uncomplicate.diamond.internal.dnnl.core — org.uncomplicate/deep-diamond-dnnl 0.34.0

arg-md clj

(arg-md pd arg)

Queries the primitive descriptor pd for the argument's memory descriptor.

args clj

(args arg-map)

batch-norm-bwd clj

(batch-norm-bwd eng hint-fwd-pd prop-kind diff-data-desc data-desc & flags)

TODO

batch-norm-bwd-args clj

(batch-norm-bwd-args diff-dst src mean variance diff-src)

(batch-norm-bwd-args diff-dst src scale shift mean variance diff-src diff-scale)

(batch-norm-bwd-args diff-dst
                     src
                     scale
                     shift
                     mean
                     variance
                     diff-src
                     diff-scale
                     workspace)

batch-norm-fwd clj

(batch-norm-fwd eng prop-kind data-desc & flags)

TODO

batch-norm-fwd-args clj

(batch-norm-fwd-args src-and-dst)

(batch-norm-fwd-args src dst)

(batch-norm-fwd-args src dst mean variance)

(batch-norm-fwd-args src dst scale shift _)

(batch-norm-fwd-args src dst scale shift mean variance)

(batch-norm-fwd-args src dst scale shift mean variance workspace)

binary clj

(binary eng alg-kind src-dst-desc)

(binary eng alg-kind src-dst-desc src1-desc)

(binary eng alg-kind src0-desc src1-desc dst-desc)

DNNL binary operation. NOTE: much slower than Neanderthal add or mul. Use only when can't avoid it.

binary-args clj

(binary-args src-and-dst src1)

(binary-args src0 src1 dst)

bwd-args clj

(bwd-args diff-dst weights diff-src)

(bwd-args src diff-dst diff-weights diff-bias)

concatenate clj

(concatenate eng dst concat-dimension src)

(concatenate eng dst concat-dimension src & srcs)

TODO

convolution-bwd-data clj

(convolution-bwd-data eng
                      hint-fwd-pd
                      alg-kind
                      diff-src-desc
                      weights-desc
                      diff-dst-desc
                      strides
                      padding)

(convolution-bwd-data eng
                      hint-fwd-pd
                      alg-kind
                      diff-src-desc
                      weights-desc
                      diff-dst-desc
                      strides
                      dilates
                      padding)

(convolution-bwd-data eng
                      hint-fwd-pd
                      alg-kind
                      diff-src-desc
                      weights-desc
                      diff-dst-desc
                      strides
                      dilates
                      padding-l
                      padding-r)

TODO

convolution-bwd-weights clj

(convolution-bwd-weights eng
                         hint-fwd-pd
                         alg-kind
                         src-desc
                         diff-weights-desc
                         diff-bias-desc
                         diff-dst-desc
                         strides
                         padding)

(convolution-bwd-weights eng
                         hint-fwd-pd
                         alg-kind
                         src-desc
                         diff-weights-desc
                         diff-bias-desc
                         diff-dst-desc
                         strides
                         dilates
                         padding)

(convolution-bwd-weights eng
                         hint-fwd-pd
                         alg-kind
                         src-desc
                         diff-weights-desc
                         diff-bias-desc
                         diff-dst-desc
                         strides
                         dilates
                         padding-l
                         padding-r)

TODO

convolution-fwd clj

(convolution-fwd eng
                 prop-kind
                 alg-kind
                 src-desc
                 weights-desc
                 bias-desc
                 dst-desc
                 strides
                 padding)

(convolution-fwd eng
                 prop-kind
                 alg-kind
                 src-desc
                 weights-desc
                 bias-desc
                 dst-desc
                 strides
                 dilates
                 padding)

(convolution-fwd eng
                 prop-kind
                 alg-kind
                 src-desc
                 weights-desc
                 bias-desc
                 dst-desc
                 strides
                 dilates
                 padding-l
                 padding-r)

TODO

data-type clj

(data-type mem-desc)

Queries the data type of a memory descriptor.

diff-dst-md clj

(diff-dst-md pd)

Queries the primitive descriptor pd for the gradient of the destination (output).

diff-src-md clj

(diff-src-md pd)

Queries the primitive descriptor pd for the gradient of the source (input).

diff-weights-md clj

(diff-weights-md pd)

Queries the primitive descriptor pd for the gradient of the weights.

dims clj

(dims mem-desc)

Queries the dimensions of a memory descriptor.

dnnl-contiguous-desc clj

(dnnl-contiguous-desc md)

dst-md clj

(dst-md pd)

Queries the primitive descriptor pd for the destination (output).

eltwise-bwd clj

(eltwise-bwd eng hint-fwd-pd alg-kind diff-desc src-desc)

(eltwise-bwd eng hint-fwd-pd alg-kind diff-desc src-desc alpha beta)

Creates a backward descriptor of an operation that is applied to every element of a tensor. Used only during the training.

• alg-kind: operation algorithm, such as :relu or :logistic (defined in [[constants/dnnl-eltwise-alg-kind]])
• diff-desc: the diff memory descriptor
• src-desc: the source memory descriptor
• dst-desc: the destination memory descriptor
• alpha, and beta: optional coefficients, depending on alg-kind.

eltwise-bwd-args clj

(eltwise-bwd-args src diff-dst diff-src)

Creates DNNL's data structure that holds arguments as required by elementwise operations.

eltwise-fwd clj

(eltwise-fwd eng prop-kind alg-kind mem-desc)

(eltwise-fwd eng prop-kind alg-kind mem-desc alpha beta)

Creates a forward descriptor of an operation that is applied to every element of a tensor.

• prop-kind: the kind of propagation: :inference, training, or :scoring (defined in [[constants/dnnl-forward-prop-kind]])
• alg-kind: operation algorithm, such as :relu or :logistic (defined in [[constants/dnnl-eltwise-alg-kind]])
• mem-desc: the descriptor that defines memory layout of the data
• alpha, and beta: optional coefficients, depending on alg-kind.

engine clj

(engine)

(engine id)

(engine id kind)

Creates an engine for the device id of the specified keyword kind.

Supported engine kinds are :cpu, :gpu, and :any. The default kind is :cpu. Engine has to be released.

Throws an ExceptionInfo if the id does not correspond to a physical device or if kind is not supported.

engine-count clj

(engine-count)

(engine-count kind)

Returns the number of physical engines of the specified kind (:cpu, :gpu, :any).

Throws an ExceptionInfo if kind is not supported.

engine-kind clj

(engine-kind eng)

Returns engine's kind as a keyword. Typical values are :gpu and :cpu.

Throws an ExceptionInfo if kind is not supported.

equal-desc? clj

(equal-desc? x y)

Compares two memory descriptors for logical equality.

Two descriptors may be equal even though the objects are not equal nor identical in the JVM sense.

execute! clj

(execute! strm p args)

Queues the operation primitive p for execution in stream strm.

Returns strm. Throws an ExceptionInfo if the DNNL stream is not valid, or the primitive cannot be executed.

extend-memory-desc-info clj macro

(extend-memory-desc-info t)

fwd-args clj

(fwd-args src-and-dst)

(fwd-args src dst)

(fwd-args src dst workspace)

(fwd-args src weights bias dst)

Creates DNNL's data structure that holds arguments as required by forward operations.

get-engine clj

(get-engine mem)

Returns the engine context of the memory object mem.

gru-bwd clj

(gru-bwd eng
         hint-fwd-pd
         direction
         src-desc
         src-iter-desc
         weights-desc
         weights-iter-desc
         bias-desc
         dst-desc
         dst-iter-desc
         diff-src-desc
         diff-src-iter-desc
         diff-weights-desc
         diff-weights-iter-desc
         diff-bias-desc
         diff-dst-desc
         diff-dst-iter-desc)

TODO

gru-fwd clj

(gru-fwd eng
         prop-kind
         direction
         src-desc
         src-iter-desc
         weights-desc
         weights-iter-desc
         bias-desc
         dst-desc
         dst-iter-desc)

TODO

inner-product-bwd clj

(inner-product-bwd eng hint-fwd-pd diff-src-desc weights-desc diff-dst-desc)

(inner-product-bwd eng
                   hint-fwd-pd
                   src-desc
                   diff-weights-desc
                   diff-bias-desc
                   diff-dst-desc)

Creates a descriptor for the backward phase of the inner product operation, for data (3-arguments) weights (5-arguments) updates.

• The gradient of data computes diff-src <- f(weights, diff-dst): diff-src-desc: descriptor of the source gradient (input) memory. weights-desc: descriptor of the weights memory. diff-dst-desc: descriptor of the destination gradient (output) memory.

• The gradient of data computes diff-weights <- f(diff-dst, src), and diff-bias <- f(diff-dst, src): src-desc: descriptor of the source (input) memory. diff-weights-desc: descriptor of the weights gradient memory. diff-bias-desc: descriptor of the bias gradient memory. diff-dst-desc: descriptor of the destination gradient (output) memory.

inner-product-fwd clj

(inner-product-fwd eng prop-kind src-desc weights-desc bias-desc dst-desc)

Creates a descriptor for the forward phase of the inner product operation, which computes dst <- src * weights + bias.

prop-kind: one of the values defined in constants/dnnl-forward-prop-kind (:inference, :training, :scoring). src-desc: descriptor of the source (input) memory. weights-desc: descriptor of the weights memory. bias-desc: descriptor of the bias memory. dst-desc: descripror of the destination (output) memory.

lstm-bwd clj

(lstm-bwd eng
          hint-fwd-pd
          direction
          src-desc
          src-iter-desc
          weights-desc
          weights-iter-desc
          bias-desc
          dst-desc
          dst-iter-desc
          diff-src-desc
          diff-src-iter-desc
          diff-weights-desc
          diff-weights-iter-desc
          diff-bias-desc
          diff-dst-desc
          diff-dst-iter-desc)

(lstm-bwd eng
          hint-fwd-pd
          direction
          src-desc
          src-iter-desc
          src-iter-c-desc
          weights-iter-peephole-projection
          bias-desc
          dst-desc
          dst-iter-desc
          dst-iter-c-desc
          diff-src-desc
          diff-src-iter-desc
          diff-src-iter-c-desc
          diff-weights-iter-peephole-projection
          diff-bias-desc
          diff-dst-desc
          diff-dst-iter-desc
          diff-dst-iter-c-desc)

TODO

lstm-fwd clj

(lstm-fwd eng
          prop-kind
          direction
          src-desc
          src-iter-desc
          weights-desc
          weights-iter-desc
          bias-desc
          dst-desc
          dst-iter-desc)

(lstm-fwd eng
          prop-kind
          direction
          src-desc
          src-iter-desc
          src-iter-c-desc
          weights-desc
          weights-iter-desc
          weights-peephole-desc
          weights-projection-desc
          bias-desc
          dst-desc
          dst-iter-desc
          dst-iter-c-desc)

TODO

memory clj

(memory eng mem-desc)

(memory eng mem-desc buf)

(memory eng mem-desc buf master)

An engine-specific memory handle for a raw buffer and a matching descriptor.

eng a DNNL engine that controls the context. mem-desc logical memory descriptor. buf JavaCPP pointer instance. master indicates whether this memory object handles the life cycle of buf.

memory-desc clj

(memory-desc)

(memory-desc dims)

(memory-desc dims format)

(memory-desc dims data-type format)

Creates an engine-agnostic, logical, description of data, based on dimensions, data type and data format.

dims is a Clojure vector of positive numbers representing dimensions in the :abcdef format, regardless of the physical layout of dimensions. data-type is a keyword that specifies one of the supported types of data, defined in [[constants/dnnl-data-type]] (:float, :int, etc.) format specifies an (optional) physical layout as a keyword, choosing one of [[constants/dnnl-format]] (:nchw, :acdeb, :any, etc.), or through strides specified as a Clojure vector of positive numbers that match logical dimensions.

Examples:

(memory-desc [2 3] :float :nc)

(memory-desc [2 3 4 5] :float [120 3 4 5])

multi-args clj

(multi-args src-and-dst)

(multi-args dst src)

(multi-args dst src0 src1)

(multi-args dst src0 src1 & srcs)

Creates DNNL's data structure that holds arguments for various operations that accept one destination and one or multiple sources.

ndims clj

(ndims mem-desc)

Queries the number of dimensions of a memory descriptor.

offset clj

(offset mem)

Gets the starting position in the buffer that the memory object mem controls.

offset! clj

(offset! mem n)

Sets the starting position in the buffer that the memory object mem controls.

pooling-bwd clj

(pooling-bwd eng
             hint-fwd-pd
             alg-kind
             diff-src-desc
             diff-dst-desc
             kernel
             strides
             padding)

(pooling-bwd eng
             hint-fwd-pd
             alg-kind
             diff-src-desc
             diff-dst-desc
             kernel
             strides
             dilates
             padding)

(pooling-bwd eng
             hint-fwd-pd
             alg-kind
             diff-src-desc
             diff-dst-desc
             kernel
             strides
             dilates
             padding-l
             padding-r)

TODO

pooling-bwd-args clj

(pooling-bwd-args diff-dst diff-src workspace)

TODO

pooling-fwd clj

(pooling-fwd eng prop-kind alg-kind src-desc dst-desc kernel strides padding)

(pooling-fwd eng
             prop-kind
             alg-kind
             src-desc
             dst-desc
             kernel
             strides
             dilates
             padding)

(pooling-fwd eng
             prop-kind
             alg-kind
             src-desc
             dst-desc
             kernel
             strides
             dilates
             padding-l
             padding-r)

TODO

primitive clj

(primitive pd)

Creates a primitive from the primitive descriptor pd.

Primitive encapsulates a pre-generated computation optimized for particular data shapes defined in the primitive descriptor. Usually, such primitive is executed many times with the data of these shapes, while the preparation cost is paid only at the time of creation.

Primitive is a function with execution context (state). In addition to immutable state such as input and output shape and data type, it could require a mutable temporary work memory buffer that is called scratchpad in DNNL terminology.

For more info about DNNL's concepts, see the official DNNL guide.

primitive-cache-capacity clj

(primitive-cache-capacity)

primitive-cache-capacity! clj

(primitive-cache-capacity! n)

query-md clj

(query-md pd what)

(query-md pd what index)

Queries the primitive descriptor pd for the property what and (optional) index index.

reduction clj

(reduction eng alg-kind src dst)

(reduction eng alg-kind src dst p epsilon)

TODO

reorder clj

(reorder eng input output)

(reorder input-eng input output-eng output)

Copies data across engines, between physical memory formats, keeping the logical structure of the tensor.

softmax-bwd clj

(softmax-bwd eng hint-fwd-pd alg-kind diff-desc src-desc axis)

TODO

softmax-bwd-args clj

(softmax-bwd-args dst diff-dst diff-src)

(softmax-bwd-args src dst diff-dst diff-src)

Creates DNNL's data structure that holds arguments as required by softmax fwd and bwd operations.

softmax-fwd clj

(softmax-fwd eng prop-kind alg-kind mem-desc axis)

TODO

src-md clj

(src-md pd)

Queries the primitive descriptor pd for the source (input).

stream clj

(stream eng & flags)

Creates a stream for executing primitive operations for engine eng.

Stream execution can be further specified by flags, defined in the constants/dnnl-stream-flags. Stream has to be released.

strides clj

(strides mem-desc)

Queries the strides of a memory descriptor.

submemory-desc clj

(submemory-desc parent-desc dim)

(submemory-desc parent-desc dims offsets)

Creates a (sub)memory section of a memory object, using the specified shape dims, and offsets vectors.

sum! clj

(sum! eng scale dst)

(sum! eng dst scale src & scale-srcs)

Scales a single dst, or sums scaled entries of more tensors elementwise.

This operation changes dst. All sources and destinations have to be of the same shape.

BEWARE: if dst and one of the srcs are identical, this source has to be the first src argument, due to how DNNL algorithm works internally, or result would be incorrect!

eng: the computing context engine scale: a floating point scale for the first source

If only a single tensor is provided, computes dst = scale * dst. dst: the source and destination tensor

Otherwise, computes dst = scale * src + scale-srcs[0] * scale-srcs[1] etc. dst: the source and destination tensor src: the first source tensor scale-srcs: a sequence of scale1, src1, scale2, src2, etc.

Example: (sum eng md 2.0 md 3.0 md)

vanilla-rnn-bwd clj

(vanilla-rnn-bwd eng
                 hint-fwd-pd
                 activation
                 direction
                 src-desc
                 src-iter-desc
                 weights-desc
                 weights-iter-desc
                 bias-desc
                 dst-desc
                 dst-iter-desc
                 diff-src-desc
                 diff-src-iter-desc
                 diff-weights-desc
                 diff-weights-iter-desc
                 diff-bias-desc
                 diff-dst-desc
                 diff-dst-iter-desc)

(vanilla-rnn-bwd eng
                 hint-fwd-pd
                 activation
                 direction
                 src-desc
                 src-iter-desc
                 weights-desc
                 weights-iter-desc
                 bias-desc
                 dst-desc
                 dst-iter-desc
                 diff-src-desc
                 diff-src-iter-desc
                 diff-weights-desc
                 diff-weights-iter-desc
                 diff-bias-desc
                 diff-dst-desc
                 diff-dst-iter-desc
                 alpha)

TODO

vanilla-rnn-fwd clj

(vanilla-rnn-fwd eng
                 prop-kind
                 activation
                 direction
                 src-desc
                 src-iter-desc
                 weights-desc
                 weights-iter-desc
                 bias-desc
                 dst-desc
                 dst-iter-desc)

(vanilla-rnn-fwd eng
                 prop-kind
                 activation
                 direction
                 src-desc
                 src-iter-desc
                 weights-desc
                 weights-iter-desc
                 bias-desc
                 dst-desc
                 dst-iter-desc
                 alpha)

TODO

wait! clj

(wait! strm)

Waits until stream s completes execution of all queued operations.

weights-md clj

(weights-md pd)

Queries the primitive descriptor pd for the weights.

workspace-md clj

(workspace-md pd)

Queries the primitive descriptor pd for the workspace (scratchpad).

zero-desc clj

zero-desc? clj

(zero-desc? mem-desc)