curioustorvald/tsvm
1
0
Fork 0
mirror of https://github.com/curioustorvald/tsvm.git synced 2026-03-07 19:51:51 +09:00
Code Issues Packages Projects Releases Wiki Activity

TAV: code cleanup, better preset selection

Browse Source
This commit is contained in:
minjaesong
2025-11-24 11:16:19 +09:00
parent dbbb471a11
commit 6132012e74
5 changed files with 93 additions and 152 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tsvm_core/src/net/torvald/tsvm/GraphicsJSR223Delegate.kt
View File

@@ -5874,7 +5874,7 @@ class GraphicsJSR223Delegate(private val vm: VM) {
val currentCo = FloatArray(coeffCount)
val currentCg = FloatArray(coeffCount)
// Delta-specific perceptual reconstruction using motion-optimized coefficients
// Delta-specific perceptual reconstruction using motion-optimised coefficients
// Estimate quality level from quantisation parameters for perceptual weighting
val estimatedQualityY = when {
qY <= 6 -> 4 // High quality

6
video_encoder/decoder_tad.c
View File

@@ -1065,12 +1065,6 @@ int main(int argc, char *argv[]) {
size_t name_len = ext - basename_start;
strncpy(output_file + dir_len, basename_start, name_len);
output_file[dir_len + name_len] = '\0';
// Replace last dot with underscore (for .qNN pattern)
/*char *last_dot = strrchr(output_file, '.');
if (last_dot && last_dot > output_file + dir_len) {
*last_dot = '_';
}*/
} else {
// No .tad extension, copy entire basename
strcpy(output_file + dir_len, basename_start);

44
video_encoder/decoder_tav.c
View File

@@ -15,7 +15,7 @@
#include <getopt.h>
#include <signal.h>
#include "decoder_tad.h" // Shared TAD decoder library
#include "tav_avx512.h" // AVX-512 SIMD optimizations
#include "tav_avx512.h" // AVX-512 SIMD optimisations
#define DECODER_VENDOR_STRING "Decoder-TAV 20251124 (avx512)"
@@ -315,7 +315,7 @@ static void dequantise_dwt_subbands_perceptual(int q_index, int q_y_global, cons
// for bright pixels, creating dark DWT-pattern blemishes
#ifdef __AVX512F__
// Use AVX-512 optimized dequantization if available (1.1x speedup against -Ofast)
// Use AVX-512 optimised dequantization if available (1.1x speedup against -Ofast)
// Check: subband has >=16 elements AND won't exceed buffer bounds
const int subband_end = subband->coeff_start + subband->coeff_count;
if (g_simd_level >= SIMD_AVX512F && subband->coeff_count >= 16 && subband_end <= coeff_count) {
@@ -2713,7 +2713,7 @@ static void print_usage(const char *prog) {
printf("Usage: %s -i input.tav -o output.mkv\n\n", prog);
printf("Options:\n");
printf(" -i <file> Input TAV file\n");
printf(" -o <file> Output MKV file (FFV1 video + PCMu8 audio)\n");
printf(" -o <file> Output MKV file (optional, auto-generated from input)\n");
printf(" -v Verbose output\n");
printf(" -h, --help Show this help\n\n");
printf("Supported features (matches TSVM decoder):\n");
@@ -2767,12 +2767,46 @@ int main(int argc, char *argv[]) {
}
}
if (!input_file || !output_file) {
fprintf(stderr, "Error: Both input and output files are required\n\n");
if (!input_file) {
fprintf(stderr, "Error: Input file is required\n\n");
print_usage(argv[0]);
return 1;
}
// Generate output filename if not provided
if (!output_file) {
size_t input_len = strlen(input_file);
output_file = malloc(input_len + 32); // Extra space for extension
// Find the last directory separator
const char *basename_start = strrchr(input_file, '/');
if (!basename_start) basename_start = strrchr(input_file, '\\');
basename_start = basename_start ? basename_start + 1 : input_file;
// Copy directory part
size_t dir_len = basename_start - input_file;
strncpy(output_file, input_file, dir_len);
// Find the .tad extension
const char *ext = strrchr(basename_start, '.');
if (ext && (strcmp(ext, ".tav") == 0 || strcmp(ext, ".mv3") == 0)) {
// Copy basename without .tav or .mv3
size_t name_len = ext - basename_start;
strncpy(output_file + dir_len, basename_start, name_len);
output_file[dir_len + name_len] = '\0';
} else {
// No .tad extension, copy entire basename
strcpy(output_file + dir_len, basename_start);
}
// Append appropriate extension
strcat(output_file, ".mkv");
if (verbose) {
printf("Auto-generated output path: %s\n", output_file);
}
}
// Create temporary audio file path
char temp_audio_file[256];
snprintf(temp_audio_file, sizeof(temp_audio_file), "/tmp/tav_audio_%d.wav", getpid());

161
video_encoder/encoder_tav.c
View File

@@ -17,7 +17,7 @@
#include <time.h>
#include <limits.h>
#include <float.h>
#include "tav_avx512.h" // AVX-512 SIMD optimizations
#include "tav_avx512.h" // AVX-512 SIMD optimisations
#define ENCODER_VENDOR_STRING "Encoder-TAV 20251124 (3d-dwt,tad,ssf-tc,cdf53-motion,avx512)"
@@ -2383,10 +2383,6 @@ static void quantise_dwt_coefficients_perceptual_per_coeff(tav_encoder_t *enc,
float *coeffs, int16_t *quantised, int size,
int base_quantiser, int width, int height,
int decomp_levels, int is_chroma, int frame_count);
static void quantise_dwt_coefficients_perceptual_per_coeff_no_normalisation(tav_encoder_t *enc,
float *coeffs, int16_t *quantised, int size,
int base_quantiser, int width, int height,
int decomp_levels, int is_chroma, int frame_count);
static size_t preprocess_coefficients_variable_layout(preprocess_mode_t preprocess_mode, int width, int height,
int16_t *coeffs_y, int16_t *coeffs_co, int16_t *coeffs_cg, int16_t *coeffs_alpha,
int coeff_count, int channel_layout, uint8_t *output_buffer);
@@ -2424,7 +2420,7 @@ static void show_usage(const char *program_name) {
printf(" --enable-delta Enable delta encoding\n");
printf(" --delta-haar N Apply N-level Haar DWT to delta coefficients (1-6, auto-enables delta)\n");
printf(" --3d-dwt Enable temporal 3D DWT (GOP-based encoding with temporal transform; the default encoding mode)\n");
printf(" --motion-coder N Temporal wavelet: 0=Haar, 1=CDF 5/3 (default: auto-select based on resolution; use 0 for older version compatibility)\n");
printf(" --motion-coder N Temporal wavelet: 0=Haar, 1=CDF 5/3 (default: auto-select based on resolution; use 0 for older version compatibility; use 1 for smoother motion)\n");
printf(" --single-pass Disable two-pass encoding with wavelet-based scene change detection (optimal GOP boundaries)\n");
// printf(" --mc-ezbc Enable MC-EZBC block-based motion compensation (requires --temporal-dwt, implies --ezbc)\n");
printf(" --ezbc Enable EZBC (Embedded Zero Block Coding) entropy coding. May help reducing file size on high-quality videos\n");
@@ -3388,37 +3384,18 @@ static void quantise_3d_dwt_coefficients(tav_encoder_t *enc,
// Q_effective = tH_base × spatial_weight
// Where spatial_weight depends on spatial frequency (LL, LH, HL, HH subbands)
// This reuses all existing perceptual weighting and dead-zone logic
//
// CRITICAL: Use no_normalisation variant when EZBC is enabled
// - EZBC mode: coefficients must be denormalised (quantise + multiply back)
// - Twobit-map/raw mode: coefficients stay normalised (quantise only)
if (enc->preprocess_mode == PREPROCESS_EZBC) {
quantise_dwt_coefficients_perceptual_per_coeff_no_normalisation(
enc,
gop_coeffs[t], // Input: spatial coefficients for this temporal subband
quantised[t], // Output: quantised spatial coefficients (denormalised for EZBC)
spatial_size, // Number of spatial coefficients
temporal_base_quantiser, // Temporally-scaled base quantiser (tH_base)
enc->width, // Frame width
enc->height, // Frame height
enc->decomp_levels, // Spatial decomposition levels (typically 6)
is_chroma, // Is chroma channel (gets additional quantisation)
enc->frame_count + t // Frame number (for any frame-dependent logic)
);
} else {
quantise_dwt_coefficients_perceptual_per_coeff(
enc,
gop_coeffs[t], // Input: spatial coefficients for this temporal subband
quantised[t], // Output: quantised spatial coefficients (normalised for twobit-map)
spatial_size, // Number of spatial coefficients
temporal_base_quantiser, // Temporally-scaled base quantiser (tH_base)
enc->width, // Frame width
enc->height, // Frame height
enc->decomp_levels, // Spatial decomposition levels (typically 6)
is_chroma, // Is chroma channel (gets additional quantisation)
enc->frame_count + t // Frame number (for any frame-dependent logic)
);
}
quantise_dwt_coefficients_perceptual_per_coeff(
enc,
gop_coeffs[t], // Input: spatial coefficients for this temporal subband
quantised[t], // Output: quantised spatial coefficients (normalised for twobit-map)
spatial_size, // Number of spatial coefficients
temporal_base_quantiser, // Temporally-scaled base quantiser (tH_base)
enc->width, // Frame width
enc->height, // Frame height
enc->decomp_levels, // Spatial decomposition levels (typically 6)
is_chroma, // Is chroma channel (gets additional quantisation)
enc->frame_count + t // Frame number (for any frame-dependent logic)
);
if (enc->verbose && (t == 0 || t == num_frames - 1)) {
printf(" Temporal subband %d: level=%d, tH_base=%d\n",
@@ -4092,13 +4069,13 @@ static size_t encode_pframe_residual(tav_encoder_t *enc, int qY) {
// EZBC mode: Quantise with perceptual weighting but no normalisation (division by quantiser)
// EZBC will compress by encoding only significant bitplanes
// fprintf(stderr, "[EZBC-QUANT-PFRAME] Using perceptual quantisation without normalisation\n");
quantise_dwt_coefficients_perceptual_per_coeff_no_normalisation(enc, residual_y_dwt, quantised_y, frame_size,
quantise_dwt_coefficients_perceptual_per_coeff(enc, residual_y_dwt, quantised_y, frame_size,
qY, enc->width, enc->height,
enc->decomp_levels, 0, 0);
quantise_dwt_coefficients_perceptual_per_coeff_no_normalisation(enc, residual_co_dwt, quantised_co, frame_size,
quantise_dwt_coefficients_perceptual_per_coeff(enc, residual_co_dwt, quantised_co, frame_size,
enc->quantiser_co, enc->width, enc->height,
enc->decomp_levels, 1, 0);
quantise_dwt_coefficients_perceptual_per_coeff_no_normalisation(enc, residual_cg_dwt, quantised_cg, frame_size,
quantise_dwt_coefficients_perceptual_per_coeff(enc, residual_cg_dwt, quantised_cg, frame_size,
enc->quantiser_cg, enc->width, enc->height,
enc->decomp_levels, 1, 0);
@@ -6431,7 +6408,7 @@ static void quantise_dwt_coefficients(float *coeffs, int16_t *quantised, int siz
effective_q = FCLAMP(effective_q, 1.0f, 4096.0f);
#ifdef __AVX512F__
// Use AVX-512 optimized version if available (2x speedup against -Ofast)
// Use AVX-512 optimised version if available (2x speedup against -Ofast)
if (g_simd_level >= SIMD_AVX512F) {
quantise_dwt_coefficients_avx512(coeffs, quantised, size, effective_q, dead_zone_threshold,
width, height, decomp_levels, is_chroma,
@@ -6739,76 +6716,6 @@ static void quantise_dwt_coefficients_perceptual_per_coeff(tav_encoder_t *enc,
}
}
// Quantisation for EZBC mode: quantises to discrete levels but doesn't normalise (shrink) values
// This reduces coefficient precision while preserving magnitude for EZBC's bitplane encoding
static void quantise_dwt_coefficients_perceptual_per_coeff_no_normalisation(tav_encoder_t *enc,
float *coeffs, int16_t *quantised, int size,
int base_quantiser, int width, int height,
int decomp_levels, int is_chroma, int frame_count) {
(void)frame_count; // Unused parameter
float effective_base_q = base_quantiser;
effective_base_q = FCLAMP(effective_base_q, 1.0f, 4096.0f);
for (int i = 0; i < size; i++) {
// Apply perceptual weight based on coefficient's position in DWT layout
float weight = get_perceptual_weight_for_position(enc, i, width, height, decomp_levels, is_chroma);
float effective_q = effective_base_q * weight;
// Step 1: Quantise - divide by quantiser to get normalised value
float quantised_val = coeffs[i] / effective_q;
// Step 2: Apply dead-zone quantisation to normalised value
if (enc->dead_zone_threshold > 0.0f && !is_chroma) {
int level = get_subband_level(i, width, height, decomp_levels);
int subband_type = get_subband_type(i, width, height, decomp_levels);
float level_threshold = 0.0f;
if (level == 1) {
// Finest level (level 1: 280x224)
if (subband_type == 3) {
// HH1: full dead-zone
level_threshold = enc->dead_zone_threshold * DEAD_ZONE_FINEST_SCALE;
} else if (subband_type == 1 || subband_type == 2) {
// LH1, HL1: half dead-zone
level_threshold = enc->dead_zone_threshold * DEAD_ZONE_FINE_SCALE;
}
} else if (level == 2) {
// Second-finest level (level 2: 140x112)
if (subband_type == 3) {
// HH2: half dead-zone
level_threshold = enc->dead_zone_threshold * DEAD_ZONE_FINE_SCALE;
}
// LH2, HL2: no dead-zone
}
// Coarser levels (3-6): no dead-zone to preserve structural information
if (fabsf(quantised_val) <= level_threshold) {
quantised_val = 0.0f;
}
}
// Step 3: Round to discrete quantisation levels
quantised_val = roundf(quantised_val); // file size explodes without rounding
// FIX: Store normalised values (not denormalised) to avoid int16_t overflow
// EZBC bitplane encoding works fine with normalised coefficients
// Denormalisation was causing bright pixels to clip at 32767
quantised[i] = (int16_t)CLAMP((int)quantised_val, -32768, 32767);
// Debug: Print LL subband coefficients (9×7 at top-left for 560×448)
/*static int debug_once = 1;
if (debug_once && i < 63 && width == 560 && !is_chroma) {
int x = i % width;
int y = i / width;
if (x < 9 && y < 7) {
fprintf(stderr, "[EZBC-QUANT-DEBUG] LL coeff[%d,%d] (idx=%d): coeff=%.1f, weight=%.3f, effective_q=%.1f, quantised_val=%.1f, stored=%d\n",
x, y, i, coeffs[i], weight, effective_q, quantised_val, quantised[i]);
if (i == 62) debug_once = 0;
}
}*/
}
}
// Serialise tile data for compression
static size_t serialise_tile_data(tav_encoder_t *enc, int tile_x, int tile_y,
@@ -6874,9 +6781,9 @@ static size_t serialise_tile_data(tav_encoder_t *enc, int tile_x, int tile_y,
if (enc->preprocess_mode == PREPROCESS_EZBC) {
// EZBC mode: Quantise with perceptual weighting but no normalisation (division by quantiser)
// fprintf(stderr, "[EZBC-QUANT-INTRA] Using perceptual quantisation without normalisation\n");
quantise_dwt_coefficients_perceptual_per_coeff_no_normalisation(enc, (float*)tile_y_data, quantised_y, tile_size, this_frame_qY, enc->temporal_gop_width, enc->temporal_gop_height, enc->decomp_levels, 0, enc->frame_count);
quantise_dwt_coefficients_perceptual_per_coeff_no_normalisation(enc, (float*)tile_co_data, quantised_co, tile_size, this_frame_qCo, enc->temporal_gop_width, enc->temporal_gop_height, enc->decomp_levels, 1, enc->frame_count);
quantise_dwt_coefficients_perceptual_per_coeff_no_normalisation(enc, (float*)tile_cg_data, quantised_cg, tile_size, this_frame_qCg, enc->temporal_gop_width, enc->temporal_gop_height, enc->decomp_levels, 1, enc->frame_count);
quantise_dwt_coefficients_perceptual_per_coeff(enc, (float*)tile_y_data, quantised_y, tile_size, this_frame_qY, enc->temporal_gop_width, enc->temporal_gop_height, enc->decomp_levels, 0, enc->frame_count);
quantise_dwt_coefficients_perceptual_per_coeff(enc, (float*)tile_co_data, quantised_co, tile_size, this_frame_qCo, enc->temporal_gop_width, enc->temporal_gop_height, enc->decomp_levels, 1, enc->frame_count);
quantise_dwt_coefficients_perceptual_per_coeff(enc, (float*)tile_cg_data, quantised_cg, tile_size, this_frame_qCg, enc->temporal_gop_width, enc->temporal_gop_height, enc->decomp_levels, 1, enc->frame_count);
// Print max abs for debug
int max_y = 0, max_co = 0, max_cg = 0;
@@ -8725,7 +8632,7 @@ static int detect_letterbox_pillarbox(tav_encoder_t *enc,
for (int x = 0; x < width; x += SAMPLE_RATE_HORZ) {
int idx = y * width + x;
// Use pre-converted YCoCg values (optimization: avoid RGB→YCoCg conversion in loop)
// Use pre-converted YCoCg values (optimisation: avoid RGB→YCoCg conversion in loop)
float yval = enc->current_frame_y[idx];
float co = enc->current_frame_co[idx];
float cg = enc->current_frame_cg[idx];
@@ -8764,7 +8671,7 @@ static int detect_letterbox_pillarbox(tav_encoder_t *enc,
for (int x = 0; x < width; x += SAMPLE_RATE_HORZ) {
int idx = y * width + x;
// Use pre-converted YCoCg values (optimization)
// Use pre-converted YCoCg values (optimisation)
float yval = enc->current_frame_y[idx];
float co = enc->current_frame_co[idx];
float cg = enc->current_frame_cg[idx];
@@ -8800,7 +8707,7 @@ static int detect_letterbox_pillarbox(tav_encoder_t *enc,
for (int y = 0; y < height; y += SAMPLE_RATE_VERT) {
int idx = y * width + x;
// Use pre-converted YCoCg values (optimization)
// Use pre-converted YCoCg values (optimisation)
float yval = enc->current_frame_y[idx];
float co = enc->current_frame_co[idx];
float cg = enc->current_frame_cg[idx];
@@ -8836,7 +8743,7 @@ static int detect_letterbox_pillarbox(tav_encoder_t *enc,
for (int y = 0; y < height; y += SAMPLE_RATE_VERT) {
int idx = y * width + x;
// Use pre-converted YCoCg values (optimization)
// Use pre-converted YCoCg values (optimisation)
float yval = enc->current_frame_y[idx];
float co = enc->current_frame_co[idx];
float cg = enc->current_frame_cg[idx];
@@ -11147,17 +11054,23 @@ int main(int argc, char *argv[]) {
// For larger videos, use Haar (better compression, smoother motion matters less)
if (enc->temporal_motion_coder == -1) {
int num_pixels = enc->width * enc->height;
if (num_pixels >= 500000) {
if (
num_pixels >= 820000 && enc->quantiser_y <= 29 ||
num_pixels >= 500000 && enc->quantiser_y <= 14 ||
num_pixels >= 340000 && enc->quantiser_y <= 7 ||
num_pixels >= 260000 && enc->quantiser_y <= 3
) {
enc->temporal_motion_coder = 0; // Haar
if (enc->verbose) {
printf("Auto-selected Haar temporal wavelet (resolution: %dx%d = %d pixels)\n",
enc->width, enc->height, num_pixels);
printf("Auto-selected Haar temporal wavelet (resolution: %dx%d = %d pixels, quantiser_y = %d)\n",
enc->width, enc->height, num_pixels, enc->quantiser_y);
}
} else {
}
else {
enc->temporal_motion_coder = 1; // CDF 5/3
if (enc->verbose) {
printf("Auto-selected CDF 5/3 temporal wavelet (resolution: %dx%d = %d pixels)\n",
enc->width, enc->height, num_pixels);
printf("Auto-selected CDF 5/3 temporal wavelet (resolution: %dx%d = %d pixels, quantiser_y = %d)\n",
enc->width, enc->height, num_pixels, enc->quantiser_y);
}
}
}

32
video_encoder/tav_avx512.h
View File

@@ -1,13 +1,13 @@
/*
* TAV AVX-512 Optimizations
* TAV AVX-512 Optimisations
*
* This file contains AVX-512 optimized versions of performance-critical functions
* This file contains AVX-512 optimised versions of performance-critical functions
* in the TAV encoder. Runtime CPU detection ensures fallback to scalar versions
* on non-AVX-512 systems.
*
* Optimized functions:
* Optimised functions:
* - 1D DWT transforms (5/3, 9/7, Haar, Bior13/7, DD4)
* - Quantization functions
* - Quantisation functions
* - RGB to YCoCg color conversion
* - 2D DWT gather/scatter operations
*
@@ -51,7 +51,7 @@ static inline void tav_simd_init(void) {
#ifdef __AVX512F__
if (cpu_has_avx512f()) {
g_simd_level = SIMD_AVX512F;
fprintf(stderr, "[TAV] AVX-512 optimizations enabled\n");
fprintf(stderr, "[TAV] AVX-512 optimisations enabled\n");
} else {
g_simd_level = SIMD_NONE;
fprintf(stderr, "[TAV] AVX-512 not available, using scalar fallback\n");
@@ -85,7 +85,7 @@ static inline __m512 _mm512_clamp_ps(__m512 v, __m512 min_val, __m512 max_val) {
}
// =============================================================================
// AVX-512 Optimized 1D DWT Forward Transforms
// AVX-512 Optimised 1D DWT Forward Transforms
// =============================================================================
// 5/3 Reversible Forward DWT with AVX-512
@@ -425,7 +425,7 @@ static inline void dwt_haar_forward_1d_avx512(float *data, int length) {
}
// =============================================================================
// AVX-512 Optimized Quantization Functions
// AVX-512 Optimised Quantisation Functions
// =============================================================================
static inline void quantise_dwt_coefficients_avx512(
@@ -488,7 +488,7 @@ static inline void quantise_dwt_coefficients_avx512(
}
}
// Perceptual quantization with per-coefficient weighting
// Perceptual quantisation with per-coefficient weighting
static inline void quantise_dwt_coefficients_perceptual_avx512(
float *coeffs, int16_t *quantised, int size,
float *weights, // Pre-computed per-coefficient weights
@@ -534,10 +534,10 @@ static inline void quantise_dwt_coefficients_perceptual_avx512(
}
// =============================================================================
// AVX-512 Optimized Dequantization Functions
// AVX-512 Optimised Dequantisation Functions
// =============================================================================
// Basic dequantization: quantised[i] * effective_q
// Basic dequantisation: quantised[i] * effective_q
static inline void dequantise_dwt_coefficients_avx512(
const int16_t *quantised, float *coeffs, int size,
float effective_q
@@ -555,7 +555,7 @@ static inline void dequantise_dwt_coefficients_avx512(
// Convert int32 to float
__m512 quant_f32 = _mm512_cvtepi32_ps(quant_i32);
// Multiply by quantizer
// Multiply by quantiser
__m512 dequant = _mm512_mul_ps(quant_f32, q_vec);
_mm512_storeu_ps(&coeffs[i], dequant);
@@ -567,7 +567,7 @@ static inline void dequantise_dwt_coefficients_avx512(
}
}
// Perceptual dequantization with per-coefficient weights
// Perceptual dequantisation with per-coefficient weights
static inline void dequantise_dwt_coefficients_perceptual_avx512(
const int16_t *quantised, float *coeffs, int size,
const float *weights, float base_quantiser
@@ -603,7 +603,7 @@ static inline void dequantise_dwt_coefficients_perceptual_avx512(
}
// =============================================================================
// AVX-512 Optimized RGB to YCoCg Conversion
// AVX-512 Optimised RGB to YCoCg Conversion
// =============================================================================
static inline void rgb_to_ycocg_avx512(const uint8_t *rgb, float *y, float *co, float *cg, int width, int height) {
@@ -658,10 +658,10 @@ static inline void rgb_to_ycocg_avx512(const uint8_t *rgb, float *y, float *co,
}
// =============================================================================
// AVX-512 Optimized 2D DWT with Gather/Scatter
// AVX-512 Optimised 2D DWT with Gather/Scatter
// =============================================================================
// Optimized column extraction using gather
// Optimised column extraction using gather
static inline void dwt_2d_extract_column_avx512(
const float *tile_data, float *column,
int x, int width, int height
@@ -688,7 +688,7 @@ static inline void dwt_2d_extract_column_avx512(
}
}
// Optimized column insertion using scatter
// Optimised column insertion using scatter
static inline void dwt_2d_insert_column_avx512(
float *tile_data, const float *column,
int x, int width, int height