X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fcell%2Fspu%2Fspu_tri.c;h=58be001be4cf32c2131b1cae686d0374c81d471e;hb=1036ef2bf468611d37b5df06fc4424f2002e3837;hp=5f908159bbf3a896e004cff403ceec60f96ef3b5;hpb=711f8a1dd94e2e1e715615d947e03015ef972326;p=mesa.git diff --git a/src/gallium/drivers/cell/spu/spu_tri.c b/src/gallium/drivers/cell/spu/spu_tri.c index 5f908159bbf..58be001be4c 100644 --- a/src/gallium/drivers/cell/spu/spu_tri.c +++ b/src/gallium/drivers/cell/spu/spu_tri.c @@ -29,12 +29,12 @@ * Triangle rendering within a tile. */ -#include #include "pipe/p_compiler.h" #include "pipe/p_format.h" #include "util/u_math.h" #include "spu_colorpack.h" #include "spu_main.h" +#include "spu_shuffle.h" #include "spu_texture.h" #include "spu_tile.h" #include "spu_tri.h" @@ -56,7 +56,7 @@ struct vertex_header { /* XXX fix this */ #undef CEILF -#define CEILF(X) ((float) (int) ((X) + 0.99999)) +#define CEILF(X) ((float) (int) ((X) + 0.99999f)) #define QUAD_TOP_LEFT 0 @@ -70,14 +70,25 @@ struct vertex_header { #define MASK_ALL 0xf +#define CHAN0 0 +#define CHAN1 1 +#define CHAN2 2 +#define CHAN3 3 + + #define DEBUG_VERTS 0 /** * Triangle edge info */ struct edge { - float dx; /**< X(v1) - X(v0), used only during setup */ - float dy; /**< Y(v1) - Y(v0), used only during setup */ + union { + struct { + float dx; /**< X(v1) - X(v0), used only during setup */ + float dy; /**< Y(v1) - Y(v0), used only during setup */ + }; + vec_float4 ds; /**< vector accessor for dx and dy */ + }; float dxdy; /**< dx/dy */ float sx, sy; /**< first sample point coord */ int lines; /**< number of lines on this edge */ @@ -102,10 +113,15 @@ struct setup_stage { * turn. Currently fixed at 4 floats, but should change in time. * Codegen will help cope with this. */ - const struct vertex_header *vmax; - const struct vertex_header *vmid; - const struct vertex_header *vmin; - const struct vertex_header *vprovoke; + union { + struct { + const struct vertex_header *vmin; + const struct vertex_header *vmid; + const struct vertex_header *vmax; + const struct vertex_header *vprovoke; + }; + qword vertex_headers; + }; struct edge ebot; struct edge etop; @@ -117,13 +133,20 @@ struct setup_stage { uint tx, ty; /**< position of current tile (x, y) */ - int cliprect_minx, cliprect_maxx, cliprect_miny, cliprect_maxy; + union { + struct { + int cliprect_minx; + int cliprect_miny; + int cliprect_maxx; + int cliprect_maxy; + }; + qword cliprect; + }; struct interp_coef coef[PIPE_MAX_SHADER_INPUTS]; struct { - int left[2]; /**< [0] = row0, [1] = row1 */ - int right[2]; + vec_int4 quad; /**< [0] = row0, [1] = row1; {left[0],left[1],right[0],right[1]} */ int y; unsigned y_flags; unsigned mask; /**< mask of MASK_BOTTOM/TOP_LEFT/RIGHT bits */ @@ -134,99 +157,97 @@ struct setup_stage { static struct setup_stage setup; -/** - * Evaluate attribute coefficients (plane equations) to compute - * attribute values for the four fragments in a quad. - * Eg: four colors will be computed (in AoS format). - */ -static INLINE void -eval_coeff(uint slot, float x, float y, vector float w, vector float result[4]) +static INLINE vector float +splatx(vector float v) { - switch (spu.vertex_info.attrib[slot].interp_mode) { - case INTERP_CONSTANT: - result[QUAD_TOP_LEFT] = - result[QUAD_TOP_RIGHT] = - result[QUAD_BOTTOM_LEFT] = - result[QUAD_BOTTOM_RIGHT] = setup.coef[slot].a0; - break; - case INTERP_LINEAR: - { - vector float dadx = setup.coef[slot].dadx; - vector float dady = setup.coef[slot].dady; - vector float topLeft = - spu_add(setup.coef[slot].a0, - spu_add(spu_mul(spu_splats(x), dadx), - spu_mul(spu_splats(y), dady))); - - result[QUAD_TOP_LEFT] = topLeft; - result[QUAD_TOP_RIGHT] = spu_add(topLeft, dadx); - result[QUAD_BOTTOM_LEFT] = spu_add(topLeft, dady); - result[QUAD_BOTTOM_RIGHT] = spu_add(spu_add(topLeft, dadx), dady); - } - break; - case INTERP_PERSPECTIVE: - { - vector float dadx = setup.coef[slot].dadx; - vector float dady = setup.coef[slot].dady; - vector float topLeft = - spu_add(setup.coef[slot].a0, - spu_add(spu_mul(spu_splats(x), dadx), - spu_mul(spu_splats(y), dady))); - - vector float wInv = spu_re(w); /* 1.0 / w */ - - result[QUAD_TOP_LEFT] = spu_mul(topLeft, wInv); - result[QUAD_TOP_RIGHT] = spu_mul(spu_add(topLeft, dadx), wInv); - result[QUAD_BOTTOM_LEFT] = spu_mul(spu_add(topLeft, dady), wInv); - result[QUAD_BOTTOM_RIGHT] = spu_mul(spu_add(spu_add(topLeft, dadx), dady), wInv); - } - break; - case INTERP_POS: - case INTERP_NONE: - break; - default: - ASSERT(0); - } + return spu_splats(spu_extract(v, CHAN0)); } - -/** - * As above, but return 4 vectors in SOA format. - * XXX this will all be re-written someday. - */ -static INLINE void -eval_coeff_soa(uint slot, float x, float y, vector float w, vector float result[4]) +static INLINE vector float +splaty(vector float v) { - eval_coeff(slot, x, y, w, result); - _transpose_matrix4x4(result, result); + return spu_splats(spu_extract(v, CHAN1)); } +static INLINE vector float +splatz(vector float v) +{ + return spu_splats(spu_extract(v, CHAN2)); +} -/** Evalute coefficients to get Z for four pixels in a quad */ static INLINE vector float -eval_z(float x, float y) +splatw(vector float v) { - const uint slot = 0; - const float dzdx = spu_extract(setup.coef[slot].dadx, 2); - const float dzdy = spu_extract(setup.coef[slot].dady, 2); - const float topLeft = spu_extract(setup.coef[slot].a0, 2) + x * dzdx + y * dzdy; - const vector float topLeftv = spu_splats(topLeft); - const vector float derivs = (vector float) { 0.0, dzdx, dzdy, dzdx + dzdy }; - return spu_add(topLeftv, derivs); + return spu_splats(spu_extract(v, CHAN3)); } -/** Evalute coefficients to get W for four pixels in a quad */ -static INLINE vector float -eval_w(float x, float y) +/** + * Setup fragment shader inputs by evaluating triangle's vertex + * attribute coefficient info. + * \param x quad x pos + * \param y quad y pos + * \param fragZ returns quad Z values + * \param fragInputs returns fragment program inputs + * Note: this code could be incorporated into the fragment program + * itself to avoid the loop and switch. + */ +static void +eval_inputs(float x, float y, vector float *fragZ, vector float fragInputs[]) { - const uint slot = 0; - const float dwdx = spu_extract(setup.coef[slot].dadx, 3); - const float dwdy = spu_extract(setup.coef[slot].dady, 3); - const float topLeft = spu_extract(setup.coef[slot].a0, 3) + x * dwdx + y * dwdy; - const vector float topLeftv = spu_splats(topLeft); - const vector float derivs = (vector float) { 0.0, dwdx, dwdy, dwdx + dwdy }; - return spu_add(topLeftv, derivs); + static const vector float deltaX = (const vector float) {0, 1, 0, 1}; + static const vector float deltaY = (const vector float) {0, 0, 1, 1}; + + const uint posSlot = 0; + const vector float pos = setup.coef[posSlot].a0; + const vector float dposdx = setup.coef[posSlot].dadx; + const vector float dposdy = setup.coef[posSlot].dady; + const vector float fragX = spu_splats(x) + deltaX; + const vector float fragY = spu_splats(y) + deltaY; + vector float fragW, wInv; + uint i; + + *fragZ = splatz(pos) + fragX * splatz(dposdx) + fragY * splatz(dposdy); + fragW = splatw(pos) + fragX * splatw(dposdx) + fragY * splatw(dposdy); + wInv = spu_re(fragW); /* 1 / w */ + + /* loop over fragment program inputs */ + for (i = 0; i < spu.vertex_info.num_attribs; i++) { + uint attr = i + 1; + enum interp_mode interp = spu.vertex_info.attrib[attr].interp_mode; + + /* constant term */ + vector float a0 = setup.coef[attr].a0; + vector float r0 = splatx(a0); + vector float r1 = splaty(a0); + vector float r2 = splatz(a0); + vector float r3 = splatw(a0); + + if (interp == INTERP_LINEAR || interp == INTERP_PERSPECTIVE) { + /* linear term */ + vector float dadx = setup.coef[attr].dadx; + vector float dady = setup.coef[attr].dady; + /* Use SPU intrinsics here to get slightly better code. + * originally: r0 += fragX * splatx(dadx) + fragY * splatx(dady); + */ + r0 = spu_madd(fragX, splatx(dadx), spu_madd(fragY, splatx(dady), r0)); + r1 = spu_madd(fragX, splaty(dadx), spu_madd(fragY, splaty(dady), r1)); + r2 = spu_madd(fragX, splatz(dadx), spu_madd(fragY, splatz(dady), r2)); + r3 = spu_madd(fragX, splatw(dadx), spu_madd(fragY, splatw(dady), r3)); + if (interp == INTERP_PERSPECTIVE) { + /* perspective term */ + r0 *= wInv; + r1 *= wInv; + r2 *= wInv; + r3 *= wInv; + } + } + fragInputs[CHAN0] = r0; + fragInputs[CHAN1] = r1; + fragInputs[CHAN2] = r2; + fragInputs[CHAN3] = r3; + fragInputs += 4; + } } @@ -252,19 +273,11 @@ emit_quad( int x, int y, mask_t mask) * Run fragment shader, execute per-fragment ops, update fb/tile. */ vector float inputs[4*4], outputs[2*4]; - vector float fragZ = eval_z((float) x, (float) y); - vector float fragW = eval_w((float) x, (float) y); vector unsigned int kill_mask; + vector float fragZ; + + eval_inputs((float) x, (float) y, &fragZ, inputs); - /* setup inputs */ -#if 0 - eval_coeff_soa(1, (float) x, (float) y, fragW, inputs); -#else - uint i; - for (i = 0; i < spu.vertex_info.num_attribs; i++) { - eval_coeff_soa(i+1, (float) x, (float) y, fragW, inputs + i * 4); - } -#endif ASSERT(spu.fragment_program); ASSERT(spu.fragment_ops); @@ -275,15 +288,20 @@ emit_quad( int x, int y, mask_t mask) /* Execute per-fragment/quad operations, including: * alpha test, z test, stencil test, blend and framebuffer writing. + * Note that there are two different fragment operations functions + * that can be called, one for front-facing fragments, and one + * for back-facing fragments. (Often the two are the same; + * but in some cases, like two-sided stenciling, they can be + * very different.) So choose the correct function depending + * on the calculated facing. */ - spu.fragment_ops(ix, iy, &spu.ctile, &spu.ztile, + spu.fragment_ops[setup.facing](ix, iy, &spu.ctile, &spu.ztile, fragZ, outputs[0*4+0], outputs[0*4+1], outputs[0*4+2], outputs[0*4+3], - mask, - setup.facing); + mask); } } } @@ -300,27 +318,6 @@ block(int x) } -/** - * Compute mask which indicates which pixels in the 2x2 quad are actually inside - * the triangle's bounds. - * The mask is a uint4 vector and each element will be 0 or 0xffffffff. - */ -static INLINE mask_t -calculate_mask(int x) -{ - /* This is a little tricky. - * Use & instead of && to avoid branches. - * Use negation to convert true/false to ~0/0 values. - */ - mask_t mask; - mask = spu_insert(-((x >= setup.span.left[0]) & (x < setup.span.right[0])), mask, 0); - mask = spu_insert(-((x+1 >= setup.span.left[0]) & (x+1 < setup.span.right[0])), mask, 1); - mask = spu_insert(-((x >= setup.span.left[1]) & (x < setup.span.right[1])), mask, 2); - mask = spu_insert(-((x+1 >= setup.span.left[1]) & (x+1 < setup.span.right[1])), mask, 3); - return mask; -} - - /** * Render a horizontal span of quads */ @@ -328,25 +325,29 @@ static void flush_spans(void) { int minleft, maxright; - int x; + + const int l0 = spu_extract(setup.span.quad, 0); + const int l1 = spu_extract(setup.span.quad, 1); + const int r0 = spu_extract(setup.span.quad, 2); + const int r1 = spu_extract(setup.span.quad, 3); switch (setup.span.y_flags) { case 0x3: /* both odd and even lines written (both quad rows) */ - minleft = MIN2(setup.span.left[0], setup.span.left[1]); - maxright = MAX2(setup.span.right[0], setup.span.right[1]); + minleft = MIN2(l0, l1); + maxright = MAX2(r0, r1); break; case 0x1: /* only even line written (quad top row) */ - minleft = setup.span.left[0]; - maxright = setup.span.right[0]; + minleft = l0; + maxright = r0; break; case 0x2: /* only odd line written (quad bottom row) */ - minleft = setup.span.left[1]; - maxright = setup.span.right[1]; + minleft = l1; + maxright = r1; break; default: @@ -384,17 +385,42 @@ flush_spans(void) ASSERT(spu.cur_ztile_status != TILE_STATUS_DEFINED); } - /* XXX this loop could be moved into the above switch cases and - * calculate_mask() could be simplified a bit... - */ - for (x = block(minleft); x <= block(maxright); x += 2) { - emit_quad( x, setup.span.y, calculate_mask( x )); + /* XXX this loop could be moved into the above switch cases... */ + + /* Setup for mask calculation */ + const vec_int4 quad_LlRr = setup.span.quad; + const vec_int4 quad_RrLl = spu_rlqwbyte(quad_LlRr, 8); + const vec_int4 quad_LLll = spu_shuffle(quad_LlRr, quad_LlRr, SHUFFLE4(A,A,B,B)); + const vec_int4 quad_RRrr = spu_shuffle(quad_RrLl, quad_RrLl, SHUFFLE4(A,A,B,B)); + + const vec_int4 twos = spu_splats(2); + + const int x = block(minleft); + vec_int4 xs = {x, x+1, x, x+1}; + + for (; spu_extract(xs, 0) <= block(maxright); xs += twos) { + /** + * Computes mask to indicate which pixels in the 2x2 quad are actually + * inside the triangle's bounds. + */ + + /* Calculate ({x,x+1,x,x+1} >= {l[0],l[0],l[1],l[1]}) */ + const mask_t gt_LLll_xs = spu_cmpgt(quad_LLll, xs); + const mask_t gte_xs_LLll = spu_nand(gt_LLll_xs, gt_LLll_xs); + + /* Calculate ({r[0],r[0],r[1],r[1]} > {x,x+1,x,x+1}) */ + const mask_t gt_RRrr_xs = spu_cmpgt(quad_RRrr, xs); + + /* Combine results to create mask */ + const mask_t mask = spu_and(gte_xs_LLll, gt_RRrr_xs); + + emit_quad(spu_extract(xs, 0), setup.span.y, mask); } setup.span.y = 0; setup.span.y_flags = 0; - setup.span.right[0] = 0; - setup.span.right[1] = 0; + /* Zero right elements */ + setup.span.quad = spu_shuffle(setup.span.quad, setup.span.quad, SHUFFLE4(A,B,0,0)); } @@ -414,6 +440,41 @@ print_vertex(const struct vertex_header *v) } #endif +/* Returns the minimum of each slot of two vec_float4s as qwords. + * i.e. return[n] = min(q0[n],q1[n]); + */ +static qword +minfq(qword q0, qword q1) +{ + const qword q0q1m = si_fcgt(q0, q1); + return si_selb(q0, q1, q0q1m); +} + +/* Returns the minimum of each slot of three vec_float4s as qwords. + * i.e. return[n] = min(q0[n],q1[n],q2[n]); + */ +static qword +min3fq(qword q0, qword q1, qword q2) +{ + return minfq(minfq(q0, q1), q2); +} + +/* Returns the maximum of each slot of two vec_float4s as qwords. + * i.e. return[n] = min(q0[n],q1[n],q2[n]); + */ +static qword +maxfq(qword q0, qword q1) { + const qword q0q1m = si_fcgt(q0, q1); + return si_selb(q1, q0, q0q1m); +} + +/* Returns the maximum of each slot of three vec_float4s as qwords. + * i.e. return[n] = min(q0[n],q1[n],q2[n]); + */ +static qword +max3fq(qword q0, qword q1, qword q2) { + return maxfq(maxfq(q0, q1), q2); +} /** * Sort vertices from top to bottom. @@ -422,9 +483,7 @@ print_vertex(const struct vertex_header *v) * \return FALSE if tri is totally outside tile, TRUE otherwise */ static boolean -setup_sort_vertices(const struct vertex_header *v0, - const struct vertex_header *v1, - const struct vertex_header *v2) +setup_sort_vertices(const qword vs) { float area, sign; @@ -437,79 +496,60 @@ setup_sort_vertices(const struct vertex_header *v0, } #endif - /* determine bottom to top order of vertices */ { - float y0 = spu_extract(v0->data[0], 1); - float y1 = spu_extract(v1->data[0], 1); - float y2 = spu_extract(v2->data[0], 1); - if (y0 <= y1) { - if (y1 <= y2) { - /* y0<=y1<=y2 */ - setup.vmin = v0; - setup.vmid = v1; - setup.vmax = v2; - sign = -1.0f; - } - else if (y2 <= y0) { - /* y2<=y0<=y1 */ - setup.vmin = v2; - setup.vmid = v0; - setup.vmax = v1; - sign = -1.0f; - } - else { - /* y0<=y2<=y1 */ - setup.vmin = v0; - setup.vmid = v2; - setup.vmax = v1; - sign = 1.0f; - } - } - else { - if (y0 <= y2) { - /* y1<=y0<=y2 */ - setup.vmin = v1; - setup.vmid = v0; - setup.vmax = v2; - sign = 1.0f; - } - else if (y2 <= y1) { - /* y2<=y1<=y0 */ - setup.vmin = v2; - setup.vmid = v1; - setup.vmax = v0; - sign = 1.0f; - } - else { - /* y1<=y2<=y0 */ - setup.vmin = v1; - setup.vmid = v2; - setup.vmax = v0; - sign = -1.0f; - } - } + /* Load the float values for various processing... */ + const qword f0 = (qword)(((const struct vertex_header*)si_to_ptr(vs))->data[0]); + const qword f1 = (qword)(((const struct vertex_header*)si_to_ptr(si_rotqbyi(vs, 4)))->data[0]); + const qword f2 = (qword)(((const struct vertex_header*)si_to_ptr(si_rotqbyi(vs, 8)))->data[0]); + + /* Check if triangle is completely outside the tile bounds + * Find the min and max x and y positions of the three poits */ + const qword minf = min3fq(f0, f1, f2); + const qword maxf = max3fq(f0, f1, f2); + + /* Compare min and max against cliprect vals */ + const qword maxsmins = si_shufb(maxf, minf, SHUFB4(A,B,a,b)); + const qword outside = si_fcgt(maxsmins, si_csflt(setup.cliprect, 0)); + + /* Use a little magic to work out of the tri is visible or not */ + if(si_to_uint(si_xori(si_gb(outside), 0xc))) return FALSE; + + /* determine bottom to top order of vertices */ + /* A table of shuffle patterns for putting vertex_header pointers into + correct order. Quite magical. */ + const qword sort_order_patterns[] = { + SHUFB4(A,B,C,C), + SHUFB4(C,A,B,C), + SHUFB4(A,C,B,C), + SHUFB4(B,C,A,C), + SHUFB4(B,A,C,C), + SHUFB4(C,B,A,C) }; + + /* Collate y values into two vectors for comparison. + Using only one shuffle constant! ;) */ + const qword y_02_ = si_shufb(f0, f2, SHUFB4(0,B,b,C)); + const qword y_10_ = si_shufb(f1, f0, SHUFB4(0,B,b,C)); + const qword y_012 = si_shufb(y_02_, f1, SHUFB4(0,B,b,C)); + const qword y_120 = si_shufb(y_10_, f2, SHUFB4(0,B,b,C)); + + /* Perform comparison: {y0,y1,y2} > {y1,y2,y0} */ + const qword compare = si_fcgt(y_012, y_120); + /* Compress the result of the comparison into 4 bits */ + const qword gather = si_gb(compare); + /* Subtract one to attain the index into the LUT. Magical. */ + const unsigned int index = si_to_uint(gather) - 1; + + /* Load the appropriate pattern and construct the desired vector. */ + setup.vertex_headers = si_shufb(vs, vs, sort_order_patterns[index]); + + /* Using the result of the comparison, set sign. + Very magical. */ + sign = ((si_to_uint(si_cntb(gather)) == 2) ? 1.0f : -1.0f); } - /* Check if triangle is completely outside the tile bounds */ - if (spu_extract(setup.vmin->data[0], 1) > setup.cliprect_maxy) - return FALSE; - if (spu_extract(setup.vmax->data[0], 1) < setup.cliprect_miny) - return FALSE; - if (spu_extract(setup.vmin->data[0], 0) < setup.cliprect_minx && - spu_extract(setup.vmid->data[0], 0) < setup.cliprect_minx && - spu_extract(setup.vmax->data[0], 0) < setup.cliprect_minx) - return FALSE; - if (spu_extract(setup.vmin->data[0], 0) > setup.cliprect_maxx && - spu_extract(setup.vmid->data[0], 0) > setup.cliprect_maxx && - spu_extract(setup.vmax->data[0], 0) > setup.cliprect_maxx) - return FALSE; - - setup.ebot.dx = spu_extract(setup.vmid->data[0], 0) - spu_extract(setup.vmin->data[0], 0); - setup.ebot.dy = spu_extract(setup.vmid->data[0], 1) - spu_extract(setup.vmin->data[0], 1); - setup.emaj.dx = spu_extract(setup.vmax->data[0], 0) - spu_extract(setup.vmin->data[0], 0); - setup.emaj.dy = spu_extract(setup.vmax->data[0], 1) - spu_extract(setup.vmin->data[0], 1); - setup.etop.dx = spu_extract(setup.vmax->data[0], 0) - spu_extract(setup.vmid->data[0], 0); - setup.etop.dy = spu_extract(setup.vmax->data[0], 1) - spu_extract(setup.vmid->data[0], 1); + setup.ebot.ds = spu_sub(setup.vmid->data[0], setup.vmin->data[0]); + setup.emaj.ds = spu_sub(setup.vmax->data[0], setup.vmin->data[0]); + setup.etop.ds = spu_sub(setup.vmax->data[0], setup.vmid->data[0]); /* * Compute triangle's area. Use 1/area to compute partial @@ -519,12 +559,17 @@ setup_sort_vertices(const struct vertex_header *v0, setup.oneOverArea = 1.0f / area; - /* The product of area * sign indicates front/back orientation (0/1) */ + /* The product of area * sign indicates front/back orientation (0/1). + * Just in case someone gets the bright idea of switching the front + * and back constants without noticing that we're assuming their + * values in this operation, also assert that the values are + * what we think they are. + */ + ASSERT(CELL_FACING_FRONT == 0); + ASSERT(CELL_FACING_BACK == 1); setup.facing = (area * sign > 0.0f) ^ (spu.rasterizer.front_winding == PIPE_WINDING_CW); - setup.vprovoke = v2; - return TRUE; } @@ -734,9 +779,11 @@ subtriangle(struct edge *eleft, struct edge *eright, unsigned lines) setup.span.y = block(_y); } - setup.span.left[_y&1] = left; - setup.span.right[_y&1] = right; - setup.span.y_flags |= 1<<(_y&1); + int offset = _y&1; + vec_int4 quad_LlRr = {left, left, right, right}; + /* Store left and right in 0 or 1 row of quad based on offset */ + setup.span.quad = spu_sel(quad_LlRr, setup.span.quad, spu_maskw(5<