#include <stdio.h>
#include <stdint.h>
#include "nouveau_class.h"
#include <nouveau_pushbuf.h>
#include <nouveau_notifier.h>

#include "screen.h"
#include "object.h"
#include "fifo.h"

#include "nv30_fpinst.h"

static void tcl_fp_color(void)
{
#if 0
000428e4  size 1, subchannel 1 (0xbeef3097),offset 0x08e4,increment
#      NV30_TCL_PRIMITIVE_3D_ACTIVE_PROGRAM = 0x07bfef01
# -- program at 0x0xad7e3000 - 0x20D000
#	0x01803e81 0x1c9dc901 0x0001c900 0x0001c900
#		INST 0: MOVX R0 (TR0.xyzw), attrib.color + END
07bfef01    NV30TCL.FP_ACTIVE_PROGRAM = DMA0=TRUE | DMA1=FALSE | OFFSET=0x07bfef00
#endif

	uint32_t *map;
	struct nouveau_bo *bo;
	struct nouveau_grobj *subch3D = grobj[NvSub3D];

	bo = screen_allocmem(0x1000);
	if (!bo) {
		return;
	}
	map = bo->map;

	printf("fragprog mapped at %p\n", map);
	/*
	 * MOV result.color, fragment.color
	 */
	map[0] = 0x01803e81;
	map[1] = 0x1c9dc901;
	map[2] = map[3] = 0x0001c900;

	/*
	NV30FP_LOCALS;
	NV30FP_SETBUF(fragprog_map);
	NV30FP_ARITH_INST_SET_DEFAULTS;
	NV30FP_ARITH(MOV, RESULT, 0, 1, 1, 1, 1);
	NV30FP_SOURCE_INPUT(0, NV30_FP_OP_INPUT_SRC_COL0, X, Y, Z, W, 0, 0, 0);
	NV30FP_LAST_INST;
	*/

	nouveau_bo_unmap(bo);

	BEGIN_RING(chan, subch3D, NV34TCL_FP_ACTIVE_PROGRAM, 1);
	OUT_RELOCd(chan, bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART |
				NOUVEAU_BO_RD | NOUVEAU_BO_LOW,
				NV34TCL_FP_ACTIVE_PROGRAM_DMA0, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_FP_CONTROL, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_FP_REG_CONTROL, 1);
	OUT_RING  (chan, (1<<16)|4);

	nouveau_bo_ref(NULL, &bo);
}

#if 0
static void tcl_fp_nop(void)
{
/*
000428e4  size 1, subchannel 1 (0xbeef3097),offset 0x08e4,increment
#      NV30_TCL_PRIMITIVE_3D_ACTIVE_PROGRAM = 0x07bfef41
# -- program at 0x0xad7e3010 - 0x20D010
#	0x00000001 0x00000000 0x00000000 0x00000000
#		INST 0: NOP (FL0.xxxx)  + END
07bfef41    NV30TCL.FP_ACTIVE_PROGRAM = DMA0=TRUE | DMA1=FALSE | OFFSET=0x07bfef40
*/

	uint32_t *map;
	struct nouveau_bo *bo;
	struct nouveau_grobj *subch3D = grobj[NvSub3D];

	bo = screen_allocmem(0x1000);
	if (!bo) {
		return;
	}
	map = bo->map;

	printf("fragprog mapped at %p\n", map);

	/*
	 * NOP
	 */
	map[0] = 1;
	map[1] =
	map[2] =
	map[3] = 0;

	nouveau_bo_unmap(bo);

	BEGIN_RING(chan, subch3D, NV34TCL_FP_ACTIVE_PROGRAM, 1);
	OUT_RELOCd(chan, bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART |
				NOUVEAU_BO_RD | NOUVEAU_BO_LOW,
				NV34TCL_FP_ACTIVE_PROGRAM_DMA0, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_FP_CONTROL, 1);
	OUT_RING  (chan, 0x40);
	BEGIN_RING(chan, subch3D, NV34TCL_FP_REG_CONTROL, 1);
	OUT_RING  (chan, (1<<16)|4);

	nouveau_bo_ref(NULL, &bo);
}
#endif

void tcl_init(void)
{
	int i;
	struct nouveau_grobj *subch3D = grobj[NvSub3D];
	struct nouveau_grobj *subchImageBlit = grobj[NvSubImageBlit];

	printf("-- TCL init\n");

	BEGIN_RING(chan, subch3D, NV34TCL_DMA_NOTIFY, 1);
	OUT_RING  (chan, notifier->handle);

	BEGIN_RING(chan, subch3D, NV34TCL_DMA_TEXTURE0, 3);
	OUT_RING  (chan, chan->vram->handle);
	OUT_RING  (chan, chan->gart->handle);
	OUT_RING  (chan, chan->vram->handle);
	BEGIN_RING(chan, subch3D, NV34TCL_DMA_COLOR0, 2);
	OUT_RING  (chan, chan->vram->handle);
	OUT_RING  (chan, chan->vram->handle);
	BEGIN_RING(chan, subch3D, NV34TCL_DMA_IN_MEMORY7, 2);
	OUT_RING  (chan, chan->vram->handle);
	OUT_RING  (chan, chan->vram->handle);
#if 0
000421a4  size 1, subchannel 1 (0xbeef3097),offset 0x01a4,increment
beef1e00    NV30TCL.DMA_FENCE = NV01_MEMORY_LOCAL_BANKED
000421a8  size 1, subchannel 1 (0xbeef3097),offset 0x01a8,increment
beef0351    NV30TCL.DMA_QUERY = NV01_MEMORY_LOCAL_BANKED
#endif
	BEGIN_RING(chan, subch3D, NV34TCL_VIEWPORT_TX_ORIGIN, 1);
	OUT_RING  (chan, 0);

	BEGIN_RING(chan, subch3D, NV34TCL_VIEWPORT_CLIP_HORIZ(0), 1);
	OUT_RING  (chan, (viewport_w-1)<<16);
	BEGIN_RING(chan, subch3D, NV34TCL_VIEWPORT_CLIP_VERT(0), 1);
	OUT_RING  (chan, (viewport_h-1)<<16);
	for (i=1; i<8; i++) {
		BEGIN_RING(chan, subch3D, NV34TCL_VIEWPORT_CLIP_HORIZ(i), 1);
		OUT_RING  (chan, 0);
		BEGIN_RING(chan, subch3D, NV34TCL_VIEWPORT_CLIP_VERT(i), 1);
		OUT_RING  (chan, 0);
	}
	BEGIN_RING(chan, subch3D, 0x2bc /* NV34TCL_VIEWPORT_CLIP_MODE */, 1);
	OUT_RING  (chan, 0);

	BEGIN_RING(chan, subch3D, 0x3b0, 1);
	OUT_RING  (chan, 0x10<<16);
	BEGIN_RING(chan, subch3D, 0x1454, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, 0x1d80, 1);
	OUT_RING  (chan, 3);

	BEGIN_RING(chan, subch3D, 0x1e98, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, 0x17e0, 3);
	OUT_RINGf (chan, 0.0);
	OUT_RINGf (chan, 0.0);
	OUT_RINGf (chan, 1.0);

	BEGIN_RING(chan, subch3D, 0x1f80, 16);
	for (i=0; i<16; i++) {
		OUT_RING(chan, (i==8) ? 0xffff : 0);
	}

	BEGIN_RING(chan, subch3D, 0x120, 3);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 1);
	OUT_RING  (chan, 2);

	BEGIN_RING(chan, subchImageBlit, 0x120, 3);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 1);
	OUT_RING  (chan, 2);

	BEGIN_RING(chan, subch3D, 0x1d88, 1);
	OUT_RING  (chan, 0x1200);

	BEGIN_RING(chan, subch3D, NV34TCL_VIEWPORT_HORIZ, 2);
	OUT_RING  (chan, viewport_w<<16);
	OUT_RING  (chan, viewport_h<<16);
	BEGIN_RING(chan, subch3D, NV34TCL_SCISSOR_HORIZ, 2);
	OUT_RING  (chan, viewport_w<<16);
	OUT_RING  (chan, viewport_h<<16);
	BEGIN_RING(chan, subch3D, NV34TCL_RT_HORIZ, 2);
	OUT_RING  (chan, viewport_w<<16);
	OUT_RING  (chan, viewport_h<<16);

	BEGIN_RING(chan, subch3D, NV34TCL_RT_FORMAT, 1);
	OUT_RING  (chan, NV34TCL_RT_FORMAT_TYPE_LINEAR | NV34TCL_RT_FORMAT_ZETA_Z24S8 | NV34TCL_RT_FORMAT_COLOR_X8R8G8B8);
	BEGIN_RING(chan, subch3D, NV34TCL_RT_ENABLE, 1);
	OUT_RING  (chan, NV34TCL_RT_ENABLE_COLOR0);

	BEGIN_RING(chan, subch3D, 0x1da4, 1);
	OUT_RING  (chan, 0);

	BEGIN_RING(chan, subch3D, NV34TCL_COLOR0_PITCH, 5);
	OUT_RING  (chan, (screen_pitch<<16)|screen_pitch);		/* zeta_pitch, color0_pitch */
	OUT_RING  (chan, screen_offset);				/* color0_offset */
	OUT_RING  (chan, screen_offset + viewport_w*(screen_bpp/8));	/* zeta_offset, on the right of color buffer */
	OUT_RING  (chan, screen_offset);				/* color1_offset */
	OUT_RING  (chan, screen_pitch);					/* color1_pitch */

	BEGIN_RING(chan, subch3D, NV34TCL_LMA_DEPTH_PITCH, 2);
	OUT_RING  (chan, screen_pitch);
	OUT_RING  (chan, screen_offset + viewport_h * screen_pitch);	/* lma below color buffer */

	BEGIN_RING(chan, subch3D, 0x234, 1);
	OUT_RING  (chan, 3);

	BEGIN_RING(chan, subch3D, NV34TCL_ALPHA_FUNC_ENABLE, 3);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0x207);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_RC_ENABLE, 1);
	OUT_RING  (chan, 0);

	for (i=0; i<4; i++) {
		BEGIN_RING(chan, subch3D, NV34TCL_TX_ENABLE(i), 1);
		OUT_RING  (chan, 0);
	}
	BEGIN_RING(chan, subch3D, NV34TCL_MULTISAMPLE_CONTROL, 1);
	OUT_RING  (chan, 0xffff<<16);

	BEGIN_RING(chan, subch3D, NV34TCL_BLEND_FUNC_ENABLE, 5);
	OUT_RING  (chan, 0);
	OUT_RING(  chan, (1<<16)|1);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0x8006);
	BEGIN_RING(chan, subch3D, NV34TCL_STENCIL_FRONT_ENABLE, 17);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0xff);
	OUT_RING  (chan, 0x207);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0xff);
	OUT_RING  (chan, 0x1e00);
	OUT_RING  (chan, 0x1e00);
	OUT_RING  (chan, 0x1e00);
	OUT_RING  (chan, 1);
	OUT_RING  (chan, 0xff);
	OUT_RING  (chan, 0x207);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0xff);
	OUT_RING  (chan, 0x1e00);
	OUT_RING  (chan, 0x1e00);
	OUT_RING  (chan, 0x1e00);
	OUT_RING  (chan, 0x1d01);
	BEGIN_RING(chan, subch3D, NV34TCL_COLOR_LOGIC_OP_ENABLE, 2);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0x1503);
	BEGIN_RING(chan, subch3D, NV34TCL_DITHER_ENABLE, 1);
	OUT_RING  (chan, 1);
	BEGIN_RING(chan, subch3D, NV34TCL_POLYGON_STIPPLE_ENABLE, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_LINE_STIPPLE_ENABLE, 2);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0xffff<<16);
	BEGIN_RING(chan, subch3D, 0x1d84, 1);
	OUT_RING  (chan, 1);
	BEGIN_RING(chan, subch3D, 0x17cc, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_POLYGON_STIPPLE_PATTERN(0), 32);
	for (i=0; i<32; i++) {
		OUT_RING  (chan, 0xffffffff);
	}
	BEGIN_RING(chan, subch3D, NV34TCL_POLYGON_OFFSET_POINT_ENABLE, 3);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_DEPTH_FUNC, 1);
	OUT_RING  (chan, 0x201);
	BEGIN_RING(chan, subch3D, NV34TCL_DEPTH_WRITE_ENABLE, 1);
	OUT_RING  (chan, 1);
	BEGIN_RING(chan, subch3D, NV34TCL_DEPTH_TEST_ENABLE, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_POLYGON_OFFSET_FACTOR, 2);
	OUT_RINGf (chan, 0.0);
	OUT_RINGf (chan, 0.0);
	BEGIN_RING(chan, subch3D, 0x1d78, 1);
	OUT_RING  (chan, 1);
	BEGIN_RING(chan, subch3D, NV34TCL_LINE_WIDTH, 2);
	OUT_RING  (chan, 8);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_LINE_STIPPLE_ENABLE, 2);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0xffff<<16);
	BEGIN_RING(chan, subch3D, NV34TCL_POLYGON_MODE_FRONT, 2);
	OUT_RING  (chan, 0x1b02);
	OUT_RING  (chan, 0x1b02);
	BEGIN_RING(chan, subch3D, NV34TCL_CULL_FACE, 2);
	OUT_RING  (chan, 0x405);
	OUT_RING  (chan, 0x901);
	BEGIN_RING(chan, subch3D, NV34TCL_POLYGON_SMOOTH_ENABLE, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_CULL_FACE_ENABLE, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_SHADE_MODEL, 1);
	OUT_RING  (chan, 0x1d01);
	BEGIN_RING(chan, subch3D, NV34TCL_POLYGON_STIPPLE_ENABLE, 1);
	OUT_RING  (chan, 0);

	BEGIN_RING(chan, subch3D, NV34TCL_VTX_ATTR_4F_X(1), 60);
	OUT_RINGf (chan, 1.0);
	OUT_RINGf (chan, 0.0);
	OUT_RINGf (chan, 0.0);
	OUT_RINGf (chan, 1.0);

	OUT_RINGf (chan, 0.0);
	OUT_RINGf (chan, 0.0);
	OUT_RINGf (chan, 1.0);
	OUT_RINGf (chan, 1.0);

	OUT_RINGf (chan, 1.0);
	OUT_RINGf (chan, 1.0);
	OUT_RINGf (chan, 1.0);
	OUT_RINGf (chan, 1.0);
	for (i=4; i<16; i++) {
		OUT_RINGf (chan, 0.0);
		OUT_RINGf (chan, 0.0);
		OUT_RINGf (chan, 0.0);
		OUT_RINGf (chan, 1.0);
	}

	BEGIN_RING(chan, subch3D, 0x145c, 1);
	OUT_RING  (chan, 1);
	BEGIN_RING(chan, subch3D, NV34TCL_COLOR_MASK, 1);
	OUT_RING  (chan, 0x01010101);
	BEGIN_RING(chan, subch3D, 0x1458, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, 0x1424, 3);
	OUT_RING  (chan, 2<<16);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_ENABLED_LIGHTS, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_NORMALIZE_ENABLE, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_POINT_SIZE, 3);
	OUT_RINGf (chan, 1.0);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0);

	BEGIN_RING(chan, subch3D, NV34TCL_TX_GEN_S(0), 32);
	for (i=0; i<32; i++) {
		OUT_RING  (chan, 0);
	}

	BEGIN_RING(chan, subch3D, NV34TCL_FOG_MODE, 1);
	OUT_RING  (chan, 0x802);
	BEGIN_RING(chan, subch3D, NV34TCL_FOG_EQUATION_CONSTANT, 3);
	OUT_RINGf (chan, 1.5);
	OUT_RINGf (chan, -0.090168);
	OUT_RINGf(chan, 0.0);
	BEGIN_RING(chan, subch3D, NV34TCL_FOG_COORD_DIST, 1);
	OUT_RING  (chan, 2);
	BEGIN_RING(chan, subch3D, NV34TCL_FOG_ENABLE, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_FOG_COLOR, 1);
	OUT_RING  (chan, 0);

	BEGIN_RING(chan, subch3D, NV34TCL_ENGINE, 1);
	OUT_RING  (chan, NV34TCL_ENGINE_FIXED);

	BEGIN_RING(chan, subch3D, NV34TCL_VIEWPORT_TRANSLATE_X, 4);
	OUT_RINGf (chan, viewport_w * 0.5);
	OUT_RINGf (chan, viewport_h * 0.5);
	OUT_RINGf (chan, 0.5);
	OUT_RINGf (chan, 0.0);
	BEGIN_RING(chan, subch3D, NV34TCL_VIEWPORT_SCALE_X, 4);
	OUT_RINGf (chan, viewport_w * 0.5);
	OUT_RINGf (chan, viewport_h * -0.5);
	OUT_RINGf (chan, 0.5);
	OUT_RINGf (chan, 0.0);

	BEGIN_RING(chan, subch3D, NV34TCL_VP_CLIP_PLANES_ENABLE, 1);
	OUT_RING  (chan, 0);

	BEGIN_RING(chan, subch3D, NV34TCL_PROJECTION_MATRIX(0), 16);
	for (i=0; i<16; i++) {
		if ((i==0) || (i==5) || (i==10) || (i==15)) {
			OUT_RINGf (chan, 1.0);
		} else {
			OUT_RING  (chan, 0);
		}
	}

	BEGIN_RING(chan, subch3D, NV34TCL_DEPTH_RANGE_NEAR, 2);
	OUT_RINGf (chan, 0.0);
	OUT_RINGf (chan, 1.0);

	BEGIN_RING(chan, subch3D, NV34TCL_TX_UNITS_ENABLE, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_TX_MATRIX_ENABLE(0), 8);
	for (i=0; i<8; i++) {
		OUT_RING  (chan, 0);
	}

	tcl_fp_color();

	/*BEGIN_RING(NvSub3D, 0x1dac, 1);
	OUT_RING(0);

	BEGIN_RING(NvSub3D, NV34TCL_VERTEX_BEGIN_END, 1);
	OUT_RING(NV34TCL_VERTEX_BEGIN_END_TRIANGLES);

	BEGIN_RING(NvSub3D, NV34TCL_VTX_ATTR_3F_X(0), 3);
	OUT_RINGf(1.0);
	OUT_RINGf(0.0);
	OUT_RINGf(0.0);
	BEGIN_RING(NvSub3D, NV34TCL_VTX_ATTR_3F_X(0), 3);
	OUT_RINGf(0.0);
	OUT_RINGf(1.0);
	OUT_RINGf(0.0);
	BEGIN_RING(NvSub3D, NV34TCL_VTX_ATTR_3F_X(0), 3);
	OUT_RINGf(0.0);
	OUT_RINGf(0.0);
	OUT_RINGf(1.0);

	BEGIN_RING(NvSub3D, NV34TCL_VERTEX_BEGIN_END, 1);
	OUT_RING(NV34TCL_VERTEX_BEGIN_END_STOP);*/

	FIRE_RING(chan);
}

void tcl_clear(void)
{
	struct nouveau_grobj *subch3D = grobj[NvSub3D];

	printf("-- TCL clear buffers\n");

	BEGIN_RING(chan, subch3D, NV34TCL_CLEAR_DEPTH_VALUE, 2);
	OUT_RING  (chan, 0);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_IDXBUF_ADDRESS, 1);
	OUT_RING  (chan, 0);
	BEGIN_RING(chan, subch3D, NV34TCL_CLEAR_BUFFERS, 1);
	OUT_RING  (chan, 0xf0);

	FIRE_RING(chan);
}