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/merges/a64_merge_int32_12x8.hpp | |
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/merges/a64_merge_int32_12x8.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/merges/a64_merge_int32_12x8.hpp | 340 |
1 files changed, 165 insertions, 175 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/merges/a64_merge_int32_12x8.hpp b/src/core/NEON/kernels/arm_gemm/merges/a64_merge_int32_12x8.hpp index 79dd1f07e3..1a51505a25 100644 --- a/src/core/NEON/kernels/arm_gemm/merges/a64_merge_int32_12x8.hpp +++ b/src/core/NEON/kernels/arm_gemm/merges/a64_merge_int32_12x8.hpp @@ -25,18 +25,16 @@ #ifdef __aarch64__ -template <> -inline void MergeResults<12, 8>(int32_t *out, const int32_t *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const int32_t alpha, const int32_t beta) -{ +template<> +inline void MergeResults<12, 8>(int32_t *out, const int32_t *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const int32_t alpha, const int32_t beta) { const int32_t *inptr = in; prefetch_6x(inptr); prefetch_6x(inptr + 96); int32x4_t alpha_value = vdupq_n_s32(alpha); - int32x4_t beta_value = vdupq_n_s32(beta); + int32x4_t beta_value = vdupq_n_s32(beta); - for(int y = y0; y < ymax; y += 8) - { + for (int y=y0; y<ymax; y+=8) { int32_t *outptr0 = out + (y * ldout) + x0; int32_t *outptr1 = outptr0 + ldout; int32_t *outptr2 = outptr1 + ldout; @@ -55,17 +53,14 @@ inline void MergeResults<12, 8>(int32_t *out, const int32_t *in, const int ldout prefetch_2x(outptr6); prefetch_2x(outptr7); - for(int i = x0; i < xmax; i += 12) - { + for (int i=x0; i<xmax; i+=12) { int32_t dummyres[12]; /* Make sure we throw away results if Y isn't a multiple of 8. * We do this by pointing the result pointer at a dummy buffer * we later discard. */ - if((y + 7) >= ymax) - { - switch((y + 7) - ymax) - { + if ((y+7) >= ymax) { + switch ((y + 7) - ymax) { case 6: outptr1 = dummyres; case 5: @@ -88,12 +83,9 @@ inline void MergeResults<12, 8>(int32_t *out, const int32_t *in, const int ldout } /* For ragged X, manually copy over the valid results. */ - if((i + 11) >= xmax) - { - for(int xi = 0; xi < 12; xi++) - { - if((i + xi) < xmax) - { + if ((i+11) >= xmax) { + for (int xi=0; xi<12; xi++) { + if ((i+xi) < xmax) { *outptr0 = (alpha * inptr[xi]) + (*outptr0 * beta); outptr0++; *outptr1 = (alpha * inptr[xi + 12]) + (*outptr1 * beta); @@ -113,177 +105,175 @@ inline void MergeResults<12, 8>(int32_t *out, const int32_t *in, const int ldout } } inptr += 96; - } - else - { + } else { /* Optimized routine to copy an entire block */ - __asm __volatile( - // Row 0 - ASM_PREFETCH("[%x[outptr1], #192]") - "ldr q3, [%x[outptr0]]\n" - "ldr q4, [%x[outptr0], #0x10]\n" - "ldr q5, [%x[outptr0], #0x20]\n" - "mul v3.4s, v3.4s, %[beta_value].4s\n" - "ldr q6, [%x[inptr]]\n" - "mul v4.4s, v4.4s, %[beta_value].4s\n" - "ldr q7, [%x[inptr], #0x10]\n" - "mul v5.4s, v5.4s, %[beta_value].4s\n" - "ldr q8, [%x[inptr], #0x20]\n" - "mla v3.4s, v6.4s, %[alpha_value].4s\n" - "ldr q0, [%x[outptr1]]\n" - "mla v4.4s, v7.4s, %[alpha_value].4s\n" - "ldr q1, [%x[outptr1], #0x10]\n" - "mla v5.4s, v8.4s, %[alpha_value].4s\n" - "ldr q2, [%x[outptr1], #0x20]\n" + __asm __volatile ( + // Row 0 + ASM_PREFETCH("[%x[outptr1], #192]") + "ldr q3, [%x[outptr0]]\n" + "ldr q4, [%x[outptr0], #0x10]\n" + "ldr q5, [%x[outptr0], #0x20]\n" + "mul v3.4s, v3.4s, %[beta_value].4s\n" + "ldr q6, [%x[inptr]]\n" + "mul v4.4s, v4.4s, %[beta_value].4s\n" + "ldr q7, [%x[inptr], #0x10]\n" + "mul v5.4s, v5.4s, %[beta_value].4s\n" + "ldr q8, [%x[inptr], #0x20]\n" + "mla v3.4s, v6.4s, %[alpha_value].4s\n" + "ldr q0, [%x[outptr1]]\n" + "mla v4.4s, v7.4s, %[alpha_value].4s\n" + "ldr q1, [%x[outptr1], #0x10]\n" + "mla v5.4s, v8.4s, %[alpha_value].4s\n" + "ldr q2, [%x[outptr1], #0x20]\n" - // Row 1 - ASM_PREFETCH("[%x[outptr2], #192]") - "mul v0.4s, v0.4s, %[beta_value].4s\n" - "ldr q6, [%x[inptr], #0x30]\n" - "str q3, [%x[outptr0]], #0x10\n" - "mul v1.4s, v1.4s, %[beta_value].4s\n" - "ldr q7, [%x[inptr], #0x40]\n" - "str q4, [%x[outptr0]], #0x10\n" - "mul v2.4s, v2.4s, %[beta_value].4s\n" - "ldr q8, [%x[inptr], #0x50]\n" - "str q5, [%x[outptr0]], #0x10\n" - "mla v0.4s, v6.4s, %[alpha_value].4s\n" - "ldr q3, [%x[outptr2]]\n" - "mla v1.4s, v7.4s, %[alpha_value].4s\n" - "ldr q4, [%x[outptr2], #0x10]\n" - "mla v2.4s, v8.4s, %[alpha_value].4s\n" - "ldr q5, [%x[outptr2], #0x20]\n" + // Row 1 + ASM_PREFETCH("[%x[outptr2], #192]") + "mul v0.4s, v0.4s, %[beta_value].4s\n" + "ldr q6, [%x[inptr], #0x30]\n" + "str q3, [%x[outptr0]], #0x10\n" + "mul v1.4s, v1.4s, %[beta_value].4s\n" + "ldr q7, [%x[inptr], #0x40]\n" + "str q4, [%x[outptr0]], #0x10\n" + "mul v2.4s, v2.4s, %[beta_value].4s\n" + "ldr q8, [%x[inptr], #0x50]\n" + "str q5, [%x[outptr0]], #0x10\n" + "mla v0.4s, v6.4s, %[alpha_value].4s\n" + "ldr q3, [%x[outptr2]]\n" + "mla v1.4s, v7.4s, %[alpha_value].4s\n" + "ldr q4, [%x[outptr2], #0x10]\n" + "mla v2.4s, v8.4s, %[alpha_value].4s\n" + "ldr q5, [%x[outptr2], #0x20]\n" - // Row 2 - ASM_PREFETCH("[%x[outptr3], #192]") - "mul v3.4s, v3.4s, %[beta_value].4s\n" - "ldr q6, [%x[inptr], #0x60]\n" - "str q0, [%x[outptr1]], #0x10\n" - "mul v4.4s, v4.4s, %[beta_value].4s\n" - "ldr q7, [%x[inptr], #0x70]\n" - "str q1, [%x[outptr1]], #0x10\n" - "mul v5.4s, v5.4s, %[beta_value].4s\n" - "ldr q8, [%x[inptr], #0x80]\n" - "str q2, [%x[outptr1]], #0x10\n" - "mla v3.4s, v6.4s, %[alpha_value].4s\n" - "ldr q0, [%x[outptr3]]\n" - "mla v4.4s, v7.4s, %[alpha_value].4s\n" - "ldr q1, [%x[outptr3], #0x10]\n" - "mla v5.4s, v8.4s, %[alpha_value].4s\n" - "ldr q2, [%x[outptr3], #0x20]\n" + // Row 2 + ASM_PREFETCH("[%x[outptr3], #192]") + "mul v3.4s, v3.4s, %[beta_value].4s\n" + "ldr q6, [%x[inptr], #0x60]\n" + "str q0, [%x[outptr1]], #0x10\n" + "mul v4.4s, v4.4s, %[beta_value].4s\n" + "ldr q7, [%x[inptr], #0x70]\n" + "str q1, [%x[outptr1]], #0x10\n" + "mul v5.4s, v5.4s, %[beta_value].4s\n" + "ldr q8, [%x[inptr], #0x80]\n" + "str q2, [%x[outptr1]], #0x10\n" + "mla v3.4s, v6.4s, %[alpha_value].4s\n" + "ldr q0, [%x[outptr3]]\n" + "mla v4.4s, v7.4s, %[alpha_value].4s\n" + "ldr q1, [%x[outptr3], #0x10]\n" + "mla v5.4s, v8.4s, %[alpha_value].4s\n" + "ldr q2, [%x[outptr3], #0x20]\n" - // Row 3 - ASM_PREFETCH("[%x[outptr4], #192]") - "mul v0.4s, v0.4s, %[beta_value].4s\n" - "ldr q6, [%x[inptr], #0x90]\n" - "str q3, [%x[outptr2]], #0x10\n" - "mul v1.4s, v1.4s, %[beta_value].4s\n" - "ldr q7, [%x[inptr], #0xa0]\n" - "str q4, [%x[outptr2]], #0x10\n" - "mul v2.4s, v2.4s, %[beta_value].4s\n" - "ldr q8, [%x[inptr], #0xb0]\n" - "str q5, [%x[outptr2]], #0x10\n" - "mla v0.4s, v6.4s, %[alpha_value].4s\n" - "ldr q3, [%x[outptr4]]\n" - "mla v1.4s, v7.4s, %[alpha_value].4s\n" - "ldr q4, [%x[outptr4], #0x10]\n" - "mla v2.4s, v8.4s, %[alpha_value].4s\n" - "ldr q5, [%x[outptr4], #0x20]\n" + // Row 3 + ASM_PREFETCH("[%x[outptr4], #192]") + "mul v0.4s, v0.4s, %[beta_value].4s\n" + "ldr q6, [%x[inptr], #0x90]\n" + "str q3, [%x[outptr2]], #0x10\n" + "mul v1.4s, v1.4s, %[beta_value].4s\n" + "ldr q7, [%x[inptr], #0xa0]\n" + "str q4, [%x[outptr2]], #0x10\n" + "mul v2.4s, v2.4s, %[beta_value].4s\n" + "ldr q8, [%x[inptr], #0xb0]\n" + "str q5, [%x[outptr2]], #0x10\n" + "mla v0.4s, v6.4s, %[alpha_value].4s\n" + "ldr q3, [%x[outptr4]]\n" + "mla v1.4s, v7.4s, %[alpha_value].4s\n" + "ldr q4, [%x[outptr4], #0x10]\n" + "mla v2.4s, v8.4s, %[alpha_value].4s\n" + "ldr q5, [%x[outptr4], #0x20]\n" - // Row 4 - ASM_PREFETCH("[%x[outptr5], #192]") - "mul v3.4s, v3.4s, %[beta_value].4s\n" - "ldr q6, [%x[inptr], #0xc0]\n" - "str q0, [%x[outptr3]], #0x10\n" - "mul v4.4s, v4.4s, %[beta_value].4s\n" - "ldr q7, [%x[inptr], #0xd0]\n" - "str q1, [%x[outptr3]], #0x10\n" - "mul v5.4s, v5.4s, %[beta_value].4s\n" - "ldr q8, [%x[inptr], #0xe0]\n" - "str q2, [%x[outptr3]], #0x10\n" - "mla v3.4s, v6.4s, %[alpha_value].4s\n" - "ldr q0, [%x[outptr5]]\n" - "mla v4.4s, v7.4s, %[alpha_value].4s\n" - "ldr q1, [%x[outptr5], #0x10]\n" - "mla v5.4s, v8.4s, %[alpha_value].4s\n" - "ldr q2, [%x[outptr5], #0x20]\n" + // Row 4 + ASM_PREFETCH("[%x[outptr5], #192]") + "mul v3.4s, v3.4s, %[beta_value].4s\n" + "ldr q6, [%x[inptr], #0xc0]\n" + "str q0, [%x[outptr3]], #0x10\n" + "mul v4.4s, v4.4s, %[beta_value].4s\n" + "ldr q7, [%x[inptr], #0xd0]\n" + "str q1, [%x[outptr3]], #0x10\n" + "mul v5.4s, v5.4s, %[beta_value].4s\n" + "ldr q8, [%x[inptr], #0xe0]\n" + "str q2, [%x[outptr3]], #0x10\n" + "mla v3.4s, v6.4s, %[alpha_value].4s\n" + "ldr q0, [%x[outptr5]]\n" + "mla v4.4s, v7.4s, %[alpha_value].4s\n" + "ldr q1, [%x[outptr5], #0x10]\n" + "mla v5.4s, v8.4s, %[alpha_value].4s\n" + "ldr q2, [%x[outptr5], #0x20]\n" - // Row 5 - ASM_PREFETCH("[%x[outptr6], #192]") - "mul v0.4s, v0.4s, %[beta_value].4s\n" - "ldr q6, [%x[inptr], #0xf0]\n" - "str q3, [%x[outptr4]], #0x10\n" - "mul v1.4s, v1.4s, %[beta_value].4s\n" - "ldr q7, [%x[inptr], #0x100]\n" - "str q4, [%x[outptr4]], #0x10\n" - "mul v2.4s, v2.4s, %[beta_value].4s\n" - "ldr q8, [%x[inptr], #0x110]\n" - "str q5, [%x[outptr4]], #0x10\n" - "mla v0.4s, v6.4s, %[alpha_value].4s\n" - "ldr q3, [%x[outptr6]]\n" - "mla v1.4s, v7.4s, %[alpha_value].4s\n" - "ldr q4, [%x[outptr6], #0x10]\n" - "mla v2.4s, v8.4s, %[alpha_value].4s\n" - "ldr q5, [%x[outptr6], #0x20]\n" + // Row 5 + ASM_PREFETCH("[%x[outptr6], #192]") + "mul v0.4s, v0.4s, %[beta_value].4s\n" + "ldr q6, [%x[inptr], #0xf0]\n" + "str q3, [%x[outptr4]], #0x10\n" + "mul v1.4s, v1.4s, %[beta_value].4s\n" + "ldr q7, [%x[inptr], #0x100]\n" + "str q4, [%x[outptr4]], #0x10\n" + "mul v2.4s, v2.4s, %[beta_value].4s\n" + "ldr q8, [%x[inptr], #0x110]\n" + "str q5, [%x[outptr4]], #0x10\n" + "mla v0.4s, v6.4s, %[alpha_value].4s\n" + "ldr q3, [%x[outptr6]]\n" + "mla v1.4s, v7.4s, %[alpha_value].4s\n" + "ldr q4, [%x[outptr6], #0x10]\n" + "mla v2.4s, v8.4s, %[alpha_value].4s\n" + "ldr q5, [%x[outptr6], #0x20]\n" - // Row 6 - ASM_PREFETCH("[%x[outptr7], #192]") - "mul v3.4s, v3.4s, %[beta_value].4s\n" - "ldr q6, [%x[inptr], #0x120]\n" - "str q0, [%x[outptr5]], #0x10\n" - "mul v4.4s, v4.4s, %[beta_value].4s\n" - "ldr q7, [%x[inptr], #0x130]\n" - "str q1, [%x[outptr5]], #0x10\n" - "mul v5.4s, v5.4s, %[beta_value].4s\n" - "ldr q8, [%x[inptr], #0x140]\n" - "str q2, [%x[outptr5]], #0x10\n" - "mla v3.4s, v6.4s, %[alpha_value].4s\n" - "ldr q0, [%x[outptr7]]\n" - "mla v4.4s, v7.4s, %[alpha_value].4s\n" - "ldr q1, [%x[outptr7], #0x10]\n" - "mla v5.4s, v8.4s, %[alpha_value].4s\n" - "ldr q2, [%x[outptr7], #0x20]\n" + // Row 6 + ASM_PREFETCH("[%x[outptr7], #192]") + "mul v3.4s, v3.4s, %[beta_value].4s\n" + "ldr q6, [%x[inptr], #0x120]\n" + "str q0, [%x[outptr5]], #0x10\n" + "mul v4.4s, v4.4s, %[beta_value].4s\n" + "ldr q7, [%x[inptr], #0x130]\n" + "str q1, [%x[outptr5]], #0x10\n" + "mul v5.4s, v5.4s, %[beta_value].4s\n" + "ldr q8, [%x[inptr], #0x140]\n" + "str q2, [%x[outptr5]], #0x10\n" + "mla v3.4s, v6.4s, %[alpha_value].4s\n" + "ldr q0, [%x[outptr7]]\n" + "mla v4.4s, v7.4s, %[alpha_value].4s\n" + "ldr q1, [%x[outptr7], #0x10]\n" + "mla v5.4s, v8.4s, %[alpha_value].4s\n" + "ldr q2, [%x[outptr7], #0x20]\n" - // Row 7 - "mul v0.4s, v0.4s, %[beta_value].4s\n" - "ldr q6, [%x[inptr], #0x150]\n" - "str q3, [%x[outptr6]], #0x10\n" - "mul v1.4s, v1.4s, %[beta_value].4s\n" - "ldr q7, [%x[inptr], #0x160]\n" - "str q4, [%x[outptr6]], #0x10\n" - "mul v2.4s, v2.4s, %[beta_value].4s\n" - "ldr q8, [%x[inptr], #0x170]\n" - "str q5, [%x[outptr6]], #0x10\n" - "mla v0.4s, v6.4s, %[alpha_value].4s\n" - "mla v1.4s, v7.4s, %[alpha_value].4s\n" - "mla v2.4s, v8.4s, %[alpha_value].4s\n" - "str q0, [%x[outptr7]], #0x10\n" - "str q1, [%x[outptr7]], #0x10\n" - "str q2, [%x[outptr7]], #0x10\n" + // Row 7 + "mul v0.4s, v0.4s, %[beta_value].4s\n" + "ldr q6, [%x[inptr], #0x150]\n" + "str q3, [%x[outptr6]], #0x10\n" + "mul v1.4s, v1.4s, %[beta_value].4s\n" + "ldr q7, [%x[inptr], #0x160]\n" + "str q4, [%x[outptr6]], #0x10\n" + "mul v2.4s, v2.4s, %[beta_value].4s\n" + "ldr q8, [%x[inptr], #0x170]\n" + "str q5, [%x[outptr6]], #0x10\n" + "mla v0.4s, v6.4s, %[alpha_value].4s\n" + "mla v1.4s, v7.4s, %[alpha_value].4s\n" + "mla v2.4s, v8.4s, %[alpha_value].4s\n" + "str q0, [%x[outptr7]], #0x10\n" + "str q1, [%x[outptr7]], #0x10\n" + "str q2, [%x[outptr7]], #0x10\n" - "add %x[inptr], %x[inptr], #0x180\n" - : [outptr0] "+r"(outptr0), - [outptr1] "+r"(outptr1), - [outptr2] "+r"(outptr2), - [outptr3] "+r"(outptr3), - [outptr4] "+r"(outptr4), - [outptr5] "+r"(outptr5), - [outptr6] "+r"(outptr6), - [outptr7] "+r"(outptr7), - [inptr] "+r"(inptr) - : [alpha_value] "w"(alpha_value), - [beta_value] "w"(beta_value) - : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8"); + "add %x[inptr], %x[inptr], #0x180\n" + : [outptr0] "+r" (outptr0), + [outptr1] "+r" (outptr1), + [outptr2] "+r" (outptr2), + [outptr3] "+r" (outptr3), + [outptr4] "+r" (outptr4), + [outptr5] "+r" (outptr5), + [outptr6] "+r" (outptr6), + [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr) + : [alpha_value] "w" (alpha_value), + [beta_value] "w" (beta_value) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8" + ); } } } } -template <> -inline void MergeResults<12, 8>(uint32_t *out, const uint32_t *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const uint32_t alpha, const uint32_t beta) -{ - // Since the above code uses only MUL and MLA instructions discard the "unsignedness" and proceed safely. - MergeResults<12, 8>(reinterpret_cast<int32_t *>(out), reinterpret_cast<const int32_t *>(in), ldout, y0, ymax, x0, xmax, static_cast<const int32_t>(alpha), static_cast<const int32_t>(beta)); +template<> +inline void MergeResults<12, 8>(uint32_t *out, const uint32_t *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const uint32_t alpha, const uint32_t beta) { + // Since the above code uses only MUL and MLA instructions discard the "unsignedness" and proceed safely. + MergeResults<12, 8>(reinterpret_cast<int32_t*>(out), reinterpret_cast<const int32_t*>(in), ldout, y0, ymax, x0, xmax, static_cast<const int32_t>(alpha), static_cast<const int32_t>(beta)); } #endif // __aarch64__ |