Optimize VAO/VBO usage for minimal state changes.

[movit] / resample_effect.cpp

diff --git a/resample_effect.cpp b/resample_effect.cpp
index 244a3e2a6081187f19962b531c7fa86dde46203b..9c8caf3bb77ed58bf2d143dcc2b3b837b81a001c 100644 (file)
--- a/resample_effect.cpp
+++ b/resample_effect.cpp
@@ -1,4 +1,6 @@
 // Three-lobed Lanczos, the most common choice.
 // Three-lobed Lanczos, the most common choice.

+// Note that if you change this, the accuracy for LANCZOS_TABLE_SIZE
+// needs to be recomputed.

 #define LANCZOS_RADIUS 3.0
 
 #include <epoxy/gl.h>
 #define LANCZOS_RADIUS 3.0
 
 #include <epoxy/gl.h>


@@ -40,15 +42,67 @@ float sinc(float x)
        }
 }
 
        }
 }
 

-float lanczos_weight(float x, float a)
+float lanczos_weight(float x)

 {
 {

-       if (fabs(x) > a) {
+       if (fabs(x) > LANCZOS_RADIUS) {

                return 0.0f;
        } else {
                return 0.0f;
        } else {

-               return sinc(M_PI * x) * sinc(M_PI * x / a);
+               return sinc(M_PI * x) * sinc((M_PI / LANCZOS_RADIUS) * x);

        }
 }
 
        }
 }
 

+// The weight function can be expensive to compute over and over again
+// (which will happen during e.g. a zoom), but it is also easy to interpolate
+// linearly. We compute the right half of the function (in the range of
+// 0..LANCZOS_RADIUS), with two guard elements for easier interpolation, and
+// linearly interpolate to get our function.
+//
+// We want to scale the table so that the maximum error is always smaller
+// than 1e-6. As per http://www-solar.mcs.st-andrews.ac.uk/~clare/Lectures/num-analysis/Numan_chap3.pdf,
+// the error for interpolating a function linearly between points [a,b] is
+//
+//   e = 1/2 (x-a)(x-b) f''(u_x)
+//
+// for some point u_x in [a,b] (where f(x) is our Lanczos function; we're
+// assuming LANCZOS_RADIUS=3 from here on). Obviously this is bounded by
+// f''(x) over the entire range. Numeric optimization shows the maximum of
+// |f''(x)| to be in x=1.09369819474562880, with the value 2.40067758733152381.
+// So if the steps between consecutive values are called d, we get
+//
+//   |e| <= 1/2 (d/2)^2 2.4007
+//   |e| <= 0.1367 d^2
+//
+// Solve for e = 1e-6 yields a step size of 0.0027, which to cover the range
+// 0..3 needs 1109 steps. We round up to the next power of two, just to be sure.
+//
+// You need to call lanczos_table_init_done before the first call to
+// lanczos_weight_cached.
+#define LANCZOS_TABLE_SIZE 2048
+bool lanczos_table_init_done = false;
+float lanczos_table[LANCZOS_TABLE_SIZE + 2];
+
+void init_lanczos_table()
+{
+       for (unsigned i = 0; i < LANCZOS_TABLE_SIZE + 2; ++i) {
+               lanczos_table[i] = lanczos_weight(float(i) * (LANCZOS_RADIUS / LANCZOS_TABLE_SIZE));
+       }
+       lanczos_table_init_done = true;
+}
+
+float lanczos_weight_cached(float x)
+{
+       x = fabs(x);
+       if (x > LANCZOS_RADIUS) {
+               return 0.0f;
+       }
+       float table_pos = x * (LANCZOS_TABLE_SIZE / LANCZOS_RADIUS);
+       int table_pos_int = int(table_pos);  // Truncate towards zero.
+       float table_pos_frac = table_pos - table_pos_int;
+       assert(table_pos < LANCZOS_TABLE_SIZE + 2);
+       return lanczos_table[table_pos_int] +
+               table_pos_frac * (lanczos_table[table_pos_int + 1] - lanczos_table[table_pos_int]);
+}
+

 // Euclid's algorithm, from Wikipedia.
 unsigned gcd(unsigned a, unsigned b)
 {
 // Euclid's algorithm, from Wikipedia.
 unsigned gcd(unsigned a, unsigned b)
 {


@@ -138,13 +192,13 @@ template<class T>
 void normalize_sum(Tap<T>* vals, unsigned num)
 {
        for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) {
 void normalize_sum(Tap<T>* vals, unsigned num)
 {
        for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) {

-               double sum = 0.0;
+               float sum = 0.0;

                for (unsigned i = 0; i < num; ++i) {
                for (unsigned i = 0; i < num; ++i) {

-                       sum += to_fp64(vals[i].weight);
+                       sum += to_fp32(vals[i].weight);

                }
                }

-               double inv_sum = 1.0 / sum;
+               float inv_sum = 1.0 / sum;

                for (unsigned i = 0; i < num; ++i) {
                for (unsigned i = 0; i < num; ++i) {

-                       vals[i].weight = from_fp64<T>(to_fp64(vals[i].weight) * inv_sum);
+                       vals[i].weight = from_fp32<T>(to_fp32(vals[i].weight) * inv_sum);

                }
        }
 }
                }
        }
 }


@@ -201,8 +255,8 @@ double compute_sum_sq_error(const Tap<float>* weights, unsigned num_weights,
        // Find the effective range of the bilinear-optimized kernel.
        // Due to rounding of the positions, this is not necessarily the same
        // as the intended range (ie., the range of the original weights).
        // Find the effective range of the bilinear-optimized kernel.
        // Due to rounding of the positions, this is not necessarily the same
        // as the intended range (ie., the range of the original weights).

-       int lower_pos = int(floor(to_fp64(bilinear_weights[0].pos) * size - 0.5));
-       int upper_pos = int(ceil(to_fp64(bilinear_weights[num_bilinear_weights - 1].pos) * size - 0.5)) + 2;
+       int lower_pos = int(floor(to_fp32(bilinear_weights[0].pos) * size - 0.5));
+       int upper_pos = int(ceil(to_fp32(bilinear_weights[num_bilinear_weights - 1].pos) * size - 0.5)) + 2;

        lower_pos = min<int>(lower_pos, lrintf(weights[0].pos * size - 0.5));
        upper_pos = max<int>(upper_pos, lrintf(weights[num_weights - 1].pos * size - 0.5) + 1);
 
        lower_pos = min<int>(lower_pos, lrintf(weights[0].pos * size - 0.5));
        upper_pos = max<int>(upper_pos, lrintf(weights[num_weights - 1].pos * size - 0.5) + 1);
 


@@ -213,7 +267,7 @@ double compute_sum_sq_error(const Tap<float>* weights, unsigned num_weights,
 
        // Now find the effective weights that result from this sampling.
        for (unsigned i = 0; i < num_bilinear_weights; ++i) {
 
        // Now find the effective weights that result from this sampling.
        for (unsigned i = 0; i < num_bilinear_weights; ++i) {

-               const float pixel_pos = to_fp64(bilinear_weights[i].pos) * size - 0.5f;
+               const float pixel_pos = to_fp32(bilinear_weights[i].pos) * size - 0.5f;

                const int x0 = int(floor(pixel_pos)) - lower_pos;
                const int x1 = x0 + 1;
                const float f = lrintf((pixel_pos - (x0 + lower_pos)) / movit_texel_subpixel_precision) * movit_texel_subpixel_precision;
                const int x0 = int(floor(pixel_pos)) - lower_pos;
                const int x1 = x0 + 1;
                const float f = lrintf((pixel_pos - (x0 + lower_pos)) / movit_texel_subpixel_precision) * movit_texel_subpixel_precision;


@@ -223,8 +277,8 @@ double compute_sum_sq_error(const Tap<float>* weights, unsigned num_weights,
                assert(x0 < upper_pos - lower_pos);
                assert(x1 < upper_pos - lower_pos);
 
                assert(x0 < upper_pos - lower_pos);
                assert(x1 < upper_pos - lower_pos);
 

-               effective_weights[x0] += to_fp64(bilinear_weights[i].weight) * (1.0 - f);
-               effective_weights[x1] += to_fp64(bilinear_weights[i].weight) * f;
+               effective_weights[x0] += to_fp32(bilinear_weights[i].weight) * (1.0 - f);
+               effective_weights[x1] += to_fp32(bilinear_weights[i].weight) * f;

        }
 
        // Subtract the desired weights to get the error.
        }
 
        // Subtract the desired weights to get the error.


@@ -397,7 +451,7 @@ SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent)
        register_float("offset", &offset);
        register_float("zoom", &zoom);
        register_uniform_sampler2d("sample_tex", &uniform_sample_tex);
        register_float("offset", &offset);
        register_float("zoom", &zoom);
        register_uniform_sampler2d("sample_tex", &uniform_sample_tex);

-       register_uniform_int("num_samples", &uniform_num_samples);  // FIXME: What about GLSL pre-1.30?
+       register_uniform_int("num_samples", &uniform_num_samples);

        register_uniform_float("num_loops", &uniform_num_loops);
        register_uniform_float("slice_height", &uniform_slice_height);
        register_uniform_float("sample_x_scale", &uniform_sample_x_scale);
        register_uniform_float("num_loops", &uniform_num_loops);
        register_uniform_float("slice_height", &uniform_slice_height);
        register_uniform_float("sample_x_scale", &uniform_sample_x_scale);


@@ -405,6 +459,12 @@ SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent)
        register_uniform_float("whole_pixel_offset", &uniform_whole_pixel_offset);
 
        glGenTextures(1, &texnum);
        register_uniform_float("whole_pixel_offset", &uniform_whole_pixel_offset);
 
        glGenTextures(1, &texnum);

+
+       if (!lanczos_table_init_done) {
+               // Could in theory race between two threads if we are unlucky,
+               // but that is harmless, since they'll write the same data.
+               init_lanczos_table();
+       }

 }
 
 SingleResamplePassEffect::~SingleResamplePassEffect()
 }
 
 SingleResamplePassEffect::~SingleResamplePassEffect()


@@ -531,7 +591,7 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str
                // Now sample <int_radius> pixels on each side around that point.
                for (int i = 0; i < src_samples; ++i) {
                        int src_y = base_src_y + i - int_radius;
                // Now sample <int_radius> pixels on each side around that point.
                for (int i = 0; i < src_samples; ++i) {
                        int src_y = base_src_y + i - int_radius;

-                       float weight = lanczos_weight(radius_scaling_factor * (src_y - center_src_y - subpixel_offset), LANCZOS_RADIUS);
+                       float weight = lanczos_weight_cached(radius_scaling_factor * (src_y - center_src_y - subpixel_offset));

                        weights[y * src_samples + i].weight = weight * radius_scaling_factor;
                        weights[y * src_samples + i].pos = (src_y + 0.5) / float(src_size);
                }
                        weights[y * src_samples + i].weight = weight * radius_scaling_factor;
                        weights[y * src_samples + i].pos = (src_y + 0.5) / float(src_size);
                }


@@ -657,10 +717,12 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const strin
        // We specifically do not want mipmaps on the input texture;
        // they break minification.
        Node *self = chain->find_node_for_effect(this);
        // We specifically do not want mipmaps on the input texture;
        // they break minification.
        Node *self = chain->find_node_for_effect(this);

-       glActiveTexture(chain->get_input_sampler(self, 0));
-       check_error();
-       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-       check_error();
+       if (chain->has_input_sampler(self, 0)) {
+               glActiveTexture(chain->get_input_sampler(self, 0));
+               check_error();
+               glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+               check_error();
+       }

 }
 
 }  // namespace movit
 }
 
 }  // namespace movit