Add new operator AddMulAdd for Neon™ backend for Float/Quantized types
This is a fused operator that merges Add + Mul + Add [+ Relu-based-Activation] layers and have an intermediate output after the first Add. It's supported for FP16/32/QASYMM8/QASYMM8_SIGNED data types.
The subsequent Add and Mul are intended for scaling and the coefficients only have one dimension (per channel).
The inputs are
- input1 : nD tensor [X, Y, Z, W, ..]
- input2 : nD tensor [X, Y, Z, W, ..]
- add_coef : 1D tensor [X]
- mul_coef : 1D tensor [X]
The outputs are
- out1 : nD tensor (intermediate output) [X, Y, Z, W, ..]
- out2 : nD tensor (final output) [X, Y, Z, W, ..]
The operation can be summarized as follows:
out1 <- input1 + input2
out2 <- Act(out1 * mul_coef + add_coef)
The activation function can be Identity, Relu, Bounded Relu or Lower/Upper Bounded Relu. The intermediate output can be skipped by providing a nullptr.
The reason of providing this operator is to be able to fuse in case of Residual network patterns and save computations by reducing memory back and forward.
Resolves: COMPMID-5463
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: I8ef577aa623b036e9a9f655cc088493fd19a6109
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9055
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/cpu/kernels/addmuladd/generic/neon/qasymm8.cpp b/src/cpu/kernels/addmuladd/generic/neon/qasymm8.cpp
new file mode 100644
index 0000000..2c0ad70
--- /dev/null
+++ b/src/cpu/kernels/addmuladd/generic/neon/qasymm8.cpp
@@ -0,0 +1,834 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/QuantizationInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Window.h"
+
+#include <cstddef>
+#include <cstdint>
+#include <limits>
+
+#ifdef __aarch64__
+namespace
+{
+void a64_add_bn_clamp_direct_u8_fp32_2x16(
+ uint8_t *out, size_t out_stride,
+ uint8_t *out_direct, size_t out_direct_stride,
+ const uint8_t *in0, size_t in0_stride,
+ const uint8_t *in1, size_t in1_stride,
+ const float *bn_mul,
+ const float *bn_add,
+ const uint8_t minval,
+ const uint8_t maxval,
+ int32_t out_zeropt, float out_scale,
+ int32_t out_direct_zeropt, float out_direct_scale,
+ int32_t in0_zeropt, float in0_scale,
+ int32_t in1_zeropt, float in1_scale,
+ size_t width, size_t height)
+{
+ float scales[4] = { in0_scale, in1_scale, 1.0f / out_scale, 1.0f / out_direct_scale };
+ struct KernelArgs
+ {
+ const float *scales;
+ int32_t in0_zeropt;
+ int32_t in1_zeropt;
+ int32_t out_zeropt;
+ int32_t out_direct_zeropt;
+ int32_t minval;
+ int32_t maxval;
+ } ka;
+ ka.scales = scales;
+ ka.in0_zeropt = in0_zeropt;
+ ka.in1_zeropt = in1_zeropt;
+ ka.out_zeropt = out_zeropt;
+ ka.out_direct_zeropt = out_direct_zeropt;
+ ka.minval = minval;
+ ka.maxval = maxval;
+
+ __asm__ __volatile__(
+ "ldr x20, [%x[args_ptr], %[offsetof_scales]]\n"
+ "ld1 { v0.4s }, [x20]\n"
+ "cmp %x[width], #0x10\n"
+ "blt 5f\n"
+ "1:" // Column loop
+ "ldr q24, [%x[bn_mul], #0x0]\n"
+ "ldr q25, [%x[bn_mul], #0x10]\n"
+ "mov x23, %x[height]\n"
+ "mov x12, %x[in0]\n"
+ "ldr q26, [%x[bn_mul], #0x20]\n"
+ "ldr q27, [%x[bn_mul], #0x30]\n"
+ "mov x11, %x[in1]\n"
+ "mov x10, %x[out]\n"
+ "ldr q28, [%x[bn_add], #0x0]\n"
+ "ldr q29, [%x[bn_add], #0x10]\n"
+ "mov x9, %x[out_direct]\n"
+ "add %x[bn_mul], %x[bn_mul], #0x40\n"
+ "ldr q30, [%x[bn_add], #0x20]\n"
+ "ldr q31, [%x[bn_add], #0x30]\n"
+ "add %x[bn_add], %x[bn_add], #0x40\n"
+ "2:" // Row loop
+ "mov x28, x12\n"
+ "ldr d4, [x28, #0x0]\n"
+ "ldr d3, [x28, #0x8]\n"
+ "add x21, x28, %x[in0_stride]\n"
+ "mov x27, x11\n"
+ "ldr d13, [x27, #0x0]\n"
+ "ldr d12, [x27, #0x8]\n"
+ "cmp x23, #0x2\n"
+ "add x12, x21, %x[in0_stride]\n"
+ "csel x21, x21, x28, GE\n"
+ "ldr d2, [x21, #0x0]\n"
+ "ldr d11, [x21, #0x8]\n"
+ "add x20, x27, %x[in1_stride]\n"
+ "add x11, x20, %x[in1_stride]\n"
+ "ldr w21, [%x[args_ptr], %[offsetof_in0_zeropt]]\n"
+ "ushll v4.8h, v4.8b, #0x0\n"
+ "csel x20, x20, x27, GE\n"
+ "ldr d10, [x20, #0x0]\n"
+ "ldr d9, [x20, #0x8]\n"
+ "ushll v3.8h, v3.8b, #0x0\n"
+ "ushll v2.8h, v2.8b, #0x0\n"
+ "ushll v11.8h, v11.8b, #0x0\n"
+ "ldr w20, [%x[args_ptr], %[offsetof_in1_zeropt]]\n"
+ "mov x26, x10\n"
+ "dup v16.8h, w21\n"
+ "ushll v13.8h, v13.8b, #0x0\n"
+ "mov x25, x9\n"
+ "add x24, x26, %x[out_stride]\n"
+ "ushll v12.8h, v12.8b, #0x0\n"
+ "ushll v10.8h, v10.8b, #0x0\n"
+ "add x22, x25, %x[out_direct_stride]\n"
+ "add x10, x24, %x[out_stride]\n"
+ "ushll v9.8h, v9.8b, #0x0\n"
+ "ssubl v1.4s, v4.4h, v16.4h\n"
+ "add x9, x22, %x[out_direct_stride]\n"
+ "csel x24, x24, x26, GE\n"
+ "ssubl2 v4.4s, v4.8h, v16.8h\n"
+ "ssubl v23.4s, v3.4h, v16.4h\n"
+ "csel x22, x22, x25, GE\n"
+ "ssubl2 v3.4s, v3.8h, v16.8h\n"
+ "ssubl v22.4s, v2.4h, v16.4h\n"
+ "ssubl2 v2.4s, v2.8h, v16.8h\n"
+ "ssubl v21.4s, v11.4h, v16.4h\n"
+ "ssubl2 v11.4s, v11.8h, v16.8h\n"
+ "dup v20.8h, w20\n"
+ "ssubl v19.4s, v13.4h, v20.4h\n"
+ "ssubl2 v13.4s, v13.8h, v20.8h\n"
+ "ssubl v18.4s, v12.4h, v20.4h\n"
+ "ssubl2 v12.4s, v12.8h, v20.8h\n"
+ "ssubl v17.4s, v10.4h, v20.4h\n"
+ "ssubl2 v10.4s, v10.8h, v20.8h\n"
+ "ssubl v16.4s, v9.4h, v20.4h\n"
+ "ssubl2 v9.4s, v9.8h, v20.8h\n"
+ "scvtf v8.4s, v1.4s\n"
+ "scvtf v7.4s, v4.4s\n"
+ "scvtf v6.4s, v23.4s\n"
+ "scvtf v5.4s, v3.4s\n"
+ "scvtf v4.4s, v22.4s\n"
+ "scvtf v3.4s, v2.4s\n"
+ "scvtf v2.4s, v21.4s\n"
+ "scvtf v1.4s, v11.4s\n"
+ "scvtf v19.4s, v19.4s\n"
+ "fmul v8.4s, v8.4s, v0.s[0]\n"
+ "fmla v8.4s, v19.4s, v0.s[1]\n"
+ "scvtf v13.4s, v13.4s\n"
+ "fmul v7.4s, v7.4s, v0.s[0]\n"
+ "fmla v7.4s, v13.4s, v0.s[1]\n"
+ "scvtf v18.4s, v18.4s\n"
+ "fmul v6.4s, v6.4s, v0.s[0]\n"
+ "fmla v6.4s, v18.4s, v0.s[1]\n"
+ "scvtf v12.4s, v12.4s\n"
+ "fmul v5.4s, v5.4s, v0.s[0]\n"
+ "fmla v5.4s, v12.4s, v0.s[1]\n"
+ "scvtf v17.4s, v17.4s\n"
+ "fmul v4.4s, v4.4s, v0.s[0]\n"
+ "fmla v4.4s, v17.4s, v0.s[1]\n"
+ "scvtf v10.4s, v10.4s\n"
+ "fmul v3.4s, v3.4s, v0.s[0]\n"
+ "fmla v3.4s, v10.4s, v0.s[1]\n"
+ "scvtf v16.4s, v16.4s\n"
+ "fmul v2.4s, v2.4s, v0.s[0]\n"
+ "fmla v2.4s, v16.4s, v0.s[1]\n"
+ "scvtf v9.4s, v9.4s\n"
+ "fmul v1.4s, v1.4s, v0.s[0]\n"
+ "fmla v1.4s, v9.4s, v0.s[1]\n"
+ "cbz %x[out_direct], 3f\n"
+ "fmul v23.4s, v8.4s, v0.s[3]\n"
+ "fmul v22.4s, v7.4s, v0.s[3]\n"
+ "ldr w20, [%x[args_ptr], %[offsetof_out_direct_zeropt]]\n"
+ "fmul v21.4s, v6.4s, v0.s[3]\n"
+ "fmul v20.4s, v5.4s, v0.s[3]\n"
+ "fmul v19.4s, v4.4s, v0.s[3]\n"
+ "fmul v18.4s, v3.4s, v0.s[3]\n"
+ "fmul v16.4s, v2.4s, v0.s[3]\n"
+ "fmul v17.4s, v1.4s, v0.s[3]\n"
+ "fcvtas v23.4s, v23.4s\n"
+ "fcvtas v22.4s, v22.4s\n"
+ "fcvtas v21.4s, v21.4s\n"
+ "fcvtas v20.4s, v20.4s\n"
+ "fcvtas v19.4s, v19.4s\n"
+ "fcvtas v18.4s, v18.4s\n"
+ "fcvtas v16.4s, v16.4s\n"
+ "fcvtas v17.4s, v17.4s\n"
+ "uzp1 v22.8h, v23.8h, v22.8h\n"
+ "uzp1 v20.8h, v21.8h, v20.8h\n"
+ "uzp1 v18.8h, v19.8h, v18.8h\n"
+ "uzp1 v17.8h, v16.8h, v17.8h\n"
+ "dup v16.8h, w20\n"
+ "add v22.8h, v22.8h, v16.8h\n"
+ "add v20.8h, v20.8h, v16.8h\n"
+ "add v18.8h, v18.8h, v16.8h\n"
+ "add v17.8h, v17.8h, v16.8h\n"
+ "movi v16.8h, #0xff\n"
+ "smin v22.8h, v22.8h, v16.8h\n"
+ "smin v20.8h, v20.8h, v16.8h\n"
+ "smin v18.8h, v18.8h, v16.8h\n"
+ "smin v17.8h, v17.8h, v16.8h\n"
+ "movi v16.8h, #0x0\n"
+ "smax v22.8h, v22.8h, v16.8h\n"
+ "smax v20.8h, v20.8h, v16.8h\n"
+ "smax v18.8h, v18.8h, v16.8h\n"
+ "smax v17.8h, v17.8h, v16.8h\n"
+ "xtn v22.8b, v22.8h\n"
+ "str d22, [x25, #0x0]\n"
+ "xtn v20.8b, v20.8h\n"
+ "xtn v18.8b, v18.8h\n"
+ "str d20, [x25, #0x8]\n"
+ "xtn v17.8b, v17.8h\n"
+ "str d18, [x22, #0x0]\n"
+ "str d17, [x22, #0x8]\n"
+ "3:" // Main loop: No direct output
+ "mov v19.16b, v28.16b\n"
+ "mov v13.16b, v29.16b\n"
+ "fmla v19.4s, v8.4s, v24.4s\n"
+ "ldr w22, [%x[args_ptr], %[offsetof_out_zeropt]]\n"
+ "mov v18.16b, v30.16b\n"
+ "mov v12.16b, v31.16b\n"
+ "fmla v13.4s, v7.4s, v25.4s\n"
+ "ldr w21, [%x[args_ptr], %[offsetof_maxval]]\n"
+ "mov v17.16b, v28.16b\n"
+ "mov v10.16b, v29.16b\n"
+ "fmla v18.4s, v6.4s, v26.4s\n"
+ "ldr w20, [%x[args_ptr], %[offsetof_minval]]\n"
+ "mov v16.16b, v30.16b\n"
+ "mov v9.16b, v31.16b\n"
+ "fmla v12.4s, v5.4s, v27.4s\n"
+ "subs x23, x23, #0x2\n"
+ "fmla v17.4s, v4.4s, v24.4s\n"
+ "fmla v10.4s, v3.4s, v25.4s\n"
+ "fmul v8.4s, v19.4s, v0.s[2]\n"
+ "fmla v16.4s, v2.4s, v26.4s\n"
+ "fmla v9.4s, v1.4s, v27.4s\n"
+ "fmul v7.4s, v13.4s, v0.s[2]\n"
+ "fmul v6.4s, v18.4s, v0.s[2]\n"
+ "fmul v5.4s, v12.4s, v0.s[2]\n"
+ "fmul v4.4s, v17.4s, v0.s[2]\n"
+ "fmul v3.4s, v10.4s, v0.s[2]\n"
+ "fmul v2.4s, v16.4s, v0.s[2]\n"
+ "fmul v1.4s, v9.4s, v0.s[2]\n"
+ "fcvtas v8.4s, v8.4s\n"
+ "fcvtas v7.4s, v7.4s\n"
+ "fcvtas v6.4s, v6.4s\n"
+ "fcvtas v5.4s, v5.4s\n"
+ "fcvtas v4.4s, v4.4s\n"
+ "fcvtas v3.4s, v3.4s\n"
+ "fcvtas v2.4s, v2.4s\n"
+ "fcvtas v1.4s, v1.4s\n"
+ "uzp1 v7.8h, v8.8h, v7.8h\n"
+ "uzp1 v5.8h, v6.8h, v5.8h\n"
+ "uzp1 v3.8h, v4.8h, v3.8h\n"
+ "uzp1 v1.8h, v2.8h, v1.8h\n"
+ "dup v16.8h, w22\n"
+ "add v7.8h, v7.8h, v16.8h\n"
+ "add v5.8h, v5.8h, v16.8h\n"
+ "add v3.8h, v3.8h, v16.8h\n"
+ "add v1.8h, v1.8h, v16.8h\n"
+ "dup v16.8h, w21\n"
+ "smin v7.8h, v7.8h, v16.8h\n"
+ "smin v5.8h, v5.8h, v16.8h\n"
+ "smin v3.8h, v3.8h, v16.8h\n"
+ "smin v1.8h, v1.8h, v16.8h\n"
+ "dup v16.8h, w20\n"
+ "smax v7.8h, v7.8h, v16.8h\n"
+ "smax v5.8h, v5.8h, v16.8h\n"
+ "smax v3.8h, v3.8h, v16.8h\n"
+ "smax v1.8h, v1.8h, v16.8h\n"
+ "xtn v7.8b, v7.8h\n"
+ "str d7, [x26, #0x0]\n"
+ "xtn v5.8b, v5.8h\n"
+ "xtn v3.8b, v3.8h\n"
+ "str d5, [x26, #0x8]\n"
+ "xtn v1.8b, v1.8h\n"
+ "str d3, [x24, #0x0]\n"
+ "str d1, [x24, #0x8]\n"
+ "bgt 2b\n"
+ "add %x[in0], %x[in0], #0x10\n"
+ "add %x[in1], %x[in1], #0x10\n"
+ "add %x[out], %x[out], #0x10\n"
+ "cbz %x[out_direct], 4f\n"
+ "add %x[out_direct], %x[out_direct], #0x10\n"
+ "4:" // No direct pointer update
+ "sub %x[width], %x[width], #0x10\n"
+ "cmp %x[width], #0x10\n"
+ "bge 1b\n"
+ "cbz %x[width], 32f\n"
+ "5:" // main loop skip
+ "ldr q24, [%x[bn_mul], #0x0]\n"
+ "ldr q25, [%x[bn_mul], #0x10]\n"
+ "mov x23, %x[height]\n"
+ "mov x12, %x[in0]\n"
+ "ldr q26, [%x[bn_mul], #0x20]\n"
+ "ldr q27, [%x[bn_mul], #0x30]\n"
+ "mov x11, %x[in1]\n"
+ "mov x10, %x[out]\n"
+ "ldr q28, [%x[bn_add], #0x0]\n"
+ "ldr q29, [%x[bn_add], #0x10]\n"
+ "mov x9, %x[out_direct]\n"
+ "add %x[bn_mul], %x[bn_mul], #0x40\n"
+ "ldr q30, [%x[bn_add], #0x20]\n"
+ "ldr q31, [%x[bn_add], #0x30]\n"
+ "add %x[bn_add], %x[bn_add], #0x40\n"
+ "6:" // tail loop: Row loop
+ "mov x28, x12\n"
+ "mov x27, x11\n"
+ "mov x26, x10\n"
+ "mov x25, x9\n"
+ "add x21, x28, %x[in0_stride]\n"
+ "add x20, x27, %x[in1_stride]\n"
+ "add x24, x26, %x[out_stride]\n"
+ "add x22, x25, %x[out_direct_stride]\n"
+ "cmp x23, #0x2\n"
+ "add x12, x21, %x[in0_stride]\n"
+ "add x11, x20, %x[in1_stride]\n"
+ "add x10, x24, %x[out_stride]\n"
+ "add x9, x22, %x[out_direct_stride]\n"
+ "csel x21, x21, x28, GE\n"
+ "csel x20, x20, x27, GE\n"
+ "csel x24, x24, x26, GE\n"
+ "csel x22, x22, x25, GE\n"
+ "tbz %x[width], #3, 10f\n"
+ "ldr d4, [x28, #0x0]\n"
+ "ldr d13, [x27, #0x0]\n"
+ "add x28, x28, #0x8\n"
+ "add x27, x27, #0x8\n"
+ "ldr d2, [x21, #0x0]\n"
+ "ldr d10, [x20, #0x0]\n"
+ "add x21, x21, #0x8\n"
+ "add x20, x20, #0x8\n"
+ "tbz %x[width], #2, 8f\n"
+ "ldr s3, [x28], #0x4\n"
+ "ldr s12, [x27], #0x4\n"
+ "ldr s11, [x21], #0x4\n"
+ "ldr s9, [x20], #0x4\n"
+ "tbz %x[width], #1, 7f\n"
+ "ld1 { v3.h }[2], [x28], #0x2\n"
+ "ld1 { v12.h }[2], [x27], #0x2\n"
+ "ld1 { v11.h }[2], [x21], #0x2\n"
+ "ld1 { v9.h }[2], [x20], #0x2\n"
+ "tbz %x[width], #0, 14f\n"
+ "ld1 { v3.b }[6], [x28], #0x1\n"
+ "ld1 { v12.b }[6], [x27], #0x1\n"
+ "ld1 { v11.b }[6], [x21], #0x1\n"
+ "ld1 { v9.b }[6], [x20], #0x1\n"
+ "b 14f\n"
+ "7:" // tail loop: unique 1: partial_0_12
+ "tbz %x[width], #0, 14f\n"
+ "ld1 { v3.b }[4], [x28], #0x1\n"
+ "ld1 { v12.b }[4], [x27], #0x1\n"
+ "ld1 { v11.b }[4], [x21], #0x1\n"
+ "ld1 { v9.b }[4], [x20], #0x1\n"
+ "b 14f\n"
+ "8:" // tail loop: unique 1: partial_1_8
+ "tbz %x[width], #1, 9f\n"
+ "ldr h3, [x28], #0x2\n"
+ "ldr h12, [x27], #0x2\n"
+ "ldr h11, [x21], #0x2\n"
+ "ldr h9, [x20], #0x2\n"
+ "tbz %x[width], #0, 14f\n"
+ "ld1 { v3.b }[2], [x28], #0x1\n"
+ "ld1 { v12.b }[2], [x27], #0x1\n"
+ "ld1 { v11.b }[2], [x21], #0x1\n"
+ "ld1 { v9.b }[2], [x20], #0x1\n"
+ "b 14f\n"
+ "9:" // tail loop: unique 1: partial_0_8
+ "tbz %x[width], #0, 14f\n"
+ "ldr b3, [x28], #0x1\n"
+ "ldr b12, [x27], #0x1\n"
+ "ldr b11, [x21], #0x1\n"
+ "ldr b9, [x20], #0x1\n"
+ "b 14f\n"
+ "10:" // tail loop: unique 1: partial_2_0
+ "tbz %x[width], #2, 12f\n"
+ "ldr s4, [x28], #0x4\n"
+ "ldr s13, [x27], #0x4\n"
+ "ldr s2, [x21], #0x4\n"
+ "ldr s10, [x20], #0x4\n"
+ "tbz %x[width], #1, 11f\n"
+ "ld1 { v4.h }[2], [x28], #0x2\n"
+ "ld1 { v13.h }[2], [x27], #0x2\n"
+ "ld1 { v2.h }[2], [x21], #0x2\n"
+ "ld1 { v10.h }[2], [x20], #0x2\n"
+ "tbz %x[width], #0, 14f\n"
+ "ld1 { v4.b }[6], [x28], #0x1\n"
+ "ld1 { v13.b }[6], [x27], #0x1\n"
+ "ld1 { v2.b }[6], [x21], #0x1\n"
+ "ld1 { v10.b }[6], [x20], #0x1\n"
+ "b 14f\n"
+ "11:" // tail loop: unique 1: partial_0_4
+ "tbz %x[width], #0, 14f\n"
+ "ld1 { v4.b }[4], [x28], #0x1\n"
+ "ld1 { v13.b }[4], [x27], #0x1\n"
+ "ld1 { v2.b }[4], [x21], #0x1\n"
+ "ld1 { v10.b }[4], [x20], #0x1\n"
+ "b 14f\n"
+ "12:" // tail loop: unique 1: partial_1_0
+ "tbz %x[width], #1, 13f\n"
+ "ldr h4, [x28], #0x2\n"
+ "ldr h13, [x27], #0x2\n"
+ "ldr h2, [x21], #0x2\n"
+ "ldr h10, [x20], #0x2\n"
+ "tbz %x[width], #0, 14f\n"
+ "ld1 { v4.b }[2], [x28], #0x1\n"
+ "ld1 { v13.b }[2], [x27], #0x1\n"
+ "ld1 { v2.b }[2], [x21], #0x1\n"
+ "ld1 { v10.b }[2], [x20], #0x1\n"
+ "b 14f\n"
+ "13:" // tail loop: unique 1: partial_0_0
+ "ldr b4, [x28], #0x1\n"
+ "ldr b13, [x27], #0x1\n"
+ "ldr b2, [x21], #0x1\n"
+ "ldr b10, [x20], #0x1\n"
+ "14:" // tail loop: unique 1: Done
+ "ldr w21, [%x[args_ptr], %[offsetof_in0_zeropt]]\n"
+ "ushll v4.8h, v4.8b, #0x0\n"
+ "ushll v3.8h, v3.8b, #0x0\n"
+ "ldr w20, [%x[args_ptr], %[offsetof_in1_zeropt]]\n"
+ "ushll v2.8h, v2.8b, #0x0\n"
+ "ushll v11.8h, v11.8b, #0x0\n"
+ "dup v16.8h, w21\n"
+ "ushll v13.8h, v13.8b, #0x0\n"
+ "ushll v12.8h, v12.8b, #0x0\n"
+ "ushll v10.8h, v10.8b, #0x0\n"
+ "ushll v9.8h, v9.8b, #0x0\n"
+ "ssubl v1.4s, v4.4h, v16.4h\n"
+ "ssubl2 v4.4s, v4.8h, v16.8h\n"
+ "ssubl v23.4s, v3.4h, v16.4h\n"
+ "ssubl2 v3.4s, v3.8h, v16.8h\n"
+ "ssubl v22.4s, v2.4h, v16.4h\n"
+ "ssubl2 v2.4s, v2.8h, v16.8h\n"
+ "ssubl v21.4s, v11.4h, v16.4h\n"
+ "ssubl2 v11.4s, v11.8h, v16.8h\n"
+ "dup v20.8h, w20\n"
+ "ssubl v19.4s, v13.4h, v20.4h\n"
+ "ssubl2 v13.4s, v13.8h, v20.8h\n"
+ "ssubl v18.4s, v12.4h, v20.4h\n"
+ "ssubl2 v12.4s, v12.8h, v20.8h\n"
+ "ssubl v17.4s, v10.4h, v20.4h\n"
+ "ssubl2 v10.4s, v10.8h, v20.8h\n"
+ "ssubl v16.4s, v9.4h, v20.4h\n"
+ "ssubl2 v9.4s, v9.8h, v20.8h\n"
+ "scvtf v8.4s, v1.4s\n"
+ "scvtf v7.4s, v4.4s\n"
+ "scvtf v6.4s, v23.4s\n"
+ "scvtf v5.4s, v3.4s\n"
+ "scvtf v4.4s, v22.4s\n"
+ "scvtf v3.4s, v2.4s\n"
+ "scvtf v2.4s, v21.4s\n"
+ "scvtf v1.4s, v11.4s\n"
+ "scvtf v19.4s, v19.4s\n"
+ "fmul v8.4s, v8.4s, v0.s[0]\n"
+ "fmla v8.4s, v19.4s, v0.s[1]\n"
+ "scvtf v13.4s, v13.4s\n"
+ "fmul v7.4s, v7.4s, v0.s[0]\n"
+ "fmla v7.4s, v13.4s, v0.s[1]\n"
+ "scvtf v18.4s, v18.4s\n"
+ "fmul v6.4s, v6.4s, v0.s[0]\n"
+ "fmla v6.4s, v18.4s, v0.s[1]\n"
+ "scvtf v12.4s, v12.4s\n"
+ "fmul v5.4s, v5.4s, v0.s[0]\n"
+ "fmla v5.4s, v12.4s, v0.s[1]\n"
+ "scvtf v17.4s, v17.4s\n"
+ "fmul v4.4s, v4.4s, v0.s[0]\n"
+ "fmla v4.4s, v17.4s, v0.s[1]\n"
+ "scvtf v10.4s, v10.4s\n"
+ "fmul v3.4s, v3.4s, v0.s[0]\n"
+ "fmla v3.4s, v10.4s, v0.s[1]\n"
+ "scvtf v16.4s, v16.4s\n"
+ "fmul v2.4s, v2.4s, v0.s[0]\n"
+ "fmla v2.4s, v16.4s, v0.s[1]\n"
+ "scvtf v9.4s, v9.4s\n"
+ "fmul v1.4s, v1.4s, v0.s[0]\n"
+ "fmla v1.4s, v9.4s, v0.s[1]\n"
+ "cbz %x[out_direct], 23f\n"
+ "fmul v23.4s, v8.4s, v0.s[3]\n"
+ "fmul v22.4s, v7.4s, v0.s[3]\n"
+ "ldr w20, [%x[args_ptr], %[offsetof_out_direct_zeropt]]\n"
+ "fmul v21.4s, v6.4s, v0.s[3]\n"
+ "fmul v20.4s, v5.4s, v0.s[3]\n"
+ "fmul v19.4s, v4.4s, v0.s[3]\n"
+ "fmul v18.4s, v3.4s, v0.s[3]\n"
+ "fmul v16.4s, v2.4s, v0.s[3]\n"
+ "fmul v17.4s, v1.4s, v0.s[3]\n"
+ "fcvtas v23.4s, v23.4s\n"
+ "fcvtas v22.4s, v22.4s\n"
+ "fcvtas v21.4s, v21.4s\n"
+ "fcvtas v20.4s, v20.4s\n"
+ "fcvtas v19.4s, v19.4s\n"
+ "fcvtas v18.4s, v18.4s\n"
+ "fcvtas v16.4s, v16.4s\n"
+ "fcvtas v17.4s, v17.4s\n"
+ "uzp1 v22.8h, v23.8h, v22.8h\n"
+ "uzp1 v20.8h, v21.8h, v20.8h\n"
+ "uzp1 v18.8h, v19.8h, v18.8h\n"
+ "uzp1 v17.8h, v16.8h, v17.8h\n"
+ "dup v16.8h, w20\n"
+ "add v22.8h, v22.8h, v16.8h\n"
+ "add v20.8h, v20.8h, v16.8h\n"
+ "add v18.8h, v18.8h, v16.8h\n"
+ "add v17.8h, v17.8h, v16.8h\n"
+ "movi v16.8h, #0xff\n"
+ "smin v22.8h, v22.8h, v16.8h\n"
+ "smin v20.8h, v20.8h, v16.8h\n"
+ "smin v18.8h, v18.8h, v16.8h\n"
+ "smin v17.8h, v17.8h, v16.8h\n"
+ "movi v16.8h, #0x0\n"
+ "smax v22.8h, v22.8h, v16.8h\n"
+ "smax v20.8h, v20.8h, v16.8h\n"
+ "smax v18.8h, v18.8h, v16.8h\n"
+ "smax v17.8h, v17.8h, v16.8h\n"
+ "xtn v22.8b, v22.8h\n"
+ "xtn v20.8b, v20.8h\n"
+ "xtn v18.8b, v18.8h\n"
+ "xtn v17.8b, v17.8h\n"
+ "tbz %x[width], #3, 18f\n"
+ "str d22, [x25, #0x0]\n"
+ "add x25, x25, #0x8\n"
+ "str d18, [x22, #0x0]\n"
+ "add x22, x22, #0x8\n"
+ "tbz %x[width], #2, 16f\n"
+ "str s20, [x25], #0x4\n"
+ "str s17, [x22], #0x4\n"
+ "tbz %x[width], #1, 15f\n"
+ "st1 { v20.h }[2], [x25], #0x2\n"
+ "st1 { v17.h }[2], [x22], #0x2\n"
+ "tbz %x[width], #0, 22f\n"
+ "st1 { v20.b }[6], [x25], #0x1\n"
+ "st1 { v17.b }[6], [x22], #0x1\n"
+ "b 22f\n"
+ "15:" // tail loop: Main loop: unique 2: partial_0_12
+ "tbz %x[width], #0, 22f\n"
+ "st1 { v20.b }[4], [x25], #0x1\n"
+ "st1 { v17.b }[4], [x22], #0x1\n"
+ "b 22f\n"
+ "16:" // tail loop: Main loop: unique 2: partial_1_8
+ "tbz %x[width], #1, 17f\n"
+ "str h20, [x25], #0x2\n"
+ "str h17, [x22], #0x2\n"
+ "tbz %x[width], #0, 22f\n"
+ "st1 { v20.b }[2], [x25], #0x1\n"
+ "st1 { v17.b }[2], [x22], #0x1\n"
+ "b 22f\n"
+ "17:" // tail loop: Main loop: unique 2: partial_0_8
+ "tbz %x[width], #0, 22f\n"
+ "str b20, [x25], #0x1\n"
+ "str b17, [x22], #0x1\n"
+ "b 22f\n"
+ "18:" // tail loop: Main loop: unique 2: partial_2_0
+ "tbz %x[width], #2, 20f\n"
+ "str s22, [x25], #0x4\n"
+ "str s18, [x22], #0x4\n"
+ "tbz %x[width], #1, 19f\n"
+ "st1 { v22.h }[2], [x25], #0x2\n"
+ "st1 { v18.h }[2], [x22], #0x2\n"
+ "tbz %x[width], #0, 22f\n"
+ "st1 { v22.b }[6], [x25], #0x1\n"
+ "st1 { v18.b }[6], [x22], #0x1\n"
+ "b 22f\n"
+ "19:" // tail loop: Main loop: unique 2: partial_0_4
+ "tbz %x[width], #0, 22f\n"
+ "st1 { v22.b }[4], [x25], #0x1\n"
+ "st1 { v18.b }[4], [x22], #0x1\n"
+ "b 22f\n"
+ "20:" // tail loop: Main loop: unique 2: partial_1_0
+ "tbz %x[width], #1, 21f\n"
+ "str h22, [x25], #0x2\n"
+ "str h18, [x22], #0x2\n"
+ "tbz %x[width], #0, 22f\n"
+ "st1 { v22.b }[2], [x25], #0x1\n"
+ "st1 { v18.b }[2], [x22], #0x1\n"
+ "b 22f\n"
+ "21:" // tail loop: Main loop: unique 2: partial_0_0
+ "str b22, [x25], #0x1\n"
+ "str b18, [x22], #0x1\n"
+ "22:" // tail loop: Main loop: unique 2: Done
+ "23:" // tail loop: Main loop: No direct output
+ "mov v19.16b, v28.16b\n"
+ "mov v13.16b, v29.16b\n"
+ "fmla v19.4s, v8.4s, v24.4s\n"
+ "ldr w22, [%x[args_ptr], %[offsetof_out_zeropt]]\n"
+ "mov v18.16b, v30.16b\n"
+ "mov v12.16b, v31.16b\n"
+ "fmla v13.4s, v7.4s, v25.4s\n"
+ "ldr w21, [%x[args_ptr], %[offsetof_maxval]]\n"
+ "mov v17.16b, v28.16b\n"
+ "mov v10.16b, v29.16b\n"
+ "fmla v18.4s, v6.4s, v26.4s\n"
+ "ldr w20, [%x[args_ptr], %[offsetof_minval]]\n"
+ "mov v16.16b, v30.16b\n"
+ "mov v9.16b, v31.16b\n"
+ "fmla v12.4s, v5.4s, v27.4s\n"
+ "fmla v17.4s, v4.4s, v24.4s\n"
+ "fmla v10.4s, v3.4s, v25.4s\n"
+ "fmul v8.4s, v19.4s, v0.s[2]\n"
+ "fmla v16.4s, v2.4s, v26.4s\n"
+ "fmla v9.4s, v1.4s, v27.4s\n"
+ "fmul v7.4s, v13.4s, v0.s[2]\n"
+ "fmul v6.4s, v18.4s, v0.s[2]\n"
+ "fmul v5.4s, v12.4s, v0.s[2]\n"
+ "fmul v4.4s, v17.4s, v0.s[2]\n"
+ "fmul v3.4s, v10.4s, v0.s[2]\n"
+ "fmul v2.4s, v16.4s, v0.s[2]\n"
+ "fmul v1.4s, v9.4s, v0.s[2]\n"
+ "fcvtas v8.4s, v8.4s\n"
+ "fcvtas v7.4s, v7.4s\n"
+ "fcvtas v6.4s, v6.4s\n"
+ "fcvtas v5.4s, v5.4s\n"
+ "fcvtas v4.4s, v4.4s\n"
+ "fcvtas v3.4s, v3.4s\n"
+ "fcvtas v2.4s, v2.4s\n"
+ "fcvtas v1.4s, v1.4s\n"
+ "uzp1 v7.8h, v8.8h, v7.8h\n"
+ "uzp1 v5.8h, v6.8h, v5.8h\n"
+ "uzp1 v3.8h, v4.8h, v3.8h\n"
+ "uzp1 v1.8h, v2.8h, v1.8h\n"
+ "dup v16.8h, w22\n"
+ "add v7.8h, v7.8h, v16.8h\n"
+ "add v5.8h, v5.8h, v16.8h\n"
+ "add v3.8h, v3.8h, v16.8h\n"
+ "add v1.8h, v1.8h, v16.8h\n"
+ "dup v16.8h, w21\n"
+ "smin v7.8h, v7.8h, v16.8h\n"
+ "smin v5.8h, v5.8h, v16.8h\n"
+ "smin v3.8h, v3.8h, v16.8h\n"
+ "smin v1.8h, v1.8h, v16.8h\n"
+ "dup v16.8h, w20\n"
+ "smax v7.8h, v7.8h, v16.8h\n"
+ "smax v5.8h, v5.8h, v16.8h\n"
+ "smax v3.8h, v3.8h, v16.8h\n"
+ "smax v1.8h, v1.8h, v16.8h\n"
+ "xtn v7.8b, v7.8h\n"
+ "xtn v5.8b, v5.8h\n"
+ "xtn v3.8b, v3.8h\n"
+ "xtn v1.8b, v1.8h\n"
+ "tbz %x[width], #3, 27f\n"
+ "str d7, [x26, #0x0]\n"
+ "add x26, x26, #0x8\n"
+ "str d3, [x24, #0x0]\n"
+ "add x24, x24, #0x8\n"
+ "tbz %x[width], #2, 25f\n"
+ "str s5, [x26], #0x4\n"
+ "str s1, [x24], #0x4\n"
+ "tbz %x[width], #1, 24f\n"
+ "st1 { v5.h }[2], [x26], #0x2\n"
+ "st1 { v1.h }[2], [x24], #0x2\n"
+ "tbz %x[width], #0, 31f\n"
+ "st1 { v5.b }[6], [x26], #0x1\n"
+ "st1 { v1.b }[6], [x24], #0x1\n"
+ "b 31f\n"
+ "24:" // tail loop: unique 3: partial_0_12
+ "tbz %x[width], #0, 31f\n"
+ "st1 { v5.b }[4], [x26], #0x1\n"
+ "st1 { v1.b }[4], [x24], #0x1\n"
+ "b 31f\n"
+ "25:" // tail loop: unique 3: partial_1_8
+ "tbz %x[width], #1, 26f\n"
+ "str h5, [x26], #0x2\n"
+ "str h1, [x24], #0x2\n"
+ "tbz %x[width], #0, 31f\n"
+ "st1 { v5.b }[2], [x26], #0x1\n"
+ "st1 { v1.b }[2], [x24], #0x1\n"
+ "b 31f\n"
+ "26:" // tail loop: unique 3: partial_0_8
+ "tbz %x[width], #0, 31f\n"
+ "str b5, [x26], #0x1\n"
+ "str b1, [x24], #0x1\n"
+ "b 31f\n"
+ "27:" // tail loop: unique 3: partial_2_0
+ "tbz %x[width], #2, 29f\n"
+ "str s7, [x26], #0x4\n"
+ "str s3, [x24], #0x4\n"
+ "tbz %x[width], #1, 28f\n"
+ "st1 { v7.h }[2], [x26], #0x2\n"
+ "st1 { v3.h }[2], [x24], #0x2\n"
+ "tbz %x[width], #0, 31f\n"
+ "st1 { v7.b }[6], [x26], #0x1\n"
+ "st1 { v3.b }[6], [x24], #0x1\n"
+ "b 31f\n"
+ "28:" // tail loop: unique 3: partial_0_4
+ "tbz %x[width], #0, 31f\n"
+ "st1 { v7.b }[4], [x26], #0x1\n"
+ "st1 { v3.b }[4], [x24], #0x1\n"
+ "b 31f\n"
+ "29:" // tail loop: unique 3: partial_1_0
+ "tbz %x[width], #1, 30f\n"
+ "str h7, [x26], #0x2\n"
+ "str h3, [x24], #0x2\n"
+ "tbz %x[width], #0, 31f\n"
+ "st1 { v7.b }[2], [x26], #0x1\n"
+ "st1 { v3.b }[2], [x24], #0x1\n"
+ "b 31f\n"
+ "30:" // tail loop: unique 3: partial_0_0
+ "str b7, [x26], #0x1\n"
+ "str b3, [x24], #0x1\n"
+ "31:" // tail loop: unique 3: Done
+ "subs x23, x23, #0x2\n"
+ "bgt 6b\n"
+ "32:" // odd columns skip
+ : [bn_add] "+&r"(bn_add), [bn_mul] "+&r"(bn_mul), [in0] "+&r"(in0), [in1] "+&r"(in1), [out] "+&r"(out), [out_direct] "+&r"(out_direct), [width] "+&r"(width)
+ : [args_ptr] "r"(&ka), [height] "r"(height), [in0_stride] "r"(in0_stride), [in1_stride] "r"(in1_stride), [offsetof_in0_zeropt] "I"(offsetof(KernelArgs, in0_zeropt)), [offsetof_in1_zeropt] "I"(offsetof(KernelArgs, in1_zeropt)), [offsetof_maxval] "I"(offsetof(KernelArgs, maxval)), [offsetof_minval] "I"(offsetof(KernelArgs, minval)), [offsetof_out_direct_zeropt] "I"(offsetof(KernelArgs, out_direct_zeropt)), [offsetof_out_zeropt] "I"(offsetof(KernelArgs, out_zeropt)), [offsetof_scales] "I"(offsetof(KernelArgs, scales)), [out_direct_stride] "r"(out_direct_stride), [out_stride] "r"(out_stride)
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28");
+}
+
+} // namespace
+
+namespace arm_compute
+{
+namespace cpu
+{
+void add_mul_add_u8_neon(const ITensor *input1, const ITensor *input2, const ITensor *bn_mul, const ITensor *bn_add,
+ ITensor *add_output, ITensor *final_output, ConvertPolicy policy, const ActivationLayerInfo &act_info, const Window &window)
+{
+ ARM_COMPUTE_UNUSED(policy);
+
+ const ITensorInfo *final_output_info = final_output->info();
+ const ITensorInfo *add_output_info = (add_output != nullptr) ? add_output->info() : nullptr;
+ const ITensorInfo *input1_info = input1->info();
+ const ITensorInfo *input2_info = input2->info();
+
+ const size_t out_stride = final_output_info->strides_in_bytes()[1];
+ const size_t out_direct_stride = (add_output != nullptr) ? add_output_info->strides_in_bytes()[1] : 0;
+ const size_t in0_stride = input1_info->strides_in_bytes()[1];
+ const size_t in1_stride = input2_info->strides_in_bytes()[1];
+
+ uint8_t minval = std::numeric_limits<uint8_t>::lowest();
+ uint8_t maxval = std::numeric_limits<uint8_t>::max();
+
+ const UniformQuantizationInfo final_output_qinfo = final_output_info->quantization_info().uniform();
+ if(act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU)
+ {
+ minval = quantize_qasymm8(0.f, final_output_qinfo);
+ }
+ else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+ {
+ minval = quantize_qasymm8(0.f, final_output_qinfo);
+ maxval = quantize_qasymm8(act_info.a(), final_output_qinfo);
+ }
+ else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+ {
+ minval = quantize_qasymm8(act_info.b(), final_output_qinfo);
+ maxval = quantize_qasymm8(act_info.a(), final_output_qinfo);
+ }
+
+ const UniformQuantizationInfo in1_qinfo = input1_info->quantization_info().uniform();
+ const UniformQuantizationInfo in2_qinfo = input2_info->quantization_info().uniform();
+ const UniformQuantizationInfo add_output_qinfo = (add_output != nullptr) ? add_output_info->quantization_info().uniform() : UniformQuantizationInfo();
+
+ const int32_t in1_offset = in1_qinfo.offset;
+ const int32_t in2_offset = in2_qinfo.offset;
+ const int32_t out_offset = final_output_qinfo.offset;
+ const int32_t out_direct_offset = add_output_qinfo.offset;
+
+ const float in1_scale = in1_qinfo.scale;
+ const float in2_scale = in2_qinfo.scale;
+ const float out_scale = final_output_qinfo.scale;
+ const float out_direct_scale = add_output_qinfo.scale;
+
+ const float *bn_mul_buffer = reinterpret_cast<float *>(bn_mul->buffer());
+ const float *bn_add_buffer = reinterpret_cast<float *>(bn_add->buffer());
+
+ // Clear X & Y dimensions on execution window as we handle manually
+ Window win = window;
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ win.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+ Iterator in1_it(input1, window);
+ Iterator in2_it(input2, window);
+ Iterator out_it(final_output, window);
+
+ const size_t width = window.num_iterations(0);
+ const size_t height = window.num_iterations(1);
+
+ if(add_output != nullptr)
+ {
+ Iterator add_out_it(add_output, window);
+ execute_window_loop(
+ win, [&](const Coordinates &)
+ {
+ a64_add_bn_clamp_direct_u8_fp32_2x16(
+ reinterpret_cast<uint8_t *>(out_it.ptr()), out_stride,
+ reinterpret_cast<uint8_t *>(add_out_it.ptr()), out_direct_stride,
+ reinterpret_cast<uint8_t *>(in1_it.ptr()), in0_stride,
+ reinterpret_cast<uint8_t *>(in2_it.ptr()), in1_stride,
+ bn_mul_buffer,
+ bn_add_buffer,
+ minval,
+ maxval,
+ out_offset, out_scale,
+ out_direct_offset, out_direct_scale,
+ in1_offset, in1_scale,
+ in2_offset, in2_scale,
+ width, height);
+ },
+ in1_it, in2_it, add_out_it, out_it);
+ }
+ else
+ {
+ execute_window_loop(
+ win, [&](const Coordinates &)
+ {
+ a64_add_bn_clamp_direct_u8_fp32_2x16(
+ reinterpret_cast<uint8_t *>(out_it.ptr()), out_stride,
+ nullptr, out_direct_stride,
+ reinterpret_cast<uint8_t *>(in1_it.ptr()), in0_stride,
+ reinterpret_cast<uint8_t *>(in2_it.ptr()), in1_stride,
+ bn_mul_buffer,
+ bn_add_buffer,
+ minval,
+ maxval,
+ out_offset, out_scale,
+ out_direct_offset, out_direct_scale,
+ in1_offset, in1_scale,
+ in2_offset, in2_scale,
+ width, height);
+ },
+ in1_it, in2_it, out_it);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __aarch64__