aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8/generic.cpp536
1 files changed, 276 insertions, 260 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8/generic.cpp
index 42e870e814..7169c8bdec 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8/generic.cpp
@@ -37,311 +37,327 @@
// Note that the intent of this is that either ablocks or bblocks will be 1
// - this construction allows the output loop to proceed in either order.
-namespace arm_gemm
-{
-void a64_sgemm_asimd_12x8_jumps(const float *Apanel, const float *Bpanel, float *Cpanel, int ablocks, int bblocks, int K, long int row_jump = 0, long int block_jump = 0)
-{
+namespace arm_gemm {
+
+void a64_sgemm_asimd_12x8_jumps(const float *Apanel, const float *Bpanel, float *Cpanel, int ablocks, int bblocks, int K, long int row_jump=0, long int block_jump=0) {
const float *a_ptr = Apanel;
- float *c_ptr = Cpanel;
+ float *c_ptr = Cpanel;
- for(int yb = 0; yb < ablocks; yb++)
- {
+ for (int yb=0; yb<ablocks; yb++) {
const float *a_ptr0 = a_ptr;
- const float *b_ptr = Bpanel;
+ const float *b_ptr = Bpanel;
- for(int xb = 0; xb < bblocks; xb++)
- {
+ for (int xb=0; xb<bblocks; xb++) {
a_ptr = a_ptr0;
// Fix up for odd lengths - set a flag if K is odd, but make
// sure we round up the iteration count.
int oddk = (K & 1);
- int k = ((K + 1) / 2) - 1;
+ int k = ((K+1)/2) - 1;
- register float32x4_t a0 asm("v0");
- register float32x4_t a1 asm("v1");
- register float32x4_t b0 asm("v2");
- register float32x4_t b1 asm("v3");
- register float32x4_t b2 asm("v4");
+ register float32x4_t a0 asm("v0");
+ register float32x4_t a1 asm("v1");
+ register float32x4_t b0 asm("v2");
+ register float32x4_t b1 asm("v3");
+ register float32x4_t b2 asm("v4");
register float32x4_t a0a asm("v5");
register float32x4_t a1a asm("v6");
- __asm __volatile(
+ __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" ASM_PREFETCH("[%[b_ptr], #64]") "movi v13.4s, #0x0\n" ASM_PREFETCH("[%[a_ptr], #64]") "movi v14.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #128]") "movi v15.4s, #0x0\n"
- ASM_PREFETCH("[%[a_ptr], #128]") "movi v16.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #192]") "movi v17.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #256]") "movi v18.4s, #0x0\n"
- ASM_PREFETCH("[%[a_ptr], #192]") "movi v19.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #320]") "movi v20.4s, #0x0\n" ASM_PREFETCH("[%[a_ptr], #256]") "movi v21.4s, #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"
+ ASM_PREFETCH("[%[a_ptr], #64]")
+ "movi v14.4s, #0x0\n"
+ ASM_PREFETCH("[%[b_ptr], #128]")
+ "movi v15.4s, #0x0\n"
+ ASM_PREFETCH("[%[a_ptr], #128]")
+ "movi v16.4s, #0x0\n"
+ ASM_PREFETCH("[%[b_ptr], #192]")
+ "movi v17.4s, #0x0\n"
+ ASM_PREFETCH("[%[b_ptr], #256]")
+ "movi v18.4s, #0x0\n"
+ ASM_PREFETCH("[%[a_ptr], #192]")
+ "movi v19.4s, #0x0\n"
+ ASM_PREFETCH("[%[b_ptr], #320]")
+ "movi v20.4s, #0x0\n"
+ ASM_PREFETCH("[%[a_ptr], #256]")
+ "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"
// Loop proper
"1:\n"
- "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
- "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
- "ldr %q[b2], [%[b_ptr], #32]\n"
- "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
- "add %[b_ptr], %[b_ptr], %[row_jump]\n"
- "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
- "ldr %q[a0a], [%[a_ptr], #32]\n"
- "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
- "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
- "ldr %q[a1a], [%[a_ptr], #48]\n"
- "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
- "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
- "ldr %q[b0], [%[b_ptr], #48]\n"
-
- "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
- "fmla v17.4s, %[b1].4s, %[a0].s[1]\n" ASM_PREFETCH("[%[a_ptr], #320]")
- "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"
- "ldr %q[b1], [%[b_ptr], #64]\n"
-
- "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
- "fmla v25.4s, %[b2].4s, %[a0].s[1]\n" ASM_PREFETCH("[%[b_ptr], #448]")
- "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"
- "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
- "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
- "ldr %q[b2], [%[b_ptr], #80]\n"
-
- "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
- "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
- "ldr %q[a0], [%[a_ptr], #64]\n"
- "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
- "add %[b_ptr], %[b_ptr], %[row_jump]\n"
- "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
- "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
- "ldr %q[a1], [%[a_ptr], #80]\n"
+ "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
+ "fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
+ "ldr %q[b2], [%[b_ptr], #32]\n"
+ "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
+ "add %[b_ptr], %[b_ptr], %[row_jump]\n"
+ "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
+ "ldr %q[a0a], [%[a_ptr], #32]\n"
+ "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
+ "fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
+ "ldr %q[a1a], [%[a_ptr], #48]\n"
+ "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
+ "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
+ "ldr %q[b0], [%[b_ptr], #48]\n"
+
+ "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
+ "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
+ ASM_PREFETCH("[%[a_ptr], #320]")
+ "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"
+ "ldr %q[b1], [%[b_ptr], #64]\n"
+
+ "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
+ "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
+ ASM_PREFETCH("[%[b_ptr], #448]")
+ "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"
+ "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
+ "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
+ "ldr %q[b2], [%[b_ptr], #80]\n"
+
+ "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
+ "fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
+ "ldr %q[a0], [%[a_ptr], #64]\n"
+ "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
+ "add %[b_ptr], %[b_ptr], %[row_jump]\n"
+ "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
+ "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
+ "ldr %q[a1], [%[a_ptr], #80]\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"
- "ldr %q[b0], [%[b_ptr], #96]\n"
-
- "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
- "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n" ASM_PREFETCH("[%[b_ptr], #512]")
- "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"
- "ldr %q[b1], [%[b_ptr], #112]\n"
-
- "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
- "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
- "add %[a_ptr], %[a_ptr], #64\n"
- "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
- "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
- "add %[b_ptr], %[b_ptr], #96\n"
- "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
- "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
- "subs %w[k], %w[k], #1\n"
- "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
- "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
- "bne 1b\n"
+ "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
+ "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
+ "ldr %q[b0], [%[b_ptr], #96]\n"
+
+ "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
+ "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
+ ASM_PREFETCH("[%[b_ptr], #512]")
+ "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"
+ "ldr %q[b1], [%[b_ptr], #112]\n"
+
+ "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
+ "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
+ "add %[a_ptr], %[a_ptr], #64\n"
+ "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
+ "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
+ "add %[b_ptr], %[b_ptr], #96\n"
+ "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
+ "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
+ "subs %w[k], %w[k], #1\n"
+ "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
+ "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
+ "bne 1b\n"
// Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
"4:\n"
// Branch to alternative tail for odd K
- "cbnz %w[oddk], 2f\n"
+ "cbnz %w[oddk], 2f\n"
// Detached final iteration (even K)
- "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
+ "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
"fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
- "ldr %q[b2], [%[b_ptr], #32]\n"
- "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
- "add %[b_ptr], %[b_ptr], %[row_jump]\n"
- "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
- "ldr %q[a0a], [%[a_ptr], #32]\n"
- "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
+ "ldr %q[b2], [%[b_ptr], #32]\n"
+ "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
+ "add %[b_ptr], %[b_ptr], %[row_jump]\n"
+ "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
+ "ldr %q[a0a], [%[a_ptr], #32]\n"
+ "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
"fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
- "ldr %q[a1a], [%[a_ptr], #48]\n"
- "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
- "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
- "ldr %q[b0], [%[b_ptr], #48]\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"
- "ldr %q[b1], [%[b_ptr], #64]\n"
-
- "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
- "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
- "add %[a_ptr], %[a_ptr], #64\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"
- "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
- "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
- "ldr %q[b2], [%[b_ptr], #80]\n"
-
- "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
- "add %[b_ptr], %[b_ptr], %[block_jump]\n"
- "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
- "add %[b_ptr], %[b_ptr], #96\n"
+ "ldr %q[a1a], [%[a_ptr], #48]\n"
+ "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
+ "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
+ "ldr %q[b0], [%[b_ptr], #48]\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"
+ "ldr %q[b1], [%[b_ptr], #64]\n"
+
+ "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
+ "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
+ "add %[a_ptr], %[a_ptr], #64\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"
+ "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
+ "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
+ "ldr %q[b2], [%[b_ptr], #80]\n"
+
+ "fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
+ "add %[b_ptr], %[b_ptr], %[block_jump]\n"
+ "fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
+ "add %[b_ptr], %[b_ptr], #96\n"
"fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
- "add %[b_ptr], %[b_ptr], %[row_jump]\n"
- "str q8, [%[c_ptr], #0]\n"
- "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
- "str q16, [%[c_ptr], #16]\n"
- "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
- "str q24, [%[c_ptr], #32]\n"
-
- "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
- "str q9, [%[c_ptr], #48]\n"
- "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
- "str q17, [%[c_ptr], #64]\n"
- "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
- "str q25, [%[c_ptr], #80]\n"
- "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
- "str q10, [%[c_ptr], #96]\n"
-
- "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
- "str q18, [%[c_ptr], #112]\n"
- "fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
- "str q26, [%[c_ptr], #128]\n"
- "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
- "str q11, [%[c_ptr], #144]\n"
-
- "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
- "str q19, [%[c_ptr], #160]\n"
- "fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
- "str q27, [%[c_ptr], #176]\n"
- "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
- "str q12, [%[c_ptr], #192]\n"
+ "add %[b_ptr], %[b_ptr], %[row_jump]\n"
+ "str q8, [%[c_ptr], #0]\n"
+ "fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
+ "str q16, [%[c_ptr], #16]\n"
+ "fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
+ "str q24, [%[c_ptr], #32]\n"
+
+ "fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
+ "str q9, [%[c_ptr], #48]\n"
+ "fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
+ "str q17, [%[c_ptr], #64]\n"
+ "fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
+ "str q25, [%[c_ptr], #80]\n"
+ "fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
+ "str q10, [%[c_ptr], #96]\n"
+
+ "fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
+ "str q18, [%[c_ptr], #112]\n"
+ "fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
+ "str q26, [%[c_ptr], #128]\n"
+ "fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
+ "str q11, [%[c_ptr], #144]\n"
+
+ "fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
+ "str q19, [%[c_ptr], #160]\n"
+ "fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
+ "str q27, [%[c_ptr], #176]\n"
+ "fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
+ "str q12, [%[c_ptr], #192]\n"
"fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
- "str q20, [%[c_ptr], #208]\n"
- "fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
- "str q28, [%[c_ptr], #224]\n"
- "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
- "str q13, [%[c_ptr], #240]\n"
-
- "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
- "str q21, [%[c_ptr], #256]\n"
- "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
- "str q29, [%[c_ptr], #272]\n"
- "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
- "str q14, [%[c_ptr], #288]\n"
-
- "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
- "str q22, [%[c_ptr], #304]\n"
- "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
- "str q30, [%[c_ptr], #320]\n"
- "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
- "str q15, [%[c_ptr], #336]\n"
-
- "b 3f\n"
+ "str q20, [%[c_ptr], #208]\n"
+ "fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
+ "str q28, [%[c_ptr], #224]\n"
+ "fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
+ "str q13, [%[c_ptr], #240]\n"
+
+ "fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
+ "str q21, [%[c_ptr], #256]\n"
+ "fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
+ "str q29, [%[c_ptr], #272]\n"
+ "fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
+ "str q14, [%[c_ptr], #288]\n"
+
+ "fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
+ "str q22, [%[c_ptr], #304]\n"
+ "fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
+ "str q30, [%[c_ptr], #320]\n"
+ "fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
+ "str q15, [%[c_ptr], #336]\n"
+
+ "b 3f\n"
// Detached final iteration (odd K)
"2:\n"
- "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
- "ldr %q[b2], [%[b_ptr], #32]\n"
- "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
- "add %[b_ptr], %[b_ptr], %[row_jump]\n"
+ "fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
+ "ldr %q[b2], [%[b_ptr], #32]\n"
+ "fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
+ "add %[b_ptr], %[b_ptr], %[row_jump]\n"
"fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
- "str q8, [%[c_ptr], #0]\n"
- "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
- "str q16, [%[c_ptr], #16]\n"
- "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
- "add %[b_ptr], %[b_ptr], #48\n"
- "add %[a_ptr], %[a_ptr], #32\n"
- "str q24, [%[c_ptr], #32]\n"
- "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
- "str q9, [%[c_ptr], #48]\n"
-
- "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
- "str q17, [%[c_ptr], #64]\n"
- "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
- "str q25, [%[c_ptr], #80]\n"
- "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
- "str q10, [%[c_ptr], #96]\n"
-
- "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
- "str q18, [%[c_ptr], #112]\n"
- "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
- "str q26, [%[c_ptr], #128]\n"
- "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
- "str q11, [%[c_ptr], #144]\n"
-
- "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
- "str q19, [%[c_ptr], #160]\n"
- "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
- "str q27, [%[c_ptr], #176]\n"
- "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
- "str q12, [%[c_ptr], #192]\n"
+ "str q8, [%[c_ptr], #0]\n"
+ "fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
+ "str q16, [%[c_ptr], #16]\n"
+ "fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
+ "add %[b_ptr], %[b_ptr], #48\n"
+ "add %[a_ptr], %[a_ptr], #32\n"
+ "str q24, [%[c_ptr], #32]\n"
+ "fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
+ "str q9, [%[c_ptr], #48]\n"
+
+ "fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
+ "str q17, [%[c_ptr], #64]\n"
+ "fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
+ "str q25, [%[c_ptr], #80]\n"
+ "fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
+ "str q10, [%[c_ptr], #96]\n"
+
+ "fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
+ "str q18, [%[c_ptr], #112]\n"
+ "fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
+ "str q26, [%[c_ptr], #128]\n"
+ "fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
+ "str q11, [%[c_ptr], #144]\n"
+
+ "fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
+ "str q19, [%[c_ptr], #160]\n"
+ "fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
+ "str q27, [%[c_ptr], #176]\n"
+ "fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
+ "str q12, [%[c_ptr], #192]\n"
"fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
- "str q20, [%[c_ptr], #208]\n"
- "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
- "str q28, [%[c_ptr], #224]\n"
- "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
- "str q13, [%[c_ptr], #240]\n"
-
- "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
- "str q21, [%[c_ptr], #256]\n"
- "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
- "str q29, [%[c_ptr], #272]\n"
- "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
- "str q14, [%[c_ptr], #288]\n"
-
- "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
- "str q22, [%[c_ptr], #304]\n"
- "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
- "str q30, [%[c_ptr], #320]\n"
- "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
- "str q15, [%[c_ptr], #336]\n"
+ "str q20, [%[c_ptr], #208]\n"
+ "fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
+ "str q28, [%[c_ptr], #224]\n"
+ "fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
+ "str q13, [%[c_ptr], #240]\n"
+
+ "fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
+ "str q21, [%[c_ptr], #256]\n"
+ "fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
+ "str q29, [%[c_ptr], #272]\n"
+ "fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
+ "str q14, [%[c_ptr], #288]\n"
+
+ "fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
+ "str q22, [%[c_ptr], #304]\n"
+ "fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
+ "str q30, [%[c_ptr], #320]\n"
+ "fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
+ "str q15, [%[c_ptr], #336]\n"
// Common tail
"3:\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),
- [b0] "+w"(b0), [b1] "+w"(b1), [b2] "+w"(b2), [k] "+r"(k)
- : [oddk] "r"(oddk), [row_jump] "r"(row_jump), [block_jump] "r"(block_jump)
- : "x20", "x21", "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");
+ "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),
+ [b0] "+w" (b0), [b1] "+w" (b1), [b2] "+w" (b2), [k] "+r" (k)
+ : [oddk] "r" (oddk), [row_jump] "r" (row_jump), [block_jump] "r" (block_jump)
+ : "x20", "x21", "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"
+ );
}
}
}
-void a64_sgemm_asimd_12x8(const float *Apanel, const float *Bpanel, float *Cpanel, int ablocks, int bblocks, int K)
-{
+void a64_sgemm_asimd_12x8(const float *Apanel, const float *Bpanel, float *Cpanel, int ablocks, int bblocks, int K) {
a64_sgemm_asimd_12x8_jumps(Apanel, Bpanel, Cpanel, ablocks, bblocks, K, 0, 0);
}
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-20 16:08:23 +0000