aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp48
1 files changed, 24 insertions, 24 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp
index 4a1c1b5638..5331f9e652 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2022-2023 Arm Limited.
+ * Copyright (c) 2022-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -54,17 +54,17 @@ void a64_ffinterleaved_bf16fp32_mmla_8x12(
"1:" // Height loop
"ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
"ldr x25, [%x[args_ptr], %[offsetof_N]]\n"
- "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
"mov x24, %x[Apanel]\n"
+ "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
"2:" // Width loop
"ldr x23, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
"ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
+ "cmp x25, #0x8\n"
+ "mov %x[Apanel], x24\n"
"add x22, x23, x20, LSL #1\n"
"add x21, x22, x20, LSL #1\n"
"add x20, x21, x20, LSL #1\n"
"str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
- "cmp x25, #0x8\n"
- "mov %x[Apanel], x24\n"
"bgt 3f\n"
"cmp x25, #0x4\n"
"mov x21, x23\n"
@@ -79,14 +79,14 @@ void a64_ffinterleaved_bf16fp32_mmla_8x12(
"movi v9.16b, #0x0\n"
"ldr q2, [%x[Apanel], #0x20]\n"
"ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
- "cmp x20, #0x2\n"
"movi v10.16b, #0x0\n"
"movi v11.16b, #0x0\n"
- "add x23, x23, #0x20\n"
"movi v12.16b, #0x0\n"
+ "add x23, x23, #0x20\n"
"movi v13.16b, #0x0\n"
- "add %x[Apanel], %x[Apanel], #0x30\n"
"movi v14.16b, #0x0\n"
+ "add %x[Apanel], %x[Apanel], #0x30\n"
+ "cmp x20, #0x2\n"
"movi v15.16b, #0x0\n"
"movi v16.16b, #0x0\n"
"movi v17.16b, #0x0\n"
@@ -217,19 +217,19 @@ void a64_ffinterleaved_bf16fp32_mmla_8x12(
"cbz x20, 6f\n"
"ldr q1, [x23, #0x0]\n"
"ldr q7, [%x[Apanel], #0x0]\n"
- ".inst 0x6e41ece8 // bfmmla v8.4s, v7.8h, v1.8h\n"
"ldr q6, [%x[Apanel], #0x10]\n"
"ldr q0, [x23, #0x10]\n"
- ".inst 0x6e40eceb // bfmmla v11.4s, v7.8h, v0.8h\n"
"ldr q5, [%x[Apanel], #0x20]\n"
"ldr q4, [%x[Apanel], #0x30]\n"
- ".inst 0x6e41ecce // bfmmla v14.4s, v6.8h, v1.8h\n"
+ "add %x[Apanel], %x[Apanel], #0x40\n"
"ldr q3, [x22, #0x0]\n"
"ldr q2, [x22, #0x10]\n"
+ ".inst 0x6e41ece8 // bfmmla v8.4s, v7.8h, v1.8h\n"
+ ".inst 0x6e40eceb // bfmmla v11.4s, v7.8h, v0.8h\n"
+ ".inst 0x6e41ecce // bfmmla v14.4s, v6.8h, v1.8h\n"
".inst 0x6e40ecd1 // bfmmla v17.4s, v6.8h, v0.8h\n"
".inst 0x6e41ecb4 // bfmmla v20.4s, v5.8h, v1.8h\n"
".inst 0x6e40ecb7 // bfmmla v23.4s, v5.8h, v0.8h\n"
- "add %x[Apanel], %x[Apanel], #0x40\n"
".inst 0x6e41ec9a // bfmmla v26.4s, v4.8h, v1.8h\n"
"ldr q1, [x21, #0x0]\n"
".inst 0x6e40ec9d // bfmmla v29.4s, v4.8h, v0.8h\n"
@@ -252,41 +252,41 @@ void a64_ffinterleaved_bf16fp32_mmla_8x12(
".inst 0x6e40ec9f // bfmmla v31.4s, v4.8h, v0.8h\n"
"6:" // multiply loop done
"subs x25, x25, #0xc\n"
- "uzp1 v0.2d, v8.2d, v11.2d\n"
+ "uzp1 v2.2d, v8.2d, v11.2d\n"
"uzp2 v8.2d, v8.2d, v11.2d\n"
"uzp1 v1.2d, v9.2d, v12.2d\n"
"uzp2 v9.2d, v9.2d, v12.2d\n"
- "str q0, [%x[Cpanel], #0x0]\n"
"uzp1 v0.2d, v10.2d, v13.2d\n"
"uzp2 v10.2d, v10.2d, v13.2d\n"
- "str q1, [%x[Cpanel], #0x10]\n"
- "str q0, [%x[Cpanel], #0x20]\n"
- "uzp1 v0.2d, v14.2d, v17.2d\n"
+ "str q2, [%x[Cpanel], #0x0]\n"
+ "uzp1 v3.2d, v14.2d, v17.2d\n"
"uzp2 v14.2d, v14.2d, v17.2d\n"
- "str q8, [%x[Cpanel], #0x30]\n"
+ "str q1, [%x[Cpanel], #0x10]\n"
"uzp1 v2.2d, v15.2d, v18.2d\n"
"uzp2 v15.2d, v15.2d, v18.2d\n"
- "str q9, [%x[Cpanel], #0x40]\n"
+ "str q0, [%x[Cpanel], #0x20]\n"
"uzp1 v17.2d, v16.2d, v19.2d\n"
"uzp2 v16.2d, v16.2d, v19.2d\n"
- "str q10, [%x[Cpanel], #0x50]\n"
+ "str q8, [%x[Cpanel], #0x30]\n"
"uzp1 v1.2d, v20.2d, v23.2d\n"
"uzp2 v20.2d, v20.2d, v23.2d\n"
- "str q0, [%x[Cpanel], #0x60]\n"
+ "str q9, [%x[Cpanel], #0x40]\n"
"uzp1 v0.2d, v21.2d, v24.2d\n"
"uzp2 v21.2d, v21.2d, v24.2d\n"
- "str q2, [%x[Cpanel], #0x70]\n"
+ "str q10, [%x[Cpanel], #0x50]\n"
"uzp1 v23.2d, v22.2d, v25.2d\n"
"uzp2 v22.2d, v22.2d, v25.2d\n"
- "str q17, [%x[Cpanel], #0x80]\n"
+ "str q3, [%x[Cpanel], #0x60]\n"
"uzp1 v19.2d, v26.2d, v29.2d\n"
"uzp2 v26.2d, v26.2d, v29.2d\n"
- "str q14, [%x[Cpanel], #0x90]\n"
+ "str q2, [%x[Cpanel], #0x70]\n"
"uzp1 v18.2d, v27.2d, v30.2d\n"
"uzp2 v27.2d, v27.2d, v30.2d\n"
- "str q15, [%x[Cpanel], #0xa0]\n"
+ "str q17, [%x[Cpanel], #0x80]\n"
"uzp1 v17.2d, v28.2d, v31.2d\n"
"uzp2 v28.2d, v28.2d, v31.2d\n"
+ "str q14, [%x[Cpanel], #0x90]\n"
+ "str q15, [%x[Cpanel], #0xa0]\n"
"str q16, [%x[Cpanel], #0xb0]\n"
"str q1, [%x[Cpanel], #0xc0]\n"
"str q0, [%x[Cpanel], #0xd0]\n"