aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp134
1 files changed, 67 insertions, 67 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp
index c5342197c1..e67d17e49a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019-2020, 2023 Arm Limited.
+ * Copyright (c) 2019-2020 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -33,46 +33,46 @@ void a64_interleaved_u8u32_mmla_8x12(
uint32_t *Cpanel, int ablocks, int bblocks, int K) {
struct KernelArgs {
+ size_t bblocks = {};
size_t K = {};
const uint8_t *Bpanel = {};
- size_t bblocks = {};
} ka;
+ ka.bblocks = bblocks;
ka.K = (K/8) - 1;
ka.Bpanel = Bpanel;
- ka.bblocks = bblocks;
__asm__ __volatile__(
"1:" // Height loop
- "ldr x23, [%x[args_ptr], %[offsetof_bblocks]]\n"
- "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n"
+ "ldr x22, [%x[args_ptr], %[offsetof_bblocks]]\n"
"mov x21, %x[Apanel]\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
"2:" // Width loop
- "ldr q4, [x22, #0x0]\n"
- "ldr q5, [x22, #0x10]\n"
+ "ldr x19, [%x[args_ptr], %[offsetof_K]]\n"
"mov %x[Apanel], x21\n"
- "ldr q0, [%x[Apanel], #0x0]\n"
- "ldr q1, [%x[Apanel], #0x10]\n"
+ "cmp x19, #0x2\n"
"movi v8.4s, #0x0\n"
- "ldr q2, [%x[Apanel], #0x20]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
- "cmp x20, #0x2\n"
"movi v9.4s, #0x0\n"
+ "ldr q4, [x20, #0x0]\n"
"movi v10.4s, #0x0\n"
- "add x22, x22, #0x20\n"
"movi v11.4s, #0x0\n"
+ "ldr q0, [%x[Apanel], #0x0]\n"
"movi v12.4s, #0x0\n"
- "add %x[Apanel], %x[Apanel], #0x30\n"
"movi v13.4s, #0x0\n"
+ "ldr q1, [%x[Apanel], #0x10]\n"
"movi v14.4s, #0x0\n"
"movi v15.4s, #0x0\n"
+ "ldr q5, [x20, #0x10]\n"
"movi v16.4s, #0x0\n"
"movi v17.4s, #0x0\n"
+ "ldr q2, [%x[Apanel], #0x20]\n"
"movi v18.4s, #0x0\n"
"movi v19.4s, #0x0\n"
+ "add x20, x20, #0x20\n"
"movi v20.4s, #0x0\n"
"movi v21.4s, #0x0\n"
+ "add %x[Apanel], %x[Apanel], #0x30\n"
"movi v22.4s, #0x0\n"
"movi v23.4s, #0x0\n"
"movi v24.4s, #0x0\n"
@@ -86,149 +86,149 @@ void a64_interleaved_u8u32_mmla_8x12(
"blt 4f\n"
"3:" // main loop head
"ldr q3, [%x[Apanel], #0x0]\n"
- "ldr q6, [x22, #0x0]\n"
".inst 0x6e84a408 // ummla v8.4s, v0.16b, v4.16b\n"
- "ldr q7, [x22, #0x10]\n"
- ".inst 0x6e85a40b // ummla v11.4s, v0.16b, v5.16b\n"
".inst 0x6e84a42e // ummla v14.4s, v1.16b, v4.16b\n"
+ ".inst 0x6e85a40b // ummla v11.4s, v0.16b, v5.16b\n"
".inst 0x6e85a431 // ummla v17.4s, v1.16b, v5.16b\n"
+ "ldr q6, [x20, #0x0]\n"
".inst 0x6e84a454 // ummla v20.4s, v2.16b, v4.16b\n"
- "sub x20, x20, #0x2\n"
".inst 0x6e85a457 // ummla v23.4s, v2.16b, v5.16b\n"
+ "ldr q7, [x20, #0x10]\n"
".inst 0x6e84a47a // ummla v26.4s, v3.16b, v4.16b\n"
- "ldr q4, [x22, #0x20]\n"
".inst 0x6e85a47d // ummla v29.4s, v3.16b, v5.16b\n"
- "ldr q5, [x22, #0x30]\n"
+ "ldr q4, [x20, #0x20]\n"
+ "ldr q5, [x20, #0x30]\n"
".inst 0x6e86a409 // ummla v9.4s, v0.16b, v6.16b\n"
- ".inst 0x6e87a40c // ummla v12.4s, v0.16b, v7.16b\n"
".inst 0x6e86a42f // ummla v15.4s, v1.16b, v6.16b\n"
- "cmp x20, #0x2\n"
- ".inst 0x6e87a432 // ummla v18.4s, v1.16b, v7.16b\n"
".inst 0x6e86a455 // ummla v21.4s, v2.16b, v6.16b\n"
- ".inst 0x6e87a458 // ummla v24.4s, v2.16b, v7.16b\n"
".inst 0x6e86a47b // ummla v27.4s, v3.16b, v6.16b\n"
- "ldr q6, [x22, #0x40]\n"
- ".inst 0x6e87a47e // ummla v30.4s, v3.16b, v7.16b\n"
- "ldr q7, [x22, #0x50]\n"
+ "ldr q6, [x20, #0x40]\n"
+ ".inst 0x6e87a40c // ummla v12.4s, v0.16b, v7.16b\n"
".inst 0x6e84a40a // ummla v10.4s, v0.16b, v4.16b\n"
+ "sub x19, x19, #0x2\n"
".inst 0x6e85a40d // ummla v13.4s, v0.16b, v5.16b\n"
+ ".inst 0x6e87a432 // ummla v18.4s, v1.16b, v7.16b\n"
"ldr q0, [%x[Apanel], #0x10]\n"
".inst 0x6e84a430 // ummla v16.4s, v1.16b, v4.16b\n"
".inst 0x6e85a433 // ummla v19.4s, v1.16b, v5.16b\n"
"ldr q1, [%x[Apanel], #0x20]\n"
+ ".inst 0x6e87a458 // ummla v24.4s, v2.16b, v7.16b\n"
+ ".inst 0x6e87a47e // ummla v30.4s, v3.16b, v7.16b\n"
+ "ldr q7, [x20, #0x50]\n"
".inst 0x6e84a456 // ummla v22.4s, v2.16b, v4.16b\n"
".inst 0x6e85a459 // ummla v25.4s, v2.16b, v5.16b\n"
"ldr q2, [%x[Apanel], #0x30]\n"
".inst 0x6e84a47c // ummla v28.4s, v3.16b, v4.16b\n"
- "ldr q4, [x22, #0x60]\n"
".inst 0x6e85a47f // ummla v31.4s, v3.16b, v5.16b\n"
"ldr q3, [%x[Apanel], #0x40]\n"
- "ldr q5, [x22, #0x70]\n"
".inst 0x6e86a408 // ummla v8.4s, v0.16b, v6.16b\n"
- ".inst 0x6e87a40b // ummla v11.4s, v0.16b, v7.16b\n"
".inst 0x6e86a42e // ummla v14.4s, v1.16b, v6.16b\n"
+ "ldr q4, [x20, #0x60]\n"
+ ".inst 0x6e87a40b // ummla v11.4s, v0.16b, v7.16b\n"
".inst 0x6e87a431 // ummla v17.4s, v1.16b, v7.16b\n"
+ "ldr q5, [x20, #0x70]\n"
".inst 0x6e86a454 // ummla v20.4s, v2.16b, v6.16b\n"
".inst 0x6e87a457 // ummla v23.4s, v2.16b, v7.16b\n"
+ "cmp x19, #0x2\n"
".inst 0x6e86a47a // ummla v26.4s, v3.16b, v6.16b\n"
- "ldr q6, [x22, #0x80]\n"
".inst 0x6e87a47d // ummla v29.4s, v3.16b, v7.16b\n"
- "ldr q7, [x22, #0x90]\n"
+ "ldr q6, [x20, #0x80]\n"
+ "ldr q7, [x20, #0x90]\n"
".inst 0x6e84a409 // ummla v9.4s, v0.16b, v4.16b\n"
- ".inst 0x6e85a40c // ummla v12.4s, v0.16b, v5.16b\n"
".inst 0x6e84a42f // ummla v15.4s, v1.16b, v4.16b\n"
- ".inst 0x6e85a432 // ummla v18.4s, v1.16b, v5.16b\n"
".inst 0x6e84a455 // ummla v21.4s, v2.16b, v4.16b\n"
- ".inst 0x6e85a458 // ummla v24.4s, v2.16b, v5.16b\n"
".inst 0x6e84a47b // ummla v27.4s, v3.16b, v4.16b\n"
- "ldr q4, [x22, #0xa0]\n"
- ".inst 0x6e85a47e // ummla v30.4s, v3.16b, v5.16b\n"
- "ldr q5, [x22, #0xb0]\n"
+ "ldr q4, [x20, #0xa0]\n"
+ ".inst 0x6e85a40c // ummla v12.4s, v0.16b, v5.16b\n"
".inst 0x6e86a40a // ummla v10.4s, v0.16b, v6.16b\n"
".inst 0x6e87a40d // ummla v13.4s, v0.16b, v7.16b\n"
+ ".inst 0x6e85a432 // ummla v18.4s, v1.16b, v5.16b\n"
"ldr q0, [%x[Apanel], #0x50]\n"
".inst 0x6e86a430 // ummla v16.4s, v1.16b, v6.16b\n"
".inst 0x6e87a433 // ummla v19.4s, v1.16b, v7.16b\n"
"ldr q1, [%x[Apanel], #0x60]\n"
+ ".inst 0x6e85a458 // ummla v24.4s, v2.16b, v5.16b\n"
+ ".inst 0x6e85a47e // ummla v30.4s, v3.16b, v5.16b\n"
+ "ldr q5, [x20, #0xb0]\n"
".inst 0x6e86a456 // ummla v22.4s, v2.16b, v6.16b\n"
".inst 0x6e87a459 // ummla v25.4s, v2.16b, v7.16b\n"
"ldr q2, [%x[Apanel], #0x70]\n"
".inst 0x6e86a47c // ummla v28.4s, v3.16b, v6.16b\n"
".inst 0x6e87a47f // ummla v31.4s, v3.16b, v7.16b\n"
"add %x[Apanel], %x[Apanel], #0x80\n"
- "add x22, x22, #0xc0\n"
+ "add x20, x20, #0xc0\n"
"bge 3b\n"
"4:" // main loop skip
"ldr q3, [%x[Apanel], #0x0]\n"
- "ldr q6, [x22, #0x0]\n"
".inst 0x6e84a408 // ummla v8.4s, v0.16b, v4.16b\n"
- "ldr q7, [x22, #0x10]\n"
- ".inst 0x6e85a40b // ummla v11.4s, v0.16b, v5.16b\n"
".inst 0x6e84a42e // ummla v14.4s, v1.16b, v4.16b\n"
+ ".inst 0x6e85a40b // ummla v11.4s, v0.16b, v5.16b\n"
".inst 0x6e85a431 // ummla v17.4s, v1.16b, v5.16b\n"
+ "ldr q6, [x20, #0x0]\n"
".inst 0x6e84a454 // ummla v20.4s, v2.16b, v4.16b\n"
- "add %x[Apanel], %x[Apanel], #0x10\n"
".inst 0x6e85a457 // ummla v23.4s, v2.16b, v5.16b\n"
+ "ldr q7, [x20, #0x10]\n"
".inst 0x6e84a47a // ummla v26.4s, v3.16b, v4.16b\n"
- "ldr q4, [x22, #0x20]\n"
".inst 0x6e85a47d // ummla v29.4s, v3.16b, v5.16b\n"
- "ldr q5, [x22, #0x30]\n"
+ "ldr q4, [x20, #0x20]\n"
+ "ldr q5, [x20, #0x30]\n"
".inst 0x6e86a409 // ummla v9.4s, v0.16b, v6.16b\n"
- ".inst 0x6e87a40c // ummla v12.4s, v0.16b, v7.16b\n"
".inst 0x6e86a42f // ummla v15.4s, v1.16b, v6.16b\n"
- "add x22, x22, #0x40\n"
- ".inst 0x6e87a432 // ummla v18.4s, v1.16b, v7.16b\n"
".inst 0x6e86a455 // ummla v21.4s, v2.16b, v6.16b\n"
- ".inst 0x6e87a458 // ummla v24.4s, v2.16b, v7.16b\n"
".inst 0x6e86a47b // ummla v27.4s, v3.16b, v6.16b\n"
- ".inst 0x6e87a47e // ummla v30.4s, v3.16b, v7.16b\n"
+ "add %x[Apanel], %x[Apanel], #0x10\n"
+ ".inst 0x6e87a40c // ummla v12.4s, v0.16b, v7.16b\n"
".inst 0x6e84a40a // ummla v10.4s, v0.16b, v4.16b\n"
+ "add x20, x20, #0x40\n"
".inst 0x6e85a40d // ummla v13.4s, v0.16b, v5.16b\n"
+ ".inst 0x6e87a432 // ummla v18.4s, v1.16b, v7.16b\n"
".inst 0x6e84a430 // ummla v16.4s, v1.16b, v4.16b\n"
".inst 0x6e85a433 // ummla v19.4s, v1.16b, v5.16b\n"
+ ".inst 0x6e87a458 // ummla v24.4s, v2.16b, v7.16b\n"
+ ".inst 0x6e87a47e // ummla v30.4s, v3.16b, v7.16b\n"
".inst 0x6e84a456 // ummla v22.4s, v2.16b, v4.16b\n"
".inst 0x6e85a459 // ummla v25.4s, v2.16b, v5.16b\n"
".inst 0x6e84a47c // ummla v28.4s, v3.16b, v4.16b\n"
".inst 0x6e85a47f // ummla v31.4s, v3.16b, v5.16b\n"
- "cbz x20, 5f\n"
- "ldr q6, [x22, #0x0]\n"
+ "cbz x19, 5f\n"
+ "ldr q6, [x20, #0x0]\n"
"ldr q0, [%x[Apanel], #0x0]\n"
".inst 0x6e86a408 // ummla v8.4s, v0.16b, v6.16b\n"
"ldr q1, [%x[Apanel], #0x10]\n"
- "ldr q7, [x22, #0x10]\n"
- ".inst 0x6e87a40b // ummla v11.4s, v0.16b, v7.16b\n"
+ "ldr q7, [x20, #0x10]\n"
+ ".inst 0x6e86a42e // ummla v14.4s, v1.16b, v6.16b\n"
"ldr q2, [%x[Apanel], #0x20]\n"
"ldr q3, [%x[Apanel], #0x30]\n"
- ".inst 0x6e86a42e // ummla v14.4s, v1.16b, v6.16b\n"
- "ldr q4, [x22, #0x20]\n"
- "ldr q5, [x22, #0x30]\n"
+ ".inst 0x6e87a40b // ummla v11.4s, v0.16b, v7.16b\n"
".inst 0x6e87a431 // ummla v17.4s, v1.16b, v7.16b\n"
".inst 0x6e86a454 // ummla v20.4s, v2.16b, v6.16b\n"
+ "ldr q4, [x20, #0x20]\n"
".inst 0x6e87a457 // ummla v23.4s, v2.16b, v7.16b\n"
- "add %x[Apanel], %x[Apanel], #0x40\n"
".inst 0x6e86a47a // ummla v26.4s, v3.16b, v6.16b\n"
- "ldr q6, [x22, #0x40]\n"
+ "ldr q5, [x20, #0x30]\n"
".inst 0x6e87a47d // ummla v29.4s, v3.16b, v7.16b\n"
- "ldr q7, [x22, #0x50]\n"
+ "ldr q6, [x20, #0x40]\n"
+ "ldr q7, [x20, #0x50]\n"
".inst 0x6e84a409 // ummla v9.4s, v0.16b, v4.16b\n"
- ".inst 0x6e85a40c // ummla v12.4s, v0.16b, v5.16b\n"
".inst 0x6e84a42f // ummla v15.4s, v1.16b, v4.16b\n"
- ".inst 0x6e85a432 // ummla v18.4s, v1.16b, v5.16b\n"
- "add x22, x22, #0x60\n"
+ "add x20, x20, #0x60\n"
".inst 0x6e84a455 // ummla v21.4s, v2.16b, v4.16b\n"
- ".inst 0x6e85a458 // ummla v24.4s, v2.16b, v5.16b\n"
".inst 0x6e84a47b // ummla v27.4s, v3.16b, v4.16b\n"
- ".inst 0x6e85a47e // ummla v30.4s, v3.16b, v5.16b\n"
+ "add %x[Apanel], %x[Apanel], #0x40\n"
+ ".inst 0x6e85a40c // ummla v12.4s, v0.16b, v5.16b\n"
".inst 0x6e86a40a // ummla v10.4s, v0.16b, v6.16b\n"
".inst 0x6e87a40d // ummla v13.4s, v0.16b, v7.16b\n"
+ ".inst 0x6e85a432 // ummla v18.4s, v1.16b, v5.16b\n"
".inst 0x6e86a430 // ummla v16.4s, v1.16b, v6.16b\n"
".inst 0x6e87a433 // ummla v19.4s, v1.16b, v7.16b\n"
+ ".inst 0x6e85a458 // ummla v24.4s, v2.16b, v5.16b\n"
+ ".inst 0x6e85a47e // ummla v30.4s, v3.16b, v5.16b\n"
".inst 0x6e86a456 // ummla v22.4s, v2.16b, v6.16b\n"
".inst 0x6e87a459 // ummla v25.4s, v2.16b, v7.16b\n"
".inst 0x6e86a47c // ummla v28.4s, v3.16b, v6.16b\n"
".inst 0x6e87a47f // ummla v31.4s, v3.16b, v7.16b\n"
"5:" // multiply loop done
- "subs x23, x23, #0x1\n"
+ "subs x22, x22, #0x1\n"
"uzp1 v4.2d, v8.2d, v11.2d\n"
"uzp2 v8.2d, v8.2d, v11.2d\n"
"uzp1 v11.2d, v9.2d, v12.2d\n"
@@ -283,7 +283,7 @@ void a64_interleaved_u8u32_mmla_8x12(
"bne 1b\n"
: [Apanel] "+&r" (Apanel), [Cpanel] "+&r" (Cpanel), [ablocks] "+&r" (ablocks)
: [args_ptr] "r" (&ka), [offsetof_Bpanel] "I" (offsetof(KernelArgs, Bpanel)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_bblocks] "I" (offsetof(KernelArgs, bblocks))
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23"
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x19", "x20", "x21", "x22"
);
}
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-20 10:42:00 +0000