Update README.md

README.md

@@ -80,11 +80,9 @@ Precompile the OpenCL programs for all available devices, and save them to the o
 `QTensorNetwork` has a qubit threshold up to which it is able to reuse more work in measurement and probability calculations, controlled by environment variable `QRACK_QTENSORNETWORK_THRESHOLD_QB`. By default, this is the same as qubit width that Qrack detects to be the maximum for state vector simulation (environment variables `QRACK_MAX_PAGING_QB` or `QRACK_MAX_CPU_QB`). Above (and not including) this threshold, `QTensorNetwork` will use techniques like restricting to "past light cones" for measurement and probablity calculation, in an attempt to reduce overall memory footprint at the cost of additional execution time.
 
 ## Maximum allocation guard
-Set the maximum allowed allocation (in MB) for the global OpenCL pool with environment variable `QRACK_MAX_ALLOC_MB`. Per OpenCL device, this sets each maximum allocation limit with the same syntax as `QRACK_QPAGER_DEVICES`. (Succesive entries in the list are MB limits numbered according to Qrack's device IDs print-out on launch.) By default, each device is capped at 3/4 of its available global memory, for stability in common use cases. This includes (VRAM) state vectors and auxiliary buffers larger than approximately `sizeof(bitCapIntOcl) * sizeof(bitCapIntOcl)`. This should also include out-of-place single duplication of any state vector. This does **not** include non-OpenCL general heap or stack allocation.
+Set the maximum allowed allocation (in MB) for the global OpenCL pool with environment variable `QRACK_MAX_ALLOC_MB`. Per OpenCL device, this sets each maximum allocation limit with the same syntax as `QRACK_QPAGER_DEVICES`. (Succesive entries in the list are MB limits numbered according to Qrack's device IDs print-out on launch.) This limit includes (VRAM) state vectors and auxiliary buffers. This should also include out-of-place single duplication of any state vector. This does **not** include non-OpenCL general heap or stack allocation.
 
-`QRACK_MAX_PAGING_QB` and `QRACK_MAX_CPU_QB` environment variables set a maximum on how many qubits can be allocated on a single `QPager` or `QEngineCPU` instance, respectively. This qubit limit is for maximum single `QPager` or `QEngineCPU` allocation, whereas `QUnit` and `QUnitMulti` might allocate _more_ qubits than this as separable subsystems, requiring that no individual separable subsystem exceeds the qubit limit environment variables. (`QEngineOCL` limits are automatically maximal according to a query Qrack makes of maximum allocation segment on a given OpenCL device.)
-
-Note that this controls total direct Qrack OpenCL buffer allocation, not total Qrack library allocation. Total OpenCL buffer allocation is **not** fully indicative of total library allocation.
+`QRACK_MAX_PAGING_QB` and `QRACK_MAX_CPU_QB` environment variables set a maximum on how many qubits can be allocated on a single `QPager` or `QEngineCPU` instance, respectively. This qubit limit is for maximum single `QPager` or `QEngineCPU` allocation, whereas `QUnit` and `QUnitMulti` might allocate _more_ qubits than this as separable subsystems, while requiring that no individual separable subsystem exceeds the qubit limit environment variables. (`QEngineOCL` limits are automatically maximal according to a query Qrack makes of maximum allocation segment on a given OpenCL device.)
 
 ## Approximation and noise options
 `QUnit` can optionally round qubit subsystems proactively or on-demand to the nearest single or double qubit eigenstate with the `QRACK_QUNIT_SEPARABILITY_THRESHOLD=[0.0 - 1.0]` environment variable, with a value between `0.0` and `1.0`. When trying to find separable subsystems, Qrack will start by making 3-axis (independent or conditional) probability measurements. Based on the probability measurements, under the assumption that the state _is_ separable, an inverse state preparation to |0> procedure is fixed. If inverse state preparation would bring any single qubit Bloch sphere projection within parameter range of the edge of the Bloch sphere, (with unit length, `1.0`,) then the subsystem will be rounded to that state, normalized, and then "uncomputed" with the corresponding (forward) state preparation, effectively "hyperpolarizing" one and two qubit separable substates by replacing entanglement with local qubit Bloch sphere extent. (If 3-axis probability is _not_ within rounding range, nothing is done directly to the substate.)