diff options
author | Anthony Barbier <anthony.barbier@arm.com> | 2018-07-03 16:22:02 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:54:10 +0000 |
commit | 5f707736413aeac77818c42838296966f8dc6761 (patch) | |
tree | b829ed3243ea5f3085f288836132416c78bc2e72 /src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp | |
parent | 7485d5a62685cb745ab50e970adb722cb71557ac (diff) | |
download | ComputeLibrary-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_gemm_u8_12x8/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp | 535 |
1 files changed, 277 insertions, 258 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp index d026dc54f3..80dd87356d 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8/generic.cpp @@ -31,309 +31,328 @@ #include "dot_toolchain_support.h" #endif -namespace arm_gemm -{ -void a64_gemm_u8_12x8(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpanel, int ablocks, int bblocks, int K) -{ +namespace arm_gemm { + +void a64_gemm_u8_12x8(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpanel, int ablocks, int bblocks, 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 init_value_k = ((W + 1) / 2) - 1; - for(int yb = 0; yb < ablocks; yb++) - { + const int oddk = (W & 1); + const int init_value_k = ((W+1)/2) - 1; + for (int yb=0; yb<ablocks; yb++) { const uint8_t *a_ptr0 = a_ptr; - const uint8_t *b_ptr = Bpanel; - for(int xb = 0; xb < bblocks; xb++) - { - a_ptr = a_ptr0; - int k = init_value_k; - register uint8x16_t a0 asm("v0"); - register uint8x16_t a1 asm("v1"); - register uint8x16_t b0 asm("v2"); - register uint8x16_t b1 asm("v3"); - register uint8x16_t b2 asm("v4"); + const uint8_t *b_ptr = Bpanel; + for (int xb=0; xb<bblocks; xb++) { + a_ptr = a_ptr0; + int k = init_value_k; + register uint8x16_t a0 asm("v0"); + register uint8x16_t a1 asm("v1"); + register uint8x16_t b0 asm("v2"); + register uint8x16_t b1 asm("v3"); + register uint8x16_t b2 asm("v4"); register uint8x16_t a0a asm("v5"); register uint8x16_t a1a asm("v6"); - __asm __volatile( + __asm __volatile ( #ifdef NO_DOT_IN_TOOLCHAIN _DECLARE_UDOT #else ".arch armv8.2-a+dotprod\n" #endif // 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" - "udot v8.4s , %[b0].16b, %[a0].4b[0]\n" - "udot v9.4s , %[b0].16b, %[a0].4b[1]\n" - - "ldr %q[b2], [%[b_ptr], #32]\n" - "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" - "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" - "ldr %q[a0a], [%[a_ptr], #32]\n" - "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" - "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" - "ldr %q[a1a], [%[a_ptr], #48]\n" - "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" - "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" - "ldr %q[b0], [%[b_ptr], #48]\n" - - "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" - "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" ASM_PREFETCH("[%[a_ptr], #320]") - "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" - "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" - "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" - "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" - "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" - "ldr %q[b1], [%[b_ptr], #64]\n" - - "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" - "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" ASM_PREFETCH("[%[b_ptr], #448]") - "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" - "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" - "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" - "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" - "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" - "ldr %q[b2], [%[b_ptr], #80]\n" - - "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" - "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n" - "ldr %q[a0], [%[a_ptr], #64]\n" - "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" - "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" - "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n" - "ldr %q[a1], [%[a_ptr], #80]\n" + "udot v8.4s , %[b0].16b, %[a0].4b[0]\n" + "udot v9.4s , %[b0].16b, %[a0].4b[1]\n" + + "ldr %q[b2], [%[b_ptr], #32]\n" + "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" + "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" + "ldr %q[a0a], [%[a_ptr], #32]\n" + "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" + "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" + "ldr %q[a1a], [%[a_ptr], #48]\n" + "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" + "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" + "ldr %q[b0], [%[b_ptr], #48]\n" + + "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" + "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" + ASM_PREFETCH("[%[a_ptr], #320]") + "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" + "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" + "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" + "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" + "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" + "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" + "ldr %q[b1], [%[b_ptr], #64]\n" + + "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" + "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" + ASM_PREFETCH("[%[b_ptr], #448]") + "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" + "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" + "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" + "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" + "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" + "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" + "ldr %q[b2], [%[b_ptr], #80]\n" + + "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" + "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n" + "ldr %q[a0], [%[a_ptr], #64]\n" + "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" + "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" + "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n" + "ldr %q[a1], [%[a_ptr], #80]\n" "udot v13.4s, %[b0].16b, %[a1a].4b[1]\n" - "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" - "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n" - "ldr %q[b0], [%[b_ptr], #96]\n" - - "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" - "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n" ASM_PREFETCH("[%[b_ptr], #512]") - "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" - "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" - "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" - "ldr %q[b1], [%[b_ptr], #112]\n" - - "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n" - "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n" - "add %[a_ptr], %[a_ptr], #64\n" - "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n" - "add %[b_ptr], %[b_ptr], #96\n" - "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" - "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n" - "subs %w[k], %w[k], #1\n" - "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n" - "udot v31.4s, %[b2].16b, %[a1a].4b[3]\n" - "bne 1b\n" + "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" + "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n" + "ldr %q[b0], [%[b_ptr], #96]\n" + + "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" + "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n" + ASM_PREFETCH("[%[b_ptr], #512]") + "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" + "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" + "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n" + "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" + "ldr %q[b1], [%[b_ptr], #112]\n" + + "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n" + "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n" + "add %[a_ptr], %[a_ptr], #64\n" + "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" + "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n" + "add %[b_ptr], %[b_ptr], #96\n" + "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" + "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n" + "subs %w[k], %w[k], #1\n" + "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n" + "udot v31.4s, %[b2].16b, %[a1a].4b[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) - "udot v8.4s , %[b0].16b, %[a0].4b[0]\n" + "udot v8.4s , %[b0].16b, %[a0].4b[0]\n" "udot v9.4s , %[b0].16b, %[a0].4b[1]\n" - "ldr %q[b2], [%[b_ptr], #32]\n" - "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" - "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" - "ldr %q[a0a], [%[a_ptr], #32]\n" - "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" + "ldr %q[b2], [%[b_ptr], #32]\n" + "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" + "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" + "ldr %q[a0a], [%[a_ptr], #32]\n" + "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" - "ldr %q[a1a], [%[a_ptr], #48]\n" - "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" - "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" - "ldr %q[b0], [%[b_ptr], #48]\n" - - "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" - "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" - "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" - "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" - "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" - "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" - "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" - "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" - "ldr %q[b1], [%[b_ptr], #64]\n" - - "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" - "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" - "add %[a_ptr], %[a_ptr], #64\n" - "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" - "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" - "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" - "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" - "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" - "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" - "ldr %q[b2], [%[b_ptr], #80]\n" - - "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" - - "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" - "add %[b_ptr], %[b_ptr], #96\n" + "ldr %q[a1a], [%[a_ptr], #48]\n" + "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" + "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" + "ldr %q[b0], [%[b_ptr], #48]\n" + + "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" + "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" + "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" + "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" + "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" + "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" + "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" + "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" + "ldr %q[b1], [%[b_ptr], #64]\n" + + "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" + "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" + "add %[a_ptr], %[a_ptr], #64\n" + "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" + "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" + "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" + "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" + "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" + "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" + "ldr %q[b2], [%[b_ptr], #80]\n" + + "udot v8.4s , %[b0].16b, %[a0a].4b[0]\n" + + "udot v16.4s, %[b1].16b, %[a0a].4b[0]\n" + "add %[b_ptr], %[b_ptr], #96\n" "udot v9.4s , %[b0].16b, %[a0a].4b[1]\n" - "str q8, [%[c_ptr], #0]\n" - "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n" - "str q16, [%[c_ptr], #16]\n" - "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n" - "str q24, [%[c_ptr], #32]\n" - - "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n" - "str q9, [%[c_ptr], #48]\n" - "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" - "str q17, [%[c_ptr], #64]\n" - "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" - "str q25, [%[c_ptr], #80]\n" - "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" - "str q10, [%[c_ptr], #96]\n" - - "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" - "str q18, [%[c_ptr], #112]\n" - "udot v19.4s, %[b1].16b, %[a0a].4b[3]\n" - "str q26, [%[c_ptr], #128]\n" - "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n" - "str q11, [%[c_ptr], #144]\n" - - "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n" - "str q19, [%[c_ptr], #160]\n" - "udot v20.4s, %[b1].16b, %[a1a].4b[0]\n" - "str q27, [%[c_ptr], #176]\n" - "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" - "str q12, [%[c_ptr], #192]\n" + "str q8, [%[c_ptr], #0]\n" + "udot v17.4s, %[b1].16b, %[a0a].4b[1]\n" + "str q16, [%[c_ptr], #16]\n" + "udot v24.4s, %[b2].16b, %[a0a].4b[0]\n" + "str q24, [%[c_ptr], #32]\n" + + "udot v25.4s, %[b2].16b, %[a0a].4b[1]\n" + "str q9, [%[c_ptr], #48]\n" + "udot v10.4s, %[b0].16b, %[a0a].4b[2]\n" + "str q17, [%[c_ptr], #64]\n" + "udot v18.4s, %[b1].16b, %[a0a].4b[2]\n" + "str q25, [%[c_ptr], #80]\n" + "udot v26.4s, %[b2].16b, %[a0a].4b[2]\n" + "str q10, [%[c_ptr], #96]\n" + + "udot v11.4s, %[b0].16b, %[a0a].4b[3]\n" + "str q18, [%[c_ptr], #112]\n" + "udot v19.4s, %[b1].16b, %[a0a].4b[3]\n" + "str q26, [%[c_ptr], #128]\n" + "udot v27.4s, %[b2].16b, %[a0a].4b[3]\n" + "str q11, [%[c_ptr], #144]\n" + + "udot v12.4s, %[b0].16b, %[a1a].4b[0]\n" + "str q19, [%[c_ptr], #160]\n" + "udot v20.4s, %[b1].16b, %[a1a].4b[0]\n" + "str q27, [%[c_ptr], #176]\n" + "udot v28.4s, %[b2].16b, %[a1a].4b[0]\n" + "str q12, [%[c_ptr], #192]\n" "udot v13.4s, %[b0].16b, %[a1a].4b[1]\n" - "str q20, [%[c_ptr], #208]\n" - "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n" - "str q28, [%[c_ptr], #224]\n" - "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n" - "str q13, [%[c_ptr], #240]\n" - - "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" - "str q21, [%[c_ptr], #256]\n" - "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n" - "str q29, [%[c_ptr], #272]\n" - "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n" - "str q14, [%[c_ptr], #288]\n" - - "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n" - "str q22, [%[c_ptr], #304]\n" - "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" - "str q30, [%[c_ptr], #320]\n" - "udot v31.4s, %[b2].16b, %[a1a].4b[3]\n" - "str q15, [%[c_ptr], #336]\n" - - "b 3f\n" + "str q20, [%[c_ptr], #208]\n" + "udot v21.4s, %[b1].16b, %[a1a].4b[1]\n" + "str q28, [%[c_ptr], #224]\n" + "udot v29.4s, %[b2].16b, %[a1a].4b[1]\n" + "str q13, [%[c_ptr], #240]\n" + + "udot v14.4s, %[b0].16b, %[a1a].4b[2]\n" + "str q21, [%[c_ptr], #256]\n" + "udot v22.4s, %[b1].16b, %[a1a].4b[2]\n" + "str q29, [%[c_ptr], #272]\n" + "udot v30.4s, %[b2].16b, %[a1a].4b[2]\n" + "str q14, [%[c_ptr], #288]\n" + + "udot v15.4s, %[b0].16b, %[a1a].4b[3]\n" + "str q22, [%[c_ptr], #304]\n" + "udot v23.4s, %[b1].16b, %[a1a].4b[3]\n" + "str q30, [%[c_ptr], #320]\n" + "udot v31.4s, %[b2].16b, %[a1a].4b[3]\n" + "str q15, [%[c_ptr], #336]\n" + + "b 3f\n" // Detached final iteration (odd K) "2:\n" - "udot v8.4s , %[b0].16b, %[a0].4b[0]\n" - "ldr %q[b2], [%[b_ptr], #32]\n" - "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" + "udot v8.4s , %[b0].16b, %[a0].4b[0]\n" + "ldr %q[b2], [%[b_ptr], #32]\n" + "udot v16.4s, %[b1].16b, %[a0].4b[0]\n" "udot v9.4s , %[b0].16b, %[a0].4b[1]\n" - "str q8, [%[c_ptr], #0]\n" - "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" - "str q16, [%[c_ptr], #16]\n" - "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" - "add %[b_ptr], %[b_ptr], #48\n" - "add %[a_ptr], %[a_ptr], #32\n" - "str q24, [%[c_ptr], #32]\n" - "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" - "str q9, [%[c_ptr], #48]\n" - - "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" - "str q17, [%[c_ptr], #64]\n" - "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" - "str q25, [%[c_ptr], #80]\n" - "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" - "str q10, [%[c_ptr], #96]\n" - - "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" - "str q18, [%[c_ptr], #112]\n" - "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" - "str q26, [%[c_ptr], #128]\n" - "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" - "str q11, [%[c_ptr], #144]\n" - - "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" - "str q19, [%[c_ptr], #160]\n" - "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" - "str q27, [%[c_ptr], #176]\n" - "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" - "str q12, [%[c_ptr], #192]\n" + "str q8, [%[c_ptr], #0]\n" + "udot v17.4s, %[b1].16b, %[a0].4b[1]\n" + "str q16, [%[c_ptr], #16]\n" + "udot v24.4s, %[b2].16b, %[a0].4b[0]\n" + "add %[b_ptr], %[b_ptr], #48\n" + "add %[a_ptr], %[a_ptr], #32\n" + "str q24, [%[c_ptr], #32]\n" + "udot v25.4s, %[b2].16b, %[a0].4b[1]\n" + "str q9, [%[c_ptr], #48]\n" + + "udot v10.4s, %[b0].16b, %[a0].4b[2]\n" + "str q17, [%[c_ptr], #64]\n" + "udot v18.4s, %[b1].16b, %[a0].4b[2]\n" + "str q25, [%[c_ptr], #80]\n" + "udot v26.4s, %[b2].16b, %[a0].4b[2]\n" + "str q10, [%[c_ptr], #96]\n" + + "udot v11.4s, %[b0].16b, %[a0].4b[3]\n" + "str q18, [%[c_ptr], #112]\n" + "udot v19.4s, %[b1].16b, %[a0].4b[3]\n" + "str q26, [%[c_ptr], #128]\n" + "udot v27.4s, %[b2].16b, %[a0].4b[3]\n" + "str q11, [%[c_ptr], #144]\n" + + "udot v12.4s, %[b0].16b, %[a1].4b[0]\n" + "str q19, [%[c_ptr], #160]\n" + "udot v20.4s, %[b1].16b, %[a1].4b[0]\n" + "str q27, [%[c_ptr], #176]\n" + "udot v28.4s, %[b2].16b, %[a1].4b[0]\n" + "str q12, [%[c_ptr], #192]\n" "udot v13.4s, %[b0].16b, %[a1].4b[1]\n" - "str q20, [%[c_ptr], #208]\n" - "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" - "str q28, [%[c_ptr], #224]\n" - "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" - "str q13, [%[c_ptr], #240]\n" - - "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" - "str q21, [%[c_ptr], #256]\n" - "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" - "str q29, [%[c_ptr], #272]\n" - "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" - "str q14, [%[c_ptr], #288]\n" - - "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" - "str q22, [%[c_ptr], #304]\n" - "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" - "str q30, [%[c_ptr], #320]\n" - "udot v31.4s, %[b2].16b, %[a1].4b[3]\n" - "str q15, [%[c_ptr], #336]\n" + "str q20, [%[c_ptr], #208]\n" + "udot v21.4s, %[b1].16b, %[a1].4b[1]\n" + "str q28, [%[c_ptr], #224]\n" + "udot v29.4s, %[b2].16b, %[a1].4b[1]\n" + "str q13, [%[c_ptr], #240]\n" + + "udot v14.4s, %[b0].16b, %[a1].4b[2]\n" + "str q21, [%[c_ptr], #256]\n" + "udot v22.4s, %[b1].16b, %[a1].4b[2]\n" + "str q29, [%[c_ptr], #272]\n" + "udot v30.4s, %[b2].16b, %[a1].4b[2]\n" + "str q14, [%[c_ptr], #288]\n" + + "udot v15.4s, %[b0].16b, %[a1].4b[3]\n" + "str q22, [%[c_ptr], #304]\n" + "udot v23.4s, %[b1].16b, %[a1].4b[3]\n" + "str q30, [%[c_ptr], #320]\n" + "udot v31.4s, %[b2].16b, %[a1].4b[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" + "str q23, [%[c_ptr], #352]\n" + "str q31, [%[c_ptr], #368]\n" + "add %[c_ptr], %[c_ptr], #384\n" #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"); + : + [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" + ); + } } } |