Adding -clbench command line option, and a new API basis_benchmark_etc1s_opencl() to determine if OpenCL encoding is worthwhile on the current machine/driver/GPU.
diff --git a/basisu_tool.cpp b/basisu_tool.cpp
index 08f84e1..9b2a43a 100644
--- a/basisu_tool.cpp
+++ b/basisu_tool.cpp
@@ -54,6 +54,7 @@
cBench,
cCompSize,
cTest,
+ cCLBench,
cSplitImage,
cCombineImages
};
@@ -365,6 +366,8 @@
m_mode = cCompSize;
else if (strcasecmp(pArg, "-test") == 0)
m_mode = cTest;
+ else if (strcasecmp(pArg, "-clbench") == 0)
+ m_mode = cCLBench;
else if (strcasecmp(pArg, "-test_dir") == 0)
{
REMAINING_ARGS_CHECK(1);
@@ -4266,7 +4269,7 @@
size_t data_size = 0;
// Test ETC1S
- flags_and_quality = (opts.m_comp_params.m_multithreading ? cFlagThreaded : 0);
+ flags_and_quality = (opts.m_comp_params.m_multithreading ? cFlagThreaded : 0) | cFlagPrintStats | cFlagPrintStatus;
void* pData = basis_compress(source_images, flags_and_quality, uastc_rdo_quality, &data_size, &stats);
if (!pData)
@@ -4293,7 +4296,7 @@
if (opencl_is_available())
{
- flags_and_quality = (opts.m_comp_params.m_multithreading ? cFlagThreaded : 0) | cFlagUseOpenCL;
+ flags_and_quality = (opts.m_comp_params.m_multithreading ? cFlagThreaded : 0) | cFlagUseOpenCL | cFlagPrintStats | cFlagPrintStatus;
pData = basis_compress(source_images, flags_and_quality, uastc_rdo_quality, &data_size, &stats);
if (!pData)
@@ -4329,7 +4332,7 @@
}
// Test UASTC
- flags_and_quality = (opts.m_comp_params.m_multithreading ? cFlagThreaded : 0) | cFlagUASTC;
+ flags_and_quality = (opts.m_comp_params.m_multithreading ? cFlagThreaded : 0) | cFlagUASTC | cFlagPrintStats | cFlagPrintStatus;
pData = basis_compress(source_images, flags_and_quality, uastc_rdo_quality, &data_size, &stats);
if (!pData)
@@ -4362,6 +4365,24 @@
return result;
}
+static bool clbench_mode(command_line_params& opts)
+{
+ BASISU_NOTE_UNUSED(opts);
+
+ bool opencl_failed = false;
+ bool use_cl = basis_benchmark_etc1s_opencl(&opencl_failed);
+ if (use_cl)
+ printf("OpenCL ETC1S encoding is faster on this machine\n");
+ else
+ {
+ if (opencl_failed)
+ printf("OpenCL failed!\n");
+ printf("CPU ETC1S encoding is faster on this machine\n");
+ }
+
+ return true;
+}
+
static int main_internal(int argc, const char **argv)
{
printf("Basis Universal GPU Texture Compressor v" BASISU_TOOL_VERSION "\nCopyright (C) 2019-2022 Binomial LLC, All rights reserved\n");
@@ -4374,7 +4395,7 @@
bool opencl_force_serialization = false;
for (int i = 1; i < argc; i++)
{
- if (strcmp(argv[i], "-opencl") == 0)
+ if ((strcmp(argv[i], "-opencl") == 0) || (strcmp(argv[i], "-clbench") == 0))
use_opencl = true;
if (strcmp(argv[i], "-opencl_serialize") == 0)
opencl_force_serialization = true;
@@ -4394,13 +4415,13 @@
#if defined(DEBUG) || defined(_DEBUG)
printf("DEBUG build\n");
#endif
-
+
if (argc == 1)
{
print_usage();
return EXIT_FAILURE;
}
-
+
command_line_params opts;
if (!opts.parse(argc, argv))
{
@@ -4413,7 +4434,7 @@
#else
printf("Multithreading: %u, Zstandard support: %u, OpenCL: %u\n", (uint32_t)opts.m_comp_params.m_multithreading, basist::basisu_transcoder_supports_ktx2_zstd(), opencl_is_available());
#endif
-
+
if (!opts.process_listing_files())
return EXIT_FAILURE;
@@ -4459,6 +4480,9 @@
case cTest:
status = test_mode(opts);
break;
+ case cCLBench:
+ status = clbench_mode(opts);
+ break;
case cSplitImage:
status = split_image_mode(opts);
break;
diff --git a/encoder/basisu_comp.cpp b/encoder/basisu_comp.cpp
index 166a1c4..41eae2b 100644
--- a/encoder/basisu_comp.cpp
+++ b/encoder/basisu_comp.cpp
@@ -1501,7 +1501,8 @@
if (m_params.m_compute_stats)
{
- printf("Slice: %u\n", slice_index);
+ if (m_params.m_print_stats)
+ printf("Slice: %u\n", slice_index);
image_stats& s = m_stats[slice_index];
@@ -1511,81 +1512,100 @@
// ---- .basis stats
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked[slice_index], 0, 3);
- em.print(".basis RGB Avg: ");
+ if (m_params.m_print_stats)
+ em.print(".basis RGB Avg: ");
s.m_basis_rgb_avg_psnr = em.m_psnr;
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked[slice_index], 0, 4);
- em.print(".basis RGBA Avg: ");
+ if (m_params.m_print_stats)
+ em.print(".basis RGBA Avg: ");
s.m_basis_rgba_avg_psnr = em.m_psnr;
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked[slice_index], 0, 1);
- em.print(".basis R Avg: ");
+ if (m_params.m_print_stats)
+ em.print(".basis R Avg: ");
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked[slice_index], 1, 1);
- em.print(".basis G Avg: ");
+ if (m_params.m_print_stats)
+ em.print(".basis G Avg: ");
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked[slice_index], 2, 1);
- em.print(".basis B Avg: ");
+ if (m_params.m_print_stats)
+ em.print(".basis B Avg: ");
if (m_params.m_uastc)
{
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked[slice_index], 3, 1);
- em.print(".basis A Avg: ");
+ if (m_params.m_print_stats)
+ em.print(".basis A Avg: ");
s.m_basis_a_avg_psnr = em.m_psnr;
}
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked[slice_index], 0, 0);
- em.print(".basis 709 Luma: ");
+ if (m_params.m_print_stats)
+ em.print(".basis 709 Luma: ");
s.m_basis_luma_709_psnr = static_cast<float>(em.m_psnr);
s.m_basis_luma_709_ssim = static_cast<float>(em.m_ssim);
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked[slice_index], 0, 0, true, true);
- em.print(".basis 601 Luma: ");
+ if (m_params.m_print_stats)
+ em.print(".basis 601 Luma: ");
s.m_basis_luma_601_psnr = static_cast<float>(em.m_psnr);
if (m_slice_descs.size() == 1)
{
const uint32_t output_size = comp_size ? (uint32_t)comp_size : (uint32_t)comp_data.size();
- debug_printf(".basis RGB PSNR per bit/texel*10000: %3.3f\n", 10000.0f * s.m_basis_rgb_avg_psnr / ((output_size * 8.0f) / (slice_desc.m_orig_width * slice_desc.m_orig_height)));
- debug_printf(".basis Luma 709 PSNR per bit/texel*10000: %3.3f\n", 10000.0f * s.m_basis_luma_709_psnr / ((output_size * 8.0f) / (slice_desc.m_orig_width * slice_desc.m_orig_height)));
+ if (m_params.m_print_stats)
+ {
+ debug_printf(".basis RGB PSNR per bit/texel*10000: %3.3f\n", 10000.0f * s.m_basis_rgb_avg_psnr / ((output_size * 8.0f) / (slice_desc.m_orig_width * slice_desc.m_orig_height)));
+ debug_printf(".basis Luma 709 PSNR per bit/texel*10000: %3.3f\n", 10000.0f * s.m_basis_luma_709_psnr / ((output_size * 8.0f) / (slice_desc.m_orig_width * slice_desc.m_orig_height)));
+ }
}
if (m_decoded_output_textures_unpacked_bc7[slice_index].get_width())
{
// ---- BC7 stats
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked_bc7[slice_index], 0, 3);
- em.print("BC7 RGB Avg: ");
+ if (m_params.m_print_stats)
+ em.print("BC7 RGB Avg: ");
s.m_bc7_rgb_avg_psnr = em.m_psnr;
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked_bc7[slice_index], 0, 4);
- em.print("BC7 RGBA Avg: ");
+ if (m_params.m_print_stats)
+ em.print("BC7 RGBA Avg: ");
s.m_bc7_rgba_avg_psnr = em.m_psnr;
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked_bc7[slice_index], 0, 1);
- em.print("BC7 R Avg: ");
+ if (m_params.m_print_stats)
+ em.print("BC7 R Avg: ");
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked_bc7[slice_index], 1, 1);
- em.print("BC7 G Avg: ");
+ if (m_params.m_print_stats)
+ em.print("BC7 G Avg: ");
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked_bc7[slice_index], 2, 1);
- em.print("BC7 B Avg: ");
+ if (m_params.m_print_stats)
+ em.print("BC7 B Avg: ");
if (m_params.m_uastc)
{
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked_bc7[slice_index], 3, 1);
- em.print("BC7 A Avg: ");
+ if (m_params.m_print_stats)
+ em.print("BC7 A Avg: ");
s.m_bc7_a_avg_psnr = em.m_psnr;
}
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked_bc7[slice_index], 0, 0);
- em.print("BC7 709 Luma: ");
+ if (m_params.m_print_stats)
+ em.print("BC7 709 Luma: ");
s.m_bc7_luma_709_psnr = static_cast<float>(em.m_psnr);
s.m_bc7_luma_709_ssim = static_cast<float>(em.m_ssim);
em.calc(m_slice_images[slice_index], m_decoded_output_textures_unpacked_bc7[slice_index], 0, 0, true, true);
- em.print("BC7 601 Luma: ");
+ if (m_params.m_print_stats)
+ em.print("BC7 601 Luma: ");
s.m_bc7_luma_601_psnr = static_cast<float>(em.m_psnr);
}
@@ -1593,16 +1613,19 @@
{
// ---- Nearly best possible ETC1S stats
em.calc(m_slice_images[slice_index], m_best_etc1s_images_unpacked[slice_index], 0, 3);
- em.print("Unquantized ETC1S RGB Avg: ");
+ if (m_params.m_print_stats)
+ em.print("Unquantized ETC1S RGB Avg: ");
s.m_best_etc1s_rgb_avg_psnr = static_cast<float>(em.m_psnr);
em.calc(m_slice_images[slice_index], m_best_etc1s_images_unpacked[slice_index], 0, 0);
- em.print("Unquantized ETC1S 709 Luma: ");
+ if (m_params.m_print_stats)
+ em.print("Unquantized ETC1S 709 Luma: ");
s.m_best_etc1s_luma_709_psnr = static_cast<float>(em.m_psnr);
s.m_best_etc1s_luma_709_ssim = static_cast<float>(em.m_ssim);
em.calc(m_slice_images[slice_index], m_best_etc1s_images_unpacked[slice_index], 0, 0, true, true);
- em.print("Unquantized ETC1S 601 Luma: ");
+ if (m_params.m_print_stats)
+ em.print("Unquantized ETC1S 601 Luma: ");
s.m_best_etc1s_luma_601_psnr = static_cast<float>(em.m_psnr);
}
}
@@ -2311,6 +2334,8 @@
}
comp_params.m_compute_stats = (pStats != nullptr);
+ comp_params.m_print_stats = (flags_and_quality & cFlagPrintStats) != 0;
+ comp_params.m_status_output = (flags_and_quality & cFlagPrintStatus) != 0;
// Create the compressor, initialize it, and process the input
basis_compressor comp;
@@ -2328,6 +2353,11 @@
return nullptr;
}
+ if ((pStats) && (comp.get_opencl_failed()))
+ {
+ pStats->m_opencl_failed = true;
+ }
+
// Get the output file data and return it to the caller
void* pFile_data = nullptr;
const uint8_vec* pFile_data_vec = comp_params.m_create_ktx2_file ? &comp.get_output_ktx2_file() : &comp.get_output_basis_file();
@@ -2388,4 +2418,108 @@
free(p);
}
+ bool basis_benchmark_etc1s_opencl(bool* pOpenCL_failed)
+ {
+ if (pOpenCL_failed)
+ *pOpenCL_failed = false;
+
+ if (!opencl_is_available())
+ {
+ error_printf("basis_benchmark_etc1s_opencl: OpenCL support must be enabled first!\n");
+ return false;
+ }
+
+ const uint32_t W = 1024, H = 1024;
+ basisu::vector<image> images;
+ image& img = images.enlarge(1)->resize(W, H);
+
+ const uint32_t NUM_RAND_LETTERS = 6000;// 40000;
+
+ rand r;
+ r.seed(200);
+
+ for (uint32_t i = 0; i < NUM_RAND_LETTERS; i++)
+ {
+ uint32_t x = r.irand(0, W - 1), y = r.irand(0, H - 1);
+ uint32_t sx = r.irand(1, 4), sy = r.irand(1, 4);
+ color_rgba c(r.byte(), r.byte(), r.byte(), 255);
+
+ img.debug_text(x, y, sx, sy, c, nullptr, false, "%c", static_cast<char>(r.irand(32, 127)));
+ }
+
+ //save_png("test.png", img);
+
+ image_stats stats;
+
+ uint32_t flags_and_quality = cFlagSRGB | cFlagThreaded | 255;
+ size_t comp_size = 0;
+
+ double best_cpu_time = 1e+9f, best_gpu_time = 1e+9f;
+
+ const uint32_t TIMES_TO_ENCODE = 2;
+ interval_timer tm;
+
+ for (uint32_t i = 0; i < TIMES_TO_ENCODE; i++)
+ {
+ tm.start();
+ void* pComp_data = basis_compress(
+ images,
+ flags_and_quality, 1.0f,
+ &comp_size,
+ &stats);
+ double cpu_time = tm.get_elapsed_secs();
+ if (!pComp_data)
+ {
+ error_printf("basis_benchmark_etc1s_opencl: basis_compress() failed (CPU)!\n");
+ return false;
+ }
+
+ best_cpu_time = minimum(best_cpu_time, cpu_time);
+
+ basis_free_data(pComp_data);
+ }
+
+ printf("Best CPU time: %3.3f\n", best_cpu_time);
+
+ for (uint32_t i = 0; i < TIMES_TO_ENCODE; i++)
+ {
+ tm.start();
+ void* pComp_data = basis_compress(
+ images,
+ flags_and_quality | cFlagUseOpenCL, 1.0f,
+ &comp_size,
+ &stats);
+
+ if (stats.m_opencl_failed)
+ {
+ error_printf("basis_benchmark_etc1s_opencl: OpenCL failed!\n");
+
+ basis_free_data(pComp_data);
+
+ if (pOpenCL_failed)
+ *pOpenCL_failed = true;
+
+ return false;
+ }
+
+ double gpu_time = tm.get_elapsed_secs();
+ if (!pComp_data)
+ {
+ error_printf("basis_benchmark_etc1s_opencl: basis_compress() failed (GPU)!\n");
+ return false;
+ }
+
+ best_gpu_time = minimum(best_gpu_time, gpu_time);
+
+ basis_free_data(pComp_data);
+ }
+
+ printf("Best GPU time: %3.3f\n", best_gpu_time);
+
+ return best_gpu_time < best_cpu_time;
+ }
+
} // namespace basisu
+
+
+
diff --git a/encoder/basisu_comp.h b/encoder/basisu_comp.h
index aa5ea6f..b6c9fef 100644
--- a/encoder/basisu_comp.h
+++ b/encoder/basisu_comp.h
@@ -92,6 +92,8 @@
m_best_etc1s_luma_709_psnr = 0.0f;
m_best_etc1s_luma_601_psnr = 0.0f;
m_best_etc1s_luma_709_ssim = 0.0f;
+
+ m_opencl_failed = false;
}
std::string m_filename;
@@ -119,6 +121,8 @@
float m_best_etc1s_luma_709_psnr;
float m_best_etc1s_luma_601_psnr;
float m_best_etc1s_luma_709_ssim;
+
+ bool m_opencl_failed;
};
template<bool def>
@@ -255,6 +259,7 @@
m_write_output_basis_files.clear();
m_compression_level.clear();
m_compute_stats.clear();
+ m_print_stats.clear();
m_check_for_alpha.clear();
m_force_alpha.clear();
m_multithreading.clear();
@@ -373,6 +378,9 @@
// Compute and display image metrics
bool_param<false> m_compute_stats;
+
+ // Print stats to stdout, if m_compute_stats is true.
+ bool_param<true> m_print_stats;
// Check to see if any input image has an alpha channel, if so then the output basis file will have alpha channels
bool_param<true> m_check_for_alpha;
@@ -583,11 +591,16 @@
cFlagYFlip = 1 << 16, // flip source image on Y axis before compression
cFlagUASTC = 1 << 17, // use UASTC compression vs. ETC1S
- cFlagUASTCRDO = 1 << 18 // use RDO postprocessing when generating UASTC files (must set uastc_rdo_quality to the quality scalar)
+ cFlagUASTCRDO = 1 << 18, // use RDO postprocessing when generating UASTC files (must set uastc_rdo_quality to the quality scalar)
+
+ cFlagPrintStats = 1 << 19, // print image stats to stdout
+ cFlagPrintStatus = 1 << 20 // print status to stdout
};
// This function accepts an array of source images.
// If more than one image is provided, it's assumed the images form a mipmap pyramid and automatic mipmap generation is disabled.
+ // Returns a pointer to the compressed .basis or .ktx2 file data. *pSize is the size of the compressed data. The returned block must be freed using basis_free_data().
+ // basisu_encoder_init() MUST be called first!
void* basis_compress(
const basisu::vector<image> &source_images,
uint32_t flags_and_quality, float uastc_rdo_quality,
@@ -604,6 +617,12 @@
// Frees the dynamically allocated file data returned by basis_compress().
void basis_free_data(void* p);
+ // Runs a short benchmark using synthetic image data to time OpenCL encoding vs. CPU encoding, with multithreading enabled.
+ // Returns true if opencl is worth using on this system, otherwise false.
+ // If pOpenCL_failed is not null, it will be set to true if OpenCL encoding failed *on this particular machine/driver/BasisU version* and the encoder falled back to CPU encoding.
+ // basisu_encoder_init() MUST be called first. If OpenCL support wasn't enabled this always returns false.
+ bool basis_benchmark_etc1s_opencl(bool *pOpenCL_failed = nullptr);
+
// Parallel compression API
struct parallel_results
{