freedreno/a6xx: convert const emit to OUT_PKT()
authorRob Clark <robdclark@chromium.org>
Wed, 29 Apr 2020 17:09:28 +0000 (10:09 -0700)
committerMarge Bot <eric+marge@anholt.net>
Thu, 30 Apr 2020 20:03:17 +0000 (20:03 +0000)
This is another hot packet.  This splits out each of the four cases
(geom vs frag, and indirect vs inline) intentionally, to avoid some
parity bit calc.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4813>

src/gallium/drivers/freedreno/a6xx/fd6_const.c
src/gallium/drivers/freedreno/a6xx/fd6_emit.h

index 0c519c2577efc3cafd4743c5221188a810fda29f..81fbf67c2df590e672038b47e74040218bb5da9e 100644 (file)
@@ -23,6 +23,7 @@
  */
 
 #include "fd6_const.h"
+#include "fd6_pack.h"
 
 #include "ir3_const.h"
 
@@ -35,43 +36,71 @@ fd6_emit_const(struct fd_ringbuffer *ring, gl_shader_stage type,
                uint32_t regid, uint32_t offset, uint32_t sizedwords,
                const uint32_t *dwords, struct pipe_resource *prsc)
 {
-       uint32_t i, sz, align_sz;
-       enum a6xx_state_src src;
-
-       debug_assert((regid % 4) == 0);
-
-       if (prsc) {
-               sz = 0;
-               src = SS6_INDIRECT;
-       } else {
-               sz = sizedwords;
-               src = SS6_DIRECT;
-       }
-
-       align_sz = align(sz, 4);
-
-       OUT_PKT7(ring, fd6_stage2opcode(type), 3 + align_sz);
-       OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(regid/4) |
-                       CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-                       CP_LOAD_STATE6_0_STATE_SRC(src) |
-                       CP_LOAD_STATE6_0_STATE_BLOCK(fd6_stage2shadersb(type)) |
-                       CP_LOAD_STATE6_0_NUM_UNIT(DIV_ROUND_UP(sizedwords, 4)));
        if (prsc) {
                struct fd_bo *bo = fd_resource(prsc)->bo;
-               OUT_RELOC(ring, bo, offset, 0, 0);
+
+               if (fd6_geom_stage(type)) {
+                       OUT_PKT(ring, CP_LOAD_STATE6_GEOM,
+                                       CP_LOAD_STATE6_0(
+                                                       .dst_off     = regid/4,
+                                                       .state_type  = ST6_CONSTANTS,
+                                                       .state_src   = SS6_INDIRECT,
+                                                       .state_block = fd6_stage2shadersb(type),
+                                                       .num_unit    = DIV_ROUND_UP(sizedwords, 4)
+                                               ),
+                                       CP_LOAD_STATE6_EXT_SRC_ADDR(
+                                                       .bo          = bo,
+                                                       .bo_offset   = offset
+                                               )
+                               );
+               } else {
+                       OUT_PKT(ring, CP_LOAD_STATE6_FRAG,
+                                       CP_LOAD_STATE6_0(
+                                                       .dst_off     = regid/4,
+                                                       .state_type  = ST6_CONSTANTS,
+                                                       .state_src   = SS6_INDIRECT,
+                                                       .state_block = fd6_stage2shadersb(type),
+                                                       .num_unit    = DIV_ROUND_UP(sizedwords, 4)
+                                               ),
+                                       CP_LOAD_STATE6_EXT_SRC_ADDR(
+                                                       .bo          = bo,
+                                                       .bo_offset   = offset
+                                               )
+                               );
+               }
        } else {
-               OUT_RING(ring, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
-               OUT_RING(ring, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
+               /* NOTE we cheat a bit here, since we know mesa is aligning
+                * the size of the user buffer to 16 bytes.  And we want to
+                * cut cycles in a hot path.
+                */
+               uint32_t align_sz = align(sizedwords, 4);
                dwords = (uint32_t *)&((uint8_t *)dwords)[offset];
-       }
 
-       for (i = 0; i < sz; i++) {
-               OUT_RING(ring, dwords[i]);
-       }
-
-       /* Zero-pad to multiple of 4 dwords */
-       for (i = sz; i < align_sz; i++) {
-               OUT_RING(ring, 0);
+               if (fd6_geom_stage(type)) {
+                       OUT_PKTBUF(ring, CP_LOAD_STATE6_GEOM, dwords, align_sz,
+                                       CP_LOAD_STATE6_0(
+                                                       .dst_off     = regid/4,
+                                                       .state_type  = ST6_CONSTANTS,
+                                                       .state_src   = SS6_DIRECT,
+                                                       .state_block = fd6_stage2shadersb(type),
+                                                       .num_unit    = DIV_ROUND_UP(sizedwords, 4)
+                                               ),
+                                       CP_LOAD_STATE6_1(),
+                                       CP_LOAD_STATE6_2()
+                               );
+               } else {
+                       OUT_PKTBUF(ring, CP_LOAD_STATE6_FRAG, dwords, align_sz,
+                                       CP_LOAD_STATE6_0(
+                                                       .dst_off     = regid/4,
+                                                       .state_type  = ST6_CONSTANTS,
+                                                       .state_src   = SS6_DIRECT,
+                                                       .state_block = fd6_stage2shadersb(type),
+                                                       .num_unit    = DIV_ROUND_UP(sizedwords, 4)
+                                               ),
+                                       CP_LOAD_STATE6_1(),
+                                       CP_LOAD_STATE6_2()
+                               );
+               }
        }
 }
 
index 32968c819d5a63eaf44fa43022a3020b5f5ef82b..2979e750908a3fd68f01b998868820c89236e3d1 100644 (file)
@@ -210,24 +210,30 @@ fd6_emit_lrz_flush(struct fd_ringbuffer *ring)
        OUT_RING(ring, LRZ_FLUSH);
 }
 
-static inline uint32_t
-fd6_stage2opcode(gl_shader_stage type)
+static inline bool
+fd6_geom_stage(gl_shader_stage type)
 {
        switch (type) {
        case MESA_SHADER_VERTEX:
        case MESA_SHADER_TESS_CTRL:
        case MESA_SHADER_TESS_EVAL:
        case MESA_SHADER_GEOMETRY:
-               return CP_LOAD_STATE6_GEOM;
+               return true;
        case MESA_SHADER_FRAGMENT:
        case MESA_SHADER_COMPUTE:
        case MESA_SHADER_KERNEL:
-               return CP_LOAD_STATE6_FRAG;
+               return false;
        default:
                unreachable("bad shader type");
        }
 }
 
+static inline uint32_t
+fd6_stage2opcode(gl_shader_stage type)
+{
+       return fd6_geom_stage(type) ? CP_LOAD_STATE6_GEOM : CP_LOAD_STATE6_FRAG;
+}
+
 static inline enum a6xx_state_block
 fd6_stage2shadersb(gl_shader_stage type)
 {