ENGR00132486 GPU: Upgrade to Rel 1.1

Upgrade to AMD "Production Release 1.1 for September 10,2010"

Signed-off-by: Richard Zhao <richard.zhao@freescale.com>

24 files changed, 410 insertions, 644 deletions

diff --git a/drivers/mxc/amd-gpu/common/gsl_cmdstream.c b/drivers/mxc/amd-gpu/common/gsl_cmdstream.c
index 338192c7ba32..4f0d107f2f67 100644
--- a/drivers/mxc/amd-gpu/common/gsl_cmdstream.c
+++ b/drivers/mxc/amd-gpu/common/gsl_cmdstream.c
@@ -20,6 +20,20 @@
 #include "gsl_hal.h"
 #include "gsl_cmdstream.h"
 
+#ifdef GSL_LOCKING_FINEGRAIN
+#define GSL_CMDSTREAM_MUTEX_CREATE()        device->cmdstream_mutex = kos_mutex_create("gsl_cmdstream"); \
+                                            if (!device->cmdstream_mutex) return (GSL_FAILURE);
+#define GSL_CMDSTREAM_MUTEX_LOCK()          kos_mutex_lock(device->cmdstream_mutex)
+#define GSL_CMDSTREAM_MUTEX_UNLOCK()        kos_mutex_unlock(device->cmdstream_mutex)
+#define GSL_CMDSTREAM_MUTEX_FREE()          kos_mutex_free(device->cmdstream_mutex); device->cmdstream_mutex = 0;
+#else
+#define GSL_CMDSTREAM_MUTEX_CREATE()
+#define GSL_CMDSTREAM_MUTEX_LOCK()
+#define GSL_CMDSTREAM_MUTEX_UNLOCK()
+#define GSL_CMDSTREAM_MUTEX_FREE()
+#endif
+
+
 //////////////////////////////////////////////////////////////////////////////
 // functions
 //////////////////////////////////////////////////////////////////////////////
@@ -27,6 +41,8 @@
 int
 kgsl_cmdstream_init(gsl_device_t *device)
 {
+    GSL_CMDSTREAM_MUTEX_CREATE();
+
     return GSL_SUCCESS;
 }
 
@@ -35,6 +51,8 @@ kgsl_cmdstream_init(gsl_device_t *device)
 int
 kgsl_cmdstream_close(gsl_device_t *device)
 {
+    GSL_CMDSTREAM_MUTEX_FREE();
+
     return GSL_SUCCESS;
 }
 
@@ -133,17 +151,22 @@ kgsl_cmdstream_memqueue_drain(gsl_device_t *device)
     gsl_memnode_t     *memnode, *nextnode, *freehead;
     gsl_timestamp_t   timestamp, ts_processed;
     gsl_memqueue_t    *memqueue = &device->memqueue;
+
+    GSL_CMDSTREAM_MUTEX_LOCK();
+
     // check head
     if (memqueue->head == NULL)
     {
+        GSL_CMDSTREAM_MUTEX_UNLOCK();
         return;
     }
     // get current EOP timestamp
     ts_processed = kgsl_cmdstream_readtimestamp0(device->id, GSL_TIMESTAMP_RETIRED);
     timestamp = memqueue->head->timestamp;
     // check head timestamp
-    if (!(((ts_processed - timestamp) >= 0) || ((ts_processed - timestamp) < -20000)))
+    if (!(((ts_processed - timestamp) >= 0) || ((ts_processed - timestamp) < -GSL_TIMESTAMP_EPSILON)))
     {
+        GSL_CMDSTREAM_MUTEX_UNLOCK();
         return;
     }
     memnode  = memqueue->head;
@@ -160,7 +183,7 @@ kgsl_cmdstream_memqueue_drain(gsl_device_t *device)
             break;
         }
         timestamp = nextnode->timestamp;
-        if (!(((ts_processed - timestamp) >= 0) || ((ts_processed - timestamp) < -20000)))
+        if (!(((ts_processed - timestamp) >= 0) || ((ts_processed - timestamp) < -GSL_TIMESTAMP_EPSILON)))
         {
             // drained up to a point
             memqueue->head = nextnode;
@@ -177,6 +200,8 @@ kgsl_cmdstream_memqueue_drain(gsl_device_t *device)
         kgsl_sharedmem_free0(&memnode->memdesc, memnode->pid);
         kos_free(memnode);
     }
+
+    GSL_CMDSTREAM_MUTEX_UNLOCK();
 }
 
 //----------------------------------------------------------------------------
@@ -190,6 +215,7 @@ kgsl_cmdstream_freememontimestamp(gsl_deviceid_t device_id, gsl_memdesc_t *memde
     (void)type; // unref. For now just use EOP timestamp
 
 	GSL_API_MUTEX_LOCK();
+	GSL_CMDSTREAM_MUTEX_LOCK();
 
 	memqueue = &device->memqueue;
 
@@ -198,6 +224,7 @@ kgsl_cmdstream_freememontimestamp(gsl_deviceid_t device_id, gsl_memdesc_t *memde
     if (!memnode)
     {
         // other solution is to idle and free which given that the upper level driver probably wont check, probably a better idea
+		GSL_CMDSTREAM_MUTEX_UNLOCK();
 		GSL_API_MUTEX_UNLOCK();
         return (GSL_FAILURE);
     }
@@ -220,6 +247,7 @@ kgsl_cmdstream_freememontimestamp(gsl_deviceid_t device_id, gsl_memdesc_t *memde
         memqueue->tail = memnode;
     }
 
+    GSL_CMDSTREAM_MUTEX_UNLOCK();
 	GSL_API_MUTEX_UNLOCK();
 
     return (GSL_SUCCESS);
@@ -228,7 +256,7 @@ kgsl_cmdstream_freememontimestamp(gsl_deviceid_t device_id, gsl_memdesc_t *memde
 static int kgsl_cmdstream_timestamp_cmp(gsl_timestamp_t ts_new, gsl_timestamp_t ts_old)
 {
 	gsl_timestamp_t ts_diff = ts_new - ts_old;
-	return (ts_diff >= 0) || (ts_diff < -20000);
+	return (ts_diff >= 0) || (ts_diff < -GSL_TIMESTAMP_EPSILON);
 }
 
 int kgsl_cmdstream_check_timestamp(gsl_deviceid_t device_id, gsl_timestamp_t timestamp)
diff --git a/drivers/mxc/amd-gpu/common/gsl_cmdwindow.c b/drivers/mxc/amd-gpu/common/gsl_cmdwindow.c
index 4d70da5b25c8..d19832d8da44 100644
--- a/drivers/mxc/amd-gpu/common/gsl_cmdwindow.c
+++ b/drivers/mxc/amd-gpu/common/gsl_cmdwindow.c
@@ -31,12 +31,31 @@
 
 
 //////////////////////////////////////////////////////////////////////////////
+// macros
+//////////////////////////////////////////////////////////////////////////////
+#ifdef GSL_LOCKING_FINEGRAIN
+#define GSL_CMDWINDOW_MUTEX_CREATE()        device->cmdwindow_mutex = kos_mutex_create("gsl_cmdwindow"); \
+                                            if (!device->cmdwindow_mutex) return (GSL_FAILURE);
+#define GSL_CMDWINDOW_MUTEX_LOCK()          kos_mutex_lock(device->cmdwindow_mutex)
+#define GSL_CMDWINDOW_MUTEX_UNLOCK()        kos_mutex_unlock(device->cmdwindow_mutex)
+#define GSL_CMDWINDOW_MUTEX_FREE()          kos_mutex_free(device->cmdwindow_mutex); device->cmdwindow_mutex = 0;
+#else
+#define GSL_CMDWINDOW_MUTEX_CREATE()
+#define GSL_CMDWINDOW_MUTEX_LOCK()
+#define GSL_CMDWINDOW_MUTEX_UNLOCK()
+#define GSL_CMDWINDOW_MUTEX_FREE()
+#endif
+
+
+//////////////////////////////////////////////////////////////////////////////
 // functions
 //////////////////////////////////////////////////////////////////////////////
 
 int
 kgsl_cmdwindow_init(gsl_device_t *device)
 {
+    GSL_CMDWINDOW_MUTEX_CREATE();
+
     return (GSL_SUCCESS);
 }
 
@@ -45,6 +64,8 @@ kgsl_cmdwindow_init(gsl_device_t *device)
 int 
 kgsl_cmdwindow_close(gsl_device_t *device)
 {
+    GSL_CMDWINDOW_MUTEX_FREE();
+
     return (GSL_SUCCESS);
 }
 
@@ -98,6 +119,8 @@ kgsl_cmdwindow_write0(gsl_deviceid_t device_id, gsl_cmdwindow_t target, unsigned
     cmdwinaddr  = ((target << GSL_CMDWINDOW_TARGET_SHIFT) & GSL_CMDWINDOW_TARGET_MASK);
     cmdwinaddr |= ((addr   << GSL_CMDWINDOW_ADDR_SHIFT)   & GSL_CMDWINDOW_ADDR_MASK);
 
+    GSL_CMDWINDOW_MUTEX_LOCK();
+
 #ifndef GSL_NO_MMU
     // set mmu pagetable
 	kgsl_mmu_setpagetable(device, GSL_CALLER_PROCESSID_GET());
@@ -109,6 +132,8 @@ kgsl_cmdwindow_write0(gsl_deviceid_t device_id, gsl_cmdwindow_t target, unsigned
     // write data
     device->ftbl.device_regwrite(device, (cmdstream)>>2, data);
 
+    GSL_CMDWINDOW_MUTEX_UNLOCK();
+
     kgsl_log_write( KGSL_LOG_GROUP_COMMAND | KGSL_LOG_LEVEL_TRACE, "<-- kgsl_cmdwindow_write. Return value %B\n", GSL_SUCCESS );
 
     return (GSL_SUCCESS);
diff --git a/drivers/mxc/amd-gpu/common/gsl_device.c b/drivers/mxc/amd-gpu/common/gsl_device.c
index c399f6da2762..6c41d3dd6dcc 100644
--- a/drivers/mxc/amd-gpu/common/gsl_device.c
+++ b/drivers/mxc/amd-gpu/common/gsl_device.c
@@ -565,6 +565,17 @@ kgsl_device_idle(gsl_deviceid_t device_id, unsigned int timeout)
 //----------------------------------------------------------------------------
 
 KGSL_API int
+kgsl_device_isidle(gsl_deviceid_t device_id)
+{
+	gsl_timestamp_t retired = kgsl_cmdstream_readtimestamp0(device_id, GSL_TIMESTAMP_RETIRED);
+	gsl_timestamp_t consumed = kgsl_cmdstream_readtimestamp0(device_id, GSL_TIMESTAMP_CONSUMED);
+	gsl_timestamp_t ts_diff = retired - consumed;
+	return (ts_diff >= 0) || (ts_diff < -GSL_TIMESTAMP_EPSILON) ? GSL_SUCCESS : GSL_FAILURE;
+}
+
+//----------------------------------------------------------------------------
+
+KGSL_API int
 kgsl_device_regread(gsl_deviceid_t device_id, unsigned int offsetwords, unsigned int *value)
 {
     int           status = GSL_FAILURE_NOTINITIALIZED;
diff --git a/drivers/mxc/amd-gpu/common/gsl_drawctxt.c b/drivers/mxc/amd-gpu/common/gsl_drawctxt.c
index afa44e618794..1e8fa1a4962e 100644
--- a/drivers/mxc/amd-gpu/common/gsl_drawctxt.c
+++ b/drivers/mxc/amd-gpu/common/gsl_drawctxt.c
@@ -24,9 +24,6 @@
 
 #ifdef GSL_BLD_YAMATO
 
-//#define DISABLE_SHADOW_WRITES
-
-/*
 //////////////////////////////////////////////////////////////////////////////
 //
 // Memory Map for Register, Constant & Instruction Shadow, and Command Buffers (34.5KB)
@@ -34,8 +31,8 @@
 // +---------------------+------------+-------------+---+---------------------+
 // | ALU Constant Shadow | Reg Shadow | C&V Buffers |Tex| Shader Instr Shadow |
 // +---------------------+------------+-------------+---+---------------------+
-//   ________________________________/               \___________________
-//  /                                                                    \
+//   ________________________________'               '___________________
+//  '                                                                    '
 // +--------------+-----------+------+-----------+------------------------+
 // | Restore Regs | Save Regs | Quad | Gmem Save | Gmem Restore | unused  |
 // +--------------+-----------+------+-----------+------------------------+
@@ -69,15 +66,11 @@
 // constants would require an additional 8KB each, for alignment.]
 //
 //////////////////////////////////////////////////////////////////////////////
-*/
 
 //////////////////////////////////////////////////////////////////////////////
 // Constants
 //////////////////////////////////////////////////////////////////////////////
 
-
-
-
 #define ALU_CONSTANTS           2048                // DWORDS
 #define NUM_REGISTERS           1024                // DWORDS
 #ifdef DISABLE_SHADOW_WRITES
@@ -117,6 +110,23 @@
 #define CONTEXT_SIZE            (SHADER_OFFSET + 3 * SHADER_SHADOW_SIZE)
 
 
+/////////////////////////////////////////////////////////////////////////////
+// macros
+//////////////////////////////////////////////////////////////////////////////
+#ifdef GSL_LOCKING_FINEGRAIN
+#define GSL_CONTEXT_MUTEX_CREATE()          device->drawctxt_mutex = kos_mutex_create("gsl_drawctxt"); \
+                                            if (!device->drawctxt_mutex) {return (GSL_FAILURE);}
+#define GSL_CONTEXT_MUTEX_LOCK()            kos_mutex_lock(device->drawctxt_mutex)
+#define GSL_CONTEXT_MUTEX_UNLOCK()          kos_mutex_unlock(device->drawctxt_mutex)
+#define GSL_CONTEXT_MUTEX_FREE()            kos_mutex_free(device->drawctxt_mutex); device->drawctxt_mutex = 0;
+#else
+#define GSL_CONTEXT_MUTEX_CREATE()
+#define GSL_CONTEXT_MUTEX_LOCK()
+#define GSL_CONTEXT_MUTEX_UNLOCK()
+#define GSL_CONTEXT_MUTEX_FREE()
+#endif
+
+
 //////////////////////////////////////////////////////////////////////////////
 // temporary work structure
 //////////////////////////////////////////////////////////////////////////////
@@ -392,6 +402,41 @@ static unsigned int gmem_copy_texcoord[TEXCOORD_LEN] = {
     0x3f800000, 0x00000000
 };
 
+#define NUM_COLOR_FORMATS   13
+
+static SurfaceFormat surface_format_table[NUM_COLOR_FORMATS] =
+{
+	FMT_4_4_4_4,		   // COLORX_4_4_4_4
+	FMT_1_5_5_5,		   // COLORX_1_5_5_5
+	FMT_5_6_5,		       // COLORX_5_6_5
+	FMT_8,			       // COLORX_8
+	FMT_8_8,		       // COLORX_8_8
+	FMT_8_8_8_8,		   // COLORX_8_8_8_8
+	FMT_8_8_8_8,		   // COLORX_S8_8_8_8
+	FMT_16_FLOAT,		   // COLORX_16_FLOAT
+	FMT_16_16_FLOAT,	   // COLORX_16_16_FLOAT
+	FMT_16_16_16_16_FLOAT, // COLORX_16_16_16_16_FLOAT
+	FMT_32_FLOAT,		   // COLORX_32_FLOAT
+	FMT_32_32_FLOAT,	   // COLORX_32_32_FLOAT
+	FMT_32_32_32_32_FLOAT, // COLORX_32_32_32_32_FLOAT
+};
+
+static unsigned int format2bytesperpixel[NUM_COLOR_FORMATS] =
+{
+    2,  // COLORX_4_4_4_4
+	2,  // COLORX_1_5_5_5
+	2,  // COLORX_5_6_5
+	1,  // COLORX_8
+	2,  // COLORX_8_8_8
+	4,  // COLORX_8_8_8_8
+	4,  // COLORX_S8_8_8_8
+	2,  // COLORX_16_FLOAT
+	4,  // COLORX_16_16_FLOAT
+	8,  // COLORX_16_16_16_16_FLOAT
+	4,  // COLORX_32_FLOAT
+	8,  // COLORX_32_32_FLOAT
+	16, // COLORX_32_32_32_32_FLOAT
+};
 
 //////////////////////////////////////////////////////////////////////////////
 // shader linkage info
@@ -426,10 +471,14 @@ config_gmemsize(gmem_shadow_t *shadow, int gmem_size)
             h *= 2;
 
     shadow->width  = w;
-    shadow->pitch  = w;
     shadow->height = h;
+    shadow->pitch  = w;
+	shadow->format = COLORX_8_8_8_8;
+    shadow->size   = shadow->pitch * shadow->height * 4;
 
-    shadow->size = shadow->pitch * shadow->height * 4;
+	shadow->gmem_width  = w;
+	shadow->gmem_height = h;
+	shadow->gmem_pitch  = w;
 }
 
 
@@ -490,7 +539,7 @@ reg_to_mem(unsigned int *cmds, gpuaddr_t dst, gpuaddr_t src, int dwords)
 
 #ifdef DISABLE_SHADOW_WRITES
 
-static void build_reg_to_mem_range(unsigned int start, unsigned int end, unsigned int** cmd, /*ctx_t *ctx, unsigned int* offset) //*/gsl_drawctxt_t *drawctxt)
+static void build_reg_to_mem_range(unsigned int start, unsigned int end, unsigned int** cmd, gsl_drawctxt_t *drawctxt)
 {
     unsigned int i = start;
 
@@ -687,16 +736,19 @@ build_gmem2sys_cmds(gsl_drawctxt_t *drawctxt, ctx_t* ctx, gmem_shadow_t *shadow)
     // change colour buffer to RGBA8888, MSAA = 1, and matching pitch
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 3);
     *cmds++ = PM4_REG(mmRB_SURFACE_INFO);
-    *cmds++ = drawctxt->context_gmem_shadow.pitch; // GMEM pitch is equal to context GMEM shadow pitch
+    *cmds++ = shadow->gmem_pitch;
 
     // RB_COLOR_INFO        Endian=none, Linear, Format=RGBA8888, Swap=0, Base=gmem_base
     if( ctx )
     {
         KOS_ASSERT((ctx->gmem_base & 0xFFF) == 0);   // gmem base assumed 4K aligned.
-        *cmds++ = (COLORX_8_8_8_8 << RB_COLOR_INFO__COLOR_FORMAT__SHIFT) | ctx->gmem_base;
+        *cmds++ = (shadow->format << RB_COLOR_INFO__COLOR_FORMAT__SHIFT) | ctx->gmem_base;
     }
     else
-        cmds++;
+    {
+		unsigned int temp = *cmds;
+		*cmds++ = (temp & ~RB_COLOR_INFO__COLOR_FORMAT_MASK) | (shadow->format << RB_COLOR_INFO__COLOR_FORMAT__SHIFT);
+    }
 
     // disable Z
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 2);
@@ -714,12 +766,12 @@ build_gmem2sys_cmds(gsl_drawctxt_t *drawctxt, ctx_t* ctx, gmem_shadow_t *shadow)
     // set the scissor to the extents of the draw surface
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 3);
     *cmds++ = PM4_REG(mmPA_SC_SCREEN_SCISSOR_TL);
-    *cmds++ = (0 << 16) | 0;
-    *cmds++ = (drawctxt->context_gmem_shadow.height << 16) | drawctxt->context_gmem_shadow.width;
+    *cmds++ = (shadow->gmem_offset_y << 16) | shadow->gmem_offset_x;
+    *cmds++ = (shadow->gmem_height << 16) | shadow->gmem_width;
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 3);
     *cmds++ = PM4_REG(mmPA_SC_WINDOW_SCISSOR_TL);
-    *cmds++ = (unsigned int) ((1U << 31) | (0 << 16) | 0);
-    *cmds++ = (drawctxt->context_gmem_shadow.height << 16) | drawctxt->context_gmem_shadow.width;
+    *cmds++ = (1U << 31) | (0 << 16) | 0;
+    *cmds++ = (shadow->height << 16) | shadow->width;
 
     // load the viewport so that z scale = clear depth and z offset = 0.0f
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 3);
@@ -732,19 +784,18 @@ build_gmem2sys_cmds(gsl_drawctxt_t *drawctxt, ctx_t* ctx, gmem_shadow_t *shadow)
     *cmds++ = PM4_REG(mmRB_COPY_CONTROL);
     *cmds++ = 0;                                                // RB_COPY_CONTROL
 
-    *cmds++ = (shadow->gmemshadow.gpuaddr+shadow->offset*4) & 0xfffff000;   // RB_COPY_DEST_BASE
-
-    *cmds++ = shadow->pitch >> 5;           // RB_COPY_DEST_PITCH
-    *cmds++ = 0x0003c058;                   // Endian=none, Linear, Format=RGBA8888,Swap=0,!Dither,MaskWrite:R=G=B=A=1
-
     {
         // Calculate the new offset based on the adjusted base
-        unsigned int addr = (shadow->gmemshadow.gpuaddr+shadow->offset*4);
-        unsigned int offset = (addr-(addr&0xfffff000))/4;
+		unsigned int bytesperpixel = format2bytesperpixel[shadow->format];
+		unsigned int addr = (shadow->gmemshadow.gpuaddr + shadow->offset * bytesperpixel);
+		unsigned int offset = (addr - (addr & 0xfffff000)) / bytesperpixel;
 
-        kos_assert( (offset & 0xfffff000) == 0 ); // Make sure we stay in offsetx field.
+		*cmds++ = addr & 0xfffff000;   // RB_COPY_DEST_BASE
+		*cmds++ = shadow->pitch >> 5;  // RB_COPY_DEST_PITCH
+		*cmds++ = 0x0003c008 | (shadow->format << RB_COPY_DEST_INFO__COPY_DEST_FORMAT__SHIFT); // Endian=none, Linear, Format=RGBA8888,Swap=0,!Dither,MaskWrite:R=G=B=A=1
 
-        *cmds++ = offset;
+        KOS_ASSERT( (offset & 0xfffff000) == 0 ); // Make sure we stay in offsetx field.
+		*cmds++ = offset;
     }
 
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 2);
@@ -851,7 +902,6 @@ build_sys2gmem_cmds(gsl_drawctxt_t *drawctxt, ctx_t* ctx, gmem_shadow_t *shadow)
 
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 2);
     *cmds++ = PM4_REG(mmSQ_INTERPOLATOR_CNTL);
-    //*cmds++ = 0x0000ffff; //mmSQ_INTERPOLATOR_CNTL
     *cmds++ = 0xffffffff; //mmSQ_INTERPOLATOR_CNTL
 
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 2);
@@ -872,38 +922,40 @@ build_sys2gmem_cmds(gsl_drawctxt_t *drawctxt, ctx_t* ctx, gmem_shadow_t *shadow)
     *cmds++ = (0x1 << 16) | (0 * 6);
     kos_memcpy(cmds, sys2gmem_tex_const, SYS2GMEM_TEX_CONST_LEN<<2);
     cmds[0] |= (shadow->pitch >> 5) << 22;
-    cmds[1] |= shadow->gmemshadow.gpuaddr;
+    cmds[1] |= shadow->gmemshadow.gpuaddr | surface_format_table[shadow->format];
     cmds[2] |= (shadow->width+shadow->offset_x-1) | (shadow->height+shadow->offset_y-1) << 13;
     cmds += SYS2GMEM_TEX_CONST_LEN;
 
     // change colour buffer to RGBA8888, MSAA = 1, and matching pitch
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 3);
     *cmds++ = PM4_REG(mmRB_SURFACE_INFO);
-    *cmds++ = drawctxt->context_gmem_shadow.pitch; // GMEM pitch is equal to context GMEM shadow pitch
-
+    *cmds++ = shadow->gmem_pitch; // GMEM pitch is equal to context GMEM shadow pitch
 
     // RB_COLOR_INFO        Endian=none, Linear, Format=RGBA8888, Swap=0, Base=gmem_base
     if( ctx )
-        *cmds++ = (COLORX_8_8_8_8 << RB_COLOR_INFO__COLOR_FORMAT__SHIFT) | ctx->gmem_base;
+	{
+        *cmds++ = (shadow->format << RB_COLOR_INFO__COLOR_FORMAT__SHIFT) | ctx->gmem_base;
+	}
     else
-        cmds++;
+	{
+		unsigned int temp = *cmds;
+		*cmds++ = (temp & ~RB_COLOR_INFO__COLOR_FORMAT_MASK) | (shadow->format << RB_COLOR_INFO__COLOR_FORMAT__SHIFT);
+	}
 
     // RB_DEPTHCONTROL
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 2);
     *cmds++ = PM4_REG(mmRB_DEPTHCONTROL);
     *cmds++ = 0;                            // disable Z
 
-
     // set the scissor to the extents of the draw surface
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 3);
     *cmds++ = PM4_REG(mmPA_SC_SCREEN_SCISSOR_TL);
     *cmds++ = (0 << 16) | 0;
-    *cmds++ = (drawctxt->context_gmem_shadow.height << 16) | drawctxt->context_gmem_shadow.width;
-
+    *cmds++ = (shadow->height << 16) | shadow->width;
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 3);
     *cmds++ = PM4_REG(mmPA_SC_WINDOW_SCISSOR_TL);
-    *cmds++ = (unsigned int) ((1U << 31) | (shadow->gmem_offset_y << 16) | shadow->gmem_offset_x);
-    *cmds++ = (drawctxt->context_gmem_shadow.height << 16) | drawctxt->context_gmem_shadow.width;
+    *cmds++ = (1U << 31) | (shadow->gmem_offset_y << 16) | shadow->gmem_offset_x;
+    *cmds++ = (shadow->gmem_height << 16) | shadow->gmem_width;
 
     // PA_CL_VTE_CNTL
     *cmds++ = pm4_type3_packet(PM4_SET_CONSTANT, 2);
@@ -966,10 +1018,6 @@ build_regrestore_cmds(gsl_drawctxt_t *drawctxt, ctx_t *ctx)
     unsigned int *start = ctx->cmd;
     unsigned int *cmd = start;
 
-
-    //*cmd++ = pm4_type3_packet(PM4_WAIT_FOR_IDLE, 1);
-    //*cmd++ = 0;
-
     // H/W Registers
     cmd++; // deferred pm4_type3_packet(PM4_LOAD_CONSTANT_CONTEXT, ???);
 #ifdef DISABLE_SHADOW_WRITES
@@ -983,7 +1031,7 @@ build_regrestore_cmds(gsl_drawctxt_t *drawctxt, ctx_t *ctx)
     cmd = reg_range(cmd, mmVGT_MAX_VTX_INDX,                mmPA_CL_VPORT_ZOFFSET);
     cmd = reg_range(cmd, mmSQ_PROGRAM_CNTL,                 mmSQ_WRAPPING_1);
     cmd = reg_range(cmd, mmRB_DEPTHCONTROL,                 mmRB_MODECONTROL);
-    cmd = reg_range(cmd, mmPA_SU_POINT_SIZE,                mmPA_SC_VIZ_QUERY/*mmVGT_ENHANCE*/);
+    cmd = reg_range(cmd, mmPA_SU_POINT_SIZE,                mmPA_SC_VIZ_QUERY);
     cmd = reg_range(cmd, mmPA_SC_LINE_CNTL,                 mmRB_COLOR_DEST_MASK);
     cmd = reg_range(cmd, mmPA_SU_POLY_OFFSET_FRONT_SCALE,   mmPA_SU_POLY_OFFSET_BACK_OFFSET);
 
@@ -1062,10 +1110,10 @@ static void set_gmem_copy_quad( gmem_shadow_t* shadow )
 
     // set vertex buffer values
 
-    gmem_copy_quad[1] = uint2float( shadow->height + shadow->gmem_offset_y );
-    gmem_copy_quad[3] = uint2float( shadow->width + shadow->gmem_offset_x );
-    gmem_copy_quad[4] = uint2float( shadow->height + shadow->gmem_offset_y );
-    gmem_copy_quad[9] = uint2float( shadow->width + shadow->gmem_offset_x );
+    gmem_copy_quad[1] = uint2float( shadow->gmem_height + shadow->gmem_offset_y );
+    gmem_copy_quad[3] = uint2float( shadow->gmem_width + shadow->gmem_offset_x );
+    gmem_copy_quad[4] = uint2float( shadow->gmem_height + shadow->gmem_offset_y );
+    gmem_copy_quad[9] = uint2float( shadow->gmem_width + shadow->gmem_offset_x );
 
     gmem_copy_quad[0] = uint2float( shadow->gmem_offset_x );
     gmem_copy_quad[6] = uint2float( shadow->gmem_offset_x );
@@ -1358,6 +1406,8 @@ create_gmem_shadow(gsl_device_t *device, gsl_drawctxt_t *drawctxt, ctx_t *ctx)
 int
 kgsl_drawctxt_init(gsl_device_t *device)
 {
+    GSL_CONTEXT_MUTEX_CREATE();
+
     return (GSL_SUCCESS);
 }
 
@@ -1369,6 +1419,8 @@ kgsl_drawctxt_init(gsl_device_t *device)
 int
 kgsl_drawctxt_close(gsl_device_t *device)
 {
+    GSL_CONTEXT_MUTEX_FREE();
+
     return (GSL_SUCCESS);
 }
 
@@ -1386,8 +1438,10 @@ kgsl_drawctxt_create(gsl_device_t* device, gsl_context_type_t type, unsigned int
 
     kgsl_device_active(device);
     
+    GSL_CONTEXT_MUTEX_LOCK();
     if (device->drawctxt_count >= GSL_CONTEXT_MAX)
     {
+        GSL_CONTEXT_MUTEX_UNLOCK();
         return (GSL_FAILURE);
     }
 
@@ -1403,6 +1457,7 @@ kgsl_drawctxt_create(gsl_device_t* device, gsl_context_type_t type, unsigned int
 
     if (index >= GSL_CONTEXT_MAX)
     {
+        GSL_CONTEXT_MUTEX_UNLOCK();
         return (GSL_FAILURE);
     }
 
@@ -1422,14 +1477,12 @@ kgsl_drawctxt_create(gsl_device_t* device, gsl_context_type_t type, unsigned int
         if (create_gpustate_shadow(device, drawctxt, &ctx) != GSL_SUCCESS)
         {
             kgsl_drawctxt_destroy(device, index);
+            GSL_CONTEXT_MUTEX_UNLOCK();
             return (GSL_FAILURE);
         }
 
-        // Save the shader instruction memory on context switching
-        drawctxt->flags |= CTXT_FLAGS_SHADER_SAVE;
-
-        if(!(flags & GSL_CONTEXT_NO_GMEM_ALLOC))
-            drawctxt->flags |= CTXT_FLAGS_GMEM_SHADOW;
+        // Save the shader instruction memory & GMEM on context switching
+        drawctxt->flags |= ( CTXT_FLAGS_SHADER_SAVE | CTXT_FLAGS_GMEM_SHADOW );
 
         // Clear out user defined GMEM shadow buffer structs
         kos_memset( drawctxt->user_gmem_shadow, 0, sizeof(gmem_shadow_t)*GSL_MAX_GMEM_SHADOW_BUFFERS );
@@ -1438,15 +1491,16 @@ kgsl_drawctxt_create(gsl_device_t* device, gsl_context_type_t type, unsigned int
         if (create_gmem_shadow(device, drawctxt, &ctx) != GSL_SUCCESS)
         {
             kgsl_drawctxt_destroy(device, index);
+            GSL_CONTEXT_MUTEX_UNLOCK();
             return (GSL_FAILURE);
         }
 
-
         KOS_ASSERT(ctx.cmd - ctx.start <= CMD_BUFFER_LEN);
     }
 
     *drawctxt_id = index;
 
+    GSL_CONTEXT_MUTEX_UNLOCK();
     return (GSL_SUCCESS);
 }
 
@@ -1460,6 +1514,8 @@ kgsl_drawctxt_destroy(gsl_device_t* device, unsigned int drawctxt_id)
 {
     gsl_drawctxt_t *drawctxt;
 
+    GSL_CONTEXT_MUTEX_LOCK();
+
     drawctxt = &device->drawctxt[drawctxt_id];
 
     if (drawctxt->flags != CTXT_FLAGS_NOT_IN_USE)
@@ -1494,6 +1550,8 @@ kgsl_drawctxt_destroy(gsl_device_t* device, unsigned int drawctxt_id)
         KOS_ASSERT(device->drawctxt_count >= 0);
     }
 
+    GSL_CONTEXT_MUTEX_UNLOCK();
+
     return (GSL_SUCCESS);
 }
 
@@ -1521,88 +1579,57 @@ kgsl_drawctxt_destroy(gsl_device_t* device, unsigned int drawctxt_id)
 //////////////////////////////////////////////////////////////////////////////
 KGSL_API int kgsl_drawctxt_bind_gmem_shadow(gsl_deviceid_t device_id, unsigned int drawctxt_id, const gsl_rect_t* gmem_rect, unsigned int shadow_x, unsigned int shadow_y, const gsl_buffer_desc_t* shadow_buffer, unsigned int buffer_id)
 {
-    gsl_device_t   *device;
-    gsl_drawctxt_t *drawctxt;
-    gmem_shadow_t  *shadow; // Shadow struct being modified
+    gsl_device_t   *device = &gsl_driver.device[device_id-1];
+    gsl_drawctxt_t *drawctxt = &device->drawctxt[drawctxt_id];
+    gmem_shadow_t  *shadow = &drawctxt->user_gmem_shadow[buffer_id];
     unsigned int    i;
 
     GSL_API_MUTEX_LOCK();
+    GSL_CONTEXT_MUTEX_LOCK();
 
-    device = &gsl_driver.device[device_id-1];       // device_id is 1 based
-
-    drawctxt = &device->drawctxt[drawctxt_id];
-
-    shadow = &drawctxt->user_gmem_shadow[buffer_id];
-
-    if( !shadow_buffer->enabled )
+	if( !shadow_buffer->enabled )
     {
         // Disable shadow
         shadow->gmemshadow.size = 0;
     }
     else
     {
-        // Binding to a buffer
-        unsigned int width, height, gmem_x, gmem_y, gmem_width, gmem_height, pixel_ratio;
+		// Sanity checks
+		KOS_ASSERT((gmem_rect->x % 2) == 0);      // Needs to be a multiple of 2
+		KOS_ASSERT((gmem_rect->y % 2) == 0);      // Needs to be a multiple of 2
+		KOS_ASSERT((gmem_rect->width % 2) == 0);  // Needs to be a multiple of 2
+		KOS_ASSERT((gmem_rect->height % 2) == 0); // Needs to be a multiple of 2
+		KOS_ASSERT((gmem_rect->pitch % 32) == 0); // Needs to be a multiple of 32
 
-        KOS_ASSERT(shadow_buffer->stride_bytes%4 == 0);
+		KOS_ASSERT((shadow_x % 2) == 0);  // Needs to be a multiple of 2
+		KOS_ASSERT((shadow_y % 2) == 0);  // Needs to be a multiple of 2
 
-        // Convert to 32bpp pixel units
-        if( shadow_buffer->bpp <= 32 )
-        {
-            KOS_ASSERT(32%shadow_buffer->bpp==0);
-            pixel_ratio = 32/shadow_buffer->bpp;
-            KOS_ASSERT(gmem_rect->x%pixel_ratio==0); // Needs to be at 32bit boundary
-            gmem_x = gmem_rect->x/pixel_ratio;
-            KOS_ASSERT(gmem_x%8==0); // Needs to be a multiple of 8
-            KOS_ASSERT(gmem_rect->y%pixel_ratio==0);  // Needs to be at 32bit boundary
-            gmem_y = gmem_rect->y/pixel_ratio;
-            KOS_ASSERT(gmem_y%8==0);  // Needs to be a multiple of 8
-            KOS_ASSERT(gmem_rect->width%pixel_ratio==0);  // Needs to be at 32bit boundary
-            gmem_width = gmem_rect->width/pixel_ratio;
-            KOS_ASSERT(gmem_rect->height%pixel_ratio==0);  // Needs to be at 32bit boundary
-            gmem_height = gmem_rect->height/pixel_ratio;
-            KOS_ASSERT(shadow_x%pixel_ratio==0);  // Needs to be at 32bit boundary
-            shadow_x = shadow_x/pixel_ratio;
-            KOS_ASSERT(shadow_x%8==0);  // Needs to be a multiple of 8
-            KOS_ASSERT(shadow_y%pixel_ratio==0);  // Needs to be at 32bit boundary
-            shadow_y = shadow_y/pixel_ratio;
-            KOS_ASSERT(shadow_y%8==0);  // Needs to be a multiple of 8
-        }
-        else
-        {
-            KOS_ASSERT(shadow_buffer->bpp==64 || shadow_buffer->bpp==128);
-            pixel_ratio = shadow_buffer->bpp/32;
-            gmem_x = gmem_rect->x*pixel_ratio;
-            KOS_ASSERT(gmem_x%8==0); // Needs to be a multiple of 8
-            gmem_y = gmem_rect->y*pixel_ratio;
-            KOS_ASSERT(gmem_y%8==0);  // Needs to be a multiple of 8
-            gmem_width = gmem_rect->width*pixel_ratio;
-            gmem_height = gmem_rect->height*pixel_ratio;
-            shadow_x = shadow_x*pixel_ratio;
-            KOS_ASSERT(shadow_x%8==0);  // Needs to be a multiple of 8
-            shadow_y = shadow_y*pixel_ratio;
-            KOS_ASSERT(shadow_y%8==0);  // Needs to be a multiple of 8
-        }
+		KOS_ASSERT(shadow_buffer->format >= COLORX_4_4_4_4);
+		KOS_ASSERT(shadow_buffer->format <= COLORX_32_32_32_32_FLOAT);
+		KOS_ASSERT((shadow_buffer->pitch % 32) == 0); // Needs to be a multiple of 32
+		KOS_ASSERT(buffer_id >= 0);
+		KOS_ASSERT(buffer_id < GSL_MAX_GMEM_SHADOW_BUFFERS);
 
-        KOS_ASSERT( buffer_id >= 0 && buffer_id < GSL_MAX_GMEM_SHADOW_BUFFERS );
+		// Set up GMEM shadow regions
+        kos_memcpy( &shadow->gmemshadow, &shadow_buffer->data, sizeof( gsl_memdesc_t ) );
+        shadow->size = shadow->gmemshadow.size;
 
-        width = gmem_width < drawctxt->context_gmem_shadow.width ? gmem_width : drawctxt->context_gmem_shadow.width;
-        height = gmem_height < drawctxt->context_gmem_shadow.height ? gmem_height : drawctxt->context_gmem_shadow.height;
+		shadow->width = shadow_buffer->width;
+		shadow->height = shadow_buffer->height;
+		shadow->pitch = shadow_buffer->pitch;
+		shadow->format = shadow_buffer->format;
 
-        drawctxt->user_gmem_shadow[buffer_id].width = width;
-        drawctxt->user_gmem_shadow[buffer_id].height = height;
-        drawctxt->user_gmem_shadow[buffer_id].pitch = shadow_buffer->stride_bytes/4;
+        shadow->offset = shadow->pitch * (shadow_y - gmem_rect->y) + shadow_x - gmem_rect->x;
 
-        kos_memcpy( &drawctxt->user_gmem_shadow[buffer_id].gmemshadow, &shadow_buffer->data, sizeof( gsl_memdesc_t ) );
-        // Calculate offset
-        drawctxt->user_gmem_shadow[buffer_id].offset = (int)shadow_buffer->stride_bytes/4*((int)shadow_y-(int)gmem_y)+(int)shadow_x-(int)gmem_x;
+        shadow->offset_x = shadow_x;
+        shadow->offset_y = shadow_y;
 
-        drawctxt->user_gmem_shadow[buffer_id].offset_x = shadow_x;
-        drawctxt->user_gmem_shadow[buffer_id].offset_y = shadow_y;
-        drawctxt->user_gmem_shadow[buffer_id].gmem_offset_x = gmem_x;
-        drawctxt->user_gmem_shadow[buffer_id].gmem_offset_y = gmem_y;
+		shadow->gmem_width = gmem_rect->width;
+		shadow->gmem_height = gmem_rect->height;
+		shadow->gmem_pitch = gmem_rect->pitch;
 
-        drawctxt->user_gmem_shadow[buffer_id].size = drawctxt->user_gmem_shadow[buffer_id].gmemshadow.size;
+        shadow->gmem_offset_x = gmem_rect->x;
+        shadow->gmem_offset_y = gmem_rect->y;
 
         // Modify quad vertices
         set_gmem_copy_quad(shadow);
@@ -1629,6 +1656,7 @@ KGSL_API int kgsl_drawctxt_bind_gmem_shadow(gsl_deviceid_t device_id, unsigned i
         }
     }
 
+    GSL_CONTEXT_MUTEX_UNLOCK();
     GSL_API_MUTEX_UNLOCK();
 
     return (GSL_SUCCESS);
@@ -1647,7 +1675,7 @@ kgsl_drawctxt_switch(gsl_device_t *device, gsl_drawctxt_t *drawctxt, gsl_flags_t
 
 	if (drawctxt != GSL_CONTEXT_NONE)
 	{
-		if(0) // flags & GSL_CONTEXT_SAVE_GMEM )
+		if( flags & GSL_CONTEXT_SAVE_GMEM )
 		{
 			// Set the flag in context so that the save is done when this context is switched out.
 			drawctxt->flags |= CTXT_FLAGS_GMEM_SAVE;
@@ -1766,6 +1794,8 @@ kgsl_drawctxt_destroyall(gsl_device_t *device)
     int             i;
     gsl_drawctxt_t  *drawctxt;
 
+    GSL_CONTEXT_MUTEX_LOCK();
+
     for (i = 0; i < GSL_CONTEXT_MAX; i++)
     {
         drawctxt = &device->drawctxt[i];
@@ -1790,6 +1820,8 @@ kgsl_drawctxt_destroyall(gsl_device_t *device)
         }
     }
 
+    GSL_CONTEXT_MUTEX_UNLOCK();
+
     return (GSL_SUCCESS);
 }
 
diff --git a/drivers/mxc/amd-gpu/common/gsl_g12.c b/drivers/mxc/amd-gpu/common/gsl_g12.c
index 14cfdb61b6a1..637e27c9e264 100644
--- a/drivers/mxc/amd-gpu/common/gsl_g12.c
+++ b/drivers/mxc/amd-gpu/common/gsl_g12.c
@@ -24,8 +24,11 @@
 #include <linux/sched.h>
 #endif
 
+#ifdef CONFIG_ARCH_MX35
+#define V3_SYNC
+#endif
+
 #ifdef GSL_BLD_G12
-#define GSL_TIMESTAMP_EPSILON   20000
 #define GSL_IRQ_TIMEOUT         200
 
 
@@ -230,6 +233,13 @@ static void kgsl_g12_updatetimestamp(gsl_device_t *device)
 	count >>= 8;
 	count &= 255;
 	device->timestamp += count;	
+#ifdef V3_SYNC
+	if (device->current_timestamp > device->timestamp)
+	{
+	    kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, 2);
+	    kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, 0);
+	}
+#endif
 	kgsl_sharedmem_write0(&device->memstore, GSL_DEVICE_MEMSTORE_OFFSET(eoptimestamp), &device->timestamp, 4, 0);
 }
 
@@ -353,7 +363,6 @@ kgsl_g12_close(gsl_device_t *device)
 
         kgsl_hal_setpowerstate(device->id, GSL_PWRFLAGS_POWER_OFF, 0);
 
-        device->ftbl.device_idle(device, GSL_TIMEOUT_NONE);
         device->flags &= ~GSL_FLAGS_INITIALIZED;
 
 #if defined(__SYMBIAN32__)
@@ -782,12 +791,20 @@ kgsl_g12_issueibcmds(gsl_device_t* device, int drawctxt_index, gpuaddr_t ibaddr,
     g_z1xx.curr = nextbuf;
 
     /* increment mark counter */
-    kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, flags);
-    kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, 0);
+#ifdef V3_SYNC
+    if (device->timestamp == device->current_timestamp)
+    {
+		kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, flags);
+		kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, 0);
+    }
+#else
+	kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, flags);
+	kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, 0);
+#endif
 
     /* increment consumed timestamp */
     device->current_timestamp++;
-
+    kgsl_sharedmem_write0(&device->memstore, GSL_DEVICE_MEMSTORE_OFFSET(soptimestamp), &device->current_timestamp, 4, 0);
     return (GSL_SUCCESS);
 }
 
@@ -947,12 +964,16 @@ static void irq_thread(void)
             /* Increase the timestamp value */
             timestamp += irq_count;
 
-            KOS_ASSERT( timestamp <= device->current_timestamp );
             /* Write the new timestamp value */
-#ifdef VG_HDK
             device->timestamp = timestamp;
-#else
             kgsl_sharedmem_write0(&device->memstore, GSL_DEVICE_MEMSTORE_OFFSET(eoptimestamp), &timestamp, 4, false);
+
+#ifdef V3_SYNC
+        	if (device->current_timestamp > device->timestamp)
+        	{
+        	    kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, 2);
+        	    kgsl_cmdwindow_write0(2, GSL_CMDWINDOW_2D, ADDR_VGV3_CONTROL, 0);
+        	}
 #endif
 
             /* Notify timestamp event */
diff --git a/drivers/mxc/amd-gpu/common/gsl_mmu.c b/drivers/mxc/amd-gpu/common/gsl_mmu.c
index 310677d926f0..1250cdb0208f 100644
--- a/drivers/mxc/amd-gpu/common/gsl_mmu.c
+++ b/drivers/mxc/amd-gpu/common/gsl_mmu.c
@@ -63,6 +63,18 @@ const unsigned int GSL_PT_PAGE_AP[4] = {(GSL_PT_PAGE_READ | GSL_PT_PAGE_WRITE),
 /////////////////////////////////////////////////////////////////////////////
 // macros
 //////////////////////////////////////////////////////////////////////////////
+#ifdef GSL_LOCKING_FINEGRAIN
+#define GSL_MMU_MUTEX_CREATE()              mmu->mutex = kos_mutex_create("gsl_mmu"); \
+                                            if (!mmu->mutex) {return (GSL_FAILURE);}
+#define GSL_MMU_LOCK()                      kos_mutex_lock(mmu->mutex)
+#define GSL_MMU_UNLOCK()                    kos_mutex_unlock(mmu->mutex)
+#define GSL_MMU_MUTEX_FREE()                kos_mutex_free(mmu->mutex); mmu->mutex = 0;
+#else
+#define GSL_MMU_MUTEX_CREATE()
+#define GSL_MMU_LOCK()
+#define GSL_MMU_UNLOCK()
+#define GSL_MMU_MUTEX_FREE()
+#endif
 
 #define GSL_PT_ENTRY_GET(va)                ((va - pagetable->va_base) >> GSL_PAGESIZE_SHIFT)
 #define GSL_PT_VIRT_GET(pte)                (pagetable->va_base + (pte * GSL_PAGESIZE))
@@ -385,6 +397,8 @@ kgsl_mmu_setpagetable(gsl_device_t *device, unsigned int pid)
     kgsl_log_write( KGSL_LOG_GROUP_MEMORY | KGSL_LOG_LEVEL_TRACE,
                     "--> gsl_pagetable_t* kgsl_mmu_setpagetable(gsl_device_t *device=0x%08x)\n", device );
 
+    GSL_MMU_LOCK();
+
     if (mmu->flags & GSL_FLAGS_STARTED)
     {
 #ifdef GSL_MMU_PAGETABLE_PERPROCESS
@@ -423,6 +437,8 @@ kgsl_mmu_setpagetable(gsl_device_t *device, unsigned int pid)
 		}
 	}
 
+    GSL_MMU_UNLOCK();
+
     kgsl_log_write( KGSL_LOG_GROUP_MEMORY | KGSL_LOG_LEVEL_TRACE, "<-- kgsl_mmu_setpagetable. Return value %B\n", status );
 
     return (status);
@@ -508,6 +524,8 @@ kgsl_mmu_init(gsl_device_t *device)
         // sub-client MMU lookups require address translation
         if ((mmu->config & ~0x1) > 0)
         {
+            GSL_MMU_MUTEX_CREATE();
+
             // make sure virtual address range is a multiple of 64Kb
             KOS_ASSERT((mmu->va_range & ((1 << 16)-1)) == 0);
 
@@ -600,9 +618,12 @@ kgsl_mmu_map(gsl_mmu_t *mmu, gpuaddr_t gpubaseaddr, const gsl_scatterlist_t *sca
     // get gpu access permissions
     ap = GSL_PT_PAGE_AP[((flags & GSL_MEMFLAGS_GPUAP_MASK) >> GSL_MEMFLAGS_GPUAP_SHIFT)];
 
+    GSL_MMU_LOCK();
+
     pagetable = kgsl_mmu_getpagetableobject(mmu, pid);
     if (!pagetable)
     {
+        GSL_MMU_UNLOCK();
         return (GSL_FAILURE);
     }
 
@@ -677,6 +698,8 @@ kgsl_mmu_map(gsl_mmu_t *mmu, gpuaddr_t gpubaseaddr, const gsl_scatterlist_t *sca
         status = GSL_FAILURE;
     }
 
+    GSL_MMU_UNLOCK();
+
     kgsl_log_write( KGSL_LOG_GROUP_MEMORY | KGSL_LOG_LEVEL_TRACE, "<-- kgsl_mmu_map. Return value %B\n", GSL_SUCCESS );
 
     return (status);
@@ -713,9 +736,12 @@ kgsl_mmu_unmap(gsl_mmu_t *mmu, gpuaddr_t gpubaseaddr, int range, unsigned int pi
         numpages++;
     }
 
+    GSL_MMU_LOCK();
+
     pagetable = kgsl_mmu_getpagetableobject(mmu, pid);
     if (!pagetable)
     {
+        GSL_MMU_UNLOCK();
         return (GSL_FAILURE);
     }
 
@@ -767,6 +793,8 @@ kgsl_mmu_unmap(gsl_mmu_t *mmu, gpuaddr_t gpubaseaddr, int range, unsigned int pi
     // invalidate tlb, debug only
 	KGSL_DEBUG(GSL_DBGFLAGS_MMU, mmu->device->ftbl.mmu_tlbinvalidate(mmu->device, gsl_cfg_mmu_reg[mmu->device->id-1].INVALIDATE, pagetable->pid));
 
+    GSL_MMU_UNLOCK();
+
     kgsl_log_write( KGSL_LOG_GROUP_MEMORY | KGSL_LOG_LEVEL_TRACE, "<-- kgsl_mmu_unmap. Return value %B\n", GSL_SUCCESS );
 
     return (status);
@@ -801,9 +829,12 @@ kgsl_mmu_getmap(gsl_mmu_t *mmu, gpuaddr_t gpubaseaddr, int range, gsl_scatterlis
         return (GSL_FAILURE);
     }
 
+    GSL_MMU_LOCK();
+
     pagetable = kgsl_mmu_getpagetableobject(mmu, pid);
     if (!pagetable)
     {
+        GSL_MMU_UNLOCK();
         return (GSL_FAILURE);
     }
 
@@ -837,6 +868,8 @@ kgsl_mmu_getmap(gsl_mmu_t *mmu, gpuaddr_t gpubaseaddr, int range, gsl_scatterlis
         scatterlist->pages[0] = GSL_PT_MAP_GETADDR(ptefirst);
     }
 
+    GSL_MMU_UNLOCK();
+
     scatterlist->contiguous = contiguous;
 
     return (GSL_SUCCESS);
@@ -906,6 +939,8 @@ kgsl_mmu_close(gsl_device_t *device)
                 kgsl_sharedmem_free0(&mmu->dummyspace, GSL_CALLER_PROCESSID_GET());
             }
 
+            GSL_MMU_MUTEX_FREE();
+
             mmu->flags &= ~GSL_FLAGS_STARTED;
             mmu->flags &= ~GSL_FLAGS_INITIALIZED;
             mmu->flags &= ~GSL_FLAGS_INITIALIZED0;
@@ -932,6 +967,8 @@ kgsl_mmu_attachcallback(gsl_mmu_t *mmu, unsigned int pid)
 
     kgsl_log_write( KGSL_LOG_GROUP_MEMORY | KGSL_LOG_LEVEL_TRACE, "--> int kgsl_mmu_attachcallback(gsl_mmu_t *mmu=0x%08x, unsigned int pid=0x%08x)\n", mmu, pid );
 
+    GSL_MMU_LOCK();
+
     if (mmu->flags & GSL_FLAGS_INITIALIZED0)
     {
         // attach to current device mmu
@@ -952,6 +989,8 @@ kgsl_mmu_attachcallback(gsl_mmu_t *mmu, unsigned int pid)
         }
     }
 
+    GSL_MMU_UNLOCK();
+
     kgsl_log_write( KGSL_LOG_GROUP_MEMORY | KGSL_LOG_LEVEL_TRACE, "<-- kgsl_mmu_attachcallback. Return value %B\n", status );
 
     return (status);
@@ -970,6 +1009,8 @@ kgsl_mmu_detachcallback(gsl_mmu_t *mmu, unsigned int pid)
 
     kgsl_log_write( KGSL_LOG_GROUP_MEMORY | KGSL_LOG_LEVEL_TRACE, "--> int kgsl_mmu_detachcallback(gsl_mmu_t *mmu=0x%08x, unsigned int pid=0x%08x)\n", mmu, pid );
 
+    GSL_MMU_LOCK();
+
     if (mmu->flags & GSL_FLAGS_INITIALIZED0)
     {
         // detach from current device mmu
@@ -990,6 +1031,8 @@ kgsl_mmu_detachcallback(gsl_mmu_t *mmu, unsigned int pid)
         }
     }
 
+    GSL_MMU_UNLOCK();
+
     kgsl_log_write( KGSL_LOG_GROUP_MEMORY | KGSL_LOG_LEVEL_TRACE, "<-- kgsl_mmu_detachcallback. Return value %B\n", status );
 
     return (status);
@@ -1005,6 +1048,8 @@ kgsl_mmu_querystats(gsl_mmu_t *mmu, gsl_mmustats_t *stats)
 
     KOS_ASSERT(stats);
 
+    GSL_MMU_LOCK();
+
     if (mmu->flags & GSL_FLAGS_STARTED)
     {
 		kos_memcpy(stats, &mmu->stats, sizeof(gsl_mmustats_t));
@@ -1014,6 +1059,8 @@ kgsl_mmu_querystats(gsl_mmu_t *mmu, gsl_mmustats_t *stats)
 		kos_memset(stats, 0, sizeof(gsl_mmustats_t));
     }
 
+    GSL_MMU_UNLOCK();
+
     return (status);
 #else
     // unreferenced formal parameters
diff --git a/drivers/mxc/amd-gpu/common/gsl_ringbuffer.c b/drivers/mxc/amd-gpu/common/gsl_ringbuffer.c
index c4b62b0174d2..ed922ef5bc26 100644
--- a/drivers/mxc/amd-gpu/common/gsl_ringbuffer.c
+++ b/drivers/mxc/amd-gpu/common/gsl_ringbuffer.c
@@ -120,6 +120,8 @@ kgsl_ringbuffer_watchdog()
 
     if (rb->flags & GSL_FLAGS_STARTED)
     {
+        GSL_RB_MUTEX_LOCK();
+
         GSL_RB_GET_READPTR(rb, &rb->rptr);
 
         // ringbuffer is currently not empty
@@ -150,6 +152,7 @@ kgsl_ringbuffer_watchdog()
             rb->watchdog.flags &= ~GSL_FLAGS_ACTIVE;
         }
 
+        GSL_RB_MUTEX_UNLOCK();
     }
 
     kgsl_log_write( KGSL_LOG_GROUP_COMMAND | KGSL_LOG_LEVEL_TRACE, "<-- kgsl_ringbuffer_watchdog.\n" );
@@ -675,6 +678,8 @@ kgsl_ringbuffer_init(gsl_device_t *device)
     rb->sizedwords       = (2 << gsl_cfg_rb_sizelog2quadwords);
     rb->blksizequadwords = gsl_cfg_rb_blksizequadwords;
 
+    GSL_RB_MUTEX_CREATE();
+
     // allocate memory for ringbuffer, needs to be double octword aligned
     // align on page from contiguous physical memory
     flags = (GSL_MEMFLAGS_ALIGNPAGE | GSL_MEMFLAGS_CONPHYS | GSL_MEMFLAGS_STRICTREQUEST);
@@ -737,6 +742,8 @@ kgsl_ringbuffer_close(gsl_ringbuffer_t *rb)
     kgsl_log_write( KGSL_LOG_GROUP_COMMAND | KGSL_LOG_LEVEL_TRACE,
                     "--> int kgsl_ringbuffer_close(gsl_ringbuffer_t *rb=0x%08x)\n", rb );
 
+    GSL_RB_MUTEX_LOCK();
+
     // stop ringbuffer
     kgsl_ringbuffer_stop(rb);
 
@@ -754,6 +761,10 @@ kgsl_ringbuffer_close(gsl_ringbuffer_t *rb)
 
     rb->flags &= ~GSL_FLAGS_INITIALIZED;
 
+    GSL_RB_MUTEX_UNLOCK();
+
+    GSL_RB_MUTEX_FREE();
+
     kos_memset(rb, 0, sizeof(gsl_ringbuffer_t));
 
     kgsl_log_write( KGSL_LOG_GROUP_COMMAND | KGSL_LOG_LEVEL_TRACE, "<-- kgsl_ringbuffer_close. Return value %B\n", GSL_SUCCESS );
@@ -874,6 +885,8 @@ kgsl_ringbuffer_issueibcmds(gsl_device_t *device, int drawctxt_index, gpuaddr_t
 
     KGSL_DEBUG(GSL_DBGFLAGS_DUMPX, dumpx_swap = kgsl_dumpx_parse_ibs(ibaddr, sizedwords));
 
+	GSL_RB_MUTEX_LOCK();
+
 	// context switch if needed
 	kgsl_drawctxt_switch(device, &device->drawctxt[drawctxt_index], flags);
 
@@ -883,6 +896,8 @@ kgsl_ringbuffer_issueibcmds(gsl_device_t *device, int drawctxt_index, gpuaddr_t
 
 	*timestamp = kgsl_ringbuffer_issuecmds(device, 0, &link[0], 3, GSL_CALLER_PROCESSID_GET());
 
+	GSL_RB_MUTEX_UNLOCK();
+
     // idle device when running in safe mode
     if (device->flags & GSL_FLAGS_SAFEMODE)
     {
diff --git a/drivers/mxc/amd-gpu/common/gsl_yamato.c b/drivers/mxc/amd-gpu/common/gsl_yamato.c
index e52d4274c6a6..658cd80f3629 100644
--- a/drivers/mxc/amd-gpu/common/gsl_yamato.c
+++ b/drivers/mxc/amd-gpu/common/gsl_yamato.c
@@ -488,6 +488,13 @@ kgsl_yamato_start(gsl_device_t *device, gsl_flags_t flags)
 int
 kgsl_yamato_stop(gsl_device_t *device)
 {
+	// HW WORKAROUND: Ringbuffer hangs during next start if it is stopped without any
+	// commands ever being submitted. To avoid this, submit a dummy wait packet.
+	unsigned int cmds[2];
+    cmds[0] = pm4_type3_packet(PM4_WAIT_FOR_IDLE, 1);
+    cmds[0] = 0;
+	kgsl_ringbuffer_issuecmds(device, 0, cmds, 2, GSL_CALLER_PROCESSID_GET());
+
     // disable rbbm interrupts
     kgsl_intr_detach(&device->intr, GSL_INTR_YDX_RBBM_READ_ERROR);
     kgsl_intr_detach(&device->intr, GSL_INTR_YDX_RBBM_DISPLAY_UPDATE);
@@ -732,6 +739,8 @@ kgsl_yamato_idle(gsl_device_t *device, unsigned int timeout)
 
     KGSL_DEBUG(GSL_DBGFLAGS_DUMPX, KGSL_DEBUG_DUMPX(BB_DUMP_REGPOLL, device->id, mmRBBM_STATUS, 0x80000000, "kgsl_yamato_idle"));
 
+    GSL_RB_MUTEX_LOCK();
+
     // first, wait until the CP has consumed all the commands in the ring buffer
     if (rb->flags & GSL_FLAGS_STARTED)
     {
@@ -754,6 +763,8 @@ kgsl_yamato_idle(gsl_device_t *device, unsigned int timeout)
         }
 
     }
+	
+    GSL_RB_MUTEX_UNLOCK();
 
     return (status);
 }
diff --git a/drivers/mxc/amd-gpu/include/api/gsl_klibapi.h b/drivers/mxc/amd-gpu/include/api/gsl_klibapi.h
index 8476f5a95969..3c08545c2b68 100644
--- a/drivers/mxc/amd-gpu/include/api/gsl_klibapi.h
+++ b/drivers/mxc/amd-gpu/include/api/gsl_klibapi.h
@@ -74,6 +74,7 @@ KGSL_API int                kgsl_driver_destroy(unsigned int pid);
 KGSL_API int                kgsl_device_start(gsl_deviceid_t device_id, gsl_flags_t flags);
 KGSL_API int                kgsl_device_stop(gsl_deviceid_t device_id);
 KGSL_API int                kgsl_device_idle(gsl_deviceid_t device_id, unsigned int timeout);
+KGSL_API int                kgsl_device_isidle(gsl_deviceid_t device_id);
 KGSL_API int                kgsl_device_getproperty(gsl_deviceid_t device_id, gsl_property_type_t type, void *value, unsigned int sizebytes);
 KGSL_API int                kgsl_device_setproperty(gsl_deviceid_t device_id, gsl_property_type_t type, void *value, unsigned int sizebytes);
 KGSL_API int                kgsl_device_regread(gsl_deviceid_t device_id, unsigned int offsetwords, unsigned int *value);
diff --git a/drivers/mxc/amd-gpu/include/api/gsl_libapi.h b/drivers/mxc/amd-gpu/include/api/gsl_libapi.h
index 7a5be862f3ee..3d359e24f572 100644
--- a/drivers/mxc/amd-gpu/include/api/gsl_libapi.h
+++ b/drivers/mxc/amd-gpu/include/api/gsl_libapi.h
@@ -65,6 +65,7 @@ GSL_API int                     gsl_library_close(void);
 GSL_API gsl_devhandle_t         gsl_device_open(gsl_deviceid_t device_id, gsl_flags_t flags);
 GSL_API int                     gsl_device_close(gsl_devhandle_t devhandle);
 GSL_API int                     gsl_device_idle(gsl_devhandle_t devhandle, unsigned int timeout);
+GSL_API int                     gsl_device_isidle(gsl_devhandle_t devhandle);
 GSL_API int                     gsl_device_getcount(void);
 GSL_API int                     gsl_device_getinfo(gsl_devhandle_t devhandle, gsl_devinfo_t *devinfo);
 GSL_API int                     gsl_device_setpowerstate(gsl_devhandle_t devhandle, gsl_flags_t flags);
diff --git a/drivers/mxc/amd-gpu/include/api/gsl_types.h b/drivers/mxc/amd-gpu/include/api/gsl_types.h
index 310c1a9f5d00..99d984966113 100644
--- a/drivers/mxc/amd-gpu/include/api/gsl_types.h
+++ b/drivers/mxc/amd-gpu/include/api/gsl_types.h
@@ -188,13 +188,10 @@
 #define GSL_TIMEOUT_NONE                        0
 #define GSL_TIMEOUT_DEFAULT                     0xFFFFFFFF
 
-#ifdef _LINUX
-#define GSL_PAGESIZE                            PAGE_SIZE
-#define GSL_PAGESIZE_SHIFT                      PAGE_SHIFT
-#else
 #define GSL_PAGESIZE                            0x1000
 #define GSL_PAGESIZE_SHIFT                      12
-#endif
+
+#define GSL_TIMESTAMP_EPSILON           20000
 
 //////////////////////////////////////////////////////////////////////////////
 // types
@@ -407,6 +404,7 @@ typedef struct _gsl_rect_t {
     unsigned int y;
     unsigned int width;
     unsigned int height;
+	unsigned int pitch;
 } gsl_rect_t;
 
 // -----------------------
@@ -414,8 +412,10 @@ typedef struct _gsl_rect_t {
 // -----------------------
 typedef struct _gsl_buffer_desc_t {
     gsl_memdesc_t data;
-    unsigned int stride_bytes;
-    unsigned int bpp;
+	unsigned int width;
+	unsigned int height;
+	unsigned int pitch;
+	unsigned int format;
     unsigned int enabled;
 } gsl_buffer_desc_t;
 
diff --git a/drivers/mxc/amd-gpu/include/gsl_buildconfig.h b/drivers/mxc/amd-gpu/include/gsl_buildconfig.h
index 01ba14e4cb66..4e6be4da7dc4 100644
--- a/drivers/mxc/amd-gpu/include/gsl_buildconfig.h
+++ b/drivers/mxc/amd-gpu/include/gsl_buildconfig.h
@@ -36,7 +36,7 @@
 #define GSL_BLD_YAMATO
 #define GSL_BLD_G12
 
-#define GSL_LOCKING_COURSEGRAIN
+#define GSL_LOCKING_COARSEGRAIN
 
 #define GSL_STATS_MEM
 #define GSL_STATS_RINGBUFFER
diff --git a/drivers/mxc/amd-gpu/include/gsl_device.h b/drivers/mxc/amd-gpu/include/gsl_device.h
index 433dc6963dfd..087dcc1acabf 100644
--- a/drivers/mxc/amd-gpu/include/gsl_device.h
+++ b/drivers/mxc/amd-gpu/include/gsl_device.h
@@ -91,12 +91,18 @@ struct _gsl_device_t {
 #ifdef GSL_BLD_YAMATO
     gsl_memregion_t   gmemspace;
     gsl_ringbuffer_t  ringbuffer;
+#ifdef GSL_LOCKING_FINEGRAIN
+    oshandle_t        drawctxt_mutex;
+#endif
     unsigned int      drawctxt_count;
     gsl_drawctxt_t    *drawctxt_active;
     gsl_drawctxt_t    drawctxt[GSL_CONTEXT_MAX];
 #endif // GSL_BLD_YAMATO
 
 #ifdef GSL_BLD_G12
+#ifdef GSL_LOCKING_FINEGRAIN
+    oshandle_t        cmdwindow_mutex;
+#endif
     unsigned int      intrcnt[GSL_G12_INTR_COUNT];
     gsl_timestamp_t   current_timestamp;
     gsl_timestamp_t   timestamp;
@@ -108,6 +114,9 @@ struct _gsl_device_t {
     oshandle_t        irqthread_event;
 #endif
 #endif // GSL_BLD_G12
+#ifdef GSL_LOCKING_FINEGRAIN
+    oshandle_t        cmdstream_mutex;
+#endif
 #ifndef _LINUX	
     oshandle_t        timestamp_event;
 #else
diff --git a/drivers/mxc/amd-gpu/include/gsl_drawctxt.h b/drivers/mxc/amd-gpu/include/gsl_drawctxt.h
index 15b8097828af..f3bc8c37f4f2 100644
--- a/drivers/mxc/amd-gpu/include/gsl_drawctxt.h
+++ b/drivers/mxc/amd-gpu/include/gsl_drawctxt.h
@@ -59,13 +59,21 @@ typedef struct _gmem_shadow_t
 
     // 256 KB GMEM surface = 4 bytes-per-pixel x 256 pixels/row x 256 rows.
     // width & height must be a multiples of 32, in case tiled textures are used.
-    unsigned int    size;       // Size of surface used to store GMEM
+    unsigned int    size;           // Size of surface used to store GMEM
     unsigned int    width;          // Width of surface used to store GMEM
     unsigned int    height;         // Height of surface used to store GMEM
     unsigned int    pitch;          // Pitch of surface used to store GMEM
+	unsigned int    format;         // Format of surface used to store GMEM
+
     int             offset;
-    unsigned int    offset_x;
-    unsigned int    offset_y;
+
+    unsigned int    offset_x;    
+	unsigned int    offset_y;
+
+	unsigned int	gmem_width;     // GMEM width
+	unsigned int	gmem_height;    // GMEM height
+	unsigned int    gmem_pitch;     // GMEM pitch
+
     unsigned int    gmem_offset_x;
     unsigned int    gmem_offset_y;
 
diff --git a/drivers/mxc/amd-gpu/include/gsl_driver.h b/drivers/mxc/amd-gpu/include/gsl_driver.h
index 42dff457dc49..9c908ce4696e 100644
--- a/drivers/mxc/amd-gpu/include/gsl_driver.h
+++ b/drivers/mxc/amd-gpu/include/gsl_driver.h
@@ -39,7 +39,7 @@
 #define GSL_CALLER_PROCESSID_GET()      kos_process_getid()
 #endif // GSL_DEDICATED_PROCESS
 
-#ifdef GSL_LOCKING_COURSEGRAIN
+#ifdef GSL_LOCKING_COARSEGRAIN
 #define GSL_API_MUTEX_CREATE()          gsl_driver.mutex = kos_mutex_create("gsl_global"); \
                                         if (!gsl_driver.mutex) {return (GSL_FAILURE);}
 #define GSL_API_MUTEX_LOCK()            kos_mutex_lock(gsl_driver.mutex)
diff --git a/drivers/mxc/amd-gpu/include/gsl_ioctl.h b/drivers/mxc/amd-gpu/include/gsl_ioctl.h
index 0f1983e4c770..6a06f3e0aec4 100644
--- a/drivers/mxc/amd-gpu/include/gsl_ioctl.h
+++ b/drivers/mxc/amd-gpu/include/gsl_ioctl.h
@@ -50,6 +50,10 @@ typedef struct _kgsl_device_idle_t {
     unsigned int    timeout;
 } kgsl_device_idle_t;
 
+typedef struct _kgsl_device_isidle_t {
+    gsl_deviceid_t  device_id;
+} kgsl_device_isidle_t;
+
 typedef struct _kgsl_device_getproperty_t {
     gsl_deviceid_t  device_id;
     gsl_property_type_t type;
@@ -209,29 +213,30 @@ typedef struct _kgsl_device_clock_t {
 #define IOCTL_KGSL_DEVICE_START                 _IOW(GSL_MAGIC, 0x20, struct _kgsl_device_start_t)
 #define IOCTL_KGSL_DEVICE_STOP                  _IOW(GSL_MAGIC, 0x21, struct _kgsl_device_stop_t)
 #define IOCTL_KGSL_DEVICE_IDLE                  _IOW(GSL_MAGIC, 0x22, struct _kgsl_device_idle_t)
-#define IOCTL_KGSL_DEVICE_GETPROPERTY           _IOWR(GSL_MAGIC, 0x23, struct _kgsl_device_getproperty_t)
-#define IOCTL_KGSL_DEVICE_SETPROPERTY           _IOW(GSL_MAGIC, 0x24, struct _kgsl_device_setproperty_t)
-#define IOCTL_KGSL_DEVICE_REGREAD               _IOWR(GSL_MAGIC, 0x25, struct _kgsl_device_regread_t)
-#define IOCTL_KGSL_DEVICE_REGWRITE              _IOW(GSL_MAGIC, 0x26, struct _kgsl_device_regwrite_t)
-#define IOCTL_KGSL_DEVICE_WAITIRQ               _IOWR(GSL_MAGIC, 0x27, struct _kgsl_device_waitirq_t)
-#define IOCTL_KGSL_CMDSTREAM_ISSUEIBCMDS        _IOWR(GSL_MAGIC, 0x28, struct _kgsl_cmdstream_issueibcmds_t)
-#define IOCTL_KGSL_CMDSTREAM_READTIMESTAMP      _IOWR(GSL_MAGIC, 0x29, struct _kgsl_cmdstream_readtimestamp_t)
-#define IOCTL_KGSL_CMDSTREAM_FREEMEMONTIMESTAMP _IOW(GSL_MAGIC, 0x2A, struct _kgsl_cmdstream_freememontimestamp_t)
-#define IOCTL_KGSL_CMDSTREAM_WAITTIMESTAMP      _IOW(GSL_MAGIC, 0x2B, struct _kgsl_cmdstream_waittimestamp_t)
-#define IOCTL_KGSL_CMDWINDOW_WRITE              _IOW(GSL_MAGIC, 0x2C, struct _kgsl_cmdwindow_write_t)
-#define IOCTL_KGSL_CONTEXT_CREATE               _IOWR(GSL_MAGIC, 0x2D, struct _kgsl_context_create_t)
-#define IOCTL_KGSL_CONTEXT_DESTROY              _IOW(GSL_MAGIC, 0x2E, struct _kgsl_context_destroy_t)
-#define IOCTL_KGSL_DRAWCTXT_BIND_GMEM_SHADOW    _IOW(GSL_MAGIC, 0x2F, struct _kgsl_drawctxt_bind_gmem_shadow_t)
-#define IOCTL_KGSL_SHAREDMEM_ALLOC              _IOWR(GSL_MAGIC, 0x30, struct _kgsl_sharedmem_alloc_t)
-#define IOCTL_KGSL_SHAREDMEM_FREE               _IOW(GSL_MAGIC, 0x31, struct _kgsl_sharedmem_free_t)
-#define IOCTL_KGSL_SHAREDMEM_READ               _IOWR(GSL_MAGIC, 0x32, struct _kgsl_sharedmem_read_t)
-#define IOCTL_KGSL_SHAREDMEM_WRITE              _IOW(GSL_MAGIC, 0x33, struct _kgsl_sharedmem_write_t)
-#define IOCTL_KGSL_SHAREDMEM_SET                _IOW(GSL_MAGIC, 0x34, struct _kgsl_sharedmem_set_t)
-#define IOCTL_KGSL_SHAREDMEM_LARGESTFREEBLOCK   _IOWR(GSL_MAGIC, 0x35, struct _kgsl_sharedmem_largestfreeblock_t)
-#define IOCTL_KGSL_SHAREDMEM_CACHEOPERATION     _IOW(GSL_MAGIC, 0x36, struct _kgsl_sharedmem_cacheoperation_t)
-#define IOCTL_KGSL_SHAREDMEM_FROMHOSTPOINTER    _IOW(GSL_MAGIC, 0x37, struct _kgsl_sharedmem_fromhostpointer_t)
-#define IOCTL_KGSL_ADD_TIMESTAMP                _IOWR(GSL_MAGIC, 0x38, struct _kgsl_add_timestamp_t)
-#define IOCTL_KGSL_DRIVER_EXIT		        _IOWR(GSL_MAGIC, 0x39, NULL)
+#define IOCTL_KGSL_DEVICE_ISIDLE                _IOR(GSL_MAGIC, 0x23, struct _kgsl_device_isidle_t)
+#define IOCTL_KGSL_DEVICE_GETPROPERTY           _IOWR(GSL_MAGIC, 0x24, struct _kgsl_device_getproperty_t)
+#define IOCTL_KGSL_DEVICE_SETPROPERTY           _IOW(GSL_MAGIC, 0x25, struct _kgsl_device_setproperty_t)
+#define IOCTL_KGSL_DEVICE_REGREAD               _IOWR(GSL_MAGIC, 0x26, struct _kgsl_device_regread_t)
+#define IOCTL_KGSL_DEVICE_REGWRITE              _IOW(GSL_MAGIC, 0x27, struct _kgsl_device_regwrite_t)
+#define IOCTL_KGSL_DEVICE_WAITIRQ               _IOWR(GSL_MAGIC, 0x28, struct _kgsl_device_waitirq_t)
+#define IOCTL_KGSL_CMDSTREAM_ISSUEIBCMDS        _IOWR(GSL_MAGIC, 0x29, struct _kgsl_cmdstream_issueibcmds_t)
+#define IOCTL_KGSL_CMDSTREAM_READTIMESTAMP      _IOWR(GSL_MAGIC, 0x2A, struct _kgsl_cmdstream_readtimestamp_t)
+#define IOCTL_KGSL_CMDSTREAM_FREEMEMONTIMESTAMP _IOW(GSL_MAGIC, 0x2B, struct _kgsl_cmdstream_freememontimestamp_t)
+#define IOCTL_KGSL_CMDSTREAM_WAITTIMESTAMP      _IOW(GSL_MAGIC, 0x2C, struct _kgsl_cmdstream_waittimestamp_t)
+#define IOCTL_KGSL_CMDWINDOW_WRITE              _IOW(GSL_MAGIC, 0x2D, struct _kgsl_cmdwindow_write_t)
+#define IOCTL_KGSL_CONTEXT_CREATE               _IOWR(GSL_MAGIC, 0x2E, struct _kgsl_context_create_t)
+#define IOCTL_KGSL_CONTEXT_DESTROY              _IOW(GSL_MAGIC, 0x2F, struct _kgsl_context_destroy_t)
+#define IOCTL_KGSL_DRAWCTXT_BIND_GMEM_SHADOW    _IOW(GSL_MAGIC, 0x30, struct _kgsl_drawctxt_bind_gmem_shadow_t)
+#define IOCTL_KGSL_SHAREDMEM_ALLOC              _IOWR(GSL_MAGIC, 0x31, struct _kgsl_sharedmem_alloc_t)
+#define IOCTL_KGSL_SHAREDMEM_FREE               _IOW(GSL_MAGIC, 0x32, struct _kgsl_sharedmem_free_t)
+#define IOCTL_KGSL_SHAREDMEM_READ               _IOWR(GSL_MAGIC, 0x33, struct _kgsl_sharedmem_read_t)
+#define IOCTL_KGSL_SHAREDMEM_WRITE              _IOW(GSL_MAGIC, 0x34, struct _kgsl_sharedmem_write_t)
+#define IOCTL_KGSL_SHAREDMEM_SET                _IOW(GSL_MAGIC, 0x35, struct _kgsl_sharedmem_set_t)
+#define IOCTL_KGSL_SHAREDMEM_LARGESTFREEBLOCK   _IOWR(GSL_MAGIC, 0x36, struct _kgsl_sharedmem_largestfreeblock_t)
+#define IOCTL_KGSL_SHAREDMEM_CACHEOPERATION     _IOW(GSL_MAGIC, 0x37, struct _kgsl_sharedmem_cacheoperation_t)
+#define IOCTL_KGSL_SHAREDMEM_FROMHOSTPOINTER    _IOW(GSL_MAGIC, 0x38, struct _kgsl_sharedmem_fromhostpointer_t)
+#define IOCTL_KGSL_ADD_TIMESTAMP                _IOWR(GSL_MAGIC, 0x39, struct _kgsl_add_timestamp_t)
+#define IOCTL_KGSL_DRIVER_EXIT		        _IOWR(GSL_MAGIC, 0x3A, NULL)
 #define IOCTL_KGSL_DEVICE_CLOCK			_IOWR(GSL_MAGIC, 0x60, struct _kgsl_device_clock_t)
 
 
diff --git a/drivers/mxc/amd-gpu/include/gsl_mmu.h b/drivers/mxc/amd-gpu/include/gsl_mmu.h
index 868c5156f290..ddb2243b58d5 100644
--- a/drivers/mxc/amd-gpu/include/gsl_mmu.h
+++ b/drivers/mxc/amd-gpu/include/gsl_mmu.h
@@ -135,6 +135,9 @@ typedef struct _gsl_tlbflushfilter_t {
 // mmu object
 // ----------
 typedef struct _gsl_mmu_t {
+#ifdef GSL_LOCKING_FINEGRAIN
+    oshandle_t            mutex;
+#endif
     unsigned int          refcnt;
     gsl_flags_t           flags;
     gsl_device_t          *device;
diff --git a/drivers/mxc/amd-gpu/include/gsl_ringbuffer.h b/drivers/mxc/amd-gpu/include/gsl_ringbuffer.h
index 6081c396f6e4..57f6297735e2 100644
--- a/drivers/mxc/amd-gpu/include/gsl_ringbuffer.h
+++ b/drivers/mxc/amd-gpu/include/gsl_ringbuffer.h
@@ -136,7 +136,9 @@ typedef struct _gsl_ringbuffer_t {
 
     gsl_device_t      *device;
     gsl_flags_t       flags;
-
+#ifdef GSL_LOCKING_FINEGRAIN
+    oshandle_t        mutex;
+#endif
     gsl_memdesc_t     buffer_desc;              // allocated memory descriptor
     gsl_memdesc_t     memptrs_desc;
 
@@ -163,6 +165,19 @@ typedef struct _gsl_ringbuffer_t {
 // macros
 //////////////////////////////////////////////////////////////////////////////
 
+#ifdef GSL_LOCKING_FINEGRAIN
+#define GSL_RB_MUTEX_CREATE()               rb->mutex = kos_mutex_create("gsl_ringbuffer"); \
+                                            if (!rb->mutex) {return (GSL_FAILURE);}
+#define GSL_RB_MUTEX_LOCK()                 kos_mutex_lock(rb->mutex)
+#define GSL_RB_MUTEX_UNLOCK()               kos_mutex_unlock(rb->mutex)
+#define GSL_RB_MUTEX_FREE()                 kos_mutex_free(rb->mutex); rb->mutex = 0;
+#else
+#define GSL_RB_MUTEX_CREATE()
+#define GSL_RB_MUTEX_LOCK()
+#define GSL_RB_MUTEX_UNLOCK()
+#define GSL_RB_MUTEX_FREE()
+#endif
+
 // ----------
 // ring write
 // ----------
diff --git a/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_mask.h b/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_mask.h
index f31b2a74d1fa..c3087908c541 100644
--- a/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_mask.h
+++ b/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_mask.h
@@ -28,20 +28,6 @@
 
 #if !defined (_yamato_MASK_HEADER)
 #define _yamato_MASK_HEADER
-/*
-*   yamato_mask.h
-*
-*   Register Spec Release:  Chip Spec 1.0
-*
-*
-*    (c) 2000 ATI Technologies Inc.  (unpublished)
-*
-*    All rights reserved.  This notice is intended as a precaution against
-*    inadvertent publication and does not imply publication or any waiver
-*    of confidentiality.  The year included in the foregoing notice is the
-*    year of creation of the work.
-*
-*/
 
 // PA_CL_VPORT_XSCALE
 #define PA_CL_VPORT_XSCALE__VPORT_XSCALE_MASK              0xffffffffL
diff --git a/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_registers.h b/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_registers.h
index 3cd315f903db..fc6b8b98f849 100644
--- a/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_registers.h
+++ b/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_registers.h
@@ -28,20 +28,6 @@
 
 #if !defined (_yamato_REG_HEADER)
 #define _yamato_REG_HEADER
-/*
-*   yamato_registers.h
-*
-*   Register Spec Release:  Chip Spec 1.0
-*
-*
-*    (c) 2000 ATI Technologies Inc.  (unpublished)
-*
-*    All rights reserved.  This notice is intended as a precaution against
-*    inadvertent publication and does not imply publication or any waiver
-*    of confidentiality.  The year included in the foregoing notice is the
-*    year of creation of the work.
-*
-*/
 
     union PA_CL_VPORT_XSCALE {
     struct {
diff --git a/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_struct.h b/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_struct.h
index 9e9c7282dcdb..d6cc2fe9abdf 100644
--- a/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_struct.h
+++ b/drivers/mxc/amd-gpu/include/reg/yamato/10/yamato_struct.h
@@ -7847,21 +7847,6 @@ typedef union {
 #if !defined (_RBBM_FIDDLE_H)
 #define _RBBM_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  rbbm_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -10977,21 +10962,6 @@ typedef union {
 #if !defined (_MH_FIDDLE_H)
 #define _MH_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  mh_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -19932,21 +19902,6 @@ typedef union {
 #if !defined (_PA_FIDDLE_H)
 #define _PA_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  pa_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -27841,21 +27796,6 @@ typedef union {
 #if !defined (_VGT_FIDDLE_H)
 #define _VGT_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  vgt_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -32985,21 +32925,6 @@ typedef union {
 #if !defined (_SQ_FIDDLE_H)
 #define _SQ_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  sq_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -42029,21 +41954,6 @@ typedef union {
 #if !defined (_SX_FIDDLE_H)
 #define _SX_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  sx_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -42064,21 +41974,6 @@ typedef union {
 #if !defined (_TP_FIDDLE_H)
 #define _TP_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  tp_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -48231,21 +48126,6 @@ typedef union {
 #if !defined (_TC_FIDDLE_H)
 #define _TC_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  tc_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -48266,21 +48146,6 @@ typedef union {
 #if !defined (_SC_FIDDLE_H)
 #define _SC_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  sc_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -48301,21 +48166,6 @@ typedef union {
 #if !defined (_BC_FIDDLE_H)
 #define _BC_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  bc_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
diff --git a/drivers/mxc/amd-gpu/include/reg/yamato/14/yamato_struct.h b/drivers/mxc/amd-gpu/include/reg/yamato/14/yamato_struct.h
index 80b9106759e3..e8402cda4463 100644
--- a/drivers/mxc/amd-gpu/include/reg/yamato/14/yamato_struct.h
+++ b/drivers/mxc/amd-gpu/include/reg/yamato/14/yamato_struct.h
@@ -7826,21 +7826,6 @@ typedef union {
 #if !defined (_RBBM_FIDDLE_H)
 #define _RBBM_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  rbbm_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -10526,21 +10511,6 @@ typedef union {
 #if !defined (_MH_FIDDLE_H)
 #define _MH_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  mh_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -19151,21 +19121,6 @@ typedef union {
 #if !defined (_PA_FIDDLE_H)
 #define _PA_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  pa_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -26983,21 +26938,6 @@ typedef union {
 #if !defined (_VGT_FIDDLE_H)
 #define _VGT_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  vgt_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -32094,21 +32034,6 @@ typedef union {
 #if !defined (_SQ_FIDDLE_H)
 #define _SQ_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  sq_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -40867,21 +40792,6 @@ typedef union {
 #if !defined (_SX_FIDDLE_H)
 #define _SX_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  sx_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -40902,21 +40812,6 @@ typedef union {
 #if !defined (_TP_FIDDLE_H)
 #define _TP_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  tp_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -47069,21 +46964,6 @@ typedef union {
 #if !defined (_TC_FIDDLE_H)
 #define _TC_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  tc_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -47104,21 +46984,6 @@ typedef union {
 #if !defined (_SC_FIDDLE_H)
 #define _SC_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  sc_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -47139,21 +47004,6 @@ typedef union {
 #if !defined (_BC_FIDDLE_H)
 #define _BC_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  bc_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
diff --git a/drivers/mxc/amd-gpu/include/reg/yamato/22/yamato_struct.h b/drivers/mxc/amd-gpu/include/reg/yamato/22/yamato_struct.h
index 78d4924ef79d..21de35591e7d 100644
--- a/drivers/mxc/amd-gpu/include/reg/yamato/22/yamato_struct.h
+++ b/drivers/mxc/amd-gpu/include/reg/yamato/22/yamato_struct.h
@@ -7848,21 +7848,6 @@ typedef union {
 #if !defined (_RBBM_FIDDLE_H)
 #define _RBBM_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  rbbm_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -10978,21 +10963,6 @@ typedef union {
 #if !defined (_MH_FIDDLE_H)
 #define _MH_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  mh_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -19933,21 +19903,6 @@ typedef union {
 #if !defined (_PA_FIDDLE_H)
 #define _PA_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  pa_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -27853,21 +27808,6 @@ typedef union {
 #if !defined (_VGT_FIDDLE_H)
 #define _VGT_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  vgt_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -32997,21 +32937,6 @@ typedef union {
 #if !defined (_SQ_FIDDLE_H)
 #define _SQ_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  sq_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -42041,21 +41966,6 @@ typedef union {
 #if !defined (_SX_FIDDLE_H)
 #define _SX_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  sx_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -42076,21 +41986,6 @@ typedef union {
 #if !defined (_TP_FIDDLE_H)
 #define _TP_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  tp_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -48243,21 +48138,6 @@ typedef union {
 #if !defined (_TC_FIDDLE_H)
 #define _TC_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  tc_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -48278,21 +48158,6 @@ typedef union {
 #if !defined (_SC_FIDDLE_H)
 #define _SC_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  sc_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
@@ -48313,21 +48178,6 @@ typedef union {
 #if !defined (_BC_FIDDLE_H)
 #define _BC_FIDDLE_H
 
-/*****************************************************************************************************************
- *
- *  bc_reg.h
- *
- *  Register Spec Release:  Block Spec 1.0
- *
- *   (c) 2000 ATI Technologies Inc.  (unpublished)
- *
- *   All rights reserved.  This notice is intended as a precaution against
- *   inadvertent publication and does not imply publication or any waiver
- *   of confidentiality.  The year included in the foregoing notice is the
- *   year of creation of the work.
- *
- *****************************************************************************************************************/
-
 /*******************************************************
  * Enums
  *******************************************************/
diff --git a/drivers/mxc/amd-gpu/platform/hal/linux/gsl_kmod.c b/drivers/mxc/amd-gpu/platform/hal/linux/gsl_kmod.c
index b83706f9c36c..c2b26d3b2a0a 100644
--- a/drivers/mxc/amd-gpu/platform/hal/linux/gsl_kmod.c
+++ b/drivers/mxc/amd-gpu/platform/hal/linux/gsl_kmod.c
@@ -131,6 +131,18 @@ static int gsl_kmod_ioctl(struct inode *inode, struct file *fd, unsigned int cmd
             kgslStatus = kgsl_device_idle(param.device_id, param.timeout);
             break;
         }
+    case IOCTL_KGSL_DEVICE_ISIDLE:
+        {
+            kgsl_device_isidle_t param;
+            if (copy_from_user(&param, (void __user *)arg, sizeof(kgsl_device_isidle_t)))
+            {
+                printk(KERN_ERR "%s: copy_from_user error\n", __func__);
+                kgslStatus = GSL_FAILURE;
+                break;
+            }
+            kgslStatus = kgsl_device_isidle(param.device_id);
+            break;
+        }
     case IOCTL_KGSL_DEVICE_GETPROPERTY:
         {
             kgsl_device_getproperty_t param;