21, x21, #0x10\n" + "fmla z31.s, z8.s, z7.s[0]\n" "add x20, x20, #0x10\n" - "addvl x7, x7, #1\n" - "ble 110f\n" - "ld1w { z13.s }, p2/Z, [x7]\n" - "fmla z24.s, z13.s, z0.s[1]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z13.s, z1.s[1]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z13.s, z2.s[1]\n" - "fmla z27.s, z13.s, z3.s[1]\n" - "fmla z28.s, z13.s, z4.s[1]\n" - "fmla z29.s, z13.s, z5.s[1]\n" - "fmla z30.s, z13.s, z6.s[1]\n" - "fmla z31.s, z13.s, z7.s[1]\n" - "ble 110f\n" - "ld1w { z14.s }, p2/Z, [x7]\n" - "fmla z24.s, z14.s, z0.s[2]\n" - "subs x15, x15, #0x1\n" - "fmla z25.s, z14.s, z1.s[2]\n" - "addvl x7, x7, #1\n" - "fmla z26.s, z14.s, z2.s[2]\n" - "fmla z27.s, z14.s, z3.s[2]\n" - "fmla z28.s, z14.s, z4.s[2]\n" - "fmla z29.s, z14.s, z5.s[2]\n" - "fmla z30.s, z14.s, z6.s[2]\n" - "fmla z31.s, z14.s, z7.s[2]\n" - "ble 110f\n" - "ld1w { z15.s }, p2/Z, [x7]\n" - "fmla z24.s, z15.s, z0.s[3]\n" - "addvl x7, x7, #1\n" - "fmla z25.s, z15.s, z1.s[3]\n" - "fmla z26.s, z15.s, z2.s[3]\n" - "fmla z27.s, z15.s, z3.s[3]\n" - "fmla z28.s, z15.s, z4.s[3]\n" - "fmla z29.s, z15.s, z5.s[3]\n" - "fmla z30.s, z15.s, z6.s[3]\n" - "fmla z31.s, z15.s, z7.s[3]\n" - "110:" // Height 8: Multiply loop: multiply skip - "prfm pldl1keep, [x14, #0x80]\n" - "add x16, x16, #0x1\n" - "prfm pldl1keep, [x12, #0x80]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "prfm pldl1keep, [x28, #0x80]\n" + "addvl x12, x12, #1\n" + "ble 102f\n" + "ld1w { z9.s }, p2/Z, [x12]\n" + "fmla z24.s, z9.s, z0.s[1]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z9.s, z1.s[1]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z9.s, z2.s[1]\n" + "fmla z27.s, z9.s, z3.s[1]\n" + "fmla z28.s, z9.s, z4.s[1]\n" + "fmla z29.s, z9.s, z5.s[1]\n" + "fmla z30.s, z9.s, z6.s[1]\n" + "fmla z31.s, z9.s, z7.s[1]\n" + "ble 102f\n" + "ld1w { z10.s }, p2/Z, [x12]\n" + "fmla z24.s, z10.s, z0.s[2]\n" + "subs x28, x28, #0x1\n" + "fmla z25.s, z10.s, z1.s[2]\n" + "addvl x12, x12, #1\n" + "fmla z26.s, z10.s, z2.s[2]\n" + "fmla z27.s, z10.s, z3.s[2]\n" + "fmla z28.s, z10.s, z4.s[2]\n" + "fmla z29.s, z10.s, z5.s[2]\n" + "fmla z30.s, z10.s, z6.s[2]\n" + "fmla z31.s, z10.s, z7.s[2]\n" + "ble 102f\n" + "ld1w { z11.s }, p2/Z, [x12]\n" + "fmla z24.s, z11.s, z0.s[3]\n" + "addvl x12, x12, #1\n" + "fmla z25.s, z11.s, z1.s[3]\n" + "fmla z26.s, z11.s, z2.s[3]\n" + "fmla z27.s, z11.s, z3.s[3]\n" + "fmla z28.s, z11.s, z4.s[3]\n" + "fmla z29.s, z11.s, z5.s[3]\n" + "fmla z30.s, z11.s, z6.s[3]\n" + "fmla z31.s, z11.s, z7.s[3]\n" + "102:" // Height 8: Multiply loop: multiply skip + "prfm pldl1keep, [x27, #0x80]\n" + "add x9, x9, #0x1\n" "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x21, #0x80]\n" "prfm pldl1keep, [x20, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "cmp x16, x19\n" - "bne 105b\n" - "tbz %x[flags], #1, 111f\n" + "cmp x9, x19\n" + "bne 97b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x26, x10, x19, LSL #2\n" + "add x25, x26, x19, LSL #2\n" + "add x24, x25, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "add x22, x23, x19, LSL #2\n" + "add x21, x22, x19, LSL #2\n" + "add x20, x21, x19, LSL #2\n" + "tbz %x[flags], #1, 103f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rw { z17.s }, p2/Z, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -1663,43 +1582,36 @@ void sve_hybrid_fp32_mla_8x1VL ( "fmax z29.s, p2/M, z29.s, z17.s\n" "fmax z30.s, p2/M, z30.s, z17.s\n" "fmax z31.s, p2/M, z31.s, z17.s\n" - "111:" // Height 8: No activation - "st1w { z24.s }, p1, [x17]\n" - "addvl x17, x17, #1\n" - "st1w { z25.s }, p1, [x13]\n" - "addvl x13, x13, #1\n" - "st1w { z26.s }, p1, [x11]\n" - "addvl x11, x11, #1\n" - "st1w { z27.s }, p1, [x9]\n" - "addvl x9, x9, #1\n" - "st1w { z28.s }, p1, [x27]\n" - "addvl x27, x27, #1\n" - "st1w { z29.s }, p1, [x25]\n" - "addvl x25, x25, #1\n" - "st1w { z30.s }, p1, [x23]\n" - "addvl x23, x23, #1\n" - "st1w { z31.s }, p1, [x21]\n" - "addvl x21, x21, #1\n" - "112:" // Height 8: Writeback done - "decw x6\n" - "cmp x6, XZR\n" - "bgt 101b\n" + "103:" // Height 8: No activation + "st1w { z24.s }, p1, [x10]\n" + "addvl x10, x10, #1\n" + "st1w { z25.s }, p1, [x26]\n" + "st1w { z26.s }, p1, [x25]\n" + "st1w { z27.s }, p1, [x24]\n" + "st1w { z28.s }, p1, [x23]\n" + "st1w { z29.s }, p1, [x22]\n" + "st1w { z30.s }, p1, [x21]\n" + "st1w { z31.s }, p1, [x20]\n" + "104:" // Height 8: Writeback done + "decw x13\n" + "cmp x13, XZR\n" + "bgt 93b\n" "subs %x[M], %x[M], #0x8\n" - "beq 114f\n" + "beq 106f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" - "tbz %x[flags], #3, 113f\n" + "tbz %x[flags], #3, 105f\n" "add x20, x20, #0x8\n" "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "b 1b\n" - "113:" // Update direct input + "105:" // Update direct input "mov x19, #0x20\n" "madd %x[input_ptr], x19, x20, %x[input_ptr]\n" "b 1b\n" - "114:" // Exit + "106:" // 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", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "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", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x13", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z16", "z17", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } |