aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp496
1 files changed, 248 insertions, 248 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp
index b9a2a3a3ef..00ec904e51 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp
@@ -54,312 +54,312 @@ void a64_sgemm_asimd_8x12_a53(const float *Apanel, const float *Bpanel, float *C
__asm __volatile (
// Initialize result registers, load initial operands, prime prefetches.
- "movi v8.4s, #0x0\n"
- "ldr %q[a0], [%[a_ptr]]\n"
- "movi v9.4s, #0x0\n"
- "ldr %q[b0], [%[b_ptr]]\n"
- "movi v10.4s, #0x0\n"
- "ldr %q[a1], [%[a_ptr], #16]\n"
- "movi v11.4s, #0x0\n"
- "ldr %q[b1], [%[b_ptr], #16]\n"
- "movi v12.4s, #0x0\n"
+ "movi v8.4s, #0x0\n"
+ "ldr %q[a0], [%[a_ptr]]\n"
+ "movi v9.4s, #0x0\n"
+ "ldr %q[b0], [%[b_ptr]]\n"
+ "movi v10.4s, #0x0\n"
+ "ldr %q[a1], [%[a_ptr], #16]\n"
+ "movi v11.4s, #0x0\n"
+ "ldr %q[b1], [%[b_ptr], #16]\n"
+ "movi v12.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #64]")
- "movi v13.4s, #0x0\n"
+ "movi v13.4s, #0x0\n"
ASM_PREFETCH("[%[a_ptr], #64]")
- "movi v14.4s, #0x0\n"
+ "movi v14.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #128]")
- "movi v15.4s, #0x0\n"
+ "movi v15.4s, #0x0\n"
ASM_PREFETCH("[%[a_ptr], #128]")
- "movi v16.4s, #0x0\n"
+ "movi v16.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #192]")
- "movi v17.4s, #0x0\n"
+ "movi v17.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #256]")
- "movi v18.4s, #0x0\n"
+ "movi v18.4s, #0x0\n"
ASM_PREFETCH("[%[a_ptr], #192]")
- "movi v19.4s, #0x0\n"
+ "movi v19.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #320]")
- "movi v20.4s, #0x0\n"
+ "movi v20.4s, #0x0\n"
ASM_PREFETCH("[%[a_ptr], #256]")
- "movi v21.4s, #0x0\n"
+ "movi v21.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #384]")
- "movi v22.4s, #0x0\n"
- "movi v23.4s, #0x0\n"
- "movi v24.4s, #0x0\n"
- "movi v25.4s, #0x0\n"
- "movi v26.4s, #0x0\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 v22.4s, #0x0\n"
+ "movi v23.4s, #0x0\n"
+ "movi v24.4s, #0x0\n"
+ "movi v25.4s, #0x0\n"
+ "movi v26.4s, #0x0\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 loop if we are doing zero iterations of it.
- "cbz %w[k], 4f\n"
+ "cbz %w[k], 4f\n"
"1:\n"
// Unroll 0
- "ldr %d[b2], [%[b_ptr], #32]\n"
+ "ldr %d[b2], [%[b_ptr], #32]\n"
"nop\n"
- "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
- "ldr x20, [%[b_ptr], #40]\n"
- "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
- "subs %w[k], %w[k], #1\n"
- "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
-
- "ldr %d[a0a], [%[a_ptr], #32]\n"
- "ins %[b2].d[1], x20\n"
- "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
- "ldr x20, [%[a_ptr], #40]\n"
- "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
- "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
-
- "ldr %d[a1a], [%[a_ptr], #48]\n"
- "ins %[a0a].d[1], x20\n"
- "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
- "ldr x20, [%[a_ptr], #56]\n"
- "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
- "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
-
- "ldr %d[b0], [%[b_ptr], #48]\n"
- "ins %[a1a].d[1], x20\n"
- "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
- "ldr x20, [%[b_ptr], #56]\n"
- "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
- "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
+ "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
+ "ldr x20, [%[b_ptr], #40]\n"
+ "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
+ "subs %w[k], %w[k], #1\n"
+ "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
+
+ "ldr %d[a0a], [%[a_ptr], #32]\n"
+ "ins %[b2].d[1], x20\n"
+ "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
+ "ldr x20, [%[a_ptr], #40]\n"
+ "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
+ "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
+
+ "ldr %d[a1a], [%[a_ptr], #48]\n"
+ "ins %[a0a].d[1], x20\n"
+ "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
+ "ldr x20, [%[a_ptr], #56]\n"
+ "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
+ "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
+
+ "ldr %d[b0], [%[b_ptr], #48]\n"
+ "ins %[a1a].d[1], x20\n"
+ "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
+ "ldr x20, [%[b_ptr], #56]\n"
+ "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
+ "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
ASM_PREFETCH("[%[a_ptr], #320]")
- "ins %[b0].d[1], x20\n"
- "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
- "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
- "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
+ "ins %[b0].d[1], x20\n"
+ "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
+ "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
+ "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
ASM_PREFETCH("[%[b_ptr], #448]")
"nop\n"
- "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
- "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
- "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
+ "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
+ "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
+ "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
- "ldr %d[b1], [%[b_ptr], #64]\n"
+ "ldr %d[b1], [%[b_ptr], #64]\n"
"nop\n"
- "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
- "ldr x20, [%[b_ptr], #72]\n"
- "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
- "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
+ "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
+ "ldr x20, [%[b_ptr], #72]\n"
+ "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
+ "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
ASM_PREFETCH("[%[b_ptr], #512]")
- "ins %[b1].d[1], x20\n"
- "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
- "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
- "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
+ "ins %[b1].d[1], x20\n"
+ "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
+ "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
+ "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
// Unroll 1
- "ldr %d[b2], [%[b_ptr], #80]\n"
+ "ldr %d[b2], [%[b_ptr], #80]\n"
"nop\n"
- "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
- "ldr x20, [%[b_ptr], #88]\n"
- "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
- "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
-
- "ldr %d[a0], [%[a_ptr], #64]\n"
- "ins %[b2].d[1], x20\n"
- "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
- "ldr x20, [%[a_ptr], #72]\n"
- "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
- "fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
-
- "ldr %d[a1], [%[a_ptr], #80]\n"
- "ins %[a0].d[1], x20\n"
- "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
- "ldr x20, [%[a_ptr], #88]\n"
- "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
- "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
-
- "ldr %d[b0], [%[b_ptr], #96]\n"
- "ins %[a1].d[1], x20\n"
- "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
- "ldr x20, [%[b_ptr], #104]\n"
- "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
- "fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
+ "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
+ "ldr x20, [%[b_ptr], #88]\n"
+ "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
+ "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
+
+ "ldr %d[a0], [%[a_ptr], #64]\n"
+ "ins %[b2].d[1], x20\n"
+ "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
+ "ldr x20, [%[a_ptr], #72]\n"
+ "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
+ "fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
+
+ "ldr %d[a1], [%[a_ptr], #80]\n"
+ "ins %[a0].d[1], x20\n"
+ "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
+ "ldr x20, [%[a_ptr], #88]\n"
+ "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
+ "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
+
+ "ldr %d[b0], [%[b_ptr], #96]\n"
+ "ins %[a1].d[1], x20\n"
+ "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
+ "ldr x20, [%[b_ptr], #104]\n"
+ "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
+ "fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
"nop\n"
- "ins %[b0].d[1], x20\n"
- "fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
- "fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
- "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
+ "ins %[b0].d[1], x20\n"
+ "fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
+ "fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
+ "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
"nop\n"
"nop\n"
- "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
- "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
- "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
+ "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
+ "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
+ "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
- "ldr %d[b1], [%[b_ptr], #112]\n"
+ "ldr %d[b1], [%[b_ptr], #112]\n"
"nop\n"
- "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
- "ldr x20, [%[b_ptr], #120]\n"
- "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
- "add %[a_ptr], %[a_ptr], #64\n"
- "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
- "add %[b_ptr], %[b_ptr], #96\n"
+ "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
+ "ldr x20, [%[b_ptr], #120]\n"
+ "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
+ "add %[a_ptr], %[a_ptr], #64\n"
+ "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
+ "add %[b_ptr], %[b_ptr], #96\n"
"nop\n"
- "ins %[b1].d[1], x20\n"
- "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
- "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
- "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
+ "ins %[b1].d[1], x20\n"
+ "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
+ "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
+ "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
- "bne 1b\n"
+ "bne 1b\n"
// Branch here if K=1 or 2. Do the right thing for odd/even at the end.
"4:\n"
- "cbnz %w[oddk], 2f\n"
+ "cbnz %w[oddk], 2f\n"
// Detached final iteration. (even K)
- "ldr %d[b2], [%[b_ptr], #32]\n"
+ "ldr %d[b2], [%[b_ptr], #32]\n"
"nop\n"
- "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
- "ldr x20, [%[b_ptr], #40]\n"
- "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
- "subs %w[k], %w[k], #1\n"
- "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
-
- "ldr %d[a0a], [%[a_ptr], #32]\n"
- "ins %[b2].d[1], x20\n"
- "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
- "ldr x20, [%[a_ptr], #40]\n"
- "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
- "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
-
- "ldr %d[a1a], [%[a_ptr], #48]\n"
- "ins %[a0a].d[1], x20\n"
- "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
- "ldr x20, [%[a_ptr], #56]\n"
- "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
- "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
-
- "ldr %d[b0], [%[b_ptr], #48]\n"
- "ins %[a1a].d[1], x20\n"
- "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
- "ldr x20, [%[b_ptr], #56]\n"
- "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
- "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
-
- "ins %[b0].d[1], x20\n"
- "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
- "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
- "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
+ "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
+ "ldr x20, [%[b_ptr], #40]\n"
+ "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
+ "subs %w[k], %w[k], #1\n"
+ "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
+
+ "ldr %d[a0a], [%[a_ptr], #32]\n"
+ "ins %[b2].d[1], x20\n"
+ "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
+ "ldr x20, [%[a_ptr], #40]\n"
+ "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
+ "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
+
+ "ldr %d[a1a], [%[a_ptr], #48]\n"
+ "ins %[a0a].d[1], x20\n"
+ "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
+ "ldr x20, [%[a_ptr], #56]\n"
+ "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
+ "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
+
+ "ldr %d[b0], [%[b_ptr], #48]\n"
+ "ins %[a1a].d[1], x20\n"
+ "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
+ "ldr x20, [%[b_ptr], #56]\n"
+ "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
+ "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
+
+ "ins %[b0].d[1], x20\n"
+ "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
+ "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
+ "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
"nop\n"
- "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
- "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
- "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
+ "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
+ "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
+ "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
- "ldr %d[b1], [%[b_ptr], #64]\n"
+ "ldr %d[b1], [%[b_ptr], #64]\n"
"nop\n"
- "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
- "ldr x20, [%[b_ptr], #72]\n"
- "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
- "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
+ "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
+ "ldr x20, [%[b_ptr], #72]\n"
+ "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
+ "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
- "ins %[b1].d[1], x20\n"
- "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
- "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
- "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
+ "ins %[b1].d[1], x20\n"
+ "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
+ "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
+ "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
- "ldr %d[b2], [%[b_ptr], #80]\n"
+ "ldr %d[b2], [%[b_ptr], #80]\n"
"nop\n"
- "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
- "ldr x20, [%[b_ptr], #88]\n"
- "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
- "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
-
- "ins %[b2].d[1], x20\n"
- "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
- "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
- "fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
- "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
- "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
- "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
- "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
- "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
- "fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
- "fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
- "fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
- "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
- "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
- "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
- "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
- "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
- "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
- "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
- "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
- "add %[a_ptr], %[a_ptr], #64\n"
- "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
- "add %[b_ptr], %[b_ptr], #96\n"
- "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
- "b 3f\n"
+ "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
+ "ldr x20, [%[b_ptr], #88]\n"
+ "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
+ "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
+
+ "ins %[b2].d[1], x20\n"
+ "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
+ "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
+ "fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
+ "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
+ "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
+ "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
+ "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
+ "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
+ "fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
+ "fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
+ "fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
+ "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
+ "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
+ "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
+ "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
+ "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
+ "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
+ "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
+ "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
+ "add %[a_ptr], %[a_ptr], #64\n"
+ "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
+ "add %[b_ptr], %[b_ptr], #96\n"
+ "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
+ "b 3f\n"
// Detached final iteration. (odd K)
"2:\n"
- "ldr %d[b2], [%[b_ptr], #32]\n"
+ "ldr %d[b2], [%[b_ptr], #32]\n"
"nop\n"
- "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
- "ldr x20, [%[b_ptr], #40]\n"
- "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
- "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
-
- "ins %[b2].d[1], x20\n"
- "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
- "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
- "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
- "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
- "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
- "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
- "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
- "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
- "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
- "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
- "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
- "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
- "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
- "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
- "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
- "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
- "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
- "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
- "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
- "add %[a_ptr], %[a_ptr], #32\n"
- "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
- "add %[b_ptr], %[b_ptr], #48\n"
- "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
+ "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
+ "ldr x20, [%[b_ptr], #40]\n"
+ "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
+ "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
+
+ "ins %[b2].d[1], x20\n"
+ "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
+ "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
+ "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
+ "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
+ "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
+ "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
+ "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
+ "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
+ "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
+ "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
+ "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
+ "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
+ "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
+ "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
+ "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
+ "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
+ "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
+ "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
+ "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
+ "add %[a_ptr], %[a_ptr], #32\n"
+ "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
+ "add %[b_ptr], %[b_ptr], #48\n"
+ "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
// Common tail
"3:\n"
- "str q8, [%[c_ptr]]\n"
- "str q16, [%[c_ptr], #16]\n"
- "str q24, [%[c_ptr], #32]\n"
- "str q9, [%[c_ptr], #48]\n"
- "str q17, [%[c_ptr], #64]\n"
- "str q25, [%[c_ptr], #80]\n"
- "str q10, [%[c_ptr], #96]\n"
- "str q18, [%[c_ptr], #112]\n"
- "str q26, [%[c_ptr], #128]\n"
- "str q11, [%[c_ptr], #144]\n"
- "str q19, [%[c_ptr], #160]\n"
- "str q27, [%[c_ptr], #176]\n"
- "str q12, [%[c_ptr], #192]\n"
- "str q20, [%[c_ptr], #208]\n"
- "str q28, [%[c_ptr], #224]\n"
- "str q13, [%[c_ptr], #240]\n"
- "str q21, [%[c_ptr], #256]\n"
- "str q29, [%[c_ptr], #272]\n"
- "str q14, [%[c_ptr], #288]\n"
- "str q22, [%[c_ptr], #304]\n"
- "str q30, [%[c_ptr], #320]\n"
- "str q15, [%[c_ptr], #336]\n"
- "str q23, [%[c_ptr], #352]\n"
- "str q31, [%[c_ptr], #368]\n"
- "add %[c_ptr], %[c_ptr], #384\n"
+ "str q8, [%[c_ptr]]\n"
+ "str q16, [%[c_ptr], #16]\n"
+ "str q24, [%[c_ptr], #32]\n"
+ "str q9, [%[c_ptr], #48]\n"
+ "str q17, [%[c_ptr], #64]\n"
+ "str q25, [%[c_ptr], #80]\n"
+ "str q10, [%[c_ptr], #96]\n"
+ "str q18, [%[c_ptr], #112]\n"
+ "str q26, [%[c_ptr], #128]\n"
+ "str q11, [%[c_ptr], #144]\n"
+ "str q19, [%[c_ptr], #160]\n"
+ "str q27, [%[c_ptr], #176]\n"
+ "str q12, [%[c_ptr], #192]\n"
+ "str q20, [%[c_ptr], #208]\n"
+ "str q28, [%[c_ptr], #224]\n"
+ "str q13, [%[c_ptr], #240]\n"
+ "str q21, [%[c_ptr], #256]\n"
+ "str q29, [%[c_ptr], #272]\n"
+ "str q14, [%[c_ptr], #288]\n"
+ "str q22, [%[c_ptr], #304]\n"
+ "str q30, [%[c_ptr], #320]\n"
+ "str q15, [%[c_ptr], #336]\n"
+ "str q23, [%[c_ptr], #352]\n"
+ "str q31, [%[c_ptr], #368]\n"
+ "add %[c_ptr], %[c_ptr], #384\n"
:
[a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
[a0] "+w" (a0), [a1] "+w" (a1), [a0a] "+w" (a0a), [a1a] "+w" (a1a),
generated by cgit v1.2.1 (git 2.23.0) at 2024-08-20 20:25:26 +0000