ac,winsys/amdgpu: align IBs the same as the kernel

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5603>

diff --git a/src/amd/common/ac_gpu_info.c b/src/amd/common/ac_gpu_info.c
index dbf5c930f4616770600f7d19bf691a2b66423ea0..cbc6ed6faa4604954681774be394ac89448378ae 100644 (file)
--- a/src/amd/common/ac_gpu_info.c
+++ b/src/amd/common/ac_gpu_info.c
@@ -561,6 +561,17 @@ bool ac_query_gpu_info(int fd, void *dev_p,
        info->num_rings[RING_VCN_ENC] = util_bitcount(vcn_enc.available_rings);
        info->num_rings[RING_VCN_JPEG] = util_bitcount(vcn_jpeg.available_rings);
 
+       /* This is "align_mask" copied from the kernel, maximums of all IP versions. */
+       info->ib_pad_dw_mask[RING_GFX] = 0xff;
+       info->ib_pad_dw_mask[RING_COMPUTE] = 0xff;
+       info->ib_pad_dw_mask[RING_DMA] = 0xf;
+       info->ib_pad_dw_mask[RING_UVD] = 0xf;
+       info->ib_pad_dw_mask[RING_VCE] = 0x3f;
+       info->ib_pad_dw_mask[RING_UVD_ENC] = 0x3f;
+       info->ib_pad_dw_mask[RING_VCN_DEC] = 0xf;
+       info->ib_pad_dw_mask[RING_VCN_ENC] = 0x3f;
+       info->ib_pad_dw_mask[RING_VCN_JPEG] = 0xf;
+
        /* The mere presence of CLEAR_STATE in the IB causes random GPU hangs
         * on GFX6. Some CLEAR_STATE cause asic hang on radeon kernel, etc.
         * SPI_VS_OUT_CONFIG. So only enable GFX7 CLEAR_STATE on amdgpu kernel.
@@ -682,7 +693,11 @@ bool ac_query_gpu_info(int fd, void *dev_p,
        /* GFX10 and maybe GFX9 need this alignment for cache coherency. */
        if (info->chip_class >= GFX9)
                ib_align = MAX2(ib_align, info->tcc_cache_line_size);
-       assert(ib_align);
+       /* The kernel pads gfx and compute IBs to 256 dwords since:
+        *   66f3b2d527154bd258a57c8815004b5964aa1cf5
+        * Do the same.
+        */
+       ib_align = MAX2(ib_align, 1024);
        info->ib_alignment = ib_align;
 
         if ((info->drm_minor >= 31 &&

diff --git a/src/amd/common/ac_gpu_info.h b/src/amd/common/ac_gpu_info.h
index 3f33ec5ae3262417835fddaa83177adeea24692e..6022a199065a6018e18bc719ba1c5b63bd31115c 100644 (file)
--- a/src/amd/common/ac_gpu_info.h
+++ b/src/amd/common/ac_gpu_info.h
@@ -59,6 +59,7 @@ struct radeon_info {
        /* Features. */
        bool                        has_graphics; /* false if the chip is compute-only */
        uint32_t                    num_rings[NUM_RING_TYPES];
+       uint32_t                    ib_pad_dw_mask[NUM_RING_TYPES];
        bool                        has_clear_state;
        bool                        has_distributed_tess;
        bool                        has_dcc_constant_encode;

diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
index 80eb0f6b7f0c6f744f08a5c17392b6dc7328c447..05e77b03325856c78e43e82c712f6f2a23b6f479 100644 (file)
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
@@ -1097,14 +1097,16 @@ static bool amdgpu_cs_check_space(struct radeon_cmdbuf *rcs, unsigned dw,
    /* This space was originally reserved. */
    rcs->current.max_dw += cs_epilog_dw;
 
-   /* Pad with NOPs and add INDIRECT_BUFFER packet */
-   while ((rcs->current.cdw & 7) != 4)
+   /* Pad with NOPs but leave 4 dwords for INDIRECT_BUFFER. */
+   uint32_t ib_pad_dw_mask = cs->ctx->ws->info.ib_pad_dw_mask[cs->ring_type];
+   while ((rcs->current.cdw & ib_pad_dw_mask) != ib_pad_dw_mask - 3)
       radeon_emit(rcs, PKT3_NOP_PAD);
 
    radeon_emit(rcs, PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0));
    radeon_emit(rcs, va);
    radeon_emit(rcs, va >> 32);
    new_ptr_ib_size = &rcs->current.buf[rcs->current.cdw++];
+   assert((rcs->current.cdw & ib_pad_dw_mask) == 0);
 
    assert((rcs->current.cdw & 7) == 0);
    assert(rcs->current.cdw <= rcs->current.max_dw);
@@ -1664,25 +1666,28 @@ static int amdgpu_cs_flush(struct radeon_cmdbuf *rcs,
    struct amdgpu_cs *cs = amdgpu_cs(rcs);
    struct amdgpu_winsys *ws = cs->ctx->ws;
    int error_code = 0;
+   uint32_t ib_pad_dw_mask = ws->info.ib_pad_dw_mask[cs->ring_type];
 
    rcs->current.max_dw += amdgpu_cs_epilog_dws(cs);
 
+   /* Pad the IB according to the mask. */
    switch (cs->ring_type) {
    case RING_DMA:
-      /* pad DMA ring to 8 DWs */
       if (ws->info.chip_class <= GFX6) {
-         while (rcs->current.cdw & 7)
+         while (rcs->current.cdw & ib_pad_dw_mask)
             radeon_emit(rcs, 0xf0000000); /* NOP packet */
+      } else {
+         while (rcs->current.cdw & ib_pad_dw_mask)
+            radeon_emit(rcs, 0x00000000); /* NOP packet */
       }
       break;
    case RING_GFX:
    case RING_COMPUTE:
-      /* pad GFX ring to 8 DWs to meet CP fetch alignment requirements */
       if (ws->info.gfx_ib_pad_with_type2) {
-         while (rcs->current.cdw & 7)
+         while (rcs->current.cdw & ib_pad_dw_mask)
             radeon_emit(rcs, PKT2_NOP_PAD);
       } else {
-         while (rcs->current.cdw & 7)
+         while (rcs->current.cdw & ib_pad_dw_mask)
             radeon_emit(rcs, PKT3_NOP_PAD);
       }
       if (cs->ring_type == RING_GFX)
@@ -1690,25 +1695,25 @@ static int amdgpu_cs_flush(struct radeon_cmdbuf *rcs,
 
       /* Also pad secondary IBs. */
       if (cs->compute_ib.ib_mapped) {
-         while (cs->compute_ib.base.current.cdw & 7)
+         while (cs->compute_ib.base.current.cdw & ib_pad_dw_mask)
             radeon_emit(&cs->compute_ib.base, PKT3_NOP_PAD);
       }
       break;
    case RING_UVD:
    case RING_UVD_ENC:
-      while (rcs->current.cdw & 15)
+      while (rcs->current.cdw & ib_pad_dw_mask)
          radeon_emit(rcs, 0x80000000); /* type2 nop packet */
       break;
    case RING_VCN_JPEG:
       if (rcs->current.cdw % 2)
          assert(0);
-      while (rcs->current.cdw & 15) {
+      while (rcs->current.cdw & ib_pad_dw_mask) {
          radeon_emit(rcs, 0x60000000); /* nop packet */
          radeon_emit(rcs, 0x00000000);
       }
       break;
    case RING_VCN_DEC:
-      while (rcs->current.cdw & 15)
+      while (rcs->current.cdw & ib_pad_dw_mask)
          radeon_emit(rcs, 0x81ff); /* nop packet */
       break;
    default: