diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_16bit.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_16bit.hpp | 131 |
1 files changed, 64 insertions, 67 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_16bit.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_16bit.hpp index 3cbc8815e3..99bb2d66bd 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_16bit.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_16bit.hpp @@ -29,17 +29,15 @@ #include "../asmlib.hpp" -template <> -template <typename T> -void TransformImpl<8, 1, false, 2, 2>::Transform(T *out, const T *in, int ldin, int y0, int ymax, int k0, int kmax) -{ - uint16_t *outptr = (uint16_t *)out; - const uint16_t *inptr = (const uint16_t *)in; +template<> +template<typename T> +void TransformImpl<8, 1, false, 2, 2>::Transform(T *out, const T *in, int ldin, int y0, int ymax, int k0, int kmax) { + uint16_t *outptr = (uint16_t *)out; + const uint16_t *inptr = (const uint16_t *)in; uint16_t zerobuff[24]; - for(int y = y0; y < ymax; y += 8) - { + for (int y=y0; y<ymax; y+=8) { const uint16_t *inptr0 = inptr + y * ldin + k0; const uint16_t *inptr1 = inptr0 + ldin; const uint16_t *inptr2 = inptr1 + ldin; @@ -58,14 +56,11 @@ void TransformImpl<8, 1, false, 2, 2>::Transform(T *out, const T *in, int ldin, prefetch_2x(inptr6); prefetch_2x(inptr7); - int x = (kmax - k0); - for(; x > 7; x -= 8) - { + int x=(kmax-k0); + for (;x>7;x-=8) { /* Cope with ragged cases by copying from a buffer of zeroes instead */ - if((y + 7) >= ymax) - { - switch((y + 7) - ymax) - { + if ((y + 7) >= ymax) { + switch ((y + 7) - ymax) { /* Everything falls through in here */ case 6: inptr1 = zerobuff; @@ -89,72 +84,74 @@ void TransformImpl<8, 1, false, 2, 2>::Transform(T *out, const T *in, int ldin, } int skippf = (x & 31); - __asm __volatile( + __asm __volatile ( // Load up 8 elements (1 vector) from each of 8 sources. - "CBNZ %w[skippf], 1f\n" ASM_PREFETCH("[%[inptr0], #128]") + "CBNZ %w[skippf], 1f\n" + ASM_PREFETCH("[%[inptr0], #128]") ASM_PREFETCH("[%[inptr1], #128]") ASM_PREFETCH("[%[inptr2], #128]") ASM_PREFETCH("[%[inptr3], #128]") "1:\n" - "LDR q0, [%[inptr0]], #16\n" // q0=A0A1A2A3A4A5A6A7 - "LDR q4, [%[inptr4]], #16\n" // q8=E0E1E2E3E4E5E6E7 - "LDR q2, [%[inptr2]], #16\n" // q4=C0C1C2C3... - "LDR q6, [%[inptr6]], #16\n" - "ZIP1 v8.8h, v0.8h, v4.8h\n" // q8=A0E0A1E1A2E2A3E3 - "ZIP2 v16.8h, v0.8h, v4.8h\n" // q16=A4E4A5E5A6E6A7E7 - "ZIP1 v9.8h, v2.8h, v6.8h\n" // q9=C0G0C1G1C2G2C3G3 - "ZIP2 v17.8h, v2.8h, v6.8h\n" // q17=C4G4C5G5C6G6C7G7 - "LDR q1, [%[inptr1]], #16\n" // q1=B0B1B2B3B4B5B6B7 - "LDR q5, [%[inptr5]], #16\n" - "LDR q3, [%[inptr3]], #16\n" // q3=D0D1D2D3.... - "LDR q7, [%[inptr7]], #16\n" - "ZIP1 v10.8h, v1.8h, v5.8h\n" // q18=B0F0B1F1B2F2B3F3 - "ZIP2 v18.8h, v1.8h, v5.8h\n" // q18=B4F4B5F5B6F6B7F7 - "ZIP1 v11.8h, v3.8h, v7.8h\n" // q19=D0H0D1H1D2H2D3H3 - "ZIP2 v19.8h, v3.8h, v7.8h\n" // q19=D4H4D5H5D6H6D7H7 - - "ZIP1 v12.8h, v8.8h, v9.8h\n" // q20=A0C0E0G0A1C1E1G1 - "ZIP2 v20.8h, v8.8h, v9.8h\n" - "ZIP1 v13.8h, v10.8h, v11.8h\n" // q21=B0D0F0H0B1I1F1H1 - "ZIP2 v21.8h, v10.8h, v11.8h\n" - - "CBNZ %w[skippf], 2f\n" ASM_PREFETCH("[%[inptr4], #112]") + "LDR q0, [%[inptr0]], #16\n" // q0=A0A1A2A3A4A5A6A7 + "LDR q4, [%[inptr4]], #16\n" // q8=E0E1E2E3E4E5E6E7 + "LDR q2, [%[inptr2]], #16\n" // q4=C0C1C2C3... + "LDR q6, [%[inptr6]], #16\n" + "ZIP1 v8.8h, v0.8h, v4.8h\n" // q8=A0E0A1E1A2E2A3E3 + "ZIP2 v16.8h, v0.8h, v4.8h\n" // q16=A4E4A5E5A6E6A7E7 + "ZIP1 v9.8h, v2.8h, v6.8h\n" // q9=C0G0C1G1C2G2C3G3 + "ZIP2 v17.8h, v2.8h, v6.8h\n" // q17=C4G4C5G5C6G6C7G7 + "LDR q1, [%[inptr1]], #16\n" // q1=B0B1B2B3B4B5B6B7 + "LDR q5, [%[inptr5]], #16\n" + "LDR q3, [%[inptr3]], #16\n" // q3=D0D1D2D3.... + "LDR q7, [%[inptr7]], #16\n" + "ZIP1 v10.8h, v1.8h, v5.8h\n" // q18=B0F0B1F1B2F2B3F3 + "ZIP2 v18.8h, v1.8h, v5.8h\n" // q18=B4F4B5F5B6F6B7F7 + "ZIP1 v11.8h, v3.8h, v7.8h\n" // q19=D0H0D1H1D2H2D3H3 + "ZIP2 v19.8h, v3.8h, v7.8h\n" // q19=D4H4D5H5D6H6D7H7 + + "ZIP1 v12.8h, v8.8h, v9.8h\n" // q20=A0C0E0G0A1C1E1G1 + "ZIP2 v20.8h, v8.8h, v9.8h\n" + "ZIP1 v13.8h, v10.8h, v11.8h\n" // q21=B0D0F0H0B1I1F1H1 + "ZIP2 v21.8h, v10.8h, v11.8h\n" + + "CBNZ %w[skippf], 2f\n" + ASM_PREFETCH("[%[inptr4], #112]") ASM_PREFETCH("[%[inptr5], #112]") ASM_PREFETCH("[%[inptr6], #112]") ASM_PREFETCH("[%[inptr7], #112]") "2:\n" - "ZIP1 v22.8h, v16.8h, v17.8h\n" - "ZIP2 v30.8h, v16.8h, v17.8h\n" - "ZIP1 v23.8h, v18.8h, v19.8h\n" - "ZIP2 v31.8h, v18.8h, v19.8h\n" - - "ZIP1 v14.8h, v12.8h, v13.8h\n" // q22=A0B0C0D0E0F0G0H0 - "ZIP2 v15.8h, v12.8h, v13.8h\n" // q23=A1B1C1D1E1F1G1H1 - "STP q14, q15, [%[outptr]], #32\n" // Write back first two elements - - "ZIP1 v0.8h, v20.8h, v21.8h\n" - "ZIP2 v1.8h, v20.8h, v21.8h\n" - "STP q0, q1, [%[outptr]], #32\n" // Write back next two elements - - "ZIP1 v2.8h, v22.8h, v23.8h\n" - "ZIP2 v3.8h, v22.8h, v23.8h\n" - "STP q2, q3, [%[outptr]], #32\n" // Write back next two elements - - "ZIP1 v4.8h, v30.8h, v31.8h\n" - "ZIP2 v5.8h, v30.8h, v31.8h\n" - "STP q4, q5, [%[outptr]], #32\n" // Write back last two elements - : [inptr0] "+r"(inptr0), [inptr1] "+r"(inptr1), [inptr2] "+r"(inptr2), [inptr3] "+r"(inptr3), - [inptr4] "+r"(inptr4), [inptr5] "+r"(inptr5), [inptr6] "+r"(inptr6), [inptr7] "+r"(inptr7), [outptr] "+r"(outptr) - : [skippf] "r"(skippf) + "ZIP1 v22.8h, v16.8h, v17.8h\n" + "ZIP2 v30.8h, v16.8h, v17.8h\n" + "ZIP1 v23.8h, v18.8h, v19.8h\n" + "ZIP2 v31.8h, v18.8h, v19.8h\n" + + "ZIP1 v14.8h, v12.8h, v13.8h\n" // q22=A0B0C0D0E0F0G0H0 + "ZIP2 v15.8h, v12.8h, v13.8h\n" // q23=A1B1C1D1E1F1G1H1 + "STP q14, q15, [%[outptr]], #32\n" // Write back first two elements + + "ZIP1 v0.8h, v20.8h, v21.8h\n" + "ZIP2 v1.8h, v20.8h, v21.8h\n" + "STP q0, q1, [%[outptr]], #32\n" // Write back next two elements + + "ZIP1 v2.8h, v22.8h, v23.8h\n" + "ZIP2 v3.8h, v22.8h, v23.8h\n" + "STP q2, q3, [%[outptr]], #32\n" // Write back next two elements + + "ZIP1 v4.8h, v30.8h, v31.8h\n" + "ZIP2 v5.8h, v30.8h, v31.8h\n" + "STP q4, q5, [%[outptr]], #32\n" // Write back last two elements + : [inptr0] "+r" (inptr0), [inptr1] "+r" (inptr1), [inptr2] "+r" (inptr2), [inptr3] "+r" (inptr3), + [inptr4] "+r" (inptr4), [inptr5] "+r" (inptr5), [inptr6] "+r" (inptr6), [inptr7] "+r" (inptr7), [outptr] "+r" (outptr) + : [skippf] "r" (skippf) : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", - "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", - "v25", "v26", "v27", "v28", "v29", "v30", "v31"); + "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", + "v25", "v26", "v27", "v28", "v29", "v30", "v31" + ); } - for(; x > 0; x--) - { + for (;x>0;x--) { *outptr++ = *inptr0++; *outptr++ = *inptr1++; *outptr++ = *inptr2++; |