diff options
Diffstat (limited to 'chapters/ewise_binary.adoc')
| -rw-r--r-- | chapters/ewise_binary.adoc | 302 |
1 files changed, 150 insertions, 152 deletions
diff --git a/chapters/ewise_binary.adoc b/chapters/ewise_binary.adoc index e9d76f8..2b8d321 100644 --- a/chapters/ewise_binary.adoc +++ b/chapters/ewise_binary.adoc @@ -28,14 +28,13 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = apply_add<in_t>(value1, value2) - tensor_write<in_t>(output, shape, index, acc) -} +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = apply_add<in_t>(value1, value2); + tensor_write<in_t>(output, shape, index, acc); ---- *Supported Data Types:* @@ -43,8 +42,8 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 32|int32 -|MI, MT|float|float +|Any|signed 32|int32_t +|MI, MT|floating-point|float_t |=== ==== ARITHMETIC_RIGHT_SHIFT @@ -59,7 +58,7 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match |Input|in_t*|input1|shape1|Input tensor |Input|in_t*|input2|shape2|Input tensor with the same rank as input1 -|Input|bool |round |- | If true then the shift is rounded +|Input|bool_t |round |- | If true then the shift is rounded |Output|in_t*|output|shape|Output tensor with broadcast shape if necessary |=== @@ -67,17 +66,17 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - assert(0 <= value2 && value2 <= 31) - in_t acc = value1 >> value2 - if (round==true && value2>0 && (value1>>(value2-1))&1!=0) { +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + assert(0 <= value2 && value2 <= 31); + in_t acc = value1 >> value2; + if (round == true && value2 > 0 && (value1 >> (value2 - 1)) & 1 != 0) { acc = acc + 1; } - acc = apply_clip(acc, minimum<in_t>, maximum<in_t>) + acc = apply_clip<in_t>(acc, minimum<in_t>, maximum<in_t>) tensor_write<in_t>(output, shape, index, acc) } ---- @@ -87,9 +86,9 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 8|int8 -|Any|signed 16|int16 -|Any|signed 32|int32 +|Any|signed 8|int8_t +|Any|signed 16|int16_t +|Any|signed 32|int32_t |=== ==== BITWISE_AND @@ -111,13 +110,13 @@ Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = value1 & value2 - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = value1 & value2; + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -126,9 +125,9 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 8|int8 -|Any|signed 16|int16 -|Any|signed 32|int32 +|Any|signed 8|int8_t +|Any|signed 16|int16_t +|Any|signed 32|int32_t |=== ==== BITWISE_OR @@ -150,13 +149,13 @@ Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = value1 | value2 - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = value1 | value2; + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -165,9 +164,9 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 8|int8 -|Any|signed 16|int16 -|Any|signed 32|int32 +|Any|signed 8|int8_t +|Any|signed 16|int16_t +|Any|signed 32|int32_t |=== ==== BITWISE_XOR @@ -189,13 +188,13 @@ Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = value1 ^ value2 - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = value1 ^ value2; + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -204,9 +203,9 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 8|int8 -|Any|signed 16|int16 -|Any|signed 32|int32 +|Any|signed 8|int8_t +|Any|signed 16|int16_t +|Any|signed 32|int32_t |=== ==== LOGICAL_AND @@ -232,13 +231,13 @@ None [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = value1 && value2 - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = value1 && value2; + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -247,7 +246,7 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|Bool|Bool +|Any|Bool|bool_t |=== ==== LOGICAL_LEFT_SHIFT @@ -269,14 +268,14 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - assert(0 <= value2 && value2 <= 31) - in_t acc = value1 << value2 - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + assert(0 <= value2 && value2 <= 31); + in_t acc = value1 << value2; + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -285,9 +284,9 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 8|int8 -|Any|signed 16|int16 -|Any|signed 32|int32 +|Any|signed 8|int8_t +|Any|signed 16|int16_t +|Any|signed 32|int32_t |=== ==== LOGICAL_RIGHT_SHIFT @@ -309,14 +308,14 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - assert(0 <= value2 && value2 <= 31) - in_t acc = (unsigned in_t)value1 >> value2 - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + assert(0 <= value2 && value2 <= 31); + in_t acc = (unsigned in_t)value1 >> value2; + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -325,9 +324,9 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 8|int8 -|Any|signed 16|int16 -|Any|signed 32|int32 +|Any|signed 8|int8_t +|Any|signed 16|int16_t +|Any|signed 32|int32_t |=== ==== LOGICAL_OR @@ -349,13 +348,13 @@ Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = value1 || value2 - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = value1 || value2; + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -364,7 +363,7 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|Bool|Bool +|Any|Bool|bool_t |=== ==== LOGICAL_XOR @@ -377,7 +376,7 @@ Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. |=== |Argument|Type|Name|Shape|Description -|Input|in_t*|input1|shape1|Input tensor from 1 to 4 dims +|Input|in_t*|input1|shape1|Input tensor |Input|in_t*|input2|shape2|Input tensor with the same rank as input1 |Output|in_t*|output|shape|Output tensor of same type as the input tensors, with broadcast shape if necessary |=== @@ -386,13 +385,13 @@ Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = value1 != value2 - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = value1 != value2; + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -401,7 +400,7 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|Bool|Bool +|Any|Bool|bool_t |=== ==== MAXIMUM @@ -423,13 +422,13 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = apply_max(value1, value2) - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = apply_max(value1, value2); + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -438,8 +437,8 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 32|int32 -|MI, MT|float|float +|Any|signed 32|int32_t +|MI, MT|floating-point|float_t |=== ==== MINIMUM @@ -461,13 +460,13 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = apply_min(value1, value2) - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = apply_min(value1, value2); + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -476,8 +475,8 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 32|int32 -|MI, MT|float|float +|Any|signed 32|int32_t +|MI, MT|floating-point|float_t |=== ==== MUL @@ -492,7 +491,7 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match |Input|in_t*|input1|shape1|Input tensor |Input|in_t*|input2|shape2|Input tensor with the same rank as input1 -|Input (MT profile) Attribute (BI/MI profiles)|uint6_t|shift|-|Result right shift (int32 data type only) +|Input (MT profile) Attribute (BI/MI profiles)|uint6_t|shift|-|Result right shift (int32_t data type only) |Output|out_t*|output|shape|Output tensor with broadcast shape if necessary |=== @@ -500,18 +499,18 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match [source,c] ---- -assert(in_t==int32_t || shift==0); -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) +assert(in_t == int32_t || shift == 0); +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); if (shift>0) { - out_t acc = apply_scale_32(value1, value2, shift) + out_t acc = apply_scale_32(value1, value2, shift); } else { out_t acc = value1 * value2; // low 32-bits of result for int32_t } - tensor_write<out_t>(output, shape, index, acc) + tensor_write<out_t>(output, shape, index, acc); } ---- @@ -519,10 +518,10 @@ for_each (index in shape) { |=== |Profile|Mode|in_t|out_t -|Any|signed 8|int8|int32 -|Any|signed 16|int16|int32 -|Any|signed 32|int32|int32 -|MI, MT|float|float|float +|Any|signed 8|int8_t|int32_t +|Any|signed 16|int16_t|int32_t +|Any|signed 32|int32_t|int32_t +|MI, MT|floating-point|float_t|float_t |=== ==== POW @@ -542,14 +541,14 @@ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match *Quantization Parameters:* -Only supported with floating point values. +Only supported with floating-point values. *Supported Data Types:* |=== |Profile|Mode|in_t -|MI, MT|float|float +|MI, MT|floating-point|float_t |=== ==== SUB @@ -571,13 +570,13 @@ Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. [source,c] ---- -for_each (index in shape) { - index1 = apply_broadcast(shape, shape1, index) - index2 = apply_broadcast(shape, shape2, index) - in_t value1 = tensor_read<in_t>(input1, shape1, index1) - in_t value2 = tensor_read<in_t>(input2, shape2, index2) - in_t acc = apply_sub<out_t>(value1, value2); - tensor_write<in_t>(output, shape, index, acc) +for_each(index in shape) { + index1 = apply_broadcast(shape, shape1, index); + index2 = apply_broadcast(shape, shape2, index); + in_t value1 = tensor_read<in_t>(input1, shape1, index1); + in_t value2 = tensor_read<in_t>(input2, shape2, index2); + in_t acc = apply_sub<in_t>(value1, value2); + tensor_write<in_t>(output, shape, index, acc); } ---- @@ -586,25 +585,25 @@ for_each (index in shape) { |=== |Profile|Mode|in_t -|Any|signed 32|int32 -|MI, MT|float|float +|Any|signed 32|int32_t +|MI, MT|floating-point|float_t |=== ==== TABLE -Interpolated table lookup operation. The int16 input is treated as a fixed-point 9.7 value. The high 9 bits are used to index into the table. The fractional 7 bits are used to interpolate based on table[index] and table[index+1]. The TABLE operator returns a 16.7 interpolated value which can then be input to the RESCALE operator to scale to the required output data type. Note that table has 513 values to handle table[index+1] when index=511. +Interpolated table lookup operation. The int16_t input is treated as a fixed-point 9.7 value. The high 9 bits are used to index into the table. The fractional 7 bits are used to interpolate based on table[index] and table[index+1]. The TABLE operator returns a 16.7 interpolated value which can then be input to the RESCALE operator to scale to the required output data type. Note that table has 513 values to handle table[index+1] when index=511. An int8_t to int8_t table lookup can be constructed in TOSA as follows: -* Use RESCALE (in_t=int8, out_t=int16, input_zp=0, scale=1<<14, shift=7) to perform a shift left of 7 and convert to int16 +* Use RESCALE (in_t=int8_t, out_t=int16_t, input_zp=0, scale=1<<14, shift=7) to perform a shift left of 7 and convert to int16_t * Use the TABLE operator to produce a fixed point 16.7 result. The lower 7 bits will be zero and only the central 256 table entries will be used. -* Use RESCALE (in_t=int32, out_t=int8, scale=1<<14, shift=28) to scale the output to int8_t range (or alternate scale as required) +* Use RESCALE (in_t=int32_t, out_t=int8_t, scale=1<<14, shift=28) to scale the output to int8_t range (or alternate scale as required) * Note that this TOSA sequence can be implemented in software as a 256 entry 8-bit lookup table. An int16_t to int16_t table lookup can be constructed in TOSA as follows: * Use the TABLE operator to produce a fixed point 16.7 interpolated result -* Use RESCALE (in_t=int32, out_t=int16, scale=1<<14, shift=21) to scale the output to int16_t range (or alternate scale as required) +* Use RESCALE (in_t=int32_t, out_t=int16_t, scale=1<<14, shift=21) to scale the output to int16_t range (or alternate scale as required) *Arguments:* @@ -624,11 +623,10 @@ None [source,c] ---- -assert(rank(shape)<=4) -for_each (index in shape) { - in_t value = tensor_read<in_t>(input, shape, index) - out_t acc = apply_lookup(table, value) - tensor_write<out_t>(output, shape, index, acc) +for_each(index in shape) { + in_t value = tensor_read<in_t>(input, shape, index); + out_t acc = apply_lookup(table, value); + tensor_write<out_t>(output, shape, index, acc); } ---- @@ -637,6 +635,6 @@ for_each (index in shape) { |=== |Profile|Mode|in_t|table_t|out_t -|Any|signed 16|int16|int16|int32 +|Any|signed 16|int16_t|int16_t|int32_t |=== |