Unroll the multiply-add loops. At the cost of a function pointer dereference we get a whole lot of conditionals eliminated. Should be a solid win once the debug asserts are gone.

ago%!(EXTRA string=11 years) · a95da9ee1d
parent 7abd4ccf34
commit a95da9ee1d
1 changed files with 90 additions and 24 deletions
--- a/stb_resample.h
+++ b/stb_resample.h
@ -490,9 +490,84 @@ static float* stbr__add_empty_ring_buffer_entry(stbr__info* stbr_info, int n)
 	return ring_buffer;
 }

+typedef void(*stbr__output_decode_coefficients)(float* output_buffer, int out_texel_index, float* decode_buffer, int decode_texel_index, int channels, float coefficient);
+
+static void stbr__output_decode_coefficients_1(float* output_buffer, int out_texel_index, float* input_buffer, int input_texel_index, int channels, float coefficient)
+{
+	STBR_DEBUG_ASSERT(channels == 1);
+
+	output_buffer[out_texel_index] += input_buffer[input_texel_index] * coefficient;
+
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index] <= 1.0f);
+}
+
+static void stbr__output_decode_coefficients_2(float* output_buffer, int out_texel_index, float* input_buffer, int input_texel_index, int channels, float coefficient)
+{
+	STBR_DEBUG_ASSERT(channels == 2);
+
+	output_buffer[out_texel_index    ] += input_buffer[input_texel_index    ] * coefficient;
+	output_buffer[out_texel_index + 1] += input_buffer[input_texel_index + 1] * coefficient;
+
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index  ] <= 1.0f);
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index+1] <= 1.0f);
+}
+
+static void stbr__output_decode_coefficients_3(float* output_buffer, int out_texel_index, float* input_buffer, int input_texel_index, int channels, float coefficient)
+{
+	STBR_DEBUG_ASSERT(channels == 3);
+
+	output_buffer[out_texel_index    ] += input_buffer[input_texel_index    ] * coefficient;
+	output_buffer[out_texel_index + 1] += input_buffer[input_texel_index + 1] * coefficient;
+	output_buffer[out_texel_index + 2] += input_buffer[input_texel_index + 2] * coefficient;
+
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index    ] <= 1.0f);
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index + 1] <= 1.0f);
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index + 2] <= 1.0f);
+}
+
+static void stbr__output_decode_coefficients_4(float* output_buffer, int out_texel_index, float* input_buffer, int input_texel_index, int channels, float coefficient)
+{
+	STBR_DEBUG_ASSERT(channels == 4);
+
+	output_buffer[out_texel_index    ] += input_buffer[input_texel_index    ] * coefficient;
+	output_buffer[out_texel_index + 1] += input_buffer[input_texel_index + 1] * coefficient;
+	output_buffer[out_texel_index + 2] += input_buffer[input_texel_index + 2] * coefficient;
+	output_buffer[out_texel_index + 3] += input_buffer[input_texel_index + 3] * coefficient;
+
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index    ] <= 1.0f);
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index + 1] <= 1.0f);
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index + 2] <= 1.0f);
+	STBR_DEBUG_ASSERT(output_buffer[out_texel_index + 3] <= 1.0f);
+}
+
+static void stbr__output_decode_coefficients_n(float* output_buffer, int out_texel_index, float* input_buffer, int input_texel_index, int channels, float coefficient)
+{
+	int c;
+	for (c = 0; c < channels; c++)
+	{
+		output_buffer[out_texel_index + c] += input_buffer[input_texel_index + c] * coefficient;
+
+		STBR_DEBUG_ASSERT(output_buffer[out_texel_index + c] <= 1.0f);
+	}
+}
+
+static stbr__output_decode_coefficients stbr__get_output_decode_coefficients_function(int channels)
+{
+	if (channels == 1)
+		return &stbr__output_decode_coefficients_1;
+	else if (channels == 2)
+		return &stbr__output_decode_coefficients_2;
+	else if (channels == 3)
+		return &stbr__output_decode_coefficients_3;
+	else if (channels == 4)
+		return &stbr__output_decode_coefficients_4;
+
+	return &stbr__output_decode_coefficients_n;
+}
+
 static void stbr__resample_horizontal_upsample(stbr__info* stbr_info, int n, float* output_buffer)
 {
-	int x, k, c;
+	int x, k;
 	int output_w = stbr_info->output_w;
 	int kernel_texel_width = stbr__get_filter_texel_width(stbr_info->filter);
 	int channels = stbr_info->channels;
@ -500,6 +575,8 @@ static void stbr__resample_horizontal_upsample(stbr__info* stbr_info, int n, flo
 	stbr__contributors* horizontal_contributors = stbr_info->horizontal_contributors;
 	float* horizontal_coefficients = stbr_info->horizontal_coefficients;

+	stbr__output_decode_coefficients output_decode_coefficients_fn = stbr__get_output_decode_coefficients_function(channels);
+
 	for (x = 0; x < output_w; x++)
 	{
 		int n0 = horizontal_contributors[x].n0;
@ -521,19 +598,14 @@ static void stbr__resample_horizontal_upsample(stbr__info* stbr_info, int n, flo
 			int in_texel_index = k * channels;
 			float coefficient = horizontal_coefficients[coefficient_index];

-			for (c = 0; c < channels; c++)
-			{
-				output_buffer[out_texel_index + c] += decode_buffer[in_texel_index + c] * coefficient;
-
-				STBR_DEBUG_ASSERT(output_buffer[out_texel_index + c] <= 1.0f);
-			}
+			output_decode_coefficients_fn(output_buffer, out_texel_index, decode_buffer, in_texel_index, channels, coefficient);
 		}
 	}
 }

 static void stbr__resample_horizontal_downsample(stbr__info* stbr_info, int n, float* output_buffer)
 {
-	int x, k, c;
+	int x, k;
 	int input_w = stbr_info->input_w;
 	int output_w = stbr_info->output_w;
 	int kernel_texel_width = stbr__get_filter_texel_width(stbr_info->filter);
@ -544,6 +616,8 @@ static void stbr__resample_horizontal_downsample(stbr__info* stbr_info, int n, f
 	int filter_texel_margin = stbr__get_filter_texel_margin(stbr_info->filter);
 	int max_x = input_w + filter_texel_margin * 2;

+	stbr__output_decode_coefficients output_decode_coefficients_fn = stbr__get_output_decode_coefficients_function(channels);
+
 	STBR_DEBUG_ASSERT(!stbr__use_width_upsampling(stbr_info));

 	for (x = 0; x < max_x; x++)
@ -565,12 +639,7 @@ static void stbr__resample_horizontal_downsample(stbr__info* stbr_info, int n, f
 			int out_texel_index = k * channels;
 			float coefficient = horizontal_coefficients[coefficient_index];

-			for (c = 0; c < channels; c++)
-			{
-				output_buffer[out_texel_index + c] += decode_buffer[in_texel_index + c] * coefficient;
-
-				STBR_DEBUG_ASSERT(output_buffer[out_texel_index + c] <= 1.0); // This would indicate that the sum of kernels for this texel doesn't add to 1.
-			}
+			output_decode_coefficients_fn(output_buffer, out_texel_index, decode_buffer, in_texel_index, channels, coefficient);
 		}
 	}
 }
@ -638,6 +707,8 @@ static void stbr__resample_vertical_upsample(stbr__info* stbr_info, int n, int i

 	int output_row_index = n * stbr_info->output_stride_bytes;

+	stbr__output_decode_coefficients output_decode_coefficients_fn = stbr__get_output_decode_coefficients_function(channels);
+
 	STBR_DEBUG_ASSERT(stbr__use_height_upsampling(stbr_info));
 	STBR_DEBUG_ASSERT(n0 >= in_first_scanline);
 	STBR_DEBUG_ASSERT(n1 <= in_last_scanline);
@ -658,8 +729,7 @@ static void stbr__resample_vertical_upsample(stbr__info* stbr_info, int n, int i
 			float* ring_buffer_entry = stbr__get_ring_buffer_scanline(k, ring_buffer, ring_buffer_begin_index, ring_buffer_first_scanline, kernel_texel_width, ring_buffer_length);
 			float coefficient = vertical_coefficients[coefficient_index];

-			for (c = 0; c < channels; c++)
-				encode_buffer[c] += ring_buffer_entry[in_texel_index + c] * coefficient;
+			output_decode_coefficients_fn(encode_buffer, 0, ring_buffer_entry, in_texel_index, channels, coefficient);
 		}

 		for (c = 0; c < channels; c++)
@ -669,7 +739,7 @@ static void stbr__resample_vertical_upsample(stbr__info* stbr_info, int n, int i

 static void stbr__resample_vertical_downsample(stbr__info* stbr_info, int n, int in_first_scanline, int in_last_scanline, float in_center_of_out)
 {
-	int x, k, c;
+	int x, k;
 	int output_w = stbr_info->output_w;
 	int output_h = stbr_info->output_h;
 	stbr__contributors* vertical_contributors = &stbr_info->vertical_contributors;
@ -690,6 +760,8 @@ static void stbr__resample_vertical_downsample(stbr__info* stbr_info, int n, int
 	int n0 = vertical_contributors->n0;
 	int n1 = vertical_contributors->n1;

+	stbr__output_decode_coefficients output_decode_coefficients_fn = stbr__get_output_decode_coefficients_function(channels);
+
 	STBR_DEBUG_ASSERT(!stbr__use_height_upsampling(stbr_info));
 	STBR_DEBUG_ASSERT(n0 >= in_first_scanline);
 	STBR_DEBUG_ASSERT(n1 <= in_last_scanline);
@ -708,13 +780,7 @@ static void stbr__resample_vertical_downsample(stbr__info* stbr_info, int n, int
 			float* ring_buffer_entry = stbr__get_ring_buffer_scanline(k, ring_buffer, ring_buffer_begin_index, ring_buffer_first_scanline, kernel_texel_width, ring_buffer_length);
 			float coefficient = vertical_coefficients[coefficient_index];

-			for (c = 0; c < channels; c++)
-			{
-				int index = in_texel_index + c;
-				ring_buffer_entry[index] += horizontal_buffer[index] * coefficient;
-
-				STBR_DEBUG_ASSERT(ring_buffer_entry[index] <= 1.0); // This would indicate that the sum of kernels for this texel doesn't add to 1.
-			}
+			output_decode_coefficients_fn(ring_buffer_entry, in_texel_index, horizontal_buffer, in_texel_index, channels, coefficient);
 		}
 	}
 }