diff options
author | Michael Tyler <michael.tyler@arm.com> | 2023-04-12 17:43:17 +0100 |
---|---|---|
committer | michael.tyler <michael.tyler@arm.com> | 2023-06-05 15:57:58 +0000 |
commit | 74921eee924625426429044decefe3673561b174 (patch) | |
tree | 654da1a95e3d42d6af8ad1ff27bb40d77b1fd8c5 /src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp | |
parent | df5d9878008be9b60586df97ebfff197abb5195e (diff) | |
download | ComputeLibrary-74921eee924625426429044decefe3673561b174.tar.gz |
Update CPU kernel implementations and guard directives
Resolves COMPMID-6023
Change-Id: I868975d14c4f98af6716726feda22405a6a4c891
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp | 751 |
1 files changed, 374 insertions, 377 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp index dfdc4ea289..093feee6ce 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifdef __ARM_FEATURE_SVE -#ifdef ARM_COMPUTE_ENABLE_SME2 + +#if defined(ARM_COMPUTE_ENABLE_SME2) #include "arm_gemm.hpp" #include "../../utils.hpp" @@ -35,11 +35,9 @@ namespace arm_gemm { void sme2_gemv_u8qa_dot_16VL ( const uint8_t *A_ptr, const uint8_t *B_ptr, uint8_t *output_ptr, size_t N, size_t K, - const Requantize32 *qp, const int32_t *col_bias, unsigned int col_base + const Requantize32 *qp, const int32_t *col_bias, unsigned int ) { - ARM_COMPUTE_UNUSED(col_base); - struct KernelArgs { const uint8_t *B_ptr = {}; size_t output_offset = {}; @@ -52,7 +50,7 @@ void sme2_gemv_u8qa_dot_16VL ( flags |= 0x20; } __asm__ __volatile__( - "ptrue p2.b\n" + "ptrue p8.b\n" ".inst 0xd503477f // SMSTART ZA\n" "cntw x28, ALL, MUL #4\n" "add x27, %x[N], x28\n" @@ -84,8 +82,8 @@ void sme2_gemv_u8qa_dot_16VL ( ".inst 0xf8b64b5a // rprfm pldonce, x22, [x26]\n" "3:" // RHS prefetch exit "mov x24, %x[col_bias]\n" - "mov z26.s, #0x0\n" - "mov z24.b, #0x1\n" + "mov z28.s, #0x0\n" + "mov z29.b, #0x1\n" "bic %x[flags], %x[flags], #0x80000000\n" "4:" // Column loop "cmp x27, #0x4\n" @@ -94,404 +92,404 @@ void sme2_gemv_u8qa_dot_16VL ( "bgt 24f\n" "beq 14f\n" "mov x23, %x[A_ptr]\n" - "mov x22, %x[K]\n" - "mov x20, %x[N]\n" "mov x21, %x[K]\n" - ".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n" + "mov x20, %x[N]\n" + "mov x22, %x[K]\n" + ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" "whilelt p1.b, XZR, x20\n" "cbz x24, 5f\n" - ".inst 0xa040c304 // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n" - ".inst 0xc0042c80 // mova za.d[x9, #0], { z4.d-z7.d }\n" + ".inst 0xa040c300 // ld1w { z0.s-z3.s }, pn8.b/Z, [x24]\n" + ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n" "b 6f\n" "5:" // Width 1: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "6:" // Width 1: setup done - "cmp x21, #0x10\n" + "cmp x22, #0x10\n" "ble 9f\n" "7:" // Width 1: Multiply loop: Main loop head - "whilelt p0.b, XZR, x21\n" - "ld1rqb { z3.b }, p0/Z, [x23]\n" + "whilelt p0.b, XZR, x22\n" + "ld1rqb { z1.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153b230 // udot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n" - "addvl x26, x26, #16\n" ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n" + ".inst 0xc151b2b0 // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[0]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151b5b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[1]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151b9b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151bdb0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n" "addvl x26, x26, #16\n" "tbnz %x[flags], #31, 8f\n" - "udot z26.s, z3.b, z24.b\n" + "udot z28.s, z1.b, z29.b\n" "8:" // Width 1: Multiply loop: unique 1: skip row sum - "sub x21, x21, #0x10\n" - "cmp x21, #0x10\n" + "sub x22, x22, #0x10\n" + "cmp x22, #0x10\n" "bgt 7b\n" "9:" // Width 1: Multiply loop: Single iteration only - "whilelt p0.b, XZR, x21\n" - "ld1rqb { z3.b }, p0/Z, [x23]\n" - "subs x21, x21, #0x4\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + "whilelt p0.b, XZR, x22\n" + "ld1rqb { z1.b }, p0/Z, [x23]\n" + "subs x22, x22, #0x4\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" "add x23, x23, #0x10\n" - ".inst 0xc153b230 // udot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n" + ".inst 0xc151b1b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n" "addvl x26, x26, #16\n" "ble 10f\n" ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - "subs x21, x21, #0x4\n" - ".inst 0xc153b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n" + "subs x22, x22, #0x4\n" + ".inst 0xc151b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[1]\n" "addvl x26, x26, #16\n" "ble 10f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - "subs x21, x21, #0x4\n" - ".inst 0xc153bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n" + ".inst 0xa0408349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n" + "subs x22, x22, #0x4\n" + ".inst 0xc151b930 // udot za.s[x9, 0], { z8.b-z11.b }, z1.b[2]\n" "addvl x26, x26, #16\n" "ble 10f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n" + ".inst 0xa0408349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151bd30 // udot za.s[x9, 0], { z8.b-z11.b }, z1.b[3]\n" "addvl x26, x26, #16\n" "10:" // Width 1: Multiply loop: multiply skip "tbnz %x[flags], #31, 11f\n" - "udot z26.s, z3.b, z24.b\n" + "udot z28.s, z1.b, z29.b\n" "11:" // Width 1: Multiply loop: unique 2: skip row sum "tbnz %x[flags], #31, 12f\n" "add x21, %x[qp], %[b_offset]\n" "mov x20, #0x4\n" - "ld1rw { z10.s }, p2/Z, [x21]\n" - "neg z10.s, p2/M, z10.s\n" + "ld1rw { z26.s }, p2/Z, [x21]\n" + "neg z26.s, p2/M, z26.s\n" "whilelt p0.s, XZR, x20\n" - "uaddv d26, p0, z26.s\n" - "mov z26.s, z26.s[0]\n" - "mul z26.s, p2/M, z26.s, z10.s\n" + "uaddv d28, p0, z28.s\n" + "mov z28.s, z28.s[0]\n" + "mul z28.s, p2/M, z28.s, z26.s\n" "orr %x[flags], %x[flags], #0x80000000\n" "12:" // Width 1: skip row sum fixup - ".inst 0xc0904b40 // addha za0.s, p2/M, p2/M, z26.s\n" + ".inst 0xc0904b80 // addha za0.s, p2/M, p2/M, z28.s\n" "add x20, %x[qp], %[per_layer_mul]\n" - "ld1rw { z5.s }, p2/Z, [x20]\n" + "ld1rw { z1.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" - ".inst 0xc0904b41 // addha za1.s, p2/M, p2/M, z26.s\n" - "ld1rw { z4.s }, p2/Z, [x20]\n" + ".inst 0xc0904b81 // addha za1.s, p2/M, p2/M, z28.s\n" + "ld1rw { z7.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[c_offset]\n" "add x21, %x[qp], %[minval]\n" - ".inst 0xc0904b42 // addha za2.s, p2/M, p2/M, z26.s\n" - "ld1rw { z6.s }, p2/Z, [x20]\n" + ".inst 0xc0904b82 // addha za2.s, p2/M, p2/M, z28.s\n" + "ld1rw { z2.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[maxval]\n" - ".inst 0xc0904b43 // addha za3.s, p2/M, p2/M, z26.s\n" - "ld1rw { z21.s }, p2/Z, [x21]\n" - ".inst 0xc0062c1c // mova { z28.d-z31.d }, za.d[x9, #0]\n" - ".inst 0xc1a5ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z5.s\n" - ".inst 0xc1a4aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z4.s\n" - "ld1rw { z16.s }, p2/Z, [x20]\n" - ".inst 0xc1a6ab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z6.s\n" - ".inst 0xc1b0cebc // sclamp { z28.s-z31.s }, z21.s, z16.s\n" - "uzp1 z28.h, z28.h, z29.h\n" - "uzp1 z29.h, z30.h, z31.h\n" - "uzp1 z28.b, z28.b, z29.b\n" - "st1b { z28.b }, p1, [x25]\n" + ".inst 0xc0904b83 // addha za3.s, p2/M, p2/M, z28.s\n" + "ld1rw { z16.s }, p2/Z, [x21]\n" + ".inst 0xc0062c0c // mova { z12.d-z15.d }, za.d[x9, #0]\n" + ".inst 0xc1a1ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n" + ".inst 0xc1a7aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n" + "ld1rw { z30.s }, p2/Z, [x20]\n" + ".inst 0xc1a2ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n" + ".inst 0xc1bece0c // sclamp { z12.s-z15.s }, z16.s, z30.s\n" + "uzp1 z12.h, z12.h, z13.h\n" + "uzp1 z19.h, z14.h, z15.h\n" + "uzp1 z12.b, z12.b, z19.b\n" + "st1b { z12.b }, p1, [x25]\n" "addvl x25, x25, #1\n" "13:" // Width 1: Output done "b 44f\n" "14:" // Width 2 "mov x23, %x[A_ptr]\n" - "mov x22, %x[K]\n" - "sub x20, %x[N], x28\n" "mov x21, %x[K]\n" - ".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n" + "sub x20, %x[N], x28\n" + "mov x22, %x[K]\n" + ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" "whilelt p1.b, XZR, x20\n" "cbz x24, 15f\n" - ".inst 0xa040c304 // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n" - ".inst 0xc0042c80 // mova za.d[x9, #0], { z4.d-z7.d }\n" - ".inst 0xa041c314 // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0042e81 // mova za.d[x9, #1], { z20.d-z23.d }\n" + ".inst 0xa040c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24]\n" + ".inst 0xc0042e00 // mova za.d[x9, #0], { z16.d-z19.d }\n" + ".inst 0xa041c318 // ld1w { z24.s-z27.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xc0042f01 // mova za.d[x9, #1], { z24.d-z27.d }\n" "b 16f\n" "15:" // Width 2: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "16:" // Width 2: setup done - "cmp x21, #0x10\n" + "cmp x22, #0x10\n" "ble 19f\n" "17:" // Width 2: Multiply loop: Main loop head - "whilelt p0.b, XZR, x21\n" - "ld1rqb { z3.b }, p0/Z, [x23]\n" + "whilelt p0.b, XZR, x22\n" + "ld1rqb { z1.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153b230 // udot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b0b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151b1b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n" + ".inst 0xa0418359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b331 // udot za.s[x9, 1], { z24.b-z27.b }, z1.b[0]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n" - ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b531 // udot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n" + ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151b630 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n" + ".inst 0xa0418355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b6b1 // udot za.s[x9, 1], { z20.b-z23.b }, z1.b[1]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b8b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151b9b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n" + ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b9b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n" - ".inst 0xa0418355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153beb1 // udot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n" + ".inst 0xa0408345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151bcb0 // udot za.s[x9, 0], { z4.b-z7.b }, z1.b[3]\n" + ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151bd31 // udot za.s[x9, 1], { z8.b-z11.b }, z1.b[3]\n" "addvl x26, x26, #16\n" "tbnz %x[flags], #31, 18f\n" - "udot z26.s, z3.b, z24.b\n" + "udot z28.s, z1.b, z29.b\n" "18:" // Width 2: Multiply loop: unique 3: skip row sum - "sub x21, x21, #0x10\n" - "cmp x21, #0x10\n" + "sub x22, x22, #0x10\n" + "cmp x22, #0x10\n" "bgt 17b\n" "19:" // Width 2: Multiply loop: Single iteration only - "whilelt p0.b, XZR, x21\n" - "ld1rqb { z3.b }, p0/Z, [x23]\n" - "subs x21, x21, #0x4\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + "whilelt p0.b, XZR, x22\n" + "ld1rqb { z1.b }, p0/Z, [x23]\n" + "subs x22, x22, #0x4\n" + ".inst 0xa0408359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n" "add x23, x23, #0x10\n" - ".inst 0xc153b230 // udot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b0b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n" + ".inst 0xc151b330 // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[0]\n" + ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b231 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n" "addvl x26, x26, #16\n" "ble 20f\n" ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - "subs x21, x21, #0x4\n" - ".inst 0xc153b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n" - ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b531 // udot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n" + "subs x22, x22, #0x4\n" + ".inst 0xc151b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[1]\n" + ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b5b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n" "addvl x26, x26, #16\n" "ble 20f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - "subs x21, x21, #0x4\n" - ".inst 0xc153bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b8b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + "subs x22, x22, #0x4\n" + ".inst 0xc151b9b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n" + ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b9b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n" "addvl x26, x26, #16\n" "ble 20f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n" - ".inst 0xa0418355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153beb1 // udot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n" + ".inst 0xa0408359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151bf30 // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[3]\n" + ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151bd31 // udot za.s[x9, 1], { z8.b-z11.b }, z1.b[3]\n" "addvl x26, x26, #16\n" "20:" // Width 2: Multiply loop: multiply skip "tbnz %x[flags], #31, 21f\n" - "udot z26.s, z3.b, z24.b\n" + "udot z28.s, z1.b, z29.b\n" "21:" // Width 2: Multiply loop: unique 4: skip row sum "tbnz %x[flags], #31, 22f\n" "add x21, %x[qp], %[b_offset]\n" "mov x20, #0x4\n" - "ld1rw { z10.s }, p2/Z, [x21]\n" - "neg z10.s, p2/M, z10.s\n" + "ld1rw { z16.s }, p2/Z, [x21]\n" + "neg z16.s, p2/M, z16.s\n" "whilelt p0.s, XZR, x20\n" - "uaddv d26, p0, z26.s\n" - "mov z26.s, z26.s[0]\n" - "mul z26.s, p2/M, z26.s, z10.s\n" + "uaddv d28, p0, z28.s\n" + "mov z28.s, z28.s[0]\n" + "mul z28.s, p2/M, z28.s, z16.s\n" "orr %x[flags], %x[flags], #0x80000000\n" "22:" // Width 2: skip row sum fixup - ".inst 0xc0904b40 // addha za0.s, p2/M, p2/M, z26.s\n" + ".inst 0xc0904b80 // addha za0.s, p2/M, p2/M, z28.s\n" "add x20, %x[qp], %[per_layer_mul]\n" - "ld1rw { z5.s }, p2/Z, [x20]\n" + "ld1rw { z6.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" - ".inst 0xc0904b41 // addha za1.s, p2/M, p2/M, z26.s\n" - "ld1rw { z4.s }, p2/Z, [x20]\n" + ".inst 0xc0904b81 // addha za1.s, p2/M, p2/M, z28.s\n" + "ld1rw { z5.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[c_offset]\n" "add x21, %x[qp], %[minval]\n" - ".inst 0xc0904b42 // addha za2.s, p2/M, p2/M, z26.s\n" - "ld1rw { z6.s }, p2/Z, [x20]\n" + ".inst 0xc0904b82 // addha za2.s, p2/M, p2/M, z28.s\n" + "ld1rw { z9.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[maxval]\n" - ".inst 0xc0904b43 // addha za3.s, p2/M, p2/M, z26.s\n" - "ld1rw { z21.s }, p2/Z, [x21]\n" - ".inst 0xc0062c1c // mova { z28.d-z31.d }, za.d[x9, #0]\n" - ".inst 0xc1a5ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z5.s\n" - ".inst 0xc0062c2c // mova { z12.d-z15.d }, za.d[x9, #1]\n" - ".inst 0xc1a5ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z5.s\n" - ".inst 0xc1a4aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z4.s\n" - "ld1rw { z16.s }, p2/Z, [x20]\n" - ".inst 0xc1a4aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z4.s\n" - ".inst 0xc1a6ab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z6.s\n" - ".inst 0xc1a6ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n" - ".inst 0xc1b0cebc // sclamp { z28.s-z31.s }, z21.s, z16.s\n" - ".inst 0xc1b0ceac // sclamp { z12.s-z15.s }, z21.s, z16.s\n" - "uzp1 z28.h, z28.h, z29.h\n" - "uzp1 z29.h, z30.h, z31.h\n" - "uzp1 z12.h, z12.h, z13.h\n" - "uzp1 z13.h, z14.h, z15.h\n" - "uzp1 z28.b, z28.b, z29.b\n" - "st1b { z28.b }, p2, [x25]\n" - "uzp1 z12.b, z12.b, z13.b\n" - "st1b { z12.b }, p1, [x25, #1, MUL VL]\n" + ".inst 0xc0904b83 // addha za3.s, p2/M, p2/M, z28.s\n" + "ld1rw { z16.s }, p2/Z, [x21]\n" + ".inst 0xc0062c18 // mova { z24.d-z27.d }, za.d[x9, #0]\n" + ".inst 0xc1a6ac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n" + ".inst 0xc0062c20 // mova { z0.d-z3.d }, za.d[x9, #1]\n" + ".inst 0xc1a6ac00 // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n" + ".inst 0xc1a5aa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z5.s\n" + "ld1rw { z21.s }, p2/Z, [x20]\n" + ".inst 0xc1a5aa20 // srshl { z0.s-z3.s }, { z0.s-z3.s }, z5.s\n" + ".inst 0xc1a9ab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z9.s\n" + ".inst 0xc1a9ab00 // add { z0.s-z3.s }, { z0.s-z3.s }, z9.s\n" + ".inst 0xc1b5ce18 // sclamp { z24.s-z27.s }, z16.s, z21.s\n" + ".inst 0xc1b5ce00 // sclamp { z0.s-z3.s }, z16.s, z21.s\n" + "uzp1 z24.h, z24.h, z25.h\n" + "uzp1 z9.h, z26.h, z27.h\n" + "uzp1 z0.h, z0.h, z1.h\n" + "uzp1 z26.h, z2.h, z3.h\n" + "uzp1 z24.b, z24.b, z9.b\n" + "st1b { z24.b }, p2, [x25]\n" + "uzp1 z0.b, z0.b, z26.b\n" + "st1b { z0.b }, p1, [x25, #1, MUL VL]\n" "addvl x25, x25, #2\n" "23:" // Width 2: Output done "b 44f\n" "24:" // Width 3 "mov x20, #0x2\n" "mov x23, %x[A_ptr]\n" - "mov x22, %x[K]\n" - "msub x20, x28, x20, %x[N]\n" "mov x21, %x[K]\n" - ".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n" + "msub x20, x28, x20, %x[N]\n" + "mov x22, %x[K]\n" + ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" "whilelt p1.b, XZR, x20\n" "cbz x24, 25f\n" - ".inst 0xa040c304 // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n" - ".inst 0xc0042c80 // mova za.d[x9, #0], { z4.d-z7.d }\n" - ".inst 0xa041c314 // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0042e81 // mova za.d[x9, #1], { z20.d-z23.d }\n" - ".inst 0xa042c314 // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n" - ".inst 0xc0042e82 // mova za.d[x9, #2], { z20.d-z23.d }\n" + ".inst 0xa040c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24]\n" + ".inst 0xc0042e00 // mova za.d[x9, #0], { z16.d-z19.d }\n" + ".inst 0xa041c30c // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xc0042d81 // mova za.d[x9, #1], { z12.d-z15.d }\n" + ".inst 0xa042c318 // ld1w { z24.s-z27.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n" + ".inst 0xc0042f02 // mova za.d[x9, #2], { z24.d-z27.d }\n" "b 26f\n" "25:" // Width 3: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "26:" // Width 3: setup done - "cmp x21, #0x10\n" + "cmp x22, #0x10\n" "ble 29f\n" "27:" // Width 3: Multiply loop: Main loop head - "whilelt p0.b, XZR, x21\n" - "ld1rqb { z3.b }, p0/Z, [x23]\n" + "whilelt p0.b, XZR, x22\n" + "ld1rqb { z1.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153b230 // udot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b0b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n" + ".inst 0xc151b230 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[0]\n" + ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b231 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n" ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b1b2 // udot za.s[x9, 2], { z12.b-z15.b }, z3.b[0]\n" + ".inst 0xc151b1b2 // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[0]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n" - ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b531 // udot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b632 // udot za.s[x9, 2], { z16.b-z19.b }, z3.b[1]\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151b5b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[1]\n" + ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b5b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n" + ".inst 0xa0428355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc151b6b2 // udot za.s[x9, 2], { z20.b-z23.b }, z1.b[1]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n" + ".inst 0xa0408349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151b930 // udot za.s[x9, 0], { z8.b-z11.b }, z1.b[2]\n" ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b8b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b9b2 // udot za.s[x9, 2], { z12.b-z15.b }, z3.b[2]\n" + ".inst 0xc151b8b1 // udot za.s[x9, 1], { z4.b-z7.b }, z1.b[2]\n" + ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc151ba32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n" - ".inst 0xa0418355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153beb1 // udot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n" + ".inst 0xa0408359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151bf30 // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[3]\n" + ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151bcb1 // udot za.s[x9, 1], { z4.b-z7.b }, z1.b[3]\n" ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153be32 // udot za.s[x9, 2], { z16.b-z19.b }, z3.b[3]\n" + ".inst 0xc151be32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[3]\n" "addvl x26, x26, #16\n" "tbnz %x[flags], #31, 28f\n" - "udot z26.s, z3.b, z24.b\n" + "udot z28.s, z1.b, z29.b\n" "28:" // Width 3: Multiply loop: unique 5: skip row sum - "sub x21, x21, #0x10\n" - "cmp x21, #0x10\n" + "sub x22, x22, #0x10\n" + "cmp x22, #0x10\n" "bgt 27b\n" "29:" // Width 3: Multiply loop: Single iteration only - "whilelt p0.b, XZR, x21\n" - "ld1rqb { z3.b }, p0/Z, [x23]\n" - "subs x21, x21, #0x4\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + "whilelt p0.b, XZR, x22\n" + "ld1rqb { z1.b }, p0/Z, [x23]\n" + "subs x22, x22, #0x4\n" + ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" "add x23, x23, #0x10\n" - ".inst 0xc153b230 // udot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b0b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b1b2 // udot za.s[x9, 2], { z12.b-z15.b }, z3.b[0]\n" + ".inst 0xc151b2b0 // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[0]\n" + ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b231 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n" + ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc151b232 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[0]\n" "addvl x26, x26, #16\n" "ble 30f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - "subs x21, x21, #0x4\n" - ".inst 0xc153b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n" - ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b531 // udot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n" + ".inst 0xa0408359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n" + "subs x22, x22, #0x4\n" + ".inst 0xc151b730 // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[1]\n" + ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b5b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n" ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b632 // udot za.s[x9, 2], { z16.b-z19.b }, z3.b[1]\n" + ".inst 0xc151b632 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[1]\n" "addvl x26, x26, #16\n" "ble 30f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - "subs x21, x21, #0x4\n" - ".inst 0xc153bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b8b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b9b2 // udot za.s[x9, 2], { z12.b-z15.b }, z3.b[2]\n" + ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + "subs x22, x22, #0x4\n" + ".inst 0xc151ba30 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n" + ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151ba31 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[2]\n" + ".inst 0xa0428355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc151bab2 // udot za.s[x9, 2], { z20.b-z23.b }, z1.b[2]\n" "addvl x26, x26, #16\n" "ble 30f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n" - ".inst 0xa0418355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153beb1 // udot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153be32 // udot za.s[x9, 2], { z16.b-z19.b }, z3.b[3]\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151bdb0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n" + ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151be31 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[3]\n" + ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc151bdb2 // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[3]\n" "addvl x26, x26, #16\n" "30:" // Width 3: Multiply loop: multiply skip "tbnz %x[flags], #31, 31f\n" - "udot z26.s, z3.b, z24.b\n" + "udot z28.s, z1.b, z29.b\n" "31:" // Width 3: Multiply loop: unique 6: skip row sum "tbnz %x[flags], #31, 32f\n" "add x21, %x[qp], %[b_offset]\n" "mov x20, #0x4\n" - "ld1rw { z10.s }, p2/Z, [x21]\n" - "neg z10.s, p2/M, z10.s\n" + "ld1rw { z16.s }, p2/Z, [x21]\n" + "neg z16.s, p2/M, z16.s\n" "whilelt p0.s, XZR, x20\n" - "uaddv d26, p0, z26.s\n" - "mov z26.s, z26.s[0]\n" - "mul z26.s, p2/M, z26.s, z10.s\n" + "uaddv d28, p0, z28.s\n" + "mov z28.s, z28.s[0]\n" + "mul z28.s, p2/M, z28.s, z16.s\n" "orr %x[flags], %x[flags], #0x80000000\n" "32:" // Width 3: skip row sum fixup - ".inst 0xc0904b40 // addha za0.s, p2/M, p2/M, z26.s\n" + ".inst 0xc0904b80 // addha za0.s, p2/M, p2/M, z28.s\n" "add x20, %x[qp], %[per_layer_mul]\n" - "ld1rw { z5.s }, p2/Z, [x20]\n" + "ld1rw { z2.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" - ".inst 0xc0904b41 // addha za1.s, p2/M, p2/M, z26.s\n" - "ld1rw { z4.s }, p2/Z, [x20]\n" + ".inst 0xc0904b81 // addha za1.s, p2/M, p2/M, z28.s\n" + "ld1rw { z1.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[c_offset]\n" "add x21, %x[qp], %[minval]\n" - ".inst 0xc0904b42 // addha za2.s, p2/M, p2/M, z26.s\n" - "ld1rw { z6.s }, p2/Z, [x20]\n" + ".inst 0xc0904b82 // addha za2.s, p2/M, p2/M, z28.s\n" + "ld1rw { z3.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[maxval]\n" - ".inst 0xc0904b43 // addha za3.s, p2/M, p2/M, z26.s\n" - "ld1rw { z21.s }, p2/Z, [x21]\n" - ".inst 0xc0062c1c // mova { z28.d-z31.d }, za.d[x9, #0]\n" - ".inst 0xc1a5ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z5.s\n" - ".inst 0xc0062c2c // mova { z12.d-z15.d }, za.d[x9, #1]\n" - ".inst 0xc1a5ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z5.s\n" - ".inst 0xc0062c40 // mova { z0.d-z3.d }, za.d[x9, #2]\n" - ".inst 0xc1a5ac00 // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z5.s\n" - ".inst 0xc1a4aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z4.s\n" - "ld1rw { z16.s }, p2/Z, [x20]\n" - ".inst 0xc1a4aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z4.s\n" - ".inst 0xc1a4aa20 // srshl { z0.s-z3.s }, { z0.s-z3.s }, z4.s\n" - ".inst 0xc1a6ab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z6.s\n" - ".inst 0xc1a6ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n" - ".inst 0xc1a6ab00 // add { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n" - ".inst 0xc1b0cebc // sclamp { z28.s-z31.s }, z21.s, z16.s\n" - ".inst 0xc1b0ceac // sclamp { z12.s-z15.s }, z21.s, z16.s\n" - "uzp1 z28.h, z28.h, z29.h\n" - ".inst 0xc1b0cea0 // sclamp { z0.s-z3.s }, z21.s, z16.s\n" - "uzp1 z29.h, z30.h, z31.h\n" + ".inst 0xc0904b83 // addha za3.s, p2/M, p2/M, z28.s\n" + "ld1rw { z16.s }, p2/Z, [x21]\n" + ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" + ".inst 0xc1a2ac08 // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n" + ".inst 0xc0062c24 // mova { z4.d-z7.d }, za.d[x9, #1]\n" + ".inst 0xc1a2ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n" + ".inst 0xc0062c4c // mova { z12.d-z15.d }, za.d[x9, #2]\n" + ".inst 0xc1a2ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n" + ".inst 0xc1a1aa28 // srshl { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n" + "ld1rw { z0.s }, p2/Z, [x20]\n" + ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" + ".inst 0xc1a1aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n" + ".inst 0xc1a3ab08 // add { z8.s-z11.s }, { z8.s-z11.s }, z3.s\n" + ".inst 0xc1a3ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" + ".inst 0xc1a3ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z3.s\n" + ".inst 0xc1a0ce08 // sclamp { z8.s-z11.s }, z16.s, z0.s\n" + ".inst 0xc1a0ce04 // sclamp { z4.s-z7.s }, z16.s, z0.s\n" + "uzp1 z8.h, z8.h, z9.h\n" + ".inst 0xc1a0ce0c // sclamp { z12.s-z15.s }, z16.s, z0.s\n" + "uzp1 z18.h, z10.h, z11.h\n" + "uzp1 z4.h, z4.h, z5.h\n" + "uzp1 z17.h, z6.h, z7.h\n" "uzp1 z12.h, z12.h, z13.h\n" - "uzp1 z13.h, z14.h, z15.h\n" - "uzp1 z0.h, z0.h, z1.h\n" - "uzp1 z1.h, z2.h, z3.h\n" - "uzp1 z28.b, z28.b, z29.b\n" - "st1b { z28.b }, p2, [x25]\n" - "uzp1 z12.b, z12.b, z13.b\n" - "st1b { z12.b }, p2, [x25, #1, MUL VL]\n" - "uzp1 z0.b, z0.b, z1.b\n" - "st1b { z0.b }, p1, [x25, #2, MUL VL]\n" + "uzp1 z16.h, z14.h, z15.h\n" + "uzp1 z8.b, z8.b, z18.b\n" + "st1b { z8.b }, p2, [x25]\n" + "uzp1 z4.b, z4.b, z17.b\n" + "st1b { z4.b }, p2, [x25, #1, MUL VL]\n" + "uzp1 z12.b, z12.b, z16.b\n" + "st1b { z12.b }, p1, [x25, #2, MUL VL]\n" "addvl x25, x25, #3\n" "33:" // Width 3: Output done "b 44f\n" "34:" // Width 4 "mov x20, #0x3\n" "mov x23, %x[A_ptr]\n" - "mov x22, %x[K]\n" - "msub x20, x28, x20, %x[N]\n" "mov x21, %x[K]\n" - ".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n" + "msub x20, x28, x20, %x[N]\n" + "mov x22, %x[K]\n" + ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" "whilelt p1.b, XZR, x20\n" "cbz x24, 35f\n" - ".inst 0xa040c304 // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n" - ".inst 0xc0042c80 // mova za.d[x9, #0], { z4.d-z7.d }\n" - ".inst 0xa041c314 // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0042e81 // mova za.d[x9, #1], { z20.d-z23.d }\n" - ".inst 0xa042c314 // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n" - ".inst 0xc0042e82 // mova za.d[x9, #2], { z20.d-z23.d }\n" + ".inst 0xa040c314 // ld1w { z20.s-z23.s }, pn8.b/Z, [x24]\n" + ".inst 0xc0042e80 // mova za.d[x9, #0], { z20.d-z23.d }\n" + ".inst 0xa041c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xc0042e01 // mova za.d[x9, #1], { z16.d-z19.d }\n" + ".inst 0xa042c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n" + ".inst 0xc0042e02 // mova za.d[x9, #2], { z16.d-z19.d }\n" ".inst 0xa043c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0xc, MUL VL]\n" ".inst 0xc0042e03 // mova za.d[x9, #3], { z16.d-z19.d }\n" "addvl x24, x24, #16\n" @@ -499,165 +497,165 @@ void sme2_gemv_u8qa_dot_16VL ( "35:" // Width 4: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "36:" // Width 4: setup done - "cmp x21, #0x10\n" + "cmp x22, #0x10\n" "ble 39f\n" "37:" // Width 4: Multiply loop: Main loop head - "whilelt p0.b, XZR, x21\n" - "ld1rqb { z3.b }, p0/Z, [x23]\n" + "whilelt p0.b, XZR, x22\n" + "ld1rqb { z1.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153b230 // udot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b0b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n" + ".inst 0xc151b230 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[0]\n" + ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b231 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n" ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b1b2 // udot za.s[x9, 2], { z12.b-z15.b }, z3.b[0]\n" + ".inst 0xc151b1b2 // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[0]\n" ".inst 0xa043834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153b1b3 // udot za.s[x9, 3], { z12.b-z15.b }, z3.b[0]\n" + ".inst 0xc151b1b3 // udot za.s[x9, 3], { z12.b-z15.b }, z1.b[0]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n" - ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b531 // udot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n" + ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151b630 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n" + ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b5b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n" ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b632 // udot za.s[x9, 2], { z16.b-z19.b }, z3.b[1]\n" - ".inst 0xa043834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153b5b3 // udot za.s[x9, 3], { z12.b-z15.b }, z3.b[1]\n" - "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b8b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b9b2 // udot za.s[x9, 2], { z12.b-z15.b }, z3.b[2]\n" - ".inst 0xa043835d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153bbb3 // udot za.s[x9, 3], { z28.b-z31.b }, z3.b[2]\n" + ".inst 0xc151b632 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[1]\n" + ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc151b633 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[1]\n" "addvl x26, x26, #16\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n" - ".inst 0xa0418355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153beb1 // udot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n" + ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151ba30 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n" + ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b9b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n" ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153be32 // udot za.s[x9, 2], { z16.b-z19.b }, z3.b[3]\n" + ".inst 0xc151ba32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n" ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153be33 // udot za.s[x9, 3], { z16.b-z19.b }, z3.b[3]\n" + ".inst 0xc151ba33 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[2]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151bdb0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n" + ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151bdb1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[3]\n" + ".inst 0xa0428359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc151bf32 // udot za.s[x9, 2], { z24.b-z27.b }, z1.b[3]\n" + ".inst 0xa0438345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc151bcb3 // udot za.s[x9, 3], { z4.b-z7.b }, z1.b[3]\n" "addvl x26, x26, #16\n" "tbnz %x[flags], #31, 38f\n" - "udot z26.s, z3.b, z24.b\n" + "udot z28.s, z1.b, z29.b\n" "38:" // Width 4: Multiply loop: unique 7: skip row sum - "sub x21, x21, #0x10\n" - "cmp x21, #0x10\n" + "sub x22, x22, #0x10\n" + "cmp x22, #0x10\n" "bgt 37b\n" "39:" // Width 4: Multiply loop: Single iteration only - "whilelt p0.b, XZR, x21\n" - "ld1rqb { z3.b }, p0/Z, [x23]\n" - "subs x21, x21, #0x4\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + "whilelt p0.b, XZR, x22\n" + "ld1rqb { z1.b }, p0/Z, [x23]\n" + "subs x22, x22, #0x4\n" + ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" "add x23, x23, #0x10\n" - ".inst 0xc153b230 // udot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b0b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b1b2 // udot za.s[x9, 2], { z12.b-z15.b }, z3.b[0]\n" - ".inst 0xa043834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153b1b3 // udot za.s[x9, 3], { z12.b-z15.b }, z3.b[0]\n" + ".inst 0xc151b1b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n" + ".inst 0xa0418359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b331 // udot za.s[x9, 1], { z24.b-z27.b }, z1.b[0]\n" + ".inst 0xa0428349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc151b132 // udot za.s[x9, 2], { z8.b-z11.b }, z1.b[0]\n" + ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc151b233 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[0]\n" "addvl x26, x26, #16\n" "ble 40f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - "subs x21, x21, #0x4\n" - ".inst 0xc153b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n" - ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b531 // udot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b632 // udot za.s[x9, 2], { z16.b-z19.b }, z3.b[1]\n" - ".inst 0xa043834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153b5b3 // udot za.s[x9, 3], { z12.b-z15.b }, z3.b[1]\n" + ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + "subs x22, x22, #0x4\n" + ".inst 0xc151b630 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n" + ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151b631 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[1]\n" + ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc151b5b2 // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[1]\n" + ".inst 0xa0438355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc151b6b3 // udot za.s[x9, 3], { z20.b-z23.b }, z1.b[1]\n" "addvl x26, x26, #16\n" "ble 40f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - "subs x21, x21, #0x4\n" - ".inst 0xc153bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153b8b1 // udot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153b9b2 // udot za.s[x9, 2], { z12.b-z15.b }, z3.b[2]\n" - ".inst 0xa043835d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153bbb3 // udot za.s[x9, 3], { z28.b-z31.b }, z3.b[2]\n" + ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + "subs x22, x22, #0x4\n" + ".inst 0xc151ba30 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n" + ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151ba31 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[2]\n" + ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc151ba32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n" + ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc151ba33 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[2]\n" "addvl x26, x26, #16\n" "ble 40f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc153beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n" - ".inst 0xa0418355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153beb1 // udot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n" + ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + ".inst 0xc151be30 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[3]\n" + ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc151be31 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[3]\n" ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc153be32 // udot za.s[x9, 2], { z16.b-z19.b }, z3.b[3]\n" + ".inst 0xc151be32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[3]\n" ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153be33 // udot za.s[x9, 3], { z16.b-z19.b }, z3.b[3]\n" + ".inst 0xc151be33 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[3]\n" "addvl x26, x26, #16\n" "40:" // Width 4: Multiply loop: multiply skip "tbnz %x[flags], #31, 41f\n" - "udot z26.s, z3.b, z24.b\n" + "udot z28.s, z1.b, z29.b\n" "41:" // Width 4: Multiply loop: unique 8: skip row sum "tbnz %x[flags], #31, 42f\n" "add x21, %x[qp], %[b_offset]\n" "mov x20, #0x4\n" - "ld1rw { z10.s }, p2/Z, [x21]\n" - "neg z10.s, p2/M, z10.s\n" + "ld1rw { z16.s }, p2/Z, [x21]\n" + "neg z16.s, p2/M, z16.s\n" "whilelt p0.s, XZR, x20\n" - "uaddv d26, p0, z26.s\n" - "mov z26.s, z26.s[0]\n" - "mul z26.s, p2/M, z26.s, z10.s\n" + "uaddv d28, p0, z28.s\n" + "mov z28.s, z28.s[0]\n" + "mul z28.s, p2/M, z28.s, z16.s\n" "orr %x[flags], %x[flags], #0x80000000\n" "42:" // Width 4: skip row sum fixup - ".inst 0xc0904b40 // addha za0.s, p2/M, p2/M, z26.s\n" + ".inst 0xc0904b80 // addha za0.s, p2/M, p2/M, z28.s\n" "add x20, %x[qp], %[per_layer_mul]\n" - "ld1rw { z5.s }, p2/Z, [x20]\n" + "ld1rw { z11.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" - ".inst 0xc0904b41 // addha za1.s, p2/M, p2/M, z26.s\n" - "ld1rw { z4.s }, p2/Z, [x20]\n" + ".inst 0xc0904b81 // addha za1.s, p2/M, p2/M, z28.s\n" + "ld1rw { z7.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[c_offset]\n" "add x21, %x[qp], %[minval]\n" - ".inst 0xc0904b42 // addha za2.s, p2/M, p2/M, z26.s\n" + ".inst 0xc0904b82 // addha za2.s, p2/M, p2/M, z28.s\n" "ld1rw { z6.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[maxval]\n" - ".inst 0xc0904b43 // addha za3.s, p2/M, p2/M, z26.s\n" - "ld1rw { z21.s }, p2/Z, [x21]\n" - ".inst 0xc0062c1c // mova { z28.d-z31.d }, za.d[x9, #0]\n" - ".inst 0xc1a5ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z5.s\n" - ".inst 0xc0062c2c // mova { z12.d-z15.d }, za.d[x9, #1]\n" - ".inst 0xc1a5ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z5.s\n" - ".inst 0xc0062c40 // mova { z0.d-z3.d }, za.d[x9, #2]\n" - ".inst 0xc1a5ac00 // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z5.s\n" - ".inst 0xc0062c68 // mova { z8.d-z11.d }, za.d[x9, #3]\n" - ".inst 0xc1a5ac08 // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z5.s\n" - ".inst 0xc1a4aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z4.s\n" - "ld1rw { z16.s }, p2/Z, [x20]\n" - ".inst 0xc1a4aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z4.s\n" - ".inst 0xc1a4aa20 // srshl { z0.s-z3.s }, { z0.s-z3.s }, z4.s\n" - ".inst 0xc1a4aa28 // srshl { z8.s-z11.s }, { z8.s-z11.s }, z4.s\n" - ".inst 0xc1a6ab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z6.s\n" + ".inst 0xc0904b83 // addha za3.s, p2/M, p2/M, z28.s\n" + "ld1rw { z3.s }, p2/Z, [x21]\n" + ".inst 0xc0062c18 // mova { z24.d-z27.d }, za.d[x9, #0]\n" + ".inst 0xc1abac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z11.s\n" + ".inst 0xc0062c30 // mova { z16.d-z19.d }, za.d[x9, #1]\n" + ".inst 0xc1abac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z11.s\n" + ".inst 0xc0062c54 // mova { z20.d-z23.d }, za.d[x9, #2]\n" + ".inst 0xc1abac14 // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z11.s\n" + ".inst 0xc0062c6c // mova { z12.d-z15.d }, za.d[x9, #3]\n" + ".inst 0xc1abac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z11.s\n" + ".inst 0xc1a7aa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z7.s\n" + "ld1rw { z31.s }, p2/Z, [x20]\n" + ".inst 0xc1a7aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z7.s\n" + ".inst 0xc1a7aa34 // srshl { z20.s-z23.s }, { z20.s-z23.s }, z7.s\n" + ".inst 0xc1a7aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n" + ".inst 0xc1a6ab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n" + ".inst 0xc1a6ab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z6.s\n" + ".inst 0xc1a6ab14 // add { z20.s-z23.s }, { z20.s-z23.s }, z6.s\n" ".inst 0xc1a6ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n" - ".inst 0xc1a6ab00 // add { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n" - ".inst 0xc1a6ab08 // add { z8.s-z11.s }, { z8.s-z11.s }, z6.s\n" - ".inst 0xc1b0cebc // sclamp { z28.s-z31.s }, z21.s, z16.s\n" - ".inst 0xc1b0ceac // sclamp { z12.s-z15.s }, z21.s, z16.s\n" - "uzp1 z28.h, z28.h, z29.h\n" - ".inst 0xc1b0cea0 // sclamp { z0.s-z3.s }, z21.s, z16.s\n" - ".inst 0xc1b0cea8 // sclamp { z8.s-z11.s }, z21.s, z16.s\n" - "uzp1 z29.h, z30.h, z31.h\n" + ".inst 0xc1bfcc78 // sclamp { z24.s-z27.s }, z3.s, z31.s\n" + ".inst 0xc1bfcc70 // sclamp { z16.s-z19.s }, z3.s, z31.s\n" + "uzp1 z24.h, z24.h, z25.h\n" + ".inst 0xc1bfcc74 // sclamp { z20.s-z23.s }, z3.s, z31.s\n" + ".inst 0xc1bfcc6c // sclamp { z12.s-z15.s }, z3.s, z31.s\n" + "uzp1 z25.h, z26.h, z27.h\n" + "uzp1 z16.h, z16.h, z17.h\n" + "uzp1 z18.h, z18.h, z19.h\n" + "uzp1 z20.h, z20.h, z21.h\n" + "uzp1 z17.h, z22.h, z23.h\n" "uzp1 z12.h, z12.h, z13.h\n" - "uzp1 z13.h, z14.h, z15.h\n" - "uzp1 z0.h, z0.h, z1.h\n" - "uzp1 z1.h, z2.h, z3.h\n" - "uzp1 z8.h, z8.h, z9.h\n" - "uzp1 z9.h, z10.h, z11.h\n" - "uzp1 z28.b, z28.b, z29.b\n" - "st1b { z28.b }, p2, [x25]\n" - "uzp1 z12.b, z12.b, z13.b\n" - "st1b { z12.b }, p2, [x25, #1, MUL VL]\n" - "uzp1 z0.b, z0.b, z1.b\n" - "uzp1 z8.b, z8.b, z9.b\n" - "st1b { z0.b }, p2, [x25, #2, MUL VL]\n" - "st1b { z8.b }, p1, [x25, #3, MUL VL]\n" + "uzp1 z30.h, z14.h, z15.h\n" + "uzp1 z24.b, z24.b, z25.b\n" + "st1b { z24.b }, p2, [x25]\n" + "uzp1 z16.b, z16.b, z18.b\n" + "st1b { z16.b }, p2, [x25, #1, MUL VL]\n" + "uzp1 z20.b, z20.b, z17.b\n" + "uzp1 z12.b, z12.b, z30.b\n" + "st1b { z20.b }, p2, [x25, #2, MUL VL]\n" + "st1b { z12.b }, p1, [x25, #3, MUL VL]\n" "addvl x25, x25, #4\n" "43:" // Width 4: Output done "subs x27, x27, #0x4\n" @@ -665,7 +663,7 @@ void sme2_gemv_u8qa_dot_16VL ( "bgt 4b\n" "44:" // Exit ".inst 0xd503467f // SMSTOP\n" - "ptrue p2.b\n" + "ptrue p8.b\n" : [N] "+&r" (N), [flags] "+&r" (flags) : [A_ptr] "r" (A_ptr), [B_ptr] "r" (B_ptr), [K] "r" (K), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [output_ptr] "r" (output_ptr), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp) : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "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" @@ -674,5 +672,4 @@ void sme2_gemv_u8qa_dot_16VL ( } // namespace arm_gemm -#endif // ARM_COMPUTE_ENABLE_SME2 -#endif // __ARM_FEATURE_SVE +#endif // defined(ARM_COMPUTE_ENABLE_SME2) |