Make AVX2 code a bit cleaner.

[bmusb] / bmusb.cpp

diff --git a/bmusb.cpp b/bmusb.cpp
index 8d0918c2215bd6b750ba015b421d50b4c3f38c72..45d1d669dd2fe0771f362aa9eed4b24df8482b6c 100644 (file)
--- a/bmusb.cpp
+++ b/bmusb.cpp
@@ -343,22 +343,29 @@ const uint8_t *add_to_frame_fastpath(FrameAllocator::Frame *current_frame, const
                        swap(out1, out2);
                }
 
-               __m256i mask_lower_byte = _mm256_set1_epi16(0x00ff);
+               __m256i shuffle_cw = _mm256_set_epi8(
+                       15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0,
+                       15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0);
                while (in < (const __m256i *)limit) {
-                       __m256i data1 = _mm256_load_si256(in);
-                       __m256i data2 = _mm256_load_si256(in + 1);
-                       __m256i data1_lo = _mm256_and_si256(data1, mask_lower_byte);
-                       __m256i data2_lo = _mm256_and_si256(data2, mask_lower_byte);
-                       __m256i data1_hi = _mm256_srli_epi16(data1, 8);
-                       __m256i data2_hi = _mm256_srli_epi16(data2, 8);
-                       __m256i lo = _mm256_packus_epi16(data1_lo, data2_lo);
-                       lo = _mm256_permute4x64_epi64(lo, 0b11011000);
-                       _mm256_storeu_si256(out1, lo);  // Store as early as possible, even if the data isn't used.
-                       __m256i hi = _mm256_packus_epi16(data1_hi, data2_hi);
-                       hi = _mm256_permute4x64_epi64(hi, 0b11011000);
-                       _mm256_storeu_si256(out2, hi);
+                       // Note: For brevity, comments show lanes as if they were 2x64-bit (they're actually 2x128).
+                       __m256i data1 = _mm256_load_si256(in);                // AaBbCcDd EeFfGgHh
+                       __m256i data2 = _mm256_load_si256(in + 1);            // IiJjKkLl MmNnOoPp
+
                        __m256i found1 = _mm256_cmpeq_epi8(data1, needle);
                        __m256i found2 = _mm256_cmpeq_epi8(data2, needle);
+
+                       data1 = _mm256_shuffle_epi8(data1, shuffle_cw);       // ABCDabcd EFGHefgh
+                       data2 = _mm256_shuffle_epi8(data2, shuffle_cw);       // IJKLijkl MNOPmnop
+               
+                       data1 = _mm256_permute4x64_epi64(data1, 0b11011000);  // ABCDEFGH abcdefgh
+                       data2 = _mm256_permute4x64_epi64(data2, 0b11011000);  // IJKLMNOP ijklmnop
+
+                       __m256i lo = _mm256_permute2x128_si256(data1, data2, 0b00100000);
+                       __m256i hi = _mm256_permute2x128_si256(data1, data2, 0b00110001);
+
+                       _mm256_storeu_si256(out1, lo);  // Store as early as possible, even if the data isn't used.
+                       _mm256_storeu_si256(out2, hi);
+
                        if (!_mm256_testz_si256(found1, found1) ||
                            !_mm256_testz_si256(found2, found2)) {
                                break;
@@ -604,8 +611,8 @@ void BMUSBCapture::start_bm_capture()
                printf("  interface %d\n", interface_number);
                const libusb_interface *interface = &config->interface[interface_number];
                for (int altsetting = 0; altsetting < interface->num_altsetting; ++altsetting) {
-                       printf("    alternate setting %d\n", altsetting);
                        const libusb_interface_descriptor *interface_desc = &interface->altsetting[altsetting];
+                       printf("    alternate setting %d\n", interface_desc->bAlternateSetting);
                        for (int endpoint_number = 0; endpoint_number < interface_desc->bNumEndpoints; ++endpoint_number) {
                                const libusb_endpoint_descriptor *endpoint = &interface_desc->endpoint[endpoint_number];
                                printf("        endpoint address 0x%02x\n", endpoint->bEndpointAddress);
@@ -675,6 +682,8 @@ void BMUSBCapture::start_bm_capture()
        //
        //    so only first 16 bits count, and 0x0100 is a mask for ok/stable signal?
        //
+       //    Bottom 16 bits of this register seem to be firmware version number (possibly not all all of them).
+       //
        //    28 and 32 seems to be analog audio input levels (one byte for each of the eight channels).
        //    however, if setting 32 with HDMI embedded audio, it is immediately overwritten back (to 0xe137002a).
        //
@@ -684,6 +693,11 @@ void BMUSBCapture::start_bm_capture()
        //    36 can be set to 0 with no apparent effect (all of this tested on both video and audio),
        //    but the driver sets it to 0x8036802a at some point.
        //
+       //    all of this is on request 214/215. other requests (192, 219,
+       //    222, 223, 224) are used for firmware upgrade. Probably best to
+       //    stay out of it unless you know what you're doing.
+       //
+       //
        // register 16:
        // first byte is 0x39 for a stable 576p60 signal, 0x2d for a stable 720p60 signal, 0x20 for no signal
        //