@@ -1,5 +1,5 @@
/*
- * Copyright(c) 2019-2021 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ * Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
@@ -40,47 +40,6 @@ typedef long HVX_VectorPair __attribute__((__vector_size__(256)))
typedef long HVX_VectorPred __attribute__((__vector_size__(128)))
__attribute__((aligned(128)));
-#define VSCATTER_16(BASE, RGN, OFF, VALS) \
- __builtin_HEXAGON_V6_vscattermh_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_MASKED(MASK, BASE, RGN, OFF, VALS) \
- __builtin_HEXAGON_V6_vscattermhq_128B(MASK, (int)BASE, RGN, OFF, VALS)
-#define VSCATTER_32(BASE, RGN, OFF, VALS) \
- __builtin_HEXAGON_V6_vscattermw_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_32_MASKED(MASK, BASE, RGN, OFF, VALS) \
- __builtin_HEXAGON_V6_vscattermwq_128B(MASK, (int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_32(BASE, RGN, OFF, VALS) \
- __builtin_HEXAGON_V6_vscattermhw_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_32_MASKED(MASK, BASE, RGN, OFF, VALS) \
- __builtin_HEXAGON_V6_vscattermhwq_128B(MASK, (int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_ACC(BASE, RGN, OFF, VALS) \
- __builtin_HEXAGON_V6_vscattermh_add_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_32_ACC(BASE, RGN, OFF, VALS) \
- __builtin_HEXAGON_V6_vscattermw_add_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_32_ACC(BASE, RGN, OFF, VALS) \
- __builtin_HEXAGON_V6_vscattermhw_add_128B((int)BASE, RGN, OFF, VALS)
-
-#define VGATHER_16(DSTADDR, BASE, RGN, OFF) \
- __builtin_HEXAGON_V6_vgathermh_128B(DSTADDR, (int)BASE, RGN, OFF)
-#define VGATHER_16_MASKED(DSTADDR, MASK, BASE, RGN, OFF) \
- __builtin_HEXAGON_V6_vgathermhq_128B(DSTADDR, MASK, (int)BASE, RGN, OFF)
-#define VGATHER_32(DSTADDR, BASE, RGN, OFF) \
- __builtin_HEXAGON_V6_vgathermw_128B(DSTADDR, (int)BASE, RGN, OFF)
-#define VGATHER_32_MASKED(DSTADDR, MASK, BASE, RGN, OFF) \
- __builtin_HEXAGON_V6_vgathermwq_128B(DSTADDR, MASK, (int)BASE, RGN, OFF)
-#define VGATHER_16_32(DSTADDR, BASE, RGN, OFF) \
- __builtin_HEXAGON_V6_vgathermhw_128B(DSTADDR, (int)BASE, RGN, OFF)
-#define VGATHER_16_32_MASKED(DSTADDR, MASK, BASE, RGN, OFF) \
- __builtin_HEXAGON_V6_vgathermhwq_128B(DSTADDR, MASK, (int)BASE, RGN, OFF)
-
-#define VSHUFF_H(V) \
- __builtin_HEXAGON_V6_vshuffh_128B(V)
-#define VSPLAT_H(X) \
- __builtin_HEXAGON_V6_lvsplath_128B(X)
-#define VAND_VAL(PRED, VAL) \
- __builtin_HEXAGON_V6_vandvrt_128B(PRED, VAL)
-#define VDEAL_H(V) \
- __builtin_HEXAGON_V6_vdealh_128B(V)
-
int err;
/* define the number of rows/cols in a square matrix */
@@ -108,22 +67,22 @@ unsigned short vscatter16_32_ref[SCATTER_BUFFER_SIZE];
unsigned short vgather16_32_ref[MATRIX_SIZE];
/* declare the arrays of offsets */
-unsigned short half_offsets[MATRIX_SIZE];
-unsigned int word_offsets[MATRIX_SIZE];
+unsigned short half_offsets[MATRIX_SIZE] __attribute__((aligned(128)));
+unsigned int word_offsets[MATRIX_SIZE] __attribute__((aligned(128)));
/* declare the arrays of values */
-unsigned short half_values[MATRIX_SIZE];
-unsigned short half_values_acc[MATRIX_SIZE];
-unsigned short half_values_masked[MATRIX_SIZE];
-unsigned int word_values[MATRIX_SIZE];
-unsigned int word_values_acc[MATRIX_SIZE];
-unsigned int word_values_masked[MATRIX_SIZE];
+unsigned short half_values[MATRIX_SIZE] __attribute__((aligned(128)));
+unsigned short half_values_acc[MATRIX_SIZE] __attribute__((aligned(128)));
+unsigned short half_values_masked[MATRIX_SIZE] __attribute__((aligned(128)));
+unsigned int word_values[MATRIX_SIZE] __attribute__((aligned(128)));
+unsigned int word_values_acc[MATRIX_SIZE] __attribute__((aligned(128)));
+unsigned int word_values_masked[MATRIX_SIZE] __attribute__((aligned(128)));
/* declare the arrays of predicates */
-unsigned short half_predicates[MATRIX_SIZE];
-unsigned int word_predicates[MATRIX_SIZE];
+unsigned short half_predicates[MATRIX_SIZE] __attribute__((aligned(128)));
+unsigned int word_predicates[MATRIX_SIZE] __attribute__((aligned(128)));
-/* make this big enough for all the intrinsics */
+/* make this big enough for all the operations */
const size_t region_len = sizeof(vtcm);
/* optionally add sync instructions */
@@ -261,164 +220,201 @@ void create_offsets_values_preds_16_32(void)
}
}
-/* scatter the 16 bit elements using intrinsics */
+/* scatter the 16 bit elements using HVX */
void vector_scatter_16(void)
{
- /* copy the offsets and values to vectors */
- HVX_Vector offsets = *(HVX_Vector *)half_offsets;
- HVX_Vector values = *(HVX_Vector *)half_values;
-
- VSCATTER_16(&vtcm.vscatter16, region_len, offsets, values);
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "v1 = vmem(%3 + #0)\n\t"
+ "vscatter(%0, m0, v0.h).h = v1\n\t"
+ : : "r"(vtcm.vscatter16), "r"(region_len),
+ "r"(half_offsets), "r"(half_values)
+ : "m0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter16);
}
-/* scatter-accumulate the 16 bit elements using intrinsics */
+/* scatter-accumulate the 16 bit elements using HVX */
void vector_scatter_16_acc(void)
{
- /* copy the offsets and values to vectors */
- HVX_Vector offsets = *(HVX_Vector *)half_offsets;
- HVX_Vector values = *(HVX_Vector *)half_values_acc;
-
- VSCATTER_16_ACC(&vtcm.vscatter16, region_len, offsets, values);
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "v1 = vmem(%3 + #0)\n\t"
+ "vscatter(%0, m0, v0.h).h += v1\n\t"
+ : : "r"(vtcm.vscatter16), "r"(region_len),
+ "r"(half_offsets), "r"(half_values_acc)
+ : "m0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter16);
}
-/* scatter the 16 bit elements using intrinsics */
+/* masked scatter the 16 bit elements using HVX */
void vector_scatter_16_masked(void)
{
- /* copy the offsets and values to vectors */
- HVX_Vector offsets = *(HVX_Vector *)half_offsets;
- HVX_Vector values = *(HVX_Vector *)half_values_masked;
- HVX_Vector pred_reg = *(HVX_Vector *)half_predicates;
- HVX_VectorPred preds = VAND_VAL(pred_reg, ~0);
-
- VSCATTER_16_MASKED(preds, &vtcm.vscatter16, region_len, offsets, values);
+ asm ("r1 = #-1\n\t"
+ "v0 = vmem(%0 + #0)\n\t"
+ "q0 = vand(v0, r1)\n\t"
+ "m0 = %2\n\t"
+ "v0 = vmem(%3 + #0)\n\t"
+ "v1 = vmem(%4 + #0)\n\t"
+ "if (q0) vscatter(%1, m0, v0.h).h = v1\n\t"
+ : : "r"(half_predicates), "r"(vtcm.vscatter16), "r"(region_len),
+ "r"(half_offsets), "r"(half_values_masked)
+ : "r1", "q0", "m0", "q0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter16);
}
-/* scatter the 32 bit elements using intrinsics */
+/* scatter the 32 bit elements using HVX */
void vector_scatter_32(void)
{
- /* copy the offsets and values to vectors */
- HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
- HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
- HVX_Vector valueslo = *(HVX_Vector *)word_values;
- HVX_Vector valueshi = *(HVX_Vector *)&word_values[MATRIX_SIZE / 2];
-
- VSCATTER_32(&vtcm.vscatter32, region_len, offsetslo, valueslo);
- VSCATTER_32(&vtcm.vscatter32, region_len, offsetshi, valueshi);
+ HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
+ HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
+ HVX_Vector *valueslo = (HVX_Vector *)word_values;
+ HVX_Vector *valueshi = (HVX_Vector *)&word_values[MATRIX_SIZE / 2];
+
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "v1 = vmem(%3 + #0)\n\t"
+ "vscatter(%0, m0, v0.w).w = v1\n\t"
+ : : "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetslo), "r"(valueslo)
+ : "m0", "v0", "v1", "memory");
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "v1 = vmem(%3 + #0)\n\t"
+ "vscatter(%0, m0, v0.w).w = v1\n\t"
+ : : "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetshi), "r"(valueshi)
+ : "m0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter32);
}
-/* scatter-acc the 32 bit elements using intrinsics */
+/* scatter-accumulate the 32 bit elements using HVX */
void vector_scatter_32_acc(void)
{
- /* copy the offsets and values to vectors */
- HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
- HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
- HVX_Vector valueslo = *(HVX_Vector *)word_values_acc;
- HVX_Vector valueshi = *(HVX_Vector *)&word_values_acc[MATRIX_SIZE / 2];
-
- VSCATTER_32_ACC(&vtcm.vscatter32, region_len, offsetslo, valueslo);
- VSCATTER_32_ACC(&vtcm.vscatter32, region_len, offsetshi, valueshi);
+ HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
+ HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
+ HVX_Vector *valueslo = (HVX_Vector *)word_values_acc;
+ HVX_Vector *valueshi = (HVX_Vector *)&word_values_acc[MATRIX_SIZE / 2];
+
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "v1 = vmem(%3 + #0)\n\t"
+ "vscatter(%0, m0, v0.w).w += v1\n\t"
+ : : "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetslo), "r"(valueslo)
+ : "m0", "v0", "v1", "memory");
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "v1 = vmem(%3 + #0)\n\t"
+ "vscatter(%0, m0, v0.w).w += v1\n\t"
+ : : "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetshi), "r"(valueshi)
+ : "m0", "v0", "v1", "memory");
sync_scatter(vtcm.vscatter32);
}
-/* scatter the 32 bit elements using intrinsics */
+/* masked scatter the 32 bit elements using HVX */
void vector_scatter_32_masked(void)
{
- /* copy the offsets and values to vectors */
- HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
- HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
- HVX_Vector valueslo = *(HVX_Vector *)word_values_masked;
- HVX_Vector valueshi = *(HVX_Vector *)&word_values_masked[MATRIX_SIZE / 2];
- HVX_Vector pred_reglo = *(HVX_Vector *)word_predicates;
- HVX_Vector pred_reghi = *(HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
- HVX_VectorPred predslo = VAND_VAL(pred_reglo, ~0);
- HVX_VectorPred predshi = VAND_VAL(pred_reghi, ~0);
-
- VSCATTER_32_MASKED(predslo, &vtcm.vscatter32, region_len, offsetslo,
- valueslo);
- VSCATTER_32_MASKED(predshi, &vtcm.vscatter32, region_len, offsetshi,
- valueshi);
+ HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
+ HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
+ HVX_Vector *valueslo = (HVX_Vector *)word_values_masked;
+ HVX_Vector *valueshi = (HVX_Vector *)&word_values_masked[MATRIX_SIZE / 2];
+ HVX_Vector *predslo = (HVX_Vector *)word_predicates;
+ HVX_Vector *predshi = (HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
+
+ asm ("r1 = #-1\n\t"
+ "v0 = vmem(%0 + #0)\n\t"
+ "q0 = vand(v0, r1)\n\t"
+ "m0 = %2\n\t"
+ "v0 = vmem(%3 + #0)\n\t"
+ "v1 = vmem(%4 + #0)\n\t"
+ "if (q0) vscatter(%1, m0, v0.w).w = v1\n\t"
+ : : "r"(predslo), "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetslo), "r"(valueslo)
+ : "r1", "q0", "m0", "q0", "v0", "v1", "memory");
+ asm ("r1 = #-1\n\t"
+ "v0 = vmem(%0 + #0)\n\t"
+ "q0 = vand(v0, r1)\n\t"
+ "m0 = %2\n\t"
+ "v0 = vmem(%3 + #0)\n\t"
+ "v1 = vmem(%4 + #0)\n\t"
+ "if (q0) vscatter(%1, m0, v0.w).w = v1\n\t"
+ : : "r"(predshi), "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetshi), "r"(valueshi)
+ : "r1", "q0", "m0", "q0", "v0", "v1", "memory");
- sync_scatter(vtcm.vscatter16);
+ sync_scatter(vtcm.vscatter32);
}
-/* scatter the 16 bit elements with 32 bit offsets using intrinsics */
+/* scatter the 16 bit elements with 32 bit offsets using HVX */
void vector_scatter_16_32(void)
{
- HVX_VectorPair offsets;
- HVX_Vector values;
-
- /* get the word offsets in a vector pair */
- offsets = *(HVX_VectorPair *)word_offsets;
-
- /* these values need to be shuffled for the scatter */
- values = *(HVX_Vector *)half_values;
- values = VSHUFF_H(values);
-
- VSCATTER_16_32(&vtcm.vscatter16_32, region_len, offsets, values);
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "v1 = vmem(%2 + #1)\n\t"
+ "v2 = vmem(%3 + #0)\n\t"
+ "v2.h = vshuff(v2.h)\n\t" /* shuffle the values for the scatter */
+ "vscatter(%0, m0, v1:0.w).h = v2\n\t"
+ : : "r"(vtcm.vscatter16_32), "r"(region_len),
+ "r"(word_offsets), "r"(half_values)
+ : "m0", "v0", "v1", "v2", "memory");
sync_scatter(vtcm.vscatter16_32);
}
-/* scatter-acc the 16 bit elements with 32 bit offsets using intrinsics */
+/* scatter-accumulate the 16 bit elements with 32 bit offsets using HVX */
void vector_scatter_16_32_acc(void)
{
- HVX_VectorPair offsets;
- HVX_Vector values;
-
- /* get the word offsets in a vector pair */
- offsets = *(HVX_VectorPair *)word_offsets;
-
- /* these values need to be shuffled for the scatter */
- values = *(HVX_Vector *)half_values_acc;
- values = VSHUFF_H(values);
-
- VSCATTER_16_32_ACC(&vtcm.vscatter16_32, region_len, offsets, values);
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "v1 = vmem(%2 + #1)\n\t"
+ "v2 = vmem(%3 + #0)\n\t" \
+ "v2.h = vshuff(v2.h)\n\t" /* shuffle the values for the scatter */
+ "vscatter(%0, m0, v1:0.w).h += v2\n\t"
+ : : "r"(vtcm.vscatter16_32), "r"(region_len),
+ "r"(word_offsets), "r"(half_values_acc)
+ : "m0", "v0", "v1", "v2", "memory");
sync_scatter(vtcm.vscatter16_32);
}
-/* masked scatter the 16 bit elements with 32 bit offsets using intrinsics */
+/* masked scatter the 16 bit elements with 32 bit offsets using HVX */
void vector_scatter_16_32_masked(void)
{
- HVX_VectorPair offsets;
- HVX_Vector values;
- HVX_Vector pred_reg;
-
- /* get the word offsets in a vector pair */
- offsets = *(HVX_VectorPair *)word_offsets;
-
- /* these values need to be shuffled for the scatter */
- values = *(HVX_Vector *)half_values_masked;
- values = VSHUFF_H(values);
-
- pred_reg = *(HVX_Vector *)half_predicates;
- pred_reg = VSHUFF_H(pred_reg);
- HVX_VectorPred preds = VAND_VAL(pred_reg, ~0);
-
- VSCATTER_16_32_MASKED(preds, &vtcm.vscatter16_32, region_len, offsets,
- values);
+ asm ("r1 = #-1\n\t"
+ "v0 = vmem(%0 + #0)\n\t"
+ "v0.h = vshuff(v0.h)\n\t" /* shuffle the predicates */
+ "q0 = vand(v0, r1)\n\t"
+ "m0 = %2\n\t"
+ "v0 = vmem(%3 + #0)\n\t"
+ "v1 = vmem(%3 + #1)\n\t"
+ "v2 = vmem(%4 + #0)\n\t" \
+ "v2.h = vshuff(v2.h)\n\t" /* shuffle the values for the scatter */
+ "if (q0) vscatter(%1, m0, v1:0.w).h = v2\n\t"
+ : : "r"(half_predicates), "r"(vtcm.vscatter16_32), "r"(region_len),
+ "r"(word_offsets), "r"(half_values_masked)
+ : "r1", "q0", "m0", "v0", "v1", "v2", "memory");
sync_scatter(vtcm.vscatter16_32);
}
-/* gather the elements from the scatter16 buffer */
+/* gather the elements from the scatter16 buffer using HVX */
void vector_gather_16(void)
{
- HVX_Vector *vgather = (HVX_Vector *)&vtcm.vgather16;
- HVX_Vector offsets = *(HVX_Vector *)half_offsets;
-
- VGATHER_16(vgather, &vtcm.vscatter16, region_len, offsets);
-
- sync_gather(vgather);
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "{ vtmp.h = vgather(%0, m0, v0.h).h\n\t"
+ " vmem(%3 + #0) = vtmp.new }\n\t"
+ : : "r"(vtcm.vscatter16), "r"(region_len),
+ "r"(half_offsets), "r"(vtcm.vgather16)
+ : "m0", "v0", "memory");
+
+ sync_gather(vtcm.vgather16);
}
static unsigned short gather_16_masked_init(void)
@@ -427,31 +423,51 @@ static unsigned short gather_16_masked_init(void)
return letter | (letter << 8);
}
+/* masked gather the elements from the scatter16 buffer using HVX */
void vector_gather_16_masked(void)
{
- HVX_Vector *vgather = (HVX_Vector *)&vtcm.vgather16;
- HVX_Vector offsets = *(HVX_Vector *)half_offsets;
- HVX_Vector pred_reg = *(HVX_Vector *)half_predicates;
- HVX_VectorPred preds = VAND_VAL(pred_reg, ~0);
-
- *vgather = VSPLAT_H(gather_16_masked_init());
- VGATHER_16_MASKED(vgather, preds, &vtcm.vscatter16, region_len, offsets);
-
- sync_gather(vgather);
+ unsigned short init = gather_16_masked_init();
+
+ asm ("v0.h = vsplat(%5)\n\t"
+ "vmem(%4 + #0) = v0\n\t" /* initialize the write area */
+ "r1 = #-1\n\t"
+ "v0 = vmem(%0 + #0)\n\t"
+ "q0 = vand(v0, r1)\n\t"
+ "m0 = %2\n\t"
+ "v0 = vmem(%3 + #0)\n\t"
+ "{ if (q0) vtmp.h = vgather(%1, m0, v0.h).h\n\t"
+ " vmem(%4 + #0) = vtmp.new }\n\t"
+ : : "r"(half_predicates), "r"(vtcm.vscatter16), "r"(region_len),
+ "r"(half_offsets), "r"(vtcm.vgather16), "r"(init)
+ : "r1", "q0", "m0", "v0", "memory");
+
+ sync_gather(vtcm.vgather16);
}
-/* gather the elements from the scatter32 buffer */
+/* gather the elements from the scatter32 buffer using HVX */
void vector_gather_32(void)
{
- HVX_Vector *vgatherlo = (HVX_Vector *)&vtcm.vgather32;
- HVX_Vector *vgatherhi =
- (HVX_Vector *)((int)&vtcm.vgather32 + (MATRIX_SIZE * 2));
- HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
- HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
-
- VGATHER_32(vgatherlo, &vtcm.vscatter32, region_len, offsetslo);
- VGATHER_32(vgatherhi, &vtcm.vscatter32, region_len, offsetshi);
+ HVX_Vector *vgatherlo = (HVX_Vector *)vtcm.vgather32;
+ HVX_Vector *vgatherhi = (HVX_Vector *)&vtcm.vgather32[MATRIX_SIZE / 2];
+ HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
+ HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
+
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "{ vtmp.w = vgather(%0, m0, v0.w).w\n\t"
+ " vmem(%3 + #0) = vtmp.new }\n\t"
+ : : "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetslo), "r"(vgatherlo)
+ : "m0", "v0", "memory");
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "{ vtmp.w = vgather(%0, m0, v0.w).w\n\t"
+ " vmem(%3 + #0) = vtmp.new }\n\t"
+ : : "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetshi), "r"(vgatherhi)
+ : "m0", "v0", "memory");
+ sync_gather(vgatherlo);
sync_gather(vgatherhi);
}
@@ -461,79 +477,88 @@ static unsigned int gather_32_masked_init(void)
return letter | (letter << 8) | (letter << 16) | (letter << 24);
}
+/* masked gather the elements from the scatter32 buffer using HVX */
void vector_gather_32_masked(void)
{
- HVX_Vector *vgatherlo = (HVX_Vector *)&vtcm.vgather32;
- HVX_Vector *vgatherhi =
- (HVX_Vector *)((int)&vtcm.vgather32 + (MATRIX_SIZE * 2));
- HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
- HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
- HVX_Vector pred_reglo = *(HVX_Vector *)word_predicates;
- HVX_VectorPred predslo = VAND_VAL(pred_reglo, ~0);
- HVX_Vector pred_reghi = *(HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
- HVX_VectorPred predshi = VAND_VAL(pred_reghi, ~0);
-
- *vgatherlo = VSPLAT_H(gather_32_masked_init());
- *vgatherhi = VSPLAT_H(gather_32_masked_init());
- VGATHER_32_MASKED(vgatherlo, predslo, &vtcm.vscatter32, region_len,
- offsetslo);
- VGATHER_32_MASKED(vgatherhi, predshi, &vtcm.vscatter32, region_len,
- offsetshi);
+ unsigned int init = gather_32_masked_init();
+ HVX_Vector *vgatherlo = (HVX_Vector *)vtcm.vgather32;
+ HVX_Vector *vgatherhi = (HVX_Vector *)&vtcm.vgather32[MATRIX_SIZE / 2];
+ HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
+ HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
+ HVX_Vector *predslo = (HVX_Vector *)word_predicates;
+ HVX_Vector *predshi = (HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
+
+ asm ("v0.h = vsplat(%5)\n\t"
+ "vmem(%4 + #0) = v0\n\t" /* initialize the write area */
+ "r1 = #-1\n\t"
+ "v0 = vmem(%0 + #0)\n\t"
+ "q0 = vand(v0, r1)\n\t"
+ "m0 = %2\n\t"
+ "v0 = vmem(%3 + #0)\n\t"
+ "{ if (q0) vtmp.w = vgather(%1, m0, v0.w).w\n\t"
+ " vmem(%4 + #0) = vtmp.new }\n\t"
+ : : "r"(predslo), "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetslo), "r"(vgatherlo), "r"(init)
+ : "r1", "q0", "m0", "v0", "memory");
+ asm ("v0.h = vsplat(%5)\n\t"
+ "vmem(%4 + #0) = v0\n\t" /* initialize the write area */
+ "r1 = #-1\n\t"
+ "v0 = vmem(%0 + #0)\n\t"
+ "q0 = vand(v0, r1)\n\t"
+ "m0 = %2\n\t"
+ "v0 = vmem(%3 + #0)\n\t"
+ "{ if (q0) vtmp.w = vgather(%1, m0, v0.w).w\n\t"
+ " vmem(%4 + #0) = vtmp.new }\n\t"
+ : : "r"(predshi), "r"(vtcm.vscatter32), "r"(region_len),
+ "r"(offsetshi), "r"(vgatherhi), "r"(init)
+ : "r1", "q0", "m0", "v0", "memory");
sync_gather(vgatherlo);
sync_gather(vgatherhi);
}
-/* gather the elements from the scatter16_32 buffer */
+/* gather the elements from the scatter16_32 buffer using HVX */
void vector_gather_16_32(void)
{
- HVX_Vector *vgather;
- HVX_VectorPair offsets;
- HVX_Vector values;
-
- /* get the vtcm address to gather from */
- vgather = (HVX_Vector *)&vtcm.vgather16_32;
-
- /* get the word offsets in a vector pair */
- offsets = *(HVX_VectorPair *)word_offsets;
-
- VGATHER_16_32(vgather, &vtcm.vscatter16_32, region_len, offsets);
-
- /* deal the elements to get the order back */
- values = *(HVX_Vector *)vgather;
- values = VDEAL_H(values);
-
- /* write it back to vtcm address */
- *(HVX_Vector *)vgather = values;
+ asm ("m0 = %1\n\t"
+ "v0 = vmem(%2 + #0)\n\t"
+ "v1 = vmem(%2 + #1)\n\t"
+ "{ vtmp.h = vgather(%0, m0, v1:0.w).h\n\t"
+ " vmem(%3 + #0) = vtmp.new }\n\t"
+ "v0 = vmem(%3 + #0)\n\t"
+ "v0.h = vdeal(v0.h)\n\t" /* deal the elements to get the order back */
+ "vmem(%3 + #0) = v0\n\t"
+ : : "r"(vtcm.vscatter16_32), "r"(region_len),
+ "r"(word_offsets), "r"(vtcm.vgather16_32)
+ : "m0", "v0", "v1", "memory");
+
+ sync_gather(vtcm.vgather16_32);
}
+/* masked gather the elements from the scatter16_32 buffer using HVX */
void vector_gather_16_32_masked(void)
{
- HVX_Vector *vgather;
- HVX_VectorPair offsets;
- HVX_Vector pred_reg;
- HVX_VectorPred preds;
- HVX_Vector values;
-
- /* get the vtcm address to gather from */
- vgather = (HVX_Vector *)&vtcm.vgather16_32;
-
- /* get the word offsets in a vector pair */
- offsets = *(HVX_VectorPair *)word_offsets;
- pred_reg = *(HVX_Vector *)half_predicates;
- pred_reg = VSHUFF_H(pred_reg);
- preds = VAND_VAL(pred_reg, ~0);
-
- *vgather = VSPLAT_H(gather_16_masked_init());
- VGATHER_16_32_MASKED(vgather, preds, &vtcm.vscatter16_32, region_len,
- offsets);
-
- /* deal the elements to get the order back */
- values = *(HVX_Vector *)vgather;
- values = VDEAL_H(values);
-
- /* write it back to vtcm address */
- *(HVX_Vector *)vgather = values;
+ unsigned short init = gather_16_masked_init();
+
+ asm ("v0.h = vsplat(%5)\n\t"
+ "vmem(%4 + #0) = v0\n\t" /* initialize the write area */
+ "r1 = #-1\n\t"
+ "v0 = vmem(%0 + #0)\n\t"
+ "v0.h = vshuff(v0.h)\n\t" /* shuffle the predicates */
+ "q0 = vand(v0, r1)\n\t"
+ "m0 = %2\n\t"
+ "v0 = vmem(%3 + #0)\n\t"
+ "v1 = vmem(%3 + #1)\n\t"
+ "{ if (q0) vtmp.h = vgather(%1, m0, v1:0.w).h\n\t"
+ " vmem(%4 + #0) = vtmp.new }\n\t"
+ "v0 = vmem(%4 + #0)\n\t"
+ "v0.h = vdeal(v0.h)\n\t" /* deal the elements to get the order back */
+ "vmem(%4 + #0) = v0\n\t"
+ : : "r"(half_predicates), "r"(vtcm.vscatter16_32), "r"(region_len),
+ "r"(word_offsets), "r"(vtcm.vgather16_32), "r"(init)
+ : "r1", "q0", "m0", "v0", "v1", "memory");
+
+ sync_gather(vtcm.vgather16_32);
}
static void check_buffer(const char *name, void *c, void *r, size_t size)
@@ -579,6 +604,7 @@ void scalar_scatter_16_acc(unsigned short *vscatter16)
}
}
+/* scatter-accumulate the 16 bit elements using C */
void check_scatter_16_acc()
{
memset(vscatter16_ref, FILL_CHAR,
@@ -589,7 +615,7 @@ void check_scatter_16_acc()
SCATTER_BUFFER_SIZE * sizeof(unsigned short));
}
-/* scatter the 16 bit elements using C */
+/* masked scatter the 16 bit elements using C */
void scalar_scatter_16_masked(unsigned short *vscatter16)
{
for (int i = 0; i < MATRIX_SIZE; i++) {
@@ -628,7 +654,7 @@ void check_scatter_32()
SCATTER_BUFFER_SIZE * sizeof(unsigned int));
}
-/* scatter the 32 bit elements using C */
+/* scatter-accumulate the 32 bit elements using C */
void scalar_scatter_32_acc(unsigned int *vscatter32)
{
for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -646,7 +672,7 @@ void check_scatter_32_acc()
SCATTER_BUFFER_SIZE * sizeof(unsigned int));
}
-/* scatter the 32 bit elements using C */
+/* masked scatter the 32 bit elements using C */
void scalar_scatter_32_masked(unsigned int *vscatter32)
{
for (int i = 0; i < MATRIX_SIZE; i++) {
@@ -667,7 +693,7 @@ void check_scatter_32_masked()
SCATTER_BUFFER_SIZE * sizeof(unsigned int));
}
-/* scatter the 32 bit elements using C */
+/* scatter the 16 bit elements with 32 bit offsets using C */
void scalar_scatter_16_32(unsigned short *vscatter16_32)
{
for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -684,7 +710,7 @@ void check_scatter_16_32()
SCATTER_BUFFER_SIZE * sizeof(unsigned short));
}
-/* scatter the 32 bit elements using C */
+/* scatter-accumulate the 16 bit elements with 32 bit offsets using C */
void scalar_scatter_16_32_acc(unsigned short *vscatter16_32)
{
for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -702,6 +728,7 @@ void check_scatter_16_32_acc()
SCATTER_BUFFER_SIZE * sizeof(unsigned short));
}
+/* masked scatter the 16 bit elements with 32 bit offsets using C */
void scalar_scatter_16_32_masked(unsigned short *vscatter16_32)
{
for (int i = 0; i < MATRIX_SIZE; i++) {
@@ -738,6 +765,7 @@ void check_gather_16()
MATRIX_SIZE * sizeof(unsigned short));
}
+/* masked gather the elements from the scatter buffer using C */
void scalar_gather_16_masked(unsigned short *vgather16)
{
for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -756,7 +784,7 @@ void check_gather_16_masked()
MATRIX_SIZE * sizeof(unsigned short));
}
-/* gather the elements from the scatter buffer using C */
+/* gather the elements from the scatter32 buffer using C */
void scalar_gather_32(unsigned int *vgather32)
{
for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -772,6 +800,7 @@ void check_gather_32(void)
MATRIX_SIZE * sizeof(unsigned int));
}
+/* masked gather the elements from the scatter32 buffer using C */
void scalar_gather_32_masked(unsigned int *vgather32)
{
for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -781,7 +810,6 @@ void scalar_gather_32_masked(unsigned int *vgather32)
}
}
-
void check_gather_32_masked(void)
{
memset(vgather32_ref, gather_32_masked_init(),
@@ -791,7 +819,7 @@ void check_gather_32_masked(void)
vgather32_ref, MATRIX_SIZE * sizeof(unsigned int));
}
-/* gather the elements from the scatter buffer using C */
+/* gather the elements from the scatter16_32 buffer using C */
void scalar_gather_16_32(unsigned short *vgather16_32)
{
for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -807,6 +835,7 @@ void check_gather_16_32(void)
MATRIX_SIZE * sizeof(unsigned short));
}
+/* masked gather the elements from the scatter16_32 buffer using C */
void scalar_gather_16_32_masked(unsigned short *vgather16_32)
{
for (int i = 0; i < MATRIX_SIZE; ++i) {
Replace __builtin_* with inline assembly The __builtin's are subject to change with different compiler releases, so might break Mark arrays as aligned when accessed as HVX vectors Clean up comments Signed-off-by: Taylor Simpson <tsimpson@quicinc.com> --- tests/tcg/hexagon/scatter_gather.c | 513 +++++++++++++++-------------- 1 file changed, 271 insertions(+), 242 deletions(-)