COMPMID-3776: Indirect GEMM
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I51a1b0f098bc3a8c408c50c92221e4df3061e12c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4343
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_u8_u8_summing.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_u8_u8_summing.hpp
new file mode 100644
index 0000000..07164d6
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_u8_u8_summing.hpp
@@ -0,0 +1,362 @@
+/*
+ * Copyright (c) 2019-2020 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.
+ */
+
+#ifdef __aarch64__
+
+template<>
+void interleave_block<8, 8, VLType::None, true>(
+ uint8_t * &out_ptr, const uint8_t * const * in, size_t width, size_t height,
+ size_t row_offset, bool first
+)
+{
+ __asm__ __volatile__(
+ "movi v5.8h, #0x0\n"
+ "ldr x27, [%x[in], #0x0]\n"
+ "mov x19, #0x0\n"
+ "movi v4.8h, #0x0\n"
+ "ldr x26, [%x[in], #0x8]\n"
+ "cmp %x[height], #0x8\n"
+ "movi v3.8h, #0x0\n"
+ "ldr x25, [%x[in], #0x10]\n"
+ "add x27, x27, %x[row_offset]\n"
+ "movi v2.8h, #0x0\n"
+ "ldr x24, [%x[in], #0x18]\n"
+ "movi v1.4s, #0x0\n"
+ "ldr x23, [%x[in], #0x20]\n"
+ "add x26, x26, %x[row_offset]\n"
+ "movi v0.4s, #0x0\n"
+ "ldr x22, [%x[in], #0x28]\n"
+ "add x25, x25, %x[row_offset]\n"
+ "movi v31.4s, #0x0\n"
+ "ldr x21, [%x[in], #0x30]\n"
+ "add x24, x24, %x[row_offset]\n"
+ "movi v30.4s, #0x0\n"
+ "ldr x20, [%x[in], #0x38]\n"
+ "add x23, x23, %x[row_offset]\n"
+ "add x22, x22, %x[row_offset]\n"
+ "add x21, x21, %x[row_offset]\n"
+ "add x20, x20, %x[row_offset]\n"
+ "beq 1f\n"
+ "mov x20, x27\n"
+ "cmp %x[height], #0x2\n"
+ "csel x26, x26, x27, GE\n"
+ "csel x25, x25, x27, GT\n"
+ "cmp %x[height], #0x4\n"
+ "csel x24, x24, x27, GE\n"
+ "csel x23, x23, x27, GT\n"
+ "cmp %x[height], #0x6\n"
+ "csel x22, x22, x27, GE\n"
+ "csel x21, x21, x27, GT\n"
+ "1:" // no_pointer_adj
+ "movi v29.4s, #0x0\n"
+ "prfm pldl1keep, [x27, #0x0]\n"
+ "movi v28.4s, #0x0\n"
+ "prfm pldl1keep, [x26, #0x0]\n"
+ "prfm pldl1keep, [x25, #0x0]\n"
+ "prfm pldl1keep, [x24, #0x0]\n"
+ "prfm pldl1keep, [x23, #0x0]\n"
+ "prfm pldl1keep, [x22, #0x0]\n"
+ "prfm pldl1keep, [x21, #0x0]\n"
+ "prfm pldl1keep, [x20, #0x0]\n"
+ "prfm pldl1keep, [x27, #0x40]\n"
+ "prfm pldl1keep, [x26, #0x40]\n"
+ "prfm pldl1keep, [x25, #0x40]\n"
+ "prfm pldl1keep, [x24, #0x40]\n"
+ "prfm pldl1keep, [x23, #0x40]\n"
+ "prfm pldl1keep, [x22, #0x40]\n"
+ "prfm pldl1keep, [x21, #0x40]\n"
+ "prfm pldl1keep, [x20, #0x40]\n"
+ "cbnz %w[first], 2f\n"
+ "sub %x[out_ptr], %x[out_ptr], #0x20\n"
+ "ld1 { v29.4s }, [%x[out_ptr]]\n"
+ "ldr q28, [%x[out_ptr], #0x10]\n"
+ "2:" // first_pass
+ "cmp %x[width], #0x10\n"
+ "blt 5f\n"
+ "3:" // Main loop head
+ "cmp x19, #0x3e\n"
+ "ble 4f\n"
+ "uadalp v1.4s, v5.8h\n"
+ "movi v5.8h, #0x0\n"
+ "uadalp v0.4s, v4.8h\n"
+ "movi v4.8h, #0x0\n"
+ "uadalp v31.4s, v3.8h\n"
+ "movi v3.8h, #0x0\n"
+ "uadalp v30.4s, v2.8h\n"
+ "movi v2.8h, #0x0\n"
+ "mov x19, #0x0\n"
+ "4:" // no_accumulate_16
+ "ldr q27, [x27], #0x10\n"
+ "prfm pldl1keep, [x27, #0x70]\n"
+ "ldr q24, [x26], #0x10\n"
+ "zip1 v26.2d, v27.2d, v24.2d\n"
+ "prfm pldl1keep, [x26, #0x70]\n"
+ "ldr q25, [x25], #0x10\n"
+ "zip2 v24.2d, v27.2d, v24.2d\n"
+ "prfm pldl1keep, [x25, #0x70]\n"
+ "ldr q21, [x24], #0x10\n"
+ "zip1 v23.2d, v25.2d, v21.2d\n"
+ "prfm pldl1keep, [x24, #0x70]\n"
+ "ldr q22, [x23], #0x10\n"
+ "zip2 v21.2d, v25.2d, v21.2d\n"
+ "prfm pldl1keep, [x23, #0x70]\n"
+ "ldr q18, [x22], #0x10\n"
+ "zip1 v20.2d, v22.2d, v18.2d\n"
+ "prfm pldl1keep, [x22, #0x70]\n"
+ "ldr q19, [x21], #0x10\n"
+ "zip2 v18.2d, v22.2d, v18.2d\n"
+ "prfm pldl1keep, [x21, #0x70]\n"
+ "ldr q16, [x20], #0x10\n"
+ "zip1 v17.2d, v19.2d, v16.2d\n"
+ "prfm pldl1keep, [x20, #0x70]\n"
+ "str q26, [%x[out_ptr], #0x0]\n"
+ "zip2 v16.2d, v19.2d, v16.2d\n"
+ "uadalp v5.8h, v26.16b\n"
+ "str q23, [%x[out_ptr], #0x10]\n"
+ "uadalp v4.8h, v23.16b\n"
+ "str q20, [%x[out_ptr], #0x20]\n"
+ "uadalp v3.8h, v20.16b\n"
+ "str q17, [%x[out_ptr], #0x30]\n"
+ "uadalp v2.8h, v17.16b\n"
+ "str q24, [%x[out_ptr], #0x40]\n"
+ "uadalp v5.8h, v24.16b\n"
+ "str q21, [%x[out_ptr], #0x50]\n"
+ "uadalp v4.8h, v21.16b\n"
+ "str q18, [%x[out_ptr], #0x60]\n"
+ "uadalp v3.8h, v18.16b\n"
+ "str q16, [%x[out_ptr], #0x70]\n"
+ "uadalp v2.8h, v16.16b\n"
+ "add x19, x19, #0x1\n"
+ "subs %x[width], %x[width], #0x10\n"
+ "cmp %x[width], #0x10\n"
+ "add %x[out_ptr], %x[out_ptr], #0x80\n"
+ "bge 3b\n"
+ "5:" // Main loop skip
+ "cbz %x[width], 14f\n"
+ "tbz %x[width], #3, 9f\n"
+ "ldr d27, [x27], #0x8\n"
+ "ldr d24, [x26], #0x8\n"
+ "ldr d25, [x25], #0x8\n"
+ "ldr d21, [x24], #0x8\n"
+ "ldr d22, [x23], #0x8\n"
+ "ldr d18, [x22], #0x8\n"
+ "ldr d19, [x21], #0x8\n"
+ "ldr d16, [x20], #0x8\n"
+ "tbz %x[width], #2, 7f\n"
+ "ld1 { v27.s }[2], [x27], #0x4\n"
+ "ld1 { v24.s }[2], [x26], #0x4\n"
+ "ld1 { v25.s }[2], [x25], #0x4\n"
+ "ld1 { v21.s }[2], [x24], #0x4\n"
+ "ld1 { v22.s }[2], [x23], #0x4\n"
+ "ld1 { v18.s }[2], [x22], #0x4\n"
+ "ld1 { v19.s }[2], [x21], #0x4\n"
+ "ld1 { v16.s }[2], [x20], #0x4\n"
+ "tbz %x[width], #1, 6f\n"
+ "ld1 { v27.h }[6], [x27], #0x2\n"
+ "ld1 { v24.h }[6], [x26], #0x2\n"
+ "ld1 { v25.h }[6], [x25], #0x2\n"
+ "ld1 { v21.h }[6], [x24], #0x2\n"
+ "ld1 { v22.h }[6], [x23], #0x2\n"
+ "ld1 { v18.h }[6], [x22], #0x2\n"
+ "ld1 { v19.h }[6], [x21], #0x2\n"
+ "ld1 { v16.h }[6], [x20], #0x2\n"
+ "mov x19, #0x2\n"
+ "tbz %x[width], #0, 13f\n"
+ "ld1 { v27.b }[14], [x27]\n"
+ "ld1 { v24.b }[14], [x26]\n"
+ "ld1 { v25.b }[14], [x25]\n"
+ "ld1 { v21.b }[14], [x24]\n"
+ "ld1 { v22.b }[14], [x23]\n"
+ "ld1 { v18.b }[14], [x22]\n"
+ "ld1 { v19.b }[14], [x21]\n"
+ "ld1 { v16.b }[14], [x20]\n"
+ "b 13f\n"
+ "6:" // odd_loads_1_12
+ "mov x19, #0x2\n"
+ "tbz %x[width], #0, 13f\n"
+ "ld1 { v27.b }[12], [x27]\n"
+ "ld1 { v24.b }[12], [x26]\n"
+ "ld1 { v25.b }[12], [x25]\n"
+ "ld1 { v21.b }[12], [x24]\n"
+ "ld1 { v22.b }[12], [x23]\n"
+ "ld1 { v18.b }[12], [x22]\n"
+ "ld1 { v19.b }[12], [x21]\n"
+ "ld1 { v16.b }[12], [x20]\n"
+ "b 13f\n"
+ "7:" // odd_loads_2_8
+ "tbz %x[width], #1, 8f\n"
+ "ld1 { v27.h }[4], [x27], #0x2\n"
+ "ld1 { v24.h }[4], [x26], #0x2\n"
+ "ld1 { v25.h }[4], [x25], #0x2\n"
+ "ld1 { v21.h }[4], [x24], #0x2\n"
+ "ld1 { v22.h }[4], [x23], #0x2\n"
+ "ld1 { v18.h }[4], [x22], #0x2\n"
+ "ld1 { v19.h }[4], [x21], #0x2\n"
+ "ld1 { v16.h }[4], [x20], #0x2\n"
+ "mov x19, #0x2\n"
+ "tbz %x[width], #0, 13f\n"
+ "ld1 { v27.b }[10], [x27]\n"
+ "ld1 { v24.b }[10], [x26]\n"
+ "ld1 { v25.b }[10], [x25]\n"
+ "ld1 { v21.b }[10], [x24]\n"
+ "ld1 { v22.b }[10], [x23]\n"
+ "ld1 { v18.b }[10], [x22]\n"
+ "ld1 { v19.b }[10], [x21]\n"
+ "ld1 { v16.b }[10], [x20]\n"
+ "b 13f\n"
+ "8:" // odd_loads_1_8
+ "mov x19, #0x1\n"
+ "tbz %x[width], #0, 13f\n"
+ "ld1 { v27.b }[8], [x27]\n"
+ "ld1 { v24.b }[8], [x26]\n"
+ "ld1 { v25.b }[8], [x25]\n"
+ "ld1 { v21.b }[8], [x24]\n"
+ "ld1 { v22.b }[8], [x23]\n"
+ "ld1 { v18.b }[8], [x22]\n"
+ "ld1 { v19.b }[8], [x21]\n"
+ "ld1 { v16.b }[8], [x20]\n"
+ "mov x19, #0x2\n"
+ "b 13f\n"
+ "9:" // odd_loads_4_0
+ "tbz %x[width], #2, 11f\n"
+ "ldr s27, [x27], #0x4\n"
+ "ldr s24, [x26], #0x4\n"
+ "ldr s25, [x25], #0x4\n"
+ "ldr s21, [x24], #0x4\n"
+ "ldr s22, [x23], #0x4\n"
+ "ldr s18, [x22], #0x4\n"
+ "ldr s19, [x21], #0x4\n"
+ "ldr s16, [x20], #0x4\n"
+ "tbz %x[width], #1, 10f\n"
+ "ld1 { v27.h }[2], [x27], #0x2\n"
+ "ld1 { v24.h }[2], [x26], #0x2\n"
+ "ld1 { v25.h }[2], [x25], #0x2\n"
+ "ld1 { v21.h }[2], [x24], #0x2\n"
+ "ld1 { v22.h }[2], [x23], #0x2\n"
+ "ld1 { v18.h }[2], [x22], #0x2\n"
+ "ld1 { v19.h }[2], [x21], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
+ "mov x19, #0x1\n"
+ "tbz %x[width], #0, 13f\n"
+ "ld1 { v27.b }[6], [x27]\n"
+ "ld1 { v24.b }[6], [x26]\n"
+ "ld1 { v25.b }[6], [x25]\n"
+ "ld1 { v21.b }[6], [x24]\n"
+ "ld1 { v22.b }[6], [x23]\n"
+ "ld1 { v18.b }[6], [x22]\n"
+ "ld1 { v19.b }[6], [x21]\n"
+ "ld1 { v16.b }[6], [x20]\n"
+ "b 13f\n"
+ "10:" // odd_loads_1_4
+ "mov x19, #0x1\n"
+ "tbz %x[width], #0, 13f\n"
+ "ld1 { v27.b }[4], [x27]\n"
+ "ld1 { v24.b }[4], [x26]\n"
+ "ld1 { v25.b }[4], [x25]\n"
+ "ld1 { v21.b }[4], [x24]\n"
+ "ld1 { v22.b }[4], [x23]\n"
+ "ld1 { v18.b }[4], [x22]\n"
+ "ld1 { v19.b }[4], [x21]\n"
+ "ld1 { v16.b }[4], [x20]\n"
+ "b 13f\n"
+ "11:" // odd_loads_2_0
+ "tbz %x[width], #1, 12f\n"
+ "ldr h27, [x27], #0x2\n"
+ "ldr h24, [x26], #0x2\n"
+ "ldr h25, [x25], #0x2\n"
+ "ldr h21, [x24], #0x2\n"
+ "ldr h22, [x23], #0x2\n"
+ "ldr h18, [x22], #0x2\n"
+ "ldr h19, [x21], #0x2\n"
+ "ldr h16, [x20], #0x2\n"
+ "mov x19, #0x1\n"
+ "tbz %x[width], #0, 13f\n"
+ "ld1 { v27.b }[2], [x27]\n"
+ "ld1 { v24.b }[2], [x26]\n"
+ "ld1 { v25.b }[2], [x25]\n"
+ "ld1 { v21.b }[2], [x24]\n"
+ "ld1 { v22.b }[2], [x23]\n"
+ "ld1 { v18.b }[2], [x22]\n"
+ "ld1 { v19.b }[2], [x21]\n"
+ "ld1 { v16.b }[2], [x20]\n"
+ "b 13f\n"
+ "12:" // odd_loads_1_0
+ "ldr b27, [x27, #0x0]\n"
+ "ldr b24, [x26, #0x0]\n"
+ "ldr b25, [x25, #0x0]\n"
+ "ldr b21, [x24, #0x0]\n"
+ "ldr b22, [x23, #0x0]\n"
+ "ldr b18, [x22, #0x0]\n"
+ "ldr b19, [x21, #0x0]\n"
+ "ldr b16, [x20, #0x0]\n"
+ "mov x19, #0x1\n"
+ "13:" // Odd load end
+ "zip1 v26.2d, v27.2d, v24.2d\n"
+ "subs x19, x19, #0x1\n"
+ "zip1 v23.2d, v25.2d, v21.2d\n"
+ "str q26, [%x[out_ptr], #0x0]\n"
+ "zip1 v20.2d, v22.2d, v18.2d\n"
+ "uadalp v5.8h, v26.16b\n"
+ "zip1 v17.2d, v19.2d, v16.2d\n"
+ "str q23, [%x[out_ptr], #0x10]\n"
+ "uadalp v4.8h, v23.16b\n"
+ "str q20, [%x[out_ptr], #0x20]\n"
+ "uadalp v3.8h, v20.16b\n"
+ "str q17, [%x[out_ptr], #0x30]\n"
+ "uadalp v2.8h, v17.16b\n"
+ "add %x[out_ptr], %x[out_ptr], #0x40\n"
+ "beq 14f\n"
+ "zip2 v24.2d, v27.2d, v24.2d\n"
+ "zip2 v21.2d, v25.2d, v21.2d\n"
+ "str q24, [%x[out_ptr], #0x0]\n"
+ "zip2 v18.2d, v22.2d, v18.2d\n"
+ "uadalp v5.8h, v24.16b\n"
+ "zip2 v16.2d, v19.2d, v16.2d\n"
+ "str q21, [%x[out_ptr], #0x10]\n"
+ "uadalp v4.8h, v21.16b\n"
+ "str q18, [%x[out_ptr], #0x20]\n"
+ "uadalp v3.8h, v18.16b\n"
+ "str q16, [%x[out_ptr], #0x30]\n"
+ "uadalp v2.8h, v16.16b\n"
+ "add %x[out_ptr], %x[out_ptr], #0x40\n"
+ "14:" // Odds skip
+ "uadalp v1.4s, v5.8h\n"
+ "uadalp v0.4s, v4.8h\n"
+ "addp v1.4s, v1.4s, v0.4s\n"
+ "uadalp v31.4s, v3.8h\n"
+ "uadalp v30.4s, v2.8h\n"
+ "add v1.4s, v1.4s, v29.4s\n"
+ "str q1, [%x[out_ptr], #0x0]\n"
+ "addp v0.4s, v31.4s, v30.4s\n"
+ "add v0.4s, v0.4s, v28.4s\n"
+ "str q0, [%x[out_ptr], #0x10]\n"
+ "add %x[out_ptr], %x[out_ptr], #0x20\n"
+ : [out_ptr] "+r" (out_ptr), [width] "+r" (width)
+ : [first] "r" (first), [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27"
+ );
+}
+
+
+#endif // __aarch64__