[spirv] Implement faster in-memory compression for shaders

Seems to be anything up to 3x as fast to decode than the previous code,
with the compression ratio being slightly worse. Encoding seems faster
as well.

src/spirv/spirv_compression.cpp

@@ -8,58 +8,76 @@ namespace dxvk {
   }
 
 
-  SpirvCompressedBuffer::SpirvCompressedBuffer(
+  SpirvCompressedBuffer::SpirvCompressedBuffer(SpirvCodeBuffer& code)
-    const SpirvCodeBuffer& code)
   : m_size(code.dwords()) {
+    // The compression (detailed below) achieves roughly 55% of the
+    // original size on average and is very consistent, so an initial
+    // estimate of roughly 58% will be accurate most of the time.
     const uint32_t* data = code.data();
+    m_code.reserve((m_size * 75) / 128);
 
-    // The compression works by eliminating leading null bytes
+    std::array<uint32_t, 16> block;
-    // from DWORDs, exploiting that SPIR-V IDs are consecutive
+    uint32_t blockMask = 0;
-    // integers that usually fall into the 16-bit range. For
+    uint32_t blockOffset = 0;
-    // each DWORD, a two-bit integer is stored which indicates
-    // the number of bytes it takes in the compressed buffer.
-    // This way, it can achieve a compression ratio of ~50%.
-    m_mask.reserve((m_size + NumMaskWords - 1) / NumMaskWords);
-    m_code.reserve((m_size + 1) / 2);
 
-    uint64_t dstWord  = 0;
+    // The algorithm used is a simple variable-to-fixed compression that
-    uint32_t dstShift = 0;
+    // encodes up to two consecutive SPIR-V tokens into one DWORD using
+    // a small number of different encodings. While not achieving great
+    // compression ratios, the main goal is to allow decompression code
+    // to be fast, with short dependency chains.
+    // Compressed tokens are stored in blocks of 16 DWORDs, each preceeded
+    // by a single DWORD which stores the layout for each DWORD, two bits
+    // each. The supported layouts, are as follows:
+    // 0x0: 1x 32-bit;  0x1: 1x 20-bit + 1x 12-bit
+    // 0x2: 2x 16-bit;  0x3: 1x 12-bit + 1x 20-bit
+    // These layouts are chosen to allow reasonably efficient encoding of
+    // opcode tokens, which usually fit into 20 bits, followed by type IDs,
+    // which tend to be low as well since most types are defined early.
+    for (size_t i = 0; i < m_size; ) {
+      if (likely(i + 1 < m_size)) {
+        uint32_t a = data[i];
+        uint32_t b = data[i + 1];
+        uint32_t schema;
+        uint32_t encode;
 
-    for (uint32_t i = 0; i < m_size; i += NumMaskWords) {
+        if (std::max(a, b) < (1u << 16)) {
-      uint64_t byteCounts = 0;
+          schema = 0x2;
-
+          encode = a | (b << 16);
-      for (uint32_t w = 0; w < NumMaskWords && i + w < m_size; w++) {
+        } else if (a < (1u << 20) && b < (1u << 12)) {
-        uint64_t word = data[i + w];
+          schema = 0x1;
-        uint64_t bytes = 0;
+          encode = a | (b << 20);
-
+        } else if (a < (1u << 12) && b < (1u << 20)) {
-        if      (word < (1 <<  8)) bytes = 0;
+          schema = 0x3;
-        else if (word < (1 << 16)) bytes = 1;
+          encode = a | (b << 12);
-        else if (word < (1 << 24)) bytes = 2;
+        } else {
-        else                       bytes = 3;
+          schema = 0x0;
-
+          encode = a;
-        byteCounts |= bytes << (2 * w);
-
-        uint32_t bits = 8 * bytes + 8;
-        uint32_t rem  = bit::pack(dstWord, dstShift, word, bits);
-
-        if (unlikely(rem != 0)) {
-          m_code.push_back(dstWord);
-
-          dstWord  = 0;
-          dstShift = 0;
-
-          bit::pack(dstWord, dstShift, word >> (bits - rem), rem);
         }
+
+        block[blockOffset] = encode;
+        blockMask |= schema << (blockOffset << 1);
+        blockOffset += 1;
+
+        i += schema ? 2 : 1;
+      } else {
+        block[blockOffset] = data[i++];
+        blockOffset += 1;
       }
 
-      m_mask.push_back(byteCounts);
+      if (unlikely(blockOffset == 16) || unlikely(i == m_size)) {
+        m_code.insert(m_code.end(), blockMask);
+        m_code.insert(m_code.end(), block.begin(), block.begin() + blockOffset);
+
+        blockMask = 0;
+        blockOffset = 0;
+      }
     }
 
-    if (dstShift)
+    // Only shrink the array if we have lots of overhead for some reason.
-      m_code.push_back(dstWord);
+    // This should only happen on shaders where our initial estimate was
-
+    // too small. In general, we want to avoid reallocation here.
-    m_mask.shrink_to_fit();
+    if (m_code.capacity() > (m_code.size() * 10) / 9)
-    m_code.shrink_to_fit();
+      m_code.shrink_to_fit();
   }
 
     
@@ -72,36 +90,31 @@ namespace dxvk {
     SpirvCodeBuffer code(m_size);
     uint32_t* data = code.data();
 
-    if (m_size == 0)
+    uint32_t srcOffset = 0;
-      return code;
+    uint32_t dstOffset = 0;
 
-    uint32_t maskIdx = 0;
+    constexpr uint32_t shiftAmounts = 0x0c101420;
-    uint32_t codeIdx = 0;
 
-    uint64_t srcWord  = m_code[codeIdx++];
+    while (dstOffset < m_size) {
-    uint32_t srcShift = 0;
+      uint32_t blockMask = m_code[srcOffset];
 
-    for (uint32_t i = 0; i < m_size; i += NumMaskWords) {
+      for (uint32_t i = 0; i < 16 && dstOffset < m_size; i++) {
-      uint64_t srcMask = m_mask[maskIdx++];
+        // Use 64-bit integers for some of the operands so we can
+        // shift by 32 bits and not handle it as a special cases
+        uint32_t schema = (blockMask >> (i << 1)) & 0x3;
+        uint32_t shift  = (shiftAmounts >> (schema << 3)) & 0xff;
+        uint64_t mask   = ~(~0ull << shift);
+        uint64_t encode = m_code[srcOffset + i + 1];
 
-      for (uint32_t w = 0; w < NumMaskWords && i + w < m_size; w++) {
+        data[dstOffset] = encode & mask;
-        uint32_t bits = 8 * ((srcMask & 3) + 1);
 
-        uint64_t word = 0;
+        if (likely(schema))
-        uint32_t rem = bit::unpack(word, srcWord, srcShift, bits);
+          data[dstOffset + 1] = encode >> shift;
 
-        if (unlikely(rem != 0)) {
+        dstOffset += schema ? 2 : 1;
-          srcWord  = m_code[codeIdx++];
-          srcShift = 0;
-
-          uint64_t tmp = 0;
-          bit::unpack(tmp, srcWord, srcShift, rem);
-          word |= tmp << (bits - rem);
-        }
-
-        data[i + w] = word;
-        srcMask >>= 2;
       }
+
+      srcOffset += 17;
     }
 
     return code;

src/spirv/spirv_compression.h

@@ -13,13 +13,12 @@ namespace dxvk {
    * to keep memory footprint low.
    */
   class SpirvCompressedBuffer {
-    constexpr static uint32_t NumMaskWords = 32;
+
   public:
 
     SpirvCompressedBuffer();
 
-    SpirvCompressedBuffer(
+    SpirvCompressedBuffer(SpirvCodeBuffer& code);
-      const SpirvCodeBuffer&  code);
     
     ~SpirvCompressedBuffer();
     
@@ -27,9 +26,12 @@ namespace dxvk {
 
   private:
 
-    uint32_t              m_size;
+    size_t                m_size;
-    std::vector<uint64_t> m_mask;
+    std::vector<uint32_t> m_code;
-    std::vector<uint64_t> m_code;
+
+    void encodeDword(uint32_t dw);
+
+    uint32_t decodeDword(size_t& offset) const;
 
   };