Use native texture buffers when available.
These were added in Metal 2.1.
diff --git a/MoltenVK/MoltenVK/API/vk_mvk_moltenvk.h b/MoltenVK/MoltenVK/API/vk_mvk_moltenvk.h
index 7365f17..27bac91 100644
--- a/MoltenVK/MoltenVK/API/vk_mvk_moltenvk.h
+++ b/MoltenVK/MoltenVK/API/vk_mvk_moltenvk.h
@@ -538,6 +538,7 @@
VkBool32 memoryBarriers; /**< If true, full memory barriers within Metal render passes are supported. */
VkBool32 multisampleLayeredRendering; /**< If true, layered rendering to multiple multi-sampled cube or texture array layers is supported. */
VkBool32 stencilFeedback; /**< If true, fragment shaders that write to [[stencil]] outputs are supported. */
+ VkBool32 textureBuffers; /**< If true, textures of type MTLTextureTypeBuffer are supported. */
} MVKPhysicalDeviceMetalFeatures;
/**
diff --git a/MoltenVK/MoltenVK/GPUObjects/MVKBuffer.mm b/MoltenVK/MoltenVK/GPUObjects/MVKBuffer.mm
index f56cd86..e60dab4 100644
--- a/MoltenVK/MoltenVK/GPUObjects/MVKBuffer.mm
+++ b/MoltenVK/MoltenVK/GPUObjects/MVKBuffer.mm
@@ -162,16 +162,25 @@
lock_guard<mutex> lock(_lock);
if (_mtlTexture) { return _mtlTexture; }
- MTLTextureDescriptor* mtlTexDesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat: _mtlPixelFormat
- width: _textureSize.width
- height: _textureSize.height
- mipmapped: NO];
- id<MTLBuffer> mtlBuff = _buffer->getMTLBuffer();
- mtlTexDesc.storageMode = mtlBuff.storageMode;
- mtlTexDesc.cpuCacheMode = mtlBuff.cpuCacheMode;
- mtlTexDesc.usage = MTLTextureUsageShaderRead;
+ MTLTextureUsage usage = MTLTextureUsageShaderRead;
if ( mvkIsAnyFlagEnabled(_buffer->getUsage(), VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT) ) {
- mtlTexDesc.usage |= MTLTextureUsageShaderWrite;
+ usage |= MTLTextureUsageShaderWrite;
+ }
+ id<MTLBuffer> mtlBuff = _buffer->getMTLBuffer();
+ MTLTextureDescriptor* mtlTexDesc;
+ if ( _device->_pMetalFeatures->textureBuffers ) {
+ mtlTexDesc = [MTLTextureDescriptor textureBufferDescriptorWithPixelFormat: _mtlPixelFormat
+ width: _textureSize.width
+ resourceOptions: (mtlBuff.cpuCacheMode << MTLResourceCPUCacheModeShift) | (mtlBuff.storageMode << MTLResourceStorageModeShift)
+ usage: usage];
+ } else {
+ mtlTexDesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat: _mtlPixelFormat
+ width: _textureSize.width
+ height: _textureSize.height
+ mipmapped: NO];
+ mtlTexDesc.storageMode = mtlBuff.storageMode;
+ mtlTexDesc.cpuCacheMode = mtlBuff.cpuCacheMode;
+ mtlTexDesc.usage = usage;
}
_mtlTexture = [_buffer->getMTLBuffer() newTextureWithDescriptor: mtlTexDesc
offset: _mtlBufferOffset
@@ -197,17 +206,24 @@
if (byteCount == VK_WHOLE_SIZE) { byteCount = _buffer->getByteCount() - pCreateInfo->offset; } // Remaining bytes in buffer
size_t blockCount = byteCount / bytesPerBlock;
- // But Metal requires the texture to be a 2D texture. Determine the number of 2D rows we need and their width.
- // Multiple rows will automatically align with PoT max texture dimension, but need to align upwards if less than full single row.
- size_t maxBlocksPerRow = _device->_pMetalFeatures->maxTextureDimension / fmtBlockSize.width;
- size_t blocksPerRow = min(blockCount, maxBlocksPerRow);
- _mtlBytesPerRow = mvkAlignByteOffset(blocksPerRow * bytesPerBlock, _device->getVkFormatTexelBufferAlignment(pCreateInfo->format, this));
+ if ( !_device->_pMetalFeatures->textureBuffers ) {
+ // But Metal requires the texture to be a 2D texture. Determine the number of 2D rows we need and their width.
+ // Multiple rows will automatically align with PoT max texture dimension, but need to align upwards if less than full single row.
+ size_t maxBlocksPerRow = _device->_pMetalFeatures->maxTextureDimension / fmtBlockSize.width;
+ size_t blocksPerRow = min(blockCount, maxBlocksPerRow);
+ _mtlBytesPerRow = mvkAlignByteOffset(blocksPerRow * bytesPerBlock, _device->getVkFormatTexelBufferAlignment(pCreateInfo->format, this));
- size_t rowCount = blockCount / blocksPerRow;
- if (blockCount % blocksPerRow) { rowCount++; }
+ size_t rowCount = blockCount / blocksPerRow;
+ if (blockCount % blocksPerRow) { rowCount++; }
- _textureSize.width = uint32_t(blocksPerRow * fmtBlockSize.width);
- _textureSize.height = uint32_t(rowCount * fmtBlockSize.height);
+ _textureSize.width = uint32_t(blocksPerRow * fmtBlockSize.width);
+ _textureSize.height = uint32_t(rowCount * fmtBlockSize.height);
+ } else {
+ // With native texture buffers we don't need to bother with any of that.
+ // We can just use a simple 1D texel array.
+ _textureSize.width = uint32_t(blockCount * fmtBlockSize.width);
+ _textureSize.height = 1;
+ }
if ( !_device->_pMetalFeatures->texelBuffers ) {
setConfigurationResult(reportError(VK_ERROR_FEATURE_NOT_PRESENT, "Texel buffers are not supported on this device."));
diff --git a/MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm b/MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm
index 30c2a4a..4334aa2 100644
--- a/MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm
+++ b/MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm
@@ -737,6 +737,7 @@
if ( [_mtlDevice supportsFeatureSet: MTLFeatureSet_iOS_GPUFamily1_v5] ) {
_metalFeatures.mslVersionEnum = MTLLanguageVersion2_1;
MVK_SET_FROM_ENV_OR_BUILD_BOOL(_metalFeatures.events, MVK_ALLOW_METAL_EVENTS);
+ _metalFeatures.textureBuffers = true;
}
if ( [_mtlDevice supportsFeatureSet: MTLFeatureSet_iOS_GPUFamily3_v1] ) {
@@ -796,6 +797,7 @@
_metalFeatures.multisampleArrayTextures = true;
MVK_SET_FROM_ENV_OR_BUILD_BOOL(_metalFeatures.events, MVK_ALLOW_METAL_EVENTS);
_metalFeatures.memoryBarriers = true;
+ _metalFeatures.textureBuffers = true;
}
if ( [_mtlDevice supportsFeatureSet: MTLFeatureSet_macOS_GPUFamily2_v1] ) {
diff --git a/MoltenVK/MoltenVK/GPUObjects/MVKPipeline.mm b/MoltenVK/MoltenVK/GPUObjects/MVKPipeline.mm
index abbdc49..84fa930 100644
--- a/MoltenVK/MoltenVK/GPUObjects/MVKPipeline.mm
+++ b/MoltenVK/MoltenVK/GPUObjects/MVKPipeline.mm
@@ -1129,6 +1129,7 @@
shaderContext.options.mslOptions.msl_version = _device->_pMetalFeatures->mslVersion;
shaderContext.options.mslOptions.texel_buffer_texture_width = _device->_pMetalFeatures->maxTextureDimension;
+ shaderContext.options.mslOptions.texture_buffer_native = _device->_pMetalFeatures->textureBuffers;
MVKPipelineLayout* layout = (MVKPipelineLayout*)pCreateInfo->layout;
layout->populateShaderConverterContext(shaderContext);
@@ -1321,6 +1322,7 @@
shaderContext.options.mslOptions.msl_version = _device->_pMetalFeatures->mslVersion;
shaderContext.options.mslOptions.texel_buffer_texture_width = _device->_pMetalFeatures->maxTextureDimension;
shaderContext.options.mslOptions.swizzle_texture_samples = _fullImageViewSwizzle;
+ shaderContext.options.mslOptions.texture_buffer_native = _device->_pMetalFeatures->textureBuffers;
MVKPipelineLayout* layout = (MVKPipelineLayout*)pCreateInfo->layout;
layout->populateShaderConverterContext(shaderContext);
