aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_native_16x4
diff options
context:
space:
mode:
authorAnthony Barbier <anthony.barbier@arm.com>2018-07-03 16:22:02 +0100
committerAnthony Barbier <anthony.barbier@arm.com>2018-11-02 16:54:10 +0000
commit5f707736413aeac77818c42838296966f8dc6761 (patch)
treeb829ed3243ea5f3085f288836132416c78bc2e72 /src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_native_16x4
parent7485d5a62685cb745ab50e970adb722cb71557ac (diff)
downloadComputeLibrary-5f707736413aeac77818c42838296966f8dc6761.tar.gz
COMPMID-1369: Revert accidental formatting of RSH's repo
Pulled latest fixes from David's repo: commit f43ebe932c84083332b0b1a0348241b69dda63a7 Author: David Mansell <David.Mansell@arm.com> Date: Tue Jul 3 18:09:01 2018 +0100 Whitespace tidying, fixed comment in gemv_batched imported from ACL. Change-Id: Ie37a623f44e90d88072236cb853ac55ac82d5f51 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/138530 Tested-by: Jenkins <bsgcomp@arm.com> Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com> Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com> Reviewed-by: David Mansell <david.mansell@arm.com> Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_native_16x4')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_native_16x4/generic.cpp1285
1 files changed, 642 insertions, 643 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_native_16x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_native_16x4/generic.cpp
index 89a16f75a0..8325b3f352 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_native_16x4/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_native_16x4/generic.cpp
@@ -29,12 +29,11 @@
#include <arm_neon.h>
-namespace arm_gemm
-{
-void a64_sgemm_native_16x4(const float *A, int lda, const float *B, int ldb, float *C, int ldc, float beta, int M, int N, int K)
-{
- const int oddk = ((K % 8) >= 4) ? 1 : 0;
- const int beta0 = (beta == 0.0f) ? 1 : 0;
+namespace arm_gemm {
+
+void a64_sgemm_native_16x4(const float *A, int lda, const float *B, int ldb, float *C, int ldc, float beta, int M, int N, int K) {
+ const int oddk = ((K % 8) >= 4) ? 1 : 0;
+ const int beta0 = (beta == 0.0f) ? 1 : 0;
const int oddones = (K % 4);
float dummy_buffer[16];
@@ -67,12 +66,12 @@ void a64_sgemm_native_16x4(const float *A, int lda, const float *B, int ldb, flo
const float *b_ptr = B + x0;
- int loops = ((K + 4) / 8) - 1;
- int odds = oddones;
+ int loops = ((K+4)/8) - 1;
+ int odds = oddones;
size_t ldbb = ldb * sizeof(float);
- __asm __volatile(
+ __asm __volatile (
"a0 .req v0\n"
"a1 .req v1\n"
"a2 .req v2\n"
@@ -107,140 +106,140 @@ void a64_sgemm_native_16x4(const float *A, int lda, const float *B, int ldb, flo
"b2aq .req q14\n"
"b3aq .req q15\n"
- "movi v16.4s, #0x0\n"
- "ldr a0q, [%[a_ptr0]]\n"
- "movi v17.4s, #0x0\n"
- "ldr b0q, [%[b_ptr]]\n"
- "movi v18.4s, #0x0\n"
- "ldr b1q, [%[b_ptr], #16]\n"
- "movi v19.4s, #0x0\n"
- "ldr b2q, [%[b_ptr], #32]\n"
- "movi v20.4s, #0x0\n"
- "ldr b3q, [%[b_ptr], #48]\n"
- "movi v21.4s, #0x0\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "ldr a1q, [%[a_ptr1]]\n"
- "movi v22.4s, #0x0\n"
- "ldr a2q, [%[a_ptr2]]\n"
- "movi v23.4s, #0x0\n"
- "ldr a3q, [%[a_ptr3]]\n"
- "movi v24.4s, #0x0\n"
- "ldr b0aq, [%[b_ptr]]\n"
- "movi v25.4s, #0x0\n"
- "ldr b1aq, [%[b_ptr], #16]\n"
- "movi v26.4s, #0x0\n"
- "ldr b2aq, [%[b_ptr], #32]\n"
- "cbz %w[beta0], 5f\n"
- "movi v27.4s, #0x0\n"
- "movi v28.4s, #0x0\n"
- "movi v29.4s, #0x0\n"
- "movi v30.4s, #0x0\n"
- "movi v31.4s, #0x0\n"
+ "movi v16.4s, #0x0\n"
+ "ldr a0q, [%[a_ptr0]]\n"
+ "movi v17.4s, #0x0\n"
+ "ldr b0q, [%[b_ptr]]\n"
+ "movi v18.4s, #0x0\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+ "movi v19.4s, #0x0\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+ "movi v20.4s, #0x0\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
+ "movi v21.4s, #0x0\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "ldr a1q, [%[a_ptr1]]\n"
+ "movi v22.4s, #0x0\n"
+ "ldr a2q, [%[a_ptr2]]\n"
+ "movi v23.4s, #0x0\n"
+ "ldr a3q, [%[a_ptr3]]\n"
+ "movi v24.4s, #0x0\n"
+ "ldr b0aq, [%[b_ptr]]\n"
+ "movi v25.4s, #0x0\n"
+ "ldr b1aq, [%[b_ptr], #16]\n"
+ "movi v26.4s, #0x0\n"
+ "ldr b2aq, [%[b_ptr], #32]\n"
+ "cbz %w[beta0], 5f\n"
+ "movi v27.4s, #0x0\n"
+ "movi v28.4s, #0x0\n"
+ "movi v29.4s, #0x0\n"
+ "movi v30.4s, #0x0\n"
+ "movi v31.4s, #0x0\n"
// Skip if no complete loops.
- "cbz %w[loops], 4f\n"
- "b 1f\n"
+ "cbz %w[loops], 4f\n"
+ "b 1f\n"
// If beta is non-zero, need to load and multiply by beta
"5:\n"
- "ld1r {v4.4s}, [%[betaptr]]\n"
- "ldr q16, [%[c_ptr0]]\n"
- "ldr q17, [%[c_ptr0], #16]\n"
- "ldr q18, [%[c_ptr0], #32]\n"
- "ldr q19, [%[c_ptr0], #48]\n"
-
- "ldr q20, [%[c_ptr1]]\n"
- "fmul v16.4s, v16.4s, v4.4s\n"
- "ldr q21, [%[c_ptr1], #16]\n"
- "fmul v17.4s, v17.4s, v4.4s\n"
- "ldr q22, [%[c_ptr1], #32]\n"
- "fmul v18.4s, v18.4s, v4.4s\n"
- "ldr q23, [%[c_ptr1], #48]\n"
- "fmul v19.4s, v19.4s, v4.4s\n"
-
- "ldr q24, [%[c_ptr2]]\n"
- "fmul v20.4s, v20.4s, v4.4s\n"
- "ldr q25, [%[c_ptr2], #16]\n"
- "fmul v21.4s, v21.4s, v4.4s\n"
- "ldr q26, [%[c_ptr2], #32]\n"
- "fmul v22.4s, v22.4s, v4.4s\n"
- "ldr q27, [%[c_ptr2], #48]\n"
- "fmul v23.4s, v23.4s, v4.4s\n"
-
- "ldr q28, [%[c_ptr3]]\n"
- "fmul v24.4s, v24.4s, v4.4s\n"
- "ldr q29, [%[c_ptr3], #16]\n"
- "fmul v25.4s, v25.4s, v4.4s\n"
- "ldr q30, [%[c_ptr3], #32]\n"
- "fmul v26.4s, v26.4s, v4.4s\n"
- "ldr q31, [%[c_ptr3], #48]\n"
- "fmul v27.4s, v27.4s, v4.4s\n"
-
- "fmul v28.4s, v28.4s, v4.4s\n"
- "fmul v29.4s, v29.4s, v4.4s\n"
- "fmul v30.4s, v30.4s, v4.4s\n"
- "fmul v31.4s, v31.4s, v4.4s\n"
-
- "cbz %w[loops], 4f\n"
+ "ld1r {v4.4s}, [%[betaptr]]\n"
+ "ldr q16, [%[c_ptr0]]\n"
+ "ldr q17, [%[c_ptr0], #16]\n"
+ "ldr q18, [%[c_ptr0], #32]\n"
+ "ldr q19, [%[c_ptr0], #48]\n"
+
+ "ldr q20, [%[c_ptr1]]\n"
+ "fmul v16.4s, v16.4s, v4.4s\n"
+ "ldr q21, [%[c_ptr1], #16]\n"
+ "fmul v17.4s, v17.4s, v4.4s\n"
+ "ldr q22, [%[c_ptr1], #32]\n"
+ "fmul v18.4s, v18.4s, v4.4s\n"
+ "ldr q23, [%[c_ptr1], #48]\n"
+ "fmul v19.4s, v19.4s, v4.4s\n"
+
+ "ldr q24, [%[c_ptr2]]\n"
+ "fmul v20.4s, v20.4s, v4.4s\n"
+ "ldr q25, [%[c_ptr2], #16]\n"
+ "fmul v21.4s, v21.4s, v4.4s\n"
+ "ldr q26, [%[c_ptr2], #32]\n"
+ "fmul v22.4s, v22.4s, v4.4s\n"
+ "ldr q27, [%[c_ptr2], #48]\n"
+ "fmul v23.4s, v23.4s, v4.4s\n"
+
+ "ldr q28, [%[c_ptr3]]\n"
+ "fmul v24.4s, v24.4s, v4.4s\n"
+ "ldr q29, [%[c_ptr3], #16]\n"
+ "fmul v25.4s, v25.4s, v4.4s\n"
+ "ldr q30, [%[c_ptr3], #32]\n"
+ "fmul v26.4s, v26.4s, v4.4s\n"
+ "ldr q31, [%[c_ptr3], #48]\n"
+ "fmul v27.4s, v27.4s, v4.4s\n"
+
+ "fmul v28.4s, v28.4s, v4.4s\n"
+ "fmul v29.4s, v29.4s, v4.4s\n"
+ "fmul v30.4s, v30.4s, v4.4s\n"
+ "fmul v31.4s, v31.4s, v4.4s\n"
+
+ "cbz %w[loops], 4f\n"
"1:\n"
// Unroll 0
- "fmla v16.4s, bb0.4s, a0.s[0]\n"
- "fmla v20.4s, bb0.4s, a1.s[0]\n"
- "ldr b3aq, [%[b_ptr], #48]\n"
- "fmla v24.4s, bb0.4s, a2.s[0]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v28.4s, bb0.4s, a3.s[0]\n"
- "ldr b0q, [%[b_ptr]]\n"
-
- "fmla v17.4s, bb1.4s, a0.s[0]\n"
- "fmla v21.4s, bb1.4s, a1.s[0]\n"
- "ldr a0aq, [%[a_ptr0], #16]\n"
- "fmla v25.4s, bb1.4s, a2.s[0]\n"
- "fmla v29.4s, bb1.4s, a3.s[0]\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, bb2.4s, a0.s[0]\n"
- "fmla v22.4s, bb2.4s, a1.s[0]\n"
- "ldr a1aq, [%[a_ptr1], #16]\n"
- "fmla v26.4s, bb2.4s, a2.s[0]\n"
- "fmla v30.4s, bb2.4s, a3.s[0]\n"
- "ldr b2q, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, bb3.4s, a0.s[0]\n"
- "fmla v23.4s, bb3.4s, a1.s[0]\n"
- "ldr a2aq, [%[a_ptr2], #16]\n"
- "fmla v27.4s, bb3.4s, a2.s[0]\n"
- "fmla v31.4s, bb3.4s, a3.s[0]\n"
- "ldr b3q, [%[b_ptr], #48]\n"
+ "fmla v16.4s, bb0.4s, a0.s[0]\n"
+ "fmla v20.4s, bb0.4s, a1.s[0]\n"
+ "ldr b3aq, [%[b_ptr], #48]\n"
+ "fmla v24.4s, bb0.4s, a2.s[0]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v28.4s, bb0.4s, a3.s[0]\n"
+ "ldr b0q, [%[b_ptr]]\n"
+
+ "fmla v17.4s, bb1.4s, a0.s[0]\n"
+ "fmla v21.4s, bb1.4s, a1.s[0]\n"
+ "ldr a0aq, [%[a_ptr0], #16]\n"
+ "fmla v25.4s, bb1.4s, a2.s[0]\n"
+ "fmla v29.4s, bb1.4s, a3.s[0]\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, bb2.4s, a0.s[0]\n"
+ "fmla v22.4s, bb2.4s, a1.s[0]\n"
+ "ldr a1aq, [%[a_ptr1], #16]\n"
+ "fmla v26.4s, bb2.4s, a2.s[0]\n"
+ "fmla v30.4s, bb2.4s, a3.s[0]\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, bb3.4s, a0.s[0]\n"
+ "fmla v23.4s, bb3.4s, a1.s[0]\n"
+ "ldr a2aq, [%[a_ptr2], #16]\n"
+ "fmla v27.4s, bb3.4s, a2.s[0]\n"
+ "fmla v31.4s, bb3.4s, a3.s[0]\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
// Unroll 1
- "fmla v16.4s, b0a.4s, a0.s[1]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v20.4s, b0a.4s, a1.s[1]\n"
- "ldr a3aq, [%[a_ptr3], #16]\n"
- "fmla v24.4s, b0a.4s, a2.s[1]\n"
- "fmla v28.4s, b0a.4s, a3.s[1]\n"
- "ldr b0aq, [%[b_ptr]]\n"
-
- "fmla v17.4s, b1a.4s, a0.s[1]\n"
- "fmla v21.4s, b1a.4s, a1.s[1]\n"
- "subs %w[loops], %w[loops], #1\n"
- "fmla v25.4s, b1a.4s, a2.s[1]\n"
- "fmla v29.4s, b1a.4s, a3.s[1]\n"
- "ldr b1aq, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, b2a.4s, a0.s[1]\n"
- "fmla v22.4s, b2a.4s, a1.s[1]\n"
- "fmla v26.4s, b2a.4s, a2.s[1]\n"
- "fmla v30.4s, b2a.4s, a3.s[1]\n"
- "ldr b2aq, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, b3a.4s, a0.s[1]\n"
- "fmla v23.4s, b3a.4s, a1.s[1]\n"
- "fmla v27.4s, b3a.4s, a2.s[1]\n"
- "fmla v31.4s, b3a.4s, a3.s[1]\n"
- "ldr b3aq, [%[b_ptr], #48]\n"
+ "fmla v16.4s, b0a.4s, a0.s[1]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v20.4s, b0a.4s, a1.s[1]\n"
+ "ldr a3aq, [%[a_ptr3], #16]\n"
+ "fmla v24.4s, b0a.4s, a2.s[1]\n"
+ "fmla v28.4s, b0a.4s, a3.s[1]\n"
+ "ldr b0aq, [%[b_ptr]]\n"
+
+ "fmla v17.4s, b1a.4s, a0.s[1]\n"
+ "fmla v21.4s, b1a.4s, a1.s[1]\n"
+ "subs %w[loops], %w[loops], #1\n"
+ "fmla v25.4s, b1a.4s, a2.s[1]\n"
+ "fmla v29.4s, b1a.4s, a3.s[1]\n"
+ "ldr b1aq, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, b2a.4s, a0.s[1]\n"
+ "fmla v22.4s, b2a.4s, a1.s[1]\n"
+ "fmla v26.4s, b2a.4s, a2.s[1]\n"
+ "fmla v30.4s, b2a.4s, a3.s[1]\n"
+ "ldr b2aq, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, b3a.4s, a0.s[1]\n"
+ "fmla v23.4s, b3a.4s, a1.s[1]\n"
+ "fmla v27.4s, b3a.4s, a2.s[1]\n"
+ "fmla v31.4s, b3a.4s, a3.s[1]\n"
+ "ldr b3aq, [%[b_ptr], #48]\n"
// Unroll 2
"fmla v16.4s, bb0.4s, a0.s[2]\n"
@@ -273,173 +272,173 @@ void a64_sgemm_native_16x4(const float *A, int lda, const float *B, int ldb, flo
"ldr b3q, [%[b_ptr], #48]\n"
// Unroll 3
- "fmla v16.4s, b0a.4s, a0.s[3]\n"
- "fmla v20.4s, b0a.4s, a1.s[3]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, b0a.4s, a2.s[3]\n"
- "fmla v28.4s, b0a.4s, a3.s[3]\n"
- "ldr b0aq, [%[b_ptr]]\n"
-
- "fmla v17.4s, b1a.4s, a0.s[3]\n"
- "fmla v21.4s, b1a.4s, a1.s[3]\n"
- "fmla v25.4s, b1a.4s, a2.s[3]\n"
- "fmla v29.4s, b1a.4s, a3.s[3]\n"
- "ldr b1aq, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, b2a.4s, a0.s[3]\n"
- "fmla v22.4s, b2a.4s, a1.s[3]\n"
- "fmla v26.4s, b2a.4s, a2.s[3]\n"
- "fmla v30.4s, b2a.4s, a3.s[3]\n"
- "ldr b2aq, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, b3a.4s, a0.s[3]\n"
- "fmla v23.4s, b3a.4s, a1.s[3]\n"
- "ldr a0q, [%[a_ptr0]]\n"
- "fmla v27.4s, b3a.4s, a2.s[3]\n"
- "fmla v31.4s, b3a.4s, a3.s[3]\n"
- "ldr b3aq, [%[b_ptr], #48]\n"
+ "fmla v16.4s, b0a.4s, a0.s[3]\n"
+ "fmla v20.4s, b0a.4s, a1.s[3]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, b0a.4s, a2.s[3]\n"
+ "fmla v28.4s, b0a.4s, a3.s[3]\n"
+ "ldr b0aq, [%[b_ptr]]\n"
+
+ "fmla v17.4s, b1a.4s, a0.s[3]\n"
+ "fmla v21.4s, b1a.4s, a1.s[3]\n"
+ "fmla v25.4s, b1a.4s, a2.s[3]\n"
+ "fmla v29.4s, b1a.4s, a3.s[3]\n"
+ "ldr b1aq, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, b2a.4s, a0.s[3]\n"
+ "fmla v22.4s, b2a.4s, a1.s[3]\n"
+ "fmla v26.4s, b2a.4s, a2.s[3]\n"
+ "fmla v30.4s, b2a.4s, a3.s[3]\n"
+ "ldr b2aq, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, b3a.4s, a0.s[3]\n"
+ "fmla v23.4s, b3a.4s, a1.s[3]\n"
+ "ldr a0q, [%[a_ptr0]]\n"
+ "fmla v27.4s, b3a.4s, a2.s[3]\n"
+ "fmla v31.4s, b3a.4s, a3.s[3]\n"
+ "ldr b3aq, [%[b_ptr], #48]\n"
// Unroll 4
- "fmla v16.4s, bb0.4s, a0a.s[0]\n"
- "fmla v20.4s, bb0.4s, a1a.s[0]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, bb0.4s, a2a.s[0]\n"
- "fmla v28.4s, bb0.4s, a3a.s[0]\n"
- "ldr b0q, [%[b_ptr]]\n"
-
- "fmla v17.4s, bb1.4s, a0a.s[0]\n"
- "fmla v21.4s, bb1.4s, a1a.s[0]\n"
- "ldr a1q, [%[a_ptr1]]\n"
- "fmla v25.4s, bb1.4s, a2a.s[0]\n"
- "fmla v29.4s, bb1.4s, a3a.s[0]\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, bb2.4s, a0a.s[0]\n"
- "fmla v22.4s, bb2.4s, a1a.s[0]\n"
- "ldr a2q, [%[a_ptr2]]\n"
- "fmla v26.4s, bb2.4s, a2a.s[0]\n"
- "fmla v30.4s, bb2.4s, a3a.s[0]\n"
- "ldr b2q, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, bb3.4s, a0a.s[0]\n"
- "fmla v23.4s, bb3.4s, a1a.s[0]\n"
- "ldr a3q, [%[a_ptr3]]\n"
- "fmla v27.4s, bb3.4s, a2a.s[0]\n"
- "fmla v31.4s, bb3.4s, a3a.s[0]\n"
- "ldr b3q, [%[b_ptr], #48]\n"
+ "fmla v16.4s, bb0.4s, a0a.s[0]\n"
+ "fmla v20.4s, bb0.4s, a1a.s[0]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, bb0.4s, a2a.s[0]\n"
+ "fmla v28.4s, bb0.4s, a3a.s[0]\n"
+ "ldr b0q, [%[b_ptr]]\n"
+
+ "fmla v17.4s, bb1.4s, a0a.s[0]\n"
+ "fmla v21.4s, bb1.4s, a1a.s[0]\n"
+ "ldr a1q, [%[a_ptr1]]\n"
+ "fmla v25.4s, bb1.4s, a2a.s[0]\n"
+ "fmla v29.4s, bb1.4s, a3a.s[0]\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, bb2.4s, a0a.s[0]\n"
+ "fmla v22.4s, bb2.4s, a1a.s[0]\n"
+ "ldr a2q, [%[a_ptr2]]\n"
+ "fmla v26.4s, bb2.4s, a2a.s[0]\n"
+ "fmla v30.4s, bb2.4s, a3a.s[0]\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, bb3.4s, a0a.s[0]\n"
+ "fmla v23.4s, bb3.4s, a1a.s[0]\n"
+ "ldr a3q, [%[a_ptr3]]\n"
+ "fmla v27.4s, bb3.4s, a2a.s[0]\n"
+ "fmla v31.4s, bb3.4s, a3a.s[0]\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
// Unroll 5
- "fmla v16.4s, b0a.4s, a0a.s[1]\n"
- "fmla v20.4s, b0a.4s, a1a.s[1]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, b0a.4s, a2a.s[1]\n"
- "fmla v28.4s, b0a.4s, a3a.s[1]\n"
- "ldr b0aq, [%[b_ptr]]\n"
-
- "fmla v17.4s, b1a.4s, a0a.s[1]\n"
- "fmla v21.4s, b1a.4s, a1a.s[1]\n"
- "fmla v25.4s, b1a.4s, a2a.s[1]\n"
- "fmla v29.4s, b1a.4s, a3a.s[1]\n"
- "ldr b1aq, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, b2a.4s, a0a.s[1]\n"
- "fmla v22.4s, b2a.4s, a1a.s[1]\n"
- "fmla v26.4s, b2a.4s, a2a.s[1]\n"
- "fmla v30.4s, b2a.4s, a3a.s[1]\n"
- "ldr b2aq, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, b3a.4s, a0a.s[1]\n"
- "fmla v23.4s, b3a.4s, a1a.s[1]\n"
- "fmla v27.4s, b3a.4s, a2a.s[1]\n"
- "fmla v31.4s, b3a.4s, a3a.s[1]\n"
- "ldr b3aq, [%[b_ptr], #48]\n"
+ "fmla v16.4s, b0a.4s, a0a.s[1]\n"
+ "fmla v20.4s, b0a.4s, a1a.s[1]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, b0a.4s, a2a.s[1]\n"
+ "fmla v28.4s, b0a.4s, a3a.s[1]\n"
+ "ldr b0aq, [%[b_ptr]]\n"
+
+ "fmla v17.4s, b1a.4s, a0a.s[1]\n"
+ "fmla v21.4s, b1a.4s, a1a.s[1]\n"
+ "fmla v25.4s, b1a.4s, a2a.s[1]\n"
+ "fmla v29.4s, b1a.4s, a3a.s[1]\n"
+ "ldr b1aq, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, b2a.4s, a0a.s[1]\n"
+ "fmla v22.4s, b2a.4s, a1a.s[1]\n"
+ "fmla v26.4s, b2a.4s, a2a.s[1]\n"
+ "fmla v30.4s, b2a.4s, a3a.s[1]\n"
+ "ldr b2aq, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, b3a.4s, a0a.s[1]\n"
+ "fmla v23.4s, b3a.4s, a1a.s[1]\n"
+ "fmla v27.4s, b3a.4s, a2a.s[1]\n"
+ "fmla v31.4s, b3a.4s, a3a.s[1]\n"
+ "ldr b3aq, [%[b_ptr], #48]\n"
// Unroll 6
- "fmla v16.4s, bb0.4s, a0a.s[2]\n"
- "fmla v20.4s, bb0.4s, a1a.s[2]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, bb0.4s, a2a.s[2]\n"
- "fmla v28.4s, bb0.4s, a3a.s[2]\n"
- "ldr b0q, [%[b_ptr]]\n"
-
- "fmla v17.4s, bb1.4s, a0a.s[2]\n"
- "fmla v21.4s, bb1.4s, a1a.s[2]\n"
- "fmla v25.4s, bb1.4s, a2a.s[2]\n"
- "fmla v29.4s, bb1.4s, a3a.s[2]\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, bb2.4s, a0a.s[2]\n"
- "fmla v22.4s, bb2.4s, a1a.s[2]\n"
- "fmla v26.4s, bb2.4s, a2a.s[2]\n"
- "fmla v30.4s, bb2.4s, a3a.s[2]\n"
- "ldr b2q, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, bb3.4s, a0a.s[2]\n"
- "fmla v23.4s, bb3.4s, a1a.s[2]\n"
- "fmla v27.4s, bb3.4s, a2a.s[2]\n"
- "fmla v31.4s, bb3.4s, a3a.s[2]\n"
- "ldr b3q, [%[b_ptr], #48]\n"
+ "fmla v16.4s, bb0.4s, a0a.s[2]\n"
+ "fmla v20.4s, bb0.4s, a1a.s[2]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, bb0.4s, a2a.s[2]\n"
+ "fmla v28.4s, bb0.4s, a3a.s[2]\n"
+ "ldr b0q, [%[b_ptr]]\n"
+
+ "fmla v17.4s, bb1.4s, a0a.s[2]\n"
+ "fmla v21.4s, bb1.4s, a1a.s[2]\n"
+ "fmla v25.4s, bb1.4s, a2a.s[2]\n"
+ "fmla v29.4s, bb1.4s, a3a.s[2]\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, bb2.4s, a0a.s[2]\n"
+ "fmla v22.4s, bb2.4s, a1a.s[2]\n"
+ "fmla v26.4s, bb2.4s, a2a.s[2]\n"
+ "fmla v30.4s, bb2.4s, a3a.s[2]\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, bb3.4s, a0a.s[2]\n"
+ "fmla v23.4s, bb3.4s, a1a.s[2]\n"
+ "fmla v27.4s, bb3.4s, a2a.s[2]\n"
+ "fmla v31.4s, bb3.4s, a3a.s[2]\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
// Unroll 7
- "fmla v16.4s, b0a.4s, a0a.s[3]\n"
- "fmla v20.4s, b0a.4s, a1a.s[3]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, b0a.4s, a2a.s[3]\n"
- "fmla v28.4s, b0a.4s, a3a.s[3]\n"
- "ldr b0aq, [%[b_ptr]]\n"
-
- "fmla v17.4s, b1a.4s, a0a.s[3]\n"
- "fmla v21.4s, b1a.4s, a1a.s[3]\n"
- "fmla v25.4s, b1a.4s, a2a.s[3]\n"
- "fmla v29.4s, b1a.4s, a3a.s[3]\n"
- "ldr b1aq, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, b2a.4s, a0a.s[3]\n"
- "fmla v22.4s, b2a.4s, a1a.s[3]\n"
- "fmla v26.4s, b2a.4s, a2a.s[3]\n"
- "fmla v30.4s, b2a.4s, a3a.s[3]\n"
- "ldr b2aq, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, b3a.4s, a0a.s[3]\n"
- "fmla v23.4s, b3a.4s, a1a.s[3]\n"
- "fmla v27.4s, b3a.4s, a2a.s[3]\n"
- "fmla v31.4s, b3a.4s, a3a.s[3]\n"
- "bne 1b\n"
+ "fmla v16.4s, b0a.4s, a0a.s[3]\n"
+ "fmla v20.4s, b0a.4s, a1a.s[3]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, b0a.4s, a2a.s[3]\n"
+ "fmla v28.4s, b0a.4s, a3a.s[3]\n"
+ "ldr b0aq, [%[b_ptr]]\n"
+
+ "fmla v17.4s, b1a.4s, a0a.s[3]\n"
+ "fmla v21.4s, b1a.4s, a1a.s[3]\n"
+ "fmla v25.4s, b1a.4s, a2a.s[3]\n"
+ "fmla v29.4s, b1a.4s, a3a.s[3]\n"
+ "ldr b1aq, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, b2a.4s, a0a.s[3]\n"
+ "fmla v22.4s, b2a.4s, a1a.s[3]\n"
+ "fmla v26.4s, b2a.4s, a2a.s[3]\n"
+ "fmla v30.4s, b2a.4s, a3a.s[3]\n"
+ "ldr b2aq, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, b3a.4s, a0a.s[3]\n"
+ "fmla v23.4s, b3a.4s, a1a.s[3]\n"
+ "fmla v27.4s, b3a.4s, a2a.s[3]\n"
+ "fmla v31.4s, b3a.4s, a3a.s[3]\n"
+ "bne 1b\n"
// Skip to here
"4:\n"
// Detached final iteration
// Unroll 0
- "fmla v16.4s, bb0.4s, a0.s[0]\n"
- "fmla v20.4s, bb0.4s, a1.s[0]\n"
- "ldr b3aq, [%[b_ptr], #48]\n"
- "fmla v24.4s, bb0.4s, a2.s[0]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v28.4s, bb0.4s, a3.s[0]\n"
- "ldr b0q, [%[b_ptr]]\n"
-
- "fmla v17.4s, bb1.4s, a0.s[0]\n"
- "cbnz %w[oddk], 2f\n" // Deal with odd K before we load a0a
- "fmla v21.4s, bb1.4s, a1.s[0]\n"
- "ldr a0aq, [%[a_ptr0], #16]\n"
- "fmla v25.4s, bb1.4s, a2.s[0]\n"
- "fmla v29.4s, bb1.4s, a3.s[0]\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, bb2.4s, a0.s[0]\n"
- "fmla v22.4s, bb2.4s, a1.s[0]\n"
- "ldr a1aq, [%[a_ptr1], #16]\n"
- "fmla v26.4s, bb2.4s, a2.s[0]\n"
- "fmla v30.4s, bb2.4s, a3.s[0]\n"
- "ldr b2q, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, bb3.4s, a0.s[0]\n"
- "fmla v23.4s, bb3.4s, a1.s[0]\n"
- "ldr a2aq, [%[a_ptr2], #16]\n"
- "fmla v27.4s, bb3.4s, a2.s[0]\n"
- "fmla v31.4s, bb3.4s, a3.s[0]\n"
- "ldr b3q, [%[b_ptr], #48]\n"
+ "fmla v16.4s, bb0.4s, a0.s[0]\n"
+ "fmla v20.4s, bb0.4s, a1.s[0]\n"
+ "ldr b3aq, [%[b_ptr], #48]\n"
+ "fmla v24.4s, bb0.4s, a2.s[0]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v28.4s, bb0.4s, a3.s[0]\n"
+ "ldr b0q, [%[b_ptr]]\n"
+
+ "fmla v17.4s, bb1.4s, a0.s[0]\n"
+ "cbnz %w[oddk], 2f\n" // Deal with odd K before we load a0a
+ "fmla v21.4s, bb1.4s, a1.s[0]\n"
+ "ldr a0aq, [%[a_ptr0], #16]\n"
+ "fmla v25.4s, bb1.4s, a2.s[0]\n"
+ "fmla v29.4s, bb1.4s, a3.s[0]\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, bb2.4s, a0.s[0]\n"
+ "fmla v22.4s, bb2.4s, a1.s[0]\n"
+ "ldr a1aq, [%[a_ptr1], #16]\n"
+ "fmla v26.4s, bb2.4s, a2.s[0]\n"
+ "fmla v30.4s, bb2.4s, a3.s[0]\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, bb3.4s, a0.s[0]\n"
+ "fmla v23.4s, bb3.4s, a1.s[0]\n"
+ "ldr a2aq, [%[a_ptr2], #16]\n"
+ "fmla v27.4s, bb3.4s, a2.s[0]\n"
+ "fmla v31.4s, bb3.4s, a3.s[0]\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
// Unroll 1
"fmla v16.4s, b0a.4s, a0.s[1]\n"
@@ -473,394 +472,394 @@ void a64_sgemm_native_16x4(const float *A, int lda, const float *B, int ldb, flo
"ldr b3aq, [%[b_ptr], #48]\n"
// Unroll 2
- "fmla v16.4s, bb0.4s, a0.s[2]\n"
- "fmla v20.4s, bb0.4s, a1.s[2]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, bb0.4s, a2.s[2]\n"
- "fmla v28.4s, bb0.4s, a3.s[2]\n"
- "ldr b0q, [%[b_ptr]]\n"
-
- "fmla v17.4s, bb1.4s, a0.s[2]\n"
- "fmla v21.4s, bb1.4s, a1.s[2]\n"
- "fmla v25.4s, bb1.4s, a2.s[2]\n"
- "fmla v29.4s, bb1.4s, a3.s[2]\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, bb2.4s, a0.s[2]\n"
- "fmla v22.4s, bb2.4s, a1.s[2]\n"
- "fmla v26.4s, bb2.4s, a2.s[2]\n"
- "fmla v30.4s, bb2.4s, a3.s[2]\n"
- "ldr b2q, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, bb3.4s, a0.s[2]\n"
- "fmla v23.4s, bb3.4s, a1.s[2]\n"
- "fmla v27.4s, bb3.4s, a2.s[2]\n"
- "fmla v31.4s, bb3.4s, a3.s[2]\n"
- "ldr b3q, [%[b_ptr], #48]\n"
+ "fmla v16.4s, bb0.4s, a0.s[2]\n"
+ "fmla v20.4s, bb0.4s, a1.s[2]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, bb0.4s, a2.s[2]\n"
+ "fmla v28.4s, bb0.4s, a3.s[2]\n"
+ "ldr b0q, [%[b_ptr]]\n"
+
+ "fmla v17.4s, bb1.4s, a0.s[2]\n"
+ "fmla v21.4s, bb1.4s, a1.s[2]\n"
+ "fmla v25.4s, bb1.4s, a2.s[2]\n"
+ "fmla v29.4s, bb1.4s, a3.s[2]\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, bb2.4s, a0.s[2]\n"
+ "fmla v22.4s, bb2.4s, a1.s[2]\n"
+ "fmla v26.4s, bb2.4s, a2.s[2]\n"
+ "fmla v30.4s, bb2.4s, a3.s[2]\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, bb3.4s, a0.s[2]\n"
+ "fmla v23.4s, bb3.4s, a1.s[2]\n"
+ "fmla v27.4s, bb3.4s, a2.s[2]\n"
+ "fmla v31.4s, bb3.4s, a3.s[2]\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
// Unroll 3
- "fmla v16.4s, b0a.4s, a0.s[3]\n"
- "fmla v20.4s, b0a.4s, a1.s[3]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, b0a.4s, a2.s[3]\n"
- "fmla v28.4s, b0a.4s, a3.s[3]\n"
- "ldr b0aq, [%[b_ptr]]\n"
-
- "fmla v17.4s, b1a.4s, a0.s[3]\n"
- "fmla v21.4s, b1a.4s, a1.s[3]\n"
- "fmla v25.4s, b1a.4s, a2.s[3]\n"
- "fmla v29.4s, b1a.4s, a3.s[3]\n"
- "ldr b1aq, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, b2a.4s, a0.s[3]\n"
- "fmla v22.4s, b2a.4s, a1.s[3]\n"
- "fmla v26.4s, b2a.4s, a2.s[3]\n"
- "fmla v30.4s, b2a.4s, a3.s[3]\n"
- "ldr b2aq, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, b3a.4s, a0.s[3]\n"
- "fmla v23.4s, b3a.4s, a1.s[3]\n"
- "fmla v27.4s, b3a.4s, a2.s[3]\n"
- "fmla v31.4s, b3a.4s, a3.s[3]\n"
- "ldr b3aq, [%[b_ptr], #48]\n"
+ "fmla v16.4s, b0a.4s, a0.s[3]\n"
+ "fmla v20.4s, b0a.4s, a1.s[3]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, b0a.4s, a2.s[3]\n"
+ "fmla v28.4s, b0a.4s, a3.s[3]\n"
+ "ldr b0aq, [%[b_ptr]]\n"
+
+ "fmla v17.4s, b1a.4s, a0.s[3]\n"
+ "fmla v21.4s, b1a.4s, a1.s[3]\n"
+ "fmla v25.4s, b1a.4s, a2.s[3]\n"
+ "fmla v29.4s, b1a.4s, a3.s[3]\n"
+ "ldr b1aq, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, b2a.4s, a0.s[3]\n"
+ "fmla v22.4s, b2a.4s, a1.s[3]\n"
+ "fmla v26.4s, b2a.4s, a2.s[3]\n"
+ "fmla v30.4s, b2a.4s, a3.s[3]\n"
+ "ldr b2aq, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, b3a.4s, a0.s[3]\n"
+ "fmla v23.4s, b3a.4s, a1.s[3]\n"
+ "fmla v27.4s, b3a.4s, a2.s[3]\n"
+ "fmla v31.4s, b3a.4s, a3.s[3]\n"
+ "ldr b3aq, [%[b_ptr], #48]\n"
// Unroll 4
- "fmla v16.4s, bb0.4s, a0a.s[0]\n"
- "fmla v20.4s, bb0.4s, a1a.s[0]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, bb0.4s, a2a.s[0]\n"
- "fmla v28.4s, bb0.4s, a3a.s[0]\n"
- "ldr b0q, [%[b_ptr]]\n"
-
- "fmla v17.4s, bb1.4s, a0a.s[0]\n"
- "fmla v21.4s, bb1.4s, a1a.s[0]\n"
- "fmla v25.4s, bb1.4s, a2a.s[0]\n"
- "fmla v29.4s, bb1.4s, a3a.s[0]\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, bb2.4s, a0a.s[0]\n"
- "fmla v22.4s, bb2.4s, a1a.s[0]\n"
- "fmla v26.4s, bb2.4s, a2a.s[0]\n"
- "fmla v30.4s, bb2.4s, a3a.s[0]\n"
- "ldr b2q, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, bb3.4s, a0a.s[0]\n"
- "fmla v23.4s, bb3.4s, a1a.s[0]\n"
- "fmla v27.4s, bb3.4s, a2a.s[0]\n"
- "fmla v31.4s, bb3.4s, a3a.s[0]\n"
- "ldr b3q, [%[b_ptr], #48]\n"
+ "fmla v16.4s, bb0.4s, a0a.s[0]\n"
+ "fmla v20.4s, bb0.4s, a1a.s[0]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, bb0.4s, a2a.s[0]\n"
+ "fmla v28.4s, bb0.4s, a3a.s[0]\n"
+ "ldr b0q, [%[b_ptr]]\n"
+
+ "fmla v17.4s, bb1.4s, a0a.s[0]\n"
+ "fmla v21.4s, bb1.4s, a1a.s[0]\n"
+ "fmla v25.4s, bb1.4s, a2a.s[0]\n"
+ "fmla v29.4s, bb1.4s, a3a.s[0]\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, bb2.4s, a0a.s[0]\n"
+ "fmla v22.4s, bb2.4s, a1a.s[0]\n"
+ "fmla v26.4s, bb2.4s, a2a.s[0]\n"
+ "fmla v30.4s, bb2.4s, a3a.s[0]\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, bb3.4s, a0a.s[0]\n"
+ "fmla v23.4s, bb3.4s, a1a.s[0]\n"
+ "fmla v27.4s, bb3.4s, a2a.s[0]\n"
+ "fmla v31.4s, bb3.4s, a3a.s[0]\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
// Unroll 5
- "fmla v16.4s, b0a.4s, a0a.s[1]\n"
- "fmla v20.4s, b0a.4s, a1a.s[1]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, b0a.4s, a2a.s[1]\n"
- "fmla v28.4s, b0a.4s, a3a.s[1]\n"
- "ldr b0aq, [%[b_ptr]]\n"
-
- "fmla v17.4s, b1a.4s, a0a.s[1]\n"
- "fmla v21.4s, b1a.4s, a1a.s[1]\n"
- "fmla v25.4s, b1a.4s, a2a.s[1]\n"
- "fmla v29.4s, b1a.4s, a3a.s[1]\n"
- "ldr b1aq, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, b2a.4s, a0a.s[1]\n"
- "fmla v22.4s, b2a.4s, a1a.s[1]\n"
- "fmla v26.4s, b2a.4s, a2a.s[1]\n"
- "fmla v30.4s, b2a.4s, a3a.s[1]\n"
- "ldr b2aq, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, b3a.4s, a0a.s[1]\n"
- "fmla v23.4s, b3a.4s, a1a.s[1]\n"
- "fmla v27.4s, b3a.4s, a2a.s[1]\n"
- "fmla v31.4s, b3a.4s, a3a.s[1]\n"
- "ldr b3aq, [%[b_ptr], #48]\n"
+ "fmla v16.4s, b0a.4s, a0a.s[1]\n"
+ "fmla v20.4s, b0a.4s, a1a.s[1]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, b0a.4s, a2a.s[1]\n"
+ "fmla v28.4s, b0a.4s, a3a.s[1]\n"
+ "ldr b0aq, [%[b_ptr]]\n"
+
+ "fmla v17.4s, b1a.4s, a0a.s[1]\n"
+ "fmla v21.4s, b1a.4s, a1a.s[1]\n"
+ "fmla v25.4s, b1a.4s, a2a.s[1]\n"
+ "fmla v29.4s, b1a.4s, a3a.s[1]\n"
+ "ldr b1aq, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, b2a.4s, a0a.s[1]\n"
+ "fmla v22.4s, b2a.4s, a1a.s[1]\n"
+ "fmla v26.4s, b2a.4s, a2a.s[1]\n"
+ "fmla v30.4s, b2a.4s, a3a.s[1]\n"
+ "ldr b2aq, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, b3a.4s, a0a.s[1]\n"
+ "fmla v23.4s, b3a.4s, a1a.s[1]\n"
+ "fmla v27.4s, b3a.4s, a2a.s[1]\n"
+ "fmla v31.4s, b3a.4s, a3a.s[1]\n"
+ "ldr b3aq, [%[b_ptr], #48]\n"
// Unroll 6
- "fmla v16.4s, bb0.4s, a0a.s[2]\n"
- "fmla v20.4s, bb0.4s, a1a.s[2]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v24.4s, bb0.4s, a2a.s[2]\n"
- "fmla v28.4s, bb0.4s, a3a.s[2]\n"
-
- "fmla v17.4s, bb1.4s, a0a.s[2]\n"
- "fmla v21.4s, bb1.4s, a1a.s[2]\n"
- "fmla v25.4s, bb1.4s, a2a.s[2]\n"
- "fmla v29.4s, bb1.4s, a3a.s[2]\n"
-
- "fmla v18.4s, bb2.4s, a0a.s[2]\n"
- "fmla v22.4s, bb2.4s, a1a.s[2]\n"
- "fmla v26.4s, bb2.4s, a2a.s[2]\n"
- "fmla v30.4s, bb2.4s, a3a.s[2]\n"
-
- "fmla v19.4s, bb3.4s, a0a.s[2]\n"
- "fmla v23.4s, bb3.4s, a1a.s[2]\n"
- "fmla v27.4s, bb3.4s, a2a.s[2]\n"
- "fmla v31.4s, bb3.4s, a3a.s[2]\n"
+ "fmla v16.4s, bb0.4s, a0a.s[2]\n"
+ "fmla v20.4s, bb0.4s, a1a.s[2]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v24.4s, bb0.4s, a2a.s[2]\n"
+ "fmla v28.4s, bb0.4s, a3a.s[2]\n"
+
+ "fmla v17.4s, bb1.4s, a0a.s[2]\n"
+ "fmla v21.4s, bb1.4s, a1a.s[2]\n"
+ "fmla v25.4s, bb1.4s, a2a.s[2]\n"
+ "fmla v29.4s, bb1.4s, a3a.s[2]\n"
+
+ "fmla v18.4s, bb2.4s, a0a.s[2]\n"
+ "fmla v22.4s, bb2.4s, a1a.s[2]\n"
+ "fmla v26.4s, bb2.4s, a2a.s[2]\n"
+ "fmla v30.4s, bb2.4s, a3a.s[2]\n"
+
+ "fmla v19.4s, bb3.4s, a0a.s[2]\n"
+ "fmla v23.4s, bb3.4s, a1a.s[2]\n"
+ "fmla v27.4s, bb3.4s, a2a.s[2]\n"
+ "fmla v31.4s, bb3.4s, a3a.s[2]\n"
// Unroll 7
- "fmla v16.4s, b0a.4s, a0a.s[3]\n"
- "fmla v17.4s, b1a.4s, a0a.s[3]\n"
- "fmla v18.4s, b2a.4s, a0a.s[3]\n"
- "fmla v19.4s, b3a.4s, a0a.s[3]\n"
- "cbnz %w[odds], 6f\n"
-
- "fmla v20.4s, b0a.4s, a1a.s[3]\n"
- "str q16, [%[c_ptr0]]\n"
- "fmla v21.4s, b1a.4s, a1a.s[3]\n"
- "str q17, [%[c_ptr0], #16]\n"
- "fmla v22.4s, b2a.4s, a1a.s[3]\n"
- "str q18, [%[c_ptr0], #32]\n"
- "fmla v23.4s, b3a.4s, a1a.s[3]\n"
- "str q19, [%[c_ptr0], #48]\n"
-
- "fmla v24.4s, b0a.4s, a2a.s[3]\n"
- "str q20, [%[c_ptr1]]\n"
- "fmla v25.4s, b1a.4s, a2a.s[3]\n"
- "str q21, [%[c_ptr1], #16]\n"
- "fmla v26.4s, b2a.4s, a2a.s[3]\n"
- "str q22, [%[c_ptr1], #32]\n"
- "fmla v27.4s, b3a.4s, a2a.s[3]\n"
- "str q23, [%[c_ptr1], #48]\n"
-
- "fmla v28.4s, b0a.4s, a3a.s[3]\n"
- "str q24, [%[c_ptr2]]\n"
- "fmla v29.4s, b1a.4s, a3a.s[3]\n"
- "str q25, [%[c_ptr2], #16]\n"
- "fmla v30.4s, b2a.4s, a3a.s[3]\n"
- "str q26, [%[c_ptr2], #32]\n"
- "fmla v31.4s, b3a.4s, a3a.s[3]\n"
- "str q27, [%[c_ptr2], #48]\n"
- "b 3f\n"
+ "fmla v16.4s, b0a.4s, a0a.s[3]\n"
+ "fmla v17.4s, b1a.4s, a0a.s[3]\n"
+ "fmla v18.4s, b2a.4s, a0a.s[3]\n"
+ "fmla v19.4s, b3a.4s, a0a.s[3]\n"
+ "cbnz %w[odds], 6f\n"
+
+ "fmla v20.4s, b0a.4s, a1a.s[3]\n"
+ "str q16, [%[c_ptr0]]\n"
+ "fmla v21.4s, b1a.4s, a1a.s[3]\n"
+ "str q17, [%[c_ptr0], #16]\n"
+ "fmla v22.4s, b2a.4s, a1a.s[3]\n"
+ "str q18, [%[c_ptr0], #32]\n"
+ "fmla v23.4s, b3a.4s, a1a.s[3]\n"
+ "str q19, [%[c_ptr0], #48]\n"
+
+ "fmla v24.4s, b0a.4s, a2a.s[3]\n"
+ "str q20, [%[c_ptr1]]\n"
+ "fmla v25.4s, b1a.4s, a2a.s[3]\n"
+ "str q21, [%[c_ptr1], #16]\n"
+ "fmla v26.4s, b2a.4s, a2a.s[3]\n"
+ "str q22, [%[c_ptr1], #32]\n"
+ "fmla v27.4s, b3a.4s, a2a.s[3]\n"
+ "str q23, [%[c_ptr1], #48]\n"
+
+ "fmla v28.4s, b0a.4s, a3a.s[3]\n"
+ "str q24, [%[c_ptr2]]\n"
+ "fmla v29.4s, b1a.4s, a3a.s[3]\n"
+ "str q25, [%[c_ptr2], #16]\n"
+ "fmla v30.4s, b2a.4s, a3a.s[3]\n"
+ "str q26, [%[c_ptr2], #32]\n"
+ "fmla v31.4s, b3a.4s, a3a.s[3]\n"
+ "str q27, [%[c_ptr2], #48]\n"
+ "b 3f\n"
// Odd K case: Just do 4 more.
"2:\n"
- "fmla v21.4s, bb1.4s, a1.s[0]\n"
- "add %[a_ptr0], %[a_ptr0], #16\n"
- "fmla v25.4s, bb1.4s, a2.s[0]\n"
- "add %[a_ptr1], %[a_ptr1], #16\n"
- "fmla v29.4s, bb1.4s, a3.s[0]\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, bb2.4s, a0.s[0]\n"
- "add %[a_ptr2], %[a_ptr2], #16\n"
- "fmla v22.4s, bb2.4s, a1.s[0]\n"
- "add %[a_ptr3], %[a_ptr3], #16\n"
- "fmla v26.4s, bb2.4s, a2.s[0]\n"
- "fmla v30.4s, bb2.4s, a3.s[0]\n"
- "ldr b2q, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, bb3.4s, a0.s[0]\n"
- "fmla v23.4s, bb3.4s, a1.s[0]\n"
- "fmla v27.4s, bb3.4s, a2.s[0]\n"
- "fmla v31.4s, bb3.4s, a3.s[0]\n"
- "ldr b3q, [%[b_ptr], #48]\n"
+ "fmla v21.4s, bb1.4s, a1.s[0]\n"
+ "add %[a_ptr0], %[a_ptr0], #16\n"
+ "fmla v25.4s, bb1.4s, a2.s[0]\n"
+ "add %[a_ptr1], %[a_ptr1], #16\n"
+ "fmla v29.4s, bb1.4s, a3.s[0]\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, bb2.4s, a0.s[0]\n"
+ "add %[a_ptr2], %[a_ptr2], #16\n"
+ "fmla v22.4s, bb2.4s, a1.s[0]\n"
+ "add %[a_ptr3], %[a_ptr3], #16\n"
+ "fmla v26.4s, bb2.4s, a2.s[0]\n"
+ "fmla v30.4s, bb2.4s, a3.s[0]\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, bb3.4s, a0.s[0]\n"
+ "fmla v23.4s, bb3.4s, a1.s[0]\n"
+ "fmla v27.4s, bb3.4s, a2.s[0]\n"
+ "fmla v31.4s, bb3.4s, a3.s[0]\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
// Unroll 1
- "fmla v16.4s, b0a.4s, a0.s[1]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v20.4s, b0a.4s, a1.s[1]\n"
- "fmla v24.4s, b0a.4s, a2.s[1]\n"
- "fmla v28.4s, b0a.4s, a3.s[1]\n"
- "ldr b0aq, [%[b_ptr]]\n"
-
- "fmla v17.4s, b1a.4s, a0.s[1]\n"
- "fmla v21.4s, b1a.4s, a1.s[1]\n"
- "fmla v25.4s, b1a.4s, a2.s[1]\n"
- "fmla v29.4s, b1a.4s, a3.s[1]\n"
- "ldr b1aq, [%[b_ptr], #16]\n"
-
- "fmla v18.4s, b2a.4s, a0.s[1]\n"
- "fmla v22.4s, b2a.4s, a1.s[1]\n"
- "fmla v26.4s, b2a.4s, a2.s[1]\n"
- "fmla v30.4s, b2a.4s, a3.s[1]\n"
- "ldr b2aq, [%[b_ptr], #32]\n"
-
- "fmla v19.4s, b3a.4s, a0.s[1]\n"
- "fmla v23.4s, b3a.4s, a1.s[1]\n"
- "fmla v27.4s, b3a.4s, a2.s[1]\n"
- "fmla v31.4s, b3a.4s, a3.s[1]\n"
- "ldr b3aq, [%[b_ptr], #48]\n"
+ "fmla v16.4s, b0a.4s, a0.s[1]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v20.4s, b0a.4s, a1.s[1]\n"
+ "fmla v24.4s, b0a.4s, a2.s[1]\n"
+ "fmla v28.4s, b0a.4s, a3.s[1]\n"
+ "ldr b0aq, [%[b_ptr]]\n"
+
+ "fmla v17.4s, b1a.4s, a0.s[1]\n"
+ "fmla v21.4s, b1a.4s, a1.s[1]\n"
+ "fmla v25.4s, b1a.4s, a2.s[1]\n"
+ "fmla v29.4s, b1a.4s, a3.s[1]\n"
+ "ldr b1aq, [%[b_ptr], #16]\n"
+
+ "fmla v18.4s, b2a.4s, a0.s[1]\n"
+ "fmla v22.4s, b2a.4s, a1.s[1]\n"
+ "fmla v26.4s, b2a.4s, a2.s[1]\n"
+ "fmla v30.4s, b2a.4s, a3.s[1]\n"
+ "ldr b2aq, [%[b_ptr], #32]\n"
+
+ "fmla v19.4s, b3a.4s, a0.s[1]\n"
+ "fmla v23.4s, b3a.4s, a1.s[1]\n"
+ "fmla v27.4s, b3a.4s, a2.s[1]\n"
+ "fmla v31.4s, b3a.4s, a3.s[1]\n"
+ "ldr b3aq, [%[b_ptr], #48]\n"
// Unroll 2
- "fmla v16.4s, bb0.4s, a0.s[2]\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v20.4s, bb0.4s, a1.s[2]\n"
- "fmla v24.4s, bb0.4s, a2.s[2]\n"
- "fmla v28.4s, bb0.4s, a3.s[2]\n"
-
- "fmla v17.4s, bb1.4s, a0.s[2]\n"
- "fmla v21.4s, bb1.4s, a1.s[2]\n"
- "fmla v25.4s, bb1.4s, a2.s[2]\n"
- "fmla v29.4s, bb1.4s, a3.s[2]\n"
-
- "fmla v18.4s, bb2.4s, a0.s[2]\n"
- "fmla v22.4s, bb2.4s, a1.s[2]\n"
- "fmla v26.4s, bb2.4s, a2.s[2]\n"
- "fmla v30.4s, bb2.4s, a3.s[2]\n"
-
- "fmla v19.4s, bb3.4s, a0.s[2]\n"
- "fmla v23.4s, bb3.4s, a1.s[2]\n"
- "fmla v27.4s, bb3.4s, a2.s[2]\n"
- "fmla v31.4s, bb3.4s, a3.s[2]\n"
+ "fmla v16.4s, bb0.4s, a0.s[2]\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v20.4s, bb0.4s, a1.s[2]\n"
+ "fmla v24.4s, bb0.4s, a2.s[2]\n"
+ "fmla v28.4s, bb0.4s, a3.s[2]\n"
+
+ "fmla v17.4s, bb1.4s, a0.s[2]\n"
+ "fmla v21.4s, bb1.4s, a1.s[2]\n"
+ "fmla v25.4s, bb1.4s, a2.s[2]\n"
+ "fmla v29.4s, bb1.4s, a3.s[2]\n"
+
+ "fmla v18.4s, bb2.4s, a0.s[2]\n"
+ "fmla v22.4s, bb2.4s, a1.s[2]\n"
+ "fmla v26.4s, bb2.4s, a2.s[2]\n"
+ "fmla v30.4s, bb2.4s, a3.s[2]\n"
+
+ "fmla v19.4s, bb3.4s, a0.s[2]\n"
+ "fmla v23.4s, bb3.4s, a1.s[2]\n"
+ "fmla v27.4s, bb3.4s, a2.s[2]\n"
+ "fmla v31.4s, bb3.4s, a3.s[2]\n"
// Unroll 3
- "fmla v16.4s, b0a.4s, a0.s[3]\n"
- "fmla v17.4s, b1a.4s, a0.s[3]\n"
- "fmla v18.4s, b2a.4s, a0.s[3]\n"
- "fmla v19.4s, b3a.4s, a0.s[3]\n"
- "cbnz %w[odds], 7f\n"
-
- "fmla v20.4s, b0a.4s, a1.s[3]\n"
- "str q16, [%[c_ptr0]]\n"
- "fmla v21.4s, b1a.4s, a1.s[3]\n"
- "str q17, [%[c_ptr0], #16]\n"
- "fmla v22.4s, b2a.4s, a1.s[3]\n"
- "str q18, [%[c_ptr0], #32]\n"
- "fmla v23.4s, b3a.4s, a1.s[3]\n"
- "str q19, [%[c_ptr0], #48]\n"
-
- "fmla v24.4s, b0a.4s, a2.s[3]\n"
- "str q20, [%[c_ptr1]]\n"
- "fmla v25.4s, b1a.4s, a2.s[3]\n"
- "str q21, [%[c_ptr1], #16]\n"
- "fmla v26.4s, b2a.4s, a2.s[3]\n"
- "str q22, [%[c_ptr1], #32]\n"
- "fmla v27.4s, b3a.4s, a2.s[3]\n"
- "str q23, [%[c_ptr1], #48]\n"
-
- "fmla v28.4s, b0a.4s, a3.s[3]\n"
- "str q24, [%[c_ptr2]]\n"
- "fmla v29.4s, b1a.4s, a3.s[3]\n"
- "str q25, [%[c_ptr2], #16]\n"
- "fmla v30.4s, b2a.4s, a3.s[3]\n"
- "str q26, [%[c_ptr2], #32]\n"
- "fmla v31.4s, b3a.4s, a3.s[3]\n"
- "str q27, [%[c_ptr2], #48]\n"
- "b 3f\n"
+ "fmla v16.4s, b0a.4s, a0.s[3]\n"
+ "fmla v17.4s, b1a.4s, a0.s[3]\n"
+ "fmla v18.4s, b2a.4s, a0.s[3]\n"
+ "fmla v19.4s, b3a.4s, a0.s[3]\n"
+ "cbnz %w[odds], 7f\n"
+
+ "fmla v20.4s, b0a.4s, a1.s[3]\n"
+ "str q16, [%[c_ptr0]]\n"
+ "fmla v21.4s, b1a.4s, a1.s[3]\n"
+ "str q17, [%[c_ptr0], #16]\n"
+ "fmla v22.4s, b2a.4s, a1.s[3]\n"
+ "str q18, [%[c_ptr0], #32]\n"
+ "fmla v23.4s, b3a.4s, a1.s[3]\n"
+ "str q19, [%[c_ptr0], #48]\n"
+
+ "fmla v24.4s, b0a.4s, a2.s[3]\n"
+ "str q20, [%[c_ptr1]]\n"
+ "fmla v25.4s, b1a.4s, a2.s[3]\n"
+ "str q21, [%[c_ptr1], #16]\n"
+ "fmla v26.4s, b2a.4s, a2.s[3]\n"
+ "str q22, [%[c_ptr1], #32]\n"
+ "fmla v27.4s, b3a.4s, a2.s[3]\n"
+ "str q23, [%[c_ptr1], #48]\n"
+
+ "fmla v28.4s, b0a.4s, a3.s[3]\n"
+ "str q24, [%[c_ptr2]]\n"
+ "fmla v29.4s, b1a.4s, a3.s[3]\n"
+ "str q25, [%[c_ptr2], #16]\n"
+ "fmla v30.4s, b2a.4s, a3.s[3]\n"
+ "str q26, [%[c_ptr2], #32]\n"
+ "fmla v31.4s, b3a.4s, a3.s[3]\n"
+ "str q27, [%[c_ptr2], #48]\n"
+ "b 3f\n"
// "Odd ones" - lead in from even
"6:\n"
- "fmla v20.4s, b0a.4s, a1a.s[3]\n"
- "fmla v21.4s, b1a.4s, a1a.s[3]\n"
- "ldr b0q, [%[b_ptr]]\n"
- "fmla v22.4s, b2a.4s, a1a.s[3]\n"
- "subs %w[odds], %w[odds], #1\n"
- "fmla v23.4s, b3a.4s, a1a.s[3]\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v24.4s, b0a.4s, a2a.s[3]\n"
- "fmla v25.4s, b1a.4s, a2a.s[3]\n"
- "ldr b2q, [%[b_ptr], #32]\n"
- "fmla v26.4s, b2a.4s, a2a.s[3]\n"
- "fmla v27.4s, b3a.4s, a2a.s[3]\n"
- "ldr b3q, [%[b_ptr], #48]\n"
-
- "fmla v28.4s, b0a.4s, a3a.s[3]\n"
- "ld1r {a0.4s}, [%[a_ptr0]], #4\n"
- "fmla v29.4s, b1a.4s, a3a.s[3]\n"
- "fmla v30.4s, b2a.4s, a3a.s[3]\n"
- "ld1r {a1.4s}, [%[a_ptr1]], #4\n"
- "fmla v31.4s, b3a.4s, a3a.s[3]\n"
-
- "fmla v16.4s, bb0.4s, a0.4s\n"
- "beq 9f\n"
- "b 8f\n"
+ "fmla v20.4s, b0a.4s, a1a.s[3]\n"
+ "fmla v21.4s, b1a.4s, a1a.s[3]\n"
+ "ldr b0q, [%[b_ptr]]\n"
+ "fmla v22.4s, b2a.4s, a1a.s[3]\n"
+ "subs %w[odds], %w[odds], #1\n"
+ "fmla v23.4s, b3a.4s, a1a.s[3]\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v24.4s, b0a.4s, a2a.s[3]\n"
+ "fmla v25.4s, b1a.4s, a2a.s[3]\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+ "fmla v26.4s, b2a.4s, a2a.s[3]\n"
+ "fmla v27.4s, b3a.4s, a2a.s[3]\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
+
+ "fmla v28.4s, b0a.4s, a3a.s[3]\n"
+ "ld1r {a0.4s}, [%[a_ptr0]], #4\n"
+ "fmla v29.4s, b1a.4s, a3a.s[3]\n"
+ "fmla v30.4s, b2a.4s, a3a.s[3]\n"
+ "ld1r {a1.4s}, [%[a_ptr1]], #4\n"
+ "fmla v31.4s, b3a.4s, a3a.s[3]\n"
+
+ "fmla v16.4s, bb0.4s, a0.4s\n"
+ "beq 9f\n"
+ "b 8f\n"
// "Odd ones" - lead in from odd
"7:\n"
- "fmla v20.4s, b0a.4s, a1.s[3]\n"
- "subs %w[odds], %w[odds], #1\n"
- "fmla v21.4s, b1a.4s, a1.s[3]\n"
- "ldr b0q, [%[b_ptr]]\n"
- "fmla v22.4s, b2a.4s, a1.s[3]\n"
- "fmla v23.4s, b3a.4s, a1.s[3]\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v24.4s, b0a.4s, a2.s[3]\n"
- "fmla v25.4s, b1a.4s, a2.s[3]\n"
- "ldr b2q, [%[b_ptr], #32]\n"
- "fmla v26.4s, b2a.4s, a2.s[3]\n"
- "fmla v27.4s, b3a.4s, a2.s[3]\n"
- "ldr b3q, [%[b_ptr], #48]\n"
-
- "fmla v28.4s, b0a.4s, a3.s[3]\n"
- "ld1r {a0.4s}, [%[a_ptr0]], #4\n"
- "fmla v29.4s, b1a.4s, a3.s[3]\n"
- "fmla v30.4s, b2a.4s, a3.s[3]\n"
- "ld1r {a1.4s}, [%[a_ptr1]], #4\n"
- "fmla v31.4s, b3a.4s, a3.s[3]\n"
-
- "fmla v16.4s, bb0.4s, a0.4s\n"
- "beq 9f\n"
+ "fmla v20.4s, b0a.4s, a1.s[3]\n"
+ "subs %w[odds], %w[odds], #1\n"
+ "fmla v21.4s, b1a.4s, a1.s[3]\n"
+ "ldr b0q, [%[b_ptr]]\n"
+ "fmla v22.4s, b2a.4s, a1.s[3]\n"
+ "fmla v23.4s, b3a.4s, a1.s[3]\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v24.4s, b0a.4s, a2.s[3]\n"
+ "fmla v25.4s, b1a.4s, a2.s[3]\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+ "fmla v26.4s, b2a.4s, a2.s[3]\n"
+ "fmla v27.4s, b3a.4s, a2.s[3]\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
+
+ "fmla v28.4s, b0a.4s, a3.s[3]\n"
+ "ld1r {a0.4s}, [%[a_ptr0]], #4\n"
+ "fmla v29.4s, b1a.4s, a3.s[3]\n"
+ "fmla v30.4s, b2a.4s, a3.s[3]\n"
+ "ld1r {a1.4s}, [%[a_ptr1]], #4\n"
+ "fmla v31.4s, b3a.4s, a3.s[3]\n"
+
+ "fmla v16.4s, bb0.4s, a0.4s\n"
+ "beq 9f\n"
// "Odd ones" - loop
"8:\n"
- "fmla v17.4s, bb1.4s, a0.4s\n"
- "ld1r {a2.4s}, [%[a_ptr2]], #4\n"
- "fmla v18.4s, bb2.4s, a0.4s\n"
- "add %[b_ptr], %[b_ptr], %[ldb]\n"
- "fmla v19.4s, bb3.4s, a0.4s\n"
- "ld1r {a3.4s}, [%[a_ptr3]], #4\n"
-
- "fmla v20.4s, bb0.4s, a1.4s\n"
- "subs %w[odds], %w[odds], #1\n"
- "fmla v21.4s, bb1.4s, a1.4s\n"
- "ld1r {a0.4s}, [%[a_ptr0]], #4\n"
- "fmla v22.4s, bb2.4s, a1.4s\n"
- "fmla v23.4s, bb3.4s, a1.4s\n"
- "ld1r {a1.4s}, [%[a_ptr1]], #4\n"
-
- "fmla v24.4s, bb0.4s, a2.4s\n"
- "fmla v28.4s, bb0.4s, a3.4s\n"
- "ldr b0q, [%[b_ptr]]\n"
- "fmla v25.4s, bb1.4s, a2.4s\n"
- "fmla v29.4s, bb1.4s, a3.4s\n"
- "ldr b1q, [%[b_ptr], #16]\n"
-
- "fmla v26.4s, bb2.4s, a2.4s\n"
- "fmla v30.4s, bb2.4s, a3.4s\n"
- "ldr b2q, [%[b_ptr], #32]\n"
- "fmla v27.4s, bb3.4s, a2.4s\n"
- "fmla v31.4s, bb3.4s, a3.4s\n"
- "ldr b3q, [%[b_ptr], #48]\n"
- "fmla v16.4s, bb0.4s, a0.4s\n"
- "bne 8b\n"
+ "fmla v17.4s, bb1.4s, a0.4s\n"
+ "ld1r {a2.4s}, [%[a_ptr2]], #4\n"
+ "fmla v18.4s, bb2.4s, a0.4s\n"
+ "add %[b_ptr], %[b_ptr], %[ldb]\n"
+ "fmla v19.4s, bb3.4s, a0.4s\n"
+ "ld1r {a3.4s}, [%[a_ptr3]], #4\n"
+
+ "fmla v20.4s, bb0.4s, a1.4s\n"
+ "subs %w[odds], %w[odds], #1\n"
+ "fmla v21.4s, bb1.4s, a1.4s\n"
+ "ld1r {a0.4s}, [%[a_ptr0]], #4\n"
+ "fmla v22.4s, bb2.4s, a1.4s\n"
+ "fmla v23.4s, bb3.4s, a1.4s\n"
+ "ld1r {a1.4s}, [%[a_ptr1]], #4\n"
+
+ "fmla v24.4s, bb0.4s, a2.4s\n"
+ "fmla v28.4s, bb0.4s, a3.4s\n"
+ "ldr b0q, [%[b_ptr]]\n"
+ "fmla v25.4s, bb1.4s, a2.4s\n"
+ "fmla v29.4s, bb1.4s, a3.4s\n"
+ "ldr b1q, [%[b_ptr], #16]\n"
+
+ "fmla v26.4s, bb2.4s, a2.4s\n"
+ "fmla v30.4s, bb2.4s, a3.4s\n"
+ "ldr b2q, [%[b_ptr], #32]\n"
+ "fmla v27.4s, bb3.4s, a2.4s\n"
+ "fmla v31.4s, bb3.4s, a3.4s\n"
+ "ldr b3q, [%[b_ptr], #48]\n"
+ "fmla v16.4s, bb0.4s, a0.4s\n"
+ "bne 8b\n"
// "Odd ones" - detached final iteration
"9:\n"
- "fmla v17.4s, bb1.4s, a0.4s\n"
- "ld1r {a2.4s}, [%[a_ptr2]], #4\n"
- "fmla v18.4s, bb2.4s, a0.4s\n"
- "fmla v19.4s, bb3.4s, a0.4s\n"
- "ld1r {a3.4s}, [%[a_ptr3]], #4\n"
-
- "fmla v20.4s, bb0.4s, a1.4s\n"
- "str q16, [%[c_ptr0]]\n"
- "fmla v21.4s, bb1.4s, a1.4s\n"
- "str q17, [%[c_ptr0], #16]\n"
- "fmla v22.4s, bb2.4s, a1.4s\n"
- "str q18, [%[c_ptr0], #32]\n"
- "fmla v23.4s, bb3.4s, a1.4s\n"
- "str q19, [%[c_ptr0], #48]\n"
-
- "fmla v24.4s, bb0.4s, a2.4s\n"
- "str q20, [%[c_ptr1]]\n"
- "fmla v25.4s, bb1.4s, a2.4s\n"
- "str q21, [%[c_ptr1], #16]\n"
- "fmla v26.4s, bb2.4s, a2.4s\n"
- "str q22, [%[c_ptr1], #32]\n"
- "fmla v27.4s, bb3.4s, a2.4s\n"
- "str q23, [%[c_ptr1], #48]\n"
-
- "fmla v28.4s, bb0.4s, a3.4s\n"
- "str q24, [%[c_ptr2]]\n"
- "fmla v29.4s, bb1.4s, a3.4s\n"
- "str q25, [%[c_ptr2], #16]\n"
- "fmla v30.4s, bb2.4s, a3.4s\n"
- "str q26, [%[c_ptr2], #32]\n"
- "fmla v31.4s, bb3.4s, a3.4s\n"
- "str q27, [%[c_ptr2], #48]\n"
+ "fmla v17.4s, bb1.4s, a0.4s\n"
+ "ld1r {a2.4s}, [%[a_ptr2]], #4\n"
+ "fmla v18.4s, bb2.4s, a0.4s\n"
+ "fmla v19.4s, bb3.4s, a0.4s\n"
+ "ld1r {a3.4s}, [%[a_ptr3]], #4\n"
+
+ "fmla v20.4s, bb0.4s, a1.4s\n"
+ "str q16, [%[c_ptr0]]\n"
+ "fmla v21.4s, bb1.4s, a1.4s\n"
+ "str q17, [%[c_ptr0], #16]\n"
+ "fmla v22.4s, bb2.4s, a1.4s\n"
+ "str q18, [%[c_ptr0], #32]\n"
+ "fmla v23.4s, bb3.4s, a1.4s\n"
+ "str q19, [%[c_ptr0], #48]\n"
+
+ "fmla v24.4s, bb0.4s, a2.4s\n"
+ "str q20, [%[c_ptr1]]\n"
+ "fmla v25.4s, bb1.4s, a2.4s\n"
+ "str q21, [%[c_ptr1], #16]\n"
+ "fmla v26.4s, bb2.4s, a2.4s\n"
+ "str q22, [%[c_ptr1], #32]\n"
+ "fmla v27.4s, bb3.4s, a2.4s\n"
+ "str q23, [%[c_ptr1], #48]\n"
+
+ "fmla v28.4s, bb0.4s, a3.4s\n"
+ "str q24, [%[c_ptr2]]\n"
+ "fmla v29.4s, bb1.4s, a3.4s\n"
+ "str q25, [%[c_ptr2], #16]\n"
+ "fmla v30.4s, bb2.4s, a3.4s\n"
+ "str q26, [%[c_ptr2], #32]\n"
+ "fmla v31.4s, bb3.4s, a3.4s\n"
+ "str q27, [%[c_ptr2], #48]\n"
"3:\n"
"str q28, [%[c_ptr3]]\n"