Add Wiener convolution test coverage for 5-tap filters
Wiener convolution filters have a maximum size of 7 taps. However,
the vast majority of Wiener filters used in a VoD encode are 5-tap
filters 0-padded to 7 taps. This patch adds test coverage for 5-tap
Wiener filters. Subsequent patches will add specialised Neon paths
for Wiener convolution using 5-tap filters.
Change-Id: Ia8722c6511f91c2b70a586a151378d4ea82ed1a9
diff --git a/test/hiprec_convolve_test_util.cc b/test/hiprec_convolve_test_util.cc
index 7dbe32a..6d7902f 100644
--- a/test/hiprec_convolve_test_util.cc
+++ b/test/hiprec_convolve_test_util.cc
@@ -26,21 +26,21 @@
static void generate_kernels(ACMRandom *rnd, InterpKernel hkernel,
InterpKernel vkernel, int kernel_type = 2) {
if (kernel_type == 0) {
- // Low possible values for filter coefficients
+ // Low possible values for filter coefficients, 7-tap kernel
hkernel[0] = hkernel[6] = vkernel[0] = vkernel[6] = WIENER_FILT_TAP0_MINV;
hkernel[1] = hkernel[5] = vkernel[1] = vkernel[5] = WIENER_FILT_TAP1_MINV;
hkernel[2] = hkernel[4] = vkernel[2] = vkernel[4] = WIENER_FILT_TAP2_MINV;
hkernel[3] = vkernel[3] = -2 * (hkernel[0] + hkernel[1] + hkernel[2]);
hkernel[7] = vkernel[7] = 0;
} else if (kernel_type == 1) {
- // Max possible values for filter coefficients
+ // Max possible values for filter coefficients, 7-tap kernel
hkernel[0] = hkernel[6] = vkernel[0] = vkernel[6] = WIENER_FILT_TAP0_MAXV;
hkernel[1] = hkernel[5] = vkernel[1] = vkernel[5] = WIENER_FILT_TAP1_MAXV;
hkernel[2] = hkernel[4] = vkernel[2] = vkernel[4] = WIENER_FILT_TAP2_MAXV;
hkernel[3] = vkernel[3] = -2 * (hkernel[0] + hkernel[1] + hkernel[2]);
hkernel[7] = vkernel[7] = 0;
- } else {
- // Randomly generated values for filter coefficients
+ } else if (kernel_type == 2) {
+ // Randomly generated values for filter coefficients, 7-tap kernel
hkernel[0] = hkernel[6] =
WIENER_FILT_TAP0_MINV +
rnd->PseudoUniform(WIENER_FILT_TAP0_MAXV + 1 - WIENER_FILT_TAP0_MINV);
@@ -64,6 +64,41 @@
rnd->PseudoUniform(WIENER_FILT_TAP2_MAXV + 2 - WIENER_FILT_TAP2_MINV);
vkernel[3] = -2 * (vkernel[0] + vkernel[1] + vkernel[2]);
vkernel[7] = 0;
+ } else if (kernel_type == 3) {
+ // Low possible values for filter coefficients, 5-tap kernel
+ hkernel[0] = hkernel[6] = vkernel[0] = vkernel[6] = 0;
+ hkernel[1] = hkernel[5] = vkernel[1] = vkernel[5] = WIENER_FILT_TAP1_MINV;
+ hkernel[2] = hkernel[4] = vkernel[2] = vkernel[4] = WIENER_FILT_TAP2_MINV;
+ hkernel[3] = vkernel[3] = -2 * (hkernel[0] + hkernel[1] + hkernel[2]);
+ hkernel[7] = vkernel[7] = 0;
+ } else if (kernel_type == 4) {
+ // Max possible values for filter coefficients, 5-tap kernel
+ hkernel[0] = hkernel[6] = vkernel[0] = vkernel[6] = 0;
+ hkernel[1] = hkernel[5] = vkernel[1] = vkernel[5] = WIENER_FILT_TAP1_MAXV;
+ hkernel[2] = hkernel[4] = vkernel[2] = vkernel[4] = WIENER_FILT_TAP2_MAXV;
+ hkernel[3] = vkernel[3] = -2 * (hkernel[0] + hkernel[1] + hkernel[2]);
+ hkernel[7] = vkernel[7] = 0;
+ } else {
+ // Randomly generated values for filter coefficients, 5-tap kernel
+ hkernel[0] = hkernel[6] = 0;
+ hkernel[1] = hkernel[5] =
+ WIENER_FILT_TAP1_MINV +
+ rnd->PseudoUniform(WIENER_FILT_TAP1_MAXV + 1 - WIENER_FILT_TAP1_MINV);
+ hkernel[2] = hkernel[4] =
+ WIENER_FILT_TAP2_MINV +
+ rnd->PseudoUniform(WIENER_FILT_TAP2_MAXV + 1 - WIENER_FILT_TAP2_MINV);
+ hkernel[3] = -2 * (hkernel[0] + hkernel[1] + hkernel[2]);
+ hkernel[7] = 0;
+
+ vkernel[0] = vkernel[6] = 0;
+ vkernel[1] = vkernel[5] =
+ WIENER_FILT_TAP1_MINV +
+ rnd->PseudoUniform(WIENER_FILT_TAP1_MAXV + 2 - WIENER_FILT_TAP1_MINV);
+ vkernel[2] = vkernel[4] =
+ WIENER_FILT_TAP2_MINV +
+ rnd->PseudoUniform(WIENER_FILT_TAP2_MAXV + 2 - WIENER_FILT_TAP2_MINV);
+ vkernel[3] = -2 * (vkernel[0] + vkernel[1] + vkernel[2]);
+ vkernel[7] = 0;
}
}
@@ -112,7 +147,7 @@
DECLARE_ALIGNED(16, InterpKernel, hkernel);
DECLARE_ALIGNED(16, InterpKernel, vkernel);
- for (int kernel_type = 0; kernel_type < 3; kernel_type++) {
+ for (int kernel_type = 0; kernel_type < 6; kernel_type++) {
generate_kernels(&rnd_, hkernel, vkernel, kernel_type);
for (i = 0; i < num_iters; ++i) {
for (k = 0; k < h; ++k)
@@ -251,7 +286,7 @@
uint8_t *input_ptr = CONVERT_TO_BYTEPTR(input.get());
uint8_t *output_ptr = CONVERT_TO_BYTEPTR(output.get());
uint8_t *output2_ptr = CONVERT_TO_BYTEPTR(output2.get());
- for (int kernel_type = 0; kernel_type < 3; kernel_type++) {
+ for (int kernel_type = 0; kernel_type < 6; kernel_type++) {
generate_kernels(&rnd_, hkernel, vkernel, kernel_type);
for (i = 0; i < num_iters; ++i) {
// Choose random locations within the source block
