Kill the division in the rANS GPU encoder.

[narabu] / rans.shader

diff --git a/rans.shader b/rans.shader
index 5b13d7ea75e25c2d63f486fb59a8c5bc54019d00..b677415952f89ef67f37c69926a71759357a229b 100644 (file)
--- a/rans.shader
+++ b/rans.shader
@@ -27,7 +27,9 @@ const uint luma_mapping[8] = {
 layout(std430, binding = 9) buffer layoutName
 {
        uint dist[4 * 256];
-       uvec2 ransdist[4 * 256];
+       uint ransfreq[4 * 256];
+       uvec4 ransdist[4 * 256];
+       uint sign_biases[4];
 };
 
 layout(std430, binding = 10) buffer outputBuf
@@ -43,6 +45,7 @@ layout(std430, binding = 11) buffer outputBuf2
 struct RansEncoder {
        uint stream_num;   // const
        uint lut_base;     // const
+       uint sign_bias;    // const
        uint rans_offset;
        uint rans;
 };
@@ -57,6 +60,7 @@ void RansEncInit(uint streamgroup_num, uint coeff_row, uint coeff_col, uint dist
 {
        enc.stream_num = streamgroup_num * 64 + coeff_row * 8 + coeff_col;
        enc.lut_base = dist_num * 256;
+       enc.sign_bias = sign_biases[dist_num];
        enc.rans_offset = enc.stream_num * STREAM_BUF_SIZE + STREAM_BUF_SIZE;  // Starts at the end.
        enc.rans = RANS_BYTE_L;
 }
@@ -64,11 +68,9 @@ void RansEncInit(uint streamgroup_num, uint coeff_row, uint coeff_col, uint dist
 void RansEncRenorm(inout uint rans, inout uint rans_offset, uint freq, uint prob_bits)
 {
        uint x_max = ((RANS_BYTE_L >> prob_bits) << 8) * freq; // this turns into a shift.
-       if (rans >= x_max) {
-               do {
-                       rans_output[--rans_offset] = uint8_t(rans & 0xff);
-                       rans >>= 8;
-               } while (rans >= x_max);
+       while (rans >= x_max) {
+               rans_output[--rans_offset] = uint8_t(rans & 0xffu);
+               rans >>= 8;
        }
 }
 
@@ -78,6 +80,27 @@ void RansEncPut(inout uint rans, inout uint rans_offset, uint start, uint freq,
        rans = ((rans / freq) << prob_bits) + (rans % freq) + start;
 }
 
+void RansEncPutSymbol(inout uint rans, inout uint rans_offset, uvec4 sym)
+{
+       uint x_max = sym.x;
+       uint rcp_freq = sym.y;
+       uint bias = sym.z;
+       uint rcp_shift = (sym.w & 0xffffu);
+       uint cmpl_freq = (sym.w >> 16);
+
+       // renormalize
+       if (rans >= x_max) {
+               do {
+                       rans_output[--rans_offset] = uint8_t(rans & 0xffu);
+                       rans >>= 8;
+               } while (rans >= x_max);
+       }
+
+       uint q, unused;
+       umulExtended(rans, rcp_freq, q, unused);
+       rans += bias + (q >> rcp_shift) * cmpl_freq;
+}
+
 void RansEncFlush(uint rans, inout uint rans_offset)
 {
        rans_offset -= 4;
@@ -92,7 +115,7 @@ int sign_extend(uint coeff, uint bits)
        return int(coeff << (32 - bits)) >> (32 - bits);
 }
 
-void encode_coeff(int signed_k, uint sign_bias, inout RansEncoder enc)
+void encode_coeff(int signed_k, inout RansEncoder enc)
 {
        uint k = abs(signed_k);
 
@@ -104,11 +127,11 @@ void encode_coeff(int signed_k, uint sign_bias, inout RansEncoder enc)
                k = ESCAPE_LIMIT;
        }
 
-       uvec2 sym = ransdist[enc.lut_base + ((k - 1) & (NUM_SYMS - 1))];
-       RansEncPut(enc.rans, enc.rans_offset, sym.x, sym.y, prob_bits + 1);
+       uvec4 sym = ransdist[enc.lut_base + ((k - 1) & (NUM_SYMS - 1))];
+       RansEncPutSymbol(enc.rans, enc.rans_offset, sym);
        
        if (signed_k < 0) {
-               enc.rans += sign_bias;
+               enc.rans += enc.sign_bias;
        }
 }
 
@@ -124,17 +147,14 @@ void encode_9_7(uint streamgroup_num, uint coeff_row, layout(r16ui) restrict rea
        RansEncInit(streamgroup_num, coeff_row, col1, dist1, enc1);
        RansEncInit(streamgroup_num, coeff_row, col2, dist2, enc2);
 
-       uint sign_bias1 = ransdist[enc1.lut_base + 255].x + ransdist[enc1.lut_base + 255].y;
-       uint sign_bias2 = ransdist[enc2.lut_base + 255].x + ransdist[enc2.lut_base + 255].y;
-
        for (uint subblock_idx = 0; subblock_idx < BLOCKS_PER_STREAM; ++subblock_idx) {
                // TODO: Use SSBOs instead of a texture?
                uint x = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) % 160;
                uint y = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) / 160;
                uint f = imageLoad(tex, ivec2(x, y * 8 + coeff_row)).x;
 
-               encode_coeff(sign_extend(f & 0x1ffu, 9), sign_bias1, enc1);
-               encode_coeff(sign_extend(f >> 9, 7), sign_bias2, enc2);
+               encode_coeff(sign_extend(f & 0x1ffu, 9), enc1);
+               encode_coeff(sign_extend(f >> 9, 7), enc2);
        }
 
        encode_end(enc1);
@@ -146,15 +166,13 @@ void encode_8(uint streamgroup_num, uint coeff_row, layout(r8i) restrict readonl
        RansEncoder enc;
        RansEncInit(streamgroup_num, coeff_row, col, dist, enc);
 
-       uint sign_bias = ransdist[enc.lut_base + 255].x + ransdist[enc.lut_base + 255].y;
-
        for (uint subblock_idx = 0; subblock_idx < BLOCKS_PER_STREAM; ++subblock_idx) {
                // TODO: Use SSBOs instead of a texture?
                uint x = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) % 160;
                uint y = (streamgroup_num * BLOCKS_PER_STREAM + subblock_idx) / 160;
                int f = imageLoad(tex, ivec2(x, y * 8 + coeff_row)).x;
 
-               encode_coeff(f, sign_bias, enc);
+               encode_coeff(f, enc);
        }
 
        encode_end(enc);