From 48b3ef89de5f21a0169d8416e3d54081f82c7bf8 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 14 Oct 2019 19:03:09 +0100 Subject: COMPMID-2577: Fuse bias addition and activation in gemm assembly kernels Change-Id: I7f52112d2d05b1ea3d3f3d4b19b8eafab05d6c44 Signed-off-by: Georgios Pinitas Reviewed-on: https://review.mlplatform.org/c/2141 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Pablo Marquez --- .../a64_smallK_hybrid_u8u32_dot_4x6/a55.cpp | 4491 ++++++++++---------- 1 file changed, 2249 insertions(+), 2242 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_4x6/a55.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_4x6/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_4x6/a55.cpp index 8bf5debc37..ceb4a3b2d3 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_4x6/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_4x6/a55.cpp @@ -25,14 +25,15 @@ #include +#include "arm_gemm.hpp" + #include #include "../../asmlib.hpp" #include "../../utils.hpp" namespace arm_gemm { -void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_t *B, uint32_t *C, int ldc, uint32_t beta, int M, int N, int K) { - UNUSED(beta); +void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_t *B, uint32_t *C, int ldc, int M, int N, int K, const uint32_t *, Activation, bool) { const long loops_count = iceildiv(N, (int)4) - 1; const long ldab = lda * sizeof(uint8_t); const long ldcb = ldc * sizeof(uint32_t); @@ -97,140 +98,155 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "add a_ptr1, %[a_ptr0], #0x0\n" "1:\n" "ldr q0, [%[a_ptr0]], #0x10\n" - "ldr q4, [a_ptr1], #0x10\n" - "ldr q8, [a_ptr2], #0x10\n" - "ldr q12, [a_ptr3], #0x10\n" - "ldr q16, [a_ptr4], #0x10\n" - "ldr q20, [a_ptr5], #0x10\n" + "ldr q3, [a_ptr1], #0x10\n" + "ldr q6, [a_ptr2], #0x10\n" + "ldr q9, [a_ptr3], #0x10\n" + "ldr q12, [a_ptr4], #0x10\n" + "ldr q15, [a_ptr5], #0x10\n" "ldr q1, [%[a_ptr0]], #0x10\n" - "ldr q5, [a_ptr1], #0x10\n" - "ldr q9, [a_ptr2], #0x10\n" - "ldr q13, [a_ptr3], #0x10\n" - "ldr q17, [a_ptr4], #0x10\n" - "ldr q21, [a_ptr5], #0x10\n" + "ldr q4, [a_ptr1], #0x10\n" + "ldr q7, [a_ptr2], #0x10\n" + "ldr q10, [a_ptr3], #0x10\n" + "ldr q13, [a_ptr4], #0x10\n" + "ldr q16, [a_ptr5], #0x10\n" "cbnz %[odds], 2f\n" "ldr s2, [%[a_ptr0]]\n" - "ldr s6, [a_ptr1]\n" - "ldr s10, [a_ptr2]\n" - "ldr s14, [a_ptr3]\n" - "ldr s18, [a_ptr4]\n" - "ldr s22, [a_ptr5]\n" + "ldr s5, [a_ptr1]\n" + "ldr s8, [a_ptr2]\n" + "ldr s11, [a_ptr3]\n" + "ldr s14, [a_ptr4]\n" + "ldr s17, [a_ptr5]\n" "b 3f\n" "2:\n" "subs %[odds], %[odds], #0x1\n" "b.ne 4f\n" "ldr b2, [%[a_ptr0]]\n" - "ldr b6, [a_ptr1]\n" - "ldr b10, [a_ptr2]\n" - "ldr b14, [a_ptr3]\n" - "ldr b18, [a_ptr4]\n" - "ldr b22, [a_ptr5]\n" + "ldr b5, [a_ptr1]\n" + "ldr b8, [a_ptr2]\n" + "ldr b11, [a_ptr3]\n" + "ldr b14, [a_ptr4]\n" + "ldr b17, [a_ptr5]\n" "b 3f\n" "4:\n" "ldr h2, [%[a_ptr0]], #0x2\n" - "ldr h6, [a_ptr1], #0x2\n" - "ldr h10, [a_ptr2], #0x2\n" - "ldr h14, [a_ptr3], #0x2\n" - "ldr h18, [a_ptr4], #0x2\n" - "ldr h22, [a_ptr5], #0x2\n" + "ldr h5, [a_ptr1], #0x2\n" + "ldr h8, [a_ptr2], #0x2\n" + "ldr h11, [a_ptr3], #0x2\n" + "ldr h14, [a_ptr4], #0x2\n" + "ldr h17, [a_ptr5], #0x2\n" "subs %[odds], %[odds], #0x1\n" "b.ne 5f\n" "b 3f\n" "5:\n" "ld1 {v2.b}[2], [%[a_ptr0]]\n" - "ld1 {v6.b}[2], [a_ptr1]\n" - "ld1 {v10.b}[2], [a_ptr2]\n" - "ld1 {v14.b}[2], [a_ptr3]\n" - "ld1 {v18.b}[2], [a_ptr4]\n" - "ld1 {v22.b}[2], [a_ptr5]\n" + "ld1 {v5.b}[2], [a_ptr1]\n" + "ld1 {v8.b}[2], [a_ptr2]\n" + "ld1 {v11.b}[2], [a_ptr3]\n" + "ld1 {v14.b}[2], [a_ptr4]\n" + "ld1 {v17.b}[2], [a_ptr5]\n" "3:\n" "movi v26.4s, #0\n" - "ldr q24, [%[b_ptr0]]\n" + "ldr q18, [%[b_ptr0]]\n" "movi v27.4s, #0\n" - "ldr q25, [%[b_ptr0], #0x10]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" "movi v28.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0x40]\n" + "ldr q20, [%[b_ptr0], #0x20]\n" "movi v29.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0x80]\n" + "ldr q21, [%[b_ptr0], #0x30]\n" "movi v30.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" + "ldr q22, [%[b_ptr0], #0x40]\n" "movi v31.4s, #0\n" + "ldr q23, [%[b_ptr0], #0x50]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" + "ldr q24, [%[b_ptr0], #0x60]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" + "ldr q25, [%[b_ptr0], #0x70]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0x40]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0x80]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x100]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" "prfm PLDL1KEEP, [a_ptr5, #0x140]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" "prfm PLDL1KEEP, [a_ptr5, #0x180]\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" + "ldr q18, [%[b_ptr0]]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" "cbz %[loops], 6f\n" - "ldr d24, [%[b_ptr0]]\n" + "ldr d18, [%[b_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + "ldr d21, [%[b_ptr0], #0x30]\n" + "ldr temploadreg1, [%[b_ptr0], #0x38]\n" + "ldr d22, [%[b_ptr0], #0x40]\n" + "ins v18.d[1], temploadreg2\n" + "ldr temploadreg2, [%[b_ptr0], #0x48]\n" + "ldr d23, [%[b_ptr0], #0x50]\n" + "ins v19.d[1], temploadreg3\n" + "ldr temploadreg3, [%[b_ptr0], #0x58]\n" + "ldr d24, [%[b_ptr0], #0x60]\n" + "ins v20.d[1], temploadreg0\n" + "ldr temploadreg0, [%[b_ptr0], #0x68]\n" + "ldr d25, [%[b_ptr0], #0x70]\n" + "ins v21.d[1], temploadreg1\n" + "ldr temploadreg1, [%[b_ptr0], #0x78]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + "ins v22.d[1], temploadreg2\n" + "ins v23.d[1], temploadreg3\n" "ins v24.d[1], temploadreg0\n" "ins v25.d[1], temploadreg1\n" "b.eq 7f\n" @@ -238,114 +254,111 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "str q26, [%[c_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" "movi v26.4s, #0\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" "prfm PSTL1KEEP, [c_ptr1, #0x40]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" + "ldr d18, [%[b_ptr0]]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" + "add %[b_ptr0], %[b_ptr0], #0x10\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" + "ins v18.d[1], temploadreg2\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + "ldr temploadreg1, [%[b_ptr0], #0x38]\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" "prfm PSTL1KEEP, [c_ptr2, #0x40]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + "ins v19.d[1], temploadreg3\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + "ldr temploadreg3, [%[b_ptr0], #0x58]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" "prfm PSTL1KEEP, [c_ptr3, #0x40]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" "prfm PSTL1KEEP, [c_ptr4, #0x40]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + "ins v20.d[1], temploadreg0\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + "ldr temploadreg0, [%[b_ptr0], #0x68]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" "prfm PSTL1KEEP, [c_ptr5, #0x40]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + "ldr d21, [%[b_ptr0], #0x30]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + "ins v21.d[1], temploadreg1\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + "ldr temploadreg1, [%[b_ptr0], #0x78]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + "ldr d22, [%[b_ptr0], #0x40]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + "ldr d23, [%[b_ptr0], #0x50]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + "ins v23.d[1], temploadreg3\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + "ldr d24, [%[b_ptr0], #0x60]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + "ldr d25, [%[b_ptr0], #0x70]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - "ins v24.d[1], temploadreg0\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + "ldr d18, [%[b_ptr0]]\n" + "ins v18.d[1], temploadreg2\n" + "ldr temploadreg2, [%[b_ptr0], #0x48]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + "ins v22.d[1], temploadreg2\n" "b.ne 8b\n" "7:\n" "str q26, [%[c_ptr0]]\n" @@ -354,83 +367,74 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" + "ldr q18, [%[b_ptr0]]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" "6:\n" "str q26, [%[c_ptr0]]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" @@ -511,369 +515,372 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "add a_ptr1, %[a_ptr0], #0x0\n" "1:\n" "ldr q0, [%[a_ptr0]], #0x10\n" - "ldr q4, [a_ptr1], #0x10\n" - "ldr q8, [a_ptr2], #0x10\n" - "ldr q12, [a_ptr3], #0x10\n" - "ldr q16, [a_ptr4], #0x10\n" - "ldr q20, [a_ptr5], #0x10\n" + "ldr q3, [a_ptr1], #0x10\n" + "ldr q6, [a_ptr2], #0x10\n" + "ldr q9, [a_ptr3], #0x10\n" + "ldr q12, [a_ptr4], #0x10\n" + "ldr q15, [a_ptr5], #0x10\n" "ldr q1, [%[a_ptr0]], #0x10\n" - "ldr q5, [a_ptr1], #0x10\n" - "ldr q9, [a_ptr2], #0x10\n" - "ldr q13, [a_ptr3], #0x10\n" - "ldr q17, [a_ptr4], #0x10\n" - "ldr q21, [a_ptr5], #0x10\n" + "ldr q4, [a_ptr1], #0x10\n" + "ldr q7, [a_ptr2], #0x10\n" + "ldr q10, [a_ptr3], #0x10\n" + "ldr q13, [a_ptr4], #0x10\n" + "ldr q16, [a_ptr5], #0x10\n" "cbnz %[odds], 2f\n" "ldr d2, [%[a_ptr0]]\n" - "ldr d6, [a_ptr1]\n" - "ldr d10, [a_ptr2]\n" - "ldr d14, [a_ptr3]\n" - "ldr d18, [a_ptr4]\n" - "ldr d22, [a_ptr5]\n" + "ldr d5, [a_ptr1]\n" + "ldr d8, [a_ptr2]\n" + "ldr d11, [a_ptr3]\n" + "ldr d14, [a_ptr4]\n" + "ldr d17, [a_ptr5]\n" "b 3f\n" "2:\n" "ldr s2, [%[a_ptr0]], #0x4\n" - "ldr s6, [a_ptr1], #0x4\n" - "ldr s10, [a_ptr2], #0x4\n" - "ldr s14, [a_ptr3], #0x4\n" - "ldr s18, [a_ptr4], #0x4\n" - "ldr s22, [a_ptr5], #0x4\n" + "ldr s5, [a_ptr1], #0x4\n" + "ldr s8, [a_ptr2], #0x4\n" + "ldr s11, [a_ptr3], #0x4\n" + "ldr s14, [a_ptr4], #0x4\n" + "ldr s17, [a_ptr5], #0x4\n" "subs %[odds], %[odds], #0x1\n" "b.ne 4f\n" "ld1 {v2.b}[4], [%[a_ptr0]]\n" - "ld1 {v6.b}[4], [a_ptr1]\n" - "ld1 {v10.b}[4], [a_ptr2]\n" - "ld1 {v14.b}[4], [a_ptr3]\n" - "ld1 {v18.b}[4], [a_ptr4]\n" - "ld1 {v22.b}[4], [a_ptr5]\n" + "ld1 {v5.b}[4], [a_ptr1]\n" + "ld1 {v8.b}[4], [a_ptr2]\n" + "ld1 {v11.b}[4], [a_ptr3]\n" + "ld1 {v14.b}[4], [a_ptr4]\n" + "ld1 {v17.b}[4], [a_ptr5]\n" "b 3f\n" "4:\n" "ld1 {v2.h}[2], [%[a_ptr0]], #2\n" - "ld1 {v6.h}[2], [a_ptr1], #2\n" - "ld1 {v10.h}[2], [a_ptr2], #2\n" - "ld1 {v14.h}[2], [a_ptr3], #2\n" - "ld1 {v18.h}[2], [a_ptr4], #2\n" - "ld1 {v22.h}[2], [a_ptr5], #2\n" + "ld1 {v5.h}[2], [a_ptr1], #2\n" + "ld1 {v8.h}[2], [a_ptr2], #2\n" + "ld1 {v11.h}[2], [a_ptr3], #2\n" + "ld1 {v14.h}[2], [a_ptr4], #2\n" + "ld1 {v17.h}[2], [a_ptr5], #2\n" "subs %[odds], %[odds], #0x1\n" "b.ne 5f\n" "b 3f\n" "5:\n" "ld1 {v2.b}[6], [%[a_ptr0]]\n" - "ld1 {v6.b}[6], [a_ptr1]\n" - "ld1 {v10.b}[6], [a_ptr2]\n" - "ld1 {v14.b}[6], [a_ptr3]\n" - "ld1 {v18.b}[6], [a_ptr4]\n" - "ld1 {v22.b}[6], [a_ptr5]\n" + "ld1 {v5.b}[6], [a_ptr1]\n" + "ld1 {v8.b}[6], [a_ptr2]\n" + "ld1 {v11.b}[6], [a_ptr3]\n" + "ld1 {v14.b}[6], [a_ptr4]\n" + "ld1 {v17.b}[6], [a_ptr5]\n" "3:\n" "movi v26.4s, #0\n" - "ldr q24, [%[b_ptr0]]\n" + "ldr q18, [%[b_ptr0]]\n" "movi v27.4s, #0\n" - "ldr q25, [%[b_ptr0], #0x10]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" "movi v28.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0x40]\n" + "ldr q20, [%[b_ptr0], #0x20]\n" "movi v29.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0x80]\n" + "ldr q21, [%[b_ptr0], #0x30]\n" "movi v30.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" + "ldr q22, [%[b_ptr0], #0x40]\n" "movi v31.4s, #0\n" + "ldr q23, [%[b_ptr0], #0x50]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" + "ldr q24, [%[b_ptr0], #0x60]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" + "ldr q25, [%[b_ptr0], #0x70]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0x40]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0x80]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x100]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" "prfm PLDL1KEEP, [a_ptr5, #0x140]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" "prfm PLDL1KEEP, [a_ptr5, #0x180]\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" + "ldr q18, [%[b_ptr0]]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + "add %[b_ptr0], %[b_ptr0], #0x20\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + ".inst 0x6fa2e27a // udot v26.4s, v19.16b, v2.4b[1]\n" + ".inst 0x6fa5e27b // udot v27.4s, v19.16b, v5.4b[1]\n" + ".inst 0x6fa8e27c // udot v28.4s, v19.16b, v8.4b[1]\n" + ".inst 0x6fabe27d // udot v29.4s, v19.16b, v11.4b[1]\n" + ".inst 0x6faee27e // udot v30.4s, v19.16b, v14.4b[1]\n" + ".inst 0x6fb1e27f // udot v31.4s, v19.16b, v17.4b[1]\n" "cbz %[loops], 6f\n" - "ldr d24, [%[b_ptr0]]\n" + "ldr d18, [%[b_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + "ldr d21, [%[b_ptr0], #0x30]\n" + "ldr temploadreg1, [%[b_ptr0], #0x38]\n" + "ldr d22, [%[b_ptr0], #0x40]\n" + "ins v18.d[1], temploadreg2\n" + "ldr temploadreg2, [%[b_ptr0], #0x48]\n" + "ldr d23, [%[b_ptr0], #0x50]\n" + "ins v19.d[1], temploadreg3\n" + "ldr temploadreg3, [%[b_ptr0], #0x58]\n" + "ldr d24, [%[b_ptr0], #0x60]\n" + "ins v20.d[1], temploadreg0\n" + "ldr temploadreg0, [%[b_ptr0], #0x68]\n" + "ldr d25, [%[b_ptr0], #0x70]\n" + "ins v21.d[1], temploadreg1\n" + "ldr temploadreg1, [%[b_ptr0], #0x78]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + "ins v22.d[1], temploadreg2\n" + "ins v23.d[1], temploadreg3\n" "ins v24.d[1], temploadreg0\n" + "ins v25.d[1], temploadreg1\n" "b.eq 7f\n" "8:\n" "str q26, [%[c_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" "movi v26.4s, #0\n" - "ins v25.d[1], temploadreg1\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" - "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + "add c_ptr2, c_ptr2, #0x10\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" - "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + "add c_ptr3, c_ptr3, #0x10\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" - "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + "add c_ptr4, c_ptr4, #0x10\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" - "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" "prfm PSTL1KEEP, [c_ptr1, #0x40]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" + "ldr d18, [%[b_ptr0]]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" "prfm PSTL1KEEP, [c_ptr2, #0x40]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" + "ins v18.d[1], temploadreg2\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" "prfm PSTL1KEEP, [c_ptr3, #0x40]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" "prfm PSTL1KEEP, [c_ptr4, #0x40]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" "prfm PSTL1KEEP, [c_ptr5, #0x40]\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + "add %[b_ptr0], %[b_ptr0], #0x20\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + "ins v19.d[1], temploadreg3\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + "ldr temploadreg1, [%[b_ptr0], #0x38]\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + "ins v20.d[1], temploadreg0\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + "ldr temploadreg0, [%[b_ptr0], #0x68]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + "ldr d21, [%[b_ptr0], #0x30]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + "ins v21.d[1], temploadreg1\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + "ldr temploadreg1, [%[b_ptr0], #0x78]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + "ldr d22, [%[b_ptr0], #0x40]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + "ldr d23, [%[b_ptr0], #0x50]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + "ldr d24, [%[b_ptr0], #0x60]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + "ldr d25, [%[b_ptr0], #0x70]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + "ldr d18, [%[b_ptr0]]\n" + ".inst 0x6fa2e27a // udot v26.4s, v19.16b, v2.4b[1]\n" + ".inst 0x6fa5e27b // udot v27.4s, v19.16b, v5.4b[1]\n" + ".inst 0x6fa8e27c // udot v28.4s, v19.16b, v8.4b[1]\n" + "ins v18.d[1], temploadreg2\n" + ".inst 0x6fabe27d // udot v29.4s, v19.16b, v11.4b[1]\n" + "ldr temploadreg2, [%[b_ptr0], #0x48]\n" + ".inst 0x6faee27e // udot v30.4s, v19.16b, v14.4b[1]\n" + ".inst 0x6fb1e27f // udot v31.4s, v19.16b, v17.4b[1]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + "ins v22.d[1], temploadreg2\n" + "ins v19.d[1], temploadreg3\n" + "ldr temploadreg3, [%[b_ptr0], #0x58]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + "ins v23.d[1], temploadreg3\n" "b.ne 8b\n" "7:\n" "str q26, [%[c_ptr0]]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" "movi v26.4s, #0\n" - "ins v25.d[1], temploadreg1\n" "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" + "ldr q18, [%[b_ptr0]]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + "add %[b_ptr0], %[b_ptr0], #0x20\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + ".inst 0x6fa2e27a // udot v26.4s, v19.16b, v2.4b[1]\n" + ".inst 0x6fa5e27b // udot v27.4s, v19.16b, v5.4b[1]\n" + ".inst 0x6fa8e27c // udot v28.4s, v19.16b, v8.4b[1]\n" + ".inst 0x6fabe27d // udot v29.4s, v19.16b, v11.4b[1]\n" + ".inst 0x6faee27e // udot v30.4s, v19.16b, v14.4b[1]\n" + ".inst 0x6fb1e27f // udot v31.4s, v19.16b, v17.4b[1]\n" "6:\n" "str q26, [%[c_ptr0]]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" @@ -954,161 +961,175 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "add a_ptr1, %[a_ptr0], #0x0\n" "1:\n" "ldr q0, [%[a_ptr0]], #0x10\n" - "ldr q4, [a_ptr1], #0x10\n" - "ldr q8, [a_ptr2], #0x10\n" - "ldr q12, [a_ptr3], #0x10\n" - "ldr q16, [a_ptr4], #0x10\n" - "ldr q20, [a_ptr5], #0x10\n" + "ldr q3, [a_ptr1], #0x10\n" + "ldr q6, [a_ptr2], #0x10\n" + "ldr q9, [a_ptr3], #0x10\n" + "ldr q12, [a_ptr4], #0x10\n" + "ldr q15, [a_ptr5], #0x10\n" "ldr q1, [%[a_ptr0]], #0x10\n" - "ldr q5, [a_ptr1], #0x10\n" - "ldr q9, [a_ptr2], #0x10\n" - "ldr q13, [a_ptr3], #0x10\n" + "ldr q4, [a_ptr1], #0x10\n" + "ldr q7, [a_ptr2], #0x10\n" + "ldr q10, [a_ptr3], #0x10\n" "ldr d2, [%[a_ptr0]], #0x8\n" - "ldr q17, [a_ptr4], #0x10\n" - "ldr d6, [a_ptr1], #0x8\n" - "ldr q21, [a_ptr5], #0x10\n" - "ldr d10, [a_ptr2], #0x8\n" - "ldr d14, [a_ptr3], #0x8\n" - "ldr d18, [a_ptr4], #0x8\n" - "ldr d22, [a_ptr5], #0x8\n" + "ldr q13, [a_ptr4], #0x10\n" + "ldr d5, [a_ptr1], #0x8\n" + "ldr q16, [a_ptr5], #0x10\n" + "ldr d8, [a_ptr2], #0x8\n" + "ldr d11, [a_ptr3], #0x8\n" + "ldr d14, [a_ptr4], #0x8\n" + "ldr d17, [a_ptr5], #0x8\n" "cbnz %[odds], 2f\n" "ld1 {v2.s}[2], [%[a_ptr0]]\n" - "ld1 {v6.s}[2], [a_ptr1]\n" - "ld1 {v10.s}[2], [a_ptr2]\n" - "ld1 {v14.s}[2], [a_ptr3]\n" - "ld1 {v18.s}[2], [a_ptr4]\n" - "ld1 {v22.s}[2], [a_ptr5]\n" + "ld1 {v5.s}[2], [a_ptr1]\n" + "ld1 {v8.s}[2], [a_ptr2]\n" + "ld1 {v11.s}[2], [a_ptr3]\n" + "ld1 {v14.s}[2], [a_ptr4]\n" + "ld1 {v17.s}[2], [a_ptr5]\n" "b 3f\n" "2:\n" "subs %[odds], %[odds], #0x1\n" "b.ne 4f\n" "ld1 {v2.b}[8], [%[a_ptr0]]\n" - "ld1 {v6.b}[8], [a_ptr1]\n" - "ld1 {v10.b}[8], [a_ptr2]\n" - "ld1 {v14.b}[8], [a_ptr3]\n" - "ld1 {v18.b}[8], [a_ptr4]\n" - "ld1 {v22.b}[8], [a_ptr5]\n" + "ld1 {v5.b}[8], [a_ptr1]\n" + "ld1 {v8.b}[8], [a_ptr2]\n" + "ld1 {v11.b}[8], [a_ptr3]\n" + "ld1 {v14.b}[8], [a_ptr4]\n" + "ld1 {v17.b}[8], [a_ptr5]\n" "b 3f\n" "4:\n" "ld1 {v2.h}[4], [%[a_ptr0]], #2\n" - "ld1 {v6.h}[4], [a_ptr1], #2\n" - "ld1 {v10.h}[4], [a_ptr2], #2\n" - "ld1 {v14.h}[4], [a_ptr3], #2\n" - "ld1 {v18.h}[4], [a_ptr4], #2\n" - "ld1 {v22.h}[4], [a_ptr5], #2\n" + "ld1 {v5.h}[4], [a_ptr1], #2\n" + "ld1 {v8.h}[4], [a_ptr2], #2\n" + "ld1 {v11.h}[4], [a_ptr3], #2\n" + "ld1 {v14.h}[4], [a_ptr4], #2\n" + "ld1 {v17.h}[4], [a_ptr5], #2\n" "subs %[odds], %[odds], #0x1\n" "b.ne 5f\n" "b 3f\n" "5:\n" "ld1 {v2.b}[10], [%[a_ptr0]]\n" - "ld1 {v6.b}[10], [a_ptr1]\n" - "ld1 {v10.b}[10], [a_ptr2]\n" - "ld1 {v14.b}[10], [a_ptr3]\n" - "ld1 {v18.b}[10], [a_ptr4]\n" - "ld1 {v22.b}[10], [a_ptr5]\n" + "ld1 {v5.b}[10], [a_ptr1]\n" + "ld1 {v8.b}[10], [a_ptr2]\n" + "ld1 {v11.b}[10], [a_ptr3]\n" + "ld1 {v14.b}[10], [a_ptr4]\n" + "ld1 {v17.b}[10], [a_ptr5]\n" "3:\n" "movi v26.4s, #0\n" - "ldr q24, [%[b_ptr0]]\n" + "ldr q18, [%[b_ptr0]]\n" "movi v27.4s, #0\n" - "ldr q25, [%[b_ptr0], #0x10]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" "movi v28.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0x40]\n" + "ldr q20, [%[b_ptr0], #0x20]\n" "movi v29.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0x80]\n" + "ldr q21, [%[b_ptr0], #0x30]\n" "movi v30.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" + "ldr q22, [%[b_ptr0], #0x40]\n" "movi v31.4s, #0\n" + "ldr q23, [%[b_ptr0], #0x50]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" + "ldr q24, [%[b_ptr0], #0x60]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" + "ldr q25, [%[b_ptr0], #0x70]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0x40]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0x80]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x100]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" "prfm PLDL1KEEP, [a_ptr5, #0x140]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" "prfm PLDL1KEEP, [a_ptr5, #0x180]\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" + "ldr q18, [%[b_ptr0]]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + "ldr q20, [%[b_ptr0], #0x20]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + "add %[b_ptr0], %[b_ptr0], #0x30\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + ".inst 0x6fa2e27a // udot v26.4s, v19.16b, v2.4b[1]\n" + ".inst 0x6fa5e27b // udot v27.4s, v19.16b, v5.4b[1]\n" + ".inst 0x6fa8e27c // udot v28.4s, v19.16b, v8.4b[1]\n" + ".inst 0x6fabe27d // udot v29.4s, v19.16b, v11.4b[1]\n" + ".inst 0x6faee27e // udot v30.4s, v19.16b, v14.4b[1]\n" + ".inst 0x6fb1e27f // udot v31.4s, v19.16b, v17.4b[1]\n" + ".inst 0x6f82ea9a // udot v26.4s, v20.16b, v2.4b[2]\n" + ".inst 0x6f85ea9b // udot v27.4s, v20.16b, v5.4b[2]\n" + ".inst 0x6f88ea9c // udot v28.4s, v20.16b, v8.4b[2]\n" + ".inst 0x6f8bea9d // udot v29.4s, v20.16b, v11.4b[2]\n" + ".inst 0x6f8eea9e // udot v30.4s, v20.16b, v14.4b[2]\n" + ".inst 0x6f91ea9f // udot v31.4s, v20.16b, v17.4b[2]\n" "cbz %[loops], 6f\n" - "ldr d24, [%[b_ptr0]]\n" + "ldr d18, [%[b_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + "ldr d21, [%[b_ptr0], #0x30]\n" + "ldr temploadreg1, [%[b_ptr0], #0x38]\n" + "ldr d22, [%[b_ptr0], #0x40]\n" + "ins v18.d[1], temploadreg2\n" + "ldr temploadreg2, [%[b_ptr0], #0x48]\n" + "ldr d23, [%[b_ptr0], #0x50]\n" + "ins v19.d[1], temploadreg3\n" + "ldr temploadreg3, [%[b_ptr0], #0x58]\n" + "ldr d24, [%[b_ptr0], #0x60]\n" + "ins v20.d[1], temploadreg0\n" + "ldr temploadreg0, [%[b_ptr0], #0x68]\n" + "ldr d25, [%[b_ptr0], #0x70]\n" + "ins v21.d[1], temploadreg1\n" + "ldr temploadreg1, [%[b_ptr0], #0x78]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + "ins v22.d[1], temploadreg2\n" + "ins v23.d[1], temploadreg3\n" "ins v24.d[1], temploadreg0\n" "ins v25.d[1], temploadreg1\n" "b.eq 7f\n" @@ -1116,132 +1137,128 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "str q26, [%[c_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" "movi v26.4s, #0\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" + "ldr d18, [%[b_ptr0]]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" "prfm PSTL1KEEP, [c_ptr1, #0x40]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" + "ins v18.d[1], temploadreg2\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" "prfm PSTL1KEEP, [c_ptr2, #0x40]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" "prfm PSTL1KEEP, [c_ptr3, #0x40]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" "prfm PSTL1KEEP, [c_ptr4, #0x40]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" "prfm PSTL1KEEP, [c_ptr5, #0x40]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + "ins v19.d[1], temploadreg3\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + "add %[b_ptr0], %[b_ptr0], #0x30\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + "ins v20.d[1], temploadreg0\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + "ldr d21, [%[b_ptr0], #0x30]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + "ldr temploadreg1, [%[b_ptr0], #0x38]\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + "ins v21.d[1], temploadreg1\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + "ldr temploadreg1, [%[b_ptr0], #0x78]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + "ldr d22, [%[b_ptr0], #0x40]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + "ldr d23, [%[b_ptr0], #0x50]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + "ldr d24, [%[b_ptr0], #0x60]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + "ldr d25, [%[b_ptr0], #0x70]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + "ldr d18, [%[b_ptr0]]\n" + ".inst 0x6fa2e27a // udot v26.4s, v19.16b, v2.4b[1]\n" + ".inst 0x6fa5e27b // udot v27.4s, v19.16b, v5.4b[1]\n" + ".inst 0x6fa8e27c // udot v28.4s, v19.16b, v8.4b[1]\n" + "ins v18.d[1], temploadreg2\n" + ".inst 0x6fabe27d // udot v29.4s, v19.16b, v11.4b[1]\n" + "ldr temploadreg2, [%[b_ptr0], #0x48]\n" + ".inst 0x6faee27e // udot v30.4s, v19.16b, v14.4b[1]\n" + ".inst 0x6fb1e27f // udot v31.4s, v19.16b, v17.4b[1]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f82ea9a // udot v26.4s, v20.16b, v2.4b[2]\n" + "ins v22.d[1], temploadreg2\n" + ".inst 0x6f85ea9b // udot v27.4s, v20.16b, v5.4b[2]\n" + ".inst 0x6f88ea9c // udot v28.4s, v20.16b, v8.4b[2]\n" + "ins v19.d[1], temploadreg3\n" + ".inst 0x6f8bea9d // udot v29.4s, v20.16b, v11.4b[2]\n" + "ldr temploadreg3, [%[b_ptr0], #0x58]\n" + ".inst 0x6f8eea9e // udot v30.4s, v20.16b, v14.4b[2]\n" + ".inst 0x6f91ea9f // udot v31.4s, v20.16b, v17.4b[2]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + "ins v23.d[1], temploadreg3\n" + "ins v20.d[1], temploadreg0\n" + "ldr temploadreg0, [%[b_ptr0], #0x68]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" "ins v24.d[1], temploadreg0\n" "b.ne 8b\n" "7:\n" @@ -1251,98 +1268,88 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" + "ldr q18, [%[b_ptr0]]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + "ldr q20, [%[b_ptr0], #0x20]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + "add %[b_ptr0], %[b_ptr0], #0x30\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + ".inst 0x6fa2e27a // udot v26.4s, v19.16b, v2.4b[1]\n" + ".inst 0x6fa5e27b // udot v27.4s, v19.16b, v5.4b[1]\n" + ".inst 0x6fa8e27c // udot v28.4s, v19.16b, v8.4b[1]\n" + ".inst 0x6fabe27d // udot v29.4s, v19.16b, v11.4b[1]\n" + ".inst 0x6faee27e // udot v30.4s, v19.16b, v14.4b[1]\n" + ".inst 0x6fb1e27f // udot v31.4s, v19.16b, v17.4b[1]\n" + ".inst 0x6f82ea9a // udot v26.4s, v20.16b, v2.4b[2]\n" + ".inst 0x6f85ea9b // udot v27.4s, v20.16b, v5.4b[2]\n" + ".inst 0x6f88ea9c // udot v28.4s, v20.16b, v8.4b[2]\n" + ".inst 0x6f8bea9d // udot v29.4s, v20.16b, v11.4b[2]\n" + ".inst 0x6f8eea9e // udot v30.4s, v20.16b, v14.4b[2]\n" + ".inst 0x6f91ea9f // udot v31.4s, v20.16b, v17.4b[2]\n" "6:\n" "str q26, [%[c_ptr0]]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" @@ -1423,174 +1430,188 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "add a_ptr1, %[a_ptr0], #0x0\n" "1:\n" "ldr q0, [%[a_ptr0]], #0x10\n" - "ldr q4, [a_ptr1], #0x10\n" - "ldr q8, [a_ptr2], #0x10\n" - "ldr q12, [a_ptr3], #0x10\n" - "ldr q16, [a_ptr4], #0x10\n" - "ldr q20, [a_ptr5], #0x10\n" + "ldr q3, [a_ptr1], #0x10\n" + "ldr q6, [a_ptr2], #0x10\n" + "ldr q9, [a_ptr3], #0x10\n" + "ldr q12, [a_ptr4], #0x10\n" + "ldr q15, [a_ptr5], #0x10\n" "ldr q1, [%[a_ptr0]], #0x10\n" - "ldr q5, [a_ptr1], #0x10\n" - "ldr q9, [a_ptr2], #0x10\n" - "ldr q13, [a_ptr3], #0x10\n" - "ldr q17, [a_ptr4], #0x10\n" - "ldr q21, [a_ptr5], #0x10\n" + "ldr q4, [a_ptr1], #0x10\n" + "ldr q7, [a_ptr2], #0x10\n" + "ldr q10, [a_ptr3], #0x10\n" + "ldr q13, [a_ptr4], #0x10\n" + "ldr q16, [a_ptr5], #0x10\n" "cbnz %[odds], 2f\n" "ldr q2, [%[a_ptr0]]\n" - "ldr q6, [a_ptr1]\n" - "ldr q10, [a_ptr2]\n" - "ldr q14, [a_ptr3]\n" - "ldr q18, [a_ptr4]\n" - "ldr q22, [a_ptr5]\n" + "ldr q5, [a_ptr1]\n" + "ldr q8, [a_ptr2]\n" + "ldr q11, [a_ptr3]\n" + "ldr q14, [a_ptr4]\n" + "ldr q17, [a_ptr5]\n" "b 3f\n" "2:\n" "ldr d2, [%[a_ptr0]], #0x8\n" - "ldr d6, [a_ptr1], #0x8\n" - "ldr d10, [a_ptr2], #0x8\n" - "ldr d14, [a_ptr3], #0x8\n" - "ldr d18, [a_ptr4], #0x8\n" - "ldr d22, [a_ptr5], #0x8\n" + "ldr d5, [a_ptr1], #0x8\n" + "ldr d8, [a_ptr2], #0x8\n" + "ldr d11, [a_ptr3], #0x8\n" + "ldr d14, [a_ptr4], #0x8\n" + "ldr d17, [a_ptr5], #0x8\n" "ld1 {v2.s}[2], [%[a_ptr0]], #4\n" - "ld1 {v6.s}[2], [a_ptr1], #4\n" - "ld1 {v10.s}[2], [a_ptr2], #4\n" - "ld1 {v14.s}[2], [a_ptr3], #4\n" - "ld1 {v18.s}[2], [a_ptr4], #4\n" - "ld1 {v22.s}[2], [a_ptr5], #4\n" + "ld1 {v5.s}[2], [a_ptr1], #4\n" + "ld1 {v8.s}[2], [a_ptr2], #4\n" + "ld1 {v11.s}[2], [a_ptr3], #4\n" + "ld1 {v14.s}[2], [a_ptr4], #4\n" + "ld1 {v17.s}[2], [a_ptr5], #4\n" "subs %[odds], %[odds], #0x1\n" "b.ne 4f\n" "ld1 {v2.b}[12], [%[a_ptr0]]\n" - "ld1 {v6.b}[12], [a_ptr1]\n" - "ld1 {v10.b}[12], [a_ptr2]\n" - "ld1 {v14.b}[12], [a_ptr3]\n" - "ld1 {v18.b}[12], [a_ptr4]\n" - "ld1 {v22.b}[12], [a_ptr5]\n" + "ld1 {v5.b}[12], [a_ptr1]\n" + "ld1 {v8.b}[12], [a_ptr2]\n" + "ld1 {v11.b}[12], [a_ptr3]\n" + "ld1 {v14.b}[12], [a_ptr4]\n" + "ld1 {v17.b}[12], [a_ptr5]\n" "b 3f\n" "4:\n" "ld1 {v2.h}[6], [%[a_ptr0]], #2\n" - "ld1 {v6.h}[6], [a_ptr1], #2\n" - "ld1 {v10.h}[6], [a_ptr2], #2\n" - "ld1 {v14.h}[6], [a_ptr3], #2\n" - "ld1 {v18.h}[6], [a_ptr4], #2\n" - "ld1 {v22.h}[6], [a_ptr5], #2\n" + "ld1 {v5.h}[6], [a_ptr1], #2\n" + "ld1 {v8.h}[6], [a_ptr2], #2\n" + "ld1 {v11.h}[6], [a_ptr3], #2\n" + "ld1 {v14.h}[6], [a_ptr4], #2\n" + "ld1 {v17.h}[6], [a_ptr5], #2\n" "subs %[odds], %[odds], #0x1\n" "b.ne 5f\n" "b 3f\n" "5:\n" "ld1 {v2.b}[14], [%[a_ptr0]]\n" - "ld1 {v6.b}[14], [a_ptr1]\n" - "ld1 {v10.b}[14], [a_ptr2]\n" - "ld1 {v14.b}[14], [a_ptr3]\n" - "ld1 {v18.b}[14], [a_ptr4]\n" - "ld1 {v22.b}[14], [a_ptr5]\n" + "ld1 {v5.b}[14], [a_ptr1]\n" + "ld1 {v8.b}[14], [a_ptr2]\n" + "ld1 {v11.b}[14], [a_ptr3]\n" + "ld1 {v14.b}[14], [a_ptr4]\n" + "ld1 {v17.b}[14], [a_ptr5]\n" "3:\n" "movi v26.4s, #0\n" - "ldr q24, [%[b_ptr0]]\n" + "ldr q18, [%[b_ptr0]]\n" "movi v27.4s, #0\n" - "ldr q25, [%[b_ptr0], #0x10]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" "movi v28.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0x40]\n" + "ldr q20, [%[b_ptr0], #0x20]\n" "movi v29.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0x80]\n" + "ldr q21, [%[b_ptr0], #0x30]\n" "movi v30.4s, #0\n" - "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" + "ldr q22, [%[b_ptr0], #0x40]\n" "movi v31.4s, #0\n" + "ldr q23, [%[b_ptr0], #0x50]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" + "ldr q24, [%[b_ptr0], #0x60]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" + "ldr q25, [%[b_ptr0], #0x70]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0x40]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0x80]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x100]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" "prfm PLDL1KEEP, [a_ptr5, #0x140]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" "prfm PLDL1KEEP, [a_ptr5, #0x180]\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" + "ldr q18, [%[b_ptr0]]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + "ldr q20, [%[b_ptr0], #0x20]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + "ldr q21, [%[b_ptr0], #0x30]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + "add %[b_ptr0], %[b_ptr0], #0x40\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + ".inst 0x6fa2e27a // udot v26.4s, v19.16b, v2.4b[1]\n" + ".inst 0x6fa5e27b // udot v27.4s, v19.16b, v5.4b[1]\n" + ".inst 0x6fa8e27c // udot v28.4s, v19.16b, v8.4b[1]\n" + ".inst 0x6fabe27d // udot v29.4s, v19.16b, v11.4b[1]\n" + ".inst 0x6faee27e // udot v30.4s, v19.16b, v14.4b[1]\n" + ".inst 0x6fb1e27f // udot v31.4s, v19.16b, v17.4b[1]\n" + ".inst 0x6f82ea9a // udot v26.4s, v20.16b, v2.4b[2]\n" + ".inst 0x6f85ea9b // udot v27.4s, v20.16b, v5.4b[2]\n" + ".inst 0x6f88ea9c // udot v28.4s, v20.16b, v8.4b[2]\n" + ".inst 0x6f8bea9d // udot v29.4s, v20.16b, v11.4b[2]\n" + ".inst 0x6f8eea9e // udot v30.4s, v20.16b, v14.4b[2]\n" + ".inst 0x6f91ea9f // udot v31.4s, v20.16b, v17.4b[2]\n" + ".inst 0x6fa2eaba // udot v26.4s, v21.16b, v2.4b[3]\n" + ".inst 0x6fa5eabb // udot v27.4s, v21.16b, v5.4b[3]\n" + ".inst 0x6fa8eabc // udot v28.4s, v21.16b, v8.4b[3]\n" + ".inst 0x6fabeabd // udot v29.4s, v21.16b, v11.4b[3]\n" + ".inst 0x6faeeabe // udot v30.4s, v21.16b, v14.4b[3]\n" + ".inst 0x6fb1eabf // udot v31.4s, v21.16b, v17.4b[3]\n" "cbz %[loops], 6f\n" - "ldr d24, [%[b_ptr0]]\n" + "ldr d18, [%[b_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + "ldr d21, [%[b_ptr0], #0x30]\n" + "ldr temploadreg1, [%[b_ptr0], #0x38]\n" + "ldr d22, [%[b_ptr0], #0x40]\n" + "ins v18.d[1], temploadreg2\n" + "ldr temploadreg2, [%[b_ptr0], #0x48]\n" + "ldr d23, [%[b_ptr0], #0x50]\n" + "ins v19.d[1], temploadreg3\n" + "ldr temploadreg3, [%[b_ptr0], #0x58]\n" + "ldr d24, [%[b_ptr0], #0x60]\n" + "ins v20.d[1], temploadreg0\n" + "ldr temploadreg0, [%[b_ptr0], #0x68]\n" + "ldr d25, [%[b_ptr0], #0x70]\n" + "ins v21.d[1], temploadreg1\n" + "ldr temploadreg1, [%[b_ptr0], #0x78]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" + "ins v22.d[1], temploadreg2\n" + "ins v23.d[1], temploadreg3\n" "ins v24.d[1], temploadreg0\n" "b.eq 7f\n" "8:\n" @@ -1598,141 +1619,137 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "subs %[loops], %[loops], #0x1\n" "movi v26.4s, #0\n" "ins v25.d[1], temploadreg1\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" - "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + "ldr temploadreg1, [%[b_ptr0], #0x38]\n" + "add c_ptr2, c_ptr2, #0x10\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" - "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + "add c_ptr3, c_ptr3, #0x10\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" - "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + "add c_ptr4, c_ptr4, #0x10\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" - "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" + "ldr d18, [%[b_ptr0]]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" "prfm PSTL1KEEP, [c_ptr1, #0x40]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" + "ins v18.d[1], temploadreg2\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" "prfm PSTL1KEEP, [c_ptr2, #0x40]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" "prfm PSTL1KEEP, [c_ptr3, #0x40]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" "prfm PSTL1KEEP, [c_ptr4, #0x40]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" "prfm PSTL1KEEP, [c_ptr5, #0x40]\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - "ins v24.d[1], temploadreg0\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - "ins v25.d[1], temploadreg1\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" - "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + "ins v19.d[1], temploadreg3\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + "ins v20.d[1], temploadreg0\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + "ldr d21, [%[b_ptr0], #0x30]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + "add %[b_ptr0], %[b_ptr0], #0x40\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + "ins v21.d[1], temploadreg1\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + "ldr temploadreg2, [%[b_ptr0], #0x8]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + "ldr temploadreg3, [%[b_ptr0], #0x18]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + "ldr temploadreg0, [%[b_ptr0], #0x28]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + "ldr temploadreg1, [%[b_ptr0], #0x38]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + "ldr d22, [%[b_ptr0], #0x40]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + "ldr d23, [%[b_ptr0], #0x50]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + "ldr d24, [%[b_ptr0], #0x60]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + "ldr d25, [%[b_ptr0], #0x70]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + "ldr d18, [%[b_ptr0]]\n" + ".inst 0x6fa2e27a // udot v26.4s, v19.16b, v2.4b[1]\n" + ".inst 0x6fa5e27b // udot v27.4s, v19.16b, v5.4b[1]\n" + ".inst 0x6fa8e27c // udot v28.4s, v19.16b, v8.4b[1]\n" + "ins v18.d[1], temploadreg2\n" + ".inst 0x6fabe27d // udot v29.4s, v19.16b, v11.4b[1]\n" + "ldr temploadreg2, [%[b_ptr0], #0x48]\n" + ".inst 0x6faee27e // udot v30.4s, v19.16b, v14.4b[1]\n" + ".inst 0x6fb1e27f // udot v31.4s, v19.16b, v17.4b[1]\n" + "ldr d19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f82ea9a // udot v26.4s, v20.16b, v2.4b[2]\n" + "ins v22.d[1], temploadreg2\n" + ".inst 0x6f85ea9b // udot v27.4s, v20.16b, v5.4b[2]\n" + ".inst 0x6f88ea9c // udot v28.4s, v20.16b, v8.4b[2]\n" + "ins v19.d[1], temploadreg3\n" + ".inst 0x6f8bea9d // udot v29.4s, v20.16b, v11.4b[2]\n" + "ldr temploadreg3, [%[b_ptr0], #0x58]\n" + ".inst 0x6f8eea9e // udot v30.4s, v20.16b, v14.4b[2]\n" + ".inst 0x6f91ea9f // udot v31.4s, v20.16b, v17.4b[2]\n" + "ldr d20, [%[b_ptr0], #0x20]\n" + ".inst 0x6fa2eaba // udot v26.4s, v21.16b, v2.4b[3]\n" + "ins v23.d[1], temploadreg3\n" + ".inst 0x6fa5eabb // udot v27.4s, v21.16b, v5.4b[3]\n" + ".inst 0x6fa8eabc // udot v28.4s, v21.16b, v8.4b[3]\n" + "ins v20.d[1], temploadreg0\n" + ".inst 0x6fabeabd // udot v29.4s, v21.16b, v11.4b[3]\n" + "ldr temploadreg0, [%[b_ptr0], #0x68]\n" + ".inst 0x6faeeabe // udot v30.4s, v21.16b, v14.4b[3]\n" + ".inst 0x6fb1eabf // udot v31.4s, v21.16b, v17.4b[3]\n" + "ldr d21, [%[b_ptr0], #0x30]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" - "ldr d25, [%[b_ptr0], #0x10]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" + "ins v21.d[1], temploadreg1\n" + "ldr temploadreg1, [%[b_ptr0], #0x78]\n" + "add %[b_ptr0], %[b_ptr0], #0x80\n" "b.ne 8b\n" "7:\n" "str q26, [%[c_ptr0]]\n" @@ -1742,105 +1759,95 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e25a // udot v26.4s, v18.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f83e25b // udot v27.4s, v18.16b, v3.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f86e25c // udot v28.4s, v18.16b, v6.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f89e25d // udot v29.4s, v18.16b, v9.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" - "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" - "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6f8ce25e // udot v30.4s, v18.16b, v12.4b[0]\n" + ".inst 0x6fa0e27a // udot v26.4s, v19.16b, v0.4b[1]\n" + ".inst 0x6f8fe25f // udot v31.4s, v18.16b, v15.4b[0]\n" + "ldr q18, [%[b_ptr0]]\n" + ".inst 0x6fa3e27b // udot v27.4s, v19.16b, v3.4b[1]\n" + ".inst 0x6fa6e27c // udot v28.4s, v19.16b, v6.4b[1]\n" + ".inst 0x6fa9e27d // udot v29.4s, v19.16b, v9.4b[1]\n" + ".inst 0x6face27e // udot v30.4s, v19.16b, v12.4b[1]\n" + ".inst 0x6fafe27f // udot v31.4s, v19.16b, v15.4b[1]\n" + "ldr q19, [%[b_ptr0], #0x10]\n" + ".inst 0x6f80ea9a // udot v26.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f83ea9b // udot v27.4s, v20.16b, v3.4b[2]\n" + ".inst 0x6f86ea9c // udot v28.4s, v20.16b, v6.4b[2]\n" + ".inst 0x6f89ea9d // udot v29.4s, v20.16b, v9.4b[2]\n" + ".inst 0x6f8cea9e // udot v30.4s, v20.16b, v12.4b[2]\n" + ".inst 0x6f8fea9f // udot v31.4s, v20.16b, v15.4b[2]\n" + "ldr q20, [%[b_ptr0], #0x20]\n" + ".inst 0x6fa0eaba // udot v26.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa3eabb // udot v27.4s, v21.16b, v3.4b[3]\n" + ".inst 0x6fa6eabc // udot v28.4s, v21.16b, v6.4b[3]\n" + ".inst 0x6fa9eabd // udot v29.4s, v21.16b, v9.4b[3]\n" + ".inst 0x6faceabe // udot v30.4s, v21.16b, v12.4b[3]\n" + ".inst 0x6fafeabf // udot v31.4s, v21.16b, v15.4b[3]\n" + "ldr q21, [%[b_ptr0], #0x30]\n" + ".inst 0x6f81e2da // udot v26.4s, v22.16b, v1.4b[0]\n" + "add %[b_ptr0], %[b_ptr0], #0x40\n" + ".inst 0x6f84e2db // udot v27.4s, v22.16b, v4.4b[0]\n" + ".inst 0x6f87e2dc // udot v28.4s, v22.16b, v7.4b[0]\n" + ".inst 0x6f8ae2dd // udot v29.4s, v22.16b, v10.4b[0]\n" + ".inst 0x6f8de2de // udot v30.4s, v22.16b, v13.4b[0]\n" + ".inst 0x6f90e2df // udot v31.4s, v22.16b, v16.4b[0]\n" + ".inst 0x6fa1e2fa // udot v26.4s, v23.16b, v1.4b[1]\n" + ".inst 0x6fa4e2fb // udot v27.4s, v23.16b, v4.4b[1]\n" + ".inst 0x6fa7e2fc // udot v28.4s, v23.16b, v7.4b[1]\n" + ".inst 0x6faae2fd // udot v29.4s, v23.16b, v10.4b[1]\n" + ".inst 0x6fade2fe // udot v30.4s, v23.16b, v13.4b[1]\n" + ".inst 0x6fb0e2ff // udot v31.4s, v23.16b, v16.4b[1]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f87eb1c // udot v28.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8aeb1d // udot v29.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8deb1e // udot v30.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f90eb1f // udot v31.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa7eb3c // udot v28.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6faaeb3d // udot v29.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fadeb3e // udot v30.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb0eb3f // udot v31.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6f82e25a // udot v26.4s, v18.16b, v2.4b[0]\n" + ".inst 0x6f85e25b // udot v27.4s, v18.16b, v5.4b[0]\n" + ".inst 0x6f88e25c // udot v28.4s, v18.16b, v8.4b[0]\n" + ".inst 0x6f8be25d // udot v29.4s, v18.16b, v11.4b[0]\n" + ".inst 0x6f8ee25e // udot v30.4s, v18.16b, v14.4b[0]\n" + ".inst 0x6f91e25f // udot v31.4s, v18.16b, v17.4b[0]\n" + ".inst 0x6fa2e27a // udot v26.4s, v19.16b, v2.4b[1]\n" + ".inst 0x6fa5e27b // udot v27.4s, v19.16b, v5.4b[1]\n" + ".inst 0x6fa8e27c // udot v28.4s, v19.16b, v8.4b[1]\n" + ".inst 0x6fabe27d // udot v29.4s, v19.16b, v11.4b[1]\n" + ".inst 0x6faee27e // udot v30.4s, v19.16b, v14.4b[1]\n" + ".inst 0x6fb1e27f // udot v31.4s, v19.16b, v17.4b[1]\n" + ".inst 0x6f82ea9a // udot v26.4s, v20.16b, v2.4b[2]\n" + ".inst 0x6f85ea9b // udot v27.4s, v20.16b, v5.4b[2]\n" + ".inst 0x6f88ea9c // udot v28.4s, v20.16b, v8.4b[2]\n" + ".inst 0x6f8bea9d // udot v29.4s, v20.16b, v11.4b[2]\n" + ".inst 0x6f8eea9e // udot v30.4s, v20.16b, v14.4b[2]\n" + ".inst 0x6f91ea9f // udot v31.4s, v20.16b, v17.4b[2]\n" + ".inst 0x6fa2eaba // udot v26.4s, v21.16b, v2.4b[3]\n" + ".inst 0x6fa5eabb // udot v27.4s, v21.16b, v5.4b[3]\n" + ".inst 0x6fa8eabc // udot v28.4s, v21.16b, v8.4b[3]\n" + ".inst 0x6fabeabd // udot v29.4s, v21.16b, v11.4b[3]\n" + ".inst 0x6faeeabe // udot v30.4s, v21.16b, v14.4b[3]\n" + ".inst 0x6fb1eabf // udot v31.4s, v21.16b, v17.4b[3]\n" "6:\n" "str q26, [%[c_ptr0]]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" @@ -1986,104 +1993,104 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" "movi v31.4s, #0\n" "prfm PLDL1KEEP, [a_ptr5, #0x100]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x140]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x180]\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "cbz %[loops], 6f\n" "ldr d24, [%[b_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" @@ -2105,142 +2112,142 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" "prfm PSTL1KEEP, [c_ptr1, #0x40]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" "prfm PSTL1KEEP, [c_ptr2, #0x40]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" "prfm PSTL1KEEP, [c_ptr3, #0x40]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" "prfm PSTL1KEEP, [c_ptr4, #0x40]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" "prfm PSTL1KEEP, [c_ptr5, #0x40]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" "ins v24.d[1], temploadreg0\n" @@ -2252,113 +2259,113 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "6:\n" "str q26, [%[c_ptr0]]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" @@ -2510,111 +2517,111 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" "movi v31.4s, #0\n" "prfm PLDL1KEEP, [a_ptr5, #0x100]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x140]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x180]\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" - ".word 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" - ".word 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" - ".word 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" - ".word 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" - ".word 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" - ".word 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + "add %[b_ptr0], %[b_ptr0], #0x20\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" + ".inst 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" + ".inst 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" + ".inst 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" + ".inst 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" + ".inst 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" "cbz %[loops], 6f\n" "ldr d24, [%[b_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" @@ -2635,152 +2642,152 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" "prfm PSTL1KEEP, [c_ptr1, #0x40]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" "prfm PSTL1KEEP, [c_ptr2, #0x40]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" "prfm PSTL1KEEP, [c_ptr3, #0x40]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" "prfm PSTL1KEEP, [c_ptr4, #0x40]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "prfm PSTL1KEEP, [c_ptr5, #0x40]\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" - ".word 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" - ".word 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" + ".inst 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" + ".inst 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" + ".inst 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" - ".word 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" - ".word 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" + ".inst 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" + ".inst 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" + ".inst 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" "b.ne 8b\n" @@ -2792,120 +2799,120 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" - ".word 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" - ".word 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" - ".word 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" - ".word 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" - ".word 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" - ".word 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + "add %[b_ptr0], %[b_ptr0], #0x20\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" + ".inst 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" + ".inst 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" + ".inst 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" + ".inst 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" + ".inst 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" "6:\n" "str q26, [%[c_ptr0]]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" @@ -3057,119 +3064,119 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" "movi v31.4s, #0\n" "prfm PLDL1KEEP, [a_ptr5, #0x100]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x140]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x180]\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" + ".inst 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" - ".word 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" - ".word 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" - ".word 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" - ".word 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" - ".word 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" - ".word 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" - ".word 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" - ".word 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" - ".word 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" - ".word 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" + ".inst 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" + ".inst 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" + ".inst 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" + ".inst 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" + ".inst 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" + ".inst 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" + ".inst 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" + ".inst 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" + ".inst 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" + ".inst 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" "cbz %[loops], 6f\n" "ldr d24, [%[b_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" @@ -3191,161 +3198,161 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" "prfm PSTL1KEEP, [c_ptr1, #0x40]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" "prfm PSTL1KEEP, [c_ptr2, #0x40]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" "prfm PSTL1KEEP, [c_ptr3, #0x40]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" "prfm PSTL1KEEP, [c_ptr4, #0x40]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" "prfm PSTL1KEEP, [c_ptr5, #0x40]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" + ".inst 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" + ".inst 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" + ".inst 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" + ".inst 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" - ".word 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" + ".inst 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" + ".inst 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" - ".word 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" - ".word 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" + ".inst 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" + ".inst 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" - ".word 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" - ".word 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" + ".inst 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" + ".inst 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" + ".inst 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" "ins v24.d[1], temploadreg0\n" @@ -3357,128 +3364,128 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" + ".inst 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" "add %[b_ptr0], %[b_ptr0], #0x10\n" - ".word 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" - ".word 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" - ".word 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" - ".word 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" - ".word 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" - ".word 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" - ".word 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" - ".word 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" - ".word 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" - ".word 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" - ".word 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" + ".inst 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" + ".inst 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" + ".inst 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" + ".inst 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" + ".inst 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" + ".inst 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" + ".inst 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" + ".inst 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" + ".inst 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" + ".inst 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" "6:\n" "str q26, [%[c_ptr0]]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" @@ -3637,126 +3644,126 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "prfm PLDL1KEEP, [a_ptr5, #0xc0]\n" "movi v31.4s, #0\n" "prfm PLDL1KEEP, [a_ptr5, #0x100]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x140]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "prfm PLDL1KEEP, [a_ptr5, #0x180]\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" - ".word 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" - ".word 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" - ".word 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" - ".word 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" - ".word 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" + ".inst 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" + ".inst 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" + ".inst 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" + ".inst 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" + ".inst 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" + ".inst 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" - ".word 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" - ".word 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" - ".word 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" - ".word 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" - ".word 0x6fa3eb3a // udot v26.4s, v25.16b, v3.4b[3]\n" - ".word 0x6fa7eb3b // udot v27.4s, v25.16b, v7.4b[3]\n" - ".word 0x6fabeb3c // udot v28.4s, v25.16b, v11.4b[3]\n" - ".word 0x6fafeb3d // udot v29.4s, v25.16b, v15.4b[3]\n" - ".word 0x6fb3eb3e // udot v30.4s, v25.16b, v19.4b[3]\n" - ".word 0x6fb7eb3f // udot v31.4s, v25.16b, v23.4b[3]\n" + ".inst 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" + "add %[b_ptr0], %[b_ptr0], #0x20\n" + ".inst 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" + ".inst 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" + ".inst 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" + ".inst 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" + ".inst 0x6fa3eb3a // udot v26.4s, v25.16b, v3.4b[3]\n" + ".inst 0x6fa7eb3b // udot v27.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6fabeb3c // udot v28.4s, v25.16b, v11.4b[3]\n" + ".inst 0x6fafeb3d // udot v29.4s, v25.16b, v15.4b[3]\n" + ".inst 0x6fb3eb3e // udot v30.4s, v25.16b, v19.4b[3]\n" + ".inst 0x6fb7eb3f // udot v31.4s, v25.16b, v23.4b[3]\n" "cbz %[loops], 6f\n" "ldr d24, [%[b_ptr0]]\n" "subs %[loops], %[loops], #0x1\n" @@ -3777,171 +3784,171 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "prfm PSTL1KEEP, [%[c_ptr0], #0x40]\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" "prfm PSTL1KEEP, [c_ptr1, #0x40]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" "prfm PSTL1KEEP, [c_ptr2, #0x40]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" "prfm PSTL1KEEP, [c_ptr3, #0x40]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" "prfm PSTL1KEEP, [c_ptr4, #0x40]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "prfm PSTL1KEEP, [c_ptr5, #0x40]\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" - ".word 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" - ".word 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" + ".inst 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" + ".inst 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" + ".inst 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" - ".word 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" - ".word 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" + ".inst 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" + ".inst 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" + ".inst 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" "ldr d25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" + ".inst 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" "ins v25.d[1], temploadreg1\n" - ".word 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" + ".inst 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" "ldr temploadreg0, [%[b_ptr0], #0x8]\n" - ".word 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" + ".inst 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" "ldr temploadreg1, [%[b_ptr0], #0x18]\n" - ".word 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" - ".word 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" + ".inst 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" + ".inst 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" "ldr d24, [%[b_ptr0]]\n" - ".word 0x6fa3eb3a // udot v26.4s, v25.16b, v3.4b[3]\n" - ".word 0x6fa7eb3b // udot v27.4s, v25.16b, v7.4b[3]\n" - ".word 0x6fabeb3c // udot v28.4s, v25.16b, v11.4b[3]\n" + ".inst 0x6fa3eb3a // udot v26.4s, v25.16b, v3.4b[3]\n" + ".inst 0x6fa7eb3b // udot v27.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6fabeb3c // udot v28.4s, v25.16b, v11.4b[3]\n" "ins v24.d[1], temploadreg0\n" - ".word 0x6fafeb3d // udot v29.4s, v25.16b, v15.4b[3]\n" - ".word 0x6fb3eb3e // udot v30.4s, v25.16b, v19.4b[3]\n" - ".word 0x6fb7eb3f // udot v31.4s, v25.16b, v23.4b[3]\n" + ".inst 0x6fafeb3d // udot v29.4s, v25.16b, v15.4b[3]\n" + ".inst 0x6fb3eb3e // udot v30.4s, v25.16b, v19.4b[3]\n" + ".inst 0x6fb7eb3f // udot v31.4s, v25.16b, v23.4b[3]\n" "ldr d25, [%[b_ptr0], #0x10]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" "b.ne 8b\n" @@ -3953,135 +3960,135 @@ void a64_smallK_hybrid_u8u32_dot_4x6_a55(const uint8_t *A, int lda, const uint8_ "str q27, [c_ptr1]\n" "add c_ptr1, c_ptr1, #0x10\n" "movi v27.4s, #0\n" - ".word 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f80e31a // udot v26.4s, v24.16b, v0.4b[0]\n" "str q28, [c_ptr2]\n" "movi v28.4s, #0\n" "add c_ptr2, c_ptr2, #0x10\n" - ".word 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" + ".inst 0x6f84e31b // udot v27.4s, v24.16b, v4.4b[0]\n" "str q29, [c_ptr3]\n" "movi v29.4s, #0\n" "add c_ptr3, c_ptr3, #0x10\n" - ".word 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" + ".inst 0x6f88e31c // udot v28.4s, v24.16b, v8.4b[0]\n" "str q30, [c_ptr4]\n" "movi v30.4s, #0\n" "add c_ptr4, c_ptr4, #0x10\n" - ".word 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" + ".inst 0x6f8ce31d // udot v29.4s, v24.16b, v12.4b[0]\n" "str q31, [c_ptr5]\n" "movi v31.4s, #0\n" "add c_ptr5, c_ptr5, #0x10\n" - ".word 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" - ".word 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" - ".word 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" + ".inst 0x6f90e31e // udot v30.4s, v24.16b, v16.4b[0]\n" + ".inst 0x6fa0e33a // udot v26.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6f94e31f // udot v31.4s, v24.16b, v20.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" - ".word 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" - ".word 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" - ".word 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" - ".word 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" + ".inst 0x6fa4e33b // udot v27.4s, v25.16b, v4.4b[1]\n" + ".inst 0x6fa8e33c // udot v28.4s, v25.16b, v8.4b[1]\n" + ".inst 0x6face33d // udot v29.4s, v25.16b, v12.4b[1]\n" + ".inst 0x6fb0e33e // udot v30.4s, v25.16b, v16.4b[1]\n" + ".inst 0x6fb4e33f // udot v31.4s, v25.16b, v20.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f80eb1a // udot v26.4s, v24.16b, v0.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" - ".word 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" - ".word 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" - ".word 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" - ".word 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" + ".inst 0x6f84eb1b // udot v27.4s, v24.16b, v4.4b[2]\n" + ".inst 0x6f88eb1c // udot v28.4s, v24.16b, v8.4b[2]\n" + ".inst 0x6f8ceb1d // udot v29.4s, v24.16b, v12.4b[2]\n" + ".inst 0x6f90eb1e // udot v30.4s, v24.16b, v16.4b[2]\n" + ".inst 0x6f94eb1f // udot v31.4s, v24.16b, v20.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" - ".word 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" - ".word 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" - ".word 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" - ".word 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" - ".word 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" + ".inst 0x6fa0eb3a // udot v26.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa4eb3b // udot v27.4s, v25.16b, v4.4b[3]\n" + ".inst 0x6fa8eb3c // udot v28.4s, v25.16b, v8.4b[3]\n" + ".inst 0x6faceb3d // udot v29.4s, v25.16b, v12.4b[3]\n" + ".inst 0x6fb0eb3e // udot v30.4s, v25.16b, v16.4b[3]\n" + ".inst 0x6fb4eb3f // udot v31.4s, v25.16b, v20.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f81e31a // udot v26.4s, v24.16b, v1.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" - ".word 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" - ".word 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" - ".word 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" - ".word 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" + ".inst 0x6f85e31b // udot v27.4s, v24.16b, v5.4b[0]\n" + ".inst 0x6f89e31c // udot v28.4s, v24.16b, v9.4b[0]\n" + ".inst 0x6f8de31d // udot v29.4s, v24.16b, v13.4b[0]\n" + ".inst 0x6f91e31e // udot v30.4s, v24.16b, v17.4b[0]\n" + ".inst 0x6f95e31f // udot v31.4s, v24.16b, v21.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" - ".word 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" - ".word 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" - ".word 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" - ".word 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" - ".word 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" + ".inst 0x6fa1e33a // udot v26.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa5e33b // udot v27.4s, v25.16b, v5.4b[1]\n" + ".inst 0x6fa9e33c // udot v28.4s, v25.16b, v9.4b[1]\n" + ".inst 0x6fade33d // udot v29.4s, v25.16b, v13.4b[1]\n" + ".inst 0x6fb1e33e // udot v30.4s, v25.16b, v17.4b[1]\n" + ".inst 0x6fb5e33f // udot v31.4s, v25.16b, v21.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f81eb1a // udot v26.4s, v24.16b, v1.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" - ".word 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" - ".word 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" - ".word 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" - ".word 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" + ".inst 0x6f85eb1b // udot v27.4s, v24.16b, v5.4b[2]\n" + ".inst 0x6f89eb1c // udot v28.4s, v24.16b, v9.4b[2]\n" + ".inst 0x6f8deb1d // udot v29.4s, v24.16b, v13.4b[2]\n" + ".inst 0x6f91eb1e // udot v30.4s, v24.16b, v17.4b[2]\n" + ".inst 0x6f95eb1f // udot v31.4s, v24.16b, v21.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" - ".word 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" - ".word 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" - ".word 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" - ".word 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" - ".word 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" + ".inst 0x6fa1eb3a // udot v26.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa5eb3b // udot v27.4s, v25.16b, v5.4b[3]\n" + ".inst 0x6fa9eb3c // udot v28.4s, v25.16b, v9.4b[3]\n" + ".inst 0x6fadeb3d // udot v29.4s, v25.16b, v13.4b[3]\n" + ".inst 0x6fb1eb3e // udot v30.4s, v25.16b, v17.4b[3]\n" + ".inst 0x6fb5eb3f // udot v31.4s, v25.16b, v21.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f82e31a // udot v26.4s, v24.16b, v2.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" - ".word 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" - ".word 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" - ".word 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" - ".word 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" + ".inst 0x6f86e31b // udot v27.4s, v24.16b, v6.4b[0]\n" + ".inst 0x6f8ae31c // udot v28.4s, v24.16b, v10.4b[0]\n" + ".inst 0x6f8ee31d // udot v29.4s, v24.16b, v14.4b[0]\n" + ".inst 0x6f92e31e // udot v30.4s, v24.16b, v18.4b[0]\n" + ".inst 0x6f96e31f // udot v31.4s, v24.16b, v22.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" - ".word 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" - ".word 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" - ".word 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" - ".word 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" - ".word 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" + ".inst 0x6fa2e33a // udot v26.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa6e33b // udot v27.4s, v25.16b, v6.4b[1]\n" + ".inst 0x6faae33c // udot v28.4s, v25.16b, v10.4b[1]\n" + ".inst 0x6faee33d // udot v29.4s, v25.16b, v14.4b[1]\n" + ".inst 0x6fb2e33e // udot v30.4s, v25.16b, v18.4b[1]\n" + ".inst 0x6fb6e33f // udot v31.4s, v25.16b, v22.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f82eb1a // udot v26.4s, v24.16b, v2.4b[2]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" - ".word 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" - ".word 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" - ".word 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" - ".word 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" + ".inst 0x6f86eb1b // udot v27.4s, v24.16b, v6.4b[2]\n" + ".inst 0x6f8aeb1c // udot v28.4s, v24.16b, v10.4b[2]\n" + ".inst 0x6f8eeb1d // udot v29.4s, v24.16b, v14.4b[2]\n" + ".inst 0x6f92eb1e // udot v30.4s, v24.16b, v18.4b[2]\n" + ".inst 0x6f96eb1f // udot v31.4s, v24.16b, v22.4b[2]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" - ".word 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" - ".word 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" - ".word 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" - ".word 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" - ".word 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" + ".inst 0x6fa2eb3a // udot v26.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa6eb3b // udot v27.4s, v25.16b, v6.4b[3]\n" + ".inst 0x6faaeb3c // udot v28.4s, v25.16b, v10.4b[3]\n" + ".inst 0x6faeeb3d // udot v29.4s, v25.16b, v14.4b[3]\n" + ".inst 0x6fb2eb3e // udot v30.4s, v25.16b, v18.4b[3]\n" + ".inst 0x6fb6eb3f // udot v31.4s, v25.16b, v22.4b[3]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" + ".inst 0x6f83e31a // udot v26.4s, v24.16b, v3.4b[0]\n" "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" - ".word 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" - ".word 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" - ".word 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" - ".word 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" + ".inst 0x6f87e31b // udot v27.4s, v24.16b, v7.4b[0]\n" + ".inst 0x6f8be31c // udot v28.4s, v24.16b, v11.4b[0]\n" + ".inst 0x6f8fe31d // udot v29.4s, v24.16b, v15.4b[0]\n" + ".inst 0x6f93e31e // udot v30.4s, v24.16b, v19.4b[0]\n" + ".inst 0x6f97e31f // udot v31.4s, v24.16b, v23.4b[0]\n" "ldr q24, [%[b_ptr0]]\n" - ".word 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" - ".word 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" - ".word 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" - ".word 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" - ".word 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" - ".word 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" + ".inst 0x6fa3e33a // udot v26.4s, v25.16b, v3.4b[1]\n" + ".inst 0x6fa7e33b // udot v27.4s, v25.16b, v7.4b[1]\n" + ".inst 0x6fabe33c // udot v28.4s, v25.16b, v11.4b[1]\n" + ".inst 0x6fafe33d // udot v29.4s, v25.16b, v15.4b[1]\n" + ".inst 0x6fb3e33e // udot v30.4s, v25.16b, v19.4b[1]\n" + ".inst 0x6fb7e33f // udot v31.4s, v25.16b, v23.4b[1]\n" "ldr q25, [%[b_ptr0], #0x10]\n" - ".word 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" - "add %[b_ptr0], %[b_ptr0], #0x20\n" - ".word 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" - ".word 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" - ".word 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" - ".word 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" - ".word 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" - ".word 0x6fa3eb3a // udot v26.4s, v25.16b, v3.4b[3]\n" - ".word 0x6fa7eb3b // udot v27.4s, v25.16b, v7.4b[3]\n" - ".word 0x6fabeb3c // udot v28.4s, v25.16b, v11.4b[3]\n" - ".word 0x6fafeb3d // udot v29.4s, v25.16b, v15.4b[3]\n" - ".word 0x6fb3eb3e // udot v30.4s, v25.16b, v19.4b[3]\n" - ".word 0x6fb7eb3f // udot v31.4s, v25.16b, v23.4b[3]\n" + ".inst 0x6f83eb1a // udot v26.4s, v24.16b, v3.4b[2]\n" + "add %[b_ptr0], %[b_ptr0], #0x20\n" + ".inst 0x6f87eb1b // udot v27.4s, v24.16b, v7.4b[2]\n" + ".inst 0x6f8beb1c // udot v28.4s, v24.16b, v11.4b[2]\n" + ".inst 0x6f8feb1d // udot v29.4s, v24.16b, v15.4b[2]\n" + ".inst 0x6f93eb1e // udot v30.4s, v24.16b, v19.4b[2]\n" + ".inst 0x6f97eb1f // udot v31.4s, v24.16b, v23.4b[2]\n" + ".inst 0x6fa3eb3a // udot v26.4s, v25.16b, v3.4b[3]\n" + ".inst 0x6fa7eb3b // udot v27.4s, v25.16b, v7.4b[3]\n" + ".inst 0x6fabeb3c // udot v28.4s, v25.16b, v11.4b[3]\n" + ".inst 0x6fafeb3d // udot v29.4s, v25.16b, v15.4b[3]\n" + ".inst 0x6fb3eb3e // udot v30.4s, v25.16b, v19.4b[3]\n" + ".inst 0x6fb7eb3f // udot v31.4s, v25.16b, v23.4b[3]\n" "6:\n" "str q26, [%[c_ptr0]]\n" "add %[c_ptr0], %[c_ptr0], #0x10\n" -- cgit v1.2.1