From f78b12ce70767de762be8530e770645daa02a585 Mon Sep 17 00:00:00 2001 From: Pablo Marquez Tello Date: Tue, 2 Jun 2026 11:24:54 +0100 Subject: [PATCH] perf: Speed up s8q NHWC max pooling Refactor the A64, SVE and SME QASYMM8_SIGNED differing-qinfo NHWC MAX pooling kernels to reduce the cost of the requantized path. The old code did: - add output offset - explicit clamp to [-128, 127] with smax / smin - pack down with several uzp1 shuffles - The new code does: - add output offset - saturating narrow directly with sqxtn / sqxtn2 from s32 -> s16 -> s8 Why this helps: - sqxtn already performs the signed saturation, so the explicit clamp is redundant. - It also lets us remove a chunk of shuffle/packing work. - So the requantized epilogue gets shorter: fewer instructions, less register traffic, less packing overhead. Signed-off-by: Pablo Marquez Tello Change-Id: I47a378aeda61de86f2d6784393ccb4a08984706c --- .../generic.cpp | 806 +++++++++--------- .../generic.cpp | 573 ++++++------- .../generic.cpp | 555 ++++++------ 3 files changed, 920 insertions(+), 1014 deletions(-) diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp index 94bf38c29b5..b25ae2fc218 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp @@ -43,77 +43,77 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( { __asm__ __volatile__( "cmp %x[n_channels], #0x40\n" - "mov x9, #0\n" - "mov x28, #0x10\n" // cntb _, ALL, #1 - "mov x27, #0x20\n" // cntb _, ALL, #2 - "mov x26, #0x30\n" // cntb _, ALL, #3 + "mov x21, #0\n" + "mov x20, #0x10\n" // cntb _, ALL, #1 + "mov x8, #0x20\n" // cntb _, ALL, #2 + "mov x7, #0x30\n" // cntb _, ALL, #3 "blt 7f\n" "1:" // 4-vectors of channels - "lsr x25, %x[n_valid_cells], #0x2\n" - "movi v6.16b, #0x80\n" + "lsr x6, %x[n_valid_cells], #0x2\n" "movi v9.16b, #0x80\n" - "mov x24, %x[inptrs]\n" "movi v8.16b, #0x80\n" + "mov x5, %x[inptrs]\n" "movi v7.16b, #0x80\n" - "cbz x25, 4f\n" - "ldp x23, x22, [x24, #0]\n" - "ldp x21, x20, [x24, #0x10]\n" - "subs x25, x25, #0x1\n" - "add x24, x24, #0x20\n" - "ldr q5, [x23, x9]\n" - "ldr q4, [x22, x9]\n" - "ldr q3, [x21, x9]\n" - "ldr q2, [x20, x9]\n" - "ldr q1, [x23, x28]\n" - "ldr q0, [x22, x28]\n" - "ldr q31, [x21, x28]\n" - "ldr q30, [x20, x28]\n" - "ldr q29, [x23, x27]\n" - "ldr q21, [x22, x27]\n" - "ldr q28, [x21, x27]\n" - "ldr q27, [x20, x27]\n" - "ldr q26, [x23, x26]\n" - "ldr q20, [x22, x26]\n" - "ldr q25, [x21, x26]\n" - "ldr q24, [x20, x26]\n" + "movi v6.16b, #0x80\n" + "cbz x6, 4f\n" + "ldp x4, x3, [x5, #0]\n" + "ldp x2, x1, [x5, #0x10]\n" + "subs x6, x6, #0x1\n" + "add x5, x5, #0x20\n" + "ldr q5, [x4, x21]\n" + "ldr q4, [x3, x21]\n" + "ldr q3, [x2, x21]\n" + "ldr q2, [x1, x21]\n" + "ldr q1, [x4, x20]\n" + "ldr q0, [x3, x20]\n" + "ldr q31, [x2, x20]\n" + "ldr q30, [x1, x20]\n" + "ldr q29, [x4, x8]\n" + "ldr q21, [x3, x8]\n" + "ldr q28, [x2, x8]\n" + "ldr q27, [x1, x8]\n" + "ldr q26, [x4, x7]\n" + "ldr q20, [x3, x7]\n" + "ldr q25, [x2, x7]\n" + "ldr q24, [x1, x7]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop "smax v23.16b, v5.16b, v4.16b\n" "smax v19.16b, v3.16b, v2.16b\n" - "ldp x23, x22, [x24, #0]\n" - "ldp x21, x20, [x24, #0x10]\n" + "ldp x4, x3, [x5, #0]\n" + "ldp x2, x1, [x5, #0x10]\n" "smax v22.16b, v1.16b, v0.16b\n" "smax v18.16b, v31.16b, v30.16b\n" - "subs x25, x25, #0x1\n" - "add x24, x24, #0x20\n" + "subs x6, x6, #0x1\n" + "add x5, x5, #0x20\n" "smax v21.16b, v29.16b, v21.16b\n" "smax v17.16b, v28.16b, v27.16b\n" - "ldr q5, [x23, x9]\n" - "ldr q4, [x22, x9]\n" - "ldr q3, [x21, x9]\n" - "ldr q2, [x20, x9]\n" + "ldr q5, [x4, x21]\n" + "ldr q4, [x3, x21]\n" + "ldr q3, [x2, x21]\n" + "ldr q2, [x1, x21]\n" "smax v20.16b, v26.16b, v20.16b\n" "smax v16.16b, v25.16b, v24.16b\n" - "ldr q1, [x23, x28]\n" - "ldr q0, [x22, x28]\n" + "ldr q1, [x4, x20]\n" + "ldr q0, [x3, x20]\n" "smax v19.16b, v23.16b, v19.16b\n" "smax v18.16b, v22.16b, v18.16b\n" - "ldr q31, [x21, x28]\n" - "ldr q30, [x20, x28]\n" + "ldr q31, [x2, x20]\n" + "ldr q30, [x1, x20]\n" "smax v17.16b, v21.16b, v17.16b\n" - "ldr q29, [x23, x27]\n" - "ldr q21, [x22, x27]\n" + "ldr q29, [x4, x8]\n" + "ldr q21, [x3, x8]\n" "smax v16.16b, v20.16b, v16.16b\n" - "ldr q28, [x21, x27]\n" - "ldr q27, [x20, x27]\n" - "smax v6.16b, v6.16b, v19.16b\n" - "smax v9.16b, v9.16b, v18.16b\n" - "ldr q26, [x23, x26]\n" - "ldr q20, [x22, x26]\n" - "smax v8.16b, v8.16b, v17.16b\n" - "ldr q25, [x21, x26]\n" - "ldr q24, [x20, x26]\n" - "smax v7.16b, v7.16b, v16.16b\n" + "ldr q28, [x2, x8]\n" + "ldr q27, [x1, x8]\n" + "smax v9.16b, v9.16b, v19.16b\n" + "smax v8.16b, v8.16b, v18.16b\n" + "ldr q26, [x4, x7]\n" + "ldr q20, [x3, x7]\n" + "smax v7.16b, v7.16b, v17.16b\n" + "ldr q25, [x2, x7]\n" + "ldr q24, [x1, x7]\n" + "smax v6.16b, v6.16b, v16.16b\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail "smax v23.16b, v5.16b, v4.16b\n" @@ -128,497 +128,461 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "smax v18.16b, v22.16b, v18.16b\n" "smax v17.16b, v21.16b, v17.16b\n" "smax v16.16b, v20.16b, v16.16b\n" - "smax v6.16b, v6.16b, v19.16b\n" - "smax v9.16b, v9.16b, v18.16b\n" - "smax v8.16b, v8.16b, v17.16b\n" - "smax v7.16b, v7.16b, v16.16b\n" + "smax v9.16b, v9.16b, v19.16b\n" + "smax v8.16b, v8.16b, v18.16b\n" + "smax v7.16b, v7.16b, v17.16b\n" + "smax v6.16b, v6.16b, v16.16b\n" "4:" // 4-vectors of channels: After loop - "ands x21, %x[n_valid_cells], #0x3\n" + "ands x2, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x20, [x24], #0x8\n" - "subs x21, x21, #0x1\n" - "ldr q19, [x20, x9]\n" - "ldr q18, [x20, x28]\n" - "ldr q17, [x20, x27]\n" - "ldr q16, [x20, x26]\n" - "smax v6.16b, v6.16b, v19.16b\n" - "smax v9.16b, v9.16b, v18.16b\n" - "smax v8.16b, v8.16b, v17.16b\n" - "smax v7.16b, v7.16b, v16.16b\n" + "ldr x1, [x5], #0x8\n" + "subs x2, x2, #0x1\n" + "ldr q19, [x1, x21]\n" + "ldr q18, [x1, x20]\n" + "ldr q17, [x1, x8]\n" + "ldr q16, [x1, x7]\n" + "smax v9.16b, v9.16b, v19.16b\n" + "smax v8.16b, v8.16b, v18.16b\n" + "smax v7.16b, v7.16b, v17.16b\n" + "smax v6.16b, v6.16b, v16.16b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" - "sxtl v23.8h, v6.8b\n" - "sxtl2 v19.8h, v6.16b\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1r { v6.4s }, [x21]\n" - "ld1r { v5.4s }, [x20]\n" - "sxtl v22.8h, v9.8b\n" - "sxtl2 v18.8h, v9.16b\n" - "sxtl v21.8h, v8.8b\n" - "sxtl2 v17.8h, v8.16b\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "ld1r { v4.4s }, [x21]\n" - "ld1r { v3.4s }, [x20]\n" - "sxtl v20.8h, v7.8b\n" - "sxtl2 v16.8h, v7.16b\n" - "neg v6.4s, v6.4s\n" - "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" - "movi v2.4s, #0x7f\n" + "add x2, %x[quant_params], %[offsetof_qp_input_offset]\n" + "sxtl v23.8h, v9.8b\n" + "sxtl v22.8h, v8.8b\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1r { v4.4s }, [x2]\n" + "ld1r { v3.4s }, [x1]\n" + "sxtl v21.8h, v7.8b\n" + "sxtl v20.8h, v6.8b\n" + "sxtl2 v27.8h, v9.16b\n" + "sxtl2 v26.8h, v8.16b\n" + "add x2, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "ld1r { v2.4s }, [x2]\n" + "ld1r { v1.4s }, [x1]\n" + "sxtl2 v25.8h, v7.16b\n" + "sxtl2 v24.8h, v6.16b\n" + "neg v4.4s, v4.4s\n" + "add x1, %x[quant_params], %[offsetof_qp_output_offset]\n" "sub %x[n_channels], %x[n_channels], #0x40\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v0.4s }, [x1]\n" "cmp %x[n_channels], #0x40\n" - "not v0.16b, v2.16b\n" - "saddw v31.4s, v6.4s, v23.4h\n" - "saddw2 v23.4s, v6.4s, v23.8h\n" - "saddw v30.4s, v6.4s, v19.4h\n" - "saddw2 v19.4s, v6.4s, v19.8h\n" - "saddw v29.4s, v6.4s, v22.4h\n" - "saddw2 v22.4s, v6.4s, v22.8h\n" - "saddw v28.4s, v6.4s, v18.4h\n" - "saddw2 v18.4s, v6.4s, v18.8h\n" - "saddw v27.4s, v6.4s, v21.4h\n" - "saddw2 v21.4s, v6.4s, v21.8h\n" - "saddw v26.4s, v6.4s, v17.4h\n" - "saddw2 v17.4s, v6.4s, v17.8h\n" - "saddw v25.4s, v6.4s, v20.4h\n" - "saddw2 v20.4s, v6.4s, v20.8h\n" - "saddw v24.4s, v6.4s, v16.4h\n" - "saddw2 v16.4s, v6.4s, v16.8h\n" - "srshl v31.4s, v31.4s, v5.4s\n" - "srshl v23.4s, v23.4s, v5.4s\n" - "srshl v30.4s, v30.4s, v5.4s\n" - "srshl v19.4s, v19.4s, v5.4s\n" - "srshl v29.4s, v29.4s, v5.4s\n" - "srshl v22.4s, v22.4s, v5.4s\n" - "srshl v28.4s, v28.4s, v5.4s\n" - "srshl v18.4s, v18.4s, v5.4s\n" - "srshl v27.4s, v27.4s, v5.4s\n" - "srshl v21.4s, v21.4s, v5.4s\n" - "srshl v26.4s, v26.4s, v5.4s\n" - "srshl v17.4s, v17.4s, v5.4s\n" - "srshl v25.4s, v25.4s, v5.4s\n" - "srshl v20.4s, v20.4s, v5.4s\n" - "srshl v24.4s, v24.4s, v5.4s\n" - "srshl v16.4s, v16.4s, v5.4s\n" - "sqrdmulh v31.4s, v31.4s, v4.4s\n" - "sqrdmulh v23.4s, v23.4s, v4.4s\n" - "sqrdmulh v30.4s, v30.4s, v4.4s\n" - "sqrdmulh v19.4s, v19.4s, v4.4s\n" - "sqrdmulh v29.4s, v29.4s, v4.4s\n" - "sqrdmulh v22.4s, v22.4s, v4.4s\n" - "sqrdmulh v28.4s, v28.4s, v4.4s\n" - "sqrdmulh v18.4s, v18.4s, v4.4s\n" - "sqrdmulh v27.4s, v27.4s, v4.4s\n" - "sqrdmulh v21.4s, v21.4s, v4.4s\n" - "sqrdmulh v26.4s, v26.4s, v4.4s\n" - "sqrdmulh v17.4s, v17.4s, v4.4s\n" - "sqrdmulh v25.4s, v25.4s, v4.4s\n" - "sqrdmulh v20.4s, v20.4s, v4.4s\n" - "sqrdmulh v24.4s, v24.4s, v4.4s\n" - "sqrdmulh v16.4s, v16.4s, v4.4s\n" + "saddw v19.4s, v4.4s, v23.4h\n" + "saddw v18.4s, v4.4s, v22.4h\n" + "saddw v17.4s, v4.4s, v21.4h\n" + "saddw v16.4s, v4.4s, v20.4h\n" + "saddw2 v31.4s, v4.4s, v23.8h\n" + "saddw v23.4s, v4.4s, v27.4h\n" + "saddw2 v30.4s, v4.4s, v22.8h\n" + "saddw v22.4s, v4.4s, v26.4h\n" + "saddw2 v29.4s, v4.4s, v21.8h\n" + "saddw v21.4s, v4.4s, v25.4h\n" + "saddw2 v28.4s, v4.4s, v20.8h\n" + "saddw v20.4s, v4.4s, v24.4h\n" + "srshl v19.4s, v19.4s, v3.4s\n" + "srshl v18.4s, v18.4s, v3.4s\n" + "srshl v17.4s, v17.4s, v3.4s\n" + "srshl v16.4s, v16.4s, v3.4s\n" + "saddw2 v27.4s, v4.4s, v27.8h\n" + "saddw2 v26.4s, v4.4s, v26.8h\n" + "saddw2 v25.4s, v4.4s, v25.8h\n" + "saddw2 v24.4s, v4.4s, v24.8h\n" + "sqrdmulh v19.4s, v19.4s, v2.4s\n" "srshl v31.4s, v31.4s, v3.4s\n" "srshl v23.4s, v23.4s, v3.4s\n" + "sqrdmulh v18.4s, v18.4s, v2.4s\n" "srshl v30.4s, v30.4s, v3.4s\n" - "srshl v19.4s, v19.4s, v3.4s\n" - "srshl v29.4s, v29.4s, v3.4s\n" "srshl v22.4s, v22.4s, v3.4s\n" + "sqrdmulh v17.4s, v17.4s, v2.4s\n" + "srshl v29.4s, v29.4s, v3.4s\n" + "srshl v21.4s, v21.4s, v3.4s\n" + "sqrdmulh v16.4s, v16.4s, v2.4s\n" "srshl v28.4s, v28.4s, v3.4s\n" - "srshl v18.4s, v18.4s, v3.4s\n" + "srshl v20.4s, v20.4s, v3.4s\n" + "srshl v19.4s, v19.4s, v1.4s\n" + "sqrdmulh v31.4s, v31.4s, v2.4s\n" + "sqrdmulh v23.4s, v23.4s, v2.4s\n" "srshl v27.4s, v27.4s, v3.4s\n" - "srshl v21.4s, v21.4s, v3.4s\n" + "srshl v18.4s, v18.4s, v1.4s\n" + "sqrdmulh v30.4s, v30.4s, v2.4s\n" + "sqrdmulh v22.4s, v22.4s, v2.4s\n" "srshl v26.4s, v26.4s, v3.4s\n" - "srshl v17.4s, v17.4s, v3.4s\n" + "srshl v17.4s, v17.4s, v1.4s\n" + "sqrdmulh v29.4s, v29.4s, v2.4s\n" + "sqrdmulh v21.4s, v21.4s, v2.4s\n" "srshl v25.4s, v25.4s, v3.4s\n" - "srshl v20.4s, v20.4s, v3.4s\n" + "srshl v16.4s, v16.4s, v1.4s\n" + "sqrdmulh v28.4s, v28.4s, v2.4s\n" + "sqrdmulh v20.4s, v20.4s, v2.4s\n" "srshl v24.4s, v24.4s, v3.4s\n" - "srshl v16.4s, v16.4s, v3.4s\n" - "add v31.4s, v31.4s, v1.4s\n" - "add v23.4s, v23.4s, v1.4s\n" - "add v30.4s, v30.4s, v1.4s\n" - "add v19.4s, v19.4s, v1.4s\n" - "add v29.4s, v29.4s, v1.4s\n" - "add v22.4s, v22.4s, v1.4s\n" - "add v28.4s, v28.4s, v1.4s\n" - "add v18.4s, v18.4s, v1.4s\n" - "add v27.4s, v27.4s, v1.4s\n" - "add v21.4s, v21.4s, v1.4s\n" - "add v26.4s, v26.4s, v1.4s\n" - "add v17.4s, v17.4s, v1.4s\n" - "add v25.4s, v25.4s, v1.4s\n" - "add v20.4s, v20.4s, v1.4s\n" - "add v24.4s, v24.4s, v1.4s\n" - "add v16.4s, v16.4s, v1.4s\n" - "smax v31.4s, v31.4s, v0.4s\n" - "smax v23.4s, v23.4s, v0.4s\n" - "smax v30.4s, v30.4s, v0.4s\n" - "smax v19.4s, v19.4s, v0.4s\n" - "smax v29.4s, v29.4s, v0.4s\n" - "smax v22.4s, v22.4s, v0.4s\n" - "smax v28.4s, v28.4s, v0.4s\n" - "smax v18.4s, v18.4s, v0.4s\n" - "smax v27.4s, v27.4s, v0.4s\n" - "smax v21.4s, v21.4s, v0.4s\n" - "smax v26.4s, v26.4s, v0.4s\n" - "smax v17.4s, v17.4s, v0.4s\n" - "smax v25.4s, v25.4s, v0.4s\n" - "smax v20.4s, v20.4s, v0.4s\n" - "smax v24.4s, v24.4s, v0.4s\n" - "smax v16.4s, v16.4s, v0.4s\n" - "smin v31.4s, v31.4s, v2.4s\n" - "smin v23.4s, v23.4s, v2.4s\n" - "smin v30.4s, v30.4s, v2.4s\n" - "smin v19.4s, v19.4s, v2.4s\n" - "smin v29.4s, v29.4s, v2.4s\n" - "smin v22.4s, v22.4s, v2.4s\n" - "smin v28.4s, v28.4s, v2.4s\n" - "smin v18.4s, v18.4s, v2.4s\n" - "smin v27.4s, v27.4s, v2.4s\n" - "smin v21.4s, v21.4s, v2.4s\n" - "smin v26.4s, v26.4s, v2.4s\n" - "smin v17.4s, v17.4s, v2.4s\n" - "smin v25.4s, v25.4s, v2.4s\n" - "smin v20.4s, v20.4s, v2.4s\n" - "smin v24.4s, v24.4s, v2.4s\n" - "smin v16.4s, v16.4s, v2.4s\n" - "uzp1 v23.16b, v31.16b, v23.16b\n" - "uzp1 v19.16b, v30.16b, v19.16b\n" - "uzp1 v22.16b, v29.16b, v22.16b\n" - "uzp1 v18.16b, v28.16b, v18.16b\n" - "uzp1 v21.16b, v27.16b, v21.16b\n" - "uzp1 v17.16b, v26.16b, v17.16b\n" - "uzp1 v20.16b, v25.16b, v20.16b\n" - "uzp1 v16.16b, v24.16b, v16.16b\n" - "uzp1 v19.16b, v23.16b, v19.16b\n" - "uzp1 v18.16b, v22.16b, v18.16b\n" - "uzp1 v17.16b, v21.16b, v17.16b\n" - "uzp1 v16.16b, v20.16b, v16.16b\n" - "str q19, [%x[outptr], x9]\n" - "add x9, x9, #0x40\n" - "str q18, [%x[outptr], x28]\n" - "add x28, x28, #0x40\n" - "str q17, [%x[outptr], x27]\n" - "add x27, x27, #0x40\n" - "str q16, [%x[outptr], x26]\n" - "add x26, x26, #0x40\n" + "srshl v31.4s, v31.4s, v1.4s\n" + "srshl v23.4s, v23.4s, v1.4s\n" + "sqrdmulh v27.4s, v27.4s, v2.4s\n" + "srshl v30.4s, v30.4s, v1.4s\n" + "srshl v22.4s, v22.4s, v1.4s\n" + "sqrdmulh v26.4s, v26.4s, v2.4s\n" + "srshl v29.4s, v29.4s, v1.4s\n" + "srshl v21.4s, v21.4s, v1.4s\n" + "sqrdmulh v25.4s, v25.4s, v2.4s\n" + "srshl v28.4s, v28.4s, v1.4s\n" + "srshl v20.4s, v20.4s, v1.4s\n" + "sqrdmulh v24.4s, v24.4s, v2.4s\n" + "add v19.4s, v19.4s, v0.4s\n" + "add v18.4s, v18.4s, v0.4s\n" + "add v17.4s, v17.4s, v0.4s\n" + "add v16.4s, v16.4s, v0.4s\n" + "srshl v27.4s, v27.4s, v1.4s\n" + "srshl v26.4s, v26.4s, v1.4s\n" + "srshl v25.4s, v25.4s, v1.4s\n" + "srshl v24.4s, v24.4s, v1.4s\n" + "add v31.4s, v31.4s, v0.4s\n" + "add v23.4s, v23.4s, v0.4s\n" + "add v30.4s, v30.4s, v0.4s\n" + "add v22.4s, v22.4s, v0.4s\n" + "add v29.4s, v29.4s, v0.4s\n" + "add v21.4s, v21.4s, v0.4s\n" + "add v28.4s, v28.4s, v0.4s\n" + "add v20.4s, v20.4s, v0.4s\n" + "sqxtn v19.4h, v19.4s\n" + "sqxtn v18.4h, v18.4s\n" + "sqxtn v17.4h, v17.4s\n" + "sqxtn v16.4h, v16.4s\n" + "add v27.4s, v27.4s, v0.4s\n" + "add v26.4s, v26.4s, v0.4s\n" + "add v25.4s, v25.4s, v0.4s\n" + "add v24.4s, v24.4s, v0.4s\n" + "sqxtn2 v19.8h, v31.4s\n" + "sqxtn v23.4h, v23.4s\n" + "sqxtn2 v18.8h, v30.4s\n" + "sqxtn v22.4h, v22.4s\n" + "sqxtn2 v17.8h, v29.4s\n" + "sqxtn v21.4h, v21.4s\n" + "sqxtn2 v16.8h, v28.4s\n" + "sqxtn v20.4h, v20.4s\n" + "sqxtn2 v23.8h, v27.4s\n" + "sqxtn2 v22.8h, v26.4s\n" + "sqxtn2 v21.8h, v25.4s\n" + "sqxtn v19.8b, v19.8h\n" + "sqxtn2 v20.8h, v24.4s\n" + "sqxtn v18.8b, v18.8h\n" + "sqxtn v17.8b, v17.8h\n" + "sqxtn v16.8b, v16.8h\n" + "sqxtn2 v19.16b, v23.8h\n" + "sqxtn2 v18.16b, v22.8h\n" + "sqxtn2 v17.16b, v21.8h\n" + "sqxtn2 v16.16b, v20.8h\n" + "str q19, [%x[outptr], x21]\n" + "add x21, x21, #0x40\n" + "str q18, [%x[outptr], x20]\n" + "add x20, x20, #0x40\n" + "str q17, [%x[outptr], x8]\n" + "add x8, x8, #0x40\n" + "str q16, [%x[outptr], x7]\n" + "add x7, x7, #0x40\n" "bge 1b\n" "cbz %x[n_channels], 43f\n" "7:" // Single vector of channels "cmp %x[n_channels], #0x10\n" "blt 14f\n" "8:" // Single vector of channels: Loop - "lsr x25, %x[n_valid_cells], #0x2\n" - "movi v6.16b, #0x80\n" - "mov x24, %x[inptrs]\n" - "cbz x25, 11f\n" - "ldp x23, x22, [x24, #0]\n" - "ldp x21, x20, [x24, #0x10]\n" - "subs x25, x25, #0x1\n" - "add x24, x24, #0x20\n" - "ldr q5, [x23, x9]\n" - "ldr q4, [x22, x9]\n" - "ldr q3, [x21, x9]\n" - "ldr q2, [x20, x9]\n" + "lsr x6, %x[n_valid_cells], #0x2\n" + "movi v9.16b, #0x80\n" + "mov x5, %x[inptrs]\n" + "cbz x6, 11f\n" + "ldp x4, x3, [x5, #0]\n" + "ldp x2, x1, [x5, #0x10]\n" + "subs x6, x6, #0x1\n" + "add x5, x5, #0x20\n" + "ldr q5, [x4, x21]\n" + "ldr q4, [x3, x21]\n" + "ldr q3, [x2, x21]\n" + "ldr q2, [x1, x21]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop "smax v17.16b, v5.16b, v4.16b\n" "smax v16.16b, v3.16b, v2.16b\n" - "ldp x23, x22, [x24, #0]\n" - "ldp x21, x20, [x24, #0x10]\n" - "subs x25, x25, #0x1\n" - "add x24, x24, #0x20\n" - "ldr q5, [x23, x9]\n" - "ldr q4, [x22, x9]\n" - "ldr q3, [x21, x9]\n" - "ldr q2, [x20, x9]\n" + "ldp x4, x3, [x5, #0]\n" + "ldp x2, x1, [x5, #0x10]\n" + "subs x6, x6, #0x1\n" + "add x5, x5, #0x20\n" + "ldr q5, [x4, x21]\n" + "ldr q4, [x3, x21]\n" + "ldr q3, [x2, x21]\n" + "ldr q2, [x1, x21]\n" "smax v16.16b, v17.16b, v16.16b\n" - "smax v6.16b, v6.16b, v16.16b\n" + "smax v9.16b, v9.16b, v16.16b\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail "smax v17.16b, v5.16b, v4.16b\n" "smax v16.16b, v3.16b, v2.16b\n" "smax v16.16b, v17.16b, v16.16b\n" - "smax v6.16b, v6.16b, v16.16b\n" + "smax v9.16b, v9.16b, v16.16b\n" "11:" // Single vector of channels: Loop: After loop - "ands x21, %x[n_valid_cells], #0x3\n" + "ands x2, %x[n_valid_cells], #0x3\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop - "ldr x20, [x24], #0x8\n" - "subs x21, x21, #0x1\n" - "ldr q16, [x20, x9]\n" - "smax v6.16b, v6.16b, v16.16b\n" + "ldr x1, [x5], #0x8\n" + "subs x2, x2, #0x1\n" + "ldr q16, [x1, x21]\n" + "smax v9.16b, v9.16b, v16.16b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" - "sxtl v17.8h, v6.8b\n" - "sxtl2 v26.8h, v6.16b\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1r { v16.4s }, [x21]\n" - "ld1r { v25.4s }, [x20]\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "ld1r { v24.4s }, [x21]\n" - "ld1r { v23.4s }, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" - "movi v22.4s, #0x7f\n" - "ld1r { v21.4s }, [x20]\n" + "add x2, %x[quant_params], %[offsetof_qp_input_offset]\n" + "sxtl v17.8h, v9.8b\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "sxtl2 v24.8h, v9.16b\n" + "ld1r { v18.4s }, [x2]\n" + "ld1r { v23.4s }, [x1]\n" + "add x2, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "ld1r { v22.4s }, [x2]\n" + "ld1r { v21.4s }, [x1]\n" + "add x1, %x[quant_params], %[offsetof_qp_output_offset]\n" "sub %x[n_channels], %x[n_channels], #0x10\n" - "neg v16.4s, v16.4s\n" - "not v20.16b, v22.16b\n" + "ld1r { v20.4s }, [x1]\n" "cmp %x[n_channels], #0x10\n" - "saddw v19.4s, v16.4s, v17.4h\n" - "saddw2 v17.4s, v16.4s, v17.8h\n" - "saddw v18.4s, v16.4s, v26.4h\n" - "saddw2 v16.4s, v16.4s, v26.8h\n" - "srshl v19.4s, v19.4s, v25.4s\n" - "srshl v17.4s, v17.4s, v25.4s\n" - "srshl v18.4s, v18.4s, v25.4s\n" - "srshl v16.4s, v16.4s, v25.4s\n" - "sqrdmulh v19.4s, v19.4s, v24.4s\n" - "sqrdmulh v17.4s, v17.4s, v24.4s\n" - "sqrdmulh v18.4s, v18.4s, v24.4s\n" - "sqrdmulh v16.4s, v16.4s, v24.4s\n" + "neg v18.4s, v18.4s\n" + "saddw v16.4s, v18.4s, v17.4h\n" + "saddw2 v19.4s, v18.4s, v17.8h\n" + "saddw v17.4s, v18.4s, v24.4h\n" + "srshl v16.4s, v16.4s, v23.4s\n" + "saddw2 v18.4s, v18.4s, v24.8h\n" "srshl v19.4s, v19.4s, v23.4s\n" "srshl v17.4s, v17.4s, v23.4s\n" + "sqrdmulh v16.4s, v16.4s, v22.4s\n" "srshl v18.4s, v18.4s, v23.4s\n" - "srshl v16.4s, v16.4s, v23.4s\n" - "add v19.4s, v19.4s, v21.4s\n" - "add v17.4s, v17.4s, v21.4s\n" - "add v18.4s, v18.4s, v21.4s\n" - "add v16.4s, v16.4s, v21.4s\n" - "smax v19.4s, v19.4s, v20.4s\n" - "smax v17.4s, v17.4s, v20.4s\n" - "smax v18.4s, v18.4s, v20.4s\n" - "smax v16.4s, v16.4s, v20.4s\n" - "smin v19.4s, v19.4s, v22.4s\n" - "smin v17.4s, v17.4s, v22.4s\n" - "smin v18.4s, v18.4s, v22.4s\n" - "smin v16.4s, v16.4s, v22.4s\n" - "uzp1 v17.16b, v19.16b, v17.16b\n" - "uzp1 v16.16b, v18.16b, v16.16b\n" - "uzp1 v16.16b, v17.16b, v16.16b\n" - "str q16, [%x[outptr], x9]\n" - "add x9, x9, #0x10\n" + "sqrdmulh v19.4s, v19.4s, v22.4s\n" + "sqrdmulh v17.4s, v17.4s, v22.4s\n" + "srshl v16.4s, v16.4s, v21.4s\n" + "sqrdmulh v18.4s, v18.4s, v22.4s\n" + "srshl v19.4s, v19.4s, v21.4s\n" + "srshl v17.4s, v17.4s, v21.4s\n" + "add v16.4s, v16.4s, v20.4s\n" + "srshl v18.4s, v18.4s, v21.4s\n" + "add v19.4s, v19.4s, v20.4s\n" + "add v17.4s, v17.4s, v20.4s\n" + "sqxtn v16.4h, v16.4s\n" + "add v18.4s, v18.4s, v20.4s\n" + "sqxtn v17.4h, v17.4s\n" + "sqxtn2 v16.8h, v19.4s\n" + "sqxtn2 v17.8h, v18.4s\n" + "sqxtn v16.8b, v16.8h\n" + "sqxtn2 v16.16b, v17.8h\n" + "str q16, [%x[outptr], x21]\n" + "add x21, x21, #0x10\n" "bge 8b\n" "cbz %x[n_channels], 43f\n" "14:" // Oddments - "lsr x25, %x[n_valid_cells], #0x2\n" - "add %x[outptr], %x[outptr], x9\n" - "movi v6.16b, #0x80\n" - "mov x24, %x[inptrs]\n" - "cbz x25, 24f\n" + "lsr x6, %x[n_valid_cells], #0x2\n" + "add %x[outptr], %x[outptr], x21\n" + "movi v9.16b, #0x80\n" + "mov x5, %x[inptrs]\n" + "cbz x6, 24f\n" "15:" // Oddments: 4 inputs loop - "ldp x23, x22, [x24, #0]\n" - "ldp x21, x20, [x24, #0x10]\n" - "add x24, x24, #0x20\n" + "ldp x4, x3, [x5, #0]\n" + "ldp x2, x1, [x5, #0x10]\n" + "add x5, x5, #0x20\n" "movi v5.16b, #0\n" "movi v4.16b, #0\n" "movi v3.16b, #0\n" "movi v2.16b, #0\n" - "add x23, x23, x9\n" - "add x22, x22, x9\n" - "add x21, x21, x9\n" - "add x20, x20, x9\n" + "add x4, x4, x21\n" + "add x3, x3, x21\n" + "add x2, x2, x21\n" + "add x1, x1, x21\n" "tbz %x[n_channels], #3, 19f\n" - "ldr d5, [x23], #0x8\n" - "ldr d4, [x22], #0x8\n" - "ldr d3, [x21], #0x8\n" - "ldr d2, [x20], #0x8\n" + "ldr d5, [x4], #0x8\n" + "ldr d4, [x3], #0x8\n" + "ldr d3, [x2], #0x8\n" + "ldr d2, [x1], #0x8\n" "tbz %x[n_channels], #2, 17f\n" - "ld1 { v5.s }[2], [x23], #0x4\n" - "ld1 { v4.s }[2], [x22], #0x4\n" - "ld1 { v3.s }[2], [x21], #0x4\n" - "ld1 { v2.s }[2], [x20], #0x4\n" + "ld1 { v5.s }[2], [x4], #0x4\n" + "ld1 { v4.s }[2], [x3], #0x4\n" + "ld1 { v3.s }[2], [x2], #0x4\n" + "ld1 { v2.s }[2], [x1], #0x4\n" "tbz %x[n_channels], #1, 16f\n" - "ld1 { v5.h }[6], [x23], #0x2\n" - "ld1 { v4.h }[6], [x22], #0x2\n" - "ld1 { v3.h }[6], [x21], #0x2\n" - "ld1 { v2.h }[6], [x20], #0x2\n" + "ld1 { v5.h }[6], [x4], #0x2\n" + "ld1 { v4.h }[6], [x3], #0x2\n" + "ld1 { v3.h }[6], [x2], #0x2\n" + "ld1 { v2.h }[6], [x1], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v5.b }[14], [x23], #0x1\n" - "ld1 { v4.b }[14], [x22], #0x1\n" - "ld1 { v3.b }[14], [x21], #0x1\n" - "ld1 { v2.b }[14], [x20], #0x1\n" + "ld1 { v5.b }[14], [x4], #0x1\n" + "ld1 { v4.b }[14], [x3], #0x1\n" + "ld1 { v3.b }[14], [x2], #0x1\n" + "ld1 { v2.b }[14], [x1], #0x1\n" "b 23f\n" "16:" // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 23f\n" - "ld1 { v5.b }[12], [x23], #0x1\n" - "ld1 { v4.b }[12], [x22], #0x1\n" - "ld1 { v3.b }[12], [x21], #0x1\n" - "ld1 { v2.b }[12], [x20], #0x1\n" + "ld1 { v5.b }[12], [x4], #0x1\n" + "ld1 { v4.b }[12], [x3], #0x1\n" + "ld1 { v3.b }[12], [x2], #0x1\n" + "ld1 { v2.b }[12], [x1], #0x1\n" "b 23f\n" "17:" // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Unset "tbz %x[n_channels], #1, 18f\n" - "ld1 { v5.h }[4], [x23], #0x2\n" - "ld1 { v4.h }[4], [x22], #0x2\n" - "ld1 { v3.h }[4], [x21], #0x2\n" - "ld1 { v2.h }[4], [x20], #0x2\n" + "ld1 { v5.h }[4], [x4], #0x2\n" + "ld1 { v4.h }[4], [x3], #0x2\n" + "ld1 { v3.h }[4], [x2], #0x2\n" + "ld1 { v2.h }[4], [x1], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v5.b }[10], [x23], #0x1\n" - "ld1 { v4.b }[10], [x22], #0x1\n" - "ld1 { v3.b }[10], [x21], #0x1\n" - "ld1 { v2.b }[10], [x20], #0x1\n" + "ld1 { v5.b }[10], [x4], #0x1\n" + "ld1 { v4.b }[10], [x3], #0x1\n" + "ld1 { v3.b }[10], [x2], #0x1\n" + "ld1 { v2.b }[10], [x1], #0x1\n" "b 23f\n" "18:" // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Unset: Bit 1: Unset "tbz %x[n_channels], #0, 23f\n" - "ld1 { v5.b }[8], [x23], #0x1\n" - "ld1 { v4.b }[8], [x22], #0x1\n" - "ld1 { v3.b }[8], [x21], #0x1\n" - "ld1 { v2.b }[8], [x20], #0x1\n" + "ld1 { v5.b }[8], [x4], #0x1\n" + "ld1 { v4.b }[8], [x3], #0x1\n" + "ld1 { v3.b }[8], [x2], #0x1\n" + "ld1 { v2.b }[8], [x1], #0x1\n" "b 23f\n" "19:" // Oddments: 4 inputs loop: Load: Bit 3: Unset "tbz %x[n_channels], #2, 21f\n" - "ldr s5, [x23], #0x4\n" - "ldr s4, [x22], #0x4\n" - "ldr s3, [x21], #0x4\n" - "ldr s2, [x20], #0x4\n" + "ldr s5, [x4], #0x4\n" + "ldr s4, [x3], #0x4\n" + "ldr s3, [x2], #0x4\n" + "ldr s2, [x1], #0x4\n" "tbz %x[n_channels], #1, 20f\n" - "ld1 { v5.h }[2], [x23], #0x2\n" - "ld1 { v4.h }[2], [x22], #0x2\n" - "ld1 { v3.h }[2], [x21], #0x2\n" - "ld1 { v2.h }[2], [x20], #0x2\n" + "ld1 { v5.h }[2], [x4], #0x2\n" + "ld1 { v4.h }[2], [x3], #0x2\n" + "ld1 { v3.h }[2], [x2], #0x2\n" + "ld1 { v2.h }[2], [x1], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v5.b }[6], [x23], #0x1\n" - "ld1 { v4.b }[6], [x22], #0x1\n" - "ld1 { v3.b }[6], [x21], #0x1\n" - "ld1 { v2.b }[6], [x20], #0x1\n" + "ld1 { v5.b }[6], [x4], #0x1\n" + "ld1 { v4.b }[6], [x3], #0x1\n" + "ld1 { v3.b }[6], [x2], #0x1\n" + "ld1 { v2.b }[6], [x1], #0x1\n" "b 23f\n" "20:" // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 23f\n" - "ld1 { v5.b }[4], [x23], #0x1\n" - "ld1 { v4.b }[4], [x22], #0x1\n" - "ld1 { v3.b }[4], [x21], #0x1\n" - "ld1 { v2.b }[4], [x20], #0x1\n" + "ld1 { v5.b }[4], [x4], #0x1\n" + "ld1 { v4.b }[4], [x3], #0x1\n" + "ld1 { v3.b }[4], [x2], #0x1\n" + "ld1 { v2.b }[4], [x1], #0x1\n" "b 23f\n" "21:" // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Unset "tbz %x[n_channels], #1, 22f\n" - "ldr h5, [x23], #0x2\n" - "ldr h4, [x22], #0x2\n" - "ldr h3, [x21], #0x2\n" - "ldr h2, [x20], #0x2\n" + "ldr h5, [x4], #0x2\n" + "ldr h4, [x3], #0x2\n" + "ldr h3, [x2], #0x2\n" + "ldr h2, [x1], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v5.b }[2], [x23], #0x1\n" - "ld1 { v4.b }[2], [x22], #0x1\n" - "ld1 { v3.b }[2], [x21], #0x1\n" - "ld1 { v2.b }[2], [x20], #0x1\n" + "ld1 { v5.b }[2], [x4], #0x1\n" + "ld1 { v4.b }[2], [x3], #0x1\n" + "ld1 { v3.b }[2], [x2], #0x1\n" + "ld1 { v2.b }[2], [x1], #0x1\n" "b 23f\n" "22:" // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset "tbz %x[n_channels], #0, 23f\n" - "ldr b5, [x23], #0x1\n" - "ldr b4, [x22], #0x1\n" - "ldr b3, [x21], #0x1\n" - "ldr b2, [x20], #0x1\n" + "ldr b5, [x4], #0x1\n" + "ldr b4, [x3], #0x1\n" + "ldr b3, [x2], #0x1\n" + "ldr b2, [x1], #0x1\n" "23:" // Oddments: 4 inputs loop: Load: Bit 3: End "smax v17.16b, v5.16b, v4.16b\n" "smax v16.16b, v3.16b, v2.16b\n" - "subs x25, x25, #0x1\n" + "subs x6, x6, #0x1\n" "smax v16.16b, v17.16b, v16.16b\n" - "smax v6.16b, v6.16b, v16.16b\n" + "smax v9.16b, v9.16b, v16.16b\n" "bgt 15b\n" "24:" // Oddments: After loop - "ands x21, %x[n_valid_cells], #0x3\n" + "ands x2, %x[n_valid_cells], #0x3\n" "beq 34f\n" "25:" // Oddments: Single input loop - "ldr x23, [x24], #0x8\n" + "ldr x4, [x5], #0x8\n" "movi v5.16b, #0\n" - "add x23, x23, x9\n" + "add x4, x4, x21\n" "tbz %x[n_channels], #3, 29f\n" - "ldr d5, [x23], #0x8\n" + "ldr d5, [x4], #0x8\n" "tbz %x[n_channels], #2, 27f\n" - "ld1 { v5.s }[2], [x23], #0x4\n" + "ld1 { v5.s }[2], [x4], #0x4\n" "tbz %x[n_channels], #1, 26f\n" - "ld1 { v5.h }[6], [x23], #0x2\n" + "ld1 { v5.h }[6], [x4], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v5.b }[14], [x23], #0x1\n" + "ld1 { v5.b }[14], [x4], #0x1\n" "b 33f\n" "26:" // Oddments: Single input loop: Load: Bit 3: Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 33f\n" - "ld1 { v5.b }[12], [x23], #0x1\n" + "ld1 { v5.b }[12], [x4], #0x1\n" "b 33f\n" "27:" // Oddments: Single input loop: Load: Bit 3: Bit 2: Unset "tbz %x[n_channels], #1, 28f\n" - "ld1 { v5.h }[4], [x23], #0x2\n" + "ld1 { v5.h }[4], [x4], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v5.b }[10], [x23], #0x1\n" + "ld1 { v5.b }[10], [x4], #0x1\n" "b 33f\n" "28:" // Oddments: Single input loop: Load: Bit 3: Bit 2: Unset: Bit 1: Unset "tbz %x[n_channels], #0, 33f\n" - "ld1 { v5.b }[8], [x23], #0x1\n" + "ld1 { v5.b }[8], [x4], #0x1\n" "b 33f\n" "29:" // Oddments: Single input loop: Load: Bit 3: Unset "tbz %x[n_channels], #2, 31f\n" - "ldr s5, [x23], #0x4\n" + "ldr s5, [x4], #0x4\n" "tbz %x[n_channels], #1, 30f\n" - "ld1 { v5.h }[2], [x23], #0x2\n" + "ld1 { v5.h }[2], [x4], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v5.b }[6], [x23], #0x1\n" + "ld1 { v5.b }[6], [x4], #0x1\n" "b 33f\n" "30:" // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 33f\n" - "ld1 { v5.b }[4], [x23], #0x1\n" + "ld1 { v5.b }[4], [x4], #0x1\n" "b 33f\n" "31:" // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Unset "tbz %x[n_channels], #1, 32f\n" - "ldr h5, [x23], #0x2\n" + "ldr h5, [x4], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v5.b }[2], [x23], #0x1\n" + "ld1 { v5.b }[2], [x4], #0x1\n" "b 33f\n" "32:" // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset "tbz %x[n_channels], #0, 33f\n" - "ldr b5, [x23], #0x1\n" + "ldr b5, [x4], #0x1\n" "33:" // Oddments: Single input loop: Load: Bit 3: End - "subs x21, x21, #0x1\n" - "smax v6.16b, v6.16b, v5.16b\n" + "subs x2, x2, #0x1\n" + "smax v9.16b, v9.16b, v5.16b\n" "bgt 25b\n" "34:" // Oddments: Single input loop: End - "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" - "sxtl v17.8h, v6.8b\n" - "sxtl2 v26.8h, v6.16b\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1r { v16.4s }, [x21]\n" - "ld1r { v25.4s }, [x20]\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "ld1r { v24.4s }, [x21]\n" - "ld1r { v23.4s }, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" - "movi v22.4s, #0x7f\n" - "ld1r { v21.4s }, [x20]\n" - "neg v16.4s, v16.4s\n" - "not v20.16b, v22.16b\n" - "saddw v19.4s, v16.4s, v17.4h\n" - "saddw2 v17.4s, v16.4s, v17.8h\n" - "saddw v18.4s, v16.4s, v26.4h\n" - "saddw2 v16.4s, v16.4s, v26.8h\n" - "srshl v19.4s, v19.4s, v25.4s\n" - "srshl v17.4s, v17.4s, v25.4s\n" - "srshl v18.4s, v18.4s, v25.4s\n" - "srshl v16.4s, v16.4s, v25.4s\n" - "sqrdmulh v19.4s, v19.4s, v24.4s\n" - "sqrdmulh v17.4s, v17.4s, v24.4s\n" - "sqrdmulh v18.4s, v18.4s, v24.4s\n" - "sqrdmulh v16.4s, v16.4s, v24.4s\n" + "add x2, %x[quant_params], %[offsetof_qp_input_offset]\n" + "sxtl v17.8h, v9.8b\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "sxtl2 v24.8h, v9.16b\n" + "ld1r { v18.4s }, [x2]\n" + "ld1r { v23.4s }, [x1]\n" + "add x2, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "ld1r { v22.4s }, [x2]\n" + "ld1r { v21.4s }, [x1]\n" + "add x1, %x[quant_params], %[offsetof_qp_output_offset]\n" + "ld1r { v20.4s }, [x1]\n" + "neg v18.4s, v18.4s\n" + "saddw v16.4s, v18.4s, v17.4h\n" + "saddw2 v19.4s, v18.4s, v17.8h\n" + "saddw v17.4s, v18.4s, v24.4h\n" + "srshl v16.4s, v16.4s, v23.4s\n" + "saddw2 v18.4s, v18.4s, v24.8h\n" "srshl v19.4s, v19.4s, v23.4s\n" "srshl v17.4s, v17.4s, v23.4s\n" + "sqrdmulh v16.4s, v16.4s, v22.4s\n" "srshl v18.4s, v18.4s, v23.4s\n" - "srshl v16.4s, v16.4s, v23.4s\n" - "add v19.4s, v19.4s, v21.4s\n" - "add v17.4s, v17.4s, v21.4s\n" - "add v18.4s, v18.4s, v21.4s\n" - "add v16.4s, v16.4s, v21.4s\n" - "smax v19.4s, v19.4s, v20.4s\n" - "smax v17.4s, v17.4s, v20.4s\n" - "smax v18.4s, v18.4s, v20.4s\n" - "smax v16.4s, v16.4s, v20.4s\n" - "smin v19.4s, v19.4s, v22.4s\n" - "smin v17.4s, v17.4s, v22.4s\n" - "smin v18.4s, v18.4s, v22.4s\n" - "smin v16.4s, v16.4s, v22.4s\n" - "uzp1 v17.16b, v19.16b, v17.16b\n" - "uzp1 v16.16b, v18.16b, v16.16b\n" - "uzp1 v16.16b, v17.16b, v16.16b\n" + "sqrdmulh v19.4s, v19.4s, v22.4s\n" + "sqrdmulh v17.4s, v17.4s, v22.4s\n" + "srshl v16.4s, v16.4s, v21.4s\n" + "sqrdmulh v18.4s, v18.4s, v22.4s\n" + "srshl v19.4s, v19.4s, v21.4s\n" + "srshl v17.4s, v17.4s, v21.4s\n" + "add v16.4s, v16.4s, v20.4s\n" + "srshl v18.4s, v18.4s, v21.4s\n" + "add v19.4s, v19.4s, v20.4s\n" + "add v17.4s, v17.4s, v20.4s\n" + "sqxtn v16.4h, v16.4s\n" + "add v18.4s, v18.4s, v20.4s\n" + "sqxtn v17.4h, v17.4s\n" + "sqxtn2 v16.8h, v19.4s\n" + "sqxtn2 v17.8h, v18.4s\n" + "sqxtn v16.8b, v16.8h\n" + "sqxtn2 v16.16b, v17.8h\n" "tbz %x[n_channels], #3, 38f\n" "st1 { v16.d }[0], [%x[outptr]], #0x8\n" "tbz %x[n_channels], #2, 36f\n" @@ -667,7 +631,7 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "43:" // End : [n_channels] "+&r" (n_channels), [outptr] "+&r" (outptr) : [inptrs] "r" (inptrs), [n_valid_cells] "r" (n_valid_cells), [offsetof_qp_input_offset] "I" (offsetof(Requantize32, input_offset)), [offsetof_qp_output_offset] "I" (offsetof(Requantize32, output_offset)), [offsetof_qp_per_layer_left_shift] "I" (offsetof(Requantize32, per_layer_left_shift)), [offsetof_qp_per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [offsetof_qp_per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [quant_params] "r" (&qp) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "x20", "x21" ); } diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp index 1c91baf41b4..d8e2fef111e 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp @@ -43,82 +43,82 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( { __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" - "mov x9, #0\n" - "cntb x28\n" - "cntb x27, ALL, MUL #2\n" - "cntb x26, ALL, MUL #3\n" - "whilelt p4.b, x9, %x[n_channels]\n" - "whilelt p3.b, x28, %x[n_channels]\n" - "whilelt p2.b, x27, %x[n_channels]\n" - "whilelt p1.b, x26, %x[n_channels]\n" + "mov x21, #0\n" + "cntb x20\n" + "cntb x8, ALL, MUL #2\n" + "cntb x7, ALL, MUL #3\n" + "whilelt p4.b, x21, %x[n_channels]\n" + "whilelt p3.b, x20, %x[n_channels]\n" + "whilelt p2.b, x8, %x[n_channels]\n" + "whilelt p1.b, x7, %x[n_channels]\n" "ptrue p0.b\n" "b.eq 7f\n" "1:" // 4-vectors of channels - "lsr x25, %x[n_valid_cells], #0x2\n" - "mov z5.b, #0x80\n" - "mov z4.b, #0x80\n" - "mov x24, %x[inptrs]\n" + "lsr x6, %x[n_valid_cells], #0x2\n" + "mov z6.b, #0x80\n" "mov z3.b, #0x80\n" + "mov x5, %x[inptrs]\n" + "mov z5.b, #0x80\n" "mov z2.b, #0x80\n" - "cbz x25, 4f\n" - "ldp x23, x22, [x24, #0]\n" - "subs x25, x25, #0x1\n" - "ldp x21, x20, [x24, #0x10]\n" - "add x24, x24, #0x20\n" - "ld1b { z1.b }, p4/Z, [x23, x9]\n" - "ld1b { z0.b }, p4/Z, [x22, x9]\n" - "ld1b { z23.b }, p4/Z, [x21, x9]\n" - "ld1b { z31.b }, p4/Z, [x20, x9]\n" - "ld1b { z30.b }, p3/Z, [x23, x28]\n" - "ld1b { z29.b }, p3/Z, [x22, x28]\n" - "ld1b { z22.b }, p3/Z, [x21, x28]\n" - "ld1b { z28.b }, p3/Z, [x20, x28]\n" - "ld1b { z17.b }, p2/Z, [x23, x27]\n" - "ld1b { z27.b }, p2/Z, [x22, x27]\n" - "ld1b { z21.b }, p2/Z, [x21, x27]\n" - "ld1b { z26.b }, p2/Z, [x20, x27]\n" - "ld1b { z16.b }, p1/Z, [x23, x26]\n" - "ld1b { z25.b }, p1/Z, [x22, x26]\n" - "ld1b { z20.b }, p1/Z, [x21, x26]\n" - "ld1b { z24.b }, p1/Z, [x20, x26]\n" + "cbz x6, 4f\n" + "ldp x4, x3, [x5, #0]\n" + "subs x6, x6, #0x1\n" + "ldp x2, x1, [x5, #0x10]\n" + "add x5, x5, #0x20\n" + "ld1b { z1.b }, p4/Z, [x4, x21]\n" + "ld1b { z0.b }, p4/Z, [x3, x21]\n" + "ld1b { z23.b }, p4/Z, [x2, x21]\n" + "ld1b { z31.b }, p4/Z, [x1, x21]\n" + "ld1b { z30.b }, p3/Z, [x4, x20]\n" + "ld1b { z29.b }, p3/Z, [x3, x20]\n" + "ld1b { z22.b }, p3/Z, [x2, x20]\n" + "ld1b { z28.b }, p3/Z, [x1, x20]\n" + "ld1b { z17.b }, p2/Z, [x4, x8]\n" + "ld1b { z27.b }, p2/Z, [x3, x8]\n" + "ld1b { z21.b }, p2/Z, [x2, x8]\n" + "ld1b { z26.b }, p2/Z, [x1, x8]\n" + "ld1b { z16.b }, p1/Z, [x4, x7]\n" + "ld1b { z25.b }, p1/Z, [x3, x7]\n" + "ld1b { z20.b }, p1/Z, [x2, x7]\n" + "ld1b { z24.b }, p1/Z, [x1, x7]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop "movprfx z19, z1\n smax z19.b, p0/M, z19.b, z0.b\n" "smax z23.b, p0/M, z23.b, z31.b\n" - "ldp x23, x22, [x24, #0]\n" - "subs x25, x25, #0x1\n" + "ldp x4, x3, [x5, #0]\n" + "subs x6, x6, #0x1\n" "movprfx z18, z30\n smax z18.b, p0/M, z18.b, z29.b\n" "smax z22.b, p0/M, z22.b, z28.b\n" - "ldp x21, x20, [x24, #0x10]\n" - "add x24, x24, #0x20\n" + "ldp x2, x1, [x5, #0x10]\n" + "add x5, x5, #0x20\n" "smax z17.b, p0/M, z17.b, z27.b\n" "smax z21.b, p0/M, z21.b, z26.b\n" "smax z16.b, p0/M, z16.b, z25.b\n" "smax z20.b, p0/M, z20.b, z24.b\n" - "ld1b { z1.b }, p4/Z, [x23, x9]\n" + "ld1b { z1.b }, p4/Z, [x4, x21]\n" "smax z19.b, p0/M, z19.b, z23.b\n" - "ld1b { z0.b }, p4/Z, [x22, x9]\n" + "ld1b { z0.b }, p4/Z, [x3, x21]\n" "smax z18.b, p0/M, z18.b, z22.b\n" - "ld1b { z23.b }, p4/Z, [x21, x9]\n" + "ld1b { z23.b }, p4/Z, [x2, x21]\n" "smax z17.b, p0/M, z17.b, z21.b\n" - "ld1b { z31.b }, p4/Z, [x20, x9]\n" + "ld1b { z31.b }, p4/Z, [x1, x21]\n" "smax z16.b, p0/M, z16.b, z20.b\n" - "ld1b { z30.b }, p3/Z, [x23, x28]\n" - "smax z5.b, p0/M, z5.b, z19.b\n" - "ld1b { z29.b }, p3/Z, [x22, x28]\n" - "smax z4.b, p0/M, z4.b, z18.b\n" - "ld1b { z22.b }, p3/Z, [x21, x28]\n" - "smax z3.b, p0/M, z3.b, z17.b\n" - "ld1b { z28.b }, p3/Z, [x20, x28]\n" + "ld1b { z30.b }, p3/Z, [x4, x20]\n" + "smax z6.b, p0/M, z6.b, z19.b\n" + "ld1b { z29.b }, p3/Z, [x3, x20]\n" + "smax z3.b, p0/M, z3.b, z18.b\n" + "ld1b { z22.b }, p3/Z, [x2, x20]\n" + "smax z5.b, p0/M, z5.b, z17.b\n" + "ld1b { z28.b }, p3/Z, [x1, x20]\n" "smax z2.b, p0/M, z2.b, z16.b\n" - "ld1b { z17.b }, p2/Z, [x23, x27]\n" - "ld1b { z27.b }, p2/Z, [x22, x27]\n" - "ld1b { z21.b }, p2/Z, [x21, x27]\n" - "ld1b { z26.b }, p2/Z, [x20, x27]\n" - "ld1b { z16.b }, p1/Z, [x23, x26]\n" - "ld1b { z25.b }, p1/Z, [x22, x26]\n" - "ld1b { z20.b }, p1/Z, [x21, x26]\n" - "ld1b { z24.b }, p1/Z, [x20, x26]\n" + "ld1b { z17.b }, p2/Z, [x4, x8]\n" + "ld1b { z27.b }, p2/Z, [x3, x8]\n" + "ld1b { z21.b }, p2/Z, [x2, x8]\n" + "ld1b { z26.b }, p2/Z, [x1, x8]\n" + "ld1b { z16.b }, p1/Z, [x4, x7]\n" + "ld1b { z25.b }, p1/Z, [x3, x7]\n" + "ld1b { z20.b }, p1/Z, [x2, x7]\n" + "ld1b { z24.b }, p1/Z, [x1, x7]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail "movprfx z19, z1\n smax z19.b, p0/M, z19.b, z0.b\n" @@ -133,282 +133,253 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "smax z18.b, p0/M, z18.b, z22.b\n" "smax z17.b, p0/M, z17.b, z21.b\n" "smax z16.b, p0/M, z16.b, z20.b\n" - "smax z5.b, p0/M, z5.b, z19.b\n" - "smax z4.b, p0/M, z4.b, z18.b\n" - "smax z3.b, p0/M, z3.b, z17.b\n" + "smax z6.b, p0/M, z6.b, z19.b\n" + "smax z3.b, p0/M, z3.b, z18.b\n" + "smax z5.b, p0/M, z5.b, z17.b\n" "smax z2.b, p0/M, z2.b, z16.b\n" "4:" // 4-vectors of channels: After loop - "ands x21, %x[n_valid_cells], #0x3\n" + "ands x2, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x20, [x24], #0x8\n" - "subs x21, x21, #0x1\n" - "ld1b { z19.b }, p4/Z, [x20, x9]\n" - "ld1b { z18.b }, p3/Z, [x20, x28]\n" - "ld1b { z17.b }, p2/Z, [x20, x27]\n" - "ld1b { z16.b }, p1/Z, [x20, x26]\n" - "smax z5.b, p0/M, z5.b, z19.b\n" - "smax z4.b, p0/M, z4.b, z18.b\n" - "smax z3.b, p0/M, z3.b, z17.b\n" + "ldr x1, [x5], #0x8\n" + "subs x2, x2, #0x1\n" + "ld1b { z19.b }, p4/Z, [x1, x21]\n" + "ld1b { z18.b }, p3/Z, [x1, x20]\n" + "ld1b { z17.b }, p2/Z, [x1, x8]\n" + "ld1b { z16.b }, p1/Z, [x1, x7]\n" + "smax z6.b, p0/M, z6.b, z19.b\n" + "smax z3.b, p0/M, z3.b, z18.b\n" + "smax z5.b, p0/M, z5.b, z17.b\n" "smax z2.b, p0/M, z2.b, z16.b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - "add x20, %x[quant_params], %[offsetof_qp_input_offset]\n" - ".inst 0x4508a0b3 // sshllb z19.h, z5.b, #0\n" - ".inst 0x4508a4b6 // sshllt z22.h, z5.b, #0\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1rw { z5.s }, p0/Z, [x20]\n" - ".inst 0x4508a095 // sshllb z21.h, z4.b, #0\n" - ".inst 0x4508a492 // sshllt z18.h, z4.b, #0\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - ".inst 0x4508a074 // sshllb z20.h, z3.b, #0\n" - ".inst 0x4508a471 // sshllt z17.h, z3.b, #0\n" - "ld1rw { z4.s }, p0/Z, [x21]\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - ".inst 0x4508a058 // sshllb z24.h, z2.b, #0\n" - ".inst 0x4508a457 // sshllt z23.h, z2.b, #0\n" - "ld1rw { z3.s }, p0/Z, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" - "neg z5.s, p0/M, z5.s\n" - "ld1rw { z2.s }, p0/Z, [x21]\n" - "mov z1.s, #0x7f\n" - "ld1rw { z16.s }, p0/Z, [x20]\n" - ".inst 0x459340a0 // saddwb z0.s, z5.s, z19.h\n" - ".inst 0x459344b3 // saddwt z19.s, z5.s, z19.h\n" - ".inst 0x459640bf // saddwb z31.s, z5.s, z22.h\n" - ".inst 0x459644be // saddwt z30.s, z5.s, z22.h\n" - ".inst 0x459540bd // saddwb z29.s, z5.s, z21.h\n" - ".inst 0x459544b6 // saddwt z22.s, z5.s, z21.h\n" - ".inst 0x459240bc // saddwb z28.s, z5.s, z18.h\n" - ".inst 0x459244b2 // saddwt z18.s, z5.s, z18.h\n" - ".inst 0x459440bb // saddwb z27.s, z5.s, z20.h\n" - ".inst 0x459444b5 // saddwt z21.s, z5.s, z20.h\n" - ".inst 0x459140b4 // saddwb z20.s, z5.s, z17.h\n" - ".inst 0x459144b1 // saddwt z17.s, z5.s, z17.h\n" - ".inst 0x459840ba // saddwb z26.s, z5.s, z24.h\n" - ".inst 0x459844b9 // saddwt z25.s, z5.s, z24.h\n" - ".inst 0x459740b8 // saddwb z24.s, z5.s, z23.h\n" - ".inst 0x459744b7 // saddwt z23.s, z5.s, z23.h\n" - ".inst 0x44828080 // srshl z0.s, p0/M, z0.s, z4.s\n" - ".inst 0x44828093 // srshl z19.s, p0/M, z19.s, z4.s\n" - ".inst 0x4482809f // srshl z31.s, p0/M, z31.s, z4.s\n" - ".inst 0x4482809e // srshl z30.s, p0/M, z30.s, z4.s\n" - ".inst 0x4482809d // srshl z29.s, p0/M, z29.s, z4.s\n" - ".inst 0x44828096 // srshl z22.s, p0/M, z22.s, z4.s\n" - ".inst 0x4482809c // srshl z28.s, p0/M, z28.s, z4.s\n" - ".inst 0x44828092 // srshl z18.s, p0/M, z18.s, z4.s\n" - ".inst 0x4482809b // srshl z27.s, p0/M, z27.s, z4.s\n" - ".inst 0x44828095 // srshl z21.s, p0/M, z21.s, z4.s\n" - ".inst 0x44828094 // srshl z20.s, p0/M, z20.s, z4.s\n" - ".inst 0x44828091 // srshl z17.s, p0/M, z17.s, z4.s\n" - ".inst 0x4482809a // srshl z26.s, p0/M, z26.s, z4.s\n" - ".inst 0x44828099 // srshl z25.s, p0/M, z25.s, z4.s\n" - ".inst 0x44828098 // srshl z24.s, p0/M, z24.s, z4.s\n" - ".inst 0x44828097 // srshl z23.s, p0/M, z23.s, z4.s\n" - ".inst 0x04a37400 // sqrdmulh z0.s, z0.s, z3.s\n" - ".inst 0x04a37673 // sqrdmulh z19.s, z19.s, z3.s\n" - ".inst 0x04a377ff // sqrdmulh z31.s, z31.s, z3.s\n" - ".inst 0x04a377de // sqrdmulh z30.s, z30.s, z3.s\n" - ".inst 0x04a377bd // sqrdmulh z29.s, z29.s, z3.s\n" - ".inst 0x04a376d6 // sqrdmulh z22.s, z22.s, z3.s\n" - ".inst 0x04a3779c // sqrdmulh z28.s, z28.s, z3.s\n" - ".inst 0x04a37652 // sqrdmulh z18.s, z18.s, z3.s\n" - ".inst 0x04a3777b // sqrdmulh z27.s, z27.s, z3.s\n" - ".inst 0x04a376b5 // sqrdmulh z21.s, z21.s, z3.s\n" - ".inst 0x04a37694 // sqrdmulh z20.s, z20.s, z3.s\n" - ".inst 0x04a37631 // sqrdmulh z17.s, z17.s, z3.s\n" - ".inst 0x04a3775a // sqrdmulh z26.s, z26.s, z3.s\n" - ".inst 0x04a37739 // sqrdmulh z25.s, z25.s, z3.s\n" - ".inst 0x04a37718 // sqrdmulh z24.s, z24.s, z3.s\n" - ".inst 0x04a376f7 // sqrdmulh z23.s, z23.s, z3.s\n" - ".inst 0x44828040 // srshl z0.s, p0/M, z0.s, z2.s\n" - ".inst 0x44828053 // srshl z19.s, p0/M, z19.s, z2.s\n" - ".inst 0x4482805f // srshl z31.s, p0/M, z31.s, z2.s\n" - ".inst 0x4482805e // srshl z30.s, p0/M, z30.s, z2.s\n" - ".inst 0x4482805d // srshl z29.s, p0/M, z29.s, z2.s\n" - ".inst 0x44828056 // srshl z22.s, p0/M, z22.s, z2.s\n" - ".inst 0x4482805c // srshl z28.s, p0/M, z28.s, z2.s\n" - ".inst 0x44828052 // srshl z18.s, p0/M, z18.s, z2.s\n" - ".inst 0x4482805b // srshl z27.s, p0/M, z27.s, z2.s\n" - ".inst 0x44828055 // srshl z21.s, p0/M, z21.s, z2.s\n" - ".inst 0x44828054 // srshl z20.s, p0/M, z20.s, z2.s\n" - ".inst 0x44828051 // srshl z17.s, p0/M, z17.s, z2.s\n" - ".inst 0x4482805a // srshl z26.s, p0/M, z26.s, z2.s\n" - ".inst 0x44828059 // srshl z25.s, p0/M, z25.s, z2.s\n" - ".inst 0x44828058 // srshl z24.s, p0/M, z24.s, z2.s\n" - ".inst 0x44828057 // srshl z23.s, p0/M, z23.s, z2.s\n" - "add z0.s, z0.s, z16.s\n" - "add z19.s, z19.s, z16.s\n" - "add z31.s, z31.s, z16.s\n" - "add z30.s, z30.s, z16.s\n" - "add z29.s, z29.s, z16.s\n" - "add z22.s, z22.s, z16.s\n" - "add z28.s, z28.s, z16.s\n" - "add z18.s, z18.s, z16.s\n" - "add z27.s, z27.s, z16.s\n" - "add z21.s, z21.s, z16.s\n" - "add z20.s, z20.s, z16.s\n" - "add z17.s, z17.s, z16.s\n" - "add z26.s, z26.s, z16.s\n" - "add z25.s, z25.s, z16.s\n" - "add z24.s, z24.s, z16.s\n" - "add z23.s, z23.s, z16.s\n" - "not z16.s, p0/M, z1.s\n" - "smax z0.s, p0/M, z0.s, z16.s\n" - "smax z19.s, p0/M, z19.s, z16.s\n" - "smax z31.s, p0/M, z31.s, z16.s\n" - "smax z30.s, p0/M, z30.s, z16.s\n" - "smax z29.s, p0/M, z29.s, z16.s\n" - "smax z22.s, p0/M, z22.s, z16.s\n" - "smax z28.s, p0/M, z28.s, z16.s\n" - "smax z18.s, p0/M, z18.s, z16.s\n" - "smax z27.s, p0/M, z27.s, z16.s\n" - "smax z21.s, p0/M, z21.s, z16.s\n" - "smax z20.s, p0/M, z20.s, z16.s\n" - "smax z17.s, p0/M, z17.s, z16.s\n" - "smax z26.s, p0/M, z26.s, z16.s\n" - "smax z25.s, p0/M, z25.s, z16.s\n" - "smax z24.s, p0/M, z24.s, z16.s\n" - "smax z23.s, p0/M, z23.s, z16.s\n" - "smin z0.s, p0/M, z0.s, z1.s\n" - "smin z19.s, p0/M, z19.s, z1.s\n" - "smin z31.s, p0/M, z31.s, z1.s\n" - "smin z30.s, p0/M, z30.s, z1.s\n" - "smin z29.s, p0/M, z29.s, z1.s\n" - "smin z22.s, p0/M, z22.s, z1.s\n" - "smin z28.s, p0/M, z28.s, z1.s\n" - "smin z18.s, p0/M, z18.s, z1.s\n" - "smin z27.s, p0/M, z27.s, z1.s\n" - "smin z21.s, p0/M, z21.s, z1.s\n" - "trn1 z19.h, z0.h, z19.h\n" - "smin z20.s, p0/M, z20.s, z1.s\n" - "smin z17.s, p0/M, z17.s, z1.s\n" - "trn1 z16.h, z31.h, z30.h\n" - "smin z26.s, p0/M, z26.s, z1.s\n" - "smin z25.s, p0/M, z25.s, z1.s\n" - "trn1 z22.h, z29.h, z22.h\n" - "smin z24.s, p0/M, z24.s, z1.s\n" - "smin z23.s, p0/M, z23.s, z1.s\n" - "trn1 z18.h, z28.h, z18.h\n" - "trn1 z21.h, z27.h, z21.h\n" - "trn1 z17.h, z20.h, z17.h\n" - "trn1 z20.b, z19.b, z16.b\n" - "trn1 z19.h, z26.h, z25.h\n" - "trn1 z16.h, z24.h, z23.h\n" - "trn1 z18.b, z22.b, z18.b\n" - "trn1 z17.b, z21.b, z17.b\n" - "st1b { z20.b }, p4, [%x[outptr], x9]\n" - "incb x9, ALL, MUL #4\n" - "trn1 z16.b, z19.b, z16.b\n" - "st1b { z18.b }, p3, [%x[outptr], x28]\n" - "incb x28, ALL, MUL #4\n" - "st1b { z17.b }, p2, [%x[outptr], x27]\n" - "incb x27, ALL, MUL #4\n" - "st1b { z16.b }, p1, [%x[outptr], x26]\n" - "incb x26, ALL, MUL #4\n" - "whilelt p1.b, x26, %x[n_channels]\n" + "add x1, %x[quant_params], %[offsetof_qp_input_offset]\n" + ".inst 0x4508a0d7 // sshllb z23.h, z6.b, #0\n" + ".inst 0x4508a076 // sshllb z22.h, z3.b, #0\n" + "add x2, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z4.s }, p0/Z, [x1]\n" + ".inst 0x4508a0b5 // sshllb z21.h, z5.b, #0\n" + ".inst 0x4508a054 // sshllb z20.h, z2.b, #0\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + ".inst 0x4508a4db // sshllt z27.h, z6.b, #0\n" + ".inst 0x4508a47a // sshllt z26.h, z3.b, #0\n" + "ld1rw { z3.s }, p0/Z, [x2]\n" + "add x2, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + ".inst 0x4508a4b9 // sshllt z25.h, z5.b, #0\n" + ".inst 0x4508a458 // sshllt z24.h, z2.b, #0\n" + "ld1rw { z2.s }, p0/Z, [x1]\n" + "add x1, %x[quant_params], %[offsetof_qp_output_offset]\n" + "neg z4.s, p0/M, z4.s\n" + "ld1rw { z1.s }, p0/Z, [x2]\n" + "ld1rw { z0.s }, p0/Z, [x1]\n" + ".inst 0x45974093 // saddwb z19.s, z4.s, z23.h\n" + ".inst 0x45964092 // saddwb z18.s, z4.s, z22.h\n" + ".inst 0x45954091 // saddwb z17.s, z4.s, z21.h\n" + ".inst 0x45944090 // saddwb z16.s, z4.s, z20.h\n" + ".inst 0x4597449f // saddwt z31.s, z4.s, z23.h\n" + ".inst 0x459b4097 // saddwb z23.s, z4.s, z27.h\n" + ".inst 0x4596449e // saddwt z30.s, z4.s, z22.h\n" + ".inst 0x459a4096 // saddwb z22.s, z4.s, z26.h\n" + ".inst 0x4595449d // saddwt z29.s, z4.s, z21.h\n" + ".inst 0x45994095 // saddwb z21.s, z4.s, z25.h\n" + ".inst 0x4594449c // saddwt z28.s, z4.s, z20.h\n" + ".inst 0x45984094 // saddwb z20.s, z4.s, z24.h\n" + ".inst 0x44828073 // srshl z19.s, p0/M, z19.s, z3.s\n" + ".inst 0x44828072 // srshl z18.s, p0/M, z18.s, z3.s\n" + ".inst 0x44828071 // srshl z17.s, p0/M, z17.s, z3.s\n" + ".inst 0x44828070 // srshl z16.s, p0/M, z16.s, z3.s\n" + ".inst 0x459b449b // saddwt z27.s, z4.s, z27.h\n" + ".inst 0x459a449a // saddwt z26.s, z4.s, z26.h\n" + ".inst 0x45994499 // saddwt z25.s, z4.s, z25.h\n" + ".inst 0x45984498 // saddwt z24.s, z4.s, z24.h\n" + ".inst 0x04a27673 // sqrdmulh z19.s, z19.s, z2.s\n" + ".inst 0x4482807f // srshl z31.s, p0/M, z31.s, z3.s\n" + ".inst 0x44828077 // srshl z23.s, p0/M, z23.s, z3.s\n" + ".inst 0x04a27652 // sqrdmulh z18.s, z18.s, z2.s\n" + ".inst 0x4482807e // srshl z30.s, p0/M, z30.s, z3.s\n" + ".inst 0x44828076 // srshl z22.s, p0/M, z22.s, z3.s\n" + ".inst 0x04a27631 // sqrdmulh z17.s, z17.s, z2.s\n" + ".inst 0x4482807d // srshl z29.s, p0/M, z29.s, z3.s\n" + ".inst 0x44828075 // srshl z21.s, p0/M, z21.s, z3.s\n" + ".inst 0x04a27610 // sqrdmulh z16.s, z16.s, z2.s\n" + ".inst 0x4482807c // srshl z28.s, p0/M, z28.s, z3.s\n" + ".inst 0x44828074 // srshl z20.s, p0/M, z20.s, z3.s\n" + ".inst 0x44828033 // srshl z19.s, p0/M, z19.s, z1.s\n" + ".inst 0x04a277ff // sqrdmulh z31.s, z31.s, z2.s\n" + ".inst 0x04a276f7 // sqrdmulh z23.s, z23.s, z2.s\n" + ".inst 0x4482807b // srshl z27.s, p0/M, z27.s, z3.s\n" + ".inst 0x44828032 // srshl z18.s, p0/M, z18.s, z1.s\n" + ".inst 0x04a277de // sqrdmulh z30.s, z30.s, z2.s\n" + ".inst 0x04a276d6 // sqrdmulh z22.s, z22.s, z2.s\n" + ".inst 0x4482807a // srshl z26.s, p0/M, z26.s, z3.s\n" + ".inst 0x44828031 // srshl z17.s, p0/M, z17.s, z1.s\n" + ".inst 0x04a277bd // sqrdmulh z29.s, z29.s, z2.s\n" + ".inst 0x04a276b5 // sqrdmulh z21.s, z21.s, z2.s\n" + ".inst 0x44828079 // srshl z25.s, p0/M, z25.s, z3.s\n" + ".inst 0x44828030 // srshl z16.s, p0/M, z16.s, z1.s\n" + ".inst 0x04a2779c // sqrdmulh z28.s, z28.s, z2.s\n" + ".inst 0x04a27694 // sqrdmulh z20.s, z20.s, z2.s\n" + ".inst 0x44828078 // srshl z24.s, p0/M, z24.s, z3.s\n" + ".inst 0x4482803f // srshl z31.s, p0/M, z31.s, z1.s\n" + ".inst 0x44828037 // srshl z23.s, p0/M, z23.s, z1.s\n" + ".inst 0x04a2777b // sqrdmulh z27.s, z27.s, z2.s\n" + ".inst 0x4482803e // srshl z30.s, p0/M, z30.s, z1.s\n" + ".inst 0x44828036 // srshl z22.s, p0/M, z22.s, z1.s\n" + ".inst 0x04a2775a // sqrdmulh z26.s, z26.s, z2.s\n" + ".inst 0x4482803d // srshl z29.s, p0/M, z29.s, z1.s\n" + ".inst 0x44828035 // srshl z21.s, p0/M, z21.s, z1.s\n" + ".inst 0x04a27739 // sqrdmulh z25.s, z25.s, z2.s\n" + ".inst 0x4482803c // srshl z28.s, p0/M, z28.s, z1.s\n" + ".inst 0x44828034 // srshl z20.s, p0/M, z20.s, z1.s\n" + ".inst 0x04a27718 // sqrdmulh z24.s, z24.s, z2.s\n" + "add z19.s, z19.s, z0.s\n" + "add z18.s, z18.s, z0.s\n" + "add z17.s, z17.s, z0.s\n" + "add z16.s, z16.s, z0.s\n" + ".inst 0x4482803b // srshl z27.s, p0/M, z27.s, z1.s\n" + ".inst 0x4482803a // srshl z26.s, p0/M, z26.s, z1.s\n" + ".inst 0x44828039 // srshl z25.s, p0/M, z25.s, z1.s\n" + ".inst 0x44828038 // srshl z24.s, p0/M, z24.s, z1.s\n" + "add z31.s, z31.s, z0.s\n" + "add z23.s, z23.s, z0.s\n" + "add z30.s, z30.s, z0.s\n" + "add z22.s, z22.s, z0.s\n" + "add z29.s, z29.s, z0.s\n" + "add z21.s, z21.s, z0.s\n" + "add z28.s, z28.s, z0.s\n" + "add z20.s, z20.s, z0.s\n" + ".inst 0x45304273 // sqxtnb z19.h, z19.s\n" + ".inst 0x45304252 // sqxtnb z18.h, z18.s\n" + ".inst 0x45304231 // sqxtnb z17.h, z17.s\n" + ".inst 0x45304210 // sqxtnb z16.h, z16.s\n" + "add z27.s, z27.s, z0.s\n" + "add z26.s, z26.s, z0.s\n" + "add z25.s, z25.s, z0.s\n" + "add z24.s, z24.s, z0.s\n" + ".inst 0x453047f3 // sqxtnt z19.h, z31.s\n" + ".inst 0x453042f7 // sqxtnb z23.h, z23.s\n" + ".inst 0x453047d2 // sqxtnt z18.h, z30.s\n" + ".inst 0x453042d6 // sqxtnb z22.h, z22.s\n" + ".inst 0x453047b1 // sqxtnt z17.h, z29.s\n" + ".inst 0x453042b5 // sqxtnb z21.h, z21.s\n" + ".inst 0x45304790 // sqxtnt z16.h, z28.s\n" + ".inst 0x45304294 // sqxtnb z20.h, z20.s\n" + ".inst 0x45304777 // sqxtnt z23.h, z27.s\n" + ".inst 0x45284273 // sqxtnb z19.b, z19.h\n" + ".inst 0x45304756 // sqxtnt z22.h, z26.s\n" + ".inst 0x45284252 // sqxtnb z18.b, z18.h\n" + ".inst 0x45304735 // sqxtnt z21.h, z25.s\n" + ".inst 0x45284231 // sqxtnb z17.b, z17.h\n" + ".inst 0x45304714 // sqxtnt z20.h, z24.s\n" + ".inst 0x45284210 // sqxtnb z16.b, z16.h\n" + ".inst 0x452846f3 // sqxtnt z19.b, z23.h\n" + ".inst 0x452846d2 // sqxtnt z18.b, z22.h\n" + ".inst 0x452846b1 // sqxtnt z17.b, z21.h\n" + ".inst 0x45284690 // sqxtnt z16.b, z20.h\n" + "st1b { z19.b }, p4, [%x[outptr], x21]\n" + "incb x21, ALL, MUL #4\n" + "st1b { z18.b }, p3, [%x[outptr], x20]\n" + "incb x20, ALL, MUL #4\n" + "st1b { z17.b }, p2, [%x[outptr], x8]\n" + "incb x8, ALL, MUL #4\n" + "st1b { z16.b }, p1, [%x[outptr], x7]\n" + "incb x7, ALL, MUL #4\n" + "whilelt p1.b, x7, %x[n_channels]\n" "b.ne 1b\n" "7:" // Single vector of channels - "whilelt p4.b, x9, %x[n_channels]\n" + "whilelt p4.b, x21, %x[n_channels]\n" "b.eq 14f\n" "8:" // Single vector of channels: Loop - "lsr x25, %x[n_valid_cells], #0x2\n" - "mov z5.b, #0x80\n" - "mov x24, %x[inptrs]\n" - "cbz x25, 11f\n" - "ldp x23, x22, [x24, #0]\n" - "subs x25, x25, #0x1\n" - "ldp x21, x20, [x24, #0x10]\n" - "add x24, x24, #0x20\n" - "ld1b { z1.b }, p4/Z, [x23, x9]\n" - "ld1b { z0.b }, p4/Z, [x22, x9]\n" - "ld1b { z23.b }, p4/Z, [x21, x9]\n" - "ld1b { z31.b }, p4/Z, [x20, x9]\n" + "lsr x6, %x[n_valid_cells], #0x2\n" + "mov z6.b, #0x80\n" + "mov x5, %x[inptrs]\n" + "cbz x6, 11f\n" + "ldp x4, x3, [x5, #0]\n" + "subs x6, x6, #0x1\n" + "ldp x2, x1, [x5, #0x10]\n" + "add x5, x5, #0x20\n" + "ld1b { z1.b }, p4/Z, [x4, x21]\n" + "ld1b { z0.b }, p4/Z, [x3, x21]\n" + "ld1b { z23.b }, p4/Z, [x2, x21]\n" + "ld1b { z31.b }, p4/Z, [x1, x21]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop "movprfx z16, z1\n smax z16.b, p0/M, z16.b, z0.b\n" "movprfx z17, z23\n smax z17.b, p0/M, z17.b, z31.b\n" - "ldp x23, x22, [x24, #0]\n" - "subs x25, x25, #0x1\n" - "ldp x21, x20, [x24, #0x10]\n" - "add x24, x24, #0x20\n" - "ld1b { z1.b }, p4/Z, [x23, x9]\n" + "ldp x4, x3, [x5, #0]\n" + "subs x6, x6, #0x1\n" + "ldp x2, x1, [x5, #0x10]\n" + "add x5, x5, #0x20\n" + "ld1b { z1.b }, p4/Z, [x4, x21]\n" "smax z16.b, p0/M, z16.b, z17.b\n" - "ld1b { z0.b }, p4/Z, [x22, x9]\n" - "ld1b { z23.b }, p4/Z, [x21, x9]\n" - "ld1b { z31.b }, p4/Z, [x20, x9]\n" - "smax z5.b, p0/M, z5.b, z16.b\n" + "ld1b { z0.b }, p4/Z, [x3, x21]\n" + "ld1b { z23.b }, p4/Z, [x2, x21]\n" + "ld1b { z31.b }, p4/Z, [x1, x21]\n" + "smax z6.b, p0/M, z6.b, z16.b\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail "movprfx z16, z1\n smax z16.b, p0/M, z16.b, z0.b\n" "movprfx z17, z23\n smax z17.b, p0/M, z17.b, z31.b\n" "smax z16.b, p0/M, z16.b, z17.b\n" - "smax z5.b, p0/M, z5.b, z16.b\n" + "smax z6.b, p0/M, z6.b, z16.b\n" "11:" // Single vector of channels: Loop: After loop - "ands x21, %x[n_valid_cells], #0x3\n" + "ands x2, %x[n_valid_cells], #0x3\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop - "ldr x20, [x24], #0x8\n" - "subs x21, x21, #0x1\n" - "ld1b { z16.b }, p4/Z, [x20, x9]\n" - "smax z5.b, p0/M, z5.b, z16.b\n" + "ldr x1, [x5], #0x8\n" + "subs x2, x2, #0x1\n" + "ld1b { z16.b }, p4/Z, [x1, x21]\n" + "smax z6.b, p0/M, z6.b, z16.b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" - ".inst 0x4508a0b1 // sshllb z17.h, z5.b, #0\n" - ".inst 0x4508a4ba // sshllt z26.h, z5.b, #0\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1rw { z16.s }, p0/Z, [x21]\n" - "add x22, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "mov z25.s, #0x7f\n" - "ld1rw { z24.s }, p0/Z, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" - "ld1rw { z23.s }, p0/Z, [x22]\n" - "neg z16.s, p0/M, z16.s\n" - "ld1rw { z22.s }, p0/Z, [x21]\n" - "ld1rw { z21.s }, p0/Z, [x20]\n" - "not z20.s, p0/M, z25.s\n" - ".inst 0x45914213 // saddwb z19.s, z16.s, z17.h\n" - ".inst 0x45914611 // saddwt z17.s, z16.s, z17.h\n" - ".inst 0x459a4212 // saddwb z18.s, z16.s, z26.h\n" - ".inst 0x459a4610 // saddwt z16.s, z16.s, z26.h\n" - ".inst 0x44828313 // srshl z19.s, p0/M, z19.s, z24.s\n" - ".inst 0x44828311 // srshl z17.s, p0/M, z17.s, z24.s\n" - ".inst 0x44828312 // srshl z18.s, p0/M, z18.s, z24.s\n" - ".inst 0x44828310 // srshl z16.s, p0/M, z16.s, z24.s\n" - ".inst 0x04b77673 // sqrdmulh z19.s, z19.s, z23.s\n" - ".inst 0x04b77631 // sqrdmulh z17.s, z17.s, z23.s\n" - ".inst 0x04b77652 // sqrdmulh z18.s, z18.s, z23.s\n" - ".inst 0x04b77610 // sqrdmulh z16.s, z16.s, z23.s\n" - ".inst 0x448282d3 // srshl z19.s, p0/M, z19.s, z22.s\n" - ".inst 0x448282d1 // srshl z17.s, p0/M, z17.s, z22.s\n" - ".inst 0x448282d2 // srshl z18.s, p0/M, z18.s, z22.s\n" - ".inst 0x448282d0 // srshl z16.s, p0/M, z16.s, z22.s\n" - "add z19.s, z19.s, z21.s\n" - "add z17.s, z17.s, z21.s\n" - "add z18.s, z18.s, z21.s\n" - "add z16.s, z16.s, z21.s\n" - "smax z19.s, p0/M, z19.s, z20.s\n" - "smax z17.s, p0/M, z17.s, z20.s\n" - "smax z18.s, p0/M, z18.s, z20.s\n" - "smax z16.s, p0/M, z16.s, z20.s\n" - "smin z19.s, p0/M, z19.s, z25.s\n" - "smin z17.s, p0/M, z17.s, z25.s\n" - "smin z18.s, p0/M, z18.s, z25.s\n" - "smin z16.s, p0/M, z16.s, z25.s\n" - "trn1 z17.h, z19.h, z17.h\n" - "trn1 z16.h, z18.h, z16.h\n" - "trn1 z16.b, z17.b, z16.b\n" - "st1b { z16.b }, p4, [%x[outptr], x9]\n" - "incb x9\n" - "whilelt p4.b, x9, %x[n_channels]\n" + "add x2, %x[quant_params], %[offsetof_qp_input_offset]\n" + ".inst 0x4508a0d1 // sshllb z17.h, z6.b, #0\n" + ".inst 0x4508a4d8 // sshllt z24.h, z6.b, #0\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z18.s }, p0/Z, [x2]\n" + "add x3, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x2, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "ld1rw { z23.s }, p0/Z, [x1]\n" + "add x1, %x[quant_params], %[offsetof_qp_output_offset]\n" + "ld1rw { z22.s }, p0/Z, [x3]\n" + "neg z18.s, p0/M, z18.s\n" + "ld1rw { z21.s }, p0/Z, [x2]\n" + "ld1rw { z20.s }, p0/Z, [x1]\n" + ".inst 0x45914250 // saddwb z16.s, z18.s, z17.h\n" + ".inst 0x45914653 // saddwt z19.s, z18.s, z17.h\n" + ".inst 0x45984251 // saddwb z17.s, z18.s, z24.h\n" + ".inst 0x45984652 // saddwt z18.s, z18.s, z24.h\n" + ".inst 0x448282f0 // srshl z16.s, p0/M, z16.s, z23.s\n" + ".inst 0x448282f3 // srshl z19.s, p0/M, z19.s, z23.s\n" + ".inst 0x448282f1 // srshl z17.s, p0/M, z17.s, z23.s\n" + ".inst 0x448282f2 // srshl z18.s, p0/M, z18.s, z23.s\n" + ".inst 0x04b67610 // sqrdmulh z16.s, z16.s, z22.s\n" + ".inst 0x04b67673 // sqrdmulh z19.s, z19.s, z22.s\n" + ".inst 0x04b67631 // sqrdmulh z17.s, z17.s, z22.s\n" + ".inst 0x04b67652 // sqrdmulh z18.s, z18.s, z22.s\n" + ".inst 0x448282b0 // srshl z16.s, p0/M, z16.s, z21.s\n" + ".inst 0x448282b3 // srshl z19.s, p0/M, z19.s, z21.s\n" + ".inst 0x448282b1 // srshl z17.s, p0/M, z17.s, z21.s\n" + ".inst 0x448282b2 // srshl z18.s, p0/M, z18.s, z21.s\n" + "add z16.s, z16.s, z20.s\n" + "add z19.s, z19.s, z20.s\n" + "add z17.s, z17.s, z20.s\n" + "add z18.s, z18.s, z20.s\n" + ".inst 0x45304210 // sqxtnb z16.h, z16.s\n" + ".inst 0x45304231 // sqxtnb z17.h, z17.s\n" + ".inst 0x45304670 // sqxtnt z16.h, z19.s\n" + ".inst 0x45304651 // sqxtnt z17.h, z18.s\n" + ".inst 0x45284210 // sqxtnb z16.b, z16.h\n" + ".inst 0x45284630 // sqxtnt z16.b, z17.h\n" + "st1b { z16.b }, p4, [%x[outptr], x21]\n" + "incb x21\n" + "whilelt p4.b, x21, %x[n_channels]\n" "b.ne 8b\n" "14:" // End ".inst 0xd503467f // SMSTOP\n" : : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [n_valid_cells] "r" (n_valid_cells), [offsetof_qp_input_offset] "I" (offsetof(Requantize32, input_offset)), [offsetof_qp_output_offset] "I" (offsetof(Requantize32, output_offset)), [offsetof_qp_per_layer_left_shift] "I" (offsetof(Requantize32, per_layer_left_shift)), [offsetof_qp_per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [offsetof_qp_per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [outptr] "r" (outptr), [quant_params] "r" (&qp) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "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" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "x20", "x21", "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" ); } diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp index 0b3f4b88ecc..f6af6f40b6f 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp @@ -42,82 +42,82 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( ) { __asm__ __volatile__( - "mov x9, #0\n" - "cntb x28\n" - "cntb x27, ALL, MUL #2\n" - "cntb x26, ALL, MUL #3\n" + "mov x21, #0\n" + "cntb x20\n" + "cntb x8, ALL, MUL #2\n" + "cntb x7, ALL, MUL #3\n" "ptrue p4.b\n" - "whilelt p3.b, x9, %x[n_channels]\n" - "whilelt p2.b, x28, %x[n_channels]\n" - "whilelt p1.b, x27, %x[n_channels]\n" - "whilelt p0.b, x26, %x[n_channels]\n" + "whilelt p3.b, x21, %x[n_channels]\n" + "whilelt p2.b, x20, %x[n_channels]\n" + "whilelt p1.b, x8, %x[n_channels]\n" + "whilelt p0.b, x7, %x[n_channels]\n" "b.eq 7f\n" "1:" // 4-vectors of channels - "lsr x25, %x[n_valid_cells], #0x2\n" + "lsr x6, %x[n_valid_cells], #0x2\n" + "mov z8.b, #0x80\n" + "mov z7.b, #0x80\n" + "mov x5, %x[inptrs]\n" "mov z6.b, #0x80\n" "mov z5.b, #0x80\n" - "mov x24, %x[inptrs]\n" - "mov z4.b, #0x80\n" - "mov z3.b, #0x80\n" - "cbz x25, 4f\n" - "ldp x23, x22, [x24, #0]\n" - "ldp x21, x20, [x24, #0x10]\n" - "subs x25, x25, #0x1\n" - "add x24, x24, #0x20\n" - "ld1b { z2.b }, p3/Z, [x23, x9]\n" - "ld1b { z1.b }, p3/Z, [x22, x9]\n" - "ld1b { z23.b }, p3/Z, [x21, x9]\n" - "ld1b { z0.b }, p3/Z, [x20, x9]\n" - "ld1b { z31.b }, p2/Z, [x23, x28]\n" - "ld1b { z30.b }, p2/Z, [x22, x28]\n" - "ld1b { z22.b }, p2/Z, [x21, x28]\n" - "ld1b { z29.b }, p2/Z, [x20, x28]\n" - "ld1b { z28.b }, p1/Z, [x23, x27]\n" - "ld1b { z27.b }, p1/Z, [x22, x27]\n" - "ld1b { z21.b }, p1/Z, [x21, x27]\n" - "ld1b { z26.b }, p1/Z, [x20, x27]\n" - "ld1b { z16.b }, p0/Z, [x23, x26]\n" - "ld1b { z25.b }, p0/Z, [x22, x26]\n" - "ld1b { z20.b }, p0/Z, [x21, x26]\n" - "ld1b { z24.b }, p0/Z, [x20, x26]\n" + "cbz x6, 4f\n" + "ldp x4, x3, [x5, #0]\n" + "ldp x2, x1, [x5, #0x10]\n" + "subs x6, x6, #0x1\n" + "add x5, x5, #0x20\n" + "ld1b { z2.b }, p3/Z, [x4, x21]\n" + "ld1b { z1.b }, p3/Z, [x3, x21]\n" + "ld1b { z23.b }, p3/Z, [x2, x21]\n" + "ld1b { z0.b }, p3/Z, [x1, x21]\n" + "ld1b { z31.b }, p2/Z, [x4, x20]\n" + "ld1b { z30.b }, p2/Z, [x3, x20]\n" + "ld1b { z22.b }, p2/Z, [x2, x20]\n" + "ld1b { z29.b }, p2/Z, [x1, x20]\n" + "ld1b { z28.b }, p1/Z, [x4, x8]\n" + "ld1b { z27.b }, p1/Z, [x3, x8]\n" + "ld1b { z21.b }, p1/Z, [x2, x8]\n" + "ld1b { z26.b }, p1/Z, [x1, x8]\n" + "ld1b { z16.b }, p0/Z, [x4, x7]\n" + "ld1b { z25.b }, p0/Z, [x3, x7]\n" + "ld1b { z20.b }, p0/Z, [x2, x7]\n" + "ld1b { z24.b }, p0/Z, [x1, x7]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop "movprfx z19, z2\n smax z19.b, p4/M, z19.b, z1.b\n" "smax z23.b, p4/M, z23.b, z0.b\n" - "ldp x23, x22, [x24, #0]\n" - "ldp x21, x20, [x24, #0x10]\n" + "ldp x4, x3, [x5, #0]\n" + "ldp x2, x1, [x5, #0x10]\n" "movprfx z18, z31\n smax z18.b, p4/M, z18.b, z30.b\n" "smax z22.b, p4/M, z22.b, z29.b\n" "movprfx z17, z28\n smax z17.b, p4/M, z17.b, z27.b\n" "smax z21.b, p4/M, z21.b, z26.b\n" "smax z16.b, p4/M, z16.b, z25.b\n" "smax z20.b, p4/M, z20.b, z24.b\n" - "ld1b { z2.b }, p3/Z, [x23, x9]\n" - "ld1b { z1.b }, p3/Z, [x22, x9]\n" + "ld1b { z2.b }, p3/Z, [x4, x21]\n" + "ld1b { z1.b }, p3/Z, [x3, x21]\n" "smax z19.b, p4/M, z19.b, z23.b\n" "smax z18.b, p4/M, z18.b, z22.b\n" - "ld1b { z23.b }, p3/Z, [x21, x9]\n" - "ld1b { z0.b }, p3/Z, [x20, x9]\n" + "ld1b { z23.b }, p3/Z, [x2, x21]\n" + "ld1b { z0.b }, p3/Z, [x1, x21]\n" "smax z17.b, p4/M, z17.b, z21.b\n" - "subs x25, x25, #0x1\n" - "ld1b { z31.b }, p2/Z, [x23, x28]\n" - "ld1b { z30.b }, p2/Z, [x22, x28]\n" + "subs x6, x6, #0x1\n" + "ld1b { z31.b }, p2/Z, [x4, x20]\n" + "ld1b { z30.b }, p2/Z, [x3, x20]\n" "smax z16.b, p4/M, z16.b, z20.b\n" - "add x24, x24, #0x20\n" - "ld1b { z22.b }, p2/Z, [x21, x28]\n" - "ld1b { z29.b }, p2/Z, [x20, x28]\n" - "smax z6.b, p4/M, z6.b, z19.b\n" - "smax z5.b, p4/M, z5.b, z18.b\n" - "ld1b { z28.b }, p1/Z, [x23, x27]\n" - "ld1b { z27.b }, p1/Z, [x22, x27]\n" - "smax z4.b, p4/M, z4.b, z17.b\n" - "ld1b { z21.b }, p1/Z, [x21, x27]\n" - "ld1b { z26.b }, p1/Z, [x20, x27]\n" - "smax z3.b, p4/M, z3.b, z16.b\n" - "ld1b { z16.b }, p0/Z, [x23, x26]\n" - "ld1b { z25.b }, p0/Z, [x22, x26]\n" - "ld1b { z20.b }, p0/Z, [x21, x26]\n" - "ld1b { z24.b }, p0/Z, [x20, x26]\n" + "add x5, x5, #0x20\n" + "ld1b { z22.b }, p2/Z, [x2, x20]\n" + "ld1b { z29.b }, p2/Z, [x1, x20]\n" + "smax z8.b, p4/M, z8.b, z19.b\n" + "smax z7.b, p4/M, z7.b, z18.b\n" + "ld1b { z28.b }, p1/Z, [x4, x8]\n" + "ld1b { z27.b }, p1/Z, [x3, x8]\n" + "smax z6.b, p4/M, z6.b, z17.b\n" + "ld1b { z21.b }, p1/Z, [x2, x8]\n" + "ld1b { z26.b }, p1/Z, [x1, x8]\n" + "smax z5.b, p4/M, z5.b, z16.b\n" + "ld1b { z16.b }, p0/Z, [x4, x7]\n" + "ld1b { z25.b }, p0/Z, [x3, x7]\n" + "ld1b { z20.b }, p0/Z, [x2, x7]\n" + "ld1b { z24.b }, p0/Z, [x1, x7]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail "movprfx z19, z2\n smax z19.b, p4/M, z19.b, z1.b\n" @@ -132,281 +132,252 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "smax z18.b, p4/M, z18.b, z22.b\n" "smax z17.b, p4/M, z17.b, z21.b\n" "smax z16.b, p4/M, z16.b, z20.b\n" - "smax z6.b, p4/M, z6.b, z19.b\n" - "smax z5.b, p4/M, z5.b, z18.b\n" - "smax z4.b, p4/M, z4.b, z17.b\n" - "smax z3.b, p4/M, z3.b, z16.b\n" + "smax z8.b, p4/M, z8.b, z19.b\n" + "smax z7.b, p4/M, z7.b, z18.b\n" + "smax z6.b, p4/M, z6.b, z17.b\n" + "smax z5.b, p4/M, z5.b, z16.b\n" "4:" // 4-vectors of channels: After loop - "ands x21, %x[n_valid_cells], #0x3\n" + "ands x2, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x20, [x24], #0x8\n" - "subs x21, x21, #0x1\n" - "ld1b { z19.b }, p3/Z, [x20, x9]\n" - "ld1b { z18.b }, p2/Z, [x20, x28]\n" - "ld1b { z17.b }, p1/Z, [x20, x27]\n" - "ld1b { z16.b }, p0/Z, [x20, x26]\n" - "smax z6.b, p4/M, z6.b, z19.b\n" - "smax z5.b, p4/M, z5.b, z18.b\n" - "smax z4.b, p4/M, z4.b, z17.b\n" - "smax z3.b, p4/M, z3.b, z16.b\n" + "ldr x1, [x5], #0x8\n" + "subs x2, x2, #0x1\n" + "ld1b { z19.b }, p3/Z, [x1, x21]\n" + "ld1b { z18.b }, p2/Z, [x1, x20]\n" + "ld1b { z17.b }, p1/Z, [x1, x8]\n" + "ld1b { z16.b }, p0/Z, [x1, x7]\n" + "smax z8.b, p4/M, z8.b, z19.b\n" + "smax z7.b, p4/M, z7.b, z18.b\n" + "smax z6.b, p4/M, z6.b, z17.b\n" + "smax z5.b, p4/M, z5.b, z16.b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" - ".inst 0x4508a0d3 // sshllb z19.h, z6.b, #0\n" - ".inst 0x4508a4d1 // sshllt z17.h, z6.b, #0\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1rw { z6.s }, p4/Z, [x21]\n" - ".inst 0x4508a0b2 // sshllb z18.h, z5.b, #0\n" - ".inst 0x4508a4b0 // sshllt z16.h, z5.b, #0\n" - "ld1rw { z5.s }, p4/Z, [x20]\n" - ".inst 0x4508a094 // sshllb z20.h, z4.b, #0\n" - ".inst 0x4508a498 // sshllt z24.h, z4.b, #0\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - ".inst 0x4508a077 // sshllb z23.h, z3.b, #0\n" - ".inst 0x4508a476 // sshllt z22.h, z3.b, #0\n" - "ld1rw { z4.s }, p4/Z, [x21]\n" - "ld1rw { z3.s }, p4/Z, [x20]\n" - "neg z6.s, p4/M, z6.s\n" - "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" - "mov z2.s, #0x7f\n" - "ld1rw { z1.s }, p4/Z, [x20]\n" - "not z0.s, p4/M, z2.s\n" - ".inst 0x459340df // saddwb z31.s, z6.s, z19.h\n" - ".inst 0x459344d3 // saddwt z19.s, z6.s, z19.h\n" - ".inst 0x459140de // saddwb z30.s, z6.s, z17.h\n" - ".inst 0x459144d1 // saddwt z17.s, z6.s, z17.h\n" - ".inst 0x459240dd // saddwb z29.s, z6.s, z18.h\n" - ".inst 0x459244d2 // saddwt z18.s, z6.s, z18.h\n" - ".inst 0x459040dc // saddwb z28.s, z6.s, z16.h\n" - ".inst 0x459044d0 // saddwt z16.s, z6.s, z16.h\n" - ".inst 0x448290bf // srshl z31.s, p4/M, z31.s, z5.s\n" - ".inst 0x448290b3 // srshl z19.s, p4/M, z19.s, z5.s\n" - ".inst 0x459440d5 // saddwb z21.s, z6.s, z20.h\n" - ".inst 0x459444d4 // saddwt z20.s, z6.s, z20.h\n" - ".inst 0x448290be // srshl z30.s, p4/M, z30.s, z5.s\n" - ".inst 0x448290b1 // srshl z17.s, p4/M, z17.s, z5.s\n" - ".inst 0x459840db // saddwb z27.s, z6.s, z24.h\n" - ".inst 0x459844da // saddwt z26.s, z6.s, z24.h\n" - ".inst 0x448290bd // srshl z29.s, p4/M, z29.s, z5.s\n" - ".inst 0x448290b2 // srshl z18.s, p4/M, z18.s, z5.s\n" - ".inst 0x459740d9 // saddwb z25.s, z6.s, z23.h\n" - ".inst 0x459744d8 // saddwt z24.s, z6.s, z23.h\n" - ".inst 0x448290bc // srshl z28.s, p4/M, z28.s, z5.s\n" - ".inst 0x448290b0 // srshl z16.s, p4/M, z16.s, z5.s\n" - ".inst 0x459640d7 // saddwb z23.s, z6.s, z22.h\n" - ".inst 0x459644d6 // saddwt z22.s, z6.s, z22.h\n" - ".inst 0x448290b5 // srshl z21.s, p4/M, z21.s, z5.s\n" - ".inst 0x448290b4 // srshl z20.s, p4/M, z20.s, z5.s\n" - ".inst 0x448290bb // srshl z27.s, p4/M, z27.s, z5.s\n" - ".inst 0x448290ba // srshl z26.s, p4/M, z26.s, z5.s\n" - ".inst 0x04a477ff // sqrdmulh z31.s, z31.s, z4.s\n" - ".inst 0x04a47673 // sqrdmulh z19.s, z19.s, z4.s\n" - ".inst 0x448290b9 // srshl z25.s, p4/M, z25.s, z5.s\n" - ".inst 0x448290b8 // srshl z24.s, p4/M, z24.s, z5.s\n" - ".inst 0x04a477de // sqrdmulh z30.s, z30.s, z4.s\n" - ".inst 0x04a47631 // sqrdmulh z17.s, z17.s, z4.s\n" - ".inst 0x448290b7 // srshl z23.s, p4/M, z23.s, z5.s\n" - ".inst 0x448290b6 // srshl z22.s, p4/M, z22.s, z5.s\n" - ".inst 0x04a477bd // sqrdmulh z29.s, z29.s, z4.s\n" - ".inst 0x04a47652 // sqrdmulh z18.s, z18.s, z4.s\n" - ".inst 0x04a4779c // sqrdmulh z28.s, z28.s, z4.s\n" - ".inst 0x04a47610 // sqrdmulh z16.s, z16.s, z4.s\n" - ".inst 0x4482907f // srshl z31.s, p4/M, z31.s, z3.s\n" + "add x2, %x[quant_params], %[offsetof_qp_input_offset]\n" + ".inst 0x4508a117 // sshllb z23.h, z8.b, #0\n" + ".inst 0x4508a0f6 // sshllb z22.h, z7.b, #0\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z4.s }, p4/Z, [x2]\n" + ".inst 0x4508a0d5 // sshllb z21.h, z6.b, #0\n" + ".inst 0x4508a0b4 // sshllb z20.h, z5.b, #0\n" + "ld1rw { z3.s }, p4/Z, [x1]\n" + ".inst 0x4508a51b // sshllt z27.h, z8.b, #0\n" + ".inst 0x4508a4fa // sshllt z26.h, z7.b, #0\n" + "add x2, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + ".inst 0x4508a4d9 // sshllt z25.h, z6.b, #0\n" + ".inst 0x4508a4b8 // sshllt z24.h, z5.b, #0\n" + "ld1rw { z2.s }, p4/Z, [x2]\n" + "ld1rw { z1.s }, p4/Z, [x1]\n" + "neg z4.s, p4/M, z4.s\n" + "add x1, %x[quant_params], %[offsetof_qp_output_offset]\n" + "ld1rw { z0.s }, p4/Z, [x1]\n" + ".inst 0x45974093 // saddwb z19.s, z4.s, z23.h\n" + ".inst 0x45964092 // saddwb z18.s, z4.s, z22.h\n" + ".inst 0x45954091 // saddwb z17.s, z4.s, z21.h\n" + ".inst 0x45944090 // saddwb z16.s, z4.s, z20.h\n" + ".inst 0x4597449f // saddwt z31.s, z4.s, z23.h\n" + ".inst 0x459b4097 // saddwb z23.s, z4.s, z27.h\n" + ".inst 0x4596449e // saddwt z30.s, z4.s, z22.h\n" ".inst 0x44829073 // srshl z19.s, p4/M, z19.s, z3.s\n" - ".inst 0x04a476b5 // sqrdmulh z21.s, z21.s, z4.s\n" - ".inst 0x04a47694 // sqrdmulh z20.s, z20.s, z4.s\n" - ".inst 0x4482907e // srshl z30.s, p4/M, z30.s, z3.s\n" - ".inst 0x44829071 // srshl z17.s, p4/M, z17.s, z3.s\n" - ".inst 0x04a4777b // sqrdmulh z27.s, z27.s, z4.s\n" - ".inst 0x04a4775a // sqrdmulh z26.s, z26.s, z4.s\n" - ".inst 0x4482907d // srshl z29.s, p4/M, z29.s, z3.s\n" + ".inst 0x459a4096 // saddwb z22.s, z4.s, z26.h\n" + ".inst 0x4595449d // saddwt z29.s, z4.s, z21.h\n" ".inst 0x44829072 // srshl z18.s, p4/M, z18.s, z3.s\n" - ".inst 0x04a47739 // sqrdmulh z25.s, z25.s, z4.s\n" - ".inst 0x04a47718 // sqrdmulh z24.s, z24.s, z4.s\n" - ".inst 0x4482907c // srshl z28.s, p4/M, z28.s, z3.s\n" + ".inst 0x44829071 // srshl z17.s, p4/M, z17.s, z3.s\n" + ".inst 0x45994095 // saddwb z21.s, z4.s, z25.h\n" + ".inst 0x4594449c // saddwt z28.s, z4.s, z20.h\n" ".inst 0x44829070 // srshl z16.s, p4/M, z16.s, z3.s\n" - ".inst 0x04a476f7 // sqrdmulh z23.s, z23.s, z4.s\n" - ".inst 0x04a476d6 // sqrdmulh z22.s, z22.s, z4.s\n" + ".inst 0x4482907f // srshl z31.s, p4/M, z31.s, z3.s\n" + ".inst 0x45984094 // saddwb z20.s, z4.s, z24.h\n" + ".inst 0x459b449b // saddwt z27.s, z4.s, z27.h\n" + ".inst 0x44829077 // srshl z23.s, p4/M, z23.s, z3.s\n" + ".inst 0x4482907e // srshl z30.s, p4/M, z30.s, z3.s\n" + ".inst 0x459a449a // saddwt z26.s, z4.s, z26.h\n" + ".inst 0x45994499 // saddwt z25.s, z4.s, z25.h\n" + ".inst 0x44829076 // srshl z22.s, p4/M, z22.s, z3.s\n" + ".inst 0x4482907d // srshl z29.s, p4/M, z29.s, z3.s\n" + ".inst 0x45984498 // saddwt z24.s, z4.s, z24.h\n" + ".inst 0x04a27673 // sqrdmulh z19.s, z19.s, z2.s\n" ".inst 0x44829075 // srshl z21.s, p4/M, z21.s, z3.s\n" + ".inst 0x4482907c // srshl z28.s, p4/M, z28.s, z3.s\n" + ".inst 0x04a27652 // sqrdmulh z18.s, z18.s, z2.s\n" + ".inst 0x04a27631 // sqrdmulh z17.s, z17.s, z2.s\n" ".inst 0x44829074 // srshl z20.s, p4/M, z20.s, z3.s\n" ".inst 0x4482907b // srshl z27.s, p4/M, z27.s, z3.s\n" + ".inst 0x04a27610 // sqrdmulh z16.s, z16.s, z2.s\n" + ".inst 0x04a277ff // sqrdmulh z31.s, z31.s, z2.s\n" ".inst 0x4482907a // srshl z26.s, p4/M, z26.s, z3.s\n" - "add z31.s, z31.s, z1.s\n" - "add z19.s, z19.s, z1.s\n" ".inst 0x44829079 // srshl z25.s, p4/M, z25.s, z3.s\n" + ".inst 0x44829033 // srshl z19.s, p4/M, z19.s, z1.s\n" + ".inst 0x04a276f7 // sqrdmulh z23.s, z23.s, z2.s\n" + ".inst 0x04a277de // sqrdmulh z30.s, z30.s, z2.s\n" ".inst 0x44829078 // srshl z24.s, p4/M, z24.s, z3.s\n" - "add z30.s, z30.s, z1.s\n" - "add z17.s, z17.s, z1.s\n" - ".inst 0x44829077 // srshl z23.s, p4/M, z23.s, z3.s\n" - ".inst 0x44829076 // srshl z22.s, p4/M, z22.s, z3.s\n" - "add z29.s, z29.s, z1.s\n" - "add z18.s, z18.s, z1.s\n" - "add z28.s, z28.s, z1.s\n" - "add z16.s, z16.s, z1.s\n" - "smax z31.s, p4/M, z31.s, z0.s\n" - "smax z19.s, p4/M, z19.s, z0.s\n" - "add z21.s, z21.s, z1.s\n" - "add z20.s, z20.s, z1.s\n" - "smax z30.s, p4/M, z30.s, z0.s\n" - "smax z17.s, p4/M, z17.s, z0.s\n" - "add z27.s, z27.s, z1.s\n" - "add z26.s, z26.s, z1.s\n" - "smax z29.s, p4/M, z29.s, z0.s\n" - "smax z18.s, p4/M, z18.s, z0.s\n" - "add z25.s, z25.s, z1.s\n" - "add z24.s, z24.s, z1.s\n" - "smax z28.s, p4/M, z28.s, z0.s\n" - "smax z16.s, p4/M, z16.s, z0.s\n" - "add z23.s, z23.s, z1.s\n" - "add z22.s, z22.s, z1.s\n" - "smax z21.s, p4/M, z21.s, z0.s\n" - "smax z20.s, p4/M, z20.s, z0.s\n" - "smax z27.s, p4/M, z27.s, z0.s\n" - "smax z26.s, p4/M, z26.s, z0.s\n" - "smax z25.s, p4/M, z25.s, z0.s\n" - "smax z24.s, p4/M, z24.s, z0.s\n" - "smax z23.s, p4/M, z23.s, z0.s\n" - "smax z22.s, p4/M, z22.s, z0.s\n" - "smin z31.s, p4/M, z31.s, z2.s\n" - "smin z19.s, p4/M, z19.s, z2.s\n" - "smin z30.s, p4/M, z30.s, z2.s\n" - "smin z17.s, p4/M, z17.s, z2.s\n" - "smin z29.s, p4/M, z29.s, z2.s\n" - "smin z18.s, p4/M, z18.s, z2.s\n" - "smin z28.s, p4/M, z28.s, z2.s\n" - "smin z16.s, p4/M, z16.s, z2.s\n" - "trn1 z19.h, z31.h, z19.h\n" - "smin z21.s, p4/M, z21.s, z2.s\n" - "smin z20.s, p4/M, z20.s, z2.s\n" - "trn1 z17.h, z30.h, z17.h\n" - "smin z27.s, p4/M, z27.s, z2.s\n" - "smin z26.s, p4/M, z26.s, z2.s\n" - "trn1 z18.h, z29.h, z18.h\n" - "smin z25.s, p4/M, z25.s, z2.s\n" - "smin z24.s, p4/M, z24.s, z2.s\n" - "trn1 z16.h, z28.h, z16.h\n" - "smin z23.s, p4/M, z23.s, z2.s\n" - "smin z22.s, p4/M, z22.s, z2.s\n" - "trn1 z21.h, z21.h, z20.h\n" - "trn1 z20.b, z19.b, z17.b\n" - "trn1 z17.h, z27.h, z26.h\n" - "trn1 z19.h, z25.h, z24.h\n" - "trn1 z18.b, z18.b, z16.b\n" - "trn1 z16.h, z23.h, z22.h\n" - "st1b { z20.b }, p3, [%x[outptr], x9]\n" - "incb x9, ALL, MUL #4\n" - "trn1 z17.b, z21.b, z17.b\n" - "trn1 z16.b, z19.b, z16.b\n" - "st1b { z18.b }, p2, [%x[outptr], x28]\n" - "incb x28, ALL, MUL #4\n" - "st1b { z17.b }, p1, [%x[outptr], x27]\n" - "incb x27, ALL, MUL #4\n" - "st1b { z16.b }, p0, [%x[outptr], x26]\n" - "incb x26, ALL, MUL #4\n" - "whilelt p0.b, x26, %x[n_channels]\n" + ".inst 0x44829032 // srshl z18.s, p4/M, z18.s, z1.s\n" + ".inst 0x04a276d6 // sqrdmulh z22.s, z22.s, z2.s\n" + ".inst 0x44829031 // srshl z17.s, p4/M, z17.s, z1.s\n" + ".inst 0x04a277bd // sqrdmulh z29.s, z29.s, z2.s\n" + ".inst 0x04a276b5 // sqrdmulh z21.s, z21.s, z2.s\n" + ".inst 0x44829030 // srshl z16.s, p4/M, z16.s, z1.s\n" + ".inst 0x04a2779c // sqrdmulh z28.s, z28.s, z2.s\n" + ".inst 0x4482903f // srshl z31.s, p4/M, z31.s, z1.s\n" + ".inst 0x04a27694 // sqrdmulh z20.s, z20.s, z2.s\n" + ".inst 0x44829037 // srshl z23.s, p4/M, z23.s, z1.s\n" + ".inst 0x04a2777b // sqrdmulh z27.s, z27.s, z2.s\n" + ".inst 0x4482903e // srshl z30.s, p4/M, z30.s, z1.s\n" + ".inst 0x44829036 // srshl z22.s, p4/M, z22.s, z1.s\n" + ".inst 0x04a2775a // sqrdmulh z26.s, z26.s, z2.s\n" + ".inst 0x4482903d // srshl z29.s, p4/M, z29.s, z1.s\n" + ".inst 0x04a27739 // sqrdmulh z25.s, z25.s, z2.s\n" + ".inst 0x44829035 // srshl z21.s, p4/M, z21.s, z1.s\n" + ".inst 0x4482903c // srshl z28.s, p4/M, z28.s, z1.s\n" + ".inst 0x04a27718 // sqrdmulh z24.s, z24.s, z2.s\n" + "add z19.s, z19.s, z0.s\n" + ".inst 0x44829034 // srshl z20.s, p4/M, z20.s, z1.s\n" + "add z18.s, z18.s, z0.s\n" + "add z17.s, z17.s, z0.s\n" + ".inst 0x4482903b // srshl z27.s, p4/M, z27.s, z1.s\n" + "add z16.s, z16.s, z0.s\n" + ".inst 0x4482903a // srshl z26.s, p4/M, z26.s, z1.s\n" + ".inst 0x44829039 // srshl z25.s, p4/M, z25.s, z1.s\n" + "add z31.s, z31.s, z0.s\n" + ".inst 0x44829038 // srshl z24.s, p4/M, z24.s, z1.s\n" + "add z23.s, z23.s, z0.s\n" + "add z30.s, z30.s, z0.s\n" + "add z22.s, z22.s, z0.s\n" + "add z29.s, z29.s, z0.s\n" + "add z21.s, z21.s, z0.s\n" + "add z28.s, z28.s, z0.s\n" + "add z20.s, z20.s, z0.s\n" + ".inst 0x45304273 // sqxtnb z19.h, z19.s\n" + ".inst 0x45304252 // sqxtnb z18.h, z18.s\n" + ".inst 0x45304231 // sqxtnb z17.h, z17.s\n" + ".inst 0x45304210 // sqxtnb z16.h, z16.s\n" + "add z27.s, z27.s, z0.s\n" + "add z26.s, z26.s, z0.s\n" + "add z25.s, z25.s, z0.s\n" + "add z24.s, z24.s, z0.s\n" + ".inst 0x453047f3 // sqxtnt z19.h, z31.s\n" + ".inst 0x453042f7 // sqxtnb z23.h, z23.s\n" + ".inst 0x453047d2 // sqxtnt z18.h, z30.s\n" + ".inst 0x453042d6 // sqxtnb z22.h, z22.s\n" + ".inst 0x453047b1 // sqxtnt z17.h, z29.s\n" + ".inst 0x453042b5 // sqxtnb z21.h, z21.s\n" + ".inst 0x45304790 // sqxtnt z16.h, z28.s\n" + ".inst 0x45304294 // sqxtnb z20.h, z20.s\n" + ".inst 0x45304777 // sqxtnt z23.h, z27.s\n" + ".inst 0x45304756 // sqxtnt z22.h, z26.s\n" + ".inst 0x45284273 // sqxtnb z19.b, z19.h\n" + ".inst 0x45304735 // sqxtnt z21.h, z25.s\n" + ".inst 0x45284252 // sqxtnb z18.b, z18.h\n" + ".inst 0x45304714 // sqxtnt z20.h, z24.s\n" + ".inst 0x45284231 // sqxtnb z17.b, z17.h\n" + ".inst 0x45284210 // sqxtnb z16.b, z16.h\n" + ".inst 0x452846f3 // sqxtnt z19.b, z23.h\n" + ".inst 0x452846d2 // sqxtnt z18.b, z22.h\n" + ".inst 0x452846b1 // sqxtnt z17.b, z21.h\n" + ".inst 0x45284690 // sqxtnt z16.b, z20.h\n" + "st1b { z19.b }, p3, [%x[outptr], x21]\n" + "incb x21, ALL, MUL #4\n" + "st1b { z18.b }, p2, [%x[outptr], x20]\n" + "incb x20, ALL, MUL #4\n" + "st1b { z17.b }, p1, [%x[outptr], x8]\n" + "incb x8, ALL, MUL #4\n" + "st1b { z16.b }, p0, [%x[outptr], x7]\n" + "incb x7, ALL, MUL #4\n" + "whilelt p0.b, x7, %x[n_channels]\n" "b.ne 1b\n" "7:" // Single vector of channels - "whilelt p3.b, x9, %x[n_channels]\n" + "whilelt p3.b, x21, %x[n_channels]\n" "b.eq 14f\n" "8:" // Single vector of channels: Loop - "lsr x25, %x[n_valid_cells], #0x2\n" - "mov z6.b, #0x80\n" - "mov x24, %x[inptrs]\n" - "cbz x25, 11f\n" - "ldp x23, x22, [x24, #0]\n" - "ldp x21, x20, [x24, #0x10]\n" - "subs x25, x25, #0x1\n" - "add x24, x24, #0x20\n" - "ld1b { z2.b }, p3/Z, [x23, x9]\n" - "ld1b { z1.b }, p3/Z, [x22, x9]\n" - "ld1b { z23.b }, p3/Z, [x21, x9]\n" - "ld1b { z0.b }, p3/Z, [x20, x9]\n" + "lsr x6, %x[n_valid_cells], #0x2\n" + "mov z8.b, #0x80\n" + "mov x5, %x[inptrs]\n" + "cbz x6, 11f\n" + "ldp x4, x3, [x5, #0]\n" + "ldp x2, x1, [x5, #0x10]\n" + "subs x6, x6, #0x1\n" + "add x5, x5, #0x20\n" + "ld1b { z2.b }, p3/Z, [x4, x21]\n" + "ld1b { z1.b }, p3/Z, [x3, x21]\n" + "ld1b { z23.b }, p3/Z, [x2, x21]\n" + "ld1b { z0.b }, p3/Z, [x1, x21]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop "movprfx z16, z2\n smax z16.b, p4/M, z16.b, z1.b\n" "movprfx z17, z23\n smax z17.b, p4/M, z17.b, z0.b\n" - "ldp x23, x22, [x24, #0]\n" - "ldp x21, x20, [x24, #0x10]\n" - "subs x25, x25, #0x1\n" - "add x24, x24, #0x20\n" + "ldp x4, x3, [x5, #0]\n" + "ldp x2, x1, [x5, #0x10]\n" + "subs x6, x6, #0x1\n" + "add x5, x5, #0x20\n" "smax z16.b, p4/M, z16.b, z17.b\n" - "ld1b { z2.b }, p3/Z, [x23, x9]\n" - "ld1b { z1.b }, p3/Z, [x22, x9]\n" - "ld1b { z23.b }, p3/Z, [x21, x9]\n" - "ld1b { z0.b }, p3/Z, [x20, x9]\n" - "smax z6.b, p4/M, z6.b, z16.b\n" + "ld1b { z2.b }, p3/Z, [x4, x21]\n" + "ld1b { z1.b }, p3/Z, [x3, x21]\n" + "ld1b { z23.b }, p3/Z, [x2, x21]\n" + "ld1b { z0.b }, p3/Z, [x1, x21]\n" + "smax z8.b, p4/M, z8.b, z16.b\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail "movprfx z16, z2\n smax z16.b, p4/M, z16.b, z1.b\n" "movprfx z17, z23\n smax z17.b, p4/M, z17.b, z0.b\n" "smax z16.b, p4/M, z16.b, z17.b\n" - "smax z6.b, p4/M, z6.b, z16.b\n" + "smax z8.b, p4/M, z8.b, z16.b\n" "11:" // Single vector of channels: Loop: After loop - "ands x21, %x[n_valid_cells], #0x3\n" + "ands x2, %x[n_valid_cells], #0x3\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop - "ldr x20, [x24], #0x8\n" - "subs x21, x21, #0x1\n" - "ld1b { z16.b }, p3/Z, [x20, x9]\n" - "smax z6.b, p4/M, z6.b, z16.b\n" + "ldr x1, [x5], #0x8\n" + "subs x2, x2, #0x1\n" + "ld1b { z16.b }, p3/Z, [x1, x21]\n" + "smax z8.b, p4/M, z8.b, z16.b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" - ".inst 0x4508a0d1 // sshllb z17.h, z6.b, #0\n" - ".inst 0x4508a4da // sshllt z26.h, z6.b, #0\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1rw { z16.s }, p4/Z, [x21]\n" - "ld1rw { z25.s }, p4/Z, [x20]\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "ld1rw { z24.s }, p4/Z, [x21]\n" - "ld1rw { z23.s }, p4/Z, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" - "mov z22.s, #0x7f\n" - "ld1rw { z21.s }, p4/Z, [x20]\n" - "neg z16.s, p4/M, z16.s\n" - "not z20.s, p4/M, z22.s\n" - ".inst 0x45914213 // saddwb z19.s, z16.s, z17.h\n" - ".inst 0x45914611 // saddwt z17.s, z16.s, z17.h\n" - ".inst 0x459a4212 // saddwb z18.s, z16.s, z26.h\n" - ".inst 0x459a4610 // saddwt z16.s, z16.s, z26.h\n" - ".inst 0x44829333 // srshl z19.s, p4/M, z19.s, z25.s\n" - ".inst 0x44829331 // srshl z17.s, p4/M, z17.s, z25.s\n" - ".inst 0x44829332 // srshl z18.s, p4/M, z18.s, z25.s\n" - ".inst 0x44829330 // srshl z16.s, p4/M, z16.s, z25.s\n" - ".inst 0x04b87673 // sqrdmulh z19.s, z19.s, z24.s\n" - ".inst 0x04b87631 // sqrdmulh z17.s, z17.s, z24.s\n" - ".inst 0x04b87652 // sqrdmulh z18.s, z18.s, z24.s\n" - ".inst 0x04b87610 // sqrdmulh z16.s, z16.s, z24.s\n" + "add x2, %x[quant_params], %[offsetof_qp_input_offset]\n" + ".inst 0x4508a111 // sshllb z17.h, z8.b, #0\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + ".inst 0x4508a518 // sshllt z24.h, z8.b, #0\n" + "ld1rw { z18.s }, p4/Z, [x2]\n" + "ld1rw { z23.s }, p4/Z, [x1]\n" + "add x2, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x1, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "ld1rw { z22.s }, p4/Z, [x2]\n" + "ld1rw { z21.s }, p4/Z, [x1]\n" + "add x1, %x[quant_params], %[offsetof_qp_output_offset]\n" + "ld1rw { z20.s }, p4/Z, [x1]\n" + "neg z18.s, p4/M, z18.s\n" + ".inst 0x45914250 // saddwb z16.s, z18.s, z17.h\n" + ".inst 0x45914653 // saddwt z19.s, z18.s, z17.h\n" + ".inst 0x45984251 // saddwb z17.s, z18.s, z24.h\n" + ".inst 0x448292f0 // srshl z16.s, p4/M, z16.s, z23.s\n" + ".inst 0x45984652 // saddwt z18.s, z18.s, z24.h\n" ".inst 0x448292f3 // srshl z19.s, p4/M, z19.s, z23.s\n" + ".inst 0x04b67610 // sqrdmulh z16.s, z16.s, z22.s\n" ".inst 0x448292f1 // srshl z17.s, p4/M, z17.s, z23.s\n" + ".inst 0x04b67673 // sqrdmulh z19.s, z19.s, z22.s\n" ".inst 0x448292f2 // srshl z18.s, p4/M, z18.s, z23.s\n" - ".inst 0x448292f0 // srshl z16.s, p4/M, z16.s, z23.s\n" - "add z19.s, z19.s, z21.s\n" - "add z17.s, z17.s, z21.s\n" - "add z18.s, z18.s, z21.s\n" - "add z16.s, z16.s, z21.s\n" - "smax z19.s, p4/M, z19.s, z20.s\n" - "smax z17.s, p4/M, z17.s, z20.s\n" - "smax z18.s, p4/M, z18.s, z20.s\n" - "smax z16.s, p4/M, z16.s, z20.s\n" - "smin z19.s, p4/M, z19.s, z22.s\n" - "smin z17.s, p4/M, z17.s, z22.s\n" - "smin z18.s, p4/M, z18.s, z22.s\n" - "smin z16.s, p4/M, z16.s, z22.s\n" - "trn1 z17.h, z19.h, z17.h\n" - "trn1 z16.h, z18.h, z16.h\n" - "trn1 z16.b, z17.b, z16.b\n" - "st1b { z16.b }, p3, [%x[outptr], x9]\n" - "incb x9\n" - "whilelt p3.b, x9, %x[n_channels]\n" + ".inst 0x448292b0 // srshl z16.s, p4/M, z16.s, z21.s\n" + ".inst 0x04b67631 // sqrdmulh z17.s, z17.s, z22.s\n" + ".inst 0x448292b3 // srshl z19.s, p4/M, z19.s, z21.s\n" + ".inst 0x04b67652 // sqrdmulh z18.s, z18.s, z22.s\n" + ".inst 0x448292b1 // srshl z17.s, p4/M, z17.s, z21.s\n" + "add z16.s, z16.s, z20.s\n" + ".inst 0x448292b2 // srshl z18.s, p4/M, z18.s, z21.s\n" + "add z19.s, z19.s, z20.s\n" + "add z17.s, z17.s, z20.s\n" + ".inst 0x45304210 // sqxtnb z16.h, z16.s\n" + "add z18.s, z18.s, z20.s\n" + ".inst 0x45304670 // sqxtnt z16.h, z19.s\n" + ".inst 0x45304231 // sqxtnb z17.h, z17.s\n" + ".inst 0x45284210 // sqxtnb z16.b, z16.h\n" + ".inst 0x45304651 // sqxtnt z17.h, z18.s\n" + ".inst 0x45284630 // sqxtnt z16.b, z17.h\n" + "st1b { z16.b }, p3, [%x[outptr], x21]\n" + "incb x21\n" + "whilelt p3.b, x21, %x[n_channels]\n" "b.ne 8b\n" "14:" // End : : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [n_valid_cells] "r" (n_valid_cells), [offsetof_qp_input_offset] "I" (offsetof(Requantize32, input_offset)), [offsetof_qp_output_offset] "I" (offsetof(Requantize32, output_offset)), [offsetof_qp_per_layer_left_shift] "I" (offsetof(Requantize32, per_layer_left_shift)), [offsetof_qp_per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [offsetof_qp_per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [outptr] "r" (outptr), [quant_params] "r" (&qp) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "x20", "x21", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); }