From: Steinar H. Gunderson Date: Tue, 17 Oct 2017 21:38:32 +0000 (+0200) Subject: Make rans.shader write uint32s, shedding the GL_NV_gpu_shader5 demand. X-Git-Url: https://git.sesse.net/?p=narabu;a=commitdiff_plain;h=0f84eb2a5fa77dbc20f7a073f83bdb674ebb7b75 Make rans.shader write uint32s, shedding the GL_NV_gpu_shader5 demand. --- diff --git a/narabu-encoder.cpp b/narabu-encoder.cpp index 36804d6..fe27064 100644 --- a/narabu-encoder.cpp +++ b/narabu-encoder.cpp @@ -227,10 +227,10 @@ int main(int argc, char **argv) glNamedBufferStorage(output_ssbo, 45 * 64 * 1024, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); check_error(); - GLuint output_offset_ssbo; - glGenBuffers(1, &output_offset_ssbo); - glBindBuffer(GL_SHADER_STORAGE_BUFFER, output_offset_ssbo); - glNamedBufferStorage(output_offset_ssbo, 45 * 64 * sizeof(uint32_t), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); + GLuint bytes_written_ssbo; + glGenBuffers(1, &bytes_written_ssbo); + glBindBuffer(GL_SHADER_STORAGE_BUFFER, bytes_written_ssbo); + glNamedBufferStorage(bytes_written_ssbo, 45 * 64 * sizeof(uint32_t), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); check_error(); // Bind SSBOs. @@ -243,7 +243,7 @@ int main(int argc, char **argv) glUseProgram(glsl_rans_program_num); glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 10, output_ssbo); - glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 11, output_offset_ssbo); + glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 11, bytes_written_ssbo); glBindBufferBase(GL_UNIFORM_BUFFER, 13, dist_ubo); glUseProgram(glsl_program_num); @@ -383,9 +383,8 @@ int main(int argc, char **argv) } // Write out the actual data. - // TODO: Do the deduplication. - const uint32_t *offsets = (const uint32_t *)glMapNamedBufferRange(output_offset_ssbo, 0, 45 * 64 * sizeof(uint32_t), GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); + const uint32_t *bytes_written = (const uint32_t *)glMapNamedBufferRange(bytes_written_ssbo, 0, 45 * 64 * sizeof(uint32_t), GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); #if 0 for (int i = 0; i < 45*64; ++i) { printf("%d,%d,%d: %u\n", i / 64, (i / 8) % 8, i % 8, 1024 * (i + 1) - offsets[i]); @@ -399,8 +398,8 @@ int main(int argc, char **argv) for (unsigned x = 0; x < 8; ++x) { for (unsigned int stream_idx = 0; stream_idx < 45; ++stream_idx) { const uint8_t *out_end = data + (stream_idx * 64 + y * 8 + x + 1) * 1024; - const uint8_t *ptr = data + offsets[stream_idx * 64 + y * 8 + x]; - uint32_t num_rans_bytes = out_end - ptr; + uint32_t num_rans_bytes = bytes_written[stream_idx * 64 + y * 8 + x]; + const uint8_t *ptr = out_end - num_rans_bytes; assert(num_rans_bytes <= 1024); if (num_rans_bytes == last_block.size() && diff --git a/rans.shader b/rans.shader index 39efc52..3e19c83 100644 --- a/rans.shader +++ b/rans.shader @@ -1,5 +1,4 @@ #version 440 -#extension GL_NV_gpu_shader5 : enable layout(local_size_x = 1) in; @@ -7,7 +6,7 @@ const uint prob_bits = 12; const uint prob_scale = 1 << prob_bits; const uint RANS_BYTE_L = (1u << 23); const uint BLOCKS_PER_STREAM = 320; -const uint STREAM_BUF_SIZE = 1024; // 1 kB per stream ought to be enough for everyone :-) +const uint STREAM_BUF_SIZE = 256; // In uint32s. 1 kB per stream ought to be enough for everyone :-) const uint NUM_SYMS = 256; const uint ESCAPE_LIMIT = NUM_SYMS - 1; @@ -26,12 +25,12 @@ const uint luma_mapping[8] = { layout(std430, binding = 10) buffer outputBuf { - uint8_t rans_output[]; + uint rans_output[]; }; layout(std430, binding = 11) buffer outputBuf2 { - uint rans_start_offset[]; + uint rans_bytes_written[]; }; layout(std140, binding = 13) uniform DistBlock @@ -41,11 +40,14 @@ layout(std140, binding = 13) uniform DistBlock }; struct RansEncoder { - uint stream_num; // const - uint lut_base; // const - uint sign_bias; // const + uint stream_num; // const + uint lut_base; // const + uint sign_bias; // const + uint rans_start_offset; // const uint rans_offset; uint rans; + uint bit_buffer; + uint bytes_in_buffer; }; layout(r16ui) uniform restrict readonly uimage2D dc_ac7_tex; @@ -56,7 +58,11 @@ layout(r8i) uniform restrict readonly iimage2D ac4_tex; void RansPutByte(uint x, inout RansEncoder enc) { - rans_output[--enc.rans_offset] = uint8_t(x); + enc.bit_buffer = (enc.bit_buffer << 8) | x; + if (++enc.bytes_in_buffer == 4) { + rans_output[--enc.rans_offset] = enc.bit_buffer; + enc.bytes_in_buffer = 0; + } } void RansEncInit(uint streamgroup_num, uint coeff_row, uint coeff_col, uint dist_num, out RansEncoder enc) @@ -65,7 +71,10 @@ void RansEncInit(uint streamgroup_num, uint coeff_row, uint coeff_col, uint dist 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_start_offset = enc.rans_offset; enc.rans = RANS_BYTE_L; + enc.bit_buffer = 0; + enc.bytes_in_buffer = 0; } void RansEncRenorm(inout RansEncoder enc, uint freq, uint prob_bits) @@ -102,12 +111,21 @@ void RansEncPutSymbol(inout RansEncoder enc, uvec4 sym) enc.rans += bias + (q >> rcp_shift) * cmpl_freq; } -void RansEncFlush(inout RansEncoder enc) +uint RansEncFlush(inout RansEncoder enc) { RansPutByte(enc.rans >> 24, enc); RansPutByte(enc.rans >> 16, enc); RansPutByte(enc.rans >> 8, enc); RansPutByte(enc.rans >> 0, enc); + + uint num_bytes_written = (enc.rans_start_offset - enc.rans_offset) * 4 + enc.bytes_in_buffer; + + // Make sure there's nothing left in the buffer. + RansPutByte(0, enc); + RansPutByte(0, enc); + RansPutByte(0, enc); + + return num_bytes_written; } int sign_extend(uint coeff, uint bits) @@ -137,8 +155,8 @@ void encode_coeff(int signed_k, inout RansEncoder enc) void encode_end(inout RansEncoder enc) { - RansEncFlush(enc); - rans_start_offset[enc.stream_num] = enc.rans_offset; + uint bytes_written = RansEncFlush(enc); + rans_bytes_written[enc.stream_num] = bytes_written; } void encode_9_7(uint streamgroup_num, uint coeff_row, layout(r16ui) restrict readonly uimage2D tex, uint col1, uint col2, uint dist1, uint dist2)