panfrost: Cleanup shader upload code

The old algorithm is still used (and the same issue -- namely, leaking
all shaders -- applies) but we're way more concise about it since we're
*only* using the routine for shaders nowadays; everything else is a
BO-proper or transient.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

diff --git a/src/gallium/drivers/panfrost/pan_allocate.c b/src/gallium/drivers/panfrost/pan_allocate.c
index 3f83e3f29c24e0fd9c6ee3c8da588b3ce79a7462..9d78fab85d6c18ff1de900254fd96335b096f637 100644 (file)
--- a/src/gallium/drivers/panfrost/pan_allocate.c
+++ b/src/gallium/drivers/panfrost/pan_allocate.c
@@ -140,73 +140,15 @@ panfrost_upload_transient(struct panfrost_context *ctx, const void *data, size_t
         return transfer.gpu;
 }
 
-// TODO: An actual allocator, perhaps
-// TODO: Multiple stacks for multiple bases?
-
-int hack_stack_bottom = 4096; /* Don't interfere with constant offsets */
-int last_offset = 0;
-
-static inline int
-pandev_allocate_offset(int *stack, size_t sz)
-{
-        /* First, align the stack bottom to something nice; it's not critical
-         * at this point if we waste a little space to do so. */
-
-        int excess = *stack & (ALIGNMENT - 1);
-
-        /* Add the secret of my */
-        if (excess)
-                *stack += ALIGNMENT - excess;
-
-        /* Finally, use the new bottom for the allocation and move down the
-         * stack */
-
-        int ret = *stack;
-        *stack += sz;
-        return ret;
-}
-
-inline mali_ptr
-pandev_upload(int cheating_offset, int *stack_bottom, mali_ptr base, void *base_map, const void *data, size_t sz, bool no_pad)
-{
-        int offset;
-
-        /* We're not positive about the sizes of all objects, but we don't want
-         * them to crash against each other either. Let the caller disable
-         * padding if they so choose, though. */
-
-        size_t padded_size = no_pad ? sz : sz * 2;
-
-        /* If no specific bottom is specified, use a global one... don't do
-         * this in production, kids */
-
-        if (!stack_bottom)
-                stack_bottom = &hack_stack_bottom;
-
-        /* Allocate space for the new GPU object, if required */
-
-        if (cheating_offset == -1) {
-                offset = pandev_allocate_offset(stack_bottom, padded_size);
-        } else {
-                offset = cheating_offset;
-                *stack_bottom = offset + sz;
-        }
-
-        /* Save last offset for sequential uploads (job descriptors) */
-        last_offset = offset + padded_size;
-
-        /* Upload it */
-        memcpy((uint8_t *) base_map + offset, data, sz);
-
-        /* Return the GPU address */
-        return base + offset;
-}
-
-/* Simplified APIs for the real driver, rather than replays */
+/* The code below is exclusively for the use of shader memory and is subject to
+ * be rewritten soon enough since it never frees the memory it allocates. Here
+ * be dragons, etc. */
 
 mali_ptr
-panfrost_upload(struct panfrost_memory *mem, const void *data, size_t sz, bool no_pad)
+panfrost_upload(struct panfrost_memory *mem, const void *data, size_t sz)
 {
+        sz = ALIGN_POT(sz, ALIGNMENT);
+
         /* Bounds check */
         if ((mem->stack_bottom + sz) >= mem->bo->size) {
                 printf("Out of memory, tried to upload %zd but only %zd available\n",
@@ -214,5 +156,9 @@ panfrost_upload(struct panfrost_memory *mem, const void *data, size_t sz, bool n
                 assert(0);
         }
 
-        return pandev_upload(-1, &mem->stack_bottom, mem->bo->gpu, mem->bo->cpu, data, sz, no_pad);
+        memcpy((uint8_t *) mem->bo->cpu + mem->stack_bottom, data, sz);
+        mali_ptr gpu = mem->bo->gpu + mem->stack_bottom;
+
+        mem->stack_bottom += sz;
+        return gpu;
 }

diff --git a/src/gallium/drivers/panfrost/pan_allocate.h b/src/gallium/drivers/panfrost/pan_allocate.h
index fcb00bfcfd1724ea35035c834f052586b3644548..43f69b4acebfa9f51fa626c24a030fceaa9967a7 100644 (file)
--- a/src/gallium/drivers/panfrost/pan_allocate.h
+++ b/src/gallium/drivers/panfrost/pan_allocate.h
@@ -62,11 +62,8 @@ struct panfrost_memory {
         int stack_bottom;
 };
 
-/* Functions for replay */
-mali_ptr pandev_upload(int cheating_offset, int *stack_bottom, mali_ptr base, void *base_map, const void *data, size_t sz, bool no_pad);
-
 /* Functions for the actual Galliumish driver */
-mali_ptr panfrost_upload(struct panfrost_memory *mem, const void *data, size_t sz, bool no_pad);
+mali_ptr panfrost_upload(struct panfrost_memory *mem, const void *data, size_t sz);
 
 struct panfrost_transfer
 panfrost_allocate_transient(struct panfrost_context *ctx, size_t sz);

diff --git a/src/gallium/drivers/panfrost/pan_assemble.c b/src/gallium/drivers/panfrost/pan_assemble.c
index de8a53ce05d7f7a928d0f500ff97b2257117c4fc..1bf32f1171b23b48bb6ddb3fe662d91a976db6b2 100644 (file)
--- a/src/gallium/drivers/panfrost/pan_assemble.c
+++ b/src/gallium/drivers/panfrost/pan_assemble.c
@@ -77,7 +77,7 @@ panfrost_shader_compile(struct panfrost_context *ctx, struct mali_shader_meta *m
          * I bet someone just thought that would be a cute pun. At least,
          * that's how I'd do it. */
 
-        meta->shader = panfrost_upload(&ctx->shaders, dst, size, true) | program.first_tag;
+        meta->shader = panfrost_upload(&ctx->shaders, dst, size) | program.first_tag;
 
         util_dynarray_fini(&program.compiled);
 

diff --git a/src/gallium/drivers/panfrost/pan_blend_shaders.c b/src/gallium/drivers/panfrost/pan_blend_shaders.c
index 02ebf8ef7fd8347daf723ee6c9b19223214558f5..a58808ba7be3d1247deae358306e3079f52d1aa2 100644 (file)
--- a/src/gallium/drivers/panfrost/pan_blend_shaders.c
+++ b/src/gallium/drivers/panfrost/pan_blend_shaders.c
@@ -178,7 +178,7 @@ panfrost_compile_blend_shader(
         uint8_t *dst = program.compiled.data;
 
         res.shader.cpu = mem_dup(dst, size);
-        res.shader.gpu = panfrost_upload(&ctx->shaders, dst, size, true);
+        res.shader.gpu = panfrost_upload(&ctx->shaders, dst, size);
 
         /* At least two work registers are needed due to an encoding quirk */
         res.work_count = MAX2(program.work_register_count, 2);