Merge branch 'runtime-image-jpeg-scale' into runtime-image-integer-draw

An image drawn at its own size needs no span maths at all.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

esphome/components/runtime_image/image_decoder.cpp

@@ -2,7 +2,6 @@
 #include "runtime_image.h"
 #include "esphome/core/log.h"
 #include <algorithm>
-#include <cmath>
 
 namespace esphome::runtime_image {
 
@@ -11,19 +10,48 @@ ESPHOME_LOG_TAG(TAG, "image_decoder");
 bool ImageDecoder::set_size(int width, int height) {
   bool success = this->image_->resize(width, height) > 0;
   this->size_valid_ = success;
-  this->x_scale_ = static_cast<double>(this->image_->get_buffer_width()) / width;
-  this->y_scale_ = static_cast<double>(this->image_->get_buffer_height()) / height;
+  this->draw_width_ = width;
+  this->draw_height_ = height;
   return success;
 }
 
+int ImageDecoder::scaled_x_(int x) const {
+  if (this->draw_width_ <= 0)
+    return 0;
+  const uint32_t buffer_width = this->image_->get_buffer_width();
+  return static_cast<int>((static_cast<uint32_t>(x) * buffer_width + this->draw_width_ - 1) / this->draw_width_);
+}
+
+int ImageDecoder::scaled_y_(int y) const {
+  if (this->draw_height_ <= 0)
+    return 0;
+  const uint32_t buffer_height = this->image_->get_buffer_height();
+  return static_cast<int>((static_cast<uint32_t>(y) * buffer_height + this->draw_height_ - 1) / this->draw_height_);
+}
+
 void ImageDecoder::draw(int x, int y, int w, int h, const Color &color) {
   if (!this->size_valid_) {
     return;
   }
-  auto width = std::min(this->image_->get_buffer_width(), static_cast<int>(std::ceil((x + w) * this->x_scale_)));
-  auto height = std::min(this->image_->get_buffer_height(), static_cast<int>(std::ceil((y + h) * this->y_scale_)));
-  for (int i = x * this->x_scale_; i < width; i++) {
-    for (int j = y * this->y_scale_; j < height; j++) {
+  // A source pixel paints the target pixels from ceil(start * scale) up to ceil(end * scale). The spans of neighbouring
+  // pixels meet without overlapping, so when an image is shrunk each target pixel is painted once, by the first
+  // source pixel that reaches it, rather than by every pixel that lands on it. Integer maths: the S3 has no double FPU.
+  const int buffer_width = this->image_->get_buffer_width();
+  const int buffer_height = this->image_->get_buffer_height();
+  int x1 = x;
+  int y1 = y;
+  int x2 = x + w;
+  int y2 = y + h;
+  if (buffer_width != this->draw_width_ || buffer_height != this->draw_height_) {
+    x1 = this->scaled_x_(x);
+    y1 = this->scaled_y_(y);
+    x2 = this->scaled_x_(x + w);
+    y2 = this->scaled_y_(y + h);
+  }
+  x2 = std::min(buffer_width, x2);
+  y2 = std::min(buffer_height, y2);
+  for (int i = x1; i < x2; i++) {
+    for (int j = y1; j < y2; j++) {
       this->image_->draw_pixel(i, j, color);
     }
   }

esphome/components/runtime_image/image_decoder.h

@@ -83,8 +83,8 @@ class ImageDecoder {
     this->expected_size_ = 0;
     this->decoded_bytes_ = 0;
     this->size_valid_ = true;
-    this->x_scale_ = 1.0;
-    this->y_scale_ = 1.0;
+    this->draw_width_ = 0;
+    this->draw_height_ = 0;
   }
 
   /**
@@ -184,8 +184,13 @@ class ImageDecoder {
   uint8_t *image_buffer_() const;
   bool image_is_big_endian_() const;
 
-  double x_scale_ = 1.0;
-  double y_scale_ = 1.0;
+  /// The first buffer column, or row, that source column x, or row y, paints: ceil(x * buffer size / draw size)
+  int scaled_x_(int x) const;
+  int scaled_y_(int y) const;
+
+  // Size of the image as the decoder draws it, which draw() spreads over the buffer
+  int draw_width_ = 0;
+  int draw_height_ = 0;
   RuntimeImage *image_;
   size_t expected_size_ = 0;  // Expected data size (0 if unknown)
   size_t decoded_bytes_ = 0;  // Bytes processed so far

esphome/components/runtime_image/jpeg_decoder.cpp

@@ -31,8 +31,8 @@ bool JpegDecoder::prepare_rgb565_scaling_(int width, int height) {
   const uint32_t target_width = this->image_->get_buffer_width();
   const uint32_t target_height = this->image_->get_buffer_height();
   // Boundary n is the first target pixel source pixel n covers, rounded up. When shrinking, the source
-  // pixel whose span lands on a target pixel paints it, so an 8 to 4 scale keeps columns 0, 2, 4 and 6;
-  // the generic draw() path rounds down and keeps the odd ones instead
+  // pixel whose span lands on a target pixel paints it, so an 8 to 4 scale keeps columns 0, 2, 4 and 6,
+  // as the generic draw() path does
   for (uint32_t x = 0; x <= static_cast<uint32_t>(width); x++) {
     this->x_boundaries_.push_back(static_cast<uint16_t>((x * target_width + width - 1) / width));
   }
@@ -138,7 +138,7 @@ int HOT JpegDecoder::decode(uint8_t *buffer, size_t size) {
   const int err = this->start_();
   if (err != 0)
     return err;
-  if (!this->jpeg_.decode(0, 0, 0))
+  if (!this->jpeg_.decode(0, 0, this->scale_option_()))
     return this->fail_();
   this->decoded_bytes_ = size;
   this->jpeg_.close();
@@ -165,10 +165,28 @@ int JpegDecoder::start_() {
   } else {
     this->jpeg_.setPixelType(RGB8888);
   }
-  if (!this->set_size(this->jpeg_.getWidth(), this->jpeg_.getHeight())) {
+  const int width = this->jpeg_.getWidth();
+  const int height = this->jpeg_.getHeight();
+  if (!this->set_size(width, height)) {
     return DECODE_ERROR_OUT_OF_MEMORY;
   }
-  if (this->direct_rgb565_ && !this->prepare_rgb565_scaling_(this->jpeg_.getWidth(), this->jpeg_.getHeight())) {
+  // JPEGDEC can decode at 1/2, 1/4 or 1/8 size for much less work. Use the smallest size that still has at least as
+  // many pixels as the buffer, so a large image shown small is not decoded in full only to be thrown away.
+  const int buffer_width = this->image_->get_buffer_width();
+  const int buffer_height = this->image_->get_buffer_height();
+  this->scale_ = 1;
+  while (this->scale_ < JPEG_SCALE_EIGHTH && width / (this->scale_ * 2) >= buffer_width &&
+         height / (this->scale_ * 2) >= buffer_height) {
+    this->scale_ *= 2;
+  }
+  const int decoded_width = (width + this->scale_ - 1) / this->scale_;
+  const int decoded_height = (height + this->scale_ - 1) / this->scale_;
+  this->draw_width_ = decoded_width;
+  this->draw_height_ = decoded_height;
+  if (this->scale_ != 1) {
+    ESP_LOGD(TAG, "Decoding at 1/%d size, %d x %d", this->scale_, decoded_width, decoded_height);
+  }
+  if (this->direct_rgb565_ && !this->prepare_rgb565_scaling_(decoded_width, decoded_height)) {
     return DECODE_ERROR_OUT_OF_MEMORY;
   }
   return 0;
@@ -252,7 +270,7 @@ int JpegDecoder::begin_source(DecodeSource *source) {
     this->source_ = nullptr;
     return err;
   }
-  if (!this->jpeg_.decodeBegin(0, 0, 0))
+  if (!this->jpeg_.decodeBegin(0, 0, this->scale_option_()))
     return this->fail_();
   return 0;
 }

esphome/components/runtime_image/jpeg_decoder.h

@@ -31,12 +31,16 @@ class JpegDecoder : public ImageDecoder {
   int HOT decode_step() override;
 
   bool uses_direct_rgb565() const { return this->direct_rgb565_; }
+  /// JPEGDEC decodes the image at 1/get_scale() of its size: 1, 2, 4 or 8
+  int get_scale() const { return this->scale_; }
   void draw_rgb565(int x, int y, int w, int h, const uint8_t *pixels);
 
  protected:
   bool prepare_rgb565_scaling_(int width, int height);
   /// Set up an opened image for decoding. Returns 0, or a DecodeError.
   int start_();
+  /// The JPEGDEC option for scale_: JPEG_SCALE_HALF, _QUARTER and _EIGHTH are the divisors themselves
+  int scale_option_() const { return this->scale_ == 1 ? 0 : this->scale_; }
   /// The DecodeError for a failed read of the source or JPEGDEC's last error, closing the image
   int fail_();
 
@@ -49,6 +53,7 @@ class JpegDecoder : public ImageDecoder {
   FixedVector<uint16_t> y_boundaries_{};
   int source_width_{0};
   int source_height_{0};
+  int scale_{1};                   // JPEGDEC decodes at 1/scale_ of the image size
   DecodeSource *source_{nullptr};  // Set while a begin_source() decode is in progress
   bool direct_rgb565_{false};
   bool source_failed_{false};

tests/components/runtime_image/test_decoder_reuse.cpp

@@ -4,9 +4,11 @@
 #include <algorithm>
 #include <array>
 #include <cstring>
+#include <utility>
 #include <vector>
 
 #include "esphome/components/runtime_image/image_decoder.h"
+#include "esphome/components/runtime_image/jpeg_decoder.h"
 #include "esphome/components/runtime_image/runtime_image.h"
 
 namespace esphome::runtime_image::testing {
@@ -362,6 +364,23 @@ static std::vector<uint8_t> pixel_bytes(TestableRuntimeImage &img) {
   return std::vector<uint8_t>(start, start + img.get_width_stride() * img.get_height());
 }
 
+// Shrinking a PNG keeps the first source pixel of each span, as the JPEG path does: the even rows and columns
+TEST(RuntimeImageDecoder, ShrunkPngKeepsTheFirstPixelOfEachSpan) {
+  TestableRuntimeImage img(PNG, image::IMAGE_TYPE_RGB, false, 2, 2);
+  ASSERT_TRUE(decode_all(img, PNG_RGB, sizeof(PNG_RGB), PNG));
+  ASSERT_EQ(img.get_width(), 2);
+  ASSERT_EQ(img.get_height(), 2);
+  for (int row = 0; row < 2; row++) {
+    for (int col = 0; col < 2; col++) {
+      const Color pixel = img.get_pixel(col, row);
+      const uint8_t *expected = PNG_RGB_EXPECTED[row * 2][col * 2];
+      EXPECT_EQ(pixel.r, expected[0]) << col << "," << row;
+      EXPECT_EQ(pixel.g, expected[1]) << col << "," << row;
+      EXPECT_EQ(pixel.b, expected[2]) << col << "," << row;
+    }
+  }
+}
+
 TEST(RuntimeImageDecoder, JpegDecoderStaysWarmAcrossDecodes) {
   TestableRuntimeImage img(JPEG);
 
@@ -377,6 +396,12 @@ TEST(RuntimeImageDecoder, JpegDecoderStaysWarmAcrossDecodes) {
   EXPECT_EQ(pixel_bytes(img), first_pixels) << "reused decoder must reproduce identical pixels";
 }
 
+// The RGB565 word at (x, y) of a decoded image, in its own byte order
+static uint16_t rgb565_word(const TestableRuntimeImage &img, int x, int y) {
+  const uint8_t *p = img.get_data_start() + (y * img.get_width() + x) * 2;
+  return p[0] | (p[1] << 8);
+}
+
 // JPEGDEC's colour conversion differs by one step between its SIMD and scalar builds, so each
 // channel is checked within one step; the byte order is still exact
 static void expect_rgb565_near(uint16_t actual, uint16_t expected, const char *order, size_t pixel) {
@@ -386,16 +411,12 @@ static void expect_rgb565_near(uint16_t actual, uint16_t expected, const char *o
 }
 
 TEST(RuntimeImageDecoder, JpegDecodesDirectlyToScaledRgb565WithBothByteOrders) {
-  // Halving the 8x8 gradient keeps its even rows and columns. The words are JPEGDEC's output for those
-  // pixels, red in the top five bits: a decode of the fixture with PIL gives (51, 7, 30) for source
-  // pixel (2, 0), which is word 0x3803 here.
-  static const uint16_t EXPECTED[4 * 4] = {
-      0x0000, 0x3803, 0x8008, 0xB80B, 0x0203, 0x3A07, 0x81EC, 0xBA0F,
-      0x03E7, 0x3C0B, 0x83F0, 0xBC13, 0x05EC, 0x3E0F, 0x85F4, 0xBE17,
-  };
+  // Halving the 8x8 gradient makes JPEGDEC decode it at half size, which averages each 2x2 block of pixels
+  TestableRuntimeImage full(JPEG, image::IMAGE_TYPE_RGB565, false, 8, 8);
   TestableRuntimeImage rgb565_little_endian(JPEG, image::IMAGE_TYPE_RGB565, false, 4, 4);
   TestableRuntimeImage rgb565_big_endian(JPEG, image::IMAGE_TYPE_RGB565, true, 4, 4);
 
+  ASSERT_TRUE(decode_all(full, JPEG_GRADIENT, sizeof(JPEG_GRADIENT)));
   ASSERT_TRUE(decode_all(rgb565_little_endian, JPEG_GRADIENT, sizeof(JPEG_GRADIENT)));
   ASSERT_TRUE(decode_all(rgb565_big_endian, JPEG_GRADIENT, sizeof(JPEG_GRADIENT)));
   ASSERT_EQ(rgb565_little_endian.get_width(), 4);
@@ -405,9 +426,24 @@ TEST(RuntimeImageDecoder, JpegDecodesDirectlyToScaledRgb565WithBothByteOrders) {
 
   const uint8_t *little = rgb565_little_endian.get_data_start();
   const uint8_t *big = rgb565_big_endian.get_data_start();
-  for (size_t i = 0; i < 4 * 4; i++) {
-    expect_rgb565_near(little[i * 2] | (little[i * 2 + 1] << 8), EXPECTED[i], "little-endian", i);
-    expect_rgb565_near((big[i * 2] << 8) | big[i * 2 + 1], EXPECTED[i], "big-endian", i);
+  for (int row = 0; row < 4; row++) {
+    for (int col = 0; col < 4; col++) {
+      int red = 0;
+      int green = 0;
+      int blue = 0;
+      for (int dy = 0; dy < 2; dy++) {
+        for (int dx = 0; dx < 2; dx++) {
+          const uint16_t word = rgb565_word(full, col * 2 + dx, row * 2 + dy);
+          red += word >> 11;
+          green += (word >> 5) & 0x3F;
+          blue += word & 0x1F;
+        }
+      }
+      const uint16_t expected = ((red + 2) / 4) << 11 | ((green + 2) / 4) << 5 | (blue + 2) / 4;
+      const size_t i = row * 4 + col;
+      expect_rgb565_near(little[i * 2] | (little[i * 2 + 1] << 8), expected, "little-endian", i);
+      expect_rgb565_near((big[i * 2] << 8) | big[i * 2 + 1], expected, "big-endian", i);
+    }
   }
 }
 
@@ -415,19 +451,25 @@ TEST(RuntimeImageDecoder, JpegDecodesDirectlyToScaledRgb565WithBothByteOrders) {
 // prepare_rgb565_scaling_ rounds: source n covers targets ceil(n * target / source) up to the next start
 static uint16_t scaled_boundary(int n, int source, int target) { return (n * target + source - 1) / source; }
 
-// The RGB565 word at (x, y) of a decoded image, in its own byte order
-static uint16_t rgb565_word(const TestableRuntimeImage &img, int x, int y) {
-  const uint8_t *p = img.get_data_start() + (y * img.get_width() + x) * 2;
-  return p[0] | (p[1] << 8);
-}
-
 // Every scaled direct decode must equal the unscaled one sampled through the boundary table,
 // for shrinking, growing and a ratio that is not a whole number
+// JPEGDEC decodes at the smallest of full, 1/2, 1/4 and 1/8 size that still covers the buffer
+TEST(RuntimeImageDecoder, JpegDecodesAtTheSmallestSizeThatCoversTheBuffer) {
+  for (auto [size, scale] :
+       {std::pair{8, 1}, std::pair{5, 1}, std::pair{4, 2}, std::pair{3, 2}, std::pair{2, 4}, std::pair{1, 8}}) {
+    TestableRuntimeImage img(JPEG, image::IMAGE_TYPE_RGB565, false, size, size);
+    ASSERT_TRUE(decode_all(img, JPEG_GRADIENT, sizeof(JPEG_GRADIENT)));
+    EXPECT_EQ(img.get_width(), size);
+    EXPECT_EQ(static_cast<JpegDecoder *>(img.decoder())->get_scale(), scale) << size << "x" << size;
+  }
+}
+
 TEST(RuntimeImageDecoder, JpegDirectRgb565ScalesBySourcePixelSpans) {
   TestableRuntimeImage full(JPEG, image::IMAGE_TYPE_RGB565, false, 8, 8);
   ASSERT_TRUE(decode_all(full, JPEG_GRADIENT, sizeof(JPEG_GRADIENT)));
 
-  for (int size : {4, 6, 16}) {
+  // Sizes JPEGDEC cannot reach by scaling, so the full size decode is spread over the buffer
+  for (int size : {5, 6, 16}) {
     TestableRuntimeImage scaled(JPEG, image::IMAGE_TYPE_RGB565, false, size, size);
     ASSERT_TRUE(decode_all(scaled, JPEG_GRADIENT, sizeof(JPEG_GRADIENT)));
     ASSERT_EQ(scaled.get_width(), size);