freedreno/a4xx: use point-coord helper

[mesa.git] / src / gallium / drivers / freedreno / a4xx / fd4_program.c

diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_program.c b/src/gallium/drivers/freedreno/a4xx/fd4_program.c
index d9ccecb06ad3f3abe7ba06fc165153fcea2cbe08..29d4dad70ed15abc4376bed373a46287e790b367 100644 (file)
--- a/src/gallium/drivers/freedreno/a4xx/fd4_program.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_program.c
@@ -64,7 +64,7 @@ emit_shader(struct fd_ringbuffer *ring, const struct ir3_shader_variant *so)
                OUT_RING(ring, CP_LOAD_STATE4_1_EXT_SRC_ADDR(0) |
                                CP_LOAD_STATE4_1_STATE_TYPE(ST4_SHADER));
        } else {
-               OUT_RELOCD(ring, so->bo, 0,
+               OUT_RELOC(ring, so->bo, 0,
                                CP_LOAD_STATE4_1_STATE_TYPE(ST4_SHADER), 0);
        }
 
@@ -281,7 +281,7 @@ fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit,
                        A4XX_SP_VS_CTRL_REG0_INOUTREGOVERLAP(0) |
                        A4XX_SP_VS_CTRL_REG0_THREADSIZE(TWO_QUADS) |
                        A4XX_SP_VS_CTRL_REG0_SUPERTHREADMODE |
-                       COND(s[VS].v->num_samp > 0, A4XX_SP_VS_CTRL_REG0_PIXLODENABLE));
+                       COND(s[VS].v->need_pixlod, A4XX_SP_VS_CTRL_REG0_PIXLODENABLE));
        OUT_RING(ring, A4XX_SP_VS_CTRL_REG1_CONSTLENGTH(s[VS].constlen) |
                        A4XX_SP_VS_CTRL_REG1_INITIALOUTSTANDING(s[VS].v->total_in));
        OUT_RING(ring, A4XX_SP_VS_PARAM_REG_POSREGID(pos_regid) |
@@ -356,12 +356,12 @@ fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit,
                                A4XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) |
                                A4XX_SP_FS_CTRL_REG0_THREADSIZE(fssz) |
                                A4XX_SP_FS_CTRL_REG0_SUPERTHREADMODE |
-                               COND(s[FS].v->num_samp > 0, A4XX_SP_FS_CTRL_REG0_PIXLODENABLE));
+                               COND(s[FS].v->need_pixlod, A4XX_SP_FS_CTRL_REG0_PIXLODENABLE));
                OUT_RING(ring, A4XX_SP_FS_CTRL_REG1_CONSTLENGTH(s[FS].constlen) |
                                0x80000000 |      /* XXX */
                                COND(s[FS].v->frag_face, A4XX_SP_FS_CTRL_REG1_FACENESS) |
                                COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG1_VARYING) |
-                               COND(s[FS].v->frag_coord, A4XX_SP_FS_CTRL_REG1_FRAGCOORD));
+                               COND(s[FS].v->fragcoord_compmask != 0, A4XX_SP_FS_CTRL_REG1_FRAGCOORD));
 
                OUT_PKT0(ring, REG_A4XX_SP_FS_OBJ_OFFSET_REG, 2);
                OUT_RING(ring, A4XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[FS].constoff) |
@@ -385,10 +385,8 @@ fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit,
        OUT_RING(ring, A4XX_RB_RENDER_CONTROL2_MSAA_SAMPLES(0) |
                        COND(s[FS].v->total_in > 0, A4XX_RB_RENDER_CONTROL2_VARYING) |
                        COND(s[FS].v->frag_face, A4XX_RB_RENDER_CONTROL2_FACENESS) |
-                       COND(s[FS].v->frag_coord, A4XX_RB_RENDER_CONTROL2_XCOORD |
-                                       A4XX_RB_RENDER_CONTROL2_YCOORD |
-                                       A4XX_RB_RENDER_CONTROL2_ZCOORD |
-                                       A4XX_RB_RENDER_CONTROL2_WCOORD));
+                       COND(s[FS].v->fragcoord_compmask != 0,
+                                       A4XX_RB_RENDER_CONTROL2_COORD_MASK(s[FS].v->fragcoord_compmask)));
 
        OUT_PKT0(ring, REG_A4XX_RB_FS_OUTPUT_REG, 1);
        OUT_RING(ring, A4XX_RB_FS_OUTPUT_REG_MRT(nr) |
@@ -465,40 +463,32 @@ fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit,
                                }
                        }
 
-                       gl_varying_slot slot = s[FS].v->inputs[j].slot;
-
-                       /* since we don't enable PIPE_CAP_TGSI_TEXCOORD: */
-                       if (slot >= VARYING_SLOT_VAR0) {
-                               unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
-                               /* Replace the .xy coordinates with S/T from the point sprite. Set
-                                * interpolation bits for .zw such that they become .01
+                       bool coord_mode = emit->sprite_coord_mode;
+                       if (ir3_point_sprite(s[FS].v, j, emit->sprite_coord_enable, &coord_mode)) {
+                               /* mask is two 2-bit fields, where:
+                                *   '01' -> S
+                                *   '10' -> T
+                                *   '11' -> 1 - T  (flip mode)
                                 */
-                               if (emit->sprite_coord_enable & texmask) {
-                                       /* mask is two 2-bit fields, where:
-                                        *   '01' -> S
-                                        *   '10' -> T
-                                        *   '11' -> 1 - T  (flip mode)
-                                        */
-                                       unsigned mask = emit->sprite_coord_mode ? 0b1101 : 0b1001;
-                                       uint32_t loc = inloc;
-                                       if (compmask & 0x1) {
-                                               vpsrepl[loc / 16] |= ((mask >> 0) & 0x3) << ((loc % 16) * 2);
-                                               loc++;
-                                       }
-                                       if (compmask & 0x2) {
-                                               vpsrepl[loc / 16] |= ((mask >> 2) & 0x3) << ((loc % 16) * 2);
-                                               loc++;
-                                       }
-                                       if (compmask & 0x4) {
-                                               /* .z <- 0.0f */
-                                               vinterp[loc / 16] |= 0b10 << ((loc % 16) * 2);
-                                               loc++;
-                                       }
-                                       if (compmask & 0x8) {
-                                               /* .w <- 1.0f */
-                                               vinterp[loc / 16] |= 0b11 << ((loc % 16) * 2);
-                                               loc++;
-                                       }
+                               unsigned mask = coord_mode ? 0b1101 : 0b1001;
+                               uint32_t loc = inloc;
+                               if (compmask & 0x1) {
+                                       vpsrepl[loc / 16] |= ((mask >> 0) & 0x3) << ((loc % 16) * 2);
+                                       loc++;
+                               }
+                               if (compmask & 0x2) {
+                                       vpsrepl[loc / 16] |= ((mask >> 2) & 0x3) << ((loc % 16) * 2);
+                                       loc++;
+                               }
+                               if (compmask & 0x4) {
+                                       /* .z <- 0.0f */
+                                       vinterp[loc / 16] |= 0b10 << ((loc % 16) * 2);
+                                       loc++;
+                               }
+                               if (compmask & 0x8) {
+                                       /* .w <- 1.0f */
+                                       vinterp[loc / 16] |= 0b11 << ((loc % 16) * 2);
+                                       loc++;
                                }
                        }
                }