From cf87f509fc23d02c56569f794a3fb59e1b8be277 Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Tue, 2 Feb 2021 14:59:09 +0000 Subject: Tweak scheduling use of SQDMULH in quantized AVG pooling Resolves COMPMID-4195 Change-Id: Ie5116c1ddddccafba40432fd4b5245bb27890a88 Signed-off-by: Michele Di Giorgio Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4997 Reviewed-by: TeresaARM Reviewed-by: Manuel Bottini Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins --- .../generic.cpp | 136 ++++++++++----------- 1 file changed, 68 insertions(+), 68 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp index 20293c0e0e..84a6acf80d 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp @@ -42,7 +42,7 @@ void sve_fp16_nhwc_avg_generic_depthfirst_impl( __asm__ __volatile__( "ptrue p0.b\n" - "ld1rh { z7.h }, p0/Z, [%x[rescale_ptr]]\n" + "ld1rh { z8.h }, p0/Z, [%x[rescale_ptr]]\n" "mov x28, #0x0\n" "cnth x27\n" "cnth x26, ALL, MUL #2\n" @@ -53,23 +53,23 @@ void sve_fp16_nhwc_avg_generic_depthfirst_impl( "whilelt p0.h, x25, %x[n_channels]\n" "b.none 7f\n" "1:" // 4-vectors of channels - "mov z6.b, #0x0\n" + "mov z7.b, #0x0\n" "mov x19, %x[inptrs]\n" - "mov z5.b, #0x0\n" + "mov z6.b, #0x0\n" "lsr x24, %x[n_valid_cells], #0x2\n" + "mov z5.b, #0x0\n" "mov z4.b, #0x0\n" - "mov z3.b, #0x0\n" "cbz x24, 4f\n" "ldp x23, x22, [x19, #0x0]\n" "ldp x21, x20, [x19, #0x10]\n" "add x19, x19, #0x20\n" "subs x24, x24, #0x1\n" - "ld1h { z2.h }, p3/Z, [x23, x28, LSL #1]\n" - "ld1h { z1.h }, p3/Z, [x22, x28, LSL #1]\n" - "ld1h { z0.h }, p3/Z, [x21, x28, LSL #1]\n" - "ld1h { z31.h }, p3/Z, [x20, x28, LSL #1]\n" - "ld1h { z30.h }, p2/Z, [x23, x27, LSL #1]\n" - "ld1h { z22.h }, p2/Z, [x22, x27, LSL #1]\n" + "ld1h { z3.h }, p3/Z, [x23, x28, LSL #1]\n" + "ld1h { z2.h }, p3/Z, [x22, x28, LSL #1]\n" + "ld1h { z1.h }, p3/Z, [x21, x28, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x20, x28, LSL #1]\n" + "ld1h { z31.h }, p2/Z, [x23, x27, LSL #1]\n" + "ld1h { z30.h }, p2/Z, [x22, x27, LSL #1]\n" "ld1h { z29.h }, p2/Z, [x21, x27, LSL #1]\n" "ld1h { z28.h }, p2/Z, [x20, x27, LSL #1]\n" "ld1h { z27.h }, p1/Z, [x23, x26, LSL #1]\n" @@ -82,47 +82,47 @@ void sve_fp16_nhwc_avg_generic_depthfirst_impl( "ld1h { z16.h }, p0/Z, [x20, x25, LSL #1]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop - "fadd z23.h, z2.h, z1.h\n" + "fadd z23.h, z3.h, z2.h\n" "ldp x23, x22, [x19, #0x0]\n" "subs x24, x24, #0x1\n" - "fadd z19.h, z0.h, z31.h\n" + "fadd z19.h, z1.h, z0.h\n" "ldp x21, x20, [x19, #0x10]\n" - "fadd z22.h, z30.h, z22.h\n" "add x19, x19, #0x20\n" + "fadd z22.h, z31.h, z30.h\n" + "ld1h { z3.h }, p3/Z, [x23, x28, LSL #1]\n" "fadd z18.h, z29.h, z28.h\n" - "ld1h { z2.h }, p3/Z, [x23, x28, LSL #1]\n" "fadd z21.h, z27.h, z21.h\n" + "ld1h { z2.h }, p3/Z, [x22, x28, LSL #1]\n" "fadd z17.h, z26.h, z17.h\n" - "ld1h { z1.h }, p3/Z, [x22, x28, LSL #1]\n" + "ld1h { z1.h }, p3/Z, [x21, x28, LSL #1]\n" "fadd z20.h, z25.h, z20.h\n" - "ld1h { z0.h }, p3/Z, [x21, x28, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x20, x28, LSL #1]\n" "fadd z16.h, z24.h, z16.h\n" - "ld1h { z31.h }, p3/Z, [x20, x28, LSL #1]\n" + "ld1h { z31.h }, p2/Z, [x23, x27, LSL #1]\n" "fadd z19.h, z23.h, z19.h\n" - "ld1h { z30.h }, p2/Z, [x23, x27, LSL #1]\n" + "ld1h { z30.h }, p2/Z, [x22, x27, LSL #1]\n" "fadd z18.h, z22.h, z18.h\n" - "ld1h { z22.h }, p2/Z, [x22, x27, LSL #1]\n" - "fadd z17.h, z21.h, z17.h\n" "ld1h { z29.h }, p2/Z, [x21, x27, LSL #1]\n" - "fadd z16.h, z20.h, z16.h\n" + "fadd z17.h, z21.h, z17.h\n" "ld1h { z28.h }, p2/Z, [x20, x27, LSL #1]\n" - "fadd z6.h, z6.h, z19.h\n" + "fadd z16.h, z20.h, z16.h\n" "ld1h { z27.h }, p1/Z, [x23, x26, LSL #1]\n" - "fadd z5.h, z5.h, z18.h\n" + "fadd z7.h, z7.h, z19.h\n" "ld1h { z21.h }, p1/Z, [x22, x26, LSL #1]\n" - "fadd z4.h, z4.h, z17.h\n" + "fadd z6.h, z6.h, z18.h\n" "ld1h { z26.h }, p1/Z, [x21, x26, LSL #1]\n" - "fadd z3.h, z3.h, z16.h\n" + "fadd z5.h, z5.h, z17.h\n" "ld1h { z17.h }, p1/Z, [x20, x26, LSL #1]\n" + "fadd z4.h, z4.h, z16.h\n" "ld1h { z25.h }, p0/Z, [x23, x25, LSL #1]\n" "ld1h { z20.h }, p0/Z, [x22, x25, LSL #1]\n" "ld1h { z24.h }, p0/Z, [x21, x25, LSL #1]\n" "ld1h { z16.h }, p0/Z, [x20, x25, LSL #1]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail - "fadd z23.h, z2.h, z1.h\n" - "fadd z19.h, z0.h, z31.h\n" - "fadd z22.h, z30.h, z22.h\n" + "fadd z23.h, z3.h, z2.h\n" + "fadd z19.h, z1.h, z0.h\n" + "fadd z22.h, z31.h, z30.h\n" "fadd z18.h, z29.h, z28.h\n" "fadd z21.h, z27.h, z21.h\n" "fadd z17.h, z26.h, z17.h\n" @@ -132,37 +132,37 @@ void sve_fp16_nhwc_avg_generic_depthfirst_impl( "fadd z18.h, z22.h, z18.h\n" "fadd z17.h, z21.h, z17.h\n" "fadd z16.h, z20.h, z16.h\n" - "fadd z6.h, z6.h, z19.h\n" - "fadd z5.h, z5.h, z18.h\n" - "fadd z4.h, z4.h, z17.h\n" - "fadd z3.h, z3.h, z16.h\n" + "fadd z7.h, z7.h, z19.h\n" + "fadd z6.h, z6.h, z18.h\n" + "fadd z5.h, z5.h, z17.h\n" + "fadd z4.h, z4.h, z16.h\n" "4:" // 4-vectors of channels: After loop "ands x20, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop "ldr x23, [x19], #0x8\n" "subs x20, x20, #0x1\n" - "ld1h { z2.h }, p3/Z, [x23, x28, LSL #1]\n" - "fadd z6.h, z6.h, z2.h\n" - "ld1h { z30.h }, p2/Z, [x23, x27, LSL #1]\n" + "ld1h { z3.h }, p3/Z, [x23, x28, LSL #1]\n" + "fadd z7.h, z7.h, z3.h\n" + "ld1h { z31.h }, p2/Z, [x23, x27, LSL #1]\n" "ld1h { z27.h }, p1/Z, [x23, x26, LSL #1]\n" - "fadd z5.h, z5.h, z30.h\n" + "fadd z6.h, z6.h, z31.h\n" "ld1h { z25.h }, p0/Z, [x23, x25, LSL #1]\n" - "fadd z4.h, z4.h, z27.h\n" - "fadd z3.h, z3.h, z25.h\n" + "fadd z5.h, z5.h, z27.h\n" + "fadd z4.h, z4.h, z25.h\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - "fmul z6.h, z6.h, z7.h\n" - "st1h { z6.h }, p3, [%x[outptr], x28, LSL #1]\n" - "fmul z5.h, z5.h, z7.h\n" + "fmul z7.h, z7.h, z8.h\n" + "st1h { z7.h }, p3, [%x[outptr], x28, LSL #1]\n" + "fmul z6.h, z6.h, z8.h\n" "inch x28, ALL, MUL #4\n" - "fmul z4.h, z4.h, z7.h\n" - "st1h { z5.h }, p2, [%x[outptr], x27, LSL #1]\n" - "fmul z3.h, z3.h, z7.h\n" + "fmul z5.h, z5.h, z8.h\n" + "st1h { z6.h }, p2, [%x[outptr], x27, LSL #1]\n" + "fmul z4.h, z4.h, z8.h\n" "inch x27, ALL, MUL #4\n" - "st1h { z4.h }, p1, [%x[outptr], x26, LSL #1]\n" + "st1h { z5.h }, p1, [%x[outptr], x26, LSL #1]\n" "inch x26, ALL, MUL #4\n" - "st1h { z3.h }, p0, [%x[outptr], x25, LSL #1]\n" + "st1h { z4.h }, p0, [%x[outptr], x25, LSL #1]\n" "inch x25, ALL, MUL #4\n" "whilelt p0.h, x25, %x[n_channels]\n" "b.any 1b\n" @@ -170,50 +170,50 @@ void sve_fp16_nhwc_avg_generic_depthfirst_impl( "whilelt p3.h, x28, %x[n_channels]\n" "b.none 14f\n" "8:" // Single vector of channels: Loop - "mov z6.b, #0x0\n" + "mov z7.b, #0x0\n" "mov x19, %x[inptrs]\n" "lsr x24, %x[n_valid_cells], #0x2\n" "cbz x24, 11f\n" "ldp x23, x22, [x19, #0x0]\n" - "subs x24, x24, #0x1\n" "ldp x21, x20, [x19, #0x10]\n" "add x19, x19, #0x20\n" - "ld1h { z2.h }, p3/Z, [x23, x28, LSL #1]\n" - "ld1h { z1.h }, p3/Z, [x22, x28, LSL #1]\n" - "ld1h { z0.h }, p3/Z, [x21, x28, LSL #1]\n" - "ld1h { z31.h }, p3/Z, [x20, x28, LSL #1]\n" + "subs x24, x24, #0x1\n" + "ld1h { z3.h }, p3/Z, [x23, x28, LSL #1]\n" + "ld1h { z2.h }, p3/Z, [x22, x28, LSL #1]\n" + "ld1h { z1.h }, p3/Z, [x21, x28, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x20, x28, LSL #1]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop - "fadd z23.h, z2.h, z1.h\n" + "fadd z23.h, z3.h, z2.h\n" "ldp x23, x22, [x19, #0x0]\n" "subs x24, x24, #0x1\n" - "fadd z19.h, z0.h, z31.h\n" + "fadd z19.h, z1.h, z0.h\n" "ldp x21, x20, [x19, #0x10]\n" - "fadd z19.h, z23.h, z19.h\n" - "ld1h { z2.h }, p3/Z, [x23, x28, LSL #1]\n" - "fadd z6.h, z6.h, z19.h\n" - "ld1h { z1.h }, p3/Z, [x22, x28, LSL #1]\n" "add x19, x19, #0x20\n" - "ld1h { z0.h }, p3/Z, [x21, x28, LSL #1]\n" - "ld1h { z31.h }, p3/Z, [x20, x28, LSL #1]\n" + "fadd z19.h, z23.h, z19.h\n" + "ld1h { z3.h }, p3/Z, [x23, x28, LSL #1]\n" + "ld1h { z2.h }, p3/Z, [x22, x28, LSL #1]\n" + "fadd z7.h, z7.h, z19.h\n" + "ld1h { z1.h }, p3/Z, [x21, x28, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x20, x28, LSL #1]\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail - "fadd z23.h, z2.h, z1.h\n" - "fadd z19.h, z0.h, z31.h\n" + "fadd z23.h, z3.h, z2.h\n" + "fadd z19.h, z1.h, z0.h\n" "fadd z19.h, z23.h, z19.h\n" - "fadd z6.h, z6.h, z19.h\n" + "fadd z7.h, z7.h, z19.h\n" "11:" // Single vector of channels: Loop: After loop "ands x20, %x[n_valid_cells], #0x3\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop "ldr x23, [x19], #0x8\n" "subs x20, x20, #0x1\n" - "ld1h { z2.h }, p3/Z, [x23, x28, LSL #1]\n" - "fadd z6.h, z6.h, z2.h\n" + "ld1h { z3.h }, p3/Z, [x23, x28, LSL #1]\n" + "fadd z7.h, z7.h, z3.h\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - "fmul z6.h, z6.h, z7.h\n" - "st1h { z6.h }, p3, [%x[outptr], x28, LSL #1]\n" + "fmul z7.h, z7.h, z8.h\n" + "st1h { z7.h }, p3, [%x[outptr], x28, LSL #1]\n" "inch x28\n" "whilelt p3.h, x28, %x[n_channels]\n" "b.any 8b\n" @@ -221,7 +221,7 @@ void sve_fp16_nhwc_avg_generic_depthfirst_impl( : : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [n_valid_cells] "r" (n_valid_cells), [outptr] "r" (outptr), [rescale_ptr] "r" (&rescale_value) - : "cc", "memory", "p0", "p1", "p2", "p3", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } -- cgit v1.2.1