kernel: Bump to 3.18.9

kernel: Align pcm512x driver with upstream
See: raspberrypi/linux#873

kernel: dts: overlay: add mz61581, fix piscreen and rpi-display
See: raspberrypi/linux#874

kernel: bcm2709: Increase the spare/free IRQs to match bcm2708
See: raspberrypi/linux#871

kernel: add support for Adafruit PiTFT
See: raspberrypi/linux#858

firmware: dispserve: Allow vsync requests from multiple clients
See: raspberrypi/userland#218

firmware: hello_fft: Update to version 3
See: http://www.aholme.co.uk/GPU_FFT/Main.htm

extra/git_hash

@@ -1 +1 @@
-91a03559bbc1b5cb2b1157150c1c698b22716d9a
+780e68130fba82a525b89e85f051c91b7a508e52

extra/uname_string

@@ -1 +1 @@
-Linux version 3.18.8+ (dc4@dc4-XPS13-9333) (gcc version 4.8.3 20140303 (prerelease) (crosstool-NG linaro-1.13.1+bzr2650 - Linaro GCC 2014.03) ) #765 PREEMPT Thu Mar 5 15:41:59 GMT 2015
+Linux version 3.18.9+ (dc4@dc4-XPS13-9333) (gcc version 4.8.3 20140303 (prerelease) (crosstool-NG linaro-1.13.1+bzr2650 - Linaro GCC 2014.03) ) #767 PREEMPT Sat Mar 7 21:41:13 GMT 2015

extra/uname_string7

@@ -1 +1 @@
-Linux version 3.18.8-v7+ (dc4@dc4-XPS13-9333) (gcc version 4.8.3 20140303 (prerelease) (crosstool-NG linaro-1.13.1+bzr2650 - Linaro GCC 2014.03) ) #765 SMP PREEMPT Thu Mar 5 15:47:06 GMT 2015
+Linux version 3.18.9-v7+ (dc4@dc4-XPS13-9333) (gcc version 4.8.3 20140303 (prerelease) (crosstool-NG linaro-1.13.1+bzr2650 - Linaro GCC 2014.03) ) #767 SMP PREEMPT Sat Mar 7 21:52:35 GMT 2015

hardfp/opt/vc/include/interface/mmal/mmal_encodings.h

@@ -39,6 +39,7 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /** \name Pre-defined video encodings */
 /* @{ */
 #define MMAL_ENCODING_H264             MMAL_FOURCC('H','2','6','4')
+#define MMAL_ENCODING_MVC              MMAL_FOURCC('M','V','C',' ')
 #define MMAL_ENCODING_H263             MMAL_FOURCC('H','2','6','3')
 #define MMAL_ENCODING_MP4V             MMAL_FOURCC('M','P','4','V')
 #define MMAL_ENCODING_MP2V             MMAL_FOURCC('M','P','2','V')

hardfp/opt/vc/src/hello_pi/hello_fft/gpu_fft.c

@@ -1,6 +1,6 @@
 /*
-BCM2835 "GPU_FFT" release 2.0
-Copyright (c) 2014, Andrew Holme.
+BCM2835 "GPU_FFT" release 3.0
+Copyright (c) 2015, Andrew Holme.
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without
@@ -36,7 +36,7 @@ typedef struct GPU_FFT_COMPLEX COMPLEX;
 
 int gpu_fft_prepare(
     int mb,         // mailbox file_desc
-    int log2_N,     // log2(FFT_length) = 8...20
+    int log2_N,     // log2(FFT_length) = 8...22
     int direction,  // GPU_FFT_FWD: fft(); GPU_FFT_REV: ifft()
     int jobs,       // number of transforms in batch
     struct GPU_FFT **fft) {

hardfp/opt/vc/src/hello_pi/hello_fft/gpu_fft.h

@@ -1,6 +1,6 @@
 /*
-BCM2835 "GPU_FFT" release 2.0
-Copyright (c) 2014, Andrew Holme.
+BCM2835 "GPU_FFT" release 3.0
+Copyright (c) 2015, Andrew Holme.
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without
@@ -51,7 +51,7 @@ struct GPU_FFT_PTR {
 
 struct GPU_FFT_BASE {
     int mb;
-    unsigned handle, size, vc_msg, vc_code, vc_unifs[GPU_FFT_QPUS];
+    unsigned handle, size, vc_msg, vc_code, vc_unifs[GPU_FFT_QPUS], peri_size;
     volatile unsigned *peri;
 };
 
@@ -63,7 +63,7 @@ struct GPU_FFT {
 
 int gpu_fft_prepare(
     int mb,         // mailbox file_desc
-    int log2_N,     // log2(FFT_length) = 8...20
+    int log2_N,     // log2(FFT_length) = 8...22
     int direction,  // GPU_FFT_FWD: fft(); GPU_FFT_REV: ifft()
     int jobs,       // number of transforms in batch
     struct GPU_FFT **fft);

hardfp/opt/vc/src/hello_pi/hello_fft/gpu_fft.txt

@@ -1,10 +1,10 @@
-BCM2835 "GPU_FFT" release 2.0 by Andrew Holme, 2014.
+BCM2835 "GPU_FFT" release 3.0 by Andrew Holme, 2015.
 
 GPU_FFT is an FFT library for the Raspberry Pi which exploits the BCM2835 SoC
 3D hardware to deliver ten times more data throughput than is possible on the
-700 MHz ARM.  Kernels are provided for all power-of-2 FFT lengths between 256
-and 2,097,152 points inclusive.  A transpose function, which also uses the 3D
-hardware, is provided to support 2-dimensional transforms.
+700 MHz ARM of the Pi 1.  Kernels are provided for all power-of-2 FFT lengths
+between 256 and 4,194,304 points inclusive.  A transpose function, which also
+uses the 3D hardware, is provided to support 2-dimensional transforms.
 
 
 *** Accuracy ***
@@ -13,32 +13,36 @@ GPU_FFT uses single-precision floats for data and twiddle factors.  The output
 is not scaled.  The relative root-mean-square (rms) error in parts-per-million
 (ppm) for different transform lengths (N) is typically:
 
-log2(N) |  8    | 9    | 10   |  11   |  12  |  13  |  14  |  15  |  16 |  17
-ppm rms |  0.27 | 0.42 | 0.50 |  0.70 |  2.3 |  4.4 |  7.6 |  9.2 |  18 |  70
+log2(N) |  8    |  9    | 10    | 11    | 12    | 13    | 14    | 15
+ppm rms |  0.33 |  0.46 |  0.52 |  0.59 |  0.78 |  0.83 |  0.92 |  0.98
 
-log2(N) |  18 |  19 |  20 |  21 |                8...17 batch of 10
-ppm rms | 100 | 180 | 360 | 720 |               18...21 batch of  1
+log2(N) | 16    | 17    | 18    | 19    | 20    |  21   | 22
+ppm rms |  1.0  |  1.3  |  1.3  |  1.4  |  1.5  |  1.5  |  1.5
+
+Accuracy has improved significantly over previous releases at the expense of a
+small (2%) performance hit; however, FFTW is still one order of magnitude more
+accurate than GPU_FFT.
 
 
 *** Throughput ***
 
 GPU_FFT 1.0 had to be invoked through a "mailbox" which added a 100us overhead
 on every call.  To mitigate this, batches of transforms could be submitted via
-a single call.  GPU_FFT 2.0 avoids this 100us overhead by poking GPU registers
+a single call.  GPU_FFT now avoids this 100us overhead by poking GPU registers
 directly from the ARM if total batch runtime will be short; but still uses the
 mailbox for longer jobs to avoid busy waiting at 100% CPU for too long.
 
 Typical per-transform runtimes for batch sizes of 1 and 10; and comparative
-figures for FFTW (FFTW_MEASURE mode) are:
+figures for FFTW (FFTW_MEASURE mode) on a Pi 1 with L2 cache enabled are:
 
-log2(N) |   8   |   9   |  10   |  11   |  12  |  13  |  14  |  15  |
-      1 | 0.036 | 0.051 | 0.070 | 0.11  | 0.24 | 0.58 |  1.2 |  3.3 |
-     10 | 0.016 | 0.027 | 0.045 | 0.095 | 0.25 | 0.61 |  1.2 |  3.2 |
-   FFTW | 0.092 | 0.22  | 0.48  | 0.95  | 3.0  | 5.1  | 12   | 31   |
+log2(N) |   8   |   9   |  10   |  11   |  12  |  13  |  14  |  15
+      1 | 0.033 | 0.049 | 0.070 | 0.12  | 0.25 | 0.61 |  1.2 |  3.5
+     10 | 0.017 | 0.029 | 0.049 | 0.11  | 0.27 | 0.66 |  1.2 |  3.3
+   FFTW | 0.092 | 0.22  | 0.48  | 0.95  | 3.0  | 5.1  | 12   | 31
 
-log2(N) |  16  |  17 |  18 |  19 |   20 |   21 |       All times in
-      1 |  6.8 |  16 |  42 |  95 |  190 |  380 |       milliseconds
-   FFTW | 83   | 180 | 560 | 670 | 1600 | 3400 |       2 sig. figs.
+log2(N) |  16  |  17 |  18 |  19 |   20 |   21 |   22        All times in
+      1 |  7.0 |  17 |  43 |  97 |  194 |  388 |  786        milliseconds
+   FFTW | 83   | 180 | 560 | 670 | 1600 | 3400 | 8800        2 sig. figs.
 
 
 *** API functions ***
@@ -57,7 +61,7 @@ log2(N) |  16  |  17 |  18 |  19 |   20 |   21 |       All times in
 
     int mb          Mailbox file descriptor obtained by calling mbox_open()
 
-    int log2_N      log2(FFT length) = 8 to 21
+    int log2_N      log2(FFT length) = 8 to 22
 
     int direction   FFT direction:  GPU_FFT_FWD for forward FFT
                                     GPU_FFT_REV for inverse FFT
@@ -105,7 +109,6 @@ as a demo with the latest Raspbian distro.  Build and run it as follows:
 
 cd /opt/vc/src/hello_pi/hello_fft
 make
-sudo mknod char_dev c 100 0
 sudo ./hello_fft.bin 12
 
 It accepts three optional command-line arguments: <log2_N> <batch> <loops>
@@ -114,13 +117,12 @@ The special character device is required for the ioctl mailbox through which
 the ARM communicates with the Videocore GPU.
 
 
-*** With Open GL ***
+*** With Open GL on Pi 1 ***
 
-GPU_FFT and Open GL will run concurrently if the GPU_FFT_MEM_* defines in
-file gpu_fft.c are changed as follows:
+GPU_FFT and Open GL will run concurrently on Pi 1 if GPU_FFT is configured not
+to use VC4 L2 cache by zeroing a define in file gpu_fft_base.c as follows:
 
-#define GPU_FFT_MEM_FLG 0x4        // cached=0xC; direct=0x4
-#define GPU_FFT_MEM_MAP 0x20000000 // cached=0x0; direct=0x20000000
+#define GPU_FFT_USE_VC4_L2_CACHE 0 // Pi 1 only: cached=1; direct=0
 
 Overall performance will probably be higher if GPU_FFT and Open GL take turns
 at using the 3D hardware.  Since eglSwapBuffers() returns immediately without

hardfp/opt/vc/src/hello_pi/hello_fft/gpu_fft_base.c

@@ -1,6 +1,6 @@
 /*
-BCM2835 "GPU_FFT" release 2.0
-Copyright (c) 2014, Andrew Holme.
+BCM2835 "GPU_FFT" release 3.0
+Copyright (c) 2015, Andrew Holme.
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without
@@ -26,7 +26,8 @@ ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
-#include "bcm_host.h"
+#include <dlfcn.h>
+
 #include "gpu_fft.h"
 #include "mailbox.h"
 
@@ -42,11 +43,48 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define V3D_DBQITE  (0xC00e2c>>2)
 #define V3D_DBQITC  (0xC00e30>>2)
 
-#define GPU_FFT_MEM_MAP 0x0 // cached=0x0; direct=0x20000000
+// Setting this define to zero on Pi 1 allows GPU_FFT and Open GL
+// to co-exist and also improves performance of longer transforms:
+#define GPU_FFT_USE_VC4_L2_CACHE 1 // Pi 1 only: cached=1; direct=0
 
 #define GPU_FFT_NO_FLUSH 1
 #define GPU_FFT_TIMEOUT 2000 // ms
 
+struct GPU_FFT_HOST {
+    unsigned mem_flg, mem_map, peri_addr, peri_size;
+};
+
+int gpu_fft_get_host_info(struct GPU_FFT_HOST *info) {
+    void *handle;
+    unsigned (*bcm_host_get_sdram_address)     (void);
+    unsigned (*bcm_host_get_peripheral_address)(void);
+    unsigned (*bcm_host_get_peripheral_size)   (void);
+
+    // Pi 1 defaults
+    info->peri_addr = 0x20000000;
+    info->peri_size = 0x01000000;
+    info->mem_flg = GPU_FFT_USE_VC4_L2_CACHE? 0xC : 0x4;
+    info->mem_map = GPU_FFT_USE_VC4_L2_CACHE? 0x0 : 0x20000000; // Pi 1 only
+
+    handle = dlopen("libbcm_host.so", RTLD_LAZY);
+    if (!handle) return -1;
+
+    *(void **) (&bcm_host_get_sdram_address)      = dlsym(handle, "bcm_host_get_sdram_address");
+    *(void **) (&bcm_host_get_peripheral_address) = dlsym(handle, "bcm_host_get_peripheral_address");
+    *(void **) (&bcm_host_get_peripheral_size)    = dlsym(handle, "bcm_host_get_peripheral_size");
+
+    if (bcm_host_get_sdram_address && bcm_host_get_sdram_address()!=0x40000000) { // Pi 2?
+        info->mem_flg = 0x4; // ARM cannot see VC4 L2 on Pi 2
+        info->mem_map = 0x0;
+    }
+
+    if (bcm_host_get_peripheral_address) info->peri_addr = bcm_host_get_peripheral_address();
+    if (bcm_host_get_peripheral_size)    info->peri_size = bcm_host_get_peripheral_size();
+
+    dlclose(handle);
+    return 0;
+}
+
 unsigned gpu_fft_base_exec_direct (
     struct GPU_FFT_BASE *base,
     int num_qpus) {
@@ -95,43 +133,46 @@ int gpu_fft_alloc (
     unsigned size,
     struct GPU_FFT_PTR *ptr) {
 
+    struct GPU_FFT_HOST host;
     struct GPU_FFT_BASE *base;
     volatile unsigned *peri;
     unsigned handle;
 
+    if (gpu_fft_get_host_info(&host)) return -5;
+
     if (qpu_enable(mb, 1)) return -1;
 
-    // Shared memory : cached=0xC; direct=0x4
-    unsigned mem_flg = bcm_host_get_sdram_address() == 0x40000000 ? 0xC : 0x4;
-    handle = mem_alloc(mb, size, 4096, mem_flg);
+    // Shared memory
+    handle = mem_alloc(mb, size, 4096, host.mem_flg);
     if (!handle) {
         qpu_enable(mb, 0);
         return -3;
     }
 
-    peri = (volatile unsigned *) mapmem(bcm_host_get_peripheral_address(), bcm_host_get_peripheral_size());
+    peri = (volatile unsigned *) mapmem(host.peri_addr, host.peri_size);
     if (!peri) {
         mem_free(mb, handle);
         qpu_enable(mb, 0);
         return -4;
     }
 
     ptr->vc = mem_lock(mb, handle);
-    ptr->arm.vptr = mapmem(BUS_TO_PHYS(ptr->vc+GPU_FFT_MEM_MAP), size);
+    ptr->arm.vptr = mapmem(BUS_TO_PHYS(ptr->vc+host.mem_map), size);
 
     base = (struct GPU_FFT_BASE *) ptr->arm.vptr;
-    base->peri   = peri;
-    base->mb     = mb;
-    base->handle = handle;
-    base->size   = size;
+    base->peri      = peri;
+    base->peri_size = host.peri_size;
+    base->mb        = mb;
+    base->handle    = handle;
+    base->size      = size;
 
     return 0;
 }
 
 void gpu_fft_base_release(struct GPU_FFT_BASE *base) {
     int mb = base->mb;
     unsigned handle = base->handle, size = base->size;
-    unmapmem((void*)base->peri, bcm_host_get_peripheral_size());
+    unmapmem((void*)base->peri, base->peri_size);
     unmapmem((void*)base, size);
     mem_unlock(mb, handle);
     mem_free(mb, handle);

hardfp/opt/vc/src/hello_pi/hello_fft/gpu_fft_shaders.c

@@ -1,6 +1,6 @@
 /*
-BCM2835 "GPU_FFT" release 2.0
-Copyright (c) 2014, Andrew Holme.
+BCM2835 "GPU_FFT" release 3.0
+Copyright (c) 2015, Andrew Holme.
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without
@@ -68,6 +68,9 @@ static unsigned int shader_1024k[] = {
 static unsigned int shader_2048k[] = {
     #include "hex/shader_2048k.hex"
 };
+static unsigned int shader_4096k[] = {
+    #include "hex/shader_4096k.hex"
+};
 
 static struct {
     unsigned int size, *code;
@@ -86,7 +89,8 @@ shaders[] = {
     {sizeof(shader_256k), shader_256k},
     {sizeof(shader_512k), shader_512k},
     {sizeof(shader_1024k), shader_1024k},
-    {sizeof(shader_2048k), shader_2048k}
+    {sizeof(shader_2048k), shader_2048k},
+    {sizeof(shader_4096k), shader_4096k}
 };
 
 unsigned int  gpu_fft_shader_size(int log2_N) {

hardfp/opt/vc/src/hello_pi/hello_fft/gpu_fft_twiddles.c

@@ -1,6 +1,6 @@
 /*
-BCM2835 "GPU_FFT" release 2.0
-Copyright (c) 2014, Andrew Holme.
+BCM2835 "GPU_FFT" release 3.0
+Copyright (c) 2015, Andrew Holme.
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without
@@ -85,6 +85,15 @@ static float *twiddles_step_32(double two_pi, float *out, double theta) {
     return twiddles_step_16(two_pi, out, 2*theta);
 }
 
+static float *twiddles_step_64(double two_pi, float *out, double theta) {
+    int i;
+    for (i=0; i<32; i++) {
+        *out++ = ALPHA(theta);
+        *out++ =  BETA(theta);
+    }
+    return twiddles_step_32(two_pi, out, 2*theta);
+}
+
 /****************************************************************************/
 
 static void twiddles_256(double two_pi, float *out) {
@@ -256,6 +265,19 @@ static void twiddles_2048k(double two_pi, float *out) {
         out = twiddles_base_32(two_pi, out, two_pi/N*q);
 }
 
+static void twiddles_4096k(double two_pi, float *out) {
+    double N=4096*1024;
+    int q;
+
+    out = twiddles_base_64(two_pi, out);
+    out = twiddles_step_64(two_pi, out, two_pi/N * 32*32);
+    out = twiddles_step_32(two_pi, out, two_pi/N * 32);
+    out = twiddles_step_32(two_pi, out, two_pi/N * GPU_FFT_QPUS);
+
+    for (q=0; q<GPU_FFT_QPUS; q++)
+        out = twiddles_base_32(two_pi, out, two_pi/N*q);
+}
+
 /****************************************************************************/
 
 static struct {
@@ -276,11 +298,12 @@ shaders[] = {
     {4, 6, 2, twiddles_256k},
     {4, 7, 2, twiddles_512k},
     {4, 8, 2, twiddles_1024k},
-    {4,10, 2, twiddles_2048k}
+    {4,10, 2, twiddles_2048k},
+    {4,12, 2, twiddles_4096k}
 };
 
 int gpu_fft_twiddle_size(int log2_N, int *shared, int *unique, int *passes) {
-    if (log2_N<8 || log2_N>21) return -1;
+    if (log2_N<8 || log2_N>22) return -1;
     *shared = shaders[log2_N-8].shared;
     *unique = shaders[log2_N-8].unique;
     *passes = shaders[log2_N-8].passes;