st/nine: Implement TEXBEM,TEXBEML and BEM

Signed-off-by: Tiziano Bacocco <[email protected]>

7 files changed, 165 insertions, 8 deletions

diff --git a/src/gallium/state_trackers/nine/device9.c b/src/gallium/state_trackers/nine/device9.c
index e0f3e398ad7..1ca04a430ea 100644
--- a/src/gallium/state_trackers/nine/device9.c
+++ b/src/gallium/state_trackers/nine/device9.c
@@ -342,8 +342,9 @@ NineDevice9_ctor( struct NineDevice9 *This,
         This->state.vs_const_f = CALLOC(This->vs_const_size, 1);
         This->state.ps_const_f = CALLOC(This->ps_const_size, 1);
         This->state.vs_lconstf_temp = CALLOC(This->vs_const_size,1);
+        This->state.ps_lconstf_temp = CALLOC(This->ps_const_size,1);
         if (!This->state.vs_const_f || !This->state.ps_const_f ||
-            !This->state.vs_lconstf_temp)
+            !This->state.vs_lconstf_temp || !This->state.ps_lconstf_temp)
             return E_OUTOFMEMORY;
 
         if (strstr(pScreen->get_name(pScreen), "AMD") ||
@@ -466,6 +467,7 @@ NineDevice9_dtor( struct NineDevice9 *This )
     FREE(This->state.vs_const_f);
     FREE(This->state.ps_const_f);
     FREE(This->state.vs_lconstf_temp);
+    FREE(This->state.ps_lconstf_temp);
 
     if (This->swapchains) {
         for (i = 0; i < This->nswapchains; ++i)
@@ -2636,6 +2638,7 @@ NineDevice9_SetTextureStageState( struct NineDevice9 *This,
                                   DWORD Value )
 {
     struct nine_state *state = This->update;
+    int bumpmap_index = -1;
 
     DBG("Stage=%u Type=%u Value=%08x\n", Stage, Type, Value);
     nine_dump_D3DTSS_value(DBG_FF, Type, Value);
@@ -2644,6 +2647,33 @@ NineDevice9_SetTextureStageState( struct NineDevice9 *This,
     user_assert(Type < Elements(state->ff.tex_stage[0]), D3DERR_INVALIDCALL);
 
     state->ff.tex_stage[Stage][Type] = Value;
+    switch (Type) {
+    case D3DTSS_BUMPENVMAT00:
+        bumpmap_index = 4 * Stage;
+        break;
+    case D3DTSS_BUMPENVMAT10:
+        bumpmap_index = 4 * Stage + 1;
+        break;
+    case D3DTSS_BUMPENVMAT01:
+        bumpmap_index = 4 * Stage + 2;
+        break;
+    case D3DTSS_BUMPENVMAT11:
+        bumpmap_index = 4 * Stage + 3;
+        break;
+    case D3DTSS_BUMPENVLSCALE:
+        bumpmap_index = 4 * 8 + 2 * Stage;
+        break;
+    case D3DTSS_BUMPENVLOFFSET:
+        bumpmap_index = 4 * 8 + 2 * Stage + 1;
+        break;
+    default:
+        break;
+    }
+
+    if (bumpmap_index >= 0) {
+        state->bumpmap_vars[bumpmap_index] = Value;
+        state->changed.group |= NINE_STATE_PS_CONST;
+    }
 
     state->changed.group |= NINE_STATE_FF_PSSTAGES;
     state->ff.changed.tex_stage[Stage][Type / 32] |= 1 << (Type % 32);
diff --git a/src/gallium/state_trackers/nine/nine_shader.c b/src/gallium/state_trackers/nine/nine_shader.c
index 754f5af6b8e..a11c4c7c35a 100644
--- a/src/gallium/state_trackers/nine/nine_shader.c
+++ b/src/gallium/state_trackers/nine/nine_shader.c
@@ -89,6 +89,15 @@ static inline const char *d3dsio_to_string(unsigned opcode);
 #define NINE_SWIZZLE4(x,y,z,w) \
    TGSI_SWIZZLE_##x, TGSI_SWIZZLE_##y, TGSI_SWIZZLE_##z, TGSI_SWIZZLE_##w
 
+#define NINE_CONSTANT_SRC(index) \
+   ureg_src_register(TGSI_FILE_CONSTANT, index)
+
+#define NINE_APPLY_SWIZZLE(src, s) \
+   ureg_swizzle(src, NINE_SWIZZLE4(s, s, s, s))
+
+#define NINE_CONSTANT_SRC_SWIZZLE(index, s) \
+   NINE_APPLY_SWIZZLE(NINE_CONSTANT_SRC(index), s)
+
 #define NINED3DSPDM_SATURATE (D3DSPDM_SATURATE >> D3DSP_DSTMOD_SHIFT)
 #define NINED3DSPDM_PARTIALP (D3DSPDM_PARTIALPRECISION >> D3DSP_DSTMOD_SHIFT)
 #define NINED3DSPDM_CENTROID (D3DSPDM_MSAMPCENTROID >> D3DSP_DSTMOD_SHIFT)
@@ -2135,12 +2144,76 @@ DECL_SPECIAL(TEXKILL)
 
 DECL_SPECIAL(TEXBEM)
 {
-    STUB(D3DERR_INVALIDCALL);
-}
+    struct ureg_program *ureg = tx->ureg;
+    struct ureg_dst dst = tx_dst_param(tx, &tx->insn.dst[0]);
+    struct ureg_dst tmp, tmp2;
+    struct ureg_src sample, m00, m01, m10, m11;
+    struct ureg_src bumpenvlscale, bumpenvloffset;
+    const int m = tx->insn.dst[0].idx;
+    const int n = tx->insn.src[0].idx;
 
-DECL_SPECIAL(TEXBEML)
-{
-    STUB(D3DERR_INVALIDCALL);
+    assert(tx->version.major == 1);
+
+    sample = ureg_DECL_sampler(ureg, m);
+    tx->info->sampler_mask |= 1 << m;
+
+    tx_texcoord_alloc(tx, m);
+
+    tmp = tx_scratch(tx);
+    tmp2 = tx_scratch(tx);
+    /*
+     * Bump-env-matrix:
+     * 00 is X
+     * 01 is Y
+     * 10 is Z
+     * 11 is W
+     */
+    nine_info_mark_const_f_used(tx->info, 8 + 8 + m/2);
+    m00 = NINE_CONSTANT_SRC_SWIZZLE(8 + m, X);
+    m01 = NINE_CONSTANT_SRC_SWIZZLE(8 + m, Y);
+    m10 = NINE_CONSTANT_SRC_SWIZZLE(8 + m, Z);
+    m11 = NINE_CONSTANT_SRC_SWIZZLE(8 + m, W);
+
+    /* These two attributes are packed as X=scale0 Y=offset0 Z=scale1 W=offset1 etc */
+    if (m % 2 == 0) {
+        bumpenvlscale = NINE_CONSTANT_SRC_SWIZZLE(8 + 8 + m / 2, X);
+        bumpenvloffset = NINE_CONSTANT_SRC_SWIZZLE(8 + 8 + m / 2, Y);
+    } else {
+        bumpenvlscale = NINE_CONSTANT_SRC_SWIZZLE(8 + 8 + m / 2, Z);
+        bumpenvloffset = NINE_CONSTANT_SRC_SWIZZLE(8 + 8 + m / 2, W);
+    }
+
+    /* u' = TextureCoordinates(stage m)u + D3DTSS_BUMPENVMAT00(stage m)*t(n)R  */
+    ureg_MAD(ureg, ureg_writemask(tmp, TGSI_WRITEMASK_X), m00,
+             NINE_APPLY_SWIZZLE(ureg_src(tx->regs.tS[n]), X), tx->regs.vT[m]);
+    /* u' = u' + D3DTSS_BUMPENVMAT10(stage m)*t(n)G */
+    ureg_MAD(ureg, ureg_writemask(tmp, TGSI_WRITEMASK_X), m10,
+             NINE_APPLY_SWIZZLE(ureg_src(tx->regs.tS[n]), Y),
+             NINE_APPLY_SWIZZLE(ureg_src(tmp), X));
+
+    /* v' = TextureCoordinates(stage m)v + D3DTSS_BUMPENVMAT01(stage m)*t(n)R */
+    ureg_MAD(ureg, ureg_writemask(tmp, TGSI_WRITEMASK_Y), m01,
+             NINE_APPLY_SWIZZLE(ureg_src(tx->regs.tS[n]), X), tx->regs.vT[m]);
+    /* v' = v' + D3DTSS_BUMPENVMAT11(stage m)*t(n)G*/
+    ureg_MAD(ureg, ureg_writemask(tmp, TGSI_WRITEMASK_Y), m11,
+             NINE_APPLY_SWIZZLE(ureg_src(tx->regs.tS[n]), Y),
+             NINE_APPLY_SWIZZLE(ureg_src(tmp), Y));
+
+    /* Now the texture coordinates are in tmp.xy */
+
+    if (tx->insn.opcode == D3DSIO_TEXBEM) {
+        ureg_TEX(ureg, dst, ps1x_sampler_type(tx->info, m), ureg_src(tmp), sample);
+    } else if (tx->insn.opcode == D3DSIO_TEXBEML) {
+        /* t(m)RGBA = t(m)RGBA * [(t(n)B * D3DTSS_BUMPENVLSCALE(stage m)) + D3DTSS_BUMPENVLOFFSET(stage m)] */
+        ureg_TEX(ureg, tmp, ps1x_sampler_type(tx->info, m), ureg_src(tmp), sample);
+        ureg_MAD(ureg, tmp2, NINE_APPLY_SWIZZLE(ureg_src(tx->regs.tS[n]), Z),
+                 bumpenvlscale, bumpenvloffset);
+        ureg_MUL(ureg, dst, ureg_src(tmp), ureg_src(tmp2));
+    }
+
+    tx->info->bumpenvmat_needed = 1;
+
+    return D3D_OK;
 }
 
 DECL_SPECIAL(TEXREG2AR)
@@ -2421,7 +2494,43 @@ DECL_SPECIAL(TEXDEPTH)
 
 DECL_SPECIAL(BEM)
 {
-    STUB(D3DERR_INVALIDCALL);
+    struct ureg_program *ureg = tx->ureg;
+    struct ureg_dst dst = tx_dst_param(tx, &tx->insn.dst[0]);
+    struct ureg_src src0 = tx_src_param(tx, &tx->insn.src[0]);
+    struct ureg_src src1 = tx_src_param(tx, &tx->insn.src[1]);
+    struct ureg_src m00, m01, m10, m11;
+    const int m = tx->insn.dst[0].idx;
+    struct ureg_dst tmp;
+    /*
+     * Bump-env-matrix:
+     * 00 is X
+     * 01 is Y
+     * 10 is Z
+     * 11 is W
+     */
+    nine_info_mark_const_f_used(tx->info, 8 + m);
+    m00 = NINE_CONSTANT_SRC_SWIZZLE(8 + m, X);
+    m01 = NINE_CONSTANT_SRC_SWIZZLE(8 + m, Y);
+    m10 = NINE_CONSTANT_SRC_SWIZZLE(8 + m, Z);
+    m11 = NINE_CONSTANT_SRC_SWIZZLE(8 + m, W);
+    /* dest.r = src0.r + D3DTSS_BUMPENVMAT00(stage n) * src1.r  */
+    ureg_MAD(ureg, ureg_writemask(tmp, TGSI_WRITEMASK_X), m00,
+             NINE_APPLY_SWIZZLE(src1, X), NINE_APPLY_SWIZZLE(src0, X));
+    /* dest.r = dest.r + D3DTSS_BUMPENVMAT10(stage n) * src1.g; */
+    ureg_MAD(ureg, ureg_writemask(tmp, TGSI_WRITEMASK_X), m10,
+             NINE_APPLY_SWIZZLE(src1, Y), NINE_APPLY_SWIZZLE(ureg_src(tmp), X));
+
+    /* dest.g = src0.g + D3DTSS_BUMPENVMAT01(stage n) * src1.r */
+    ureg_MAD(ureg, ureg_writemask(tmp, TGSI_WRITEMASK_Y), m01,
+             NINE_APPLY_SWIZZLE(src1, X), src0);
+    /* dest.g = dest.g + D3DTSS_BUMPENVMAT11(stage n) * src1.g */
+    ureg_MAD(ureg, ureg_writemask(tmp, TGSI_WRITEMASK_Y), m11,
+             NINE_APPLY_SWIZZLE(src1, Y), NINE_APPLY_SWIZZLE(ureg_src(tmp), Y));
+    ureg_MOV(ureg, ureg_writemask(dst, TGSI_WRITEMASK_XY), ureg_src(tmp));
+
+    tx->info->bumpenvmat_needed = 1;
+
+    return D3D_OK;
 }
 
 DECL_SPECIAL(TEXLD)
@@ -2616,7 +2725,7 @@ struct sm1_op_info inst_table[] =
     _OPI(TEX,          TEX, V(0,0), V(0,0), V(1,4), V(1,4), 1, 1, SPECIAL(TEXLD_14)),
     _OPI(TEX,          TEX, V(0,0), V(0,0), V(2,0), V(3,0), 1, 2, SPECIAL(TEXLD)),
     _OPI(TEXBEM,       TEX, V(0,0), V(0,0), V(0,0), V(1,3), 1, 1, SPECIAL(TEXBEM)),
-    _OPI(TEXBEML,      TEX, V(0,0), V(0,0), V(0,0), V(1,3), 1, 1, SPECIAL(TEXBEML)),
+    _OPI(TEXBEML,      TEX, V(0,0), V(0,0), V(0,0), V(1,3), 1, 1, SPECIAL(TEXBEM)),
     _OPI(TEXREG2AR,    TEX, V(0,0), V(0,0), V(0,0), V(1,3), 1, 1, SPECIAL(TEXREG2AR)),
     _OPI(TEXREG2GB,    TEX, V(0,0), V(0,0), V(0,0), V(1,3), 1, 1, SPECIAL(TEXREG2GB)),
     _OPI(TEXM3x2PAD,   TEX, V(0,0), V(0,0), V(0,0), V(1,3), 1, 1, SPECIAL(TEXM3x2PAD)),
@@ -3023,6 +3132,8 @@ tx_ctor(struct shader_translator *tx, struct nine_shader_info *info)
     info->lconstf.data = NULL;
     info->lconstf.ranges = NULL;
 
+    info->bumpenvmat_needed = 0;
+
     for (i = 0; i < Elements(tx->regs.rL); ++i) {
         tx->regs.rL[i] = ureg_dst_undef();
     }
@@ -3233,6 +3344,7 @@ nine_translate_shader(struct NineDevice9 *device, struct nine_shader_info *info)
                        info->const_int_slots > 0 ?
                            max_const_f + info->const_int_slots :
                                info->const_float_slots;
+
     info->const_used_size = sizeof(float[4]) * slot_max; /* slots start from 1 */
 
     for (s = 0; s < slot_max; s++)
diff --git a/src/gallium/state_trackers/nine/nine_shader.h b/src/gallium/state_trackers/nine/nine_shader.h
index ec256c153a9..f2b1e8bf6e2 100644
--- a/src/gallium/state_trackers/nine/nine_shader.h
+++ b/src/gallium/state_trackers/nine/nine_shader.h
@@ -68,6 +68,7 @@ struct nine_shader_info
     unsigned const_bool_slots;
 
     struct nine_lconstf lconstf; /* out, NOTE: members to be free'd by user */
+    uint8_t bumpenvmat_needed;
 };
 
 static inline void
diff --git a/src/gallium/state_trackers/nine/nine_state.c b/src/gallium/state_trackers/nine/nine_state.c
index 4bf5908ebe9..403cd23c0a5 100644
--- a/src/gallium/state_trackers/nine/nine_state.c
+++ b/src/gallium/state_trackers/nine/nine_state.c
@@ -586,6 +586,7 @@ update_ps_constants_userbuf(struct NineDevice9 *device)
     struct nine_state *state = &device->state;
     struct pipe_context *pipe = device->pipe;
     struct pipe_constant_buffer cb;
+    int i;
     cb.buffer = NULL;
     cb.buffer_offset = 0;
     cb.buffer_size = device->state.ps->const_used_size;
@@ -606,6 +607,14 @@ update_ps_constants_userbuf(struct NineDevice9 *device)
         state->changed.ps_const_b = 0;
     }
 
+    /* Upload special constants needed to implement PS1.x instructions like TEXBEM,TEXBEML and BEM */
+    if (device->state.ps->bumpenvmat_needed) {
+        memcpy(device->state.ps_lconstf_temp, cb.user_buffer, cb.buffer_size);
+        memcpy(&device->state.ps_lconstf_temp[4 * 8], &device->state.bumpmap_vars, sizeof(device->state.bumpmap_vars));
+
+        cb.user_buffer = device->state.ps_lconstf_temp;
+    }
+
     pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, 0, &cb);
 
     if (device->state.changed.ps_const_f) {
@@ -1152,6 +1161,7 @@ nine_state_set_defaults(struct NineDevice9 *device, const D3DCAPS9 *caps,
     }
     state->ff.tex_stage[0][D3DTSS_COLOROP] = D3DTOP_MODULATE;
     state->ff.tex_stage[0][D3DTSS_ALPHAOP] = D3DTOP_SELECTARG1;
+    memset(&state->bumpmap_vars, 0, sizeof(state->bumpmap_vars));
 
     for (s = 0; s < Elements(state->samp); ++s) {
         memcpy(&state->samp[s], nine_samp_state_defaults,
diff --git a/src/gallium/state_trackers/nine/nine_state.h b/src/gallium/state_trackers/nine/nine_state.h
index cac9af66532..89a5fd992c4 100644
--- a/src/gallium/state_trackers/nine/nine_state.h
+++ b/src/gallium/state_trackers/nine/nine_state.h
@@ -148,6 +148,8 @@ struct nine_state
     float *ps_const_f;
     int    ps_const_i[NINE_MAX_CONST_I][4];
     BOOL   ps_const_b[NINE_MAX_CONST_B];
+    float *ps_lconstf_temp;
+    uint32_t bumpmap_vars[48];
     uint32_t ps_key;
 
     struct {
diff --git a/src/gallium/state_trackers/nine/pixelshader9.c b/src/gallium/state_trackers/nine/pixelshader9.c
index 3f176a312bf..010c4f4bd28 100644
--- a/src/gallium/state_trackers/nine/pixelshader9.c
+++ b/src/gallium/state_trackers/nine/pixelshader9.c
@@ -72,6 +72,7 @@ NinePixelShader9_ctor( struct NinePixelShader9 *This,
     This->sampler_mask = info.sampler_mask;
     This->rt_mask = info.rt_mask;
     This->const_used_size = info.const_used_size;
+    This->bumpenvmat_needed = info.bumpenvmat_needed;
     /* no constant relative addressing for ps */
     assert(info.lconstf.data == NULL);
     assert(info.lconstf.ranges == NULL);
diff --git a/src/gallium/state_trackers/nine/pixelshader9.h b/src/gallium/state_trackers/nine/pixelshader9.h
index 6dad1d1ee76..9715d9044ce 100644
--- a/src/gallium/state_trackers/nine/pixelshader9.h
+++ b/src/gallium/state_trackers/nine/pixelshader9.h
@@ -41,6 +41,7 @@ struct NinePixelShader9
 
     unsigned const_used_size; /* in bytes */
 
+    uint8_t bumpenvmat_needed;
     uint16_t sampler_mask;
     uint16_t sampler_mask_shadow;
     uint8_t rt_mask;