diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/a55r1.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/a55r1.cpp | 335 |
1 files changed, 190 insertions, 145 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/a55r1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/a55r1.cpp index f8fafbdf84..994aea65f7 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/a55r1.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/a55r1.cpp @@ -31,40 +31,37 @@ #include "dot_toolchain_support.h" #endif -namespace arm_gemm -{ -void a64_gemm_u8_12x8_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpanel, const int ablocks, const int bblocks, const int K) -{ +namespace arm_gemm { + +void a64_gemm_u8_12x8_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpanel, const int ablocks, const int bblocks, const int K) { const uint8_t *a_ptr = Apanel; - uint32_t *c_ptr = Cpanel; + uint32_t *c_ptr = Cpanel; // We divide K by 4 because the udot instruction processes 4 elements at a time. - const int W = K / 4; + const int W = K/4; // Fix up for odd lengths - set a flag if K is odd, but make // sure we round up the iteration count. - const int oddk = (W & 1); - const int k_iters = ((W + 1) / 2) - 1; + const int oddk = (W & 1); + const int k_iters = ((W+1)/2) - 1; - for(int yb = 0; yb < ablocks; yb++) - { + for (int yb=0; yb<ablocks; yb++) { const uint8_t *a_ptr0 = a_ptr; - const uint8_t *b_ptr = Bpanel; + const uint8_t *b_ptr = Bpanel; - for(int xb = 0; xb < bblocks; xb++) - { + for (int xb=0; xb<bblocks; xb++) { a_ptr = a_ptr0; int k = k_iters; - register int32x4_t a0 asm("v0"); - register int32x4_t a1 asm("v1"); - register int32x4_t b0 asm("v2"); - register int32x4_t b1 asm("v3"); - register int32x4_t b2 asm("v4"); + register int32x4_t a0 asm("v0"); + register int32x4_t a1 asm("v1"); + register int32x4_t b0 asm("v2"); + register int32x4_t b1 asm("v3"); + register int32x4_t b2 asm("v4"); register int32x4_t a0a asm("v5"); register int32x4_t a1a asm("v6"); - __asm __volatile( + __asm __volatile ( #ifdef NO_DOT_IN_TOOLCHAIN _DECLARE_UDOT #else @@ -79,22 +76,39 @@ void a64_gemm_u8_12x8_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32 "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 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" - "movi v19.4s, #0x0\n" ASM_PREFETCH("[%[a_ptr], #192]") + "movi v19.4s, #0x0\n" + ASM_PREFETCH("[%[a_ptr], #192]") "movi v20.4s, #0x0\n" - "movi v21.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #320]") + "movi v21.4s, #0x0\n" + ASM_PREFETCH("[%[b_ptr], #320]") "movi v22.4s, #0x0\n" - "movi v23.4s, #0x0\n" ASM_PREFETCH("[%[a_ptr], #256]") + "movi v23.4s, #0x0\n" + ASM_PREFETCH("[%[a_ptr], #256]") "movi v24.4s, #0x0\n" - "movi v25.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #384]") + "movi v25.4s, #0x0\n" + ASM_PREFETCH("[%[b_ptr], #384]") "movi v26.4s, #0x0\n" - "movi v27.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #448]") + "movi v27.4s, #0x0\n" + ASM_PREFETCH("[%[b_ptr], #448]") "movi v28.4s, #0x0\n" - "movi v29.4s, #0x0\n" ASM_PREFETCH("[%[a_ptr], #384]") + "movi v29.4s, #0x0\n" + ASM_PREFETCH("[%[a_ptr], #384]") "movi v30.4s, #0x0\n" - "movi v31.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #512]") + "movi v31.4s, #0x0\n" + ASM_PREFETCH("[%[b_ptr], #512]") // The loop is offset by these two instructions which must // always be executed. @@ -105,102 +119,105 @@ void a64_gemm_u8_12x8_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32 "cbz %w[k], 4f\n" "1:\n" - "udot v9.4s , %[b0].16b, %[a0].4b[1]\n" - "ldr x20, [%[b_ptr], #40]\n" - "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" - "subs %w[k], %w[k], #1\n" - "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" - "ldr %d[a0a], [%[a_ptr], #32]\n" - - "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" + "udot v9.4s , %[b0].16b, %[a0].4b[1]\n" + "ldr x20, [%[b_ptr], #40]\n" + "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" + "subs %w[k], %w[k], #1\n" + "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" + "ldr %d[a0a], [%[a_ptr], #32]\n" + + "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" "ins %[b2].d[1], x20\n" - "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" + "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" "ldr x20, [%[a_ptr], #40]\n" - "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" - "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" - "ldr %d[a1a], [%[a_ptr], #48]\n" + "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" + "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" + "ldr %d[a1a], [%[a_ptr], #48]\n" - "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" + "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" "ins %[a0a].d[1], x20\n" - "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" + "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" "ldr x20, [%[a_ptr], #56]\n" - "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" - "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" - "ldr %d[b0], [%[b_ptr], #48]\n" + "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" + "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" + "ldr %d[b0], [%[b_ptr], #48]\n" - "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" + "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" "ins %[a1a].d[1], x20\n" - "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" + "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" "ldr x20, [%[b_ptr], #56]\n" - "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" - "ldr %d[b1], [%[b_ptr], #64]\n" + "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" + "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" + "ldr %d[b1], [%[b_ptr], #64]\n" - "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" + "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" "ins %[b0].d[1], x20\n" - "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" + "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" "ldr x20, [%[b_ptr], #72]\n" - "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" ASM_PREFETCH("[%[a_ptr], #448]") + "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" + "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" + ASM_PREFETCH("[%[a_ptr], #448]") - "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" - "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" ASM_PREFETCH("[%[b_ptr], #576]") - "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" - "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" + "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" + "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" + ASM_PREFETCH("[%[b_ptr], #576]") + "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" + "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" - // Unroll 1 - "ldr %d[b2], [%[b_ptr], #80]\n" + // Unroll 1 + "ldr %d[b2], [%[b_ptr], #80]\n" - "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" + "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" "ins %[b1].d[1], x20\n" - "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n" + "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n" "ldr x20, [%[b_ptr], #88]\n" - "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" - "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" - "ldr %d[a0], [%[a_ptr], #64]\n" + "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" + "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" + "ldr %d[a0], [%[a_ptr], #64]\n" - "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n" + "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n" "ins %[b2].d[1], x20\n" "udot v13.4s, %[b0].16b, %[a1a].4b[1]\n" "ldr x20, [%[a_ptr], #72]\n" - "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" - "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n" - "ldr %d[a1], [%[a_ptr], #80]\n" + "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" + "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n" + "ldr %d[a1], [%[a_ptr], #80]\n" - "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" + "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" "ins %[a0].d[1], x20\n" - "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n" + "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n" "ldr x20, [%[a_ptr], #88]\n" - "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" - "udot v19.4s, %[b1].16b, %[a0a].4b[3]\n" - "ldr %d[b0], [%[b_ptr], #96]\n" + "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" + "udot v19.4s, %[b1].16b, %[a0a].4b[3]\n" + "ldr %d[b0], [%[b_ptr], #96]\n" - "udot v20.4s, %[b1].16b, %[a1a].4b[0]\n" + "udot v20.4s, %[b1].16b, %[a1a].4b[0]\n" "ins %[a1].d[1], x20\n" - "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n" + "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n" "ldr x20, [%[b_ptr], #104]\n" - "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" - "ldr %d[b1], [%[b_ptr], #112]\n" + "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n" + "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" + "ldr %d[b1], [%[b_ptr], #112]\n" - "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n" + "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n" "ins %[b0].d[1], x20\n" - "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n" + "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n" "ldr x20, [%[b_ptr], #120]\n" - "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n" - "add %[a_ptr], %[a_ptr], #64\n" - - "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" ASM_PREFETCH("[%[b_ptr], #640]") - "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n" - "add %[b_ptr], %[b_ptr], #96\n" - "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n" + "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" + "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n" + "add %[a_ptr], %[a_ptr], #64\n" + + "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" + ASM_PREFETCH("[%[b_ptr], #640]") + "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n" + "add %[b_ptr], %[b_ptr], #96\n" + "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n" "ins %[b1].d[1], x20\n" - "udot v31.4s, %[b2].16b, %[a1a].4b[3]\n" + "udot v31.4s, %[b2].16b, %[a1a].4b[3]\n" "ldr %d[b2], [%[b_ptr], #32]\n" "udot v8.4s , %[b0].16b, %[a0].4b[0]\n" - "b.ne 1b\n" + "b.ne 1b\n" // Branch here if K=1 or 2. Do the right thing for odd/even at the end. "4:\n" @@ -212,71 +229,83 @@ void a64_gemm_u8_12x8_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32 "cbnz %w[oddk], 2f\n" // Even K continuation - "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" - "ldr %d[a0a], [%[a_ptr], #32]\n" + "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" + "ldr %d[a0a], [%[a_ptr], #32]\n" - "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" + "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" "ins %[b2].d[1], x20\n" "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" "ldr x20, [%[a_ptr], #40]\n" - "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" ASM_PREFETCHW("[%[c_ptr]]") - "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" - "ldr %d[a1a], [%[a_ptr], #48]\n" + "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" + ASM_PREFETCHW("[%[c_ptr]]") + "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" + "ldr %d[a1a], [%[a_ptr], #48]\n" - "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" + "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" "ins %[a0a].d[1], x20\n" - "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" + "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" "ldr x20, [%[a_ptr], #56]\n" - "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" - "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" - "ldr %d[b0], [%[b_ptr], #48]\n" + "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" + "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" + "ldr %d[b0], [%[b_ptr], #48]\n" - "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" + "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" "ins %[a1a].d[1], x20\n" - "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" + "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" "ldr x20, [%[b_ptr], #56]\n" - "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" ASM_PREFETCHW("[%[c_ptr], #64]") - "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" - - "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" - "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" ASM_PREFETCHW("[%[c_ptr], #128]") - "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" - "ldr %d[b1], [%[b_ptr], #64]\n" - - "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" + "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" + ASM_PREFETCHW("[%[c_ptr], #64]") + "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" + + "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" + "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" + ASM_PREFETCHW("[%[c_ptr], #128]") + "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" + "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" + "ldr %d[b1], [%[b_ptr], #64]\n" + + "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" "ins %[b0].d[1], x20\n" - "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" + "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" "ldr x20, [%[b_ptr], #72]\n" - "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" ASM_PREFETCHW("[%[c_ptr], #192]") - "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" - "ldr %d[b2], [%[b_ptr], #80]\n" + "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" + ASM_PREFETCHW("[%[c_ptr], #192]") + "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" + "ldr %d[b2], [%[b_ptr], #80]\n" - "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" + "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" "ins %[b1].d[1], x20\n" "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n" "ldr x20, [%[b_ptr], #88]\n" - "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" + "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" "ins %[b2].d[1], x20\n" - "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" ASM_PREFETCHW("[%[c_ptr], #256]") + "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" + ASM_PREFETCHW("[%[c_ptr], #256]") "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n" "udot v13.4s, %[b0].16b, %[a1a].4b[1]\n" - "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" ASM_PREFETCHW("[%[c_ptr], #320]") + "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" + ASM_PREFETCHW("[%[c_ptr], #320]") "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n" - "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" ASM_PREFETCHWL2("[%[c_ptr], #384]") + "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" + ASM_PREFETCHWL2("[%[c_ptr], #384]") "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n" - "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" ASM_PREFETCHWL2("[%[c_ptr], #448]") + "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" + ASM_PREFETCHWL2("[%[c_ptr], #448]") "udot v19.4s, %[b1].16b, %[a0a].4b[3]\n" "udot v20.4s, %[b1].16b, %[a1a].4b[0]\n" - "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n" ASM_PREFETCHWL2("[%[c_ptr], #512]") + "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n" + ASM_PREFETCHWL2("[%[c_ptr], #512]") "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" ASM_PREFETCHWL2("[%[c_ptr], #576]") + "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" + ASM_PREFETCHWL2("[%[c_ptr], #576]") "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n" "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n" - "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" ASM_PREFETCHWL2("[%[c_ptr], #640]") + "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" + ASM_PREFETCHWL2("[%[c_ptr], #640]") "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n" - "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" ASM_PREFETCHWL2("[%[c_ptr], #704]") + "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" + ASM_PREFETCHWL2("[%[c_ptr], #704]") "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n" "add %[a_ptr], %[a_ptr], #64\n" "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n" @@ -286,27 +315,41 @@ void a64_gemm_u8_12x8_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32 // Odd K continuation "2:\n" - "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" ASM_PREFETCHW("[%[c_ptr]]") + "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" + ASM_PREFETCHW("[%[c_ptr]]") "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" "ins %[b2].d[1], x20\n" - "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" ASM_PREFETCHW("[%[c_ptr], #64]") + "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" + ASM_PREFETCHW("[%[c_ptr], #64]") "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" "add %[a_ptr], %[a_ptr], #32\n" - "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" ASM_PREFETCHW("[%[c_ptr], #128]") + "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" + ASM_PREFETCHW("[%[c_ptr], #128]") "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" "add %[b_ptr], %[b_ptr], #48\n" - "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" ASM_PREFETCHW("[%[c_ptr], #192]") + "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" + ASM_PREFETCHW("[%[c_ptr], #192]") "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" - "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" ASM_PREFETCHW("[%[c_ptr], #256]") + "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" + ASM_PREFETCHW("[%[c_ptr], #256]") "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" - "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" ASM_PREFETCHW("[%[c_ptr], #320]") + "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" + ASM_PREFETCHW("[%[c_ptr], #320]") "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" ASM_PREFETCHWL2("[%[c_ptr], #384]") + "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" + ASM_PREFETCHWL2("[%[c_ptr], #384]") "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" - "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" ASM_PREFETCHWL2("[%[c_ptr], #448]") + "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" + ASM_PREFETCHWL2("[%[c_ptr], #448]") "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" ASM_PREFETCHWL2("[%[c_ptr], #512]") "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" ASM_PREFETCHWL2("[%[c_ptr], #576]") "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" - ASM_PREFETCHWL2("[%[c_ptr], #640]") "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" ASM_PREFETCHWL2("[%[c_ptr], #704]") + "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" + ASM_PREFETCHWL2("[%[c_ptr], #512]") + "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" + ASM_PREFETCHWL2("[%[c_ptr], #576]") + "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" + ASM_PREFETCHWL2("[%[c_ptr], #640]") + "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" + ASM_PREFETCHWL2("[%[c_ptr], #704]") "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" // Common tail @@ -340,13 +383,15 @@ void a64_gemm_u8_12x8_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32 #ifdef NO_DOT_IN_TOOLCHAIN ".purgem udot\n" #endif - : - [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) - : "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", "memory"); + : + [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) + : "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", "memory" + ); + } } } |