xvmc: init one-time mc context once

Don't setup one-time mc context everytime, as the content is always
unchanged. And several structs got packed layout inside to ease static
state initialization.
(cherry picked from commit b4d8ca8b38e495b56bb3b4143e5dfe91ee651f15)

3 files changed, 423 insertions, 400 deletions

diff --git a/src/xvmc/i915_structs.h b/src/xvmc/i915_structs.h
index a25d7349..7e786e78 100644
--- a/src/xvmc/i915_structs.h
+++ b/src/xvmc/i915_structs.h
@@ -28,6 +28,8 @@
 #ifndef _I915_STRUCTS_H
 #define _I915_STRUCTS_H
 
+#include <stdint.h>
+
 /* MI_INSTRUCTION */
 #define CMD_MI          0x00
 
@@ -532,6 +534,9 @@ struct i915_3dstate_sampler_state
         unsigned sampler_masker : 16;
         unsigned pad0 : 16;
     } dw1;
+    /* we always use two samplers for mc */
+    struct texture_sampler sampler0;
+    struct texture_sampler sampler1;
 };
 
 struct arithmetic_inst
@@ -641,17 +646,48 @@ union shader_inst
     struct declaration_inst d;
 };
 
+struct i915_3dstate_pixel_shader_header {
+    unsigned length : 9;
+    unsigned pad0 : 6;
+    unsigned retain : 1;
+    unsigned opcode : 13;
+    unsigned type : 3;
+};
+
 struct i915_3dstate_pixel_shader_program
 {
-    struct {
-        unsigned length : 9;
-        unsigned pad0 : 6;
-        unsigned retain : 1;
-        unsigned opcode : 13;
-        unsigned type : 3;
-    } dw0;
-
-    // union shader_inst *insts;
+    struct i915_3dstate_pixel_shader_header shader0;
+    /* mov oC, c0.0000 */
+    uint32_t inst0[3];
+
+    struct i915_3dstate_pixel_shader_header shader1;
+    /* dcl t0.xy */
+    /* dcl t1.xy */
+    /* dcl_2D s0 */
+    /* texld r0, t0, s0 */
+    /* mov oC, r0 */
+    uint32_t inst1[3*5];
+
+    struct i915_3dstate_pixel_shader_header shader2;
+    /* dcl t2.xy */
+    /* dcl t3.xy */
+    /* dcl_2D s1 */
+    /* texld r0, t2, s1 */
+    /* mov oC, r0 */
+    uint32_t inst2[3*5];
+
+    struct i915_3dstate_pixel_shader_header shader3;
+    /* dcl t0.xy */
+    /* dcl t1.xy */
+    /* dcl t2.xy */
+    /* dcl t3.xy */
+    /* dcl_2D s0 */
+    /* dcl_2D s1 */
+    /* texld r0, t0, s0 */
+    /* texld r0, t2, s1 */
+    /* add r0, r0, r1*/
+    /* mov oC, r0 */
+    uint32_t inst3[3*10];
 };
 
 #define REG_CR0         0x00000001
@@ -707,7 +743,8 @@ struct i915_3dstate_pixel_shader_constants
     struct {
         unsigned reg_mask;
     } dw1;
-    // struct shader_constant *consts;
+    /* we only need one constant */
+    struct shader_constant value;
 };
 
 #define BLOCK_SIS       0x01
diff --git a/src/xvmc/i915_xvmc.c b/src/xvmc/i915_xvmc.c
index bd31de44..45acf9b2 100644
--- a/src/xvmc/i915_xvmc.c
+++ b/src/xvmc/i915_xvmc.c
@@ -81,6 +81,372 @@ static int findOverlap(unsigned int width, unsigned int height,
 }
 #endif
 
+static void i915_inst_arith(unsigned int *inst,
+                            unsigned int op,
+                            unsigned int dest,
+                            unsigned int mask,
+                            unsigned int saturate,
+                            unsigned int src0, unsigned int src1, unsigned int src2)
+{
+    dest = UREG(GET_UREG_TYPE(dest), GET_UREG_NR(dest));
+    *inst = (op | A0_DEST(dest) | mask | saturate | A0_SRC0(src0));
+    inst++;
+    *inst = (A1_SRC0(src0) | A1_SRC1(src1));
+    inst++;
+    *inst = (A2_SRC1(src1) | A2_SRC2(src2));
+}
+
+static void i915_inst_decl(unsigned int *inst,
+                           unsigned int type,
+                           unsigned int nr,
+                           unsigned int d0_flags)
+{
+    unsigned int reg = UREG(type, nr);
+
+    *inst = (D0_DCL | D0_DEST(reg) | d0_flags);
+    inst++;
+    *inst = D1_MBZ;
+    inst++;
+    *inst = D2_MBZ;
+}
+
+static void i915_inst_texld(unsigned int *inst,
+                              unsigned int op,
+                              unsigned int dest,
+                              unsigned int coord,
+                              unsigned int sampler)
+{
+   dest = UREG(GET_UREG_TYPE(dest), GET_UREG_NR(dest));
+   *inst = (op | T0_DEST(dest) | T0_SAMPLER(sampler));
+   inst++;
+   *inst = T1_ADDRESS_REG(coord);
+   inst++;
+   *inst = T2_MBZ;
+}
+
+static void i915_emit_batch(void *data, int size, int flag)
+{
+    intelBatchbufferData(data, size, flag);
+}
+
+/* one time context initialization buffer */
+static uint32_t *one_time_load_state_imm1;
+static uint32_t *one_time_load_indirect;
+static int one_time_load_state_imm1_size, one_time_load_indirect_size;
+
+static void i915_mc_one_time_context_init(XvMCContext *context)
+{
+    unsigned int dest, src0, src1, src2;
+    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
+    int i;
+    struct i915_3dstate_sampler_state *sampler_state;
+    struct i915_3dstate_pixel_shader_program *pixel_shader_program;
+    struct i915_3dstate_pixel_shader_constants *pixel_shader_constants;
+
+    /* sampler static state */
+    sampler_state = (struct i915_3dstate_sampler_state *)pI915XvMC->ssb.map;
+    /* pixel shader static state */
+    pixel_shader_program = (struct i915_3dstate_pixel_shader_program *)pI915XvMC->psp.map;
+    /* pixel shader contant static state */
+    pixel_shader_constants = (struct i915_3dstate_pixel_shader_constants *)pI915XvMC->psc.map;
+
+    memset(sampler_state, 0, sizeof(*sampler_state));
+    sampler_state->dw0.type = CMD_3D;
+    sampler_state->dw0.opcode = OPC_3DSTATE_SAMPLER_STATE;
+    sampler_state->dw0.length = 6;
+    sampler_state->dw1.sampler_masker = SAMPLER_SAMPLER0 | SAMPLER_SAMPLER1;
+
+    sampler_state->sampler0.ts0.reverse_gamma = 0;
+    sampler_state->sampler0.ts0.planar2packet = 0;
+    sampler_state->sampler0.ts0.color_conversion = 0;
+    sampler_state->sampler0.ts0.chromakey_index = 0;
+    sampler_state->sampler0.ts0.base_level = 0;
+    sampler_state->sampler0.ts0.mip_filter = MIPFILTER_NONE;        /* NONE */
+    sampler_state->sampler0.ts0.mag_filter = MAPFILTER_LINEAR;      /* LINEAR */
+    sampler_state->sampler0.ts0.min_filter = MAPFILTER_LINEAR;      /* LINEAR */
+    sampler_state->sampler0.ts0.lod_bias = 0;       /* 0.0 */
+    sampler_state->sampler0.ts0.shadow_enable = 0;
+    sampler_state->sampler0.ts0.max_anisotropy = ANISORATIO_2;
+    sampler_state->sampler0.ts0.shadow_function = PREFILTEROP_ALWAYS;
+    sampler_state->sampler0.ts1.min_lod = 0;        /* 0.0 Maximum Mip Level */
+    sampler_state->sampler0.ts1.kill_pixel = 0;
+    sampler_state->sampler0.ts1.keyed_texture_filter = 0;
+    sampler_state->sampler0.ts1.chromakey_enable = 0;
+    sampler_state->sampler0.ts1.tcx_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler0.ts1.tcy_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler0.ts1.tcz_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler0.ts1.normalized_coor = 0;
+    sampler_state->sampler0.ts1.map_index = 0;
+    sampler_state->sampler0.ts1.east_deinterlacer = 0;
+    sampler_state->sampler0.ts2.default_color = 0;
+
+    sampler_state->sampler1.ts0.reverse_gamma = 0;
+    sampler_state->sampler1.ts0.planar2packet = 0;
+    sampler_state->sampler1.ts0.color_conversion = 0;
+    sampler_state->sampler1.ts0.chromakey_index = 0;
+    sampler_state->sampler1.ts0.base_level = 0;
+    sampler_state->sampler1.ts0.mip_filter = MIPFILTER_NONE;        /* NONE */
+    sampler_state->sampler1.ts0.mag_filter = MAPFILTER_LINEAR;      /* LINEAR */
+    sampler_state->sampler1.ts0.min_filter = MAPFILTER_LINEAR;      /* LINEAR */
+    sampler_state->sampler1.ts0.lod_bias = 0;       /* 0.0 */
+    sampler_state->sampler1.ts0.shadow_enable = 0;
+    sampler_state->sampler1.ts0.max_anisotropy = ANISORATIO_2;
+    sampler_state->sampler1.ts0.shadow_function = PREFILTEROP_ALWAYS;
+    sampler_state->sampler1.ts1.min_lod = 0;        /* 0.0 Maximum Mip Level */
+    sampler_state->sampler1.ts1.kill_pixel = 0;
+    sampler_state->sampler1.ts1.keyed_texture_filter = 0;
+    sampler_state->sampler1.ts1.chromakey_enable = 0;
+    sampler_state->sampler1.ts1.tcx_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler1.ts1.tcy_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler1.ts1.tcz_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler1.ts1.normalized_coor = 0;
+    sampler_state->sampler1.ts1.map_index = 1;
+    sampler_state->sampler1.ts1.east_deinterlacer = 0;
+    sampler_state->sampler1.ts2.default_color = 0;
+
+    memset(pixel_shader_program, 0, sizeof(*pixel_shader_program));
+    pixel_shader_program->shader0.type = CMD_3D;
+    pixel_shader_program->shader0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
+    pixel_shader_program->shader0.retain = 1;
+    pixel_shader_program->shader0.length = 2; /* 1 inst */
+    i = 0;
+
+    dest = UREG(REG_TYPE_OC, 0);
+    src0 = UREG(REG_TYPE_CONST, 0);
+    src1 = 0;
+    src2 = 0;
+    i915_inst_arith(&pixel_shader_program->inst0[i], A0_MOV,
+	    dest, A0_DEST_CHANNEL_ALL, A0_DEST_SATURATE, src0, src1, src2);
+
+    pixel_shader_program->shader1.type = CMD_3D;
+    pixel_shader_program->shader1.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
+    pixel_shader_program->shader1.retain = 1;
+    pixel_shader_program->shader1.length = 14; /* 5 inst */
+    i = 0;
+    /* dcl t0.xy */
+    i915_inst_decl(&pixel_shader_program->inst1[i], REG_TYPE_T, T_TEX0, D0_CHANNEL_XY);
+    i+=3;
+    /* dcl t1.xy */
+    i915_inst_decl(&pixel_shader_program->inst1[i], REG_TYPE_T, T_TEX1, D0_CHANNEL_XY);
+    /* dcl_2D s0 */
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst1[i], REG_TYPE_S, 0, D0_SAMPLE_TYPE_2D);
+    /* texld r0, t0, s0 */
+    i += 3;
+    dest = UREG(REG_TYPE_R, 0);
+    src0 = UREG(REG_TYPE_T, 0); /* COORD */
+    src1 = UREG(REG_TYPE_S, 0); /* SAMPLER */
+    i915_inst_texld(&pixel_shader_program->inst1[i], T0_TEXLD, dest, src0, src1);
+    /* mov oC, r0 */
+    i += 3;
+    dest = UREG(REG_TYPE_OC, 0);
+    src0 = UREG(REG_TYPE_R, 0);
+    src1 = src2 = 0;
+    i915_inst_arith(&pixel_shader_program->inst1[i], A0_MOV, dest, A0_DEST_CHANNEL_ALL,
+                    A0_DEST_SATURATE, src0, src1, src2);
+
+
+    pixel_shader_program->shader2.type = CMD_3D;
+    pixel_shader_program->shader2.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
+    pixel_shader_program->shader2.retain = 1;
+    pixel_shader_program->shader2.length = 14; /* 5 inst */
+    i = 0;
+    /* dcl t2.xy */
+    i915_inst_decl(&pixel_shader_program->inst2[i], REG_TYPE_T, T_TEX2, D0_CHANNEL_XY);
+    /* dcl t3.xy */
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst2[i], REG_TYPE_T, T_TEX3, D0_CHANNEL_XY);
+    /* dcl_2D s1 */
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst2[i], REG_TYPE_S, 1, D0_SAMPLE_TYPE_2D);
+    /* texld r0, t2, s1 */
+    i += 3;
+    dest = UREG(REG_TYPE_R, 0);
+    src0 = UREG(REG_TYPE_T, 2); /* COORD */
+    src1 = UREG(REG_TYPE_S, 1); /* SAMPLER */
+    i915_inst_texld(&pixel_shader_program->inst2[i], T0_TEXLD, dest, src0, src1);
+    /* mov oC, r0 */
+    i += 3;
+    dest = UREG(REG_TYPE_OC, 0);
+    src0 = UREG(REG_TYPE_R, 0);
+    src1 = src2 = 0;
+    i915_inst_arith(&pixel_shader_program->inst2[i], A0_MOV, dest, A0_DEST_CHANNEL_ALL,
+                    A0_DEST_SATURATE, src0, src1, src2);
+
+    /* Shader 3 */
+    pixel_shader_program->shader3.type = CMD_3D;
+    pixel_shader_program->shader3.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
+    pixel_shader_program->shader3.retain = 1;
+    pixel_shader_program->shader3.length = 29; /* 10 inst */
+    i = 0;
+    /* dcl t0.xy */
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_T, T_TEX0, D0_CHANNEL_XY);
+    /* dcl t1.xy */
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_T, T_TEX1, D0_CHANNEL_XY);
+    /* dcl t2.xy */
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_T, T_TEX2, D0_CHANNEL_XY);
+    /* dcl t3.xy */
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_T, T_TEX3, D0_CHANNEL_XY);
+    /* dcl_2D s0 */
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_S, 0, D0_SAMPLE_TYPE_2D);
+    /* dcl_2D s1 */
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_S, 1, D0_SAMPLE_TYPE_2D);
+    /* texld r0, t0, s0 */
+    i += 3;
+    dest = UREG(REG_TYPE_R, 0);
+    src0 = UREG(REG_TYPE_T, 0); /* COORD */
+    src1 = UREG(REG_TYPE_S, 0); /* SAMPLER */
+    i915_inst_texld(&pixel_shader_program->inst3[i], T0_TEXLD, dest, src0, src1);
+    /* texld r1, t2, s1 */
+    i += 3;
+    dest = UREG(REG_TYPE_R, 1);
+    src0 = UREG(REG_TYPE_T, 2); /* COORD */
+    src1 = UREG(REG_TYPE_S, 1); /* SAMPLER */
+    i915_inst_texld(&pixel_shader_program->inst3[i], T0_TEXLD, dest, src0, src1);
+    /* add r0, r0, r1 */
+    i += 3;
+    dest = UREG(REG_TYPE_R, 0);
+    src0 = UREG(REG_TYPE_R, 0);
+    src1 = UREG(REG_TYPE_R, 1);
+    src2 = 0;
+    i915_inst_arith(&pixel_shader_program->inst3[i], A0_ADD, dest, A0_DEST_CHANNEL_ALL,
+                    0 /* A0_DEST_SATURATE */, src0, src1, src2);
+    /* mul oC, r0, c0 */
+    i += 3;
+    dest = UREG(REG_TYPE_OC, 0);
+    src0 = UREG(REG_TYPE_R, 0);
+    src1 = UREG(REG_TYPE_CONST, 0);
+    src2 = 0;
+    i915_inst_arith(&pixel_shader_program->inst3[i], A0_MUL, dest, A0_DEST_CHANNEL_ALL,
+                    A0_DEST_SATURATE, src0, src1, src2);
+
+    memset(pixel_shader_constants, 0, sizeof(*pixel_shader_constants));
+    pixel_shader_constants->dw0.type = CMD_3D;
+    pixel_shader_constants->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_CONSTANTS;
+    pixel_shader_constants->dw0.length = 4;
+    pixel_shader_constants->dw1.reg_mask = REG_CR0;
+    pixel_shader_constants->value.x = 0.5;
+    pixel_shader_constants->value.y = 0.5;
+    pixel_shader_constants->value.z = 0.5;
+    pixel_shader_constants->value.w = 0.5;
+
+}
+
+static void i915_mc_one_time_state_init(XvMCContext *context)
+{
+    struct s3_dword *s3 = NULL;
+    struct s6_dword *s6 = NULL;
+    dis_state *dis = NULL;
+    ssb_state *ssb = NULL;
+    psp_state *psp = NULL;
+    psc_state *psc = NULL;
+    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
+    int mem_select = 1;
+    struct i915_3dstate_load_state_immediate_1 *load_state_immediate_1;
+    struct i915_3dstate_load_indirect *load_indirect;
+
+    /* 3DSTATE_LOAD_STATE_IMMEDIATE_1 */
+    one_time_load_state_imm1_size = sizeof(*load_state_immediate_1) + sizeof(*s3) + sizeof(*s6);
+    one_time_load_state_imm1 = calloc(1, one_time_load_state_imm1_size);
+    load_state_immediate_1 = (struct i915_3dstate_load_state_immediate_1 *)one_time_load_state_imm1;
+    load_state_immediate_1->dw0.type = CMD_3D;
+    load_state_immediate_1->dw0.opcode = OPC_3DSTATE_LOAD_STATE_IMMEDIATE_1;
+    load_state_immediate_1->dw0.load_s3 = 1;
+    load_state_immediate_1->dw0.load_s6 = 1;
+    load_state_immediate_1->dw0.length = (one_time_load_state_imm1_size >> 2) - 2;
+
+    s3 = (struct s3_dword *)(++load_state_immediate_1);
+    s3->set0_pcd = 1;
+    s3->set1_pcd = 1;
+    s3->set2_pcd = 1;
+    s3->set3_pcd = 1;
+    s3->set4_pcd = 1;
+    s3->set5_pcd = 1;
+    s3->set6_pcd = 1;
+    s3->set7_pcd = 1;
+
+    s6 = (struct s6_dword *)(++s3);
+    s6->alpha_test_enable = 0;
+    s6->alpha_test_function = 0;
+    s6->alpha_reference_value = 0;
+    s6->depth_test_enable = 1;
+    s6->depth_test_function = 0;
+    s6->color_buffer_blend = 0;
+    s6->color_blend_function = 0;
+    s6->src_blend_factor = 1;
+    s6->dest_blend_factor = 1;
+    s6->depth_buffer_write = 0;
+    s6->color_buffer_write = 1;
+    s6->triangle_pv = 0;
+
+    /* 3DSTATE_LOAD_INDIRECT */
+    one_time_load_indirect_size = sizeof(*load_indirect) + sizeof(*dis) + sizeof(*ssb) + sizeof(*psp) + sizeof(*psc);
+    one_time_load_indirect = calloc(1, one_time_load_indirect_size);
+    load_indirect = (struct i915_3dstate_load_indirect *)one_time_load_indirect;
+    load_indirect->dw0.type = CMD_3D;
+    load_indirect->dw0.opcode = OPC_3DSTATE_LOAD_INDIRECT;
+    load_indirect->dw0.block_mask = BLOCK_DIS | BLOCK_SSB | BLOCK_PSP | BLOCK_PSC;
+    load_indirect->dw0.length = (one_time_load_indirect_size >> 2) - 2;
+
+    if (pI915XvMC->deviceID == PCI_CHIP_I915_G ||
+        pI915XvMC->deviceID == PCI_CHIP_I915_GM ||
+        pI915XvMC->deviceID == PCI_CHIP_I945_G ||
+        pI915XvMC->deviceID == PCI_CHIP_I945_GM)
+        mem_select = 0;
+
+    load_indirect->dw0.mem_select = mem_select;
+
+    /* Dynamic indirect state buffer */
+    dis = (dis_state *)(++load_indirect);
+    dis->dw0.valid = 0;
+    dis->dw0.reset = 0;
+    dis->dw0.buffer_address = 0;
+
+    /* Sample state buffer */
+    ssb = (ssb_state *)(++dis);
+    ssb->dw0.valid = 1;
+    ssb->dw0.force = 1;
+    ssb->dw1.length = 7; /* 8 - 1 */
+
+    if (mem_select)
+        ssb->dw0.buffer_address = (pI915XvMC->ssb.offset >> 2);
+    else
+        ssb->dw0.buffer_address = (pI915XvMC->ssb.bus_addr >> 2);
+
+    /* Pixel shader program buffer */
+    psp = (psp_state *)(++ssb);
+    psp->dw0.valid = 1;
+    psp->dw0.force = 1;
+    psp->dw1.length = 66; /* 4 + 16 + 16 + 31 - 1 */
+
+    if (mem_select)
+        psp->dw0.buffer_address = (pI915XvMC->psp.offset >> 2);
+    else
+        psp->dw0.buffer_address = (pI915XvMC->psp.bus_addr >> 2);
+
+    /* Pixel shader constant buffer */
+    psc = (psc_state *)(++psp);
+    psc->dw0.valid = 1;
+    psc->dw0.force = 1;
+    psc->dw1.length = 5; /* 6 - 1 */
+
+    if (mem_select)
+        psc->dw0.buffer_address = (pI915XvMC->psc.offset >> 2);
+    else
+        psc->dw0.buffer_address = (pI915XvMC->psc.bus_addr >> 2);
+}
+
+static void i915_mc_one_time_state_emit(void)
+{
+    i915_emit_batch(one_time_load_state_imm1, one_time_load_state_imm1_size, 0);
+    i915_emit_batch(one_time_load_indirect, one_time_load_indirect_size, 0);
+}
+
 static void i915_flush(int map, int render)
 {
     struct i915_mi_flush mi_flush;
@@ -582,392 +948,6 @@ static void i915_mc_mpeg_macroblock_2fbmv(XvMCContext *context, XvMCMacroBlock *
     intelBatchbufferData(&macroblock_2fbmv, sizeof(macroblock_2fbmv), 0);
 }
 
-/* for MC context initialization */
-static void i915_mc_sampler_state_buffer(XvMCContext *context)
-{
-    struct i915_3dstate_sampler_state *sampler_state;
-    struct texture_sampler *ts;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-
-    /* 3DSATE_SAMPLER_STATE */
-    sampler_state = (struct i915_3dstate_sampler_state *)pI915XvMC->ssb.map;
-    memset(sampler_state, 0, sizeof(*sampler_state));
-    sampler_state->dw0.type = CMD_3D;
-    sampler_state->dw0.opcode = OPC_3DSTATE_SAMPLER_STATE;
-    sampler_state->dw0.length = 6;
-    sampler_state->dw1.sampler_masker = SAMPLER_SAMPLER0 | SAMPLER_SAMPLER1;
-
-    /* Sampler 0 */
-    ts = (struct texture_sampler *)(++sampler_state);
-    memset(ts, 0, sizeof(*ts));
-    ts->ts0.reverse_gamma = 0;
-    ts->ts0.planar2packet = 0;
-    ts->ts0.color_conversion = 0;
-    ts->ts0.chromakey_index = 0;
-    ts->ts0.base_level = 0;
-    ts->ts0.mip_filter = MIPFILTER_NONE;        /* NONE */
-    ts->ts0.mag_filter = MAPFILTER_LINEAR;      /* LINEAR */
-    ts->ts0.min_filter = MAPFILTER_LINEAR;      /* LINEAR */
-    ts->ts0.lod_bias = 0;       /* 0.0 */
-    ts->ts0.shadow_enable = 0;
-    ts->ts0.max_anisotropy = ANISORATIO_2;
-    ts->ts0.shadow_function = PREFILTEROP_ALWAYS;
-    ts->ts1.min_lod = 0;        /* 0.0 Maximum Mip Level */
-    ts->ts1.kill_pixel = 0;
-    ts->ts1.keyed_texture_filter = 0;
-    ts->ts1.chromakey_enable = 0;
-    ts->ts1.tcx_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.tcy_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.tcz_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.normalized_coor = 0;
-    ts->ts1.map_index = 0;
-    ts->ts1.east_deinterlacer = 0;
-    ts->ts2.default_color = 0;
-
-    /* Sampler 1 */
-    ++ts;
-    memset(ts, 0, sizeof(*ts));
-    ts->ts0.reverse_gamma = 0;
-    ts->ts0.planar2packet = 0;
-    ts->ts0.color_conversion = 0;
-    ts->ts0.chromakey_index = 0;
-    ts->ts0.base_level = 0;
-    ts->ts0.mip_filter = MIPFILTER_NONE;        /* NONE */
-    ts->ts0.mag_filter = MAPFILTER_LINEAR;      /* LINEAR */
-    ts->ts0.min_filter = MAPFILTER_LINEAR;      /* LINEAR */
-    ts->ts0.lod_bias = 0;       /* 0.0 */
-    ts->ts0.shadow_enable = 0;
-    ts->ts0.max_anisotropy = ANISORATIO_2;
-    ts->ts0.shadow_function = PREFILTEROP_ALWAYS;
-    ts->ts1.min_lod = 0;        /* 0.0 Maximum Mip Level */
-    ts->ts1.kill_pixel = 0;
-    ts->ts1.keyed_texture_filter = 0;
-    ts->ts1.chromakey_enable = 0;
-    ts->ts1.tcx_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.tcy_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.tcz_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.normalized_coor = 0;
-    ts->ts1.map_index = 1;
-    ts->ts1.east_deinterlacer = 0;
-    ts->ts2.default_color = 0;
-}
-
-static void i915_inst_arith(unsigned int *inst,
-                            unsigned int op,
-                            unsigned int dest,
-                            unsigned int mask,
-                            unsigned int saturate,
-                            unsigned int src0, unsigned int src1, unsigned int src2)
-{
-    dest = UREG(GET_UREG_TYPE(dest), GET_UREG_NR(dest));
-    *inst = (op | A0_DEST(dest) | mask | saturate | A0_SRC0(src0));
-    inst++;
-    *inst = (A1_SRC0(src0) | A1_SRC1(src1));
-    inst++;
-    *inst = (A2_SRC1(src1) | A2_SRC2(src2));
-}
-
-static void i915_inst_decl(unsigned int *inst,
-                           unsigned int type,
-                           unsigned int nr,
-                           unsigned int d0_flags)
-{
-    unsigned int reg = UREG(type, nr);
-
-    *inst = (D0_DCL | D0_DEST(reg) | d0_flags);
-    inst++;
-    *inst = D1_MBZ;
-    inst++;
-    *inst = D2_MBZ;
-}
-
-static void i915_inst_texld(unsigned int *inst,
-                              unsigned int op,
-                              unsigned int dest,
-                              unsigned int coord,
-                              unsigned int sampler)
-{
-   dest = UREG(GET_UREG_TYPE(dest), GET_UREG_NR(dest));
-   *inst = (op | T0_DEST(dest) | T0_SAMPLER(sampler));
-   inst++;
-   *inst = T1_ADDRESS_REG(coord);
-   inst++;
-   *inst = T2_MBZ;
-}
-
-static void i915_mc_pixel_shader_program_buffer(XvMCContext *context)
-{
-    struct i915_3dstate_pixel_shader_program *pixel_shader_program;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-    unsigned int *inst;
-    unsigned int dest, src0, src1, src2;
-
-    /* Shader 0 */
-    pixel_shader_program = (struct i915_3dstate_pixel_shader_program *)pI915XvMC->psp.map;
-    memset(pixel_shader_program, 0, sizeof(*pixel_shader_program));
-    pixel_shader_program->dw0.type = CMD_3D;
-    pixel_shader_program->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
-    pixel_shader_program->dw0.retain = 1;
-    pixel_shader_program->dw0.length = 2;
-    /* mov oC, c0.0000 */
-    inst = (unsigned int*)(++pixel_shader_program);
-    dest = UREG(REG_TYPE_OC, 0);
-    src0 = UREG(REG_TYPE_CONST, 0);
-    src1 = 0;
-    src2 = 0;
-    i915_inst_arith(inst, A0_MOV, dest, A0_DEST_CHANNEL_ALL,
-                    A0_DEST_SATURATE, src0, src1, src2);
-    inst += 3;
-
-    /* Shader 1 */
-    pixel_shader_program = (struct i915_3dstate_pixel_shader_program *)inst;
-    memset(pixel_shader_program, 0, sizeof(*pixel_shader_program));
-    pixel_shader_program->dw0.type = CMD_3D;
-    pixel_shader_program->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
-    pixel_shader_program->dw0.retain = 1;
-    pixel_shader_program->dw0.length = 14;
-    /* dcl t0.xy */
-    inst = (unsigned int*)(++pixel_shader_program);
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX0, D0_CHANNEL_XY);
-    /* dcl t1.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX1, D0_CHANNEL_XY);
-    /* dcl_2D s0 */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_S, 0, D0_SAMPLE_TYPE_2D);
-    /* texld r0, t0, s0 */
-    inst += 3;
-    dest = UREG(REG_TYPE_R, 0);
-    src0 = UREG(REG_TYPE_T, 0); /* COORD */
-    src1 = UREG(REG_TYPE_S, 0); /* SAMPLER */
-    i915_inst_texld(inst, T0_TEXLD, dest, src0, src1);
-    /* mov oC, r0 */
-    inst += 3;
-    dest = UREG(REG_TYPE_OC, 0);
-    src0 = UREG(REG_TYPE_R, 0);
-    src1 = src2 = 0;
-    i915_inst_arith(inst, A0_MOV, dest, A0_DEST_CHANNEL_ALL,
-                    A0_DEST_SATURATE, src0, src1, src2);
-    inst += 3;
-
-    /* Shader 2 */
-    pixel_shader_program = (struct i915_3dstate_pixel_shader_program *)inst;
-    memset(pixel_shader_program, 0, sizeof(*pixel_shader_program));
-    pixel_shader_program->dw0.type = CMD_3D;
-    pixel_shader_program->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
-    pixel_shader_program->dw0.retain = 1;
-    pixel_shader_program->dw0.length = 14;
-    /* dcl t2.xy */
-    inst = (unsigned int*)(++pixel_shader_program);
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX2, D0_CHANNEL_XY);
-    /* dcl t3.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX3, D0_CHANNEL_XY);
-    /* dcl_2D s1 */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_S, 1, D0_SAMPLE_TYPE_2D);
-    /* texld r0, t2, s1 */
-    inst += 3;
-    dest = UREG(REG_TYPE_R, 0);
-    src0 = UREG(REG_TYPE_T, 2); /* COORD */
-    src1 = UREG(REG_TYPE_S, 1); /* SAMPLER */
-    i915_inst_texld(inst, T0_TEXLD, dest, src0, src1);
-    /* mov oC, r0 */
-    inst += 3;
-    dest = UREG(REG_TYPE_OC, 0);
-    src0 = UREG(REG_TYPE_R, 0);
-    src1 = src2 = 0;
-    i915_inst_arith(inst, A0_MOV, dest, A0_DEST_CHANNEL_ALL,
-                    A0_DEST_SATURATE, src0, src1, src2);
-    inst += 3;
-
-    /* Shader 3 */
-    pixel_shader_program = (struct i915_3dstate_pixel_shader_program *)inst;
-    memset(pixel_shader_program, 0, sizeof(*pixel_shader_program));
-    pixel_shader_program->dw0.type = CMD_3D;
-    pixel_shader_program->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
-    pixel_shader_program->dw0.retain = 1;
-    pixel_shader_program->dw0.length = 29;
-    /* dcl t0.xy */
-    inst = (unsigned int*)(++pixel_shader_program);
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX0, D0_CHANNEL_XY);
-    /* dcl t1.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX1, D0_CHANNEL_XY);
-    /* dcl t2.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX2, D0_CHANNEL_XY);
-    /* dcl t3.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX3, D0_CHANNEL_XY);
-    /* dcl_2D s0 */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_S, 0, D0_SAMPLE_TYPE_2D);
-    /* dcl_2D s1 */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_S, 1, D0_SAMPLE_TYPE_2D);
-    /* texld r0, t0, s0 */
-    inst += 3;
-    dest = UREG(REG_TYPE_R, 0);
-    src0 = UREG(REG_TYPE_T, 0); /* COORD */
-    src1 = UREG(REG_TYPE_S, 0); /* SAMPLER */
-    i915_inst_texld(inst, T0_TEXLD, dest, src0, src1);
-    /* texld r1, t2, s1 */
-    inst += 3;
-    dest = UREG(REG_TYPE_R, 1);
-    src0 = UREG(REG_TYPE_T, 2); /* COORD */
-    src1 = UREG(REG_TYPE_S, 1); /* SAMPLER */
-    i915_inst_texld(inst, T0_TEXLD, dest, src0, src1);
-    /* add r0, r0, r1 */
-    inst += 3;
-    dest = UREG(REG_TYPE_R, 0);
-    src0 = UREG(REG_TYPE_R, 0);
-    src1 = UREG(REG_TYPE_R, 1);
-    src2 = 0;
-    i915_inst_arith(inst, A0_ADD, dest, A0_DEST_CHANNEL_ALL,
-                    0 /* A0_DEST_SATURATE */, src0, src1, src2);
-    /* mul oC, r0, c0 */
-    inst += 3;
-    dest = UREG(REG_TYPE_OC, 0);
-    src0 = UREG(REG_TYPE_R, 0);
-    src1 = UREG(REG_TYPE_CONST, 0);
-    src2 = 0;
-    i915_inst_arith(inst, A0_MUL, dest, A0_DEST_CHANNEL_ALL,
-                    A0_DEST_SATURATE, src0, src1, src2);
-    inst += 3;
-}
-
-static void i915_mc_pixel_shader_constants_buffer(XvMCContext *context)
-{
-    struct i915_3dstate_pixel_shader_constants *pixel_shader_constants;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-    float *value;
-
-    pixel_shader_constants = (struct i915_3dstate_pixel_shader_constants *)pI915XvMC->psc.map;
-    memset(pixel_shader_constants, 0, sizeof(*pixel_shader_constants));
-    pixel_shader_constants->dw0.type = CMD_3D;
-    pixel_shader_constants->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_CONSTANTS;
-    pixel_shader_constants->dw0.length = 4;
-    pixel_shader_constants->dw1.reg_mask = REG_CR0;
-    value = (float *)(++pixel_shader_constants);
-    *(value++) = 0.5;
-    *(value++) = 0.5;
-    *(value++) = 0.5;
-    *(value++) = 0.5;
-}
-
-static void i915_mc_one_time_state_initialization(XvMCContext *context)
-{
-    struct i915_3dstate_load_state_immediate_1 *load_state_immediate_1 = NULL;
-    struct s3_dword *s3 = NULL;
-    struct s6_dword *s6 = NULL;
-    struct i915_3dstate_load_indirect *load_indirect = NULL;
-    dis_state *dis = NULL;
-    ssb_state *ssb = NULL;
-    psp_state *psp = NULL;
-    psc_state *psc = NULL;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-    unsigned int size;
-    void *base = NULL;
-    int mem_select = 1;
-
-    /* 3DSTATE_LOAD_STATE_IMMEDIATE_1 */
-    size = sizeof(*load_state_immediate_1) + sizeof(*s3) + sizeof(*s6);
-    base = calloc(1, size);
-    load_state_immediate_1 = (struct i915_3dstate_load_state_immediate_1 *)base;
-    load_state_immediate_1->dw0.type = CMD_3D;
-    load_state_immediate_1->dw0.opcode = OPC_3DSTATE_LOAD_STATE_IMMEDIATE_1;
-    load_state_immediate_1->dw0.load_s3 = 1;
-    load_state_immediate_1->dw0.load_s6 = 1;
-    load_state_immediate_1->dw0.length = (size >> 2) - 2;
-
-    s3 = (struct s3_dword *)(++load_state_immediate_1);
-    s3->set0_pcd = 1;
-    s3->set1_pcd = 1;
-    s3->set2_pcd = 1;
-    s3->set3_pcd = 1;
-    s3->set4_pcd = 1;
-    s3->set5_pcd = 1;
-    s3->set6_pcd = 1;
-    s3->set7_pcd = 1;
-
-    s6 = (struct s6_dword *)(++s3);
-    s6->alpha_test_enable = 0;
-    s6->alpha_test_function = 0;
-    s6->alpha_reference_value = 0;
-    s6->depth_test_enable = 1;
-    s6->depth_test_function = 0;
-    s6->color_buffer_blend = 0;
-    s6->color_blend_function = 0;
-    s6->src_blend_factor = 1;
-    s6->dest_blend_factor = 1;
-    s6->depth_buffer_write = 0;
-    s6->color_buffer_write = 1;
-    s6->triangle_pv = 0;
-
-    intelBatchbufferData(base, size, 0);
-    free(base);
-
-    /* 3DSTATE_LOAD_INDIRECT */
-    size = sizeof(*load_indirect) + sizeof(*dis) + sizeof(*ssb) + sizeof(*psp) + sizeof(*psc);
-    base = calloc(1, size);
-    load_indirect = (struct i915_3dstate_load_indirect *)base;
-    load_indirect->dw0.type = CMD_3D;
-    load_indirect->dw0.opcode = OPC_3DSTATE_LOAD_INDIRECT;
-    load_indirect->dw0.block_mask = BLOCK_DIS | BLOCK_SSB | BLOCK_PSP | BLOCK_PSC;
-    load_indirect->dw0.length = (size >> 2) - 2;
-
-    if (pI915XvMC->deviceID == PCI_CHIP_I915_G ||
-        pI915XvMC->deviceID == PCI_CHIP_I915_GM ||
-        pI915XvMC->deviceID == PCI_CHIP_I945_G ||
-        pI915XvMC->deviceID == PCI_CHIP_I945_GM)
-        mem_select = 0;
-
-    load_indirect->dw0.mem_select = mem_select;
-
-    /* DIS */
-    dis = (dis_state *)(++load_indirect);
-    dis->dw0.valid = 0;
-    dis->dw0.reset = 0;
-    dis->dw0.buffer_address = 0;
-
-    /* SSB */
-    ssb = (ssb_state *)(++dis);
-    ssb->dw0.valid = 1;
-    ssb->dw0.force = 1;
-    ssb->dw1.length = 7; /* 8 - 1 */
-
-    if (mem_select)
-        ssb->dw0.buffer_address = (pI915XvMC->ssb.offset >> 2);
-    else
-        ssb->dw0.buffer_address = (pI915XvMC->ssb.bus_addr >> 2);
-
-    /* PSP */
-    psp = (psp_state *)(++ssb);
-    psp->dw0.valid = 1;
-    psp->dw0.force = 1;
-    psp->dw1.length = 66; /* 4 + 16 + 16 + 31 - 1 */
-
-    if (mem_select)
-        psp->dw0.buffer_address = (pI915XvMC->psp.offset >> 2);
-    else
-        psp->dw0.buffer_address = (pI915XvMC->psp.bus_addr >> 2);
-
-    /* PSC */
-    psc = (psc_state *)(++psp);
-    psc->dw0.valid = 1;
-    psc->dw0.force = 1;
-    psc->dw1.length = 5; /* 6 - 1 */
-
-    if (mem_select)
-        psc->dw0.buffer_address = (pI915XvMC->psc.offset >> 2);
-    else
-        psc->dw0.buffer_address = (pI915XvMC->psc.bus_addr >> 2);
-
-    intelBatchbufferData(base, size, 0);
-    free(base);
-}
-
 #if 0
 static void i915_mc_invalidate_subcontext_buffers(XvMCContext *context, unsigned int mask)
 {
@@ -1683,6 +1663,11 @@ static Status i915_xvmc_mc_create_context(Display *display, XvMCContext *context
     pI915XvMC->last_flip = 0;
     pI915XvMC->port = context->port;
     pI915XvMC->ref = 1;
+
+    /* preinit state buffers */
+    i915_mc_one_time_context_init(context);
+    i915_mc_one_time_state_init(context);
+
     return Success;
 }
 
@@ -1695,6 +1680,9 @@ static int i915_xvmc_mc_destroy_context(Display *display, XvMCContext *context)
 
     /* Pass Control to the X server to destroy the drm_context_t */
     i915_release_resource(display,context);
+
+    free(one_time_load_state_imm1);
+    free(one_time_load_indirect);
     return Success;
 }
 
@@ -1937,10 +1925,7 @@ static int i915_xvmc_mc_render_surface(Display *display, XvMCContext *context,
     // i915_mc_invalidate_subcontext_buffers(context, BLOCK_SIS | BLOCK_DIS | BLOCK_SSB
     // | BLOCK_MSB | BLOCK_PSP | BLOCK_PSC);
 
-    i915_mc_sampler_state_buffer(context);
-    i915_mc_pixel_shader_program_buffer(context);
-    i915_mc_pixel_shader_constants_buffer(context);
-    i915_mc_one_time_state_initialization(context);
+    i915_mc_one_time_state_emit();
 
     i915_mc_static_indirect_state_buffer(context, target_surface,
                                          picture_structure, flags,
diff --git a/src/xvmc/intel_xvmc.h b/src/xvmc/intel_xvmc.h
index 4d8b6055..c4dfd770 100644
--- a/src/xvmc/intel_xvmc.h
+++ b/src/xvmc/intel_xvmc.h
@@ -38,6 +38,7 @@
 #include <string.h>
 #include <assert.h>
 #include <signal.h>
+#include <stdint.h>
 
 #include <xf86drm.h>
 #include "i830_common.h"