diff -ruN /Applications/Xcode_15.4.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArray.h /Applications/Xcode_16.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArray.h--- /Applications/Xcode_15.4.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArray.h2024-04-13 20:23:59+++ /Applications/Xcode_16.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArray.h2024-05-30 02:38:51@@ -10,3 +10,6 @@ #import <MPSNDArray/MPSNDArrayMatrixMultiplication.h> #import <MPSNDArray/MPSNDArrayStridedSlice.h> #import <MPSNDArray/MPSNDArrayGather.h>+#import <MPSNDArray/MPSNDArrayIdentity.h>+#import <MPSNDArray/MPSNDArrayQuantization.h>+#import <MPSNDArray/MPSNDArrayQuantizedMatrixMultiplication.h>diff -ruN /Applications/Xcode_15.4.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayIdentity.h /Applications/Xcode_16.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayIdentity.h--- /Applications/Xcode_15.4.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayIdentity.h1970-01-01 01:00:00+++ /Applications/Xcode_16.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayIdentity.h2024-05-30 02:38:54@@ -0,0 +1,65 @@+//+//  MPSNDArrayIdentity.h+//  MPS+//+//  Created by Dhruv Saksena on 8/3/19.+//  Copyright © 2019 Apple. All rights reserved.+//++#ifndef MPSNDArrayIdentity_h+#define MPSNDArrayIdentity_h++#import <MPSNDArray/MPSNDArrayKernel.h>++#pragma mark - Base class Identity Kernels++/*!+ *  @class      MPSNDArrayIdentityKernel+ *+ *  @dependency This depends on Metal.framework.+ *  @abstract   An efficient kernel to handle copies, transposed-copies and reshapes.+ *+ *+ */+MPS_CLASS_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) )+@interface MPSNDArrayIdentity : MPSNDArrayUnaryKernel++-(nonnull instancetype) initWithDevice: (nonnull id <MTLDevice>) device NS_DESIGNATED_INITIALIZER;++/*! @abstract   Do a reshape operation, either by trying to alias the array, returning an arrayview, or by copying.+ *  @param      cmdBuf          The command buffer into which to encode the kernel, or to create a temporary array alias.+ *  @param      sourceArray    Source array. If this function returns a non-nil result, then the readCount of `sourceArray` is decremented.+ *  @param      shape       The new shape, given in TF dimension ordering (as always with MPSShape).+ *  @param      destinationArray    If not nil, then the result of reshape will be copied to this. Shape of `destinationArray` must match `shape`.+ *  @result     If `destinationArray` is not nil, then `destinationArray`. Otherwise aliasing is tried, and if aliasing is not possible+ *              due to existing slices or transposes nil is returned. If aliasing is successful, then a new arrayview of `sourceArray`+ *              is returned; If `sourceArray` is a `MPSTemporaryArray` then a `MPSTemporaryArray` is returned referencing the same data,+ *              otherwise a `MPSNDArray` type result is returned.+ *   */+-(MPSNDArray * __nullable) reshapeWithCommandBuffer: (__nullable id <MTLCommandBuffer>) cmdBuf+                                        sourceArray: (MPSNDArray * __nonnull) sourceArray+                                              shape: (MPSShape * __nonnull) shape+                                   destinationArray: (MPSNDArray * __nullable) destinationArray;+// Variant with MPS dimension-ordering+-(MPSNDArray * __nullable) reshapeWithCommandBuffer: (__nullable id <MTLCommandBuffer>) cmdBuf+                                        sourceArray: (MPSNDArray * __nonnull) sourceArray+                                     dimensionCount: (NSUInteger) numberOfDimensions+                                     dimensionSizes: (NSUInteger*__nonnull) dimensionSizes+                                   destinationArray: (MPSNDArray * __nullable) destinationArray;++-(MPSNDArray * __nullable) reshapeWithCommandEncoder: (__nullable id <MTLComputeCommandEncoder>) encoder+                                       commandBuffer: (__nullable id <MTLCommandBuffer>) cmdBuf+                                        sourceArray: (MPSNDArray * __nonnull) sourceArray+                                              shape: (MPSShape * __nonnull) shape+                                   destinationArray: (MPSNDArray * __nullable) destinationArray;+// Variant with MPS dimension-ordering+-(MPSNDArray * __nullable) reshapeWithCommandEncoder: (__nullable id <MTLComputeCommandEncoder>) encoder+                                       commandBuffer: (__nullable id <MTLCommandBuffer>) cmdBuf+                                        sourceArray: (MPSNDArray * __nonnull) sourceArray+                                     dimensionCount: (NSUInteger) numberOfDimensions+                                     dimensionSizes: (NSUInteger*__nonnull) dimensionSizes+                                   destinationArray: (MPSNDArray * __nullable) destinationArray;++@end    // MPSNDArrayIdentity++#endif /* MPSNDArrayIdentity_h */diff -ruN /Applications/Xcode_15.4.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayQuantization.h /Applications/Xcode_16.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayQuantization.h--- /Applications/Xcode_15.4.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayQuantization.h1970-01-01 01:00:00+++ /Applications/Xcode_16.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayQuantization.h2024-05-30 05:07:59@@ -0,0 +1,125 @@+//+//  MPSNDArrayQuantization.h+//  MPSNDArray+//+//  Created by Saharsh Oza on 4/24/23.+//  Copyright © 2023 Apple. All rights reserved.+//++#ifndef MPSNDArrayQuantization_h+#define MPSNDArrayQuantization_h++#import <MPSCore/MPSCoreTypes.h>+#import <MPSCore/MPSKernel.h>+#import <MPSCore/MPSNDArray.h>++/*! @enum       MPSNDArrayQuantizationScheme+ *  @discussion  Which quantization mode is applied to an MPSNDArray+ *+ *  @constant   MPSNDArrayQuantizationTypeNone              No quantization - inputs are read and outputs are produced in the original datatype.+ *  @constant   MPSNDArrayQuantizationTypeAffine           Dequantization with: y = scale( input - zeroPoint) + minValue+ *  @constant   MPSNDArrayQuantizationTypeLUT           Dequantization with y = lookuptable[ input ]+ */++#if defined(DOXYGEN)+    typedef enum MPSNDArrayQuantizationScheme+#else+    typedef NS_OPTIONS(NSUInteger, MPSNDArrayQuantizationScheme)+#endif+{+    MPSNDArrayQuantizationTypeNone MPS_ENUM_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) ) MPS_SWIFT_NAME( none ) = 0,+    MPSNDArrayQuantizationTypeAffine MPS_ENUM_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) ),+    MPSNDArrayQuantizationTypeLUT MPS_ENUM_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) ),+}+#if defined(DOXYGEN)+    MPSNDArrayQuantizationScheme+#endif+;+++/*!+ *  @class      MPSNDArrayQuantizationDescriptor+ *+ *  @dependency This depends on Metal.framework.+ *  @abstract   Common methods for quantization descriptors+ *+ *+ */+MPS_CLASS_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) )+@interface MPSNDArrayQuantizationDescriptor : NSObject <NSCopying>++/*! @property  quantizationDataType+ *  @abstract  The datatype to use with quantization - the default is MPSDataTypeUint8+ */+@property (readonly, nonatomic) MPSDataType quantizationDataType;++/*! @property  quantizationScheme+ *  @abstract  The quantization scheme for this descriptor. The default is MPSNDArrayQuantizationTypeNone.+ */+@property (readonly, nonatomic) MPSNDArrayQuantizationScheme quantizationScheme;++@end++/*!+ *  @class      MPSNDArrayAffineQuantizationDescriptor+ *+ *  @dependency This depends on Metal.framework.+ *  @abstract   Describes an affine quantization scheme+ *+ */+MPS_CLASS_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) )+@interface MPSNDArrayAffineQuantizationDescriptor : MPSNDArrayQuantizationDescriptor++/*! @property  hasZeroPoint+ *  @abstract  If yes then asymmetric quantization is used. See MPSNDArrayQuantizationScheme.+ */+@property (readwrite, nonatomic) BOOL hasZeroPoint;++/*! @property  hasMinValue+ *  @abstract  If yes then offset is used. See MPSNDArrayQuantizationScheme.+ */+@property (readwrite, nonatomic) BOOL hasMinValue;+++- (nonnull instancetype) init;++/*! @abstract   Initializes an affine quantization descriptor.+ *  @param      quantizationDataType    Which quantized datatype is used.+ *  @param      hasZeroPoint        A flag indicating that a zero-point input is expected.+ *  @param      hasMinValue        A flag indicating that a minimum value input is expected.+ *  @result     A new quantization descriptor.+ *   */+- (nonnull instancetype) initWithDataType: (MPSDataType) quantizationDataType+                             hasZeroPoint: (BOOL) hasZeroPoint+                              hasMinValue: (BOOL) hasMinValue;++@end+++/*!+ *  @class      MPSNDArrayLUTQuantizationDescriptor+ *+ *  @dependency This depends on Metal.framework.+ *  @abstract   Describes a lookup-table based quantization scheme+ *+ */+MPS_CLASS_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) )+@interface MPSNDArrayLUTQuantizationDescriptor : MPSNDArrayQuantizationDescriptor++/*! @abstract   Initializes a scalar lookup-table quantization descriptor.+ *  @param      quantizationDataType    Which quantized datatype is used.+ *  @result     A new quantization descriptor.+ *   */+- (nonnull instancetype) initWithDataType: (MPSDataType) quantizationDataType;++/*! @abstract   Initializes a vector lookup-table quantization descriptor.+ *  @param      quantizationDataType    Which quantized datatype is used.+ *  @param      vectorAxis        The quantization vector axis - this axis will receive the vector component in the destination.+ *  @result     A new quantization descriptor.+ *   */+- (nonnull instancetype) initWithDataType: (MPSDataType) quantizationDataType+                               vectorAxis: (NSUInteger) vectorAxis;++@end++#endif /* MPSNDArrayQuantization_h */diff -ruN /Applications/Xcode_15.4.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayQuantizedMatrixMultiplication.h /Applications/Xcode_16.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayQuantizedMatrixMultiplication.h--- /Applications/Xcode_15.4.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayQuantizedMatrixMultiplication.h1970-01-01 01:00:00+++ /Applications/Xcode_16.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShaders.framework/Frameworks/MPSNDArray.framework/Headers/MPSNDArrayQuantizedMatrixMultiplication.h2024-05-30 03:40:26@@ -0,0 +1,127 @@+//+//  MPSNDArrayQuantizedMatrixMultiplication.h+//  MPSNDArray+//+//  Copyright © 2023 Apple. All rights reserved.+//++#ifndef MPSNDArrayQuantizedMatrixMultiplication_h+#define MPSNDArrayQuantizedMatrixMultiplication_h++#import <MPSNDArray/MPSNDArrayMatrixMultiplication.h>+#import <MPSNDArray/MPSNDArrayQuantization.h>++/*!+ *  @class      MPSNDArrayQuantizedMatrixMultiplication+ *+ *  @dependency This depends on Metal.framework.+ *  @abstract   A quantized matrix multiplication kernel: C = AB, where each input A and B can be quantized.+ *  @discussion The kernel works with 2-8 inputs, order of inputs: First all LHS inputs, then all RHS inputs.+ *              The order of inputs for LUT based LHS or RHS: 1) quantized input 2) Lookup Table.+ *              The order of inputs for affine LHS or RHS: 1) quantized input 2) scale 3) zeropoint 4) minValue.+ *              The full order of inputs for the encode methods is:+ *                   `[LHS, RHS, <LHS quantization inputs>, <RHS quantization inputs>]`,+ *              where `LHS` is the left input (quantized or float) `RHS` is the right input (quantized or float) and+ *              `<LHS quantization inputs>` are the auxiliary quantization inputs for the LHS array (scales, zeropoints etc).+ *              and `<RHS quantization inputs>` are the auxiliary quantization input for the RHS array.+ *              The inputs are provided as a compacted `NSArray<MPSNDArray *>`, for example for computing+ *              `C = A * B^T` where `A` is quantized with a LUT and `B` is quantized with affine quantization that+ *              uses scale and minValue the array of inputs is:+ *              ` [ Aq, Bq^T, ALUT, BScale^T, BMin^T ] `.+ *              NOTE: For affine scale, zeropoint and minValue must have same transposes as quantized input.+ *+ */+MPS_CLASS_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) )+@interface MPSNDArrayQuantizedMatrixMultiplication : MPSNDArrayMatrixMultiplication++-(nonnull instancetype) initWithDevice:(nonnull id<MTLDevice>)device+                           sourceCount:(NSUInteger) sourceCount NS_UNAVAILABLE;++/*! @abstract   Initializes a quantized matrix multiplication kernel.+ *  @param      leftQuantizationDescriptor    The quantization definition for the LHS input.+ *  @param      rightQuantizationDescriptor    The quantization definition for the RHS input.+ *  @result     A new valid quantized matrix multiplication kernel.+ *   */+- (nonnull instancetype) initWithDevice: (nonnull id<MTLDevice>) device+             leftQuantizationDescriptor: (MPSNDArrayQuantizationDescriptor* _Nullable) leftQuantizationDescriptor+            rightQuantizationDescriptor: (MPSNDArrayQuantizationDescriptor* _Nullable) rightQuantizationDescriptor;++@end++/*!+ *  @class      MPSNDArrayLUTDequantize+ *+ *  @dependency This depends on Metal.framework.+ *  @abstract   A kernel which dequantizes a lookup-table based NDArray.+ *  @discussion The kernel works with 2 inputs: 1) The quantized input, 2) The LookUp table array.+ *+ */++MPS_CLASS_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) )+@interface MPSNDArrayLUTDequantize : MPSNDArrayMultiaryKernel++-(nonnull instancetype) initWithDevice: (nonnull id <MTLDevice>) device+NS_DESIGNATED_INITIALIZER;++-(nonnull instancetype) initWithDevice:(nonnull id<MTLDevice>)device+                           sourceCount:(NSUInteger) sourceCount NS_UNAVAILABLE;++@end++/*!+ *  @class      MPSNDArrayVectorLUTDequantize+ *+ *  @dependency This depends on Metal.framework.+ *  @abstract   A kernel which dequantizes a lookup-table based NDArray with vector LUT support.+ *  @discussion The kernel works with 2 inputs: 1) The quantized input, 2) The LookUp table array.+ *+ */+MPS_CLASS_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) )+@interface MPSNDArrayVectorLUTDequantize : MPSNDArrayMultiaryKernel++/*! @property  vectorAxis+ *  @abstract  Which axis in the destination will receive the vector component, must be less than 4.+ */+@property(readwrite, nonatomic) NSUInteger vectorAxis;++/*! @abstract   Initializes a kernel for vector-based LUT dequantization.+ *  @param      device    The Metal device to be used with this kernel.+ *  @param      axis        The vector axis in the output.+ *  @result     A new vector LUT dequantization kernel.+ *   */+-(nonnull instancetype) initWithDevice: (nonnull id <MTLDevice>) device+                                  axis: (NSUInteger) axis+NS_DESIGNATED_INITIALIZER;++-(nonnull instancetype) initWithDevice:(nonnull id<MTLDevice>)device+                           sourceCount:(NSUInteger) sourceCount NS_UNAVAILABLE;++@end++/*!+ *  @class      MPSNDArrayAffineInt4Dequantize+ *+ *  @dependency This depends on Metal.framework.+ *  @abstract   A kernel which dequantizes an input with affine quantization scheme.+ *  @discussion The kernel works with 2-4 inputs, order of inputs: 1) quantized input, 2) scale, 3) zeropoint, 4) minValue+ *+ */+MPS_CLASS_AVAILABLE_STARTING( macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0), xros(2.0) )+@interface MPSNDArrayAffineInt4Dequantize : MPSNDArrayMultiaryKernel++-(nonnull instancetype) initWithDevice: (nonnull id <MTLDevice>) device NS_UNAVAILABLE;++/*! @abstract   Initializes a kernel for 4-bit affine dequantization.+ *  @param      device    The Metal device to be used with this kernel.+ *  @param      quantizationDescriptor        Describes the quantization scheme.+ *  @result     A new vector LUT dequantization kernel.+ *   */+-(nonnull instancetype) initWithDevice:(nonnull id<MTLDevice>)device+                quantizationDescriptor:(MPSNDArrayAffineQuantizationDescriptor* __nonnull) quantizationDescriptor;++-(nonnull instancetype) initWithDevice:(nonnull id<MTLDevice>)device+                           sourceCount:(NSUInteger) sourceCount NS_UNAVAILABLE;++@end++#endif /* MPSNDArrayQuantizedMatrixMultiplication_h */