From 89d37a04aa9ba618bb1c6d3024d2262feb1ee09e Mon Sep 17 00:00:00 2001
From: Josh Bailey <josh@vandervecken.com>
Date: Thu, 23 Nov 2023 03:33:10 +0000
Subject: [PATCH] not enough copies.

---
 lib/libvkfft.cc         | 4 ++--
 lib/vkfft_impl.cc       | 4 ++--
 lib/vkfft_short_impl.cc | 4 ++--
 3 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/lib/libvkfft.cc b/lib/libvkfft.cc
index 9a5df01b..eb925a68 100644
--- a/lib/libvkfft.cc
+++ b/lib/libvkfft.cc
@@ -117,10 +117,10 @@ VkFFTResult _transferDataToCPU(char *cpu_arr) {
     return VKFFT_ERROR_MALLOC_FAILED;
   if (_shift) {
     const size_t halfFftBufferSize = fftBufferSize / 2;
-    for (int i = 0; i < vkConfiguration.numberBatches; ++i) {
+    for (int i = 0; i < vkConfiguration.numberBatches;
+         ++i, cpu_arr += fftBufferSize, data += fftBufferSize) {
       memcpy(cpu_arr + halfFftBufferSize, data, halfFftBufferSize);
       memcpy(cpu_arr, data + halfFftBufferSize, halfFftBufferSize);
-      cpu_arr += fftBufferSize;
     }
   } else {
     memcpy(cpu_arr, data, stagingBufferSize);
diff --git a/lib/vkfft_impl.cc b/lib/vkfft_impl.cc
index 1b38172c..adba0464 100644
--- a/lib/vkfft_impl.cc
+++ b/lib/vkfft_impl.cc
@@ -230,9 +230,9 @@ int vkfft_impl::work(int noutput_items, gr_vector_const_void_star &input_items,
   const gr_complex *const in =
       reinterpret_cast<const gr_complex *const>(input_items[0]);
   gr_complex *const out = reinterpret_cast<gr_complex *const>(output_items[0]);
+  size_t buffer_index = 0;
 
-  for (int i = 0; i < noutput_items; ++i) {
-    const int buffer_index = i * vlen_;
+  for (int i = 0; i < noutput_items; ++i, buffer_index += vlen_) {
     vkfft_offload((char *)&in[buffer_index], (char *)&out[buffer_index]);
   }
 
diff --git a/lib/vkfft_short_impl.cc b/lib/vkfft_short_impl.cc
index fbaad7ba..ca8a85d2 100644
--- a/lib/vkfft_short_impl.cc
+++ b/lib/vkfft_short_impl.cc
@@ -246,9 +246,9 @@ int vkfft_short_impl::work(int noutput_items,
       reinterpret_cast<const std::int16_t *const>(input_items[0]);
   gr_complex *const out = reinterpret_cast<gr_complex *const>(output_items[0]);
   auto *buffer = input_buffer_.get();
+  size_t buffer_index = 0;
 
-  for (int i = 0; i < noutput_items; ++i) {
-    const int buffer_index = i * vlen_ * 2;
+  for (int i = 0; i < noutput_items; ++i, buffer_index += vlen_ * 2) {
     _converter->conv(&in[buffer_index], &buffer[0], vlen_);
     vkfft_offload((char *)&buffer[0], (char *)&out[buffer_index]);
   }