Merge pull request #1793 from billhollings/fix-fastmath-regression

Fix app performance regression triggered by the previous introduction of VK_KHR_shader_float_controls.

Docs/Whats_New.md

@@ -20,6 +20,7 @@ Released TBD
 
 - Fix crash on descriptor update with out-of-bounds descriptor count data.
 - Fix Metal buffer index binding overrides for push constants and attachment clearing.
+- Fix app performance regression triggered by the previous introduction of `VK_KHR_shader_float_controls`.
 - Work around `MTLCounterSet` crash on additional Intel Iris Plus Graphics devices.
 - Fix mistaken YCBCR format support indication.
 - Document new linkage model used by *Xcode 14* and later, and how to link **MoltenVK**

MoltenVK/MoltenVK/API/vk_mvk_moltenvk.h

@@ -122,6 +122,14 @@ typedef enum MVKPrefillMetalCommandBuffersStyle {
 	MVK_CONFIG_PREFILL_METAL_COMMAND_BUFFERS_STYLE_MAX_ENUM                          = 0x7FFFFFFF
 } MVKPrefillMetalCommandBuffersStyle;
 
+/** Identifies when Metal shaders will be compiled with the fast math option. */
+typedef enum MVKConfigFastMath {
+	MVK_CONFIG_FAST_MATH_NEVER     = 0,  /**< Metal shaders will never be compiled with the fast math option. */
+	MVK_CONFIG_FAST_MATH_ALWAYS    = 1,  /**< Metal shaders will always be compiled with the fast math option. */
+	MVK_CONFIG_FAST_MATH_ON_DEMAND = 2,  /**< Metal shaders will be compiled with the fast math option, unless the shader includes execution modes that require it to be compiled without fast math. */
+	MVK_CONFIG_FAST_MATH_MAX_ENUM  = 0x7FFFFFFF
+} MVKConfigFastMath;
+
 /**
  * MoltenVK configuration settings.
  *
@@ -540,18 +548,30 @@ typedef struct {
 	uint32_t defaultGPUCaptureScopeQueueIndex;
 
 	/**
-	 * Corresponds to the fastMathEnabled property of MTLCompileOptions.
-	 * Setting it may cause the Metal Compiler to optimize floating point operations
-	 * in ways that may violate the IEEE 754 standard.
+	 * Identifies when Metal shaders will be compiled with the Metal fastMathEnabled property
+	 * enabled. For shaders compiled with the Metal fastMathEnabled property enabled, shader
+	 * floating point math is significantly faster, but it may cause the Metal Compiler to
+	 * optimize floating point operations in ways that may violate the IEEE 754 standard.
 	 *
-	 * Must be changed before creating a VkDevice, for the change to take effect.
+	 * Enabling Metal fast math can dramatically improve shader performance, and has little
+	 * practical effect on the numerical accuracy of most shaders. As such, disabling fast
+	 * math should be done carefully and deliberately. For most applications, always enabling
+	 * fast math, by setting the value of this property to MVK_CONFIG_FAST_MATH_ALWAYS,
+	 * is the preferred choice.
+	 *
+	 * Apps that have specific accuracy and handling needs for particular shaders, may elect to
+	 * set the value of this property to MVK_CONFIG_FAST_MATH_ON_DEMAND, so that fast math will
+	 * be disabled when compiling shaders that request capabilities such as SignedZeroInfNanPreserve.
+	 *
+	 * The value of this parameter may be changed at any time during application runtime,
+	 * and the changed value will be applied to future Metal shader compilations.
 	 *
 	 * The initial value or this parameter is set by the
 	 * MVK_CONFIG_FAST_MATH_ENABLED
 	 * runtime environment variable or MoltenVK compile-time build setting.
-	 * If neither is set, the value of this parameter defaults to true.
+	 * If neither is set, the value of this parameter defaults to MVK_CONFIG_FAST_MATH_ALWAYS.
 	 */
-	VkBool32 fastMathEnabled;
+	MVKConfigFastMath fastMathEnabled;
 
 	/**
 	 * Controls the level of logging performned by MoltenVK.

MoltenVK/MoltenVK/GPUObjects/MVKDevice.h

@@ -715,12 +715,12 @@ public:
 
 	/**
 	 * Returns an autoreleased options object to be used when compiling MSL shaders.
-	 * The useFastMath parameter is and-combined with MVKConfiguration::fastMathEnabled
+	 * The requestFastMath parameter is combined with the value of MVKConfiguration::fastMathEnabled
 	 * to determine whether to enable fast math optimizations in the compiled shader.
 	 * The preserveInvariance parameter indicates that the shader requires the position
 	 * output invariance across invocations (typically for the position output).
 	 */
-	MTLCompileOptions* getMTLCompileOptions(bool useFastMath = true, bool preserveInvariance = false);
+	MTLCompileOptions* getMTLCompileOptions(bool requestFastMath = true, bool preserveInvariance = false);
 
 	/** Returns the Metal vertex buffer index to use for the specified vertex attribute binding number.  */
 	uint32_t getMetalBufferIndexForVertexAttributeBinding(uint32_t binding);

MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm

@@ -4255,10 +4255,11 @@ id<MTLBuffer> MVKDevice::getDummyBlitMTLBuffer() {
 	return _dummyBlitMTLBuffer;
 }
 
-MTLCompileOptions* MVKDevice::getMTLCompileOptions(bool useFastMath, bool preserveInvariance) {
+MTLCompileOptions* MVKDevice::getMTLCompileOptions(bool requestFastMath, bool preserveInvariance) {
 	MTLCompileOptions* mtlCompOpt = [MTLCompileOptions new];
 	mtlCompOpt.languageVersion = _pMetalFeatures->mslVersionEnum;
-	mtlCompOpt.fastMathEnabled = useFastMath && mvkConfig().fastMathEnabled;
+	mtlCompOpt.fastMathEnabled = (mvkConfig().fastMathEnabled == MVK_CONFIG_FAST_MATH_ALWAYS ||
+								  (mvkConfig().fastMathEnabled == MVK_CONFIG_FAST_MATH_ON_DEMAND && requestFastMath));
 #if MVK_XCODE_12
 	if ([mtlCompOpt respondsToSelector: @selector(setPreserveInvariance:)]) {
 		[mtlCompOpt setPreserveInvariance: preserveInvariance];

MoltenVK/MoltenVK/Utility/MVKEnvironment.cpp

@@ -45,7 +45,7 @@ static void mvkInitConfigFromEnvVars() {
 	MVK_SET_FROM_ENV_OR_BUILD_BOOL  (evCfg.fullImageViewSwizzle,                   MVK_CONFIG_FULL_IMAGE_VIEW_SWIZZLE);
 	MVK_SET_FROM_ENV_OR_BUILD_BOOL  (evCfg.defaultGPUCaptureScopeQueueFamilyIndex, MVK_CONFIG_DEFAULT_GPU_CAPTURE_SCOPE_QUEUE_FAMILY_INDEX);
 	MVK_SET_FROM_ENV_OR_BUILD_BOOL  (evCfg.defaultGPUCaptureScopeQueueIndex,       MVK_CONFIG_DEFAULT_GPU_CAPTURE_SCOPE_QUEUE_INDEX);
-	MVK_SET_FROM_ENV_OR_BUILD_BOOL  (evCfg.fastMathEnabled,                        MVK_CONFIG_FAST_MATH_ENABLED);
+	MVK_SET_FROM_ENV_OR_BUILD_INT32 (evCfg.fastMathEnabled,                        MVK_CONFIG_FAST_MATH_ENABLED);
 	MVK_SET_FROM_ENV_OR_BUILD_INT32 (evCfg.logLevel,                               MVK_CONFIG_LOG_LEVEL);
 	MVK_SET_FROM_ENV_OR_BUILD_INT32 (evCfg.traceVulkanCalls,                       MVK_CONFIG_TRACE_VULKAN_CALLS);
 	MVK_SET_FROM_ENV_OR_BUILD_BOOL  (evCfg.forceLowPowerGPU,                       MVK_CONFIG_FORCE_LOW_POWER_GPU);

MoltenVK/MoltenVK/Utility/MVKEnvironment.h

@@ -173,9 +173,9 @@ void mvkSetConfig(const MVKConfiguration& mvkConfig);
 #   define MVK_CONFIG_FULL_IMAGE_VIEW_SWIZZLE    0
 #endif
 
-/** Set the fastMathEnabled Metal Compiler option. Enabled by default. */
+/** Set the fastMathEnabled Metal Compiler option. Set to always use fast math by default. */
 #ifndef MVK_CONFIG_FAST_MATH_ENABLED
-#   define MVK_CONFIG_FAST_MATH_ENABLED 1
+#   define MVK_CONFIG_FAST_MATH_ENABLED    MVK_CONFIG_FAST_MATH_ALWAYS
 #endif
 
 /** Set the logging level: */