diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_16x4/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_16x4/generic.cpp | 285 |
1 files changed, 175 insertions, 110 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_16x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_16x4/generic.cpp index 504769b9f0..d60d08a672 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_16x4/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_16x4/generic.cpp @@ -25,20 +25,41 @@ #include <algorithm> +#include "arm_gemm.hpp" #include "../../asmlib.hpp" #include "../../utils.hpp" namespace arm_gemm { -void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, int ldc, float beta, int M, int N, int K) { - const long beta0 = (beta == 0.0f); +void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, int ldc, int M, int N, int K, const float *bias, Activation act, bool append) { const int K_stride = K; const long loops_count = ((K + 4) / 8) - 1; K -= loops_count * 8; const long regs_count = (K / 4) - 1; K -= (regs_count + 1) * 4; const long blocks_count = K / 1; + float nullbias[16]; + if (!append && !bias) { + memset(nullbias, 0, (16 * sizeof(float))); + } + float minval = - static_cast<float>(std::numeric_limits<float>::infinity()); + float maxval = static_cast<float>(std::numeric_limits<float>::infinity()); + const float * const minptr = &minval; + const float * const maxptr = &maxval; + + switch(act.type) + { + default: + case Activation::Type::None: + break; + case Activation::Type::BoundedReLU: + maxval = static_cast<float>(act.param1); + /* fall through */ + case Activation::Type::ReLU: + minval = 0.0f; + break; + } for (int y=0; y<M; y+=4) { const float * const a_ptr0_base = A + (y * lda); @@ -48,7 +69,6 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, for (int x0=0; x0<N; x0+=16ul) { const long width = std::min((unsigned long)N-x0, 16ul); - const float *betaptr = β long loops = loops_count; long regs = regs_count; long blocks = blocks_count; @@ -58,7 +78,7 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, float result_buffer[64]; const unsigned long ldcb = (use_result_buffer ? 16 : ldc) * sizeof(float); float *c_ptr_real = c_ptr0; - if (use_result_buffer && !beta0) { + if (use_result_buffer && append) { for(int cy=0; cy<std::min(M-y, 4); cy++) { for(unsigned int cx=0; cx<width; cx++) { result_buffer[cy * 16 + cx] = c_ptr_real[cy * ldc + cx]; @@ -68,21 +88,22 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, if (use_result_buffer) { c_ptr0 = result_buffer; } + const float *biasptr = bias ? bias+x0 : nullbias; switch(M-y) { case 1: __asm __volatile ( - "cbz %[beta0], 1f\n" - "movi v16.4s, #0\n" + "cbnz %[append], 1f\n" + "ldr q16, [%[biasptr]]\n" + "ldr q17, [%[biasptr], #0x10]\n" + "ldr q18, [%[biasptr], #0x20]\n" + "ldr q19, [%[biasptr], #0x30]\n" "ldr q0, [%[a_ptr0]]\n" - "movi v17.4s, #0\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q8, [%[b_ptr0]]\n" - "movi v18.4s, #0\n" "ldr q9, [%[b_ptr0], #0x10]\n" - "movi v19.4s, #0\n" "ldr q10, [%[b_ptr0], #0x20]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q12, [%[b_ptr0], #0x40]\n" "ldr q13, [%[b_ptr0], #0x50]\n" "ldr q14, [%[b_ptr0], #0x60]\n" @@ -90,21 +111,16 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, "cbz %[loops], 2f\n" "b 3f\n" "1:\n" - "ld1r {v15.4s}, [%[betaptr]]\n" "ldr q16, [%[c_ptr0]]\n" "ldr q17, [%[c_ptr0], #0x10]\n" "ldr q18, [%[c_ptr0], #0x20]\n" "ldr q19, [%[c_ptr0], #0x30]\n" - "fmul v16.4s, v16.4s, v15.4s\n" "ldr q0, [%[a_ptr0]]\n" - "fmul v17.4s, v17.4s, v15.4s\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q8, [%[b_ptr0]]\n" - "fmul v18.4s, v18.4s, v15.4s\n" "ldr q9, [%[b_ptr0], #0x10]\n" - "fmul v19.4s, v19.4s, v15.4s\n" "ldr q10, [%[b_ptr0], #0x20]\n" "ldr q11, [%[b_ptr0], #0x30]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q12, [%[b_ptr0], #0x40]\n" "ldr q13, [%[b_ptr0], #0x50]\n" "ldr q14, [%[b_ptr0], #0x60]\n" @@ -291,13 +307,23 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, "fmla v19.4s, v11.4s, v0.s[0]\n" "b.ne 7b\n" "6:\n" + "ld1r {v14.4s}, [%[minptr]]\n" + "ld1r {v15.4s}, [%[maxptr]]\n" + "fmax v16.4s, v16.4s, v14.4s\n" + "fmax v17.4s, v17.4s, v14.4s\n" + "fmax v18.4s, v18.4s, v14.4s\n" + "fmax v19.4s, v19.4s, v14.4s\n" + "fmin v16.4s, v16.4s, v15.4s\n" + "fmin v17.4s, v17.4s, v15.4s\n" + "fmin v18.4s, v18.4s, v15.4s\n" + "fmin v19.4s, v19.4s, v15.4s\n" "str q16, [%[c_ptr0]]\n" "str q17, [%[c_ptr0], #0x10]\n" "str q18, [%[c_ptr0], #0x20]\n" "str q19, [%[c_ptr0], #0x30]\n" "add %[c_ptr0], %[c_ptr0], #0x40\n" : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks) - : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb) + : [width] "r" (width), [append] "r" (static_cast<uint64_t>(append)), [lda] "r" (ldab), [ldc] "r" (ldcb), [biasptr] "r" (biasptr), [minptr] "r" (minptr), [maxptr] "r" (maxptr) : "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", "cc", "memory" ); break; @@ -307,55 +333,46 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, "c_ptr1 .req X1\n" "add a_ptr1, %[a_ptr0], %[lda]\n" "add c_ptr1, %[c_ptr0], %[ldc]\n" - "cbz %[beta0], 1f\n" - "movi v16.4s, #0\n" + "cbnz %[append], 1f\n" + "ldr q16, [%[biasptr]]\n" + "ldr q17, [%[biasptr], #0x10]\n" + "ldr q18, [%[biasptr], #0x20]\n" + "ldr q19, [%[biasptr], #0x30]\n" + "mov v20.16b, v16.16b\n" "ldr q0, [%[a_ptr0]]\n" - "movi v17.4s, #0\n" + "mov v21.16b, v17.16b\n" "ldr q1, [a_ptr1]\n" - "movi v18.4s, #0\n" + "mov v22.16b, v18.16b\n" "ldr q8, [%[b_ptr0]]\n" - "movi v19.4s, #0\n" + "mov v23.16b, v19.16b\n" "ldr q9, [%[b_ptr0], #0x10]\n" - "movi v20.4s, #0\n" "ldr q10, [%[b_ptr0], #0x20]\n" - "movi v21.4s, #0\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q11, [%[b_ptr0], #0x30]\n" - "movi v22.4s, #0\n" + "add a_ptr1, a_ptr1, #0x10\n" "ldr q12, [%[b_ptr0], #0x40]\n" - "movi v23.4s, #0\n" "ldr q13, [%[b_ptr0], #0x50]\n" "ldr q14, [%[b_ptr0], #0x60]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" - "add a_ptr1, a_ptr1, #0x10\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" "cbz %[loops], 2f\n" "b 3f\n" "1:\n" - "ld1r {v15.4s}, [%[betaptr]]\n" "ldr q16, [%[c_ptr0]]\n" "ldr q17, [%[c_ptr0], #0x10]\n" "ldr q18, [%[c_ptr0], #0x20]\n" "ldr q19, [%[c_ptr0], #0x30]\n" - "fmul v16.4s, v16.4s, v15.4s\n" "ldr q20, [c_ptr1]\n" - "fmul v17.4s, v17.4s, v15.4s\n" "ldr q21, [c_ptr1, #0x10]\n" - "fmul v18.4s, v18.4s, v15.4s\n" "ldr q22, [c_ptr1, #0x20]\n" - "fmul v19.4s, v19.4s, v15.4s\n" "ldr q23, [c_ptr1, #0x30]\n" - "fmul v20.4s, v20.4s, v15.4s\n" "ldr q0, [%[a_ptr0]]\n" - "fmul v21.4s, v21.4s, v15.4s\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q1, [a_ptr1]\n" - "fmul v22.4s, v22.4s, v15.4s\n" + "add a_ptr1, a_ptr1, #0x10\n" "ldr q8, [%[b_ptr0]]\n" - "fmul v23.4s, v23.4s, v15.4s\n" "ldr q9, [%[b_ptr0], #0x10]\n" "ldr q10, [%[b_ptr0], #0x20]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q11, [%[b_ptr0], #0x30]\n" - "add a_ptr1, a_ptr1, #0x10\n" "ldr q12, [%[b_ptr0], #0x40]\n" "ldr q13, [%[b_ptr0], #0x50]\n" "ldr q14, [%[b_ptr0], #0x60]\n" @@ -635,9 +652,27 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, "fmla v23.4s, v11.4s, v1.s[0]\n" "b.ne 7b\n" "6:\n" + "ld1r {v14.4s}, [%[minptr]]\n" + "ld1r {v15.4s}, [%[maxptr]]\n" + "fmax v16.4s, v16.4s, v14.4s\n" + "fmax v17.4s, v17.4s, v14.4s\n" + "fmax v18.4s, v18.4s, v14.4s\n" + "fmax v19.4s, v19.4s, v14.4s\n" + "fmin v16.4s, v16.4s, v15.4s\n" + "fmin v17.4s, v17.4s, v15.4s\n" + "fmin v18.4s, v18.4s, v15.4s\n" + "fmin v19.4s, v19.4s, v15.4s\n" "str q16, [%[c_ptr0]]\n" + "fmax v20.4s, v20.4s, v14.4s\n" + "fmax v21.4s, v21.4s, v14.4s\n" + "fmax v22.4s, v22.4s, v14.4s\n" "str q17, [%[c_ptr0], #0x10]\n" + "fmax v23.4s, v23.4s, v14.4s\n" + "fmin v20.4s, v20.4s, v15.4s\n" + "fmin v21.4s, v21.4s, v15.4s\n" "str q18, [%[c_ptr0], #0x20]\n" + "fmin v22.4s, v22.4s, v15.4s\n" + "fmin v23.4s, v23.4s, v15.4s\n" "str q19, [%[c_ptr0], #0x30]\n" "add %[c_ptr0], %[c_ptr0], #0x40\n" "str q20, [c_ptr1]\n" @@ -647,7 +682,7 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, ".unreq a_ptr1\n" ".unreq c_ptr1\n" : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks) - : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb) + : [width] "r" (width), [append] "r" (static_cast<uint64_t>(append)), [lda] "r" (ldab), [ldc] "r" (ldcb), [biasptr] "r" (biasptr), [minptr] "r" (minptr), [maxptr] "r" (maxptr) : "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", "x0", "x1", "cc", "memory" ); break; @@ -661,71 +696,58 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, "add c_ptr1, %[c_ptr0], %[ldc]\n" "add a_ptr2, a_ptr1, %[lda]\n" "add c_ptr2, c_ptr1, %[ldc]\n" - "cbz %[beta0], 1f\n" - "movi v16.4s, #0\n" + "cbnz %[append], 1f\n" + "ldr q16, [%[biasptr]]\n" + "ldr q17, [%[biasptr], #0x10]\n" + "ldr q18, [%[biasptr], #0x20]\n" + "ldr q19, [%[biasptr], #0x30]\n" + "mov v20.16b, v16.16b\n" "ldr q0, [%[a_ptr0]]\n" - "movi v17.4s, #0\n" + "mov v21.16b, v17.16b\n" "ldr q1, [a_ptr1]\n" - "movi v18.4s, #0\n" + "mov v22.16b, v18.16b\n" "ldr q2, [a_ptr2]\n" - "movi v19.4s, #0\n" + "mov v23.16b, v19.16b\n" "ldr q8, [%[b_ptr0]]\n" - "movi v20.4s, #0\n" + "mov v24.16b, v16.16b\n" "ldr q9, [%[b_ptr0], #0x10]\n" - "movi v21.4s, #0\n" + "mov v25.16b, v17.16b\n" "ldr q10, [%[b_ptr0], #0x20]\n" - "movi v22.4s, #0\n" + "mov v26.16b, v18.16b\n" "ldr q11, [%[b_ptr0], #0x30]\n" - "movi v23.4s, #0\n" + "mov v27.16b, v19.16b\n" "ldr q12, [%[b_ptr0], #0x40]\n" - "movi v24.4s, #0\n" "ldr q13, [%[b_ptr0], #0x50]\n" - "movi v25.4s, #0\n" - "ldr q14, [%[b_ptr0], #0x60]\n" - "movi v26.4s, #0\n" "add %[a_ptr0], %[a_ptr0], #0x10\n" - "movi v27.4s, #0\n" + "ldr q14, [%[b_ptr0], #0x60]\n" "add a_ptr1, a_ptr1, #0x10\n" "add a_ptr2, a_ptr2, #0x10\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" "cbz %[loops], 2f\n" "b 3f\n" "1:\n" - "ld1r {v15.4s}, [%[betaptr]]\n" "ldr q16, [%[c_ptr0]]\n" "ldr q17, [%[c_ptr0], #0x10]\n" "ldr q18, [%[c_ptr0], #0x20]\n" "ldr q19, [%[c_ptr0], #0x30]\n" - "fmul v16.4s, v16.4s, v15.4s\n" "ldr q20, [c_ptr1]\n" - "fmul v17.4s, v17.4s, v15.4s\n" "ldr q21, [c_ptr1, #0x10]\n" - "fmul v18.4s, v18.4s, v15.4s\n" "ldr q22, [c_ptr1, #0x20]\n" - "fmul v19.4s, v19.4s, v15.4s\n" "ldr q23, [c_ptr1, #0x30]\n" - "fmul v20.4s, v20.4s, v15.4s\n" "ldr q24, [c_ptr2]\n" - "fmul v21.4s, v21.4s, v15.4s\n" "ldr q25, [c_ptr2, #0x10]\n" - "fmul v22.4s, v22.4s, v15.4s\n" "ldr q26, [c_ptr2, #0x20]\n" - "fmul v23.4s, v23.4s, v15.4s\n" "ldr q27, [c_ptr2, #0x30]\n" - "fmul v24.4s, v24.4s, v15.4s\n" "ldr q0, [%[a_ptr0]]\n" - "fmul v25.4s, v25.4s, v15.4s\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q1, [a_ptr1]\n" - "fmul v26.4s, v26.4s, v15.4s\n" + "add a_ptr1, a_ptr1, #0x10\n" "ldr q2, [a_ptr2]\n" - "fmul v27.4s, v27.4s, v15.4s\n" + "add a_ptr2, a_ptr2, #0x10\n" "ldr q8, [%[b_ptr0]]\n" "ldr q9, [%[b_ptr0], #0x10]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q10, [%[b_ptr0], #0x20]\n" - "add a_ptr1, a_ptr1, #0x10\n" "ldr q11, [%[b_ptr0], #0x30]\n" - "add a_ptr2, a_ptr2, #0x10\n" "ldr q12, [%[b_ptr0], #0x40]\n" "ldr q13, [%[b_ptr0], #0x50]\n" "ldr q14, [%[b_ptr0], #0x60]\n" @@ -1098,13 +1120,39 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, "fmla v27.4s, v11.4s, v2.s[0]\n" "b.ne 7b\n" "6:\n" + "ld1r {v14.4s}, [%[minptr]]\n" + "ld1r {v15.4s}, [%[maxptr]]\n" + "fmax v16.4s, v16.4s, v14.4s\n" + "fmax v17.4s, v17.4s, v14.4s\n" + "fmax v18.4s, v18.4s, v14.4s\n" + "fmax v19.4s, v19.4s, v14.4s\n" + "fmin v16.4s, v16.4s, v15.4s\n" + "fmin v17.4s, v17.4s, v15.4s\n" + "fmin v18.4s, v18.4s, v15.4s\n" + "fmin v19.4s, v19.4s, v15.4s\n" "str q16, [%[c_ptr0]]\n" + "fmax v20.4s, v20.4s, v14.4s\n" + "fmax v21.4s, v21.4s, v14.4s\n" + "fmax v22.4s, v22.4s, v14.4s\n" "str q17, [%[c_ptr0], #0x10]\n" + "fmax v23.4s, v23.4s, v14.4s\n" + "fmin v20.4s, v20.4s, v15.4s\n" + "fmin v21.4s, v21.4s, v15.4s\n" "str q18, [%[c_ptr0], #0x20]\n" + "fmin v22.4s, v22.4s, v15.4s\n" + "fmin v23.4s, v23.4s, v15.4s\n" + "fmax v24.4s, v24.4s, v14.4s\n" "str q19, [%[c_ptr0], #0x30]\n" + "fmax v25.4s, v25.4s, v14.4s\n" "add %[c_ptr0], %[c_ptr0], #0x40\n" + "fmax v26.4s, v26.4s, v14.4s\n" "str q20, [c_ptr1]\n" + "fmin v24.4s, v24.4s, v15.4s\n" + "fmin v25.4s, v25.4s, v15.4s\n" + "fmax v27.4s, v27.4s, v14.4s\n" "str q21, [c_ptr1, #0x10]\n" + "fmin v26.4s, v26.4s, v15.4s\n" + "fmin v27.4s, v27.4s, v15.4s\n" "str q22, [c_ptr1, #0x20]\n" "str q23, [c_ptr1, #0x30]\n" "str q24, [c_ptr2]\n" @@ -1116,7 +1164,7 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, ".unreq c_ptr1\n" ".unreq c_ptr2\n" : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks) - : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb) + : [width] "r" (width), [append] "r" (static_cast<uint64_t>(append)), [lda] "r" (ldab), [ldc] "r" (ldcb), [biasptr] "r" (biasptr), [minptr] "r" (minptr), [maxptr] "r" (maxptr) : "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", "x0", "x1", "x2", "x3", "cc", "memory" ); break; @@ -1135,87 +1183,70 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, "add c_ptr2, c_ptr1, %[ldc]\n" "add a_ptr3, a_ptr2, %[lda]\n" "add c_ptr3, c_ptr2, %[ldc]\n" - "cbz %[beta0], 1f\n" - "movi v16.4s, #0\n" + "cbnz %[append], 1f\n" + "ldr q16, [%[biasptr]]\n" + "ldr q17, [%[biasptr], #0x10]\n" + "ldr q18, [%[biasptr], #0x20]\n" + "ldr q19, [%[biasptr], #0x30]\n" + "mov v20.16b, v16.16b\n" "ldr q0, [%[a_ptr0]]\n" - "movi v17.4s, #0\n" + "mov v21.16b, v17.16b\n" "ldr q1, [a_ptr1]\n" - "movi v18.4s, #0\n" + "mov v22.16b, v18.16b\n" "ldr q2, [a_ptr2]\n" - "movi v19.4s, #0\n" + "mov v23.16b, v19.16b\n" "ldr q3, [a_ptr3]\n" - "movi v20.4s, #0\n" + "mov v24.16b, v16.16b\n" "ldr q8, [%[b_ptr0]]\n" - "movi v21.4s, #0\n" + "mov v25.16b, v17.16b\n" "ldr q9, [%[b_ptr0], #0x10]\n" - "movi v22.4s, #0\n" + "mov v26.16b, v18.16b\n" "ldr q10, [%[b_ptr0], #0x20]\n" - "movi v23.4s, #0\n" + "mov v27.16b, v19.16b\n" "ldr q11, [%[b_ptr0], #0x30]\n" - "movi v24.4s, #0\n" + "mov v28.16b, v16.16b\n" "ldr q12, [%[b_ptr0], #0x40]\n" - "movi v25.4s, #0\n" + "mov v29.16b, v17.16b\n" "ldr q13, [%[b_ptr0], #0x50]\n" - "movi v26.4s, #0\n" + "mov v30.16b, v18.16b\n" "ldr q14, [%[b_ptr0], #0x60]\n" - "movi v27.4s, #0\n" + "mov v31.16b, v19.16b\n" "add %[a_ptr0], %[a_ptr0], #0x10\n" - "movi v28.4s, #0\n" "add a_ptr1, a_ptr1, #0x10\n" - "movi v29.4s, #0\n" "add a_ptr2, a_ptr2, #0x10\n" - "movi v30.4s, #0\n" "add a_ptr3, a_ptr3, #0x10\n" - "movi v31.4s, #0\n" "add %[b_ptr0], %[b_ptr0], #0x80\n" "cbz %[loops], 2f\n" "b 3f\n" "1:\n" - "ld1r {v15.4s}, [%[betaptr]]\n" "ldr q16, [%[c_ptr0]]\n" "ldr q17, [%[c_ptr0], #0x10]\n" "ldr q18, [%[c_ptr0], #0x20]\n" "ldr q19, [%[c_ptr0], #0x30]\n" - "fmul v16.4s, v16.4s, v15.4s\n" "ldr q20, [c_ptr1]\n" - "fmul v17.4s, v17.4s, v15.4s\n" "ldr q21, [c_ptr1, #0x10]\n" - "fmul v18.4s, v18.4s, v15.4s\n" "ldr q22, [c_ptr1, #0x20]\n" - "fmul v19.4s, v19.4s, v15.4s\n" "ldr q23, [c_ptr1, #0x30]\n" - "fmul v20.4s, v20.4s, v15.4s\n" "ldr q24, [c_ptr2]\n" - "fmul v21.4s, v21.4s, v15.4s\n" "ldr q25, [c_ptr2, #0x10]\n" - "fmul v22.4s, v22.4s, v15.4s\n" "ldr q26, [c_ptr2, #0x20]\n" - "fmul v23.4s, v23.4s, v15.4s\n" "ldr q27, [c_ptr2, #0x30]\n" - "fmul v24.4s, v24.4s, v15.4s\n" "ldr q28, [c_ptr3]\n" - "fmul v25.4s, v25.4s, v15.4s\n" "ldr q29, [c_ptr3, #0x10]\n" - "fmul v26.4s, v26.4s, v15.4s\n" "ldr q30, [c_ptr3, #0x20]\n" - "fmul v27.4s, v27.4s, v15.4s\n" "ldr q31, [c_ptr3, #0x30]\n" - "fmul v28.4s, v28.4s, v15.4s\n" "ldr q0, [%[a_ptr0]]\n" - "fmul v29.4s, v29.4s, v15.4s\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q1, [a_ptr1]\n" - "fmul v30.4s, v30.4s, v15.4s\n" + "add a_ptr1, a_ptr1, #0x10\n" "ldr q2, [a_ptr2]\n" - "fmul v31.4s, v31.4s, v15.4s\n" + "add a_ptr2, a_ptr2, #0x10\n" "ldr q3, [a_ptr3]\n" + "add a_ptr3, a_ptr3, #0x10\n" "ldr q8, [%[b_ptr0]]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" "ldr q9, [%[b_ptr0], #0x10]\n" - "add a_ptr1, a_ptr1, #0x10\n" "ldr q10, [%[b_ptr0], #0x20]\n" - "add a_ptr2, a_ptr2, #0x10\n" "ldr q11, [%[b_ptr0], #0x30]\n" - "add a_ptr3, a_ptr3, #0x10\n" "ldr q12, [%[b_ptr0], #0x40]\n" "ldr q13, [%[b_ptr0], #0x50]\n" "ldr q14, [%[b_ptr0], #0x60]\n" @@ -1681,16 +1712,50 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, "fmla v31.4s, v11.4s, v3.s[0]\n" "b.ne 7b\n" "6:\n" + "ld1r {v14.4s}, [%[minptr]]\n" + "ld1r {v15.4s}, [%[maxptr]]\n" + "fmax v16.4s, v16.4s, v14.4s\n" + "fmax v17.4s, v17.4s, v14.4s\n" + "fmax v18.4s, v18.4s, v14.4s\n" + "fmax v19.4s, v19.4s, v14.4s\n" + "fmin v16.4s, v16.4s, v15.4s\n" + "fmin v17.4s, v17.4s, v15.4s\n" + "fmin v18.4s, v18.4s, v15.4s\n" + "fmin v19.4s, v19.4s, v15.4s\n" "str q16, [%[c_ptr0]]\n" + "fmax v20.4s, v20.4s, v14.4s\n" + "fmax v21.4s, v21.4s, v14.4s\n" + "fmax v22.4s, v22.4s, v14.4s\n" "str q17, [%[c_ptr0], #0x10]\n" + "fmax v23.4s, v23.4s, v14.4s\n" + "fmin v20.4s, v20.4s, v15.4s\n" + "fmin v21.4s, v21.4s, v15.4s\n" "str q18, [%[c_ptr0], #0x20]\n" + "fmin v22.4s, v22.4s, v15.4s\n" + "fmin v23.4s, v23.4s, v15.4s\n" + "fmax v24.4s, v24.4s, v14.4s\n" "str q19, [%[c_ptr0], #0x30]\n" + "fmax v25.4s, v25.4s, v14.4s\n" "add %[c_ptr0], %[c_ptr0], #0x40\n" + "fmax v26.4s, v26.4s, v14.4s\n" "str q20, [c_ptr1]\n" + "fmin v24.4s, v24.4s, v15.4s\n" + "fmin v25.4s, v25.4s, v15.4s\n" + "fmax v27.4s, v27.4s, v14.4s\n" "str q21, [c_ptr1, #0x10]\n" + "fmin v26.4s, v26.4s, v15.4s\n" + "fmax v28.4s, v28.4s, v14.4s\n" + "fmax v29.4s, v29.4s, v14.4s\n" "str q22, [c_ptr1, #0x20]\n" + "fmin v27.4s, v27.4s, v15.4s\n" + "fmax v30.4s, v30.4s, v14.4s\n" + "fmin v28.4s, v28.4s, v15.4s\n" "str q23, [c_ptr1, #0x30]\n" + "fmin v29.4s, v29.4s, v15.4s\n" + "fmax v31.4s, v31.4s, v14.4s\n" + "fmin v30.4s, v30.4s, v15.4s\n" "str q24, [c_ptr2]\n" + "fmin v31.4s, v31.4s, v15.4s\n" "str q25, [c_ptr2, #0x10]\n" "str q26, [c_ptr2, #0x20]\n" "str q27, [c_ptr2, #0x30]\n" @@ -1705,7 +1770,7 @@ void a64_hybrid_fp32_mla_16x4(const float *A, int lda, const float *B, float *C, ".unreq c_ptr2\n" ".unreq c_ptr3\n" : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [blocks] "+r" (blocks) - : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [lda] "r" (ldab), [ldc] "r" (ldcb) + : [width] "r" (width), [append] "r" (static_cast<uint64_t>(append)), [lda] "r" (ldab), [ldc] "r" (ldcb), [biasptr] "r" (biasptr), [minptr] "r" (minptr), [maxptr] "r" (maxptr) : "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", "x0", "x1", "x2", "x3", "x4", "x5", "cc", "memory" ); break; |