aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp80
1 files changed, 40 insertions, 40 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp
index bab687a9b4..1513d378ca 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021, 2023 Arm Limited.
+ * Copyright (c) 2021, 2023-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -54,18 +54,18 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
"2:" // Width loop
"ldp q4, q5, [x22], #0x20\n"
"mov %x[Apanel], x21\n"
- "ld1 { v0.8h }, [%x[Apanel]], #0x10\n"
- "ld1 { v1.8h }, [%x[Apanel]], #0x10\n"
- "movi v8.16b, #0x0\n"
- "ld1 { v2.8h }, [%x[Apanel]], #0x10\n"
"ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
- "cmp x20, #0x2\n"
+ "movi v8.16b, #0x0\n"
"movi v9.16b, #0x0\n"
"movi v10.16b, #0x0\n"
"movi v11.16b, #0x0\n"
+ "ld1 { v0.8h }, [%x[Apanel]], #0x10\n"
"movi v12.16b, #0x0\n"
"movi v13.16b, #0x0\n"
+ "ld1 { v1.8h }, [%x[Apanel]], #0x10\n"
+ "cmp x20, #0x2\n"
"movi v14.16b, #0x0\n"
+ "ld1 { v2.8h }, [%x[Apanel]], #0x10\n"
"movi v15.16b, #0x0\n"
"movi v16.16b, #0x0\n"
"movi v17.16b, #0x0\n"
@@ -97,7 +97,7 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
".inst 0x6e44ecda // bfmmla v26.4s, v6.8h, v4.8h\n"
"cmp x20, #0x2\n"
".inst 0x6e45ecdd // bfmmla v29.4s, v6.8h, v5.8h\n"
- "ldp q4, q5, [x22], #0x20\n"
+ "ldp q5, q4, [x22], #0x20\n"
".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n"
".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
".inst 0x6e43ec2f // bfmmla v15.4s, v1.8h, v3.8h\n"
@@ -106,28 +106,28 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
".inst 0x6e47ec58 // bfmmla v24.4s, v2.8h, v7.8h\n"
".inst 0x6e43ecdb // bfmmla v27.4s, v6.8h, v3.8h\n"
".inst 0x6e47ecde // bfmmla v30.4s, v6.8h, v7.8h\n"
- "ldp q7, q3, [x22], #0x20\n"
- ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
- ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n"
+ "ldp q3, q7, [x22], #0x20\n"
+ ".inst 0x6e45ec0a // bfmmla v10.4s, v0.8h, v5.8h\n"
+ ".inst 0x6e44ec0d // bfmmla v13.4s, v0.8h, v4.8h\n"
"ld1 { v0.8h }, [%x[Apanel]], #0x10\n"
- ".inst 0x6e44ec30 // bfmmla v16.4s, v1.8h, v4.8h\n"
- ".inst 0x6e45ec33 // bfmmla v19.4s, v1.8h, v5.8h\n"
+ ".inst 0x6e45ec30 // bfmmla v16.4s, v1.8h, v5.8h\n"
+ ".inst 0x6e44ec33 // bfmmla v19.4s, v1.8h, v4.8h\n"
+ ".inst 0x6e45ec56 // bfmmla v22.4s, v2.8h, v5.8h\n"
"ld1 { v1.8h }, [%x[Apanel]], #0x10\n"
- ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
- ".inst 0x6e45ec59 // bfmmla v25.4s, v2.8h, v5.8h\n"
+ ".inst 0x6e44ec59 // bfmmla v25.4s, v2.8h, v4.8h\n"
+ ".inst 0x6e45ecdc // bfmmla v28.4s, v6.8h, v5.8h\n"
"ld1 { v2.8h }, [%x[Apanel]], #0x10\n"
- ".inst 0x6e44ecdc // bfmmla v28.4s, v6.8h, v4.8h\n"
- ".inst 0x6e45ecdf // bfmmla v31.4s, v6.8h, v5.8h\n"
+ ".inst 0x6e44ecdf // bfmmla v31.4s, v6.8h, v4.8h\n"
+ ".inst 0x6e43ec08 // bfmmla v8.4s, v0.8h, v3.8h\n"
"ld1 { v6.8h }, [%x[Apanel]], #0x10\n"
- ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n"
+ ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n"
"ldp q4, q5, [x22], #0x20\n"
- ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n"
- ".inst 0x6e47ec2e // bfmmla v14.4s, v1.8h, v7.8h\n"
- ".inst 0x6e43ec31 // bfmmla v17.4s, v1.8h, v3.8h\n"
- ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n"
- ".inst 0x6e43ec57 // bfmmla v23.4s, v2.8h, v3.8h\n"
- ".inst 0x6e47ecda // bfmmla v26.4s, v6.8h, v7.8h\n"
- ".inst 0x6e43ecdd // bfmmla v29.4s, v6.8h, v3.8h\n"
+ ".inst 0x6e43ec2e // bfmmla v14.4s, v1.8h, v3.8h\n"
+ ".inst 0x6e47ec31 // bfmmla v17.4s, v1.8h, v7.8h\n"
+ ".inst 0x6e43ec54 // bfmmla v20.4s, v2.8h, v3.8h\n"
+ ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n"
+ ".inst 0x6e43ecda // bfmmla v26.4s, v6.8h, v3.8h\n"
+ ".inst 0x6e47ecdd // bfmmla v29.4s, v6.8h, v7.8h\n"
"ldp q7, q3, [x22], #0x20\n"
".inst 0x6e44ec09 // bfmmla v9.4s, v0.8h, v4.8h\n"
".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n"
@@ -143,11 +143,11 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
"ld1 { v0.8h }, [%x[Apanel]], #0x10\n"
".inst 0x6e47ec30 // bfmmla v16.4s, v1.8h, v7.8h\n"
".inst 0x6e43ec33 // bfmmla v19.4s, v1.8h, v3.8h\n"
- "ld1 { v1.8h }, [%x[Apanel]], #0x10\n"
".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n"
+ "ld1 { v1.8h }, [%x[Apanel]], #0x10\n"
".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\n"
- "ld1 { v2.8h }, [%x[Apanel]], #0x10\n"
".inst 0x6e47ecdc // bfmmla v28.4s, v6.8h, v7.8h\n"
+ "ld1 { v2.8h }, [%x[Apanel]], #0x10\n"
".inst 0x6e43ecdf // bfmmla v31.4s, v6.8h, v3.8h\n"
"bge 3b\n"
"4:" // main loop skip
@@ -182,9 +182,9 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
"ldp q1, q0, [x22], #0x20\n"
"ld1 { v7.8h }, [%x[Apanel]], #0x10\n"
"ld1 { v6.8h }, [%x[Apanel]], #0x10\n"
- ".inst 0x6e41ece8 // bfmmla v8.4s, v7.8h, v1.8h\n"
"ld1 { v5.8h }, [%x[Apanel]], #0x10\n"
"ld1 { v4.8h }, [%x[Apanel]], #0x10\n"
+ ".inst 0x6e41ece8 // bfmmla v8.4s, v7.8h, v1.8h\n"
".inst 0x6e40eceb // bfmmla v11.4s, v7.8h, v0.8h\n"
"ldp q3, q2, [x22], #0x20\n"
".inst 0x6e41ecce // bfmmla v14.4s, v6.8h, v1.8h\n"
@@ -212,41 +212,41 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
".inst 0x6e40ec9f // bfmmla v31.4s, v4.8h, v0.8h\n"
"5:" // multiply loop done
"subs x23, x23, #0x1\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"