diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_4VLx4/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_4VLx4/generic.cpp | 954 |
1 files changed, 528 insertions, 426 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_4VLx4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_4VLx4/generic.cpp index 1b2d7852a8..855d27a151 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_4VLx4/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_4VLx4/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 Arm Limited. + * Copyright (c) 2018-2019 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,37 +25,58 @@ #include <algorithm> +#include "arm_gemm.hpp" #include "../../asmlib.hpp" #include "../../utils.hpp" namespace arm_gemm { -void sve_hybrid_fp32_mla_4VLx4(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 sve_hybrid_fp32_mla_4VLx4(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 leftovers = K; + float nullbias[256]; + if (!append && !bias) { + memset(nullbias, 0, (4 * get_vector_length<float>() * 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); const unsigned long ldab = lda * sizeof(float); float *c_ptr0 = C + (y * ldc); - const unsigned long ldcb = ldc * sizeof(float); for (int x0=0; x0<N; x0+=(4 * get_vector_length<float>())) { const long width = std::min((unsigned long)N-x0, (4 * get_vector_length<float>())); - const float *betaptr = β long loops = loops_count; long regs = regs_count; long temp = 0; long blocks = leftovers; const float *a_ptr0 = a_ptr0_base; const float *b_ptr0 = B + (K_stride * x0); + const unsigned long ldcb = ldc * sizeof(float); + const float *biasptr = bias ? bias+x0 : nullbias; switch(M-y) { case 1: @@ -69,269 +90,283 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "whilelt p2.s, %[temp], %[width]\n" "incw %[temp], all, mul #1\n" "whilelt p3.s, %[temp], %[width]\n" - "cbz %[beta0], 1f\n" - "mov z16.s, #0\n" + "cbnz %[append], 1f\n" + "ld1w z16.s, p0/z, [%[biasptr]]\n" + "ld1w z17.s, p1/z, [%[biasptr], #1, MUL VL]\n" + "ld1w z18.s, p2/z, [%[biasptr], #2, MUL VL]\n" + "ld1w z19.s, p3/z, [%[biasptr], #3, MUL VL]\n" "ld1rqw z0.s, p7/z, [%[a_ptr0]]\n" - "mov z17.s, #0\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "mov z18.s, #0\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "mov z19.s, #0\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" - "b 2f\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" + "cbz %[loops], 2f\n" + "b 3f\n" "1:\n" - "ld1rw z15.s, p7/z, [%[betaptr]]\n" "ld1w z16.s, p0/z, [%[c_ptr0]]\n" "ld1w z17.s, p1/z, [%[c_ptr0], #1, MUL VL]\n" "ld1w z18.s, p2/z, [%[c_ptr0], #2, MUL VL]\n" "ld1w z19.s, p3/z, [%[c_ptr0], #3, MUL VL]\n" "ld1rqw z0.s, p7/z, [%[a_ptr0]]\n" "add %[a_ptr0], %[a_ptr0], #0x10\n" - "fmul z16.s, p7/m, z16.s, z15.s\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "fmul z17.s, p7/m, z17.s, z15.s\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "fmul z18.s, p7/m, z18.s, z15.s\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "fmul z19.s, p7/m, z19.s, z15.s\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "2:\n" "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "addvl %[b_ptr0], %[b_ptr0], #8\n" - "cbz %[loops], 3f\n" - "4:\n" + "cbz %[loops], 2f\n" + "3:\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z17.s, z9.s, z0.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" "fmla z18.s, z10.s, z0.s[0]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z19.s, z11.s, z0.s[0]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "add %[a_ptr0], %[a_ptr0], #0x20\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z17.s, z13.s, z0.s[1]\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z18.s, z14.s, z0.s[1]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "fmla z19.s, z15.s, z0.s[1]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" - "subs %[loops], %[loops], #0x1\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z17.s, z9.s, z0.s[2]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z18.s, z10.s, z0.s[2]\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "subs %[loops], %[loops], #0x1\n" "fmla z19.s, z11.s, z0.s[2]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" "fmla z16.s, z12.s, z0.s[3]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "add %[a_ptr0], %[a_ptr0], #0x20\n" "fmla z17.s, z13.s, z0.s[3]\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" "fmla z18.s, z14.s, z0.s[3]\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" "fmla z19.s, z15.s, z0.s[3]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z4.s[0]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" - "ld1rqw z0.s, p7/z, [%[a_ptr0], #-0x10]\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" "fmla z17.s, z9.s, z4.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" "fmla z18.s, z10.s, z4.s[0]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" "fmla z19.s, z11.s, z4.s[0]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z16.s, z12.s, z4.s[1]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" + "ld1rqw z0.s, p7/z, [%[a_ptr0], #-0x10]\n" "fmla z17.s, z13.s, z4.s[1]\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z18.s, z14.s, z4.s[1]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z19.s, z15.s, z4.s[1]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z16.s, z8.s, z4.s[2]\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "fmla z17.s, z9.s, z4.s[2]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "fmla z18.s, z10.s, z4.s[2]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z19.s, z11.s, z4.s[2]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z16.s, z12.s, z4.s[3]\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" + "fmla z17.s, z13.s, z4.s[3]\n" + "fmla z18.s, z14.s, z4.s[3]\n" + "fmla z19.s, z15.s, z4.s[3]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "fmla z17.s, z13.s, z4.s[3]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "fmla z18.s, z14.s, z4.s[3]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" - "fmla z19.s, z15.s, z4.s[3]\n" "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" - "b.ne 4b\n" - "3:\n" - "cbz %[regs], 5f\n" + "b.ne 3b\n" + "2:\n" + "cbz %[regs], 4f\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z17.s, z9.s, z0.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" "fmla z18.s, z10.s, z0.s[0]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z19.s, z11.s, z0.s[0]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z17.s, z13.s, z0.s[1]\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z18.s, z14.s, z0.s[1]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "fmla z19.s, z15.s, z0.s[1]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z17.s, z9.s, z0.s[2]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z18.s, z10.s, z0.s[2]\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" "fmla z19.s, z11.s, z0.s[2]\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" "fmla z16.s, z12.s, z0.s[3]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" "fmla z17.s, z13.s, z0.s[3]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" "fmla z18.s, z14.s, z0.s[3]\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" "fmla z19.s, z15.s, z0.s[3]\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z4.s[0]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" "fmla z17.s, z9.s, z4.s[0]\n" - "ld1rqw z0.s, p6/z, [%[a_ptr0], #0x10]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" "fmla z18.s, z10.s, z4.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" "fmla z19.s, z11.s, z4.s[0]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z16.s, z12.s, z4.s[1]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z17.s, z13.s, z4.s[1]\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z18.s, z14.s, z4.s[1]\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z19.s, z15.s, z4.s[1]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z16.s, z8.s, z4.s[2]\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "fmla z17.s, z9.s, z4.s[2]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "fmla z18.s, z10.s, z4.s[2]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z19.s, z11.s, z4.s[2]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z16.s, z12.s, z4.s[3]\n" + "ld1rqw z0.s, p6/z, [%[a_ptr0], #0x10]\n" "fmla z17.s, z13.s, z4.s[3]\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" "fmla z18.s, z14.s, z4.s[3]\n" + "addvl %[a_ptr0], %[a_ptr0], #2\n" "fmla z19.s, z15.s, z4.s[3]\n" - "cbz %[blocks], 6f\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "cbz %[blocks], 5f\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z16.s, z8.s, z0.s[0]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z17.s, z9.s, z0.s[0]\n" "fmla z18.s, z10.s, z0.s[0]\n" "fmla z19.s, z11.s, z0.s[0]\n" - "b.eq 6f\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "b.eq 5f\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z17.s, z13.s, z0.s[1]\n" "fmla z18.s, z14.s, z0.s[1]\n" "fmla z19.s, z15.s, z0.s[1]\n" - "b.eq 6f\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "b.eq 5f\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" "fmla z17.s, z9.s, z0.s[2]\n" "fmla z18.s, z10.s, z0.s[2]\n" "fmla z19.s, z11.s, z0.s[2]\n" - "b 6f\n" - "5:\n" + "b 5f\n" + "4:\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p6/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z17.s, z9.s, z0.s[0]\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z18.s, z10.s, z0.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z19.s, z11.s, z0.s[0]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z17.s, z13.s, z0.s[1]\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "fmla z18.s, z14.s, z0.s[1]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "fmla z19.s, z15.s, z0.s[1]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z17.s, z9.s, z0.s[2]\n" + "ld1rqw z4.s, p6/z, [%[a_ptr0]]\n" "fmla z18.s, z10.s, z0.s[2]\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" "fmla z19.s, z11.s, z0.s[2]\n" + "addvl %[a_ptr0], %[a_ptr0], #1\n" "fmla z16.s, z12.s, z0.s[3]\n" "fmla z17.s, z13.s, z0.s[3]\n" "fmla z18.s, z14.s, z0.s[3]\n" "fmla z19.s, z15.s, z0.s[3]\n" - "cbz %[blocks], 6f\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "cbz %[blocks], 5f\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z16.s, z8.s, z4.s[0]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z17.s, z9.s, z4.s[0]\n" "fmla z18.s, z10.s, z4.s[0]\n" "fmla z19.s, z11.s, z4.s[0]\n" - "b.eq 6f\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "b.eq 5f\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z16.s, z12.s, z4.s[1]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z17.s, z13.s, z4.s[1]\n" "fmla z18.s, z14.s, z4.s[1]\n" "fmla z19.s, z15.s, z4.s[1]\n" - "b.eq 6f\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "b.eq 5f\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z4.s[2]\n" "fmla z17.s, z9.s, z4.s[2]\n" "fmla z18.s, z10.s, z4.s[2]\n" "fmla z19.s, z11.s, z4.s[2]\n" - "6:\n" + "5:\n" + "ld1rw z14.s, p7/z, [%[minptr]]\n" + "ld1rw z15.s, p7/z, [%[maxptr]]\n" + "fmax z16.s, p7/m, z16.s, z14.s\n" + "fmax z17.s, p7/m, z17.s, z14.s\n" + "fmax z18.s, p7/m, z18.s, z14.s\n" + "fmax z19.s, p7/m, z19.s, z14.s\n" + "fmin z16.s, p7/m, z16.s, z15.s\n" + "fmin z17.s, p7/m, z17.s, z15.s\n" + "fmin z18.s, p7/m, z18.s, z15.s\n" + "fmin z19.s, p7/m, z19.s, z15.s\n" "st1w z16.s, p0, [%[c_ptr0]]\n" "st1w z17.s, p1, [%[c_ptr0], #1, MUL VL]\n" "st1w z18.s, p2, [%[c_ptr0], #2, MUL VL]\n" "st1w z19.s, p3, [%[c_ptr0], #3, MUL VL]\n" "addvl %[c_ptr0], %[c_ptr0], #4\n" : [a_ptr0] "+r" (a_ptr0), [b_ptr0] "+r" (b_ptr0), [c_ptr0] "+r" (c_ptr0), [loops] "+r" (loops), [regs] "+r" (regs), [temp] "+r" (temp), [blocks] "+r" (blocks) - : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [leftovers] "r" (leftovers), [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), [leftovers] "r" (leftovers) : "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31", "cc", "memory" ); break; @@ -350,83 +385,74 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "whilelt p2.s, %[temp], %[width]\n" "incw %[temp], all, mul #1\n" "whilelt p3.s, %[temp], %[width]\n" - "cbz %[beta0], 1f\n" - "mov z16.s, #0\n" + "cbnz %[append], 1f\n" + "ld1w z16.s, p0/z, [%[biasptr]]\n" + "ld1w z17.s, p1/z, [%[biasptr], #1, MUL VL]\n" + "ld1w z18.s, p2/z, [%[biasptr], #2, MUL VL]\n" + "ld1w z19.s, p3/z, [%[biasptr], #3, MUL VL]\n" + "mov z20.d, z16.d\n" "ld1rqw z0.s, p7/z, [%[a_ptr0]]\n" - "mov z17.s, #0\n" + "mov z21.d, z17.d\n" "ld1rqw z1.s, p7/z, [a_ptr1]\n" - "mov z18.s, #0\n" + "mov z22.d, z18.d\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "mov z19.s, #0\n" + "mov z23.d, z19.d\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "mov z20.s, #0\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "mov z21.s, #0\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "mov z22.s, #0\n" + "add a_ptr1, a_ptr1, #0x10\n" "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" - "mov z23.s, #0\n" "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" - "add a_ptr1, a_ptr1, #0x10\n" - "b 2f\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" + "cbz %[loops], 2f\n" + "b 3f\n" "1:\n" - "ld1rw z15.s, p7/z, [%[betaptr]]\n" "ld1w z16.s, p0/z, [%[c_ptr0]]\n" "ld1w z17.s, p1/z, [%[c_ptr0], #1, MUL VL]\n" "ld1w z18.s, p2/z, [%[c_ptr0], #2, MUL VL]\n" "ld1w z19.s, p3/z, [%[c_ptr0], #3, MUL VL]\n" "ld1w z20.s, p0/z, [c_ptr1]\n" - "fmul z16.s, p7/m, z16.s, z15.s\n" "ld1w z21.s, p1/z, [c_ptr1, #1, MUL VL]\n" - "fmul z17.s, p7/m, z17.s, z15.s\n" "ld1w z22.s, p2/z, [c_ptr1, #2, MUL VL]\n" - "fmul z18.s, p7/m, z18.s, z15.s\n" "ld1w z23.s, p3/z, [c_ptr1, #3, MUL VL]\n" - "fmul z19.s, p7/m, z19.s, z15.s\n" "ld1rqw z0.s, p7/z, [%[a_ptr0]]\n" - "fmul z20.s, p7/m, z20.s, z15.s\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ld1rqw z1.s, p7/z, [a_ptr1]\n" - "fmul z21.s, p7/m, z21.s, z15.s\n" + "add a_ptr1, a_ptr1, #0x10\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "fmul z22.s, p7/m, z22.s, z15.s\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "fmul z23.s, p7/m, z23.s, z15.s\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" - "add a_ptr1, a_ptr1, #0x10\n" "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" - "2:\n" "addvl %[b_ptr0], %[b_ptr0], #8\n" - "cbz %[loops], 3f\n" - "4:\n" + "cbz %[loops], 2f\n" + "3:\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" - "ld1rqw z5.s, p7/z, [a_ptr1]\n" + "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" "fmla z17.s, z9.s, z0.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1rqw z5.s, p7/z, [a_ptr1]\n" "fmla z21.s, z9.s, z1.s[0]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z18.s, z10.s, z0.s[0]\n" - "add %[a_ptr0], %[a_ptr0], #0x20\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z22.s, z10.s, z1.s[0]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z19.s, z11.s, z0.s[0]\n" - "add a_ptr1, a_ptr1, #0x20\n" + "subs %[loops], %[loops], #0x1\n" "fmla z23.s, z11.s, z1.s[0]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" - "subs %[loops], %[loops], #0x1\n" + "add %[a_ptr0], %[a_ptr0], #0x20\n" "fmla z20.s, z12.s, z1.s[1]\n" "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "fmla z17.s, z13.s, z0.s[1]\n" + "add a_ptr1, a_ptr1, #0x20\n" "fmla z21.s, z13.s, z1.s[1]\n" "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "fmla z18.s, z14.s, z0.s[1]\n" @@ -440,12 +466,12 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z1.s[2]\n" "fmla z17.s, z9.s, z0.s[2]\n" "fmla z21.s, z9.s, z1.s[2]\n" - "fmla z18.s, z10.s, z0.s[2]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "fmla z22.s, z10.s, z1.s[2]\n" + "fmla z18.s, z10.s, z0.s[2]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "fmla z19.s, z11.s, z0.s[2]\n" + "fmla z22.s, z10.s, z1.s[2]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "fmla z19.s, z11.s, z0.s[2]\n" "fmla z23.s, z11.s, z1.s[2]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z12.s, z0.s[3]\n" @@ -491,12 +517,12 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z5.s[2]\n" "fmla z17.s, z9.s, z4.s[2]\n" "fmla z21.s, z9.s, z5.s[2]\n" - "fmla z18.s, z10.s, z4.s[2]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "fmla z22.s, z10.s, z5.s[2]\n" + "fmla z18.s, z10.s, z4.s[2]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "fmla z19.s, z11.s, z4.s[2]\n" + "fmla z22.s, z10.s, z5.s[2]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "fmla z19.s, z11.s, z4.s[2]\n" "fmla z23.s, z11.s, z5.s[2]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z12.s, z4.s[3]\n" @@ -510,19 +536,19 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" "fmla z19.s, z15.s, z4.s[3]\n" "fmla z23.s, z15.s, z5.s[3]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" - "b.ne 4b\n" - "3:\n" - "cbz %[regs], 5f\n" + "b.ne 3b\n" + "2:\n" + "cbz %[regs], 4f\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" - "ld1rqw z5.s, p7/z, [a_ptr1]\n" + "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" "fmla z17.s, z9.s, z0.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1rqw z5.s, p7/z, [a_ptr1]\n" "fmla z21.s, z9.s, z1.s[0]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z18.s, z10.s, z0.s[0]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z22.s, z10.s, z1.s[0]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z19.s, z11.s, z0.s[0]\n" @@ -545,12 +571,12 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z1.s[2]\n" "fmla z17.s, z9.s, z0.s[2]\n" "fmla z21.s, z9.s, z1.s[2]\n" - "fmla z18.s, z10.s, z0.s[2]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "fmla z22.s, z10.s, z1.s[2]\n" + "fmla z18.s, z10.s, z0.s[2]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "fmla z19.s, z11.s, z0.s[2]\n" + "fmla z22.s, z10.s, z1.s[2]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "fmla z19.s, z11.s, z0.s[2]\n" "fmla z23.s, z11.s, z1.s[2]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z12.s, z0.s[3]\n" @@ -571,9 +597,11 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z5.s[0]\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z17.s, z9.s, z4.s[0]\n" + "addvl %[a_ptr0], %[a_ptr0], #2\n" "fmla z21.s, z9.s, z5.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z18.s, z10.s, z4.s[0]\n" + "addvl a_ptr1, a_ptr1, #2\n" "fmla z22.s, z10.s, z5.s[0]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z19.s, z11.s, z4.s[0]\n" @@ -592,6 +620,7 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z15.s, z5.s[1]\n" "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z16.s, z8.s, z4.s[2]\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" "fmla z20.s, z8.s, z5.s[2]\n" "fmla z17.s, z9.s, z4.s[2]\n" "fmla z21.s, z9.s, z5.s[2]\n" @@ -607,14 +636,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z22.s, z14.s, z5.s[3]\n" "fmla z19.s, z15.s, z4.s[3]\n" "fmla z23.s, z15.s, z5.s[3]\n" - "cbz %[blocks], 6f\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "cbz %[blocks], 5f\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z16.s, z8.s, z0.s[0]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" "fmla z17.s, z9.s, z0.s[0]\n" "fmla z21.s, z9.s, z1.s[0]\n" @@ -622,13 +650,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z22.s, z10.s, z1.s[0]\n" "fmla z19.s, z11.s, z0.s[0]\n" "fmla z23.s, z11.s, z1.s[0]\n" - "b.eq 6f\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "b.eq 5f\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z20.s, z12.s, z1.s[1]\n" "fmla z17.s, z13.s, z0.s[1]\n" "fmla z21.s, z13.s, z1.s[1]\n" @@ -636,11 +664,12 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z22.s, z14.s, z1.s[1]\n" "fmla z19.s, z15.s, z0.s[1]\n" "fmla z23.s, z15.s, z1.s[1]\n" - "b.eq 6f\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "b.eq 5f\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" "fmla z20.s, z8.s, z1.s[2]\n" "fmla z17.s, z9.s, z0.s[2]\n" @@ -649,23 +678,26 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z22.s, z10.s, z1.s[2]\n" "fmla z19.s, z11.s, z0.s[2]\n" "fmla z23.s, z11.s, z1.s[2]\n" - "b 6f\n" - "5:\n" + "b 5f\n" + "4:\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p6/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" - "ld1rqw z5.s, p6/z, [a_ptr1]\n" - "fmla z17.s, z9.s, z0.s[0]\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "fmla z17.s, z9.s, z0.s[0]\n" + "ld1rqw z4.s, p6/z, [%[a_ptr0]]\n" "fmla z21.s, z9.s, z1.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z18.s, z10.s, z0.s[0]\n" + "ld1rqw z5.s, p6/z, [a_ptr1]\n" "fmla z22.s, z10.s, z1.s[0]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z19.s, z11.s, z0.s[0]\n" + "addvl %[a_ptr0], %[a_ptr0], #1\n" "fmla z23.s, z11.s, z1.s[0]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" + "addvl a_ptr1, a_ptr1, #1\n" "fmla z20.s, z12.s, z1.s[1]\n" "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "fmla z17.s, z13.s, z0.s[1]\n" @@ -678,6 +710,7 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z15.s, z1.s[1]\n" "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" "fmla z20.s, z8.s, z1.s[2]\n" "fmla z17.s, z9.s, z0.s[2]\n" "fmla z21.s, z9.s, z1.s[2]\n" @@ -693,14 +726,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z22.s, z14.s, z1.s[3]\n" "fmla z19.s, z15.s, z0.s[3]\n" "fmla z23.s, z15.s, z1.s[3]\n" - "cbz %[blocks], 6f\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "cbz %[blocks], 5f\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z16.s, z8.s, z4.s[0]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z20.s, z8.s, z5.s[0]\n" "fmla z17.s, z9.s, z4.s[0]\n" "fmla z21.s, z9.s, z5.s[0]\n" @@ -708,13 +740,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z22.s, z10.s, z5.s[0]\n" "fmla z19.s, z11.s, z4.s[0]\n" "fmla z23.s, z11.s, z5.s[0]\n" - "b.eq 6f\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "b.eq 5f\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z16.s, z12.s, z4.s[1]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z20.s, z12.s, z5.s[1]\n" "fmla z17.s, z13.s, z4.s[1]\n" "fmla z21.s, z13.s, z5.s[1]\n" @@ -722,11 +754,12 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z22.s, z14.s, z5.s[1]\n" "fmla z19.s, z15.s, z4.s[1]\n" "fmla z23.s, z15.s, z5.s[1]\n" - "b.eq 6f\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "b.eq 5f\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z4.s[2]\n" "fmla z20.s, z8.s, z5.s[2]\n" "fmla z17.s, z9.s, z4.s[2]\n" @@ -735,10 +768,28 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z22.s, z10.s, z5.s[2]\n" "fmla z19.s, z11.s, z4.s[2]\n" "fmla z23.s, z11.s, z5.s[2]\n" - "6:\n" + "5:\n" + "ld1rw z14.s, p7/z, [%[minptr]]\n" + "ld1rw z15.s, p7/z, [%[maxptr]]\n" + "fmax z16.s, p7/m, z16.s, z14.s\n" + "fmax z17.s, p7/m, z17.s, z14.s\n" + "fmax z18.s, p7/m, z18.s, z14.s\n" + "fmax z19.s, p7/m, z19.s, z14.s\n" + "fmin z16.s, p7/m, z16.s, z15.s\n" + "fmin z17.s, p7/m, z17.s, z15.s\n" + "fmin z18.s, p7/m, z18.s, z15.s\n" + "fmin z19.s, p7/m, z19.s, z15.s\n" "st1w z16.s, p0, [%[c_ptr0]]\n" + "fmax z20.s, p7/m, z20.s, z14.s\n" + "fmax z21.s, p7/m, z21.s, z14.s\n" + "fmax z22.s, p7/m, z22.s, z14.s\n" "st1w z17.s, p1, [%[c_ptr0], #1, MUL VL]\n" + "fmax z23.s, p7/m, z23.s, z14.s\n" + "fmin z20.s, p7/m, z20.s, z15.s\n" + "fmin z21.s, p7/m, z21.s, z15.s\n" "st1w z18.s, p2, [%[c_ptr0], #2, MUL VL]\n" + "fmin z22.s, p7/m, z22.s, z15.s\n" + "fmin z23.s, p7/m, z23.s, z15.s\n" "st1w z19.s, p3, [%[c_ptr0], #3, MUL VL]\n" "addvl %[c_ptr0], %[c_ptr0], #4\n" "st1w z20.s, p0, [c_ptr1]\n" @@ -748,7 +799,7 @@ void sve_hybrid_fp32_mla_4VLx4(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), [temp] "+r" (temp), [blocks] "+r" (blocks) - : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [leftovers] "r" (leftovers), [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), [leftovers] "r" (leftovers) : "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31", "x0", "x1", "cc", "memory" ); break; @@ -760,110 +811,97 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "c_ptr2 .req X3\n" "add a_ptr1, %[a_ptr0], %[lda]\n" "add c_ptr1, %[c_ptr0], %[ldc]\n" + "add a_ptr2, a_ptr1, %[lda]\n" + "add c_ptr2, c_ptr1, %[ldc]\n" "whilelt p6.s, %[temp], %[leftovers]\n" "whilelt p0.s, %[temp], %[width]\n" "incw %[temp], all, mul #1\n" - "add a_ptr2, a_ptr1, %[lda]\n" - "add c_ptr2, c_ptr1, %[ldc]\n" "ptrue p7.s\n" "whilelt p1.s, %[temp], %[width]\n" "incw %[temp], all, mul #1\n" "whilelt p2.s, %[temp], %[width]\n" "incw %[temp], all, mul #1\n" "whilelt p3.s, %[temp], %[width]\n" - "cbz %[beta0], 1f\n" - "mov z16.s, #0\n" + "cbnz %[append], 1f\n" + "ld1w z16.s, p0/z, [%[biasptr]]\n" + "ld1w z17.s, p1/z, [%[biasptr], #1, MUL VL]\n" + "ld1w z18.s, p2/z, [%[biasptr], #2, MUL VL]\n" + "ld1w z19.s, p3/z, [%[biasptr], #3, MUL VL]\n" + "mov z20.d, z16.d\n" "ld1rqw z0.s, p7/z, [%[a_ptr0]]\n" - "mov z17.s, #0\n" + "mov z21.d, z17.d\n" "ld1rqw z1.s, p7/z, [a_ptr1]\n" - "mov z18.s, #0\n" + "mov z22.d, z18.d\n" "ld1rqw z2.s, p7/z, [a_ptr2]\n" - "mov z19.s, #0\n" + "mov z23.d, z19.d\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "mov z20.s, #0\n" + "mov z24.d, z16.d\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "mov z21.s, #0\n" + "mov z25.d, z17.d\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "mov z22.s, #0\n" + "mov z26.d, z18.d\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "mov z23.s, #0\n" + "mov z27.d, z19.d\n" "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" - "mov z24.s, #0\n" "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" - "mov z25.s, #0\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" - "mov z26.s, #0\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" - "mov z27.s, #0\n" "add %[a_ptr0], %[a_ptr0], #0x10\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "add a_ptr1, a_ptr1, #0x10\n" "add a_ptr2, a_ptr2, #0x10\n" - "b 2f\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" + "cbz %[loops], 2f\n" + "b 3f\n" "1:\n" - "ld1rw z15.s, p7/z, [%[betaptr]]\n" "ld1w z16.s, p0/z, [%[c_ptr0]]\n" "ld1w z17.s, p1/z, [%[c_ptr0], #1, MUL VL]\n" "ld1w z18.s, p2/z, [%[c_ptr0], #2, MUL VL]\n" "ld1w z19.s, p3/z, [%[c_ptr0], #3, MUL VL]\n" "ld1w z20.s, p0/z, [c_ptr1]\n" - "fmul z16.s, p7/m, z16.s, z15.s\n" "ld1w z21.s, p1/z, [c_ptr1, #1, MUL VL]\n" - "fmul z17.s, p7/m, z17.s, z15.s\n" "ld1w z22.s, p2/z, [c_ptr1, #2, MUL VL]\n" - "fmul z18.s, p7/m, z18.s, z15.s\n" "ld1w z23.s, p3/z, [c_ptr1, #3, MUL VL]\n" - "fmul z19.s, p7/m, z19.s, z15.s\n" "ld1w z24.s, p0/z, [c_ptr2]\n" - "fmul z20.s, p7/m, z20.s, z15.s\n" "ld1w z25.s, p1/z, [c_ptr2, #1, MUL VL]\n" - "fmul z21.s, p7/m, z21.s, z15.s\n" "ld1w z26.s, p2/z, [c_ptr2, #2, MUL VL]\n" - "fmul z22.s, p7/m, z22.s, z15.s\n" "ld1w z27.s, p3/z, [c_ptr2, #3, MUL VL]\n" - "fmul z23.s, p7/m, z23.s, z15.s\n" "ld1rqw z0.s, p7/z, [%[a_ptr0]]\n" - "fmul z24.s, p7/m, z24.s, z15.s\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ld1rqw z1.s, p7/z, [a_ptr1]\n" - "fmul z25.s, p7/m, z25.s, z15.s\n" + "add a_ptr1, a_ptr1, #0x10\n" "ld1rqw z2.s, p7/z, [a_ptr2]\n" - "fmul z26.s, p7/m, z26.s, z15.s\n" + "add a_ptr2, a_ptr2, #0x10\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "fmul z27.s, p7/m, z27.s, z15.s\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "add a_ptr1, a_ptr1, #0x10\n" "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" - "add a_ptr2, a_ptr2, #0x10\n" "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" - "2:\n" "addvl %[b_ptr0], %[b_ptr0], #8\n" - "cbz %[loops], 3f\n" - "4:\n" + "cbz %[loops], 2f\n" + "3:\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" - "ld1rqw z5.s, p7/z, [a_ptr1]\n" + "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" "fmla z24.s, z8.s, z2.s[0]\n" - "ld1rqw z6.s, p7/z, [a_ptr2]\n" + "ld1rqw z5.s, p7/z, [a_ptr1]\n" "fmla z17.s, z9.s, z0.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1rqw z6.s, p7/z, [a_ptr2]\n" "fmla z21.s, z9.s, z1.s[0]\n" - "add %[a_ptr0], %[a_ptr0], #0x20\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z25.s, z9.s, z2.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z18.s, z10.s, z0.s[0]\n" - "add a_ptr1, a_ptr1, #0x20\n" + "subs %[loops], %[loops], #0x1\n" "fmla z22.s, z10.s, z1.s[0]\n" - "add a_ptr2, a_ptr2, #0x20\n" + "add %[a_ptr0], %[a_ptr0], #0x20\n" "fmla z26.s, z10.s, z2.s[0]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z19.s, z11.s, z0.s[0]\n" - "subs %[loops], %[loops], #0x1\n" + "add a_ptr1, a_ptr1, #0x20\n" "fmla z23.s, z11.s, z1.s[0]\n" + "add a_ptr2, a_ptr2, #0x20\n" "fmla z27.s, z11.s, z2.s[0]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" @@ -887,8 +925,8 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z1.s[2]\n" "fmla z24.s, z8.s, z2.s[2]\n" "fmla z17.s, z9.s, z0.s[2]\n" - "fmla z21.s, z9.s, z1.s[2]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "fmla z21.s, z9.s, z1.s[2]\n" "fmla z25.s, z9.s, z2.s[2]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" "fmla z18.s, z10.s, z0.s[2]\n" @@ -955,8 +993,8 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z5.s[2]\n" "fmla z24.s, z8.s, z6.s[2]\n" "fmla z17.s, z9.s, z4.s[2]\n" - "fmla z21.s, z9.s, z5.s[2]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "fmla z21.s, z9.s, z5.s[2]\n" "fmla z25.s, z9.s, z6.s[2]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" "fmla z18.s, z10.s, z4.s[2]\n" @@ -982,19 +1020,19 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z19.s, z15.s, z4.s[3]\n" "fmla z23.s, z15.s, z5.s[3]\n" "fmla z27.s, z15.s, z6.s[3]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" - "b.ne 4b\n" - "3:\n" - "cbz %[regs], 5f\n" + "b.ne 3b\n" + "2:\n" + "cbz %[regs], 4f\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" - "ld1rqw z5.s, p7/z, [a_ptr1]\n" + "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" "fmla z24.s, z8.s, z2.s[0]\n" - "ld1rqw z6.s, p7/z, [a_ptr2]\n" + "ld1rqw z5.s, p7/z, [a_ptr1]\n" "fmla z17.s, z9.s, z0.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1rqw z6.s, p7/z, [a_ptr2]\n" "fmla z21.s, z9.s, z1.s[0]\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z25.s, z9.s, z2.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z18.s, z10.s, z0.s[0]\n" @@ -1026,8 +1064,8 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z1.s[2]\n" "fmla z24.s, z8.s, z2.s[2]\n" "fmla z17.s, z9.s, z0.s[2]\n" - "fmla z21.s, z9.s, z1.s[2]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "fmla z21.s, z9.s, z1.s[2]\n" "fmla z25.s, z9.s, z2.s[2]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" "fmla z18.s, z10.s, z0.s[2]\n" @@ -1059,10 +1097,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z16.s, z8.s, z4.s[0]\n" "ld1rqw z2.s, p6/z, [a_ptr2, #0x10]\n" "fmla z20.s, z8.s, z5.s[0]\n" + "addvl %[a_ptr0], %[a_ptr0], #2\n" "fmla z24.s, z8.s, z6.s[0]\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z17.s, z9.s, z4.s[0]\n" + "addvl a_ptr1, a_ptr1, #2\n" "fmla z21.s, z9.s, z5.s[0]\n" + "addvl a_ptr2, a_ptr2, #2\n" "fmla z25.s, z9.s, z6.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z18.s, z10.s, z4.s[0]\n" @@ -1090,6 +1131,7 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z27.s, z15.s, z6.s[1]\n" "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z16.s, z8.s, z4.s[2]\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" "fmla z20.s, z8.s, z5.s[2]\n" "fmla z24.s, z8.s, z6.s[2]\n" "fmla z17.s, z9.s, z4.s[2]\n" @@ -1113,14 +1155,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z19.s, z15.s, z4.s[3]\n" "fmla z23.s, z15.s, z5.s[3]\n" "fmla z27.s, z15.s, z6.s[3]\n" - "cbz %[blocks], 6f\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "cbz %[blocks], 5f\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z16.s, z8.s, z0.s[0]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" "fmla z24.s, z8.s, z2.s[0]\n" "fmla z17.s, z9.s, z0.s[0]\n" @@ -1132,13 +1173,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z19.s, z11.s, z0.s[0]\n" "fmla z23.s, z11.s, z1.s[0]\n" "fmla z27.s, z11.s, z2.s[0]\n" - "b.eq 6f\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "b.eq 5f\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z20.s, z12.s, z1.s[1]\n" "fmla z24.s, z12.s, z2.s[1]\n" "fmla z17.s, z13.s, z0.s[1]\n" @@ -1150,11 +1191,12 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z19.s, z15.s, z0.s[1]\n" "fmla z23.s, z15.s, z1.s[1]\n" "fmla z27.s, z15.s, z2.s[1]\n" - "b.eq 6f\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "b.eq 5f\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" "fmla z20.s, z8.s, z1.s[2]\n" "fmla z24.s, z8.s, z2.s[2]\n" @@ -1167,24 +1209,28 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z19.s, z11.s, z0.s[2]\n" "fmla z23.s, z11.s, z1.s[2]\n" "fmla z27.s, z11.s, z2.s[2]\n" - "b 6f\n" - "5:\n" + "b 5f\n" + "4:\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p6/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" - "ld1rqw z5.s, p6/z, [a_ptr1]\n" + "ld1rqw z4.s, p6/z, [%[a_ptr0]]\n" "fmla z24.s, z8.s, z2.s[0]\n" - "ld1rqw z6.s, p6/z, [a_ptr2]\n" - "fmla z17.s, z9.s, z0.s[0]\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "fmla z17.s, z9.s, z0.s[0]\n" + "ld1rqw z5.s, p6/z, [a_ptr1]\n" "fmla z21.s, z9.s, z1.s[0]\n" + "ld1rqw z6.s, p6/z, [a_ptr2]\n" "fmla z25.s, z9.s, z2.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z18.s, z10.s, z0.s[0]\n" + "addvl %[a_ptr0], %[a_ptr0], #1\n" "fmla z22.s, z10.s, z1.s[0]\n" + "addvl a_ptr1, a_ptr1, #1\n" "fmla z26.s, z10.s, z2.s[0]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z19.s, z11.s, z0.s[0]\n" + "addvl a_ptr2, a_ptr2, #1\n" "fmla z23.s, z11.s, z1.s[0]\n" "fmla z27.s, z11.s, z2.s[0]\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" @@ -1205,6 +1251,7 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z27.s, z15.s, z2.s[1]\n" "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" "fmla z20.s, z8.s, z1.s[2]\n" "fmla z24.s, z8.s, z2.s[2]\n" "fmla z17.s, z9.s, z0.s[2]\n" @@ -1228,14 +1275,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z19.s, z15.s, z0.s[3]\n" "fmla z23.s, z15.s, z1.s[3]\n" "fmla z27.s, z15.s, z2.s[3]\n" - "cbz %[blocks], 6f\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "cbz %[blocks], 5f\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z16.s, z8.s, z4.s[0]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z20.s, z8.s, z5.s[0]\n" "fmla z24.s, z8.s, z6.s[0]\n" "fmla z17.s, z9.s, z4.s[0]\n" @@ -1247,13 +1293,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z19.s, z11.s, z4.s[0]\n" "fmla z23.s, z11.s, z5.s[0]\n" "fmla z27.s, z11.s, z6.s[0]\n" - "b.eq 6f\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "b.eq 5f\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z16.s, z12.s, z4.s[1]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z20.s, z12.s, z5.s[1]\n" "fmla z24.s, z12.s, z6.s[1]\n" "fmla z17.s, z13.s, z4.s[1]\n" @@ -1265,11 +1311,12 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z19.s, z15.s, z4.s[1]\n" "fmla z23.s, z15.s, z5.s[1]\n" "fmla z27.s, z15.s, z6.s[1]\n" - "b.eq 6f\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "b.eq 5f\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z4.s[2]\n" "fmla z20.s, z8.s, z5.s[2]\n" "fmla z24.s, z8.s, z6.s[2]\n" @@ -1282,14 +1329,40 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z19.s, z11.s, z4.s[2]\n" "fmla z23.s, z11.s, z5.s[2]\n" "fmla z27.s, z11.s, z6.s[2]\n" - "6:\n" + "5:\n" + "ld1rw z14.s, p7/z, [%[minptr]]\n" + "ld1rw z15.s, p7/z, [%[maxptr]]\n" + "fmax z16.s, p7/m, z16.s, z14.s\n" + "fmax z17.s, p7/m, z17.s, z14.s\n" + "fmax z18.s, p7/m, z18.s, z14.s\n" + "fmax z19.s, p7/m, z19.s, z14.s\n" + "fmin z16.s, p7/m, z16.s, z15.s\n" + "fmin z17.s, p7/m, z17.s, z15.s\n" + "fmin z18.s, p7/m, z18.s, z15.s\n" + "fmin z19.s, p7/m, z19.s, z15.s\n" "st1w z16.s, p0, [%[c_ptr0]]\n" + "fmax z20.s, p7/m, z20.s, z14.s\n" + "fmax z21.s, p7/m, z21.s, z14.s\n" + "fmax z22.s, p7/m, z22.s, z14.s\n" "st1w z17.s, p1, [%[c_ptr0], #1, MUL VL]\n" + "fmax z23.s, p7/m, z23.s, z14.s\n" + "fmin z20.s, p7/m, z20.s, z15.s\n" + "fmin z21.s, p7/m, z21.s, z15.s\n" "st1w z18.s, p2, [%[c_ptr0], #2, MUL VL]\n" + "fmin z22.s, p7/m, z22.s, z15.s\n" + "fmin z23.s, p7/m, z23.s, z15.s\n" + "fmax z24.s, p7/m, z24.s, z14.s\n" "st1w z19.s, p3, [%[c_ptr0], #3, MUL VL]\n" + "fmax z25.s, p7/m, z25.s, z14.s\n" "addvl %[c_ptr0], %[c_ptr0], #4\n" + "fmax z26.s, p7/m, z26.s, z14.s\n" "st1w z20.s, p0, [c_ptr1]\n" + "fmin z24.s, p7/m, z24.s, z15.s\n" + "fmin z25.s, p7/m, z25.s, z15.s\n" + "fmax z27.s, p7/m, z27.s, z14.s\n" "st1w z21.s, p1, [c_ptr1, #1, MUL VL]\n" + "fmin z26.s, p7/m, z26.s, z15.s\n" + "fmin z27.s, p7/m, z27.s, z15.s\n" "st1w z22.s, p2, [c_ptr1, #2, MUL VL]\n" "st1w z23.s, p3, [c_ptr1, #3, MUL VL]\n" "st1w z24.s, p0, [c_ptr2]\n" @@ -1301,7 +1374,7 @@ void sve_hybrid_fp32_mla_4VLx4(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), [temp] "+r" (temp), [blocks] "+r" (blocks) - : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [leftovers] "r" (leftovers), [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), [leftovers] "r" (leftovers) : "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31", "x0", "x1", "x2", "x3", "cc", "memory" ); break; @@ -1316,132 +1389,115 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "c_ptr3 .req X5\n" "add a_ptr1, %[a_ptr0], %[lda]\n" "add c_ptr1, %[c_ptr0], %[ldc]\n" + "add a_ptr2, a_ptr1, %[lda]\n" + "add c_ptr2, c_ptr1, %[ldc]\n" + "add a_ptr3, a_ptr2, %[lda]\n" + "add c_ptr3, c_ptr2, %[ldc]\n" "whilelt p6.s, %[temp], %[leftovers]\n" "whilelt p0.s, %[temp], %[width]\n" "incw %[temp], all, mul #1\n" - "add a_ptr2, a_ptr1, %[lda]\n" - "add c_ptr2, c_ptr1, %[ldc]\n" "ptrue p7.s\n" "whilelt p1.s, %[temp], %[width]\n" - "add a_ptr3, a_ptr2, %[lda]\n" - "add c_ptr3, c_ptr2, %[ldc]\n" "incw %[temp], all, mul #1\n" "whilelt p2.s, %[temp], %[width]\n" "incw %[temp], all, mul #1\n" "whilelt p3.s, %[temp], %[width]\n" - "cbz %[beta0], 1f\n" - "mov z16.s, #0\n" + "cbnz %[append], 1f\n" + "ld1w z16.s, p0/z, [%[biasptr]]\n" + "ld1w z17.s, p1/z, [%[biasptr], #1, MUL VL]\n" + "ld1w z18.s, p2/z, [%[biasptr], #2, MUL VL]\n" + "ld1w z19.s, p3/z, [%[biasptr], #3, MUL VL]\n" + "mov z20.d, z16.d\n" "ld1rqw z0.s, p7/z, [%[a_ptr0]]\n" - "mov z17.s, #0\n" + "mov z21.d, z17.d\n" "ld1rqw z1.s, p7/z, [a_ptr1]\n" - "mov z18.s, #0\n" + "mov z22.d, z18.d\n" "ld1rqw z2.s, p7/z, [a_ptr2]\n" - "mov z19.s, #0\n" + "mov z23.d, z19.d\n" "ld1rqw z3.s, p7/z, [a_ptr3]\n" - "mov z20.s, #0\n" + "mov z24.d, z16.d\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "mov z21.s, #0\n" + "mov z25.d, z17.d\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "mov z22.s, #0\n" + "mov z26.d, z18.d\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "mov z23.s, #0\n" + "mov z27.d, z19.d\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "mov z24.s, #0\n" + "mov z28.d, z16.d\n" "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" - "mov z25.s, #0\n" + "mov z29.d, z17.d\n" "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" - "mov z26.s, #0\n" + "mov z30.d, z18.d\n" "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" - "mov z27.s, #0\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" - "mov z28.s, #0\n" + "mov z31.d, z19.d\n" "add %[a_ptr0], %[a_ptr0], #0x10\n" - "mov z29.s, #0\n" "add a_ptr1, a_ptr1, #0x10\n" - "mov z30.s, #0\n" "add a_ptr2, a_ptr2, #0x10\n" - "mov z31.s, #0\n" "add a_ptr3, a_ptr3, #0x10\n" - "b 2f\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" + "cbz %[loops], 2f\n" + "b 3f\n" "1:\n" - "ld1rw z15.s, p7/z, [%[betaptr]]\n" "ld1w z16.s, p0/z, [%[c_ptr0]]\n" "ld1w z17.s, p1/z, [%[c_ptr0], #1, MUL VL]\n" "ld1w z18.s, p2/z, [%[c_ptr0], #2, MUL VL]\n" "ld1w z19.s, p3/z, [%[c_ptr0], #3, MUL VL]\n" "ld1w z20.s, p0/z, [c_ptr1]\n" - "fmul z16.s, p7/m, z16.s, z15.s\n" "ld1w z21.s, p1/z, [c_ptr1, #1, MUL VL]\n" - "fmul z17.s, p7/m, z17.s, z15.s\n" "ld1w z22.s, p2/z, [c_ptr1, #2, MUL VL]\n" - "fmul z18.s, p7/m, z18.s, z15.s\n" "ld1w z23.s, p3/z, [c_ptr1, #3, MUL VL]\n" - "fmul z19.s, p7/m, z19.s, z15.s\n" "ld1w z24.s, p0/z, [c_ptr2]\n" - "fmul z20.s, p7/m, z20.s, z15.s\n" "ld1w z25.s, p1/z, [c_ptr2, #1, MUL VL]\n" - "fmul z21.s, p7/m, z21.s, z15.s\n" "ld1w z26.s, p2/z, [c_ptr2, #2, MUL VL]\n" - "fmul z22.s, p7/m, z22.s, z15.s\n" "ld1w z27.s, p3/z, [c_ptr2, #3, MUL VL]\n" - "fmul z23.s, p7/m, z23.s, z15.s\n" "ld1w z28.s, p0/z, [c_ptr3]\n" - "fmul z24.s, p7/m, z24.s, z15.s\n" "ld1w z29.s, p1/z, [c_ptr3, #1, MUL VL]\n" - "fmul z25.s, p7/m, z25.s, z15.s\n" "ld1w z30.s, p2/z, [c_ptr3, #2, MUL VL]\n" - "fmul z26.s, p7/m, z26.s, z15.s\n" "ld1w z31.s, p3/z, [c_ptr3, #3, MUL VL]\n" - "fmul z27.s, p7/m, z27.s, z15.s\n" "ld1rqw z0.s, p7/z, [%[a_ptr0]]\n" - "fmul z28.s, p7/m, z28.s, z15.s\n" + "add %[a_ptr0], %[a_ptr0], #0x10\n" "ld1rqw z1.s, p7/z, [a_ptr1]\n" - "fmul z29.s, p7/m, z29.s, z15.s\n" + "add a_ptr1, a_ptr1, #0x10\n" "ld1rqw z2.s, p7/z, [a_ptr2]\n" - "fmul z30.s, p7/m, z30.s, z15.s\n" + "add a_ptr2, a_ptr2, #0x10\n" "ld1rqw z3.s, p7/z, [a_ptr3]\n" - "fmul z31.s, p7/m, z31.s, z15.s\n" + "add a_ptr3, a_ptr3, #0x10\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "add %[a_ptr0], %[a_ptr0], #0x10\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "add a_ptr1, a_ptr1, #0x10\n" "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" - "add a_ptr2, a_ptr2, #0x10\n" "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" - "add a_ptr3, a_ptr3, #0x10\n" "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" - "2:\n" "addvl %[b_ptr0], %[b_ptr0], #8\n" - "cbz %[loops], 3f\n" - "4:\n" + "cbz %[loops], 2f\n" + "3:\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" - "ld1rqw z5.s, p7/z, [a_ptr1]\n" + "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" "fmla z24.s, z8.s, z2.s[0]\n" - "ld1rqw z6.s, p7/z, [a_ptr2]\n" + "ld1rqw z5.s, p7/z, [a_ptr1]\n" "fmla z28.s, z8.s, z3.s[0]\n" - "ld1rqw z7.s, p7/z, [a_ptr3]\n" + "ld1rqw z6.s, p7/z, [a_ptr2]\n" "fmla z17.s, z9.s, z0.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1rqw z7.s, p7/z, [a_ptr3]\n" "fmla z21.s, z9.s, z1.s[0]\n" - "add %[a_ptr0], %[a_ptr0], #0x20\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z25.s, z9.s, z2.s[0]\n" - "add a_ptr1, a_ptr1, #0x20\n" + "subs %[loops], %[loops], #0x1\n" "fmla z29.s, z9.s, z3.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z18.s, z10.s, z0.s[0]\n" - "add a_ptr2, a_ptr2, #0x20\n" + "add %[a_ptr0], %[a_ptr0], #0x20\n" "fmla z22.s, z10.s, z1.s[0]\n" - "add a_ptr3, a_ptr3, #0x20\n" + "add a_ptr1, a_ptr1, #0x20\n" "fmla z26.s, z10.s, z2.s[0]\n" - "subs %[loops], %[loops], #0x1\n" + "add a_ptr2, a_ptr2, #0x20\n" "fmla z30.s, z10.s, z3.s[0]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z19.s, z11.s, z0.s[0]\n" + "add a_ptr3, a_ptr3, #0x20\n" "fmla z23.s, z11.s, z1.s[0]\n" "fmla z27.s, z11.s, z2.s[0]\n" "fmla z31.s, z11.s, z3.s[0]\n" @@ -1471,8 +1527,8 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z1.s[2]\n" "fmla z24.s, z8.s, z2.s[2]\n" "fmla z28.s, z8.s, z3.s[2]\n" - "fmla z17.s, z9.s, z0.s[2]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "fmla z17.s, z9.s, z0.s[2]\n" "fmla z21.s, z9.s, z1.s[2]\n" "fmla z25.s, z9.s, z2.s[2]\n" "fmla z29.s, z9.s, z3.s[2]\n" @@ -1556,8 +1612,8 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z5.s[2]\n" "fmla z24.s, z8.s, z6.s[2]\n" "fmla z28.s, z8.s, z7.s[2]\n" - "fmla z17.s, z9.s, z4.s[2]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "fmla z17.s, z9.s, z4.s[2]\n" "fmla z21.s, z9.s, z5.s[2]\n" "fmla z25.s, z9.s, z6.s[2]\n" "fmla z29.s, z9.s, z7.s[2]\n" @@ -1591,21 +1647,21 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z15.s, z5.s[3]\n" "fmla z27.s, z15.s, z6.s[3]\n" "fmla z31.s, z15.s, z7.s[3]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" - "b.ne 4b\n" - "3:\n" - "cbz %[regs], 5f\n" + "b.ne 3b\n" + "2:\n" + "cbz %[regs], 4f\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" - "ld1rqw z5.s, p7/z, [a_ptr1]\n" + "ld1rqw z4.s, p7/z, [%[a_ptr0]]\n" "fmla z24.s, z8.s, z2.s[0]\n" - "ld1rqw z6.s, p7/z, [a_ptr2]\n" + "ld1rqw z5.s, p7/z, [a_ptr1]\n" "fmla z28.s, z8.s, z3.s[0]\n" - "ld1rqw z7.s, p7/z, [a_ptr3]\n" + "ld1rqw z6.s, p7/z, [a_ptr2]\n" "fmla z17.s, z9.s, z0.s[0]\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "ld1rqw z7.s, p7/z, [a_ptr3]\n" "fmla z21.s, z9.s, z1.s[0]\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z25.s, z9.s, z2.s[0]\n" "fmla z29.s, z9.s, z3.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" @@ -1644,8 +1700,8 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z20.s, z8.s, z1.s[2]\n" "fmla z24.s, z8.s, z2.s[2]\n" "fmla z28.s, z8.s, z3.s[2]\n" - "fmla z17.s, z9.s, z0.s[2]\n" "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "fmla z17.s, z9.s, z0.s[2]\n" "fmla z21.s, z9.s, z1.s[2]\n" "fmla z25.s, z9.s, z2.s[2]\n" "fmla z29.s, z9.s, z3.s[2]\n" @@ -1686,11 +1742,15 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z16.s, z8.s, z4.s[0]\n" "ld1rqw z3.s, p6/z, [a_ptr3, #0x10]\n" "fmla z20.s, z8.s, z5.s[0]\n" + "addvl %[a_ptr0], %[a_ptr0], #2\n" "fmla z24.s, z8.s, z6.s[0]\n" + "addvl a_ptr1, a_ptr1, #2\n" "fmla z28.s, z8.s, z7.s[0]\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "fmla z17.s, z9.s, z4.s[0]\n" + "addvl a_ptr2, a_ptr2, #2\n" "fmla z21.s, z9.s, z5.s[0]\n" + "addvl a_ptr3, a_ptr3, #2\n" "fmla z25.s, z9.s, z6.s[0]\n" "fmla z29.s, z9.s, z7.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" @@ -1725,6 +1785,7 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z31.s, z15.s, z7.s[1]\n" "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z16.s, z8.s, z4.s[2]\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" "fmla z20.s, z8.s, z5.s[2]\n" "fmla z24.s, z8.s, z6.s[2]\n" "fmla z28.s, z8.s, z7.s[2]\n" @@ -1756,14 +1817,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z15.s, z5.s[3]\n" "fmla z27.s, z15.s, z6.s[3]\n" "fmla z31.s, z15.s, z7.s[3]\n" - "cbz %[blocks], 6f\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "cbz %[blocks], 5f\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z16.s, z8.s, z0.s[0]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" "fmla z24.s, z8.s, z2.s[0]\n" "fmla z28.s, z8.s, z3.s[0]\n" @@ -1779,13 +1839,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z11.s, z1.s[0]\n" "fmla z27.s, z11.s, z2.s[0]\n" "fmla z31.s, z11.s, z3.s[0]\n" - "b.eq 6f\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "b.eq 5f\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z16.s, z12.s, z0.s[1]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z20.s, z12.s, z1.s[1]\n" "fmla z24.s, z12.s, z2.s[1]\n" "fmla z28.s, z12.s, z3.s[1]\n" @@ -1801,11 +1861,12 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z15.s, z1.s[1]\n" "fmla z27.s, z15.s, z2.s[1]\n" "fmla z31.s, z15.s, z3.s[1]\n" - "b.eq 6f\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "b.eq 5f\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" "fmla z20.s, z8.s, z1.s[2]\n" "fmla z24.s, z8.s, z2.s[2]\n" @@ -1822,25 +1883,30 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z11.s, z1.s[2]\n" "fmla z27.s, z11.s, z2.s[2]\n" "fmla z31.s, z11.s, z3.s[2]\n" - "b 6f\n" - "5:\n" + "b 5f\n" + "4:\n" "fmla z16.s, z8.s, z0.s[0]\n" - "ld1rqw z4.s, p6/z, [%[a_ptr0]]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" "fmla z20.s, z8.s, z1.s[0]\n" - "ld1rqw z5.s, p6/z, [a_ptr1]\n" + "ld1rqw z4.s, p6/z, [%[a_ptr0]]\n" "fmla z24.s, z8.s, z2.s[0]\n" - "ld1rqw z6.s, p6/z, [a_ptr2]\n" + "ld1rqw z5.s, p6/z, [a_ptr1]\n" "fmla z28.s, z8.s, z3.s[0]\n" - "ld1rqw z7.s, p6/z, [a_ptr3]\n" - "fmla z17.s, z9.s, z0.s[0]\n" "ld1w z8.s, p7/z, [%[b_ptr0]]\n" + "fmla z17.s, z9.s, z0.s[0]\n" + "ld1rqw z6.s, p6/z, [a_ptr2]\n" "fmla z21.s, z9.s, z1.s[0]\n" + "ld1rqw z7.s, p6/z, [a_ptr3]\n" "fmla z25.s, z9.s, z2.s[0]\n" + "addvl %[a_ptr0], %[a_ptr0], #1\n" "fmla z29.s, z9.s, z3.s[0]\n" "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" "fmla z18.s, z10.s, z0.s[0]\n" + "addvl a_ptr1, a_ptr1, #1\n" "fmla z22.s, z10.s, z1.s[0]\n" + "addvl a_ptr2, a_ptr2, #1\n" "fmla z26.s, z10.s, z2.s[0]\n" + "addvl a_ptr3, a_ptr3, #1\n" "fmla z30.s, z10.s, z3.s[0]\n" "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z19.s, z11.s, z0.s[0]\n" @@ -1869,6 +1935,7 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z31.s, z15.s, z3.s[1]\n" "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z16.s, z8.s, z0.s[2]\n" + "addvl %[b_ptr0], %[b_ptr0], #8\n" "fmla z20.s, z8.s, z1.s[2]\n" "fmla z24.s, z8.s, z2.s[2]\n" "fmla z28.s, z8.s, z3.s[2]\n" @@ -1900,14 +1967,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z15.s, z1.s[3]\n" "fmla z27.s, z15.s, z2.s[3]\n" "fmla z31.s, z15.s, z3.s[3]\n" - "cbz %[blocks], 6f\n" - "addvl %[b_ptr0], %[b_ptr0], #16\n" + "cbz %[blocks], 5f\n" + "ld1w z8.s, p7/z, [%[b_ptr0]]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" "fmla z16.s, z8.s, z4.s[0]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" "fmla z20.s, z8.s, z5.s[0]\n" "fmla z24.s, z8.s, z6.s[0]\n" "fmla z28.s, z8.s, z7.s[0]\n" @@ -1923,13 +1989,13 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z11.s, z5.s[0]\n" "fmla z27.s, z11.s, z6.s[0]\n" "fmla z31.s, z11.s, z7.s[0]\n" - "b.eq 6f\n" - "ld1w z12.s, p7/z, [%[b_ptr0], #-4, MUL VL]\n" + "b.eq 5f\n" + "ld1w z12.s, p7/z, [%[b_ptr0], #4, MUL VL]\n" "subs %[blocks], %[blocks], #0x1\n" - "ld1w z13.s, p7/z, [%[b_ptr0], #-3, MUL VL]\n" - "ld1w z14.s, p7/z, [%[b_ptr0], #-2, MUL VL]\n" - "ld1w z15.s, p7/z, [%[b_ptr0], #-1, MUL VL]\n" + "ld1w z13.s, p7/z, [%[b_ptr0], #5, MUL VL]\n" + "ld1w z14.s, p7/z, [%[b_ptr0], #6, MUL VL]\n" "fmla z16.s, z12.s, z4.s[1]\n" + "ld1w z15.s, p7/z, [%[b_ptr0], #7, MUL VL]\n" "fmla z20.s, z12.s, z5.s[1]\n" "fmla z24.s, z12.s, z6.s[1]\n" "fmla z28.s, z12.s, z7.s[1]\n" @@ -1945,11 +2011,12 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z15.s, z5.s[1]\n" "fmla z27.s, z15.s, z6.s[1]\n" "fmla z31.s, z15.s, z7.s[1]\n" - "b.eq 6f\n" - "ld1w z8.s, p7/z, [%[b_ptr0]]\n" - "ld1w z9.s, p7/z, [%[b_ptr0], #1, MUL VL]\n" - "ld1w z10.s, p7/z, [%[b_ptr0], #2, MUL VL]\n" - "ld1w z11.s, p7/z, [%[b_ptr0], #3, MUL VL]\n" + "b.eq 5f\n" + "addvl %[b_ptr0], %[b_ptr0], #16\n" + "ld1w z8.s, p7/z, [%[b_ptr0], #-8, MUL VL]\n" + "ld1w z9.s, p7/z, [%[b_ptr0], #-7, MUL VL]\n" + "ld1w z10.s, p7/z, [%[b_ptr0], #-6, MUL VL]\n" + "ld1w z11.s, p7/z, [%[b_ptr0], #-5, MUL VL]\n" "fmla z16.s, z8.s, z4.s[2]\n" "fmla z20.s, z8.s, z5.s[2]\n" "fmla z24.s, z8.s, z6.s[2]\n" @@ -1966,17 +2033,51 @@ void sve_hybrid_fp32_mla_4VLx4(const float *A, int lda, const float *B, float *C "fmla z23.s, z11.s, z5.s[2]\n" "fmla z27.s, z11.s, z6.s[2]\n" "fmla z31.s, z11.s, z7.s[2]\n" - "6:\n" + "5:\n" + "ld1rw z14.s, p7/z, [%[minptr]]\n" + "ld1rw z15.s, p7/z, [%[maxptr]]\n" + "fmax z16.s, p7/m, z16.s, z14.s\n" + "fmax z17.s, p7/m, z17.s, z14.s\n" + "fmax z18.s, p7/m, z18.s, z14.s\n" + "fmax z19.s, p7/m, z19.s, z14.s\n" + "fmin z16.s, p7/m, z16.s, z15.s\n" + "fmin z17.s, p7/m, z17.s, z15.s\n" + "fmin z18.s, p7/m, z18.s, z15.s\n" + "fmin z19.s, p7/m, z19.s, z15.s\n" "st1w z16.s, p0, [%[c_ptr0]]\n" + "fmax z20.s, p7/m, z20.s, z14.s\n" + "fmax z21.s, p7/m, z21.s, z14.s\n" + "fmax z22.s, p7/m, z22.s, z14.s\n" "st1w z17.s, p1, [%[c_ptr0], #1, MUL VL]\n" + "fmax z23.s, p7/m, z23.s, z14.s\n" + "fmin z20.s, p7/m, z20.s, z15.s\n" + "fmin z21.s, p7/m, z21.s, z15.s\n" "st1w z18.s, p2, [%[c_ptr0], #2, MUL VL]\n" + "fmin z22.s, p7/m, z22.s, z15.s\n" + "fmin z23.s, p7/m, z23.s, z15.s\n" + "fmax z24.s, p7/m, z24.s, z14.s\n" "st1w z19.s, p3, [%[c_ptr0], #3, MUL VL]\n" + "fmax z25.s, p7/m, z25.s, z14.s\n" "addvl %[c_ptr0], %[c_ptr0], #4\n" + "fmax z26.s, p7/m, z26.s, z14.s\n" "st1w z20.s, p0, [c_ptr1]\n" + "fmin z24.s, p7/m, z24.s, z15.s\n" + "fmin z25.s, p7/m, z25.s, z15.s\n" + "fmax z27.s, p7/m, z27.s, z14.s\n" "st1w z21.s, p1, [c_ptr1, #1, MUL VL]\n" + "fmin z26.s, p7/m, z26.s, z15.s\n" + "fmax z28.s, p7/m, z28.s, z14.s\n" + "fmax z29.s, p7/m, z29.s, z14.s\n" "st1w z22.s, p2, [c_ptr1, #2, MUL VL]\n" + "fmin z27.s, p7/m, z27.s, z15.s\n" + "fmax z30.s, p7/m, z30.s, z14.s\n" + "fmin z28.s, p7/m, z28.s, z15.s\n" "st1w z23.s, p3, [c_ptr1, #3, MUL VL]\n" + "fmin z29.s, p7/m, z29.s, z15.s\n" + "fmax z31.s, p7/m, z31.s, z14.s\n" + "fmin z30.s, p7/m, z30.s, z15.s\n" "st1w z24.s, p0, [c_ptr2]\n" + "fmin z31.s, p7/m, z31.s, z15.s\n" "st1w z25.s, p1, [c_ptr2, #1, MUL VL]\n" "st1w z26.s, p2, [c_ptr2, #2, MUL VL]\n" "st1w z27.s, p3, [c_ptr2, #3, MUL VL]\n" @@ -1991,11 +2092,12 @@ void sve_hybrid_fp32_mla_4VLx4(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), [temp] "+r" (temp), [blocks] "+r" (blocks) - : [betaptr] "r" (betaptr), [width] "r" (width), [beta0] "r" (beta0), [leftovers] "r" (leftovers), [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), [leftovers] "r" (leftovers) : "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31", "x0", "x1", "x2", "x3", "x4", "x5", "cc", "memory" ); break; } + } } } |