aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp3474
1 files changed, 1737 insertions, 1737 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp
index 87f46bb261..6301fa5657 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_16x4/generic.cpp
@@ -74,1755 +74,1755 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint
switch(M-y) {
case 1:
__asm __volatile (
- "cbz %[beta0], 1f\n"
- "movi v16.4s, #0\n"
- "ldr q0, [%[a_ptr0]]\n"
- "movi v17.4s, #0\n"
- "ldr q8, [%[b_ptr0]]\n"
- "movi v18.4s, #0\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "movi v19.4s, #0\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "b 3f\n"
- "1:\n"
- "ld1r {v15.4s}, [%[betaptr]]\n"
- "ldr q16, [%[c_ptr0]]\n"
- "ldr q17, [%[c_ptr0], #0x10]\n"
- "ldr q18, [%[c_ptr0], #0x20]\n"
- "ldr q19, [%[c_ptr0], #0x30]\n"
- "mul v16.4s, v16.4s, v15.4s\n"
- "ldr q0, [%[a_ptr0]]\n"
- "mul v17.4s, v17.4s, v15.4s\n"
- "ldr q8, [%[b_ptr0]]\n"
- "mul v18.4s, v18.4s, v15.4s\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "mul v19.4s, v19.4s, v15.4s\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "3:\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- "subs %[loops], %[loops], #0x1\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- "add %[a_ptr0], %[a_ptr0], #0x20\n"
- ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
- "ldr q0, [%[a_ptr0], #-0x10]\n"
- ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- "b.ne 3b\n"
- "2:\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- "prfm PSTL1KEEP, [%[c_ptr0]]\n"
- "cbz %[regs], 4f\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
- "b 5f\n"
- "4:\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- "5:\n"
- "cbz %[blocks], 6f\n"
- "7:\n"
- "ldr q8, [%[b_ptr0]]\n"
- "subs %[blocks], %[blocks], #0x1\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr s0, [%[a_ptr0]]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add %[a_ptr0], %[a_ptr0], #0x4\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add %[b_ptr0], %[b_ptr0], #0x40\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- "b.ne 7b\n"
- "6:\n"
- "cbz %[odds], 8f\n"
- "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
- "subs %[odds], %[odds], #0x1\n"
- "b.eq 9f\n"
- "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
- "subs %[odds], %[odds], #0x1\n"
- "b.eq 9f\n"
- "ld1 {v0.b}[2], [%[a_ptr0]]\n"
- "9:\n"
- "ldr q8, [%[b_ptr0]]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- "8:\n"
- "str q16, [%[c_ptr0]]\n"
- "str q17, [%[c_ptr0], #0x10]\n"
- "str q18, [%[c_ptr0], #0x20]\n"
- "str q19, [%[c_ptr0], #0x30]\n"
- "add %[c_ptr0], %[c_ptr0], #0x40\n"
- : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
- : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "cc", "memory"
+ "cbz %[beta0], 1f\n"
+ "movi v16.4s, #0\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "movi v17.4s, #0\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "movi v18.4s, #0\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "movi v19.4s, #0\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "b 3f\n"
+ "1:\n"
+ "ld1r {v15.4s}, [%[betaptr]]\n"
+ "ldr q16, [%[c_ptr0]]\n"
+ "ldr q17, [%[c_ptr0], #0x10]\n"
+ "ldr q18, [%[c_ptr0], #0x20]\n"
+ "ldr q19, [%[c_ptr0], #0x30]\n"
+ "mul v16.4s, v16.4s, v15.4s\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "mul v17.4s, v17.4s, v15.4s\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "mul v18.4s, v18.4s, v15.4s\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "mul v19.4s, v19.4s, v15.4s\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "3:\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ "subs %[loops], %[loops], #0x1\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x20\n"
+ ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
+ "ldr q0, [%[a_ptr0], #-0x10]\n"
+ ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ "b.ne 3b\n"
+ "2:\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0]]\n"
+ "cbz %[regs], 4f\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
+ "b 5f\n"
+ "4:\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ "5:\n"
+ "cbz %[blocks], 6f\n"
+ "7:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "subs %[blocks], %[blocks], #0x1\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr s0, [%[a_ptr0]]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x4\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ "b.ne 7b\n"
+ "6:\n"
+ "cbz %[odds], 8f\n"
+ "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[2], [%[a_ptr0]]\n"
+ "9:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ "8:\n"
+ "str q16, [%[c_ptr0]]\n"
+ "str q17, [%[c_ptr0], #0x10]\n"
+ "str q18, [%[c_ptr0], #0x20]\n"
+ "str q19, [%[c_ptr0], #0x30]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x40\n"
+ : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
+ : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
+ : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "cc", "memory"
);
break;
case 2:
__asm __volatile (
- "a_ptr1 .req X0\n"
- "c_ptr1 .req X1\n"
- "add a_ptr1, %[a_ptr0], %[lda]\n"
- "add c_ptr1, %[c_ptr0], %[ldc]\n"
- "cbz %[beta0], 1f\n"
- "movi v16.4s, #0\n"
- "ldr q0, [%[a_ptr0]]\n"
- "movi v17.4s, #0\n"
- "ldr q1, [a_ptr1]\n"
- "movi v18.4s, #0\n"
- "ldr q8, [%[b_ptr0]]\n"
- "movi v19.4s, #0\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "movi v20.4s, #0\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "movi v21.4s, #0\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "movi v22.4s, #0\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "movi v23.4s, #0\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "b 3f\n"
- "1:\n"
- "ld1r {v15.4s}, [%[betaptr]]\n"
- "ldr q16, [%[c_ptr0]]\n"
- "ldr q17, [%[c_ptr0], #0x10]\n"
- "ldr q18, [%[c_ptr0], #0x20]\n"
- "ldr q19, [%[c_ptr0], #0x30]\n"
- "mul v16.4s, v16.4s, v15.4s\n"
- "ldr q20, [c_ptr1]\n"
- "mul v17.4s, v17.4s, v15.4s\n"
- "ldr q21, [c_ptr1, #0x10]\n"
- "mul v18.4s, v18.4s, v15.4s\n"
- "ldr q22, [c_ptr1, #0x20]\n"
- "mul v19.4s, v19.4s, v15.4s\n"
- "ldr q23, [c_ptr1, #0x30]\n"
- "mul v20.4s, v20.4s, v15.4s\n"
- "ldr q0, [%[a_ptr0]]\n"
- "mul v21.4s, v21.4s, v15.4s\n"
- "ldr q1, [a_ptr1]\n"
- "mul v22.4s, v22.4s, v15.4s\n"
- "ldr q8, [%[b_ptr0]]\n"
- "mul v23.4s, v23.4s, v15.4s\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "3:\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- "subs %[loops], %[loops], #0x1\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
- ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- "add %[a_ptr0], %[a_ptr0], #0x20\n"
- ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- "add a_ptr1, a_ptr1, #0x20\n"
- ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
- ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q0, [%[a_ptr0], #-0x10]\n"
- ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q1, [a_ptr1, #-0x10]\n"
- ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
- "b.ne 3b\n"
- "2:\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- "prfm PSTL1KEEP, [%[c_ptr0]]\n"
- "prfm PSTL1KEEP, [c_ptr1]\n"
- "cbz %[regs], 4f\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
- ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
- "b 5f\n"
- "4:\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
- "5:\n"
- "cbz %[blocks], 6f\n"
- "7:\n"
- "ldr q8, [%[b_ptr0]]\n"
- "subs %[blocks], %[blocks], #0x1\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr s0, [%[a_ptr0]]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add %[a_ptr0], %[a_ptr0], #0x4\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add %[b_ptr0], %[b_ptr0], #0x40\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr s1, [a_ptr1]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x4\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- "b.ne 7b\n"
- "6:\n"
- "cbz %[odds], 8f\n"
- "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
- "ld1 {v1.b}[0], [a_ptr1], #1\n"
- "subs %[odds], %[odds], #0x1\n"
- "b.eq 9f\n"
- "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
- "ld1 {v1.b}[1], [a_ptr1], #1\n"
- "subs %[odds], %[odds], #0x1\n"
- "b.eq 9f\n"
- "ld1 {v0.b}[2], [%[a_ptr0]]\n"
- "ld1 {v1.b}[2], [a_ptr1]\n"
- "9:\n"
- "ldr q8, [%[b_ptr0]]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- "8:\n"
- "str q16, [%[c_ptr0]]\n"
- "str q17, [%[c_ptr0], #0x10]\n"
- "str q18, [%[c_ptr0], #0x20]\n"
- "str q19, [%[c_ptr0], #0x30]\n"
- "add %[c_ptr0], %[c_ptr0], #0x40\n"
- "str q20, [c_ptr1]\n"
- "str q21, [c_ptr1, #0x10]\n"
- "str q22, [c_ptr1, #0x20]\n"
- "str q23, [c_ptr1, #0x30]\n"
- ".unreq a_ptr1\n"
- ".unreq c_ptr1\n"
- : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
- : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "cc", "memory"
+ "a_ptr1 .req X0\n"
+ "c_ptr1 .req X1\n"
+ "add a_ptr1, %[a_ptr0], %[lda]\n"
+ "add c_ptr1, %[c_ptr0], %[ldc]\n"
+ "cbz %[beta0], 1f\n"
+ "movi v16.4s, #0\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "movi v17.4s, #0\n"
+ "ldr q1, [a_ptr1]\n"
+ "movi v18.4s, #0\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "movi v19.4s, #0\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "movi v20.4s, #0\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "movi v21.4s, #0\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "movi v22.4s, #0\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "movi v23.4s, #0\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "b 3f\n"
+ "1:\n"
+ "ld1r {v15.4s}, [%[betaptr]]\n"
+ "ldr q16, [%[c_ptr0]]\n"
+ "ldr q17, [%[c_ptr0], #0x10]\n"
+ "ldr q18, [%[c_ptr0], #0x20]\n"
+ "ldr q19, [%[c_ptr0], #0x30]\n"
+ "mul v16.4s, v16.4s, v15.4s\n"
+ "ldr q20, [c_ptr1]\n"
+ "mul v17.4s, v17.4s, v15.4s\n"
+ "ldr q21, [c_ptr1, #0x10]\n"
+ "mul v18.4s, v18.4s, v15.4s\n"
+ "ldr q22, [c_ptr1, #0x20]\n"
+ "mul v19.4s, v19.4s, v15.4s\n"
+ "ldr q23, [c_ptr1, #0x30]\n"
+ "mul v20.4s, v20.4s, v15.4s\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "mul v21.4s, v21.4s, v15.4s\n"
+ "ldr q1, [a_ptr1]\n"
+ "mul v22.4s, v22.4s, v15.4s\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "mul v23.4s, v23.4s, v15.4s\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "3:\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ "subs %[loops], %[loops], #0x1\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
+ ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x20\n"
+ ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ "add a_ptr1, a_ptr1, #0x20\n"
+ ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
+ ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q0, [%[a_ptr0], #-0x10]\n"
+ ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q1, [a_ptr1, #-0x10]\n"
+ ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
+ "b.ne 3b\n"
+ "2:\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0]]\n"
+ "prfm PSTL1KEEP, [c_ptr1]\n"
+ "cbz %[regs], 4f\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
+ ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
+ "b 5f\n"
+ "4:\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
+ "5:\n"
+ "cbz %[blocks], 6f\n"
+ "7:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "subs %[blocks], %[blocks], #0x1\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr s0, [%[a_ptr0]]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x4\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr s1, [a_ptr1]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x4\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ "b.ne 7b\n"
+ "6:\n"
+ "cbz %[odds], 8f\n"
+ "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[0], [a_ptr1], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[1], [a_ptr1], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[2], [%[a_ptr0]]\n"
+ "ld1 {v1.b}[2], [a_ptr1]\n"
+ "9:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ "8:\n"
+ "str q16, [%[c_ptr0]]\n"
+ "str q17, [%[c_ptr0], #0x10]\n"
+ "str q18, [%[c_ptr0], #0x20]\n"
+ "str q19, [%[c_ptr0], #0x30]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x40\n"
+ "str q20, [c_ptr1]\n"
+ "str q21, [c_ptr1, #0x10]\n"
+ "str q22, [c_ptr1, #0x20]\n"
+ "str q23, [c_ptr1, #0x30]\n"
+ ".unreq a_ptr1\n"
+ ".unreq c_ptr1\n"
+ : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
+ : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
+ : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "cc", "memory"
);
break;
case 3:
__asm __volatile (
- "a_ptr1 .req X0\n"
- "a_ptr2 .req X1\n"
- "c_ptr1 .req X2\n"
- "c_ptr2 .req X3\n"
- "add a_ptr1, %[a_ptr0], %[lda]\n"
- "add c_ptr1, %[c_ptr0], %[ldc]\n"
- "add a_ptr2, a_ptr1, %[lda]\n"
- "add c_ptr2, c_ptr1, %[ldc]\n"
- "cbz %[beta0], 1f\n"
- "movi v16.4s, #0\n"
- "ldr q0, [%[a_ptr0]]\n"
- "movi v17.4s, #0\n"
- "ldr q1, [a_ptr1]\n"
- "movi v18.4s, #0\n"
- "ldr q2, [a_ptr2]\n"
- "movi v19.4s, #0\n"
- "ldr q8, [%[b_ptr0]]\n"
- "movi v20.4s, #0\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "movi v21.4s, #0\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "movi v22.4s, #0\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "movi v23.4s, #0\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "movi v24.4s, #0\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "movi v25.4s, #0\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "movi v26.4s, #0\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "movi v27.4s, #0\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "b 3f\n"
- "1:\n"
- "ld1r {v15.4s}, [%[betaptr]]\n"
- "ldr q16, [%[c_ptr0]]\n"
- "ldr q17, [%[c_ptr0], #0x10]\n"
- "ldr q18, [%[c_ptr0], #0x20]\n"
- "ldr q19, [%[c_ptr0], #0x30]\n"
- "mul v16.4s, v16.4s, v15.4s\n"
- "ldr q20, [c_ptr1]\n"
- "mul v17.4s, v17.4s, v15.4s\n"
- "ldr q21, [c_ptr1, #0x10]\n"
- "mul v18.4s, v18.4s, v15.4s\n"
- "ldr q22, [c_ptr1, #0x20]\n"
- "mul v19.4s, v19.4s, v15.4s\n"
- "ldr q23, [c_ptr1, #0x30]\n"
- "mul v20.4s, v20.4s, v15.4s\n"
- "ldr q24, [c_ptr2]\n"
- "mul v21.4s, v21.4s, v15.4s\n"
- "ldr q25, [c_ptr2, #0x10]\n"
- "mul v22.4s, v22.4s, v15.4s\n"
- "ldr q26, [c_ptr2, #0x20]\n"
- "mul v23.4s, v23.4s, v15.4s\n"
- "ldr q27, [c_ptr2, #0x30]\n"
- "mul v24.4s, v24.4s, v15.4s\n"
- "ldr q0, [%[a_ptr0]]\n"
- "mul v25.4s, v25.4s, v15.4s\n"
- "ldr q1, [a_ptr1]\n"
- "mul v26.4s, v26.4s, v15.4s\n"
- "ldr q2, [a_ptr2]\n"
- "mul v27.4s, v27.4s, v15.4s\n"
- "ldr q8, [%[b_ptr0]]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "3:\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "subs %[loops], %[loops], #0x1\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- "add %[a_ptr0], %[a_ptr0], #0x20\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x20\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- "add a_ptr2, a_ptr2, #0x20\n"
- ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
- ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr2, #0x40]\n"
- ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q0, [%[a_ptr0], #-0x10]\n"
- ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q1, [a_ptr1, #-0x10]\n"
- ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q2, [a_ptr2, #-0x10]\n"
- ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
- ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
- ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
- ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
- ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
- ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
- ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
- ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
- ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
- ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n"
- "b.ne 3b\n"
- "2:\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- "prfm PSTL1KEEP, [%[c_ptr0]]\n"
- "prfm PSTL1KEEP, [c_ptr1]\n"
- "prfm PSTL1KEEP, [c_ptr2]\n"
- "cbz %[regs], 4f\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
- ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
- ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
- ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
- ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
- ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
- ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
- ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
- ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
- ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
- ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
- ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n"
- ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n"
- ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n"
- ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n"
- ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n"
- ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n"
- ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n"
- ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
- ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n"
- "b 5f\n"
- "4:\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
- ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
- ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
- "5:\n"
- "cbz %[blocks], 6f\n"
- "7:\n"
- "ldr q8, [%[b_ptr0]]\n"
- "subs %[blocks], %[blocks], #0x1\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr s0, [%[a_ptr0]]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add %[a_ptr0], %[a_ptr0], #0x4\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add %[b_ptr0], %[b_ptr0], #0x40\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr s1, [a_ptr1]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x4\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr s2, [a_ptr2]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- "add a_ptr2, a_ptr2, #0x4\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- "b.ne 7b\n"
- "6:\n"
- "cbz %[odds], 8f\n"
- "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
- "ld1 {v1.b}[0], [a_ptr1], #1\n"
- "ld1 {v2.b}[0], [a_ptr2], #1\n"
- "subs %[odds], %[odds], #0x1\n"
- "b.eq 9f\n"
- "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
- "ld1 {v1.b}[1], [a_ptr1], #1\n"
- "ld1 {v2.b}[1], [a_ptr2], #1\n"
- "subs %[odds], %[odds], #0x1\n"
- "b.eq 9f\n"
- "ld1 {v0.b}[2], [%[a_ptr0]]\n"
- "ld1 {v1.b}[2], [a_ptr1]\n"
- "ld1 {v2.b}[2], [a_ptr2]\n"
- "9:\n"
- "ldr q8, [%[b_ptr0]]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- "8:\n"
- "str q16, [%[c_ptr0]]\n"
- "str q17, [%[c_ptr0], #0x10]\n"
- "str q18, [%[c_ptr0], #0x20]\n"
- "str q19, [%[c_ptr0], #0x30]\n"
- "add %[c_ptr0], %[c_ptr0], #0x40\n"
- "str q20, [c_ptr1]\n"
- "str q21, [c_ptr1, #0x10]\n"
- "str q22, [c_ptr1, #0x20]\n"
- "str q23, [c_ptr1, #0x30]\n"
- "str q24, [c_ptr2]\n"
- "str q25, [c_ptr2, #0x10]\n"
- "str q26, [c_ptr2, #0x20]\n"
- "str q27, [c_ptr2, #0x30]\n"
- ".unreq a_ptr1\n"
- ".unreq a_ptr2\n"
- ".unreq c_ptr1\n"
- ".unreq c_ptr2\n"
- : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
- : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "x2", "x3", "cc", "memory"
+ "a_ptr1 .req X0\n"
+ "a_ptr2 .req X1\n"
+ "c_ptr1 .req X2\n"
+ "c_ptr2 .req X3\n"
+ "add a_ptr1, %[a_ptr0], %[lda]\n"
+ "add c_ptr1, %[c_ptr0], %[ldc]\n"
+ "add a_ptr2, a_ptr1, %[lda]\n"
+ "add c_ptr2, c_ptr1, %[ldc]\n"
+ "cbz %[beta0], 1f\n"
+ "movi v16.4s, #0\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "movi v17.4s, #0\n"
+ "ldr q1, [a_ptr1]\n"
+ "movi v18.4s, #0\n"
+ "ldr q2, [a_ptr2]\n"
+ "movi v19.4s, #0\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "movi v20.4s, #0\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "movi v21.4s, #0\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "movi v22.4s, #0\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "movi v23.4s, #0\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "movi v24.4s, #0\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "movi v25.4s, #0\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "movi v26.4s, #0\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "movi v27.4s, #0\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "b 3f\n"
+ "1:\n"
+ "ld1r {v15.4s}, [%[betaptr]]\n"
+ "ldr q16, [%[c_ptr0]]\n"
+ "ldr q17, [%[c_ptr0], #0x10]\n"
+ "ldr q18, [%[c_ptr0], #0x20]\n"
+ "ldr q19, [%[c_ptr0], #0x30]\n"
+ "mul v16.4s, v16.4s, v15.4s\n"
+ "ldr q20, [c_ptr1]\n"
+ "mul v17.4s, v17.4s, v15.4s\n"
+ "ldr q21, [c_ptr1, #0x10]\n"
+ "mul v18.4s, v18.4s, v15.4s\n"
+ "ldr q22, [c_ptr1, #0x20]\n"
+ "mul v19.4s, v19.4s, v15.4s\n"
+ "ldr q23, [c_ptr1, #0x30]\n"
+ "mul v20.4s, v20.4s, v15.4s\n"
+ "ldr q24, [c_ptr2]\n"
+ "mul v21.4s, v21.4s, v15.4s\n"
+ "ldr q25, [c_ptr2, #0x10]\n"
+ "mul v22.4s, v22.4s, v15.4s\n"
+ "ldr q26, [c_ptr2, #0x20]\n"
+ "mul v23.4s, v23.4s, v15.4s\n"
+ "ldr q27, [c_ptr2, #0x30]\n"
+ "mul v24.4s, v24.4s, v15.4s\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "mul v25.4s, v25.4s, v15.4s\n"
+ "ldr q1, [a_ptr1]\n"
+ "mul v26.4s, v26.4s, v15.4s\n"
+ "ldr q2, [a_ptr2]\n"
+ "mul v27.4s, v27.4s, v15.4s\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "3:\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q6, [a_ptr2]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "subs %[loops], %[loops], #0x1\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x20\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x20\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ "add a_ptr2, a_ptr2, #0x20\n"
+ ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
+ ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr2, #0x40]\n"
+ ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q0, [%[a_ptr0], #-0x10]\n"
+ ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
+ "ldr q1, [a_ptr1, #-0x10]\n"
+ ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q2, [a_ptr2, #-0x10]\n"
+ ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
+ ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
+ ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
+ ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
+ ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
+ ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n"
+ "b.ne 3b\n"
+ "2:\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0]]\n"
+ "prfm PSTL1KEEP, [c_ptr1]\n"
+ "prfm PSTL1KEEP, [c_ptr2]\n"
+ "cbz %[regs], 4f\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q6, [a_ptr2]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
+ ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
+ ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
+ ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
+ ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
+ ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
+ ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
+ ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n"
+ ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n"
+ ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n"
+ ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n"
+ ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n"
+ ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n"
+ ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n"
+ ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n"
+ "b 5f\n"
+ "4:\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
+ ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
+ "5:\n"
+ "cbz %[blocks], 6f\n"
+ "7:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "subs %[blocks], %[blocks], #0x1\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr s0, [%[a_ptr0]]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x4\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr s1, [a_ptr1]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x4\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr s2, [a_ptr2]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x4\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ "b.ne 7b\n"
+ "6:\n"
+ "cbz %[odds], 8f\n"
+ "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[0], [a_ptr1], #1\n"
+ "ld1 {v2.b}[0], [a_ptr2], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[1], [a_ptr1], #1\n"
+ "ld1 {v2.b}[1], [a_ptr2], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[2], [%[a_ptr0]]\n"
+ "ld1 {v1.b}[2], [a_ptr1]\n"
+ "ld1 {v2.b}[2], [a_ptr2]\n"
+ "9:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ "8:\n"
+ "str q16, [%[c_ptr0]]\n"
+ "str q17, [%[c_ptr0], #0x10]\n"
+ "str q18, [%[c_ptr0], #0x20]\n"
+ "str q19, [%[c_ptr0], #0x30]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x40\n"
+ "str q20, [c_ptr1]\n"
+ "str q21, [c_ptr1, #0x10]\n"
+ "str q22, [c_ptr1, #0x20]\n"
+ "str q23, [c_ptr1, #0x30]\n"
+ "str q24, [c_ptr2]\n"
+ "str q25, [c_ptr2, #0x10]\n"
+ "str q26, [c_ptr2, #0x20]\n"
+ "str q27, [c_ptr2, #0x30]\n"
+ ".unreq a_ptr1\n"
+ ".unreq a_ptr2\n"
+ ".unreq c_ptr1\n"
+ ".unreq c_ptr2\n"
+ : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
+ : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
+ : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "x2", "x3", "cc", "memory"
);
break;
default:
case 4:
__asm __volatile (
- "a_ptr1 .req X0\n"
- "a_ptr2 .req X1\n"
- "a_ptr3 .req X2\n"
- "c_ptr1 .req X3\n"
- "c_ptr2 .req X4\n"
- "c_ptr3 .req X5\n"
- "add a_ptr1, %[a_ptr0], %[lda]\n"
- "add c_ptr1, %[c_ptr0], %[ldc]\n"
- "add a_ptr2, a_ptr1, %[lda]\n"
- "add c_ptr2, c_ptr1, %[ldc]\n"
- "add a_ptr3, a_ptr2, %[lda]\n"
- "add c_ptr3, c_ptr2, %[ldc]\n"
- "cbz %[beta0], 1f\n"
- "movi v16.4s, #0\n"
- "ldr q0, [%[a_ptr0]]\n"
- "movi v17.4s, #0\n"
- "ldr q1, [a_ptr1]\n"
- "movi v18.4s, #0\n"
- "ldr q2, [a_ptr2]\n"
- "movi v19.4s, #0\n"
- "ldr q3, [a_ptr3]\n"
- "movi v20.4s, #0\n"
- "ldr q8, [%[b_ptr0]]\n"
- "movi v21.4s, #0\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "movi v22.4s, #0\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "movi v23.4s, #0\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "movi v24.4s, #0\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "movi v25.4s, #0\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "movi v26.4s, #0\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "movi v27.4s, #0\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "movi v28.4s, #0\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "movi v29.4s, #0\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- "movi v30.4s, #0\n"
- "add a_ptr3, a_ptr3, #0x10\n"
- "movi v31.4s, #0\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "b 3f\n"
- "1:\n"
- "ld1r {v15.4s}, [%[betaptr]]\n"
- "ldr q16, [%[c_ptr0]]\n"
- "ldr q17, [%[c_ptr0], #0x10]\n"
- "ldr q18, [%[c_ptr0], #0x20]\n"
- "ldr q19, [%[c_ptr0], #0x30]\n"
- "mul v16.4s, v16.4s, v15.4s\n"
- "ldr q20, [c_ptr1]\n"
- "mul v17.4s, v17.4s, v15.4s\n"
- "ldr q21, [c_ptr1, #0x10]\n"
- "mul v18.4s, v18.4s, v15.4s\n"
- "ldr q22, [c_ptr1, #0x20]\n"
- "mul v19.4s, v19.4s, v15.4s\n"
- "ldr q23, [c_ptr1, #0x30]\n"
- "mul v20.4s, v20.4s, v15.4s\n"
- "ldr q24, [c_ptr2]\n"
- "mul v21.4s, v21.4s, v15.4s\n"
- "ldr q25, [c_ptr2, #0x10]\n"
- "mul v22.4s, v22.4s, v15.4s\n"
- "ldr q26, [c_ptr2, #0x20]\n"
- "mul v23.4s, v23.4s, v15.4s\n"
- "ldr q27, [c_ptr2, #0x30]\n"
- "mul v24.4s, v24.4s, v15.4s\n"
- "ldr q28, [c_ptr3]\n"
- "mul v25.4s, v25.4s, v15.4s\n"
- "ldr q29, [c_ptr3, #0x10]\n"
- "mul v26.4s, v26.4s, v15.4s\n"
- "ldr q30, [c_ptr3, #0x20]\n"
- "mul v27.4s, v27.4s, v15.4s\n"
- "ldr q31, [c_ptr3, #0x30]\n"
- "mul v28.4s, v28.4s, v15.4s\n"
- "ldr q0, [%[a_ptr0]]\n"
- "mul v29.4s, v29.4s, v15.4s\n"
- "ldr q1, [a_ptr1]\n"
- "mul v30.4s, v30.4s, v15.4s\n"
- "ldr q2, [a_ptr2]\n"
- "mul v31.4s, v31.4s, v15.4s\n"
- "ldr q3, [a_ptr3]\n"
- "ldr q8, [%[b_ptr0]]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add a_ptr3, a_ptr3, #0x10\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- "cbz %[loops], 2f\n"
- "3:\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q7, [a_ptr3]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- "subs %[loops], %[loops], #0x1\n"
- ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- "add %[a_ptr0], %[a_ptr0], #0x20\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x20\n"
- ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- "add a_ptr2, a_ptr2, #0x20\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- "add a_ptr3, a_ptr3, #0x20\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
- ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr2, #0x40]\n"
- ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
- "prfm PLDL1KEEP, [a_ptr3, #0x40]\n"
- ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
- ".word 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
- ".word 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
- ".word 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
- ".word 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
- ".word 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
- ".word 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
- ".word 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
- ".word 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
- ".word 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
- ".word 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
- ".word 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- "ldr q0, [%[a_ptr0], #-0x10]\n"
- ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
- "ldr q1, [a_ptr1, #-0x10]\n"
- ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
- "ldr q2, [a_ptr2, #-0x10]\n"
- ".word 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
- "ldr q3, [a_ptr3, #-0x10]\n"
- ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
- ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n"
- ".word 0x6f87e11c // udot v28.4s, v8.16b, v7.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
- ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n"
- ".word 0x6f87e13d // udot v29.4s, v9.16b, v7.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
- ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n"
- ".word 0x6f87e15e // udot v30.4s, v10.16b, v7.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
- ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n"
- ".word 0x6f87e17f // udot v31.4s, v11.16b, v7.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
- ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n"
- ".word 0x6fa7e19c // udot v28.4s, v12.16b, v7.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
- ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n"
- ".word 0x6fa7e1bd // udot v29.4s, v13.16b, v7.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
- ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n"
- ".word 0x6fa7e1de // udot v30.4s, v14.16b, v7.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
- ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n"
- ".word 0x6fa7e1ff // udot v31.4s, v15.16b, v7.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n"
- ".word 0x6f87e91c // udot v28.4s, v8.16b, v7.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n"
- ".word 0x6f87e93d // udot v29.4s, v9.16b, v7.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n"
- ".word 0x6f87e95e // udot v30.4s, v10.16b, v7.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n"
- ".word 0x6f87e97f // udot v31.4s, v11.16b, v7.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n"
- ".word 0x6fa7e99c // udot v28.4s, v12.16b, v7.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n"
- ".word 0x6fa7e9bd // udot v29.4s, v13.16b, v7.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n"
- ".word 0x6fa7e9de // udot v30.4s, v14.16b, v7.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
- ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n"
- ".word 0x6fa7e9ff // udot v31.4s, v15.16b, v7.4b[3]\n"
- "b.ne 3b\n"
- "2:\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- "prfm PSTL1KEEP, [%[c_ptr0]]\n"
- "prfm PSTL1KEEP, [c_ptr1]\n"
- "prfm PSTL1KEEP, [c_ptr2]\n"
- "prfm PSTL1KEEP, [c_ptr3]\n"
- "cbz %[regs], 4f\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr q4, [%[a_ptr0]]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- "ldr q5, [a_ptr1]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- "ldr q6, [a_ptr2]\n"
- ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
- "ldr q7, [a_ptr3]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- "add %[a_ptr0], %[a_ptr0], #0x10\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x10\n"
- ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "add a_ptr2, a_ptr2, #0x10\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- "add a_ptr3, a_ptr3, #0x10\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
- ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
- ".word 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
- ".word 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
- ".word 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
- ".word 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x100\n"
- ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
- ".word 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n"
- "ldr q8, [%[b_ptr0], #-0x80]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
- ".word 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n"
- "ldr q9, [%[b_ptr0], #-0x70]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
- ".word 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n"
- "ldr q10, [%[b_ptr0], #-0x60]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
- ".word 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n"
- "ldr q11, [%[b_ptr0], #-0x50]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
- ".word 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n"
- "ldr q12, [%[b_ptr0], #-0x40]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
- ".word 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n"
- "ldr q13, [%[b_ptr0], #-0x30]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
- ".word 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n"
- "ldr q14, [%[b_ptr0], #-0x20]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
- ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
- ".word 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n"
- "ldr q15, [%[b_ptr0], #-0x10]\n"
- ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
- ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
- ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n"
- ".word 0x6f87e11c // udot v28.4s, v8.16b, v7.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
- ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
- ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n"
- ".word 0x6f87e13d // udot v29.4s, v9.16b, v7.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
- ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
- ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n"
- ".word 0x6f87e15e // udot v30.4s, v10.16b, v7.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
- ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
- ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n"
- ".word 0x6f87e17f // udot v31.4s, v11.16b, v7.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
- ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
- ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n"
- ".word 0x6fa7e19c // udot v28.4s, v12.16b, v7.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
- ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
- ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n"
- ".word 0x6fa7e1bd // udot v29.4s, v13.16b, v7.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
- ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
- ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n"
- ".word 0x6fa7e1de // udot v30.4s, v14.16b, v7.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
- ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
- ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n"
- ".word 0x6fa7e1ff // udot v31.4s, v15.16b, v7.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
- ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n"
- ".word 0x6f87e91c // udot v28.4s, v8.16b, v7.4b[2]\n"
- ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
- ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
- ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n"
- ".word 0x6f87e93d // udot v29.4s, v9.16b, v7.4b[2]\n"
- ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
- ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
- ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n"
- ".word 0x6f87e95e // udot v30.4s, v10.16b, v7.4b[2]\n"
- ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
- ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
- ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n"
- ".word 0x6f87e97f // udot v31.4s, v11.16b, v7.4b[2]\n"
- ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
- ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
- ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n"
- ".word 0x6fa7e99c // udot v28.4s, v12.16b, v7.4b[3]\n"
- ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
- ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
- ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n"
- ".word 0x6fa7e9bd // udot v29.4s, v13.16b, v7.4b[3]\n"
- ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
- ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
- ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n"
- ".word 0x6fa7e9de // udot v30.4s, v14.16b, v7.4b[3]\n"
- ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
- ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
- ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n"
- ".word 0x6fa7e9ff // udot v31.4s, v15.16b, v7.4b[3]\n"
- "b 5f\n"
- "4:\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
- "ldr q8, [%[b_ptr0]]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
- ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
- ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
- ".word 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n"
- "ldr q12, [%[b_ptr0], #0x40]\n"
- ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
- ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
- ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
- ".word 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n"
- "ldr q13, [%[b_ptr0], #0x50]\n"
- ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
- ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
- ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
- ".word 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n"
- "ldr q14, [%[b_ptr0], #0x60]\n"
- ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
- ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
- ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
- ".word 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n"
- "ldr q15, [%[b_ptr0], #0x70]\n"
- ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
- "add %[b_ptr0], %[b_ptr0], #0x80\n"
- ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
- ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
- ".word 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n"
- ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
- ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
- ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
- ".word 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n"
- ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
- ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
- ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
- ".word 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n"
- ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
- ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
- ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
- ".word 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n"
- ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
- ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
- ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
- ".word 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n"
- ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
- ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
- ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
- ".word 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n"
- ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
- ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
- ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
- ".word 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n"
- ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
- ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
- ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
- ".word 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n"
- "5:\n"
- "cbz %[blocks], 6f\n"
- "7:\n"
- "ldr q8, [%[b_ptr0]]\n"
- "subs %[blocks], %[blocks], #0x1\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr s0, [%[a_ptr0]]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "add %[a_ptr0], %[a_ptr0], #0x4\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- "add %[b_ptr0], %[b_ptr0], #0x40\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- "ldr s1, [a_ptr1]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- "add a_ptr1, a_ptr1, #0x4\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- "ldr s2, [a_ptr2]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- "add a_ptr2, a_ptr2, #0x4\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- "ldr s3, [a_ptr3]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- "add a_ptr3, a_ptr3, #0x4\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
- ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
- "b.ne 7b\n"
- "6:\n"
- "cbz %[odds], 8f\n"
- "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
- "ld1 {v1.b}[0], [a_ptr1], #1\n"
- "ld1 {v2.b}[0], [a_ptr2], #1\n"
- "ld1 {v3.b}[0], [a_ptr3], #1\n"
- "subs %[odds], %[odds], #0x1\n"
- "b.eq 9f\n"
- "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
- "ld1 {v1.b}[1], [a_ptr1], #1\n"
- "ld1 {v2.b}[1], [a_ptr2], #1\n"
- "ld1 {v3.b}[1], [a_ptr3], #1\n"
- "subs %[odds], %[odds], #0x1\n"
- "b.eq 9f\n"
- "ld1 {v0.b}[2], [%[a_ptr0]]\n"
- "ld1 {v1.b}[2], [a_ptr1]\n"
- "ld1 {v2.b}[2], [a_ptr2]\n"
- "ld1 {v3.b}[2], [a_ptr3]\n"
- "9:\n"
- "ldr q8, [%[b_ptr0]]\n"
- "ldr q9, [%[b_ptr0], #0x10]\n"
- "ldr q10, [%[b_ptr0], #0x20]\n"
- "ldr q11, [%[b_ptr0], #0x30]\n"
- ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
- ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
- ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
- ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
- ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
- ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
- ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
- ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
- ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
- ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
- ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
- ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
- ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
- ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
- ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
- "8:\n"
- "str q16, [%[c_ptr0]]\n"
- "str q17, [%[c_ptr0], #0x10]\n"
- "str q18, [%[c_ptr0], #0x20]\n"
- "str q19, [%[c_ptr0], #0x30]\n"
- "add %[c_ptr0], %[c_ptr0], #0x40\n"
- "str q20, [c_ptr1]\n"
- "str q21, [c_ptr1, #0x10]\n"
- "str q22, [c_ptr1, #0x20]\n"
- "str q23, [c_ptr1, #0x30]\n"
- "str q24, [c_ptr2]\n"
- "str q25, [c_ptr2, #0x10]\n"
- "str q26, [c_ptr2, #0x20]\n"
- "str q27, [c_ptr2, #0x30]\n"
- "str q28, [c_ptr3]\n"
- "str q29, [c_ptr3, #0x10]\n"
- "str q30, [c_ptr3, #0x20]\n"
- "str q31, [c_ptr3, #0x30]\n"
- ".unreq a_ptr1\n"
- ".unreq a_ptr2\n"
- ".unreq a_ptr3\n"
- ".unreq c_ptr1\n"
- ".unreq c_ptr2\n"
- ".unreq c_ptr3\n"
- : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
- : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
- : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "x2", "x3", "x4", "x5", "cc", "memory"
+ "a_ptr1 .req X0\n"
+ "a_ptr2 .req X1\n"
+ "a_ptr3 .req X2\n"
+ "c_ptr1 .req X3\n"
+ "c_ptr2 .req X4\n"
+ "c_ptr3 .req X5\n"
+ "add a_ptr1, %[a_ptr0], %[lda]\n"
+ "add c_ptr1, %[c_ptr0], %[ldc]\n"
+ "add a_ptr2, a_ptr1, %[lda]\n"
+ "add c_ptr2, c_ptr1, %[ldc]\n"
+ "add a_ptr3, a_ptr2, %[lda]\n"
+ "add c_ptr3, c_ptr2, %[ldc]\n"
+ "cbz %[beta0], 1f\n"
+ "movi v16.4s, #0\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "movi v17.4s, #0\n"
+ "ldr q1, [a_ptr1]\n"
+ "movi v18.4s, #0\n"
+ "ldr q2, [a_ptr2]\n"
+ "movi v19.4s, #0\n"
+ "ldr q3, [a_ptr3]\n"
+ "movi v20.4s, #0\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "movi v21.4s, #0\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "movi v22.4s, #0\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "movi v23.4s, #0\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "movi v24.4s, #0\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "movi v25.4s, #0\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "movi v26.4s, #0\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "movi v27.4s, #0\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "movi v28.4s, #0\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "movi v29.4s, #0\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ "movi v30.4s, #0\n"
+ "add a_ptr3, a_ptr3, #0x10\n"
+ "movi v31.4s, #0\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "b 3f\n"
+ "1:\n"
+ "ld1r {v15.4s}, [%[betaptr]]\n"
+ "ldr q16, [%[c_ptr0]]\n"
+ "ldr q17, [%[c_ptr0], #0x10]\n"
+ "ldr q18, [%[c_ptr0], #0x20]\n"
+ "ldr q19, [%[c_ptr0], #0x30]\n"
+ "mul v16.4s, v16.4s, v15.4s\n"
+ "ldr q20, [c_ptr1]\n"
+ "mul v17.4s, v17.4s, v15.4s\n"
+ "ldr q21, [c_ptr1, #0x10]\n"
+ "mul v18.4s, v18.4s, v15.4s\n"
+ "ldr q22, [c_ptr1, #0x20]\n"
+ "mul v19.4s, v19.4s, v15.4s\n"
+ "ldr q23, [c_ptr1, #0x30]\n"
+ "mul v20.4s, v20.4s, v15.4s\n"
+ "ldr q24, [c_ptr2]\n"
+ "mul v21.4s, v21.4s, v15.4s\n"
+ "ldr q25, [c_ptr2, #0x10]\n"
+ "mul v22.4s, v22.4s, v15.4s\n"
+ "ldr q26, [c_ptr2, #0x20]\n"
+ "mul v23.4s, v23.4s, v15.4s\n"
+ "ldr q27, [c_ptr2, #0x30]\n"
+ "mul v24.4s, v24.4s, v15.4s\n"
+ "ldr q28, [c_ptr3]\n"
+ "mul v25.4s, v25.4s, v15.4s\n"
+ "ldr q29, [c_ptr3, #0x10]\n"
+ "mul v26.4s, v26.4s, v15.4s\n"
+ "ldr q30, [c_ptr3, #0x20]\n"
+ "mul v27.4s, v27.4s, v15.4s\n"
+ "ldr q31, [c_ptr3, #0x30]\n"
+ "mul v28.4s, v28.4s, v15.4s\n"
+ "ldr q0, [%[a_ptr0]]\n"
+ "mul v29.4s, v29.4s, v15.4s\n"
+ "ldr q1, [a_ptr1]\n"
+ "mul v30.4s, v30.4s, v15.4s\n"
+ "ldr q2, [a_ptr2]\n"
+ "mul v31.4s, v31.4s, v15.4s\n"
+ "ldr q3, [a_ptr3]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add a_ptr3, a_ptr3, #0x10\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ "cbz %[loops], 2f\n"
+ "3:\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
+ "ldr q6, [a_ptr2]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q7, [a_ptr3]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ "subs %[loops], %[loops], #0x1\n"
+ ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "prfm PLDL1KEEP, [%[a_ptr0], #0x40]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x20\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x20\n"
+ ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x20\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ "add a_ptr3, a_ptr3, #0x20\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ "prfm PLDL1KEEP, [a_ptr1, #0x40]\n"
+ ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr2, #0x40]\n"
+ ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
+ "prfm PLDL1KEEP, [a_ptr3, #0x40]\n"
+ ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
+ ".word 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
+ ".word 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
+ ".word 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
+ ".word 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
+ ".word 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
+ ".word 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
+ ".word 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
+ ".word 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
+ ".word 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
+ ".word 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
+ ".word 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ "ldr q0, [%[a_ptr0], #-0x10]\n"
+ ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
+ "ldr q1, [a_ptr1, #-0x10]\n"
+ ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
+ "ldr q2, [a_ptr2, #-0x10]\n"
+ ".word 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
+ "ldr q3, [a_ptr3, #-0x10]\n"
+ ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
+ ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n"
+ ".word 0x6f87e11c // udot v28.4s, v8.16b, v7.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
+ ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n"
+ ".word 0x6f87e13d // udot v29.4s, v9.16b, v7.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
+ ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n"
+ ".word 0x6f87e15e // udot v30.4s, v10.16b, v7.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
+ ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n"
+ ".word 0x6f87e17f // udot v31.4s, v11.16b, v7.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
+ ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n"
+ ".word 0x6fa7e19c // udot v28.4s, v12.16b, v7.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n"
+ ".word 0x6fa7e1bd // udot v29.4s, v13.16b, v7.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n"
+ ".word 0x6fa7e1de // udot v30.4s, v14.16b, v7.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n"
+ ".word 0x6fa7e1ff // udot v31.4s, v15.16b, v7.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n"
+ ".word 0x6f87e91c // udot v28.4s, v8.16b, v7.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n"
+ ".word 0x6f87e93d // udot v29.4s, v9.16b, v7.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n"
+ ".word 0x6f87e95e // udot v30.4s, v10.16b, v7.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n"
+ ".word 0x6f87e97f // udot v31.4s, v11.16b, v7.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n"
+ ".word 0x6fa7e99c // udot v28.4s, v12.16b, v7.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n"
+ ".word 0x6fa7e9bd // udot v29.4s, v13.16b, v7.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n"
+ ".word 0x6fa7e9de // udot v30.4s, v14.16b, v7.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n"
+ ".word 0x6fa7e9ff // udot v31.4s, v15.16b, v7.4b[3]\n"
+ "b.ne 3b\n"
+ "2:\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ "prfm PSTL1KEEP, [%[c_ptr0]]\n"
+ "prfm PSTL1KEEP, [c_ptr1]\n"
+ "prfm PSTL1KEEP, [c_ptr2]\n"
+ "prfm PSTL1KEEP, [c_ptr3]\n"
+ "cbz %[regs], 4f\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr q4, [%[a_ptr0]]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ "ldr q5, [a_ptr1]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ "ldr q6, [a_ptr2]\n"
+ ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
+ "ldr q7, [a_ptr3]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x10\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x10\n"
+ ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x10\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ "add a_ptr3, a_ptr3, #0x10\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
+ ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
+ ".word 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
+ ".word 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
+ ".word 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
+ ".word 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x100\n"
+ ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
+ ".word 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n"
+ "ldr q8, [%[b_ptr0], #-0x80]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
+ ".word 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n"
+ "ldr q9, [%[b_ptr0], #-0x70]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
+ ".word 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n"
+ "ldr q10, [%[b_ptr0], #-0x60]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
+ ".word 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n"
+ "ldr q11, [%[b_ptr0], #-0x50]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
+ ".word 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n"
+ "ldr q12, [%[b_ptr0], #-0x40]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
+ ".word 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n"
+ "ldr q13, [%[b_ptr0], #-0x30]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
+ ".word 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n"
+ "ldr q14, [%[b_ptr0], #-0x20]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
+ ".word 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n"
+ "ldr q15, [%[b_ptr0], #-0x10]\n"
+ ".word 0x6f84e110 // udot v16.4s, v8.16b, v4.4b[0]\n"
+ ".word 0x6f85e114 // udot v20.4s, v8.16b, v5.4b[0]\n"
+ ".word 0x6f86e118 // udot v24.4s, v8.16b, v6.4b[0]\n"
+ ".word 0x6f87e11c // udot v28.4s, v8.16b, v7.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f84e131 // udot v17.4s, v9.16b, v4.4b[0]\n"
+ ".word 0x6f85e135 // udot v21.4s, v9.16b, v5.4b[0]\n"
+ ".word 0x6f86e139 // udot v25.4s, v9.16b, v6.4b[0]\n"
+ ".word 0x6f87e13d // udot v29.4s, v9.16b, v7.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f84e152 // udot v18.4s, v10.16b, v4.4b[0]\n"
+ ".word 0x6f85e156 // udot v22.4s, v10.16b, v5.4b[0]\n"
+ ".word 0x6f86e15a // udot v26.4s, v10.16b, v6.4b[0]\n"
+ ".word 0x6f87e15e // udot v30.4s, v10.16b, v7.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f84e173 // udot v19.4s, v11.16b, v4.4b[0]\n"
+ ".word 0x6f85e177 // udot v23.4s, v11.16b, v5.4b[0]\n"
+ ".word 0x6f86e17b // udot v27.4s, v11.16b, v6.4b[0]\n"
+ ".word 0x6f87e17f // udot v31.4s, v11.16b, v7.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa4e190 // udot v16.4s, v12.16b, v4.4b[1]\n"
+ ".word 0x6fa5e194 // udot v20.4s, v12.16b, v5.4b[1]\n"
+ ".word 0x6fa6e198 // udot v24.4s, v12.16b, v6.4b[1]\n"
+ ".word 0x6fa7e19c // udot v28.4s, v12.16b, v7.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa4e1b1 // udot v17.4s, v13.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1b5 // udot v21.4s, v13.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1b9 // udot v25.4s, v13.16b, v6.4b[1]\n"
+ ".word 0x6fa7e1bd // udot v29.4s, v13.16b, v7.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa4e1d2 // udot v18.4s, v14.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1d6 // udot v22.4s, v14.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1da // udot v26.4s, v14.16b, v6.4b[1]\n"
+ ".word 0x6fa7e1de // udot v30.4s, v14.16b, v7.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa4e1f3 // udot v19.4s, v15.16b, v4.4b[1]\n"
+ ".word 0x6fa5e1f7 // udot v23.4s, v15.16b, v5.4b[1]\n"
+ ".word 0x6fa6e1fb // udot v27.4s, v15.16b, v6.4b[1]\n"
+ ".word 0x6fa7e1ff // udot v31.4s, v15.16b, v7.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f84e910 // udot v16.4s, v8.16b, v4.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x6f85e914 // udot v20.4s, v8.16b, v5.4b[2]\n"
+ ".word 0x6f86e918 // udot v24.4s, v8.16b, v6.4b[2]\n"
+ ".word 0x6f87e91c // udot v28.4s, v8.16b, v7.4b[2]\n"
+ ".word 0x6f84e931 // udot v17.4s, v9.16b, v4.4b[2]\n"
+ ".word 0x6f85e935 // udot v21.4s, v9.16b, v5.4b[2]\n"
+ ".word 0x6f86e939 // udot v25.4s, v9.16b, v6.4b[2]\n"
+ ".word 0x6f87e93d // udot v29.4s, v9.16b, v7.4b[2]\n"
+ ".word 0x6f84e952 // udot v18.4s, v10.16b, v4.4b[2]\n"
+ ".word 0x6f85e956 // udot v22.4s, v10.16b, v5.4b[2]\n"
+ ".word 0x6f86e95a // udot v26.4s, v10.16b, v6.4b[2]\n"
+ ".word 0x6f87e95e // udot v30.4s, v10.16b, v7.4b[2]\n"
+ ".word 0x6f84e973 // udot v19.4s, v11.16b, v4.4b[2]\n"
+ ".word 0x6f85e977 // udot v23.4s, v11.16b, v5.4b[2]\n"
+ ".word 0x6f86e97b // udot v27.4s, v11.16b, v6.4b[2]\n"
+ ".word 0x6f87e97f // udot v31.4s, v11.16b, v7.4b[2]\n"
+ ".word 0x6fa4e990 // udot v16.4s, v12.16b, v4.4b[3]\n"
+ ".word 0x6fa5e994 // udot v20.4s, v12.16b, v5.4b[3]\n"
+ ".word 0x6fa6e998 // udot v24.4s, v12.16b, v6.4b[3]\n"
+ ".word 0x6fa7e99c // udot v28.4s, v12.16b, v7.4b[3]\n"
+ ".word 0x6fa4e9b1 // udot v17.4s, v13.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9b5 // udot v21.4s, v13.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9b9 // udot v25.4s, v13.16b, v6.4b[3]\n"
+ ".word 0x6fa7e9bd // udot v29.4s, v13.16b, v7.4b[3]\n"
+ ".word 0x6fa4e9d2 // udot v18.4s, v14.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9d6 // udot v22.4s, v14.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9da // udot v26.4s, v14.16b, v6.4b[3]\n"
+ ".word 0x6fa7e9de // udot v30.4s, v14.16b, v7.4b[3]\n"
+ ".word 0x6fa4e9f3 // udot v19.4s, v15.16b, v4.4b[3]\n"
+ ".word 0x6fa5e9f7 // udot v23.4s, v15.16b, v5.4b[3]\n"
+ ".word 0x6fa6e9fb // udot v27.4s, v15.16b, v6.4b[3]\n"
+ ".word 0x6fa7e9ff // udot v31.4s, v15.16b, v7.4b[3]\n"
+ "b 5f\n"
+ "4:\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6fa0e190 // udot v16.4s, v12.16b, v0.4b[1]\n"
+ ".word 0x6fa1e194 // udot v20.4s, v12.16b, v1.4b[1]\n"
+ ".word 0x6fa2e198 // udot v24.4s, v12.16b, v2.4b[1]\n"
+ ".word 0x6fa3e19c // udot v28.4s, v12.16b, v3.4b[1]\n"
+ "ldr q12, [%[b_ptr0], #0x40]\n"
+ ".word 0x6fa0e1b1 // udot v17.4s, v13.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1b5 // udot v21.4s, v13.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1b9 // udot v25.4s, v13.16b, v2.4b[1]\n"
+ ".word 0x6fa3e1bd // udot v29.4s, v13.16b, v3.4b[1]\n"
+ "ldr q13, [%[b_ptr0], #0x50]\n"
+ ".word 0x6fa0e1d2 // udot v18.4s, v14.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1d6 // udot v22.4s, v14.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1da // udot v26.4s, v14.16b, v2.4b[1]\n"
+ ".word 0x6fa3e1de // udot v30.4s, v14.16b, v3.4b[1]\n"
+ "ldr q14, [%[b_ptr0], #0x60]\n"
+ ".word 0x6fa0e1f3 // udot v19.4s, v15.16b, v0.4b[1]\n"
+ ".word 0x6fa1e1f7 // udot v23.4s, v15.16b, v1.4b[1]\n"
+ ".word 0x6fa2e1fb // udot v27.4s, v15.16b, v2.4b[1]\n"
+ ".word 0x6fa3e1ff // udot v31.4s, v15.16b, v3.4b[1]\n"
+ "ldr q15, [%[b_ptr0], #0x70]\n"
+ ".word 0x6f80e910 // udot v16.4s, v8.16b, v0.4b[2]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x80\n"
+ ".word 0x6f81e914 // udot v20.4s, v8.16b, v1.4b[2]\n"
+ ".word 0x6f82e918 // udot v24.4s, v8.16b, v2.4b[2]\n"
+ ".word 0x6f83e91c // udot v28.4s, v8.16b, v3.4b[2]\n"
+ ".word 0x6f80e931 // udot v17.4s, v9.16b, v0.4b[2]\n"
+ ".word 0x6f81e935 // udot v21.4s, v9.16b, v1.4b[2]\n"
+ ".word 0x6f82e939 // udot v25.4s, v9.16b, v2.4b[2]\n"
+ ".word 0x6f83e93d // udot v29.4s, v9.16b, v3.4b[2]\n"
+ ".word 0x6f80e952 // udot v18.4s, v10.16b, v0.4b[2]\n"
+ ".word 0x6f81e956 // udot v22.4s, v10.16b, v1.4b[2]\n"
+ ".word 0x6f82e95a // udot v26.4s, v10.16b, v2.4b[2]\n"
+ ".word 0x6f83e95e // udot v30.4s, v10.16b, v3.4b[2]\n"
+ ".word 0x6f80e973 // udot v19.4s, v11.16b, v0.4b[2]\n"
+ ".word 0x6f81e977 // udot v23.4s, v11.16b, v1.4b[2]\n"
+ ".word 0x6f82e97b // udot v27.4s, v11.16b, v2.4b[2]\n"
+ ".word 0x6f83e97f // udot v31.4s, v11.16b, v3.4b[2]\n"
+ ".word 0x6fa0e990 // udot v16.4s, v12.16b, v0.4b[3]\n"
+ ".word 0x6fa1e994 // udot v20.4s, v12.16b, v1.4b[3]\n"
+ ".word 0x6fa2e998 // udot v24.4s, v12.16b, v2.4b[3]\n"
+ ".word 0x6fa3e99c // udot v28.4s, v12.16b, v3.4b[3]\n"
+ ".word 0x6fa0e9b1 // udot v17.4s, v13.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9b5 // udot v21.4s, v13.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9b9 // udot v25.4s, v13.16b, v2.4b[3]\n"
+ ".word 0x6fa3e9bd // udot v29.4s, v13.16b, v3.4b[3]\n"
+ ".word 0x6fa0e9d2 // udot v18.4s, v14.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9d6 // udot v22.4s, v14.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9da // udot v26.4s, v14.16b, v2.4b[3]\n"
+ ".word 0x6fa3e9de // udot v30.4s, v14.16b, v3.4b[3]\n"
+ ".word 0x6fa0e9f3 // udot v19.4s, v15.16b, v0.4b[3]\n"
+ ".word 0x6fa1e9f7 // udot v23.4s, v15.16b, v1.4b[3]\n"
+ ".word 0x6fa2e9fb // udot v27.4s, v15.16b, v2.4b[3]\n"
+ ".word 0x6fa3e9ff // udot v31.4s, v15.16b, v3.4b[3]\n"
+ "5:\n"
+ "cbz %[blocks], 6f\n"
+ "7:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "subs %[blocks], %[blocks], #0x1\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr s0, [%[a_ptr0]]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "add %[a_ptr0], %[a_ptr0], #0x4\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ "add %[b_ptr0], %[b_ptr0], #0x40\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ "ldr s1, [a_ptr1]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ "add a_ptr1, a_ptr1, #0x4\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ "ldr s2, [a_ptr2]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ "add a_ptr2, a_ptr2, #0x4\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ "ldr s3, [a_ptr3]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ "add a_ptr3, a_ptr3, #0x4\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
+ ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
+ "b.ne 7b\n"
+ "6:\n"
+ "cbz %[odds], 8f\n"
+ "ld1 {v0.b}[0], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[0], [a_ptr1], #1\n"
+ "ld1 {v2.b}[0], [a_ptr2], #1\n"
+ "ld1 {v3.b}[0], [a_ptr3], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[1], [%[a_ptr0]], #1\n"
+ "ld1 {v1.b}[1], [a_ptr1], #1\n"
+ "ld1 {v2.b}[1], [a_ptr2], #1\n"
+ "ld1 {v3.b}[1], [a_ptr3], #1\n"
+ "subs %[odds], %[odds], #0x1\n"
+ "b.eq 9f\n"
+ "ld1 {v0.b}[2], [%[a_ptr0]]\n"
+ "ld1 {v1.b}[2], [a_ptr1]\n"
+ "ld1 {v2.b}[2], [a_ptr2]\n"
+ "ld1 {v3.b}[2], [a_ptr3]\n"
+ "9:\n"
+ "ldr q8, [%[b_ptr0]]\n"
+ "ldr q9, [%[b_ptr0], #0x10]\n"
+ "ldr q10, [%[b_ptr0], #0x20]\n"
+ "ldr q11, [%[b_ptr0], #0x30]\n"
+ ".word 0x6f80e110 // udot v16.4s, v8.16b, v0.4b[0]\n"
+ ".word 0x6f81e114 // udot v20.4s, v8.16b, v1.4b[0]\n"
+ ".word 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
+ ".word 0x6f83e11c // udot v28.4s, v8.16b, v3.4b[0]\n"
+ ".word 0x6f80e131 // udot v17.4s, v9.16b, v0.4b[0]\n"
+ ".word 0x6f81e135 // udot v21.4s, v9.16b, v1.4b[0]\n"
+ ".word 0x6f82e139 // udot v25.4s, v9.16b, v2.4b[0]\n"
+ ".word 0x6f83e13d // udot v29.4s, v9.16b, v3.4b[0]\n"
+ ".word 0x6f80e152 // udot v18.4s, v10.16b, v0.4b[0]\n"
+ ".word 0x6f81e156 // udot v22.4s, v10.16b, v1.4b[0]\n"
+ ".word 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
+ ".word 0x6f83e15e // udot v30.4s, v10.16b, v3.4b[0]\n"
+ ".word 0x6f80e173 // udot v19.4s, v11.16b, v0.4b[0]\n"
+ ".word 0x6f81e177 // udot v23.4s, v11.16b, v1.4b[0]\n"
+ ".word 0x6f82e17b // udot v27.4s, v11.16b, v2.4b[0]\n"
+ ".word 0x6f83e17f // udot v31.4s, v11.16b, v3.4b[0]\n"
+ "8:\n"
+ "str q16, [%[c_ptr0]]\n"
+ "str q17, [%[c_ptr0], #0x10]\n"
+ "str q18, [%[c_ptr0], #0x20]\n"
+ "str q19, [%[c_ptr0], #0x30]\n"
+ "add %[c_ptr0], %[c_ptr0], #0x40\n"
+ "str q20, [c_ptr1]\n"
+ "str q21, [c_ptr1, #0x10]\n"
+ "str q22, [c_ptr1, #0x20]\n"
+ "str q23, [c_ptr1, #0x30]\n"
+ "str q24, [c_ptr2]\n"
+ "str q25, [c_ptr2, #0x10]\n"
+ "str q26, [c_ptr2, #0x20]\n"
+ "str q27, [c_ptr2, #0x30]\n"
+ "str q28, [c_ptr3]\n"
+ "str q29, [c_ptr3, #0x10]\n"
+ "str q30, [c_ptr3, #0x20]\n"
+ "str q31, [c_ptr3, #0x30]\n"
+ ".unreq a_ptr1\n"
+ ".unreq a_ptr2\n"
+ ".unreq a_ptr3\n"
+ ".unreq c_ptr1\n"
+ ".unreq c_ptr2\n"
+ ".unreq c_ptr3\n"
+ : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks), [odds] "+r" (odds)
+ : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb)
+ : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x0", "x1", "x2", "x3", "x4", "x5", "cc", "memory"
);
break;
}
@@ -1839,4 +1839,4 @@ void a64_hybrid_u8u32_dot_16x4(const uint8_t *A, int lda, const uint8_t *B, uint
} // namespace arm_gemm
-#endif // __aarch64__ \ No newline at end of file
+#endif // __aarch64__